University of New Mexico

UNM Digital Repository

Mechanical Engineering ETDs Engineering ETDs

Fall 12-15-2017

Co-simulation Framework for Power Distribution

System Analysis with Humans in the Loop
Victor H. Ayon
University of New Mexico - Main Campus

Follow this and additional works at: https://digitalrepository.unm.edu/me_etds

Part of the Mechanical Engineering Commons, and the Power and Energy Commons

Recommended Citation
Ayon, Victor H.. "Co-simulation Framework for Power Distribution System Analysis with Humans in the Loop." (2017).
https://digitalrepository.unm.edu/me_etds/141

This Thesis is brought to you for free and open access by the Engineering ETDs at UNM Digital Repository. It has been accepted for inclusion in
Mechanical Engineering ETDs by an authorized administrator of UNM Digital Repository. For more information, please contact [email protected].
Candidate

Department

This thesis is approved, and it is acceptable in quality and form for publication:

Approved by the Thesis Committee:

, Chairperson
 Co-simulation Framework for Power
Distribution System Analysis with
Humans in the Loop

by

Victor H. Ayon

A.S., Central Arizona College, 2010

B.S., Mechanical Engineering, University of New Mexico, 2014

THESIS

Submitted in Partial Fulfillment of the

Requirements for the Degree of

Master of Science
Mechanical Engineering

The University of New Mexico

Albuquerque, New Mexico

December, 2017
Dedication

To my beautiful wife, Saydra, for her unconditional love, support, and

encouragement. She motivates me and makes me a better person.
She is my rock, my best friend, and my soulmate.

“You never fail until you stop trying.”

– Albert Einstein

iii
Acknowledgments

I would like to thank my advisor, Professor Andrea Mammoli, for his support and
guidance throughout this journey. I would also like to thank my friend and colleague,
Matthew Robinson, who was always there to lend a helping hand with support and
encouragement.
 Co-simulation Framework for Power
Distribution System Analysis with
Humans in the Loop

by

Victor H. Ayon

A.S., Central Arizona College, 2010

B.S., Mechanical Engineering, University of New Mexico, 2014

M.S., Mechanical Engineering, University of New Mexico, 2017

Abstract

Recent developments in the area of power generation have led to an increased pen-
etration of storage and distributed energy resources (DER) in power distribution
systems. As a result, new and enhanced energy management systems will be nec-
essary as the deployment of DERs, as well as the need to control loads, continues
to increase in the coming years. Advanced management systems are especially im-
portant to achieve resilient power delivery during emergency situations. During a
blackout for example, a section of a distribution feeder could island to operate as a
microgrid to ensure critical services such as water, food, and medical care remain
online. The development of such management systems will also require the ability
to integrate human behavior models with power flow simulators, as technological
advances lead to more customer-owned devices with the potential to be used to bal-
ance the power flow of a feeder. This thesis describes a co-simulation framework
that combines a bottom-up residential load generator, a load aggregator for real-
time (RT) residential demand response (DR), a utility-scale battery controller, and

v
the GridLAB-D distribution system simulator. The behavior of a distribution feeder
model is analyzed under di↵erent scenarios. The model is based on an existing feeder
located in Los Alamos, NM, which serves residential customers and a set of critical
loads including a hospital, a supermarket, and a water treatment plant. This feeder
also hosts a utility-scale solar array and battery storage that are used to operate
the feeder as a microgrid. Additionally, a real-time simulation is described in which
real-time residential demand response is implemented on a virtual community using
the load generator and aggregator. The simulation is part of a project invested in
the development of a modern microgrid control system employing a virtual power
plant approach and a model predictive controller to optimize the use of resources
within a distribution feeder. The capabilities to study power distribution systems
with humans in the loop using these platforms is showcased here. Furthermore, their
potential as instrumental tools in the development and design of new technology
essential to improve grid resilience is also discussed.

vi
Contents

List of Figures x

List of Tables xiii

Glossary xiv

1 Introduction 1

1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Description of co-simulation framework . . . . . . . . . . . . . . . . . 4

1.3 Description of real-time simulation . . . . . . . . . . . . . . . . . . . 6

2 The Models 9

2.1 Circuit 16 feeder model . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1.1 From WindMil to GridLAB-D . . . . . . . . . . . . . . . . . . 9

2.1.2 Model validation . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.1.3 Critical loads . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

vii
Contents

2.2 Residential load generator & aggregator . . . . . . . . . . . . . . . . 15

2.2.1 Residential load generator . . . . . . . . . . . . . . . . . . . . 15

2.2.2 Residential load aggregator . . . . . . . . . . . . . . . . . . . 20

2.3 Battery controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Simulation Setup 25

3.1 GridLAB-D co-simulation . . . . . . . . . . . . . . . . . . . . . . . . 25

3.1.1 Matlab as co-simulation interface . . . . . . . . . . . . . . . . 25

3.1.2 FNCS as co-simulation interface . . . . . . . . . . . . . . . . . 28

3.2 Real-time simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4 Simulation Results 37

4.1 GridLAB-D co-simulation . . . . . . . . . . . . . . . . . . . . . . . . 37

4.1.1 Matlab interface . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.1.2 FNCS interface . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.2 Real-time simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

5 Discussion & Conclusions 61

5.1 GridLAB-D co-simulation . . . . . . . . . . . . . . . . . . . . . . . . 61

5.2 Real-time simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

viii
Contents

Appendices 67

A Matlab interface models & code 68

B FNCS interface models & code 69

C Real-time simulation models & code 70

References 71

ix
List of Figures

1 Circuit 16 on a map of Los Alamos, NM . . . . . . . . . . . . . . . . 10

2 Configuration of the Los Alamos Circuit 16 feeder . . . . . . . . . . 11

3 Circuit 16 voltage profiles . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Water pump stochastic characterization . . . . . . . . . . . . . . . . 13

5 Power demand of critical loads on Circuit 16 . . . . . . . . . . . . . 14

6 Load profiles at one-minute and hourly resolutions . . . . . . . . . . 15

7 Probabilistic characterization of the use of an electric range . . . . . 17

8 Typical electric loads by appliance for a single meter . . . . . . . . . 18

9 Total feeder load resulting from 1000 houses . . . . . . . . . . . . . 19

10 Real-time residential DR example . . . . . . . . . . . . . . . . . . . 21

11 Battery control flowchart . . . . . . . . . . . . . . . . . . . . . . . . 23

12 Example of battery control performance . . . . . . . . . . . . . . . . 24

13 Co-simulation data flow diagram with Matlab interface . . . . . . . 26

14 Co-simulation data flow diagram with FNCS interface . . . . . . . . 30

x
List of Figures

15 Raspberry Pi 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

16 Real-time simulation setup and model interaction with the database 32

17 Real-time simulation input & output data flow . . . . . . . . . . . . 34

18 Artificial outside temperature represented by a sine wave . . . . . . 38

19 Loads on Circuit 16, uncontrolled conditions using Matlab interface 39

20 Performance of battery control on Circuit 16 . . . . . . . . . . . . . 41

21 Real-time DR of 50 domestic hot water heaters . . . . . . . . . . . . 43

22 Temperature recorded on a mild summer day in Phoenix, AZ . . . . 44

23 Loads on Circuit 16, uncontrolled conditions using FNCS interface . 45

24 Voltage behavior on Circuit 16 during high DER penetration . . . . 46

25 Real-time DR performance controlling 1600 AC units . . . . . . . . 48

26 Circuit 16 load history, baseline and DR . . . . . . . . . . . . . . . . 49

27 Circuit 16 voltage profiles during DR rebound, baseline and DR . . 50

28 States of the AC units participating in demand response . . . . . . . 51

29 Barranca Mesa voltage history, baseline and DR . . . . . . . . . . . 52

30 Circuit 16 load history with DR & battery control . . . . . . . . . . 53

31 Circuit 16 voltage during DR rebound, DR vs DR & battery control 55

32 Barranca Mesa voltage history, DR vs DR & battery control . . . . 56

33 Performance of real-time residential demand response during a test

period of 22 hours . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

xi
List of Figures

34 Example of control signals and response of residential demand re-

sponse control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

35 Simulated residential load used in the MRI real-time simulation . . . 59

xii
List of Tables

1 Sample TCL status data received by the aggregator . . . . . . . . . 33

2 Sample data transferred to the database by the load generator . . . 35

xiii
Glossary

DER Distributed energy resources.

RT Real-time.

DR Demand response.

PV Photovoltaics.

kW kilo-Watt.

MW Mega-Watt.

kWh kilo-Watt-hour.

FNCS The Framework for Networked Co-Simulation software.

PNNL Pacific Northwest National Laboratory.

MRI Mitsubishi Research Institute.

UNM University of New Mexico.

NEDO Japan’s New Energy and Industrial Technology Development Orga-

nization.

HIL Hardware-in-the-Loop.

xiv
Glossary

EV Electric Vehicle.

VPP Virtual Power Plant.

MPC Model Predictive Controller.

BESS Battery energy storage system.

LADPU Los Alamos Department of Public Utilities.

GIS Geographic Information System.

PDF Probability density function.

HVAC Heating, ventilation, and air conditioning.

MS Thermal mass of the conditioned space by a thermostat.

T (t) Temperature of the conditioned space by a thermostat.

Ta Ambient temperature.

t Time.

QL Thermal losses from structure.

QR Thermal gain from air conditioner or heat pump.

K1 Function of the building’s thermal insulation.

COP Coefficient of performance.

PAC Power of compressor.

⇤ State function that indicates whether an air conditioner or heat

pump is on or o↵.

TL Lower deadband for temperature control.

xv
Glossary

TU Upper deadband for temperature control.

Tamb Outside temperature.

QoS Quality of service.

TCL Thermostatically controllable loads.

AC Air conditioner.

POA Plane of array.

scp Secure Copy.

SOC State of charge.

xvi
Chapter 1

Introduction

1.1 Overview

The modernization of the power grid that is currently taking place all over the world
has triggered a shift from a century-old paradigm that focuses in centralized power
generation and long distance transmission, to a modernized approach that concen-
trates in localized and distributed power generation. The power grid is composed of
three major parts; generation, transmission, and distribution & consumption. Tra-
ditionally, power is generated exclusively at power plants and sent to consumers
through transmission and distribution lines. However, the power grid is currently
undergoing a technological revolution in which power is generated at distributed lo-
cations and near to the consumer. Microgrid systems, capable of operating without
any support from the main grid have emerged as a result of an increased penetration
of distributed energy resources (DER) in distribution level systems. These include
utility-scale photovoltaics (PV) and battery energy storage systems capable of deliv-
ering 500kW or more, as well as customer-owned rooftop PV, electric vehicles, and
other small-scale storage systems including demand response (DR). In the case of

1
Chapter 1. Introduction

a major disruptive event, such as a hurricane disabling large sections of the power
grid, the ability to divide a distribution network into smaller microgrids suggests
that local resources could be used to support critical services such as food, water,
and medical care, while damaged infrastructure is repaired. For example, the devas-
tating conditions experienced in Puerto Rico after hurricane Maria destroyed most
of the island’s transmission infrastructure in late September, 2017, could have been
significantly reduced by the use of distribution-level microgrids. Water distribution
infrastructure, hospitals, and supermarkets could have been given priority to ensure
residents had the means to remain hydrated and nourished with the required medical
care. The ability to provide critical services during disruptive events indicates that
the resilience of the power grid can be improved by considerably reducing the mag-
nitude and duration of these events with proper planning and microgrid resources.
Furthermore, DER devices used in microgrids continue to become more economically
attractive due to advances in technology in addition to incentives, rebates, and tax
credit programs provided by governments. This decrease in cost suggests that the
deployment of DERs will continue to increase in the future, and with it, the need for
more advanced management systems capable to manage operations from a few kW
to several MW.
A well-coordinated deployment of DERs coupled with the development and imple-
mentation of control systems could lead to multiple advantages. For instance, invert-
ers could be used to provide local voltage support by adjusting their power factor
in order to reduce stress on devices such as load tap changers, potentially increasing
their lifetime. Peak loads on substation transformers could be reduced by coordi-
nating the action of storage devices that would lead to deferral of upgrades and
replacements. Also, the ability of a distribution feeder to rapidly respond to load ad-
justment requests could provide frequency support to the transmission grid, thereby
reducing the likelihood of blackouts and brownouts. Distribution feeders with the
ability to function as microgrids could island sections of a feeder to provide service

2
Chapter 1. Introduction

to critical loads in the case that a blackout does occur. Also, the load on a distribu-
tion feeder could be managed by implementing more sophisticated demand response
programs, in which financial incentives are o↵ered to customers willing to reducing
or shift their electricity usage during events of high power consumption or other grid
constraints.
A well-designed control framework could enable new business models that include
and engage customers in the management of residential loads. However, despite cur-
rent standards and guidelines for the implementation of such framework, it is widely
acknowledged that a full deployment is not yet feasible. Further research is needed in
the areas of communication, controls, optimization, architecture, and human behav-
ior modeling. While advances have been made in the management and optimization
of DERs in the recent years, human behavior research is limited due to the difficulty
to test new techniques in realistic situations using real customers and their homes.
Especially, studies that examine models of the electric infrastructure in combination
with models of how human use electricity are close to non-existent.
This thesis describes the work behind two publications by the author in which a
tool to conduct research in the areas of DER management and human behavior
models is introduced. The simulation framework provides the ability to model elec-
tric distribution systems while considering the e↵ects of human power consumption
described by their interactions with household appliances [1, 2]. The framework inte-
grates the GridLAB-D agent-based distribution system simulation platform [3] with
a human-behavior based residential load synthesis framework within a co-simulation
environment. The tool is used to demonstrate how behavior models can be used to
study the e↵ects of the human factor in distribution level load management. Real-
time (RT) demand response implemented with a residential load aggregator and a
utility-scale battery controller are also demonstrated in this work. The real-time
demand response technique described here employs device-level demand response in
which customers grant utilities the ability to remotely adjust their energy consump-

3
Chapter 1. Introduction

tion by modifying the settings on internet-enabled devices such as smart thermostats

and other similar control switches. However, in traditional demand response the de-
vices are set on a static schedule with settings intended to lower consumption over
the duration of a peak event. Here, demand response is used to control the activity
of a feet of smart devices in real time to shape their aggregated power consumption
in respond to some external control signal. In a di↵erent application, the residential
load generator was incorporated into a real-time microgrid control system simulation
as a proxy for a 200-home residential community with real-time demand response
capabilities enabled by the residential load aggregator. The contributions by the
author related to the implementation of the models within the real-time simulation
are presented in this thesis as well.

1.2 Description of co-simulation framework

The use of DERs, and the use of electricity in general, is tightly related to how
humans consume electricity across all application levels. As the technology to gen-
erate power from various distributed resources continues to improve, its economical
appeal also continues to increase among customers in the utility, commercial, and
residential sectors. A well-coordinated deployment of DERs coupled to the develop-
ment of management systems will be crucial to maximize the benefits that these new
technologies can provide. However, new DER management systems must be tested
and analyzed to examine potential negative e↵ects that they might impose on the
grid. These test must be in such a way that avoids damaging expensive equipment
or risks the quality of service to customers. The strong relationship between electric
distribution systems and human behavior reveals the necessity for a simulation tool
to conduct research that considers the e↵ects of DER penetration on the electrical
infrastructure, while also considering the critical role that human power consump-

4
Chapter 1. Introduction

tion and behavior plays within these systems. Part of the main focus of this thesis is
a co-simulation framework that combines human behavior models with simulations
of power distribution systems with the purpose of analyzing how human activities
impact power management in distribution level systems. While models exist that
simulate both of these domains individually, a combined model did not exist be-
fore the co-simulation tool presented here was published in. [1, 2]. Additionally,
this framework is intended to serve as a tool in the development of new technology
that will make the power grid more efficient, reliable, but most importantly, more
resilient. Besides considering the e↵ects of human behavior on the grid, it makes
it possible to examine the performance of new control systems, and to analyze the
e↵ects imposed on electric infrastructure that would normally be neglected without a
power flow simulator. For example, it provides the ability to precisely locate sections
of a distribution feeder at risk of violating industry standards or experiencing faults,
and to test potential solutions based on the insight gained from the results.
The co-simulation framework uses a residential load generator to simulate residen-
tial power consumption based on statistical representations of human interactions
with household appliances. Residential loads are then passed to the GridLAB-D
distribution system environment where power flow calculations are performed. Data
transfer and time synchronization between co-simulators is accomplished using two
di↵erent methods. The first employs the GridLAB-D Matlab Link directive function
which uses a Matlab link control file to establish a connection to an external Matlab
instance [4], allowing GridLAB-D to execute specific Matlab commands before main
object and power flow calculations are performed during each time-step. The second
method is an enhanced version of the first, taking advantage of FNCS (Framework
for Networked Co-Simulation), a companion software package developed at Pacific
Northwest National Laboratory (PNNL) that allows GridLAB-D to interact with
other software as co-simulators [5]. For both methods, a typical scenario is exam-
ined first to confirm the co-simulations represent real distribution feeder behavior.

5
Chapter 1. Introduction

The scenario consists of a distribution feeder that hosts a utility-scale PV array and
battery system, a residential community, and critical loads comprised of a hospital,
supermarket, and water distribution infrastructure. Once the realistic behavior of
the co-simulations is established, they are used to examine scenarios that consider
the implementation of the residential load aggregator for demand response, and man-
agement of resources using the battery controller to manage the load on the feeder.

1.3 Description of real-time simulation

The increasing penetration of DERs has enabled power distribution systems to de-
velop the ability to island sections, even entire feeders, within their system for micro-
grid operation. Microgrids host enough local power generating resources to support
the power consumption of its customers, giving them the ability to disconnect from
the main grid and operate without its support. While many microgrid systems ex-
ist today all over the world, the development of microgrid control systems is not
mature. Most of the current controllers handle islanding requirements and provide
management of generating resources and load balancing for resiliency and back-up
purposes. However, the concept of the microgrid is changing to an advanced concept
that includes all the essential elements of a large-scale grid, such as the ability to
balance electrical demand with resources, maintain power quality, schedule the dis-
patch of resources, and preserve grid reliability in terms of adequacy and security.
Therefore, modern microgrids require controllers with the capability to perform these
functions [6].
A recent study conducted under a research project funded by the Mitsubishi Re-
search Institute (MRI) with collaborative e↵orts from the University of New Mexico
(UNM) and Sandia National Laboratories, investigates the application of a modern

6
Chapter 1. Introduction

microgrid control system in the Mesa del Sol community located in Albuquerque,
NM. Mesa del Sol is the home of multiple DERs such as a utility-scale PV and bat-
tery storage system, the commercial-scale NEDO Microgrid that consists of a PV
array, battery storage, fuel cell, and other DERs. The NEDO Microgrid is the legacy
of a NEDO Microgrid Demonstration Project carried out by Japan’s NEDO (New
Energy and Industrial Technology Development Organization), the US Department
of Energy, UNM, the local utility, several national laboratories, and many additional
entities. After its completion, the microgrid was transferred to UNM and is now op-
erated as a research testbed. The microgrid control system study considers real-life
DERs currently available in the Mesa del Sol area as Hardware-in-the-Loop (HIL)
components, as well as virtual models representing future resources such as residen-
tial demand response and electric vehicle (EV) charging stations.
In this case, the residential load generator was integrated into a real-time, hardware-
in-loop, simulation as a virtual proxy for the existing Mesa del Sol 200-home resi-
dential community. In addition, the use of the residential load aggregator enables
demand response capabilities for the virtual community. A virtual version of the
Mesa del Sol distribution feeder was also used within the Opal RT real-time simula-
tor to perform real-time, high resolution power flow simulation of the system. The
microgrid control system links real and virtual DERs and uses a virtual power plant
(VPP) approach, paired with a model predictive controller (MPC), for optimized
management of resources. The VPP generates stochastically optimized day-ahead
DER schedules based on load predictions and other parameters, while the MPC pro-
vides real-time dispatch of resources by deviating from the VPP schedules for an
optimized use of resources based on current conditions [7]. However, an in-depth
consideration of this microgrid control system falls beyond the scope of this thesis,
and therefore, only the implementation and results related to the residential load
generator and load aggregator will be discussed here. These include loading the
models into their system platforms, requirements for real-time execution, communi-

7
Chapter 1. Introduction

cation between system components, and results of the simulated community as well
as its response to demand response directives.
The real-time simulation platform showcased in this work demonstrates the abil-
ity to incorporate existing energy resources within a partially virtual environment
that fosters the development of future infrastructure essential for modern microgrids.
Most importantly, this platform enables the ability to test new microgrid control al-
gorithms without risking equipment damage or quality of service to the customer.
At the same time, it also provides the ability to gain valuable insight related to the
introduction of new elements and their impact on existing electric infrastructure.
Future use of this tool will help develop technology intended to make microgrid op-
erations more e↵ective, and to improve grid resilience by providing critical services
in the event of severe disruptive events.

8
Chapter 2

The Models

2.1 Circuit 16 feeder model

Circuit 16, one of the feeders in the Los Alamos, NM (Fig. 1) power distribution
network, was chosen for this study due to its utility-scale DERs and the diversity of
its loads. It hosts a 1 MW PV array and a substation-sited battery energy storage
system (BESS) composed of a 1 MW / 7.2 MWh NaS battery and a 0.8 MW /
1.2 MWh lead-acid battery. In addition to the utility-scale DERs, it serves a wide
variety of loads including residential, commercial, and critical loads such as water
pumps that contribute to the water distribution system of Los Alamos. Circuit 16
provides power to a total 1600 residential customers, while commercial consumers
include two grade-level schools, a sports complex, and a golf course.

2.1.1 From WindMil to GridLAB-D

The Los Alamos Department of Public Utilities (LADPU), the local Los Alamos mu-
nicipal utility, maintains a high fidelity model of the local power distribution system.

9
Chapter 2. The Models

Fig. 1: The Circuit 16 feeder, outlined in orange, serves the Northeast area of Los
Alamos, NM, covering mainly the Barranca and North Mesas. The battery storage
and PV array are located at the local landfill and they interconnect to Circuit 16
at the substation, outlined in blue. Customers served by Circuit 16 include a 1600-
home smart community, apartment complex, middle & elementary schools, a sports
complex, a golf course, water distribution infrastructure, among others.

The model is based on Geographic Information System (GIS) data and accurate
equipment specifications obtained from assets in the field and modeled in WindMil,
a power utility software designed to model and analyze electric distribution networks
for system planning and operations [8]. A total of 594 loads are defined in the model,
representing the transformer-level power consumption of over 1600 customers served
by this feeder. The reactance and resistance ratio, R/X, defined for the distribution
lines within the model range from 0.025 to 2.36 and they distribute power to end
users via overhead and underground lines with three-phase, bi-phase, and single-
phase configurations depending on the type of load served.
The WindMil Circuit 16 model, shown in Fig. 2, was translated to a GridLAB-D

10
Chapter 2. The Models

Fig. 2: Configuration of the Los Alamos Circuit 16 model, operated by the Los
Alamos Department of Public Utilities. Note the presence of water supply well
pumps (P1-P4), several pressure booster stations (B1-B5) and a water filter station
(F). A 1MW PV array (PV), and 2 MW / 8 MWh battery storage (BESS) are
located at the feeder head. Hospital (H) and supermarket (S) loads are also shown.
Hundreds of residential loads are distributed at the feeder end nodes.

model using the export function within WindMil, allowing to translate the model
from one platform to the other. Before exporting, Circuit 16 was isolated by sup-
pressing the other feeders within the Los Alamos power distribution network model,
ensuring that only the feeder of interest was translated to the GridLAB-D platform.
However, this led to a few errors in the GridLAB-D model. For example, a switch
object and its child node were created with mismatching phases. This was solved by
defining the objects with matching phases. In addition, switches located where Cir-
cuit 16 attaches to other feeders were generated in a closed state leading to nodes that
were not created during the translation process because they were defined within the
suppressed feeders. Once these objects were identified and removed the GridLAB-D
Circuit 16 model executed without errors.

11
Chapter 2. The Models

2.1.2 Model validation

The Circuit 16 model is defined with a constant voltage of 66,395 V at the source node
and stepped down via a transformer to 7,967 V at the head of the feeder. A snapshot
of the voltage profile across the GridLAB-D Circuit 16 model is illustrated in Fig. 3.
The graph on the left shows the voltage profile when serving only static loads. To
assess how PV penetration a↵ects the feeder’s voltage profile, a solar generation
resource was introduced near the end of a branch (at location B3 as shown in Fig. 2)
to represent a 1 MW PV array. As expected, the PV penetration caused the voltage
profile across the feeder to rise for all phases, but is especially noticeable on phase
A. Further inspection of Fig. 3 also shows that the voltage increase is highest at the
injection point. The case with no PV experienced minimum and maximum voltages
of 99.25% and 102.37% of the nominal voltage, respectively, while the case with PV
experienced a minimum voltage of 100.38% and a maximum of 102.86% of nominal.
The fact that the voltage experienced in both cases is within ±5% of the nominal
voltage confirms the Circuit 16 model meets ANSI C84.1 voltage range specifications,
which states that the voltage throughout a distribution feeder must stay within ±5%

Fig. 3: Voltage on each phase as a function of distance from the head of the feeder.
On the left is the case of power injected at the feeder head only. Power injected at
feeder head and at the end of a branch by a solar object (at B3 for testing purposes,
especially noticeable on phase A) is on the right.

12
Chapter 2. The Models

of the nominal voltage [9] specified for the feeder.

2.1.3 Critical loads

The hypothesis of this work is that during

emergency situations, a functioning critical
infrastructure could be accomplished using
distribution level DERs in combination with
a distribution management system. As a re-
sult, this could reduce the stress imposed
on transportation systems and emergency in-
frastructure during evacuation procedures by
enabling a ‘shelter in place’ strategy. Since
water, food, and medical care are basic needs
for a community they are also considered in
this thesis in an e↵ort to replicate realistic
loads on the feeder. Therefore, critical loads
are considered to be a hospital, supermarket,
and water distribution infrastructure.
The water system load was developed us-
ing five months of operation data collected
by LADPU from water pumps in the Los
Alamos water system. These pumps con- Fig. 4: Stochastic characterization of
tribute to the Los Alamos water supply di- a water well pump in Los Alamos wa-
rectly and via a number of water storage ter system.
tanks. A stochastic approach was used to
generate models that simulate the water treatment load associated with Circuit 16.
These models are characterized by probability density functions (PDF) correspond-

13
Chapter 2. The Models

ing to the number of events per day, the duration of the events, and the start-time of
the events experienced by each pump. The PDFs for the number of events per day
are described by discrete probabilities, while load durations and event start-times are
characterized by Rayleigh distributions and Gaussian Kernel Density Estimations,
respectively. As an example, the probability density functions for a water well pump
are illustrated in Fig. 4. In total, there are four water well pumps and five pressure
booster stations included in the Circuit 16 model. The location of the water pumps
on the feeder is indicated in Fig. 2, P1-4 and B1-5. Typical pump operation is shown
in Fig. 5, with normal operation of the water pumps at night, when demand is low.
The water is pumped from the wells to storage tanks distributed throughout the
community, and dispensed during the day. High-resolution data for the hospital and
supermarket are not available. Therefore, data for a hospital and a supermarket in
Santa Fe, NM, a location with very similar climatic conditions, were used instead.
The data correspond to the power consumption of these buildings during July 8,

Fig. 5: Power demand of the critical loads for the Circuit 16 feeder, for July 8, 2016.
Note that these loads, while supported by Circuit 16, are shared by the entire Los
Alamos community.

14
Chapter 2. The Models

2016. The electric demand profiles of these critical loads are also shown in Fig. 5.

2.2 Residential load generator & aggregator

The residential load generator & aggregator softwares were developed as part of a mu-
tual research project shared with this thesis. The contributions made by the author
to these models led to the ability to integrate them into multi-system simulations,
operation in real-time, and enhanced input and output data, as well as their ability
to communicate with other system components. The residential load generator and
aggregator are introduced and described next, with a focus on the details related to
this work.

2.2.1 Residential load generator

To study the e↵ects of the human factor

on a distribution feeder, it is necessary to
implement human behavior models that
represent the interactions of people with
the household appliances in their homes.
The results from an hourly load sim-
ulation obtained using the framework

Fig. 6: Measured residential load, at one- described by Hendron and Engerbrecht

minute resolution, compared to hourly [11] is compared to measured electric
building load simulation data. load data at one-minute resolution in
Fig. 6. It is evident that hourly reso-
lution data neglects considerable detail from the power consumption of a building.
Hence, it is not reasonable to use hourly data in situations with dispatchable DERs

15
Chapter 2. The Models

as it is often done in distribution simulations, given that dispatch of resources for

power balancing must occur at least at a one-second resolution. Also, using data
averaged temporally or topologically may lead to unrealistically smooth power flow
and corresponding power quality metrics. In addition, local bottlenecks related to
electricity infrastructure may not be revealed. This is especially true with high pen-
etration of DERs, including PV generation, demand response, and storage.
Residential loads are associated with the use of household appliances that consume
power ranging from a few Watts to a few kW. Here, a statistical approach is used
to represent the interactions of human users with appliances and to account for the
details of their operation, including the physical response of the house thermal man-
agement system. The use of a specific appliance is characterized by the number
of events, the start-time and the duration, each described by a probability density
function drawn from customer usage patterns. Statistical characterization of each
appliance is also associated with a demographic cluster to account for di↵erent use
of appliances by di↵erent people, namely working singles, working couples, families,
and retired people. This is similar to the method described in the work of Fis-
cher et al. [12]. The appliance package simulated for each home is composed of an
electric range, refrigerator, lighting, HVAC, water heater, and a clothes dryer. The
HVAC, water heater, and refrigerator models also contain thermodynamic systems,
while the dryers and lights are considered ON/OFF loads, and electric ranges use
a Markov-chain approach. An example of statistics for electric range operation is
shown in Fig. 7. In this case, start-times correspond to meal times with peaks that
indicate breakfast, lunch, and dinner. The magnitude of the peaks is associated with
the di↵erent lifestyles in each demographic cluster. For instance, strong breakfast,
lunch, and dinner peaks are specified for families, working singles and couples due to
routine-driven lifestyles, while retired people usually enjoy unconstrained schedules.
Durations PDFs are shorter for families during weekdays and longer for other clus-
ters due to limited time to prepare meals. Longer durations are assumed uniformly

16
Chapter 2. The Models

Fig. 7: Probabilistic characterization of the use of an electric range, in terms of prob-

ability density function for start-time, probability density function for duration, and
discrete probability for number of use events. Note that PDFs and event probabilities
are specified for four demographic categories.

for all categories during the weekend. The number of events is relatively high for
families during weekdays, but retired people cook the most. However, the number of
cooking events increases significantly for families on weekends due to a more people
eating at di↵erent times of the day.
The interaction between the user and the physical nature of the service provided by
the appliance is specific to the appliance. For instance, the interaction of the user
with the air conditioning system is described by the temperature setpoints and dead-
bands set by the user on the system’s thermostat. The following equations describe
the interaction between the load and the physical system:
dT
MS = Q̇L Q̇R , (1)
dt
Q̇L = K1 [Ta T (t)], (2)

Q̇R = ⇤ ⇥ COP ⇥ PAC , (3)

17
Chapter 2. The Models

where MS is the thermal capacity of the conditioned space, T (t) is the space tem-
perature, Ta is the ambient temperature, t is time, QL and QR are the thermal losses
from the structure and thermal gain from the air conditioner or heat pump, respec-
tively, and K1 is a constant related to the building’s thermal insulation. Also, the
coefficient of performance is denoted by COP , the power of the compressor is PAC ,
and ⇤ is a state function that indicates whether the air conditioner or heat pump is
ON or OFF. In its basic form, the air conditioning or heat pump is controlled by the
switching logic

if T (t) < TL then ⇤ = 0, (4)

if T (t) > TU then ⇤ = 1, (5)

where TL and TU are the lower and upper deadbands for the temperature control.
In addition to the air conditioner system, the loads of a refrigerator, a heat-pump
type water heater, a clothes dryer, an electric cooking range and lighting are also
considered. A similar thermal model to the air conditioner is used to describe the
operation of the refrigerator. However, the user interactions with the refrigerator

Fig. 8: Typical electric loads by appliance for a single meter, for a period of one
week. The loads are clothes dryer (1), air conditioner (2), domestic hot water (3),
refrigerator (4), cooking range (5) and lights (6).

18
Chapter 2. The Models

are described as a function of open-door events instead of temperature setpoints.

Similarly, the water heater is a thermal load with human interactions modeled as a
function of water draws describing activities such as showers and dishwashing. The
clothes dryer and the lights are modeled simply by their start-time and duration. If
the event is active, a load is specified, otherwise there is no load. To simulate multiple
lights in a house, multiple overlapping events are possible for lighting, while only one
dryer event can be active at any given time with no overlap. In the case of the lights,
the power of the LED light bulb is selected from a list of possible values. Furthermore,
the Markov chain approach to simulate the lighting load used by Widén et al. [13],
is employed here to model the power consumption of the electric cooking range. The
typical power consumption of the household appliances considered here are shown in
Fig. 8. Note that some loads, such as the clothes dryer, are high-power and relatively
infrequent, while others are low-power but very frequent, such as the refrigerator.
As a result, the total energy consumption by all the appliances is comparable. In
addition, while the use of some appliances such as the refrigerator compressor are

Fig. 9: Total feeder load resulting from 1000 houses, also indicating the aggregated
contribution from each appliance. The total load for an individual meter (42) is also
shown for comparison. The load duration curve (shown in red) indicates significant
opportunity for shifting and deferring loads.

19
Chapter 2. The Models

deferrable without detectable negative e↵ects to the quality of service (QoS) to the
customer, others such as the use of the cooking range cannot be deferred, but must
still be characterized.
The total load of a residential community of 1000 houses is shown in Fig. 9 as an
example. Also, the total load for a single home is shown in black for comparison with
the total load of the community. It is evident that while the load of an individual
meter displays significant variation in time with high variance, the total feeder load
is relatively smooth and predictable. Also note that the power consumption of the
individual home shown in this example is qualitatively very similar to the measured
load shown in Fig. 6. Finally, the load duration curve of the 1000 house community
indicates a number of interesting features of the load. First, a substantial peak load
greater than 1.5MW occurs for less than approximately 10% of the time and could be
easily removed by demand response in combination with storage. Second, the base
load is only approximately 15% of the peak load. Third, the absence of a significant
plateau indicates that there is opportunity to shift loads in time with appropriate
control.

2.2.2 Residential load aggregator

In this study, a residential load aggregator manages the real-time demand response
of the system. The appliances controlled by the aggregator are thermostatically
controllable loads (TCL) such as the air conditioner (AC) and water-heating units
in every home. For example, a 200m2 house has about 0.15kWh of stored energy
per °C associated with the air mass alone, coupled with roughly 3kW of power
controllable through a home thermostat. The load aggregator aggregates the total
power consumption of the TCL units and calculates the capacity based on the state
of the units in the system. The total power available to be decreased or increased
through demand response depends on the number of units that are ON and OFF,

20
Chapter 2. The Models

Fig. 10: Aggregated AC DR successfully following artificial sine wave signal. Total
consumption of 200 AC units in in purple, filtered load and signal in yellow and
green, respectively.

respectively. In this framework, instead of controlling each thermostat individually,

the load aggregator broadcasts a single control signal to the entire fleet of thermostats
in the system. The signal is in the form of a number that represents the probability of
an available compressor switching from its current state. Furthermore, each “smart”
thermostat decides to switch based on its current state and an internally generated
random number. The broadcast signal is generated based on the error between
the current power of the TCL system and a desired power setting that can vary
dynamically. In other words, the system can be set to follow a static load shape
on a schedule set by the utility to reduce the load on a feeder during peak hours,
or it can be set to dynamically follow a changing signal such as solar irradiance.
Fig. 10 shows the response of a TCL system composed of 200 thermostats following
a demand response signal in the form of a sinusoid with a period of 20 minutes and
amplitude of 50kW. This simulation test shows the system successfully following the
artificial signal over a period of 20,000 seconds, or roughly 5.5 hours. The purple
line shows the actual aggregated power consumption of the 200 AC units, which is

21
Chapter 2. The Models

then passed through a bandpass recursive time series digital filter [14] to obtain the
yellow curve for comparison with the sine wave. The filtering limits the response to
only a specific band of frequencies for comparison with the demand response signal.

2.3 Battery controller

Utility-scale DERs such as batteries and PV arrays are vital resources in cases when
the grid fails during emergency situation. They must have the ability to operate
independently of the main grid in order to keep critical loads online. For example, in
the case of a community such as Los Alamos where utility-scale DERs are available,
batteries must be able to charge strictly from power generated by the PV array when
power from the grid is not available. Also, power dispatch should be in such a way
that optimizes the use of resources regardless of the state of the grid. The battery
controller described here was designed to maintain the load at a point of interest,
such as a transformer, at or below a predefined target power value by managing the
power dispatched from available resources within the GridLAB-D powerflow simu-
lation environment. It also enables the battery to charge directly from the power
generated by an accompanying solar array. This battery controller was specifically
created for this study since these capabilities were not available in GridLAB-D. Sim-
ilar load following control schemes exist today in which batteries are set to charge
or discharge at a constant rate when the load at the point of interest reaches certain
thresholds. However, none combine the ability to charge a battery strictly from an
accompanying solar array, manage the dispatch of available resources, and vary the
discharge rate of the batteries according to the power needed to reach and remain
at a target value.
The battery controller algorithm for this study was internally imbedded within the
source code of the inverter object in GridLAB-D to which both the solar and battery

22
Chapter 2. The Models

Fig. 11: Logic flowchart for battery control used to maintain the load at a point
of interest on a feeder at or below a predefined target power value. The feedback
control loop minimizes the di↵erence between the current load at the point of interest
and the target load. During the day, the controller dispatches PV and the battery if
necessary; PV surplus is used to charge the battery. Only the battery is dispatched
when necessary during the night, since PV power in not available.

objects connect to. The control algorithm uses the GridLAB-D simulation time han-
dling process to create a feedback loop used to minimize the error between the target
power value and the load on the point of interest. This method leads to the controller
sensing the magnitude of the load as it was one second in the past, resulting in a
one-second lag in controller action. However, this lag is negligible at the distribution
level due to the relatively smooth load experienced by a substation. The controller
dispatches PV during the day while charging the battery from PV surplus, and dis-
charged only if it is necessary. During the night, only the battery is dispatched when
necessary since PV power is not available. At this point, the battery is only allowed

23
Chapter 2. The Models

to charge from a PV array; however, the ability to also charge from the main grid is
within the scope of the future work discussed in Section 5.3. The flowchart in Fig. 11
shows the logic behind the battery controller. As a test, a simulation was conducted
to examine the controller’s performance. The test model consists of a single home
connected to a bus to which a 500kWh battery and a solar array are interconnected
for testing purposes. The PV and battery are defined with a rated power of 20kW
and 10kW respectively. Fig. 12 shows the controller successfully maintaining the
load on the primary meter at the predefined target load of 5kW for a period of two
days, mainly due to the oversized capacity of the resources. The spikes above and
below the target load are due to the one-second lag of the system and the resolution
of the data collected. In a future version of this controller, the lag will be eliminated
by calculating the power demand on the point of interest and forcing GridLAB-D to
iterate the power flow solution for the time-step in which controller action takes place.

Fig. 12: Example of battery control performance on a test model exhibiting oversized
DERs for testing purposes. The controller successfully holds the load on the primary
meter at the target load of 5kW. The spikes above and below the target load are due
to the one-second lag and resolution of the data.

24
Chapter 3

Simulation Setup

3.1 GridLAB-D co-simulation

A co-simulation that engages multiple incompatible models or software, requires an

interface with the ability to handle communication and simulation-time synchro-
nization between the softwares. Two co-simulation interface methods used to en-
able data transfer and synchronization between the C++-based GridLAB-D and the
FORTRAN -native residential load generator and aggregator, are described in this
section. The first method uses a GridLAB-D directive that enables communication
with Matlab. The second method employs FNCS, a software especially designed to
serve as the interface for co-simulations. Their setup and flow of information are
described next.

3.1.1 Matlab as co-simulation interface

GridLAB-D’s Matlab Link directive is used to integrate the residential load generator
and aggregator with the GridLAB-D platform. This link directive enables Matlab

25
Chapter 3. Simulation Setup

to execute specific commands on every time-step, before GridLAB-D performs main

object and power flow calculations. This allows for the loads to be updated by
the load generator in synchronicity with GridLAB-D. Generally, external loads are
provided to GridLAB-D through players that read load-shapes from static files and
apply the load to load objects. In the case of demand response, the load must
evolve in parallel with the GrdiLAB-D simulation in response to controller directives
issued by the aggregator. Here, Matlab fills the role of data communicator and
synchronizer, while GridLAB-D acts as the master clock by utilizing its internal
clock. Moreover, the load generator and aggregator are considered a single simulator
within this setup, as illustrated in Fig. 13. That is, data is from the Matlab interface
is received only by the aggregator, and the output data is sent back to the interface
only by the load generator. Note that the communication between the aggregator
and generator takes place independently, or outside, of the co-simulation interface.
Instead, they exchange data via .csv files that also serve as communication signals.

Fig. 13: Block diagram of information flow (solid black lines) between GridLAB-D
and several input components, including non-controllable variable loads (via players)
and controllable residential loads (via Matlab interface). Control signals are shown
by dashed red lines. Note that the load generator and aggregator are considered a
single simulator indicated by the blue dashed boxed.

26
Chapter 3. Simulation Setup

In this exchange, the load generator sends the status of each unit participating in
demand response and the aggregator provides the control signal for them to act upon.
The co-simulation process is the following:

1) GridLAB-D initializes the simulation and prepares the environment for the first
time-step, t = 1.

2) Matlab receives a signal from GridLAB-D to execute and begins by directing the
load aggregator to run. It reads the status of the participating DR units and
creates a demand response control signal (initial statuses are used on t = 1. Also,
control signal is zero when demand response is disabled). The control signal is
then stored in a .csv file for the load generator and its filename serves as a unique
identifier.

3) The creation of the file containing the DR control signal also indicates the load
generator to execute. It begins by registering the control signal and deleting the
file. It continues by creating load values for the load objects defined in the Circuit
16 model and storing them in a .csv file with its unique identifier in the filename.
A second .csv file is created containing status of DR units for the load aggregator
to use in the next time-step. This is followed by a return signal sent to Matlab
indicating that the loads have been created and stored successfully.

4) Matlab reads and imports the load values from the load file into GridLAB-D,
updating the load values of the load objects.

5) Object and powerflow calculations are performed in the GridLAB-D environment

for the current time-step t.

6) GridLAB-D advances in time to the next simulation time-step, t = t + 1.

7) Steps (2)-(6) are repeated until the simulation ends.

27
Chapter 3. Simulation Setup

Here, the signals exchanged between GridLAB-D, Matlab, the load generator, and
the aggregator are crucial for synchronization of the overall simulation. They en-
sure that steps (2)-(6) are performed only once per time-step and not every time
GridLAB-D iterates when solving powerflow across the feeder. Once again, the load
generator and aggregator are treated as a single simulator unit within this setup,
as indicated by the blue dashed box in Fig. 13. To simulate a community with de-
mand response capabilities within Circuit 16, 50 residential meters are simulated by
the load generator and aggregated to 10 load values that represent the number of
residential transformers in this community. To account for the 594 residential loads
originally defined in the model, a static load-shape is used for the remaining 584
fixed residential loads that are defined within the GridLAB-D model. Scale factors
are used to ensure that the maximum power consumption of each load object under
the use of the load-shape, is comparable to their original constant power consump-
tion defined in the model. Static load-shapes are also used for the load associated
with critical infrastructure and solar irradiance, and are fed into GridLAB-D using
conventional player objects.

3.1.2 FNCS as co-simulation interface

The FNCS co-simulation interface is an enhanced version of the Matlab interface as

they were both built with the same purpose. The FNCS interface allows for a higher
data transfer capacity and speed with stronger reliability and performance. In a
similar way as in the Matlab interface setup, the GridLAB-D Circuit 16 model is
integrated with the residential load generator and aggregator within a co-simulation.
Here, the FNCS software allows for a co-simulation in which residential loads are
updated by the load generator in synchronicity with GridLAB-D and the load aggre-
gator. Similar to the Matlab setup, loads must evolve in parallel with the GridLAB-D
simulation in response to demand response directives issued and managed by the ag-

28
Chapter 3. Simulation Setup

gregator. In this setup, the FNCS broker represents the central server to which all
simulators connect to in order to synchronize in time and exchange messages with
other simulators [15]. The broker acts as the master clock, simulation synchronizer,
and data communicator between the three co-simulators. The simulation process is
the following:

1) The co-simulators including GridLAB-D itself, are initialized and they prepare
their environment for the first time-step, t = 1. This includes connecting with
the FNCS broker and subscribing to publications that will be made by other
co-simulators.

2) The FNCS broker receives the status of each simulator and sends the signal to
execute for the current time-step only when all co-simulators are ready.

3) The simulators receive the signal from the broker and begin by requesting the
published values to their subscriptions from the broker. This happens at the
beginning of each time-step (initial conditions are used for t = 1).

4) The simulators execute for the current time-step, t, and publish their output to the
FNCS broker, which then brokers the publications to the respective co-simulator
subscription.

5) Steps (2)-(4) are repeated until the simulation ends.

It is important to note that, unlike in the Matlab interface setup, the load generator
and aggregator are treated as individual simulators and communicate only through
FNCS, as illustrated in Fig. 14. Furthermore, all simulators obtain values to their
subscriptions at the beginning of each time-step as indicated by step 3. Here, only
the solar irradiance and the critical loads associated with the hospital, supermarket,
and water distribution are static load-shape fed into GridLAB-D by player objects,

29
Chapter 3. Simulation Setup

Fig. 14: Block diagram of information flow between FNCS and co-simulators, includ-
ing GridLAB-D, residential load generator and aggregator. Critical loads and plane
of array irradiance are read by GridLAB-D via players.

similar to the Matlab setup. Due to the higher data transfer capacity, the load gen-
erator simulates a total of 1600 homes aggregated to 594 values that represent the
number of residential homes and transformers within the Circuit 16 feeder, respec-
tively. The load generator publishes the 594 aggregated values for the GridLAB-D
Circuit 16 model in addition to TCL status of each home for the aggregator. In
return, the aggregator publishes the control signal for the load generator indicating
the demand response directive. Finally, the Circuit 16 model publishes the solar
irradiance used by the aggregator to generate the demand response signal. Overall,
a total of 2,196 values are transferred between three co-simulators each simulated
second.

30
Chapter 3. Simulation Setup

3.2 Real-time simulation

The residential load generator and aggregator

were integrated into the real-time simulation de-
scribed in Section 1.3 as a virtual proxy of a de-
mand response enabled community of 200 homes.
To simulate the physical separation of the sim-
ulated residential infrastructure and the residen-
tial load aggregator in real life, the models are
executed on individual Raspberry Pi 3 comput- Fig. 15: Raspberry Pi 3, credit-
ers (Fig. 15). These are credit-card-sized single card-sized single board computers
board computers that run on Linux operating running Linux operating systems.
systems and o↵er Internet connectivity [16]. The
simulation employs a central database as the interface between real and virtual com-
ponents responsible for real-time information used for operations optimization and
control. Communication between the models and with the database is performed via
the Internet to replicate communication of these devices in the field. Furthermore,
the models interact with the database using a Python-based interface developed by
Sandia National Laboratories, while communication between models takes place us-
ing the scp (Secure Copy) command line tool. This tool copies files across remote
hosts using a secured encrypted connection over a network [17, 18]. Fig. 16 illustrates
the simulation setup between the load generator, the aggregator, and the database,
as well as the total flow of information that takes place on every time-step of the
simulation.
The simulation was set up to run in real-time with 1-second resolution in order to
accurately represent the power consumption of the Mesa del Sol 200-home residen-
tial community, along with demand response controllability of TCL units described
in Section 2.2.2. At the beginning of each time-step, the residential load genera-

31
Chapter 3. Simulation Setup

(a) Real-time simulation setup. (b) Real-time data transfer diagram.

Fig. 16: Real-time simulation setup showing the models and their interaction with
the database. (a) The residential load generator and aggregator reside in individually
independent Raspberry Pi 3s which communicate with the centralized database.
(b) Input and output data flow is shown with green and red arrows, respectively.

tor and the aggregator request the current outside temperature and a MPC setting
value from the database, respectively. Recall from Section 1.3, that the MPC pro-
vides real-time dispatch of resources by deviating from schedules provided by the
VPP. Moreover, the load generator and the aggregator exchange TCL data critical
for demand response. In this data exchange, the load generator receives a control
signal from the aggregator as an input and returns the current statuses of the TCLs
to the aggregator. The aggregator takes the TCL statuses and the MPC setting it
receives from the database as inputs, and determines the control signal based on
these values. Samples of the TCL status sent to the load aggregator are illustrated
in Table 1, where the first column represents the meter number of each house, the
second and third columns indicate whether the AC unit of the home is ON and if it
is available to switch states (from ON to OFF, and vice versa), respectively. A value
of 0 indicates that the AC is OFF and a value of 1 indicates it is ON. Similarly, 0
signifies that a given AC unit is not available to switch states, and the opposite is
signified by a value of 1. The last three columns inform the aggregator of the overall

32
Chapter 3. Simulation Setup

meter id ac on ac switchable ac soc lower db higher db

1 0 1 0.288380 16.000000 26.000000
2 0 1 0.426664 16.000000 26.000000
3 0 1 0.339106 16.000000 26.000000
4 0 1 0.372218 16.000000 26.000000
5 0 0 0.033459 16.000000 26.000000
... ... ... ... ... ...
196 0 1 0.365487 21.742659 23.742659
197 0 1 0.219987 21.784323 23.784323
198 0 1 0.086933 21.990416 23.990416
199 1 0 0.245799 16.000000 26.000000
200 0 1 0.066387 16.000000 26.000000

Table 1: Sample of TCL status data passed from the residential load generator to
the aggregator for 200 air conditioners. The data include the identification number
for each meter, the current state of the unit (ac on), availability to switch states
(ac switchable), the state of charge of the unit (ac soc), and the lower and upper
deadbands (lower db & higher db).

state of charge, as well as the lower and higher deadband setpoints of each unit.
The flow of input data at the beginning of each time-step is depicted in Fig. 17a
and the data flow at the end of each execution between the load generator, the ag-
gregator, and the database is illustrated in Fig. 17b. The load generator aggregates
the power consumption of the 200 homes into 38 load values that correspond to the
38 residential transformers located in the Mesa del Sol community. The total power
consumption of the community is also calculated and sent to the database along
with the 38 transformer values. Each value is given a database load identification
number (load id) along with the timestamp of the current time-step of the simula-
tion. Table 2 shows an example of this data in its final format before it is sent to
the database. Note that the first row, load id=2, corresponds to the total power
consumption of the community and the rest correspond to the transformer values.
Also, load id=1 is associated with the load of an existing commercial building in
the area. At the same time, the aggregator sends the TCL’s current state of charge,

33
Chapter 3. Simulation Setup

(a) Input. (b) Output.

Fig. 17: Real-time simulation input & output data flow. (a) The load generator and
aggregator request data from the database as their input. They also exchange TCL
data for DR. (b) The simulators send their output to the database at the end of each
time-step.

power, and capacity to the database to be used by the MPC and VPP.
In order to make the models execute in real-time it was necessary to synchronize
the simulation with the current time, while also ensuring that the models execute
a time-step every second. Synchronizing with real-time requires each model to keep
track of the current date, time (with a resolution of 1-second), and the current day of
the week. The day of the week is important due to the di↵erence in residential power
consumption behavior during weekdays, when people are usually at work, then on
weekends, when people are more likely to be at home. At the start of the simulation
each model requests the clock value of its host machine which provides the current
date and time. The models then take a di↵erence of days between the current date
and a reference date for which the weekday is known. Here, the reference date used
is Sunday January 1st, 2017. With this information and the fact that there are seven
days in a week allows the models to calculate the current day of the week and to
represent it with an integer from 1-7, where 1 represents Sunday. Moreover, the time
obtained from the clock query indicates the start hour, minute, and second of the

34
Chapter 3. Simulation Setup

load id datetime measured value

2 2017-08-23 11:47:03 126269.070
3 2017-08-23 11:47:03 10334.182
4 2017-08-23 11:47:03 4540.000
5 2017-08-23 11:47:03 758.000
... ... ...
36 2017-08-23 11:47:03 532.000
37 2017-08-23 11:47:03 7290.000
38 2017-08-23 11:47:03 6600.182
39 2017-08-23 11:47:03 5342.182
40 2017-08-23 11:47:03 2350.182

Table 2: Sample data transferred to the database by the residential load generator,
representing the total aggregated load (load id=2), and followed by individual trans-
former loads, for use in the Opal RT real-time simulator. Note that load id=1 is
associated with the load of a building outside of the scope of this work.

simulation, as well as the number of time-steps left for the current day. In order to
keep track of time and generate the correct timestamp, the clock value is requested
from the host machine on every time-step. To ensure that times remain uniform
across components, each machine obtains the global time from a network time pro-
tocol server via the internet.
Ensuring that the models execute every second required the development of a trigger
that indicates to the models when to run. The Python-based trigger used here is in-
dependent of the simulation and its components, and runs quietly in the background
keeping track of the computer’s clock. It creates a trigger file that signals the models
to execute each time the seconds digit of the clock changes. Each model is associ-
ated with an individual trigger that lives on the same machine and is responsible for
signaling its corresponding model to execute for the current time-step. At the end of
the execution process, the models delete the trigger file in sub-second time intervals.
Note that the trigger file is created regardless of what is happening within the sim-
ulation; therefore, proper error checks and error-handling logic were necessary and
are crucial to ensure that the simulation keeps running even if fatal errors are en-

35
Chapter 3. Simulation Setup

countered. These error-handling procedures were designed with field deployment in

mind. For example, the load aggregator would normally crash if it was signaled to
execute a time-step, and the current TCL status data was not available. Instead,
the models are instructed to reuse the data they had on the previous time-step if the
new input data is missing or if was not updated in time.

36
Chapter 4

Simulation Results

4.1 GridLAB-D co-simulation

Several scenarios were examined using both of the GridLAB-D co-simulation inter-
face setups described in Section 3.1. In both cases, a preliminary simulation was
performed to confirm that the combined models represent real distribution feeder
behavior. Once realistic behavior is established, scenarios are performed using the
battery control and residential demand response controllability. All scenarios con-
sidered using such co-simulation interfaces are simulated to take place during July 8,
2016, the day corresponding to the load-shapes used for the critical loads. In addi-
tion, a hot summer day is ideal to stimulate the real-time demand response control
of AC units and observe its performance.

4.1.1 Matlab interface

Three cases are considered using the Matlab interface. First, the typical ‘fixed sched-
ule’ approach, in which the utility battery is charged at night and discharged during

37
Chapter 4. Simulation Results

peak hours, with no additional controls. In the second case, the battery controller
from Section 2.3 is introduced. It uses load feedback from the substation transformer
to dispatch the utility-scale battery in order to limit the peak load on the feeder. In
the third case, an example of demand response is demonstrated by aggregation of
household water heaters. The residential load is composed of a static load-shape fed
into 584 load objects, and scaled such that the peak power consumption of each load
is equal in magnitude to the constant power value with which they were originally
defined in the model. In addition, 50 residential homes are aggregated to 10 load
values that are passed to GridLAB-D for the total 594 load objects. They repre-
sent the number of homes and transformers located within a section of a simulated
community with demand response capabilities. The solar irradiance and the criti-
cal loads associated with the hospital, supermarket, and water distribution are also
static load-shape read by GridLAB-D using conventional player objects. Outside
temperature is represented by a sinusoid, Tamb = 28 8sin hour⇤⇡
12
, with characteris-
tics similar to the natural temperature cycle experienced throughout the day. The

Fig. 18: Artificial outside temperature representing a hot summer day in Los Alamos,
NM. Maximum artificial temperature is comparable to the maximum recorded tem-
perature for Los Alamos. The temperature profile is represented by a sine wave.

38
Chapter 4. Simulation Results

temperature is set to range between 20-36°C to represent a maximum temperature

similar to the highest recorded temperature in Los Alamos, NM of 35°C in 1998
(Fig. 18). Note that the purpose of the Matlab co-simulation setup was mainly to
establish the ability to integrate models that cannot operate satisfactorily together
(incompatible models) under typical computer conditions. The goal is to perform
co-simulations that consider aspects from di↵erent domains, such as human behavior
and power flow principles, in order to simulate accurate power distribution behavior
and to ensure that the simulated infrastructure operates within physical limits.

A. Uncontrolled conditions

Power distribution feeders equipped with DERs such as PV and batteries usually dis-
charge the batteries during peak load hours to reduce the stress on the transformer,
while charging takes place at night when wholesale energy cost is low. This mode
of battery operation is shown in Fig. 19. The typical battery schedule indicates the
battery to charge from midnight until 07:00 at 1/3 of its power capacity (⇠ 540kW)

Fig. 19: History of net load at the feeder head, in comparison with critical loads,
residential loads and battery charge / discharge on July 8, 2016.

39
Chapter 4. Simulation Results

using power from the grid. It then dispatches power at 1/4 capacity (500kW) from
07:00-10:00, in order to aid with the typical morning peak associated with consump-
tion as people prepare for work; shower, breakfast, etc. The battery is set to idle for
two hours before dispatching power again from 12:00-15:00 at about 556kW to as-
sist with the typical lunch peak. This is followed by another five-hour idling period
while most people are at work. Furthermore, another peak occurs around dinner
time associated with people returning home from work, cooking dinner, doing laun-
dry, etc. Thus, the battery is set to dispatch power from 20:00-21:00 at 56% capacity
(⇠ 1.1MW) in order to lessen the load on the feeder associated with this behavior. It
is also important to note from Fig. 19, the intermittency of the PV array associated
with a relatively cloudy day. An intermittent irradiance profile was used in an e↵ort
to deviate from ideal solar conditions in order to observe interesting characteristics
of the system that might not be easily apparent under ideal conditions.
A closer look at the history of the loads shown in Fig. 19, indicate that while the
residential load consumption of 1600 houses is low during the night, it dominates
during the day. Comparing the size of the residential load to the critical loads dur-
ing the day confirms the potential for aggregated residential demand response as an
e↵ective tool to ensure that power is available to serve critical loads. Moreover, the
loads associated with pumps used to distribute water to the community are high at
night to take advantage of low energy costs and also to manage total feeder load
e↵ectively. While this is a good strategy for normal conditions, moving pump loads
to the daytime, when PV power is available, may be a preferable strategy.

B. Control with battery

In some applications, more complex control algorithms are used instead of fixed
schedules to maximize the benefits of a battery. A demonstration of this kind of
operation is described in the work of Lavrova et al. [19, 20], in which a framework of

40
Chapter 4. Simulation Results

control techniques was used to optimize the benefits of battery in conjunction with
PV at the distribution scale.
The performance of the GridLAB-D built-in battery controller described in Sec-
tion 2.3, is demonstrated here. Results from the co-simulation illustrate the dispatch
of PV and battery power managed by the inverter. Note that all remains identical to
the previous case except for the operation of the battery. Here, the main goal of the
controlled operation of the battery is to maintain the load at the substation trans-
former at or below 2MW for the duration of the day. The controller dispatches PV
and if necessary the battery to ensure that the transformer load remains below the
predefined threshold as shown in Fig. 20. The initial battery SOC (state of charge)
is assumed to be 0.9 at midnight, resulting from control actions implemented the
previous day. It is evident that the control is e↵ective until shortly after noon, when
battery capacity runs out after reaching its reserve SOC of 25%, in part due to lack
of PV power owing to partly cloudy conditions. After that, loads are supported pri-

Fig. 20: Performance of battery control during the day of July 8, 2016. The battery
control the load at the substation transformer e↵ectively at the pre-set maximum
value of 2MW until the minimum SOC of 0.25 is reached at around noon. After
that, loads are met via the PV and the grid only.

41
Chapter 4. Simulation Results

marily by the grid, with partial PV support as dictated by sky conditions. While this
is generally not a problem when the grid connection exists (unless maximum trans-
former load is violated), operation of critical infrastructure in emergency conditions
could not be supported. The insight gained from these results highlight the value
of this co-simulation framework as a tool to study the e↵ects of more sophisticated
control algorithms on the electrical infrastructure.

C. Control with DR

There are two possible methods to support critical infrastructure operations. One
is the availability of stand-by fossil generation in combination with larger battery
storage. The other is curtailment of non-critical loads. The former option is expensive
and not likely to be implemented even if the value of resilience were built in to the
infrastructure cost. The second option is more cost-e↵ective, but would require the
implementation of a control infrastructure to manage the partial operation of non-
critical loads. While the design of such an infrastructure is outside the scope of this
thesis, an example of how the simulation framework used here could assist in the
design process is given below.
The residential load aggregator is introduced here to perform demand response on
50 residential homes by controlling their water heating units simulated within the
load generator. The aggregated load of the 50 meters is shown in Fig. 21. The
initial state of the heat pump water heaters is OFF, with an initial SOC normally
distributed around 0.25 with a standard deviation of 0.05, and a SOC deadband of
0.9 to 1.0. The minimum deadtime conditions of compressor operation are met after
5 minutes, indicating the heaters to turn ON and begin the charging process. In the
meantime, the GridLAB-D controller sends a signal to the load aggregator indicating
to reduce residential water heating loads. However, this is prevented internally by the
aggregator due to the SOC value being below the deadband minimum. At 30 minutes,

42
Chapter 4. Simulation Results

the internal deadband minimum is changed to 0.4, allowing the aggregator to turn
OFF heaters that are ON according to the control signal. The aggregated control
results in an immediate load reduction of 50kW within 1 minute, and a reduction
of almost 100kW for the 50 houses within 10 minutes. With higher participation,
this e↵ect would be magnified accordingly. For example, a load reduction of about
500kW would be possible with demand response control of a 250-house community.
In the case of an unexpected emergency, a utility could employ this technique to
reduce the load on a feeder while a backup generator is initiated for support. Similar
algorithms could be implemented for laundry loads and other non-essential appliances
to reduce load during emergency conditions. An example of air conditioning control
is discussed in Sections 4.1.2.B and 4.1.2.C.

Fig. 21: Example of aggregated DR of 50 domestic hot water heaters. At 30 minutes,

the lower deadband limit for the state of charge is changed from 0.9 to 0.4, allowing
the external signal from the aggregator to turn o↵ available heat pump compressors.

43
Chapter 4. Simulation Results

4.1.2 FNCS interface

Here the FNCS interface was used to examine three cases. The first consists of typi-
cal operating conditions used to establish realistic distribution feeder behavior of the
system, similar to the uncontrolled case performed using the Matlab interface. In the
second case, demand response is enabled using solar irradiance as a signal to control
a fleet of air conditioner systems within 1600 homes. The third case demonstrates
the performance of demand response coupled with the battery controller. The main
di↵erence between the Matlab interface and FNCS, is the amount of data shared by
the simulators. In this case the load generator simulates a total of 1600 homes aggre-
gated to 594 values that represent the number of residential homes and transformers
within the Circuit 16 feeder, respectively. Also, demand response is applied to the
entire residential community, as opposed to only a small section of a community as
seen in Section 4.1.1.C. Furthermore, temperature data recorded on a mild summer
day in Phoenix, AZ, shown in Fig. 22, was used here to induce stronger demand
response e↵ects. However, similar to the Matlab setup, solar irradiance and critical

Fig. 22: Temperature recorded on a mild summer day in Phoenix, AZ. It is used
here to intensify the e↵ects of demand response control of 1600 AC units.

44
Chapter 4. Simulation Results

loads remain as static shapes.

A. Baseline

In order to establish a performance baseline for the system, the same typical mode
of operation for distribution feeders described in Section 4.1.1.A is used here. Note
that the battery is set on the same dispatch schedule discussed in Section 4.1.1.A,
indicated by the red line in Fig. 23. Moreover, PV generation, as well as the loads
associated with critical infrastructure such as the hospital, supermarket, and water
pumps, remains the same as in the Matlab benchmark scenario. However, in this
case the entire residential load is simulated by the load generator composed of 1600
homes. The residential load results in a typical profile with low consumption during
the night, and high power consumption during the day.
To examine the e↵ects of DER penetration to the feeder, a voltage profile was gener-
ated during maximum PV generation and battery dispatch totaling roughly 1.6MW

Fig. 23: History of net load at the feeder head, in comparison with critical loads,
residential loads and battery charge/discharge on July 8, 2016 using the FNCS in-
terface.

45
Chapter 4. Simulation Results

(a) Node voltage at 12:55. (b) Voltage history of a branch.

Fig. 24: Voltage behavior on Circuit 16 during baseline case. (a) Profile of node
voltage as a function of distance from the source node experienced by Circuit 16
during high PV penetration and uncontrolled conditions, 12:55. (b) Voltage history
of a node located midway on the branch below location B3.

provided by DERs. Fig. 24a shows the voltage at every node as a function of their
distance from the source node at 12:55. The maximum and minimum voltage val-
ues are 104.2% and 101.1% of the nominal voltage of 7967.43V, respectively. While
these values are within the ANSI C84.1 voltage range specifications, the maximum
voltage value is within 0.8% of violating the upper voltage limit. It is evident from
Fig. 24a, that this is due to DER dispatch since the maximum voltage is experienced
at the substation where DERs reside. This suggests that DER penetration is ap-
proaching its limit due to high PV generation. At this point, it would be necessary
to lower the support from the battery to avoid any violations. Ideally, the battery
would be set to charge in order to take advantage of the power generated by the PV
array. Furthermore, this points out the importance of more sophisticated controls
that automate battery activity in order to replace fixed schedules and limit human
interaction, while improving performance and anticipating potential violations to in-
dustry specifications.
Additionally, to examine how other parts of the feeder behave over time, the voltage
history of a node located midway on Barranca Mesa, the branch below location B3

46
Chapter 4. Simulation Results

in Fig. 2, is shown in Fig. 24b. This location gives an interesting insight into the
physical behavior of the system due to its distance from the source node and the fact
that it is the heaviest populated area in the feeder, consuming more than a third of
the total load. Its profile reveals that the voltage at this location remained between
ANSI C84.1 voltage specifications and experienced a maximum voltage of 103.7%
and a minimum of 99.5% of the nominal voltage.

B. Residential DR

In this case, residential demand response is used to control a TCL system of 1600
AC units with a maximum power capacity of 400kW. The system is set to respond
to a control signal generated from solar irradiance and operates between 9:00AM-
5:00PM, the period of the day when the sun is the most intense. Note that the loads
associated with the critical infrastructure and the power provided by the DERs re-
mains as stated in the baseline case including the battery.
The demand response performance is illustrated in Fig. 25 with the aggregated
power consumption of the AC units represented in purple, while the filtered AC
consumption and the filtered solar irradiance are represented in yellow and green,
respectively. Here, the filtered solar irradiance is the desired output of the system,
and thus, is used as the demand response control signal. Similarly, the filtered AC
consumption represents the system’s response to the signal. The filtering of these
components is done using the bandpass recursive time series digital filter discussed
in Section 2.2.2. It allows for a proper comparison between the control signal and
the output of the system within the same frequency domain. It is evident from in-
specting Fig. 25, that the response of the system is in strong alignment with the
desired output, indicating that the demand response control was successful during
the control period of this simulation. However, a closer inspection of the filtered AC
consumption at 12:00, shows an example of a case when the TCL system runs out of

47
Chapter 4. Simulation Results

Fig. 25: Aggregated DR successfully following signal from solar irradiance between
9AM-5PM. Total consumption of 1600 AC units in purple, filtered load and signal
in yellow and green, respectively.

energy and is unable to meet the control demand. This happens because there are
not enough AC units available to turn OFF, in other words, the AC units that are
ON cannot change states. Doing so, would cause them to violate their thermostat
setpoint compromising the quality of service to the customer.
The e↵ects of demand response on the distribution feeder and on the overall residen-
tial power consumption are observable in Fig. 26 compared to the baseline results.
As expected, both figures show identical profiles from midnight until 09:00 when de-
mand response is enabled. Once DR is active, the load reduction on the transformer
is clear, especially notable around 09:20 when demand response reduces the load by
almost 1MW. The strong agreement in behavior from figures 25 and 26, indicates
that the substantial peak at this time of the day was a direct response of AC units
reacting to the rising outside temperature. The absence of this peak during demand
response further supports this observation since the AC power consumption is the
only varying element between both scenarios, indicating that this peak was success-
fully eliminated by demand response. Furthermore, it is also evident from Fig. 26

48
Chapter 4. Simulation Results

(a) Substation Transformer load comparison (b) Residential load comparison between
between baseline and DR case. baseline and DR case.

Fig. 26: Circuit 16 load comparisons between baseline and DR cases. (a) Load on
the substation transformer under uncontrolled conditions & with demand response.
(b) Residential load resulting from baseline and demand response cases.

that the load was reduced accordingly in response to the solar irradiance signal for
the duration of the DR control period. However, an interesting and unexpected
phenomenon is observed after demand response is disabled at 17:00. The residential
load experiences high-amplitude oscillations, known as the demand response rebound
e↵ect, for about two hours before settling to a normal load level with minor oscil-
lations. The first oscillation immediately after stopping DR, results in a spike of
roughly 2.5MW in a matter of 20 minutes, causing a fuse in the middle branch of
the feeder to blow. In addition, a substantial voltage drop across the feeder was
experienced in response to the massive load increase. The voltage profiles shown in
Fig. 27, show a comparison between the voltage at each node across the feeder as a
function of distance from the source node for the baseline and the case with demand
response at the time of peak load (17:20). It is evident from comparing Figs. 27a and
27b that this phenomenon also had substantial e↵ects on the voltage levels in the
entire Circuit 16 feeder. In fact, phase C experienced a voltage of 95.32% of nominal;
within 0.32% of violating the ANSI C84.1 voltage range specifications denoted by
the red lines in Fig. 27.
This phenomenon results from the deactivation of the demand response control. At

49
Chapter 4. Simulation Results

(a) Circuit 16 voltage profile in baseline case. (b) Circuit 16 voltage profile in DR case.

Fig. 27: Circuit 16 voltage profiles as a function of distance from the source node
recorded at 17:20 from baseline and DR cases. Note the red lines denoting the voltage
range limits from ANSI C84.1 specifications. (a) The voltage profile during typical
operating conditions exhibits normal operating voltage levels. (b) Voltage profile
during a wild load oscillation with a peak of about 2.5MW experienced 20 minutes
after DR was disabled.

the end of the control period, almost all of the AC units available for load reduction
in the system are used, as indicated by the green curve on Fig. 28. In other words,
most of the AC units in the system are either already OFF, or they are not allowed
to turn OFF due to their temperature setpoints and constraints. Therefore, once
the AC units are not required to adhere to a control signal, they return to a relaxed
state within a comfortable temperature between their setpoints, similar to when the
load is removed from a loaded spring. Fig. 28 shows that while the number of units
available to turn OFF remains low and relatively constant throughout the control
period, the number of units ON increased. It also shows that right before DR was
disabled, a total of about 400 AC units were active and 1200 were OFF, 800 of which
were available to turn ON. However, only a small fraction of the active units were
available for load reduction. After DR was disabled, 644 of the 800 units available
to turn ON became active. At the 4kW consumption per unit defined in the model,
this amounts to a total load of 2.576MW, which is in strong agreement with the
spike observed in Fig. 26. Moreover, this occurrence coincided with the maximum

50
Chapter 4. Simulation Results

Fig. 28: States of the AC units participating in demand response. The graph shows
the number of AC units that are available to turn ON and OFF, as well as the
number of AC units that are ON throughout the day. Note the spike of units ON
after DR is disabled after 17:00.

outside temperature, which may have led to an amplified e↵ect of this behavior.
The fact that a high number of units became active at the same time means that
most of them, in addition to the ones that were already ON, reached a comfortable
temperature and became inactive at a similar time, thus, leading to the substantial
drop in load observed at 18:00. However, the distribution of setpoint values among
thermostats caused their synchronized operation to normalize to a relatively con-
stant number of operating units, similar to the behavior experienced during the first
three hours the simulation. A potential solution to this behavior is to ramp-down
the removal of demand response control, or to disable control by grouping AC units
and releasing them in waves. Another option is to implement load forecasting and
more sophisticated controls such as an MPC, in order to minimize or eliminate the
negative e↵ects of demand response.
The e↵ects imposed on the voltage experienced by Barranca Mesa after demand re-
sponse was disabled are shown in Fig. 29 in comparison to the voltage history of the
same point obtained during the baseline case. At the time of peak consumption after

51
Chapter 4. Simulation Results

demand response was disabled, Barranca Mesa experienced a voltage of 96.49% of

nominal, a di↵erence of 3% from the case under normal operating conditions. The
results from this simulation further highlight the value of combining control scheme
with power flow properties into a single co-simulation. For instance, simulating de-
mand response without a power flow simulator would neglect the e↵ects of rebound
on the electrical infrastructure. It would be impossible to examine potential voltage
violations or understand the locations of faults in the system. This tool grants the
ability to test the performance of new control algorithms while considering aspects
of human behavior, and to analyze their e↵ects on the grid. It also provides the
ability to pinpoint locations where faults and violations might occur, allowing to
devise potential solutions based on the insight gained from the results, and to test
those solutions using the same simulation framework.

(a) Voltage history of a branch in baseline (b) Voltage history of a branch in demand
case. response case.

Fig. 29: Comparison of voltage history of Barranca Mesa branch during baseline
and DR cases. (a) Voltage history of a node located midway on the branch below
location B3 under baseline conditions. (b) Voltage history of the same node on the
branch under location B3. Note the substantial voltage drop at 17:20 after DR is
disabled.

52
Chapter 4. Simulation Results

C. Battery control & DR

Demand response, as well as the battery controller were used together in this case
to maintain the load at the substation transformer at or below 3.5MW. As in Sec-
tion 4.1.1.B, the initial battery SOC is assumed to be 0.9 at midnight from control
actions performed the previous day. The controller dispatches PV and the battery if
necessary to keep the load on the transformer at the target value. The same control
of 1600 AC residential units via demand response is used here with the same control
schedule between 9:00AM-5:00PM, with the same control signal created from solar
irradiance. As a consequence, the residential load remained identical to the load seen
in Section 4.1.2.B including the oscillating phenomenon experienced after demand
response is disabled. Here, the main grid supplies power to the feeder, while the
demand response lowers the residential consumption according to its control signal,

Fig. 30: History of load on substation transformer during support by the battery
controller and residential demand response. Residential load, battery dispatch, and
PV generation are also shown for comparison. Battery control successfully dispatches
PV and battery power to maintain the load on the feeder at the target load of 3.5MW,
until the load increase experienced after the DR control period drains the battery to
its reserve SOC of 25%.

53
Chapter 4. Simulation Results

and the controller minimizes the di↵erence between the load on the transformer and
the target load by dispatching PV and battery resources.
The overall load on the substation transformer in comparison to the residential and
battery load profiles including PV power generation are shown in Fig. 30. Note that
the PV output and the residential profiles are identical to cases in the previous sec-
tions. However, battery output is varied by the controller according to the power
needed on the feeder to maintain the load at the target level. Fig. 30 shows that
the load on the transformer remains below the predefined threshold of 3.5MW until
09:00 when the load begins to increase as a result of the typical morning peak. From
this point forward, the controller e↵ectively dispatches PV and battery resources in
order to maintain the load at the target value for a period of over eight hours. How-
ever, the increased power consumption associated with the synchronized activation
of the 644 AC units after demand response is disabled overpowers the support DER
support. Dictated by the lack of sunlight at this time of the day, only the battery
is dispatched, in order to counter the e↵ects of the sudden load increase experienced
by the transformer. The battery sends power to the feeder at 2MW for almost 20
minutes until it reaches its reserve SOC of 25%, allowing the load to only reach about
5MW. This is more than 1.5MW less than in the previous case at the same point in
time. Once the battery reaches its reserve SOC level and becomes inactive, the load
on the transformer spikes up again due to the high consumption of the AC units that
remains. However, the highest load experienced during the entire event is 1MW less
than that experienced during the case without battery control.
While the DERs in this scenario were not enough to mitigate the e↵ects of the de-
mand response rebound, the action of the battery was enough to lessen the disruption
to the feeder voltage levels at the time of highest consumption that took place at
17:20. The voltage at every node in the residential demand response only case dis-
cussed in the previous section, in comparison to the voltage profile obtained in this
scenario are shown in Fig. 31. With battery support managed by the controller, the

54
Chapter 4. Simulation Results

(a) Circuit 16 voltage profile in demand re- (b) Circuit 16 voltage profile in DR & bat-
sponse case. tery control case.

Fig. 31: Circuit 16 voltage profiles as a function of distance from the source node
recorded at 17:20 from DR-only and DR & battery control cases. Note the red lines
denoting the voltage range limits from ANSI C84.1 specifications. (a) Voltage profile
during consumption spike of about 2.5MW experienced 20 minutes after DR was
disabled during DR only case. (b) Voltage profile at same time of peak consumption
after DR was disabled with battery support. The lowest voltage experienced with
battery support was 96.14% compared to 95.32% during DR only case.

lowest voltage value experienced on the feeder was recorded on phase C at 96.14%
of the nominal voltage. This is a 0.82% di↵erence from the lowest voltage value
observed in Section 4.1.2.B (95.32% of nominal). A similar impact was felt in the
Barranca Mesa branch during demand response rebound with battery support, as
shown in Fig. 32. While the voltage drop remains substantial, it is noticeably less
drastic when compared to the unsupported case. Furthermore, the lowest voltage
value experienced at this location on the feeder was recorded at 97.33% of the nom-
inal voltage, a di↵erence of 0.84% and 2.17% from the case with demand response
only and the case under normal operating conditions, respectively.
The results of this simulation clearly indicate that it is possible to reduce the dis-
ruption caused by demand response rebound, or other similar events, by employing
battery support with more sophisticated controls. For example, a load forecaster or
MPC would take data from this simulation such as the load on the feeder and battery
SOC, to predict similar behavior in the future and optimize the use of resources to

55
Chapter 4. Simulation Results

(a) Barranca Mesa voltage history during (b) Barranca Mesa voltage history during
DR-only case. DR & battery control support.

Fig. 32: Comparison of impact imposed on Barranca Mesa voltage after deactivation
of demand response during DR-only and DR & battery control cases. (a) Barranca
Mesa voltage with only demand response support. Voltage levels drop to 95.32%,
almost violating ANSI C84.1 voltage limit specifications. (b) Voltage experienced by
Barranca Mesa during support by demand response and battery control. The voltage
drop in this case is remarkably less drastic due to added support by the battery.

minimize the e↵ects associated with disruptive events. Additionally, tools such as
Fig. 31 provide the ability to identify sections of a feeder at risk of violating voltage
limit specifications. In this case, it is evident that the location of the feeder corre-
sponding to the nodes on phase C that experience the lowest voltage levels, would
benefit from injection of reactive power or additional support from DERs. In fact, it
is theoretically possible to generate a constant voltage profile across the feeder with
enough control and DER support at strategic locations on the system. This valuable
insight would remain obscure without the ability of this framework to analyze how
power flow is a↵ected by certain control schemes.

56
Chapter 4. Simulation Results

4.2 Real-time simulation

The performance of the real-time simulation setup described in Section 3.2 is intro-
duced here. The residential load generator is used to simulate the residential load
of the 200-home community located in Mesa del Sol. Real-time residential demand
response capabilities are enabled by the load aggregator, who controls the aggregated
power consumption of the air conditioner system in each home using control signals
provided by a microgrid controller. The controller optimizes the use of all resources
in the system discussed in Section 1.3 for the minimum cost of operation. However,
the requests from the microgrid controller to the aggregator, were not very eventful.
The setpoint for the TCL generally remained at zero due to the existence of more
cost-e↵ective resources that were used to satisfy the demand. Therefore, a synthetic
signal was created to examine the real-time demand response behavior of the system.

600000 400000
filtered AC consumption filtered AC consumption
filtered setpoint filtered setpoint
500000 aggregated AC consumption aggregated AC consumption
setpoint setpoint
300000
400000

300000 200000
power (W)

power (W)

200000
100000
100000

0 0

-100000
-100000
-200000

-300000 -200000
14 16 18 20 22 00 02 04 06 08 10 12 15 16 17 18
time (hours) time (hours)

(a) Performance of DR control during a 22- (b) Closeup at DR performance during a 3-

hour test period. hour period.

Fig. 33: Performance of real-time residential demand response during a test period
of 22 hours using a synthetic control tracking signal. (a) Aggregated TCL reaction
to demand response control. The response of the system is as expected; the filtered
TCL response tracks the filtered signal over the course of the entire control period.
(b) Focused view at the performance of DR control during 3 hours of operation.
Note the inability to track the control signal accurately after hour 17, due to lack of
available resources.

57
Chapter 4. Simulation Results

The signal created varies hourly with requests to decrease and increase power con-
sumption at di↵erent magnitudes for a period of nine hours. The setpoint remains
at zero for the following 13 hours, for a total test period of 22 hours as shown in
Fig. 33a. It is evident from inspecting Fig. 33a, that the TCL resource e↵ectively
follows the filtered control signal for the majority of the test period. However, in cer-
tain cases the TCL’s “thermal battery” runs out of energy preventing the TCL from
fulfilling the control request, similar to 12:00 in Fig. 25. This behavior is observed
in hour 17:00 of Fig. 33b where a requests to decrease power consumption cannot
be satisfied as it would force house temperatures to increase above the deadband
compromising the customer’s comfort. This drawback is common among demand
response schemes in which residential customers participate with the condition that

800000 0.8
broadcast signal aggregated AC consumption
setpoint
600000 0.6

400000 0.4

200000 0.2
power (W)

signal
0 0

-200000 -0.2

-400000 -0.4

-600000 -0.6

-800000 -0.8
13 14 15 16 17 18
time (hours)

Fig. 34: Example of control signals and response of residential demand response
control. The graph shows the synthetic control signal sent by the microgrid controller
to the aggregator, with the associated control signal broadcasted by the aggregator to
the participating TCL units. The aggregated response of the system to the broadcast
is represented by the green line. Note the shifts of consumption in response to the
control signals.

58
Chapter 4. Simulation Results

quality of service be the priority. Moreover, the negative impact of this limitation
could be minimized with a more complex policy where financial incentives are of-
fered to flexible customers who are willing to accept remote controllability of their
temperature setpoint.
The relationship between the control signal issued by the microgrid controller, the
signal broadcasted by the load aggregator to the TCL units, and their response is
illustrated in Fig. 34. It shows the setpoint control signal issued by the microgrid
controller, the broadcast signal sent by the aggregator to the TCLs and their aggre-
gated response of the system. A closer inspection of Fig. 34 reveals the relationship
between the data exchanged by the controller and the aggregator, and the response
of the TCLs. The broadcast signal displays a heavy bias towards the control signal
as expected, especially noticeable during hours 14, 16, and 17.
The real-time results from the simulation, as explained in Section 3.2, are sent to a
central database by each model. These results are showcased at a publicly available

Fig. 35: Simulated residential load used in the MRI real-time simulation. The load is
simulated by the load generator and used as a proxy to represent the power consump-
tion of a 200-home residential community in a microgrid controller study performed
on a feeder in the Mesa del Sol community in Albuquerque, NM. This figure was
generated using real-time data obtained from the study’s database and posted on a
UNM website located at 129.24.68.15.

59
Chapter 4. Simulation Results

research website located at the IP address 129.24.68.15. Results include the total
load of the virtual Mesa del Sol residential community and other resources on the
system, such as PV arrays, batteries, fuel cell, and other components that lie beyond
the scope of this work. A snapshot of the community’s power consumption obtained
from the website is shown in Fig. 35. The graph illustrates the load history of the res-
idential community in a period of two hours between 7:30-9:30AM with five-minute
resolution. Note that the graph shows the morning ramp-down that takes place after
people leave their homes for work.
Here, the ability to simulate real-life communication over a network between residen-
tial infrastructure and a control center was successfully demonstrated. It was shown
that the physical separation of these components is negligible and that remote con-
trol of residential loads is possible using the internet and internet-enabled appliances
such as smart thermostats. The successful integration of these components into the
real-time microgrid controller simulation with HIL equipment demonstrates the abil-
ity to use this framework as a powerful tool to conduct research that considers the
e↵ects of control algorithms on existing electric infrastructure.

60
Chapter 5

Discussion & Conclusions

5.1 GridLAB-D co-simulation

The main purpose of the GridLAB-D co-simulation framework described in this the-
sis, is to provide the means for collaboration between a diverse group of models that
are originally standalone, in order to achieve a combined realistic representation of a
power distribution system’s behavior. This is accomplished by providing a platform
with a direct line of communication and mutual synchronized collaboration between
models. The platform enables dynamic results that evolve in response to the interac-
tions between co-simulators, while each model operates within its individual domain.
Here, the co-simulation framework combines the GridLAB-D power flow engine to
simulate the distribution of power within a feeder, with the abilities of a residential
load generator to accurately simulate the residential load of a community based on
human behavior models, and to perform demand response control of household ap-
pliances using a load aggregator.
The Matlab interface co-simulation setup demonstrated the potential of the frame-
work by successfully incorporating human behavior models with the GridLAB-D

61
Chapter 5. Discussion & Conclusions

power distribution simulator. It set a benchmark performance of the models, and

also provided proof of concept for the GridLAB-D built-in battery controller and
demand response capabilities of the load aggregator. However, the Matlab inter-
face was unreliable and it permitted only a small amount of data to be transferred
between models. It also required a vast amount of computational resources that re-
sulted in extremely slow execution times. These limitations made clear the necessity
for a more robust and powerful replacement of the Matlab interface.
Developed by Pacific Northwest National Laboratory, FNCS was designed with the
sole purpose to serve as a co-simulation interface for software that cannot operate
satisfactorily together under typical computer conditions. FNCS enabled the ability
to transfer more than 200 times the amount of data between co-simulators, twice
as fast execution times, and more efficient and reliable simulation executions. The
results from the FNCS baseline simulation showed similar results to the uncontrolled
scenario obtained with the Matlab interface, indicating that the simulated infrastruc-
ture behaved within physical limits of the system. Furthermore, the enhanced capa-
bilities of the FNCS interface demonstrated the ability to employ demand response
control of an entire feeder in combination with utility-scale battery controls with the
objective to reduce the stress on the power distribution infrastructure. While the
design of a demand response infrastructure is beyond the scope of this work, the
capabilities to simulate control of an aggregated system of household appliances was
successfully demonstrated. Also, although the unexpected behavior associated with
the deactivation of demand response control was analyzed and discussed, an in-depth
examination of this phenomenon is aligned with the development of demand response
infrastructure, and thus, it too lies outside of the scope of this thesis. However, the
simulation framework showcased here has the potential to be a powerful tool in the
development and design of future demand response algorithms and infrastructure, in
an e↵ort to improve the resiliency of the power grid.

62
Chapter 5. Discussion & Conclusions

5.2 Real-time simulation

The development of modern microgrid controllers, such as the one demonstrated in

the Mesa del Sol area in Albuquerque, NM, requires the ability to test the system’s
behavior under di↵erent circumstances. Testing is especially crucial in the develop-
ment of demand response algorithms due to the risks involved in deploying systems
that have the potential to disrupt the quality of service experienced by the customers
if not designed properly. This requires testing communities with real smart appli-
ances and the infrastructure necessary to conduct test. The high upfront cost and
risks associated with this type of testing method means that it is often avoided. In
addition, since this technology is relatively young, in most cases the communities
are not equipped with demand response capabilities. A more cost e↵ective and con-
venient method is to perform computer simulations using virtual models to analyze
the behavior of a system before physically deploying resources, as was done for the
Mesa del Sol microgrid controller discussed in Section 1.3.
Here, the residential load generator and aggregator were incorporated into a real-
time simulation that tests the capabilities of the micogrid controller to manage and
optimize the use of DERs on a distribution feeder. The simulation incorporates both
virtual and real elements including a commercial-scale microgrid, utility-scale PV
and BESS, residential demand response, and VPP and MPC for optimization. The
load generator was used as a proxy for the 200-home Mesa del Sol community with
simulated demand response capabilities implemented with the use of the aggrega-
tor. Both of these models were successfully installed onto individual Raspberry Pi
computers to simulate the physical separation of the residential infrastructure and
load aggregator in real life. In addition, communication between the models and the
microgrid controller database was employed via the Internet, successfully replicating
communication of similar devices in the field. The results discussed in Section 4.2
show the e↵ective performance of the aggregated control and response of the TCL

63
Chapter 5. Discussion & Conclusions

system to the control signal issued by the microgrid controller for a period of nine
hours. The HIL real-time simulation platform showcased in this thesis demonstrates
the ability to incorporate existing distributed energy resources within a partially
virtual environment that fosters the development of future infrastructure essential
for modern microgrids. Most importantly, this platform enables the ability to test
new microgrid controller algorithms without risking equipment damage or quality
of service to the customer. At the same time, it also provides the ability to gain
valuable insight related to the introduction of new elements and their impact on ex-
isting electric infrastructure. These capabilities are essential in the development of
microgrid technology that will improve grid resilience by providing critical services
in the event of severe disruptive events.

5.3 Future Work

While the models and simulation frameworks described in this work generated suc-
cessful results, improvements to some of the components is still necessary to make
them more e↵ective. These include, enhancements to the residential load generator
that simulate changes in human behavior due to outside influence. In this case, be-
havior demand response will be introduced to replicate economic incentives granted
by the utility to customers participating in load reduction programs. Additional
control testing is also required alongside the residential load aggregator to determine
an appropriate solution to rebound behavior experienced after demand response con-
trol was disabled. A potential solution to this problem is the introduction of a load
forecaster or MPC to predict disruptive events while optimizing the use of resources.
The addition of a communication infrastructure simulator to the FNCS co-simulation
framework is also on the scope of future work for this research project. The ability
to combine aspects of communication to the framework will allow to consider char-

64
Chapter 5. Discussion & Conclusions

acteristics of demand response communication between smart appliances and the

aggregator that are currently neglected. This addition will also allow to consider
aspects of cybersecurity associated with the power grid.
Currently, the GridLAB-D built-in battery controller is only capable to charge from
power generated by an accompanying PV array through a shared inverter. Therefore,
in order to represent real-life system configurations and capabilities, the battery con-
troller should include the ability to charge from the grid as well as directly from PV
through individual inverters. However, its performance illustrated in Sections 4.1.1.B
and 4.1.2.C confirmed that the control logic used for dispatching PV and battery re-
sources is e↵ective.
The real-time communication method between the load generator and the aggre-
gator via the Internet also requires improvements. Considered “brute force,” the
current communication method is secure, but very inefficient. The models exchange
information by copying and pasting files between machines over a network, and are
opened and read on every time-step. This could be improved by implementing a
standardized communication protocol such as DNP3, which also provide high levels
of security, comparable to the Secure Copy command line tool.

5.4 Conclusions

If he were alive, Thomas Edison would recognize most of the power grid infrastructure
used today despite the technological advances made to electric power infrastructure
over the last 150 years. For instance, most of the world still distributes electricity
from power stations to its customers over power lines mounted on wooden electric
poles, much like in the days of Edison. However, the power grid is currently under-
going a major transformation due to recent technological advances, especially in the
areas of computer science, telecommunications, and electrical engineering, as well as

65
Chapter 5. Discussion & Conclusions

an increase penetration of distributed energy resources that is driving the greatest

leap forward in the history of the power industry. This is forcing industry stake-
holders to adapt new business models that modernize the power grid infrastructure
in a way that considers endusers not only as customers, but also as potential con-
tributors to the operations and management of power grid resources. The increasing
penetration of residential DERs, as well as emerging internet enabled appliances
such as smart thermostats, indicates that residential communities can be considered
low-scale power plants with capabilities to o↵set loads using demand response on
their appliances. However, the standards and guidelines in existence today require
further research and development before full deployment of such infrastructure is
feasible. Also, the development of this technology demands studies that consider
human behavior to examine the e↵ects of the human factor in distribution-level load
management and demand response.
The work presented in this thesis highlights two co-simulation platforms that com-
bine distribution system simulators with a residential load synthesis framework based
on human behavior models. The results obtained from the various simulations per-
formed with each platform successfully established their performance and showed
their potential as tools to conduct research in the e↵orts to develop a modernized
power grid infrastructure. Even though improvements to the models are necessary
for enhanced behavior of the systems, the co-simulation platforms’ performance was
e↵ective, unlocking endless possibilities for future co-simulations that combine addi-
tional models regardless of programming language or compatibility. As the electric
power industry evolves with innovation, it will require similar innovative methods to
study and develop the technology necessary to create the power grid of the future,
one that is more efficient, reliable, resilient, and smart.

66
Appendices

A Matlab interface models & code 68

B FNCS interface models & code 69

C Real-time simulation models & code 70

67
Appendix A

Matlab interface models & code

The models and related files used in the co-simulations performed with the Matlab
interface can be found in the following github repository.
https://github.com/victorayon/masters-thesis/tree/master/Matlab-interface

68
Appendix B

FNCS interface models & code

The models and related files used in the co-simulations performed with the FNCS
interface can be found in the following github repository.
https://github.com/victorayon/masters-thesis/tree/master/FNCS-interface

69
Appendix C

Real-time simulation models &

code

The models and related files used in the real-time simulation can be found in the
following github repository.
https://github.com/victorayon/masters-thesis/tree/master/Real-time-simulation

70
References

[1] V. Ayon, M. Robinson, A. Mammoli, A. Fisher, and J. Fuller, “Integration of

Bottom-up Statistical Models of Loads on a Residential Feeder with the GridLAB-
D Distribution System Simulator, and Applications,” Innovative Smart Grid
Technologies (ISGT) Europe, IEEE, Sep. 2017.

[2] V. Ayon, M. Robinson, and A. Mammoli, “Simulation of Real-time Demand-

response to Support a Residential Distribution Feeder in Microgrid Mode,” Grid
of the Future, CIGRE US National Committee, Oct. 2017.

[3] D. P. Chassin, J. C. Fuller, and N. Djilali, “Gridlab-d: An agent-based simulation

framework for smart grids,” Journal of Applied Mathematics, vol. 2014, 2014.

[4] “Matlab link,” http://gridlab-d.shoutwiki.com/wiki/Matlab_link, ac-

cessed: 2017-09-26.

[5] “Chapter 9 - Interfacing with External Software,” http://gridlab-d.

shoutwiki.com/wiki/Chapter_9_-_Interfacing_with_External_Software,
accessed: 2017-08-18

[6] D. Ton and J. Reilly, “Microgrid Controller Initiatives,” IEEE Power & Energy,
no. 4, vol. 15, pp. 24-25, Jun. 2017

[7] A. Mammoli et al., “Validation of Measures to Expand Applications of Micro-Grid

Control System,” University of New Mexico & Sandia National Laboratories, for
Mitsubishi Research Institute, Sep. 2017.

[8] “Milsoft engineering analysis WindMil,” https://www.milsoft.com/

utilitysolutions/upgrades/milsoft-engineering-analysis-ea-windmil,
accessed: 2017-03-11.

[9] “C84. 1-1982,” American National Standard for Electric Power Systems and
Equipment-Voltage Ratings (60Hz), aNSI Standard.

71
References

[10] A. Mammoli, M. Robinson, V. Ayon, M. Hombrados-Herrera, M. Martinez-

Ramon, “A simulation framework to develop control and forecasting tools for
aggregated residential energy resources,” Applied Energy, 2017, under review.

[11] R. Hendron and C. Engebrecht, Building America house simulation protocols.

National Renewable Energy Laboratory Golden, CO, 2010.

[12] D. Fischer, A. Härtl, and B. Wille-Haussmann “Model for electric load profiles
with high time resolution for german households,” Energy and Buildings, vol. 92,
pp. 170-179, 2015.

[13] J. Widén, A. M. Nilsson, and E. Wäckelgård, “A combined markovchain and

bottom-up approach to modelling of domestic lighting demand,” Energy and
Buildings, vol. 41, no. 10, pp. 1001-1012, 2009.

[14] W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery, “Digital Filter-
ing in the Time Domain,” in Numerical Recipes in Fortran, 2nd ed. Cambridge,
England: CUP, 1986-1992, ch.13, sec. 5, pp. 551-556

[15] “FNCS-Tutorial/demo-large/,” https://github.com/GridOPTICS/

FNCS-Tutorial/tree/master/demo-large, accessed: 2017-08-20.

[16] “1. WHAT IS A RASPBERRY PI?” https://www.raspberrypi.org/help/

faqs/#introWhatIs, accessed: 2017-08-05.

[17] “12 scp command examples to transfer files on Linux” https:

//www.linux.com/blog/12-scp-command-examples-transfer-files-linux,
accessed: 2017-08-05.

[18] “SSH COMMAND” https://www.ssh.com/ssh/command/, accessed: 2017-08-

05.

[19] O. Lavrova, F. Cheng, S. Abdollahy, H. Barsun, A. Mammoli, D. Dreisigmayer,

S. Willard, B. Arellano, and C. Van Zeyl, “Analysis of battery storage utilization
for load shifting and peak smoothing on a distribution feeder in new mexico,” in
Innovative Smart Grid Technologies (ISGT), 2012 IEEE PES. IEEE, 2012, pp.
1?6.

[20] F. Cheng, S. Willard, J. Hawkins, B. Arellano, O. Lavrova, and A. Mammoli,

“Applying battery energy storage to enhance the benefits of photovoltaics,” in
Energytech, 2012 IEEE. IEEE, 2012, pp. 1?5.

72