Sub-track II –

Power & Energy Presentation

1
AI based smart solar
arc-fault circuit
interrupter solution

Matteo MARAVITA
STMicroelectronics
 ST reference designs for smart solar systems

Smart
Monitor
meter

Inverter AC to grid

MPPT
INV.
AFCI
RSD
Bidirect PV micro-inverter (PLC)
MPPT DC-DC
MLPE (module-level power electronic) on AC
Communication
DC to battery

PV monitor (PLC) Inverter string optimizer (MPPT) Inverter arc-fault circuit

interrupter (AFCI)

PV optimizer (PLC) Inverter DC-AC converter Inverter rapid shutdown (DC PLC)

PV rapid shutdown (PLC) Inverter bidirectional DC-DC Inverter monitor (AC PLC)

MLPE (module-level power electronics) on DC Inverter side modules (string-inverter as example)

3
 The risks of arc-fault
Factors that may cause arcs

• Aging
• Weather variation
• Harsh environment
• Human factors
Possibly
1500 V/10 A DC

Inverter

4
 Arc-fault detection standard–UL1699B

Standard for safety

• Photovoltaic (PV) DC arc-fault circuit protection

5
 STM32 + AI detector
Field proven and future oriented system for AFCI

Inverter
BOOST
MPPT
#1

AFCI
BOOST Channel 1
MPPT Host-

Neural network
interface
#2 Channel 2 to DSP
Detection
DC-AC
Channel 3 Algorithm
BOOST
To RSD
MPPT Channel 4
#3

RSD
BOOST
MPPT
#4

6
 Challenges in AFCI system design
Inverter noise
36 kHz 72 kHz

Noise frequency generated by the inverter

submerges arc frequency signal.

Weather impact

Rapid shutdown
131 kHz 143 kHz

7
 Possible issues in AFCI data processing
Poor connection

Model mis-match Data storage mis-match

Customer
prototype
Inverter noise
HW/FW
Frame-loss by USB

Board noise or crosstalk

Filter-chain performance

Transformer saturation

Field recording issue example

8
 ST reference design for AFCI
Turnkey HW/SW reference design Turnkey STM32 embedded AI

All STM32
MCUs

9
AFCI HW/FW reference design
 AFCI board hardware block diagram
microSD

ADC To PC (micro-USB)
ADC
STM32H7B3
SPI

Filter by TSV772 (30-100 KHz), 4th order ADC

From Filter by TSV772 (30-200 KHz), 2nd order ADC To PC (micro-USB)

STRIN
Filter by TSV772 (30-100 KHz), 2nd order ADC
G input STM32G473
Filter by TSV772 (30-100 KHz), STM32 OP ADC To host DSP
UAR
T
DAC output DAC
UAR
AFCI board T

RSD comm.
module
• The primary target of STM32H7B3 is to drive microSD card.
11
• STM32G473 or STM32H7B3 should be sufficient for customer AFCI product.
 AFCI hardware board v1.0
OP: TSV772 RESET button
MCU: STM32H7B3
Input CH4
Input CH2 for LED5
STM32H7 LED4

Micro-USB to PC for
STM32H7B3
Input CH3 (optional)
Input CH1 for
STM32H7 MCU: STM32G473

Input CH2 Micro-USB to PC

for STM32G473

Input CH1 DC-input (5 V)

(Optional)

Output
(player)

LED2
12
LED3
 AFCI filter design
The 2nd band stop filters are bypassed in current board by jumpers

2nd order, To ADC

RC band Amplifier Band stop Band stop 2nd order RC
CH1 CT filter x11
HPF 30 kHz,
18 kHz optional 100 kHz LPF LPF
MCU OP

RC band Amplifier 2nd order Band stop 2nd order RC 2nd order, HPF To ADC
CH2 CT filter x11 HPF 30 kHz 19-20 kHz 100 kHz LPF LPF 30 kHz, MCU OP

RC band Amplifier 2nd order Band stop Band stop 2nd order RC MCU OP To ADC
CH3 CT filter x11 HPF 30 kHz 18 kHz optional 200 kHz LPF LPF voltage follower

CH4 RC band Amplifier 4th order Band stop Band stop 4th order RC MCU OP To ADC
CT filter x11 HPF 30 kHz 19-20 kHz optional 100 kHz LPF LPF voltage follower

13
 AFCI filter design

Frequency response of hardware filters

CH1 CH2

18 kHz 19 kHz

CH3 CH4

18 kHz 19 kHz

14
 Connection in system

WIN32 console application:

• Recorder mode: mainly used for data collection for model training.
• Detector mode: mainly used for NN model accuracy verification or ARC
detection.
• Control mode: mainly used for firmware update, reset device, stop current
mode etc.
• Others …

AFCI GUI tool

15
 AFCI solution use-cases
Solution function bock: In-lab evaluation mode:
• Install in customer system for evaluation and competitor comparison.
• Enable parallel mode (4 channels), enable rapid shutdown (RSD) on arc-fault detected.
DC-in USB-A RSD UART
Channel 1 To host DSP

Channel 1 Host UART

Host UART
Channel 2

AFCI board Channel 2

AFCI RSD UART
Channel 3
Micro USB Channel 3 board
Micro USB
Channel 4
Models
Channel 4

Data gathering mode: In-field evaluation mode:

• Read current, and possible voltage, and send to PC through USB. • Support customer field-trial.
• AFCI board powered by PC USB. • Record special data for further analysis.
From inverter

To host DSP
Channel 1
Channel 1 DC-in Host UART

Channel 2
AFCI Micro USB Channel 2
AFCI RSD UART
Channel 3 board
Channel 3 board
USB
Models
Channel 4 microSD
Channel 4

16
 AFCI GUI tool makes development easier

• Support detector mode

• Support recorder mode
• Support control command
• Support NVM configuration
• Support logging

17
 ST arc fault detection lab
Main instruments in AFCI lab

1. DC electronic load.
2. Arc generator
3. DC power supply

18
AI reference design
 Overall AI development workflow

1 Data preparation 2 Data science 3 Model implementation

Model selection Model library

Data acquisition Data processing Model validation Model inference
and training creation

Automated edge AI software

All STM32
Edge AI toolkit MCUs

20
 Step by step for AFCI project

Step 1 Step 2 Step 3 Step 4 Step 5 Step 6

Confirm AI model
Collect dataset Signal check Model training Verify in lab
requirements implementation

Collect dataset Test AI model

Check custom Check signal Train model Implement h5
by using AFCI result on MCU
application quality by using with different file into MCU
v1.0 board and edge side and
scenarios with signal check current dataset, side by using
WIN32 console check if FP or
our expert tool get h5 file CubeAI
tool FN.

Several rounds data collection to reach good performance

21
 ST packages for AFCI

AFCI HW v1.0 WIN32 console Signal check AFCI AI model AFCI FW project
board application tool (exe file) training for STM32
Board including • Recorder (exe file) • Datalogging
STM32G4 and • Detector For analyzing For training the
• AI model real
STM32H7 for data • FW update signal quality. dataset for AI
time processing
collection and model.
testing. • Post-processing

User manual and guidelines

22
 Signal checking
• Checking signal with signal check tool is very important before training neural network.
Good examples

Normal signal Arc signal

Bad examples

23
 AI model training
• Collecting dataset by using Labeled data Neural network
AFCI GUI with two classes. architecture
Class1
Training set (Normal)
• Normal status SoftMax

• Arc status Validation set

Class2
• Split dataset Test set
(Abnormal)

• Training set
• Validation set
• Test set

24
 AI model implementation
WIN32 STM32 STM32

AI library (.a) UART driver

Model file (.h5)
Evaluation Host-Interface2 Remote control (to RSD
Host-Interface1
(To customer DSP, mater, wireless
project (To Win-PC)
customer protocol) modules, etc.)

AI training
Data buffering
Application AI model
Managemen AI library (.a)
live-upgrade
t
Sample data
AI validation storage

USB-Key
NVM ADC + DMA Sampling
driver

Runtime on STM32

Input data
Arc detection
from CT AI model result Postprocessing
result
sensor
25
 Takeaways
The AI based arc fault circuit interrupter (AFCI) contributes the safe
and sustainable development

Turnkey ref. design to deploy state-of-art AFCI solution efficiently

Deep-learning-based classifier to address high detection accuracy

High performance edge computation with STM32G4 or STM32H7

Flexible and accurate signal processing with TSV7xx amplifiers

Mature support materials and process with experienced team

26
27
 Find out more at www.st.com

© STMicroelectronics - All rights reserved.

ST logo is a trademark or a registered trademark of STMicroelectronics International NV or its affiliates in the EU and/or other countries.
For additional information about ST trademarks, please refer to www.st.com/trademarks.
All other product or service names are the property of their respective owners.