2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC)

Banff Center, Banff, Canada, October 5-8, 2017

Long Short Term Memory Networks for Short-Term

Electric Load Forecasting
Apurva Narayan∗ and Keith W. Hipel†
∗
Department of Electrical and Computer Engineering,
† Department of Systems Design Engineering,

University of Waterloo, Waterloo, ON Canada N2L 3G1

Email: ∗ [email protected], † [email protected]

Abstract—Short-term electricity demand forecasting is critical wavelets [5] have been deployed and shown encouraging
to utility companies. It plays a key role in the operation of results. Among these, ANNs have performed well given their
power industry. It becomes all the more important and critical ability to handle the non-linear relationships between the load
with increasing penetration of renewable energy sources. Short- and various underlying factors. The demand profiles are gen-
term load forecasting enables power companies to make informed erally taken as sequences over time. Hence, it is a sequence-
business decisions in real-time. Demand patterns are extremely
prediction task given only the historical and current infor-
complex due to market deregulation and other environmental
factors. Although there has been extensive research in the area mation about the sequence. Recently, another class of ANN
of short-term electrical load forecasting, difficulties in implemen- models have emerged, often referred to as deep learning. These
tation and lack of transparency in results has been cited as a models are similar to ANN with increased depth and spe-
main challenge. cialized approaches for training these networks are currently
being developed. Most of these models fall under the category
Deep neural architectures have recently shown their ability
to mine complex underlying patterns in various domains. In our of unsupervised learning. Unsupervised feature learning [6]
work, we present a deep recurrent neural architecture to unearth performs this function by learning feature representations from
the complex patterns underlying the regional demand profiles unlabeled data. Usually, these layers of representations are
without specific insights from the utilities. The model learns from stacked to create deep networks that are capable of modeling
historical data patterns. We show that deep recurrent neural complex structures in the data. Unsupervised feature learning
network with long-short term memory architecture presents a and deep learning have presented a success story for feature
robust methodology for accurate short term load forecasting with representation for static data [1].
the ability to adapt and learn the underlying complex features
over time. In most cases it matches the performance of the latest In the context of short-term electrical load forecasting
state-of-the-art techniques and even supercedes it in a few cases. (hourly time scale or even shorter) the problem can be
considered as a sequence prediction problem. A specialized
I. I NTRODUCTION ANN architecture referred to as recurrent neural network
Short-term load forecasting is a critical component for (RNN) has been successful in predicting sequences accurately.
the reliable and secure operation of large power systems. Moreover, RNNs have advanced to an extent where they have
High accuracy in load forecasting for power systems improves memory which has the ability to learn as the data arrives and
efficiency of operation. The power systems are required to have shown considerable success in the domain of learning
operate not only efficiently by utilizing the available resources relationships and text prediction based on context and are
but also by maintaining system security and reliability. Load called the Long-Short-Term Memories (LSTM). Electrical load
forecasting is crucial for power companies in meeting their can be considered as a sequence (time series) which can be
objectives. modeled using LSTM to accurately predict the future demand.
The problem of short-term load forecasting has tradition- In this paper we develop a LSTM-Recurrent Neural Net-
ally been viewed as a modeling problem, where load has to work for short term electrical load forecasting. The model is
be modeled as a function of time of day, day of the week, general enough and can be adapted for other time series where
weather, and other social factors. The majority of operational sufficient historical data is available. The results obtained are
and control decisions such as load dispatch, reliability analysis, promising and present the generalized nature of LSTMs for
and maintenance planning are all based on the load forecasts. time series forecasting problems.
Electrical load can be represented as a time-series. Time- The remainder of the paper is organized as follows, Section
series analysis is a vast field for sequential data analysis. II presents a literature review of the recent work in electrical
Every data set has unique properties associated with their load forecasting. In Section III, we present a deep learning
source that make them challenging to analyze and model. neural network architecture for short-term electrical load fore-
Pre-existing assumptions about data, such as noise levels, casting. In Section IV, we explain the various experiments.
redundancies, and temporal dependencies are often used in Lastly, in Section V we present our conclusions and directions
the chosen model or feature representation [1]. The alternate for future research.
approach to using hand-crafted features is to automatically II. BACKGROUND
learn them by interacting with the data.
Short-term load forecasting is a well studied problem. Ac-
Modern machine learning techniques, such as expert curate forecasts for loads are useful in the planning and opera-
systems [2], Artificial Neural Network (ANN) [3], [4], tion of large [7] and micro-power systems [8]. There have been

978-1-5386-1645-1/17/$31.00 ©2017 IEEE 2573

Authorized licensed use limited to: The University of British Columbia Library. Downloaded on January 14,2022 at 21:27:43 UTC from IEEE Xplore. Restrictions apply.
numerous models proposed for accurate modeling and predic- mented for short term electrical load forecasting, which is
tion of demand. Electricity demand forecasting is considered a a challenging task given the complex and large number of
time-series modeling problem [9]. Various time series models underlying features affecting the process. Recent advances in
such as auto regressive moving average (ARMA), generalized deep learning have proven to be useful in pattern identification
auto regressive conditional heteroskedasticity (GARCH), and for such complex cases. Till now no literature can be found
intervention time series models have been used to model regarding deep learning algorithm being used for modeling
electrical load forecasting [10], [11]. Artificial neural networks and forecasting short term electrical load at the regional level
(ANN) and other machine learning algorithms have proven for varying time scales. Our work uses the fundamentals of
successful in various tasks such as classification, regression deep learning to automatically identify the complex features in
and time series modeling [12]. Besides ANN, support vector electrical load and predict accurately and robustly at varying
regression is known as a strong predictor for achieving global time scales using LSTM-recurrent neural networks.
optimum solutions.
It is quite evident that no research has been done in
In [13], the authors use support vector regression (SVR) for short term load forecasting using deep learning techniques.
short-term load forecasting with two additional improvements The power and expressibility of these techniques still remains
in procedure for generation of model inputs and subsequent untouched. Our work here tries to explore the power of deep
model input selection using feature selection algorithms. An- networks in context of short term electrical load forecasting.
other recent work [13] combines price and load forecasting
using a hybrid time-series and adaptive wavelet neural net- III. R ECURRENT N EURAL N ETWORK
work. In [14], they evaluate the effectiveness of some of the
newest designed algorithms in machine learning to train typical Recurrent neural networks with Long Short-Term Memory
radial basis function (RBF) networks for 24-h electric load have emerged as a reliable tool for sequential data series
forecasting: support vector regression (SVR), extreme learning modeling, analysis, and forecasting [22]. Usually, techniques
machines (ELMs), and decay RBF neural networks (DRNNs). solving problems associated with sequential data such as
A comprehensive review of various tools for short term load language, and audio etc. used hand-crafted features. LSTMs
forecasting has been done in [11]. are found to be effective at capturing long-term temporal
dependencies without suffering from the optimization hurdles
ANNs were extensively used for regression, classification, that plague simple recurrent networks (SRNs) [23], and they
and time series modeling until the mid 1990, but then got left have been used to advance the state of the art for many difficult
behind with the advent of other novel regression methods. In problems.
2006, the interest in neural network research was rekindled by
Hinton et. al [15]. They showed that much better performance The key component of a LSTM architecture is a memory
could be achieved using neural networks with multiple hidden cell which retains its state over time, and non-linear gating
layers or deep networks. Numerous efforts have been made to units which regulate the information flow in the cell.
use the power of deep neural networks for time series modeling
and forecasting. LSTMs enable backpropagation of error across the network
and in time. A controlled error propagation allows for networks
In [16], a deep belief network with multiple restricted to learn over large time steps thereby enabling relational
Boltzmann machines is proposed for time series forecasting. learning across vast time differences.
The researchers optimized the model’s performance using par-
ticle swarm optimization (PSO) and present superiority of their The basic unit in the hidden layer of a LSTM network
approach over standard feed forward neural networks and other is the memory block. A memory block contains one or more
statistical models such as ARIMA models. In [17], the authors memory cells and a pair of adaptive, multiplicative gating units
conducted simulations to compare deep learning architectures which gate input and output to all cells in the block. Memory
with standard neural networks for time series forecasting. blocks allow cells to share the same gates, thus reducing the
In [18], several machine learning algorithms are presented to number of adaptive parameters.
address the time series forecasting problem, such as multi-layer
perceptron, Bayesian networks, K-nearest neighbor regression,
support vector regression, and Gaussian processes. Whereas,
in [19] the researchers presented the impact and usefulness of
local learning techniques in dealing with temporal data.
Superficial methods, such as ensemble methods, where the
goal is to improve the performance of the “unstable” predictors
(decision trees and neural networks) have been used in [20].
Research in forecasting the reliability of a mining machine
using ensemble methods is presented in [21]. The authors used
least-square support vector machine with parameter estimation
using genetic algorithm and they used standard benchmark data
sets for reliability forecasting and fault prediction.
Overall, it is found that numerous variants of the above
mentioned machine learning algorithms have been experi-
Fig. 1: LSTM Block with SRN as a hidden layer
2574

Authorized licensed use limited to: The University of British Columbia Library. Downloaded on January 14,2022 at 21:27:43 UTC from IEEE Xplore. Restrictions apply.
 The equations for LSTM under consideration are split into The input vector to the network is given as xt at time t,
two parts: forward pass and the back-propagation through time. σ, g, h are point-wise non-linear functions with σ(x) = 1+e1−x
being the logistic sigmoid, tangent hyperbolic is used for block
Forward Pass input and output activation functions respectively. We denote
the point-wise multiplication of two vectors by . The first
Let N be the number of LSTM blocks and M be the number phase of LSTM training is the forward pass which is presented
of inputs. We have the following weights: The alphabets W , in Algorithm 1 as adapted from [24].
R, p, and b denote the weight associated with input, recurrent,
peephole, or bias connections. The subscripts z, s, f , and o Algorithm 1 Forward Pass
refer to the weights connecting the input gate, blocking gate,
1: procedure FORWARD - PASS
forget gate, and the output gate, respectively.
2: z t = Wz xt + Rz y t−1 + bz
• Input Weights: Wz , Ws , Wf , Wo ∈ RN ×M 3: z t = g(z t ) block input
t
4: i = Wi xt + Ri y t−1 + pi  ct−1 + bi
• Recurrent Weights: Rz , Rs , Rf , Ro ∈ RN ×M 5:
t
it = σ(i ) input gate
t t t−1 t−1
N f = Wf x + Rf y + pf  c + bf
• Peephole Weights: ps , pf , po ∈ R 6:
t
7: f t = σ(f ) forget gate
• Bias Weights: bz , bs , bf , bo ∈ RN
8: ct = z t  it + ct−1  f t cell
All of the arrows entering the memory block are comprised 9: ot = Wo xt + Ro y t−1 + po  ct + bo
of the recurrent connection weights and the input weights. 10: ot = σ(ot ) output gate
11: y t = h(ct )  ot block output
A recurrent neural network with a memory block M, can 12: end procedure
be represented as shown in Figure 2. Here, xt represents the
input, ht represents the output, M denotes the memory block.
The second part of the algorithm is the backpropagation
The curved arrow connecting the memory block to itself is the
through time, we consider here ∆t is a vector of deltas
recurrent connection.
(or gradients) passed down from the previous layer. Let us
consider E to be the loss function, and is generally referred
∂E
to as ∂y t , which does not include the recurrent dependencies.

Therefore, there is a need for a special procedure to evaluate

the deltas inside the LSTM block presented in Algorithm 2 as
adapted from [24].

Algorithm 2 Backpropagation Through Time

1: procedure BACKPROP
2: δy t = ∆t +RzT δz t+1 +RiT δit+1 +RfT δf t+1 +RoT δot+1
3: δot = δy t  h(ct )  σ 0 (ot )
Fig. 2: RNN with a memory block 4: δct = δy t  ot  h0 (ct ) + po  δot + pi  δit+1 + pf 
δf t+1 + δct+1 + δct+1  f t+1
To visualize the RNN for sequence learning at time in- t
5: δf t = δct  ct−1  σ 0 (f )
stance t, it can be unfolded into a network with horizontal t
connections as shown in Figure 3. The subscripts 0, 1, ..., 6: δit = δct  z t  σ 0 (i )
t − 1, t represent the time lag and are a result of unfolding the 7: δz t = δct  it  g 0 (z t )
recurrent connection to the memory cell. 8: end procedure

The output at any time instant from the LSTM-RNN is Finally, the deltas for the inputs are only required if there
dependent on previous inputs as the horizontal connection or is a training layer below them and is given by the Equation 1.
recurrent connection weights govern the strength of depen-
dence. The colored blocks in Figure 3 show one possible case δxt = WzT δz t + WiT δit + WfT δf t + WoT δot (1)
where the output ht is dependent on xt−1 , ..., x1 , and x0 .
The gradients of the weights are calculated based on the
Equation 2 - 7.

T
X
t
δWk∈z,i,f,o = hδk∈z,i,f,o , xt i (2)
t=0
T
X −1
t+1
δRk∈z,i,f,o = hδk∈z,i,f,o , yt i (3)
t=0
XT
t
δbk∈z,i,f,o = δk∈z,i,f,o (4)
Fig. 3: Unfolded RNN with LSTM memory block where the
t=0
output at time t (red) is dependent on previous inputs (blue)
2575

Authorized licensed use limited to: The University of British Columbia Library. Downloaded on January 14,2022 at 21:27:43 UTC from IEEE Xplore. Restrictions apply.
 T
X −1
δpi = ct  δit+1 (5)
t=0

T
X −1
δpf = ct  δf t+1 (6)
t=0

T
X
δpo = ct  δot (7)
t=0

There are numerous variants of LSTM which have been Fig. 5: Hourly power demand for 1-day in Ontario, Canada
studied in the literature such as: no input gate, no forget gate,
no output gate, no input activation function, no output activa- The data can be analyzed on an hourly time scale but large
tion function, coupled input and forget gate, no peepholes, and scale (annual) visualization of historical data reveals other
full gate recurrence [24]. It is observed through experiments hidden characteristics which remain unknown at small scale
that for our purpose of electrical load forecasting, LSTM in (daily). As in Figure 6 we see a pattern in the energy demand
its vanilla form performs well. during varying seasons on visualizing annual hourly data.

In the next section, we present the application of LSTM

networks to the task of sequential data modeling and time
series forecasting in our case study for short-term electrical
load for the Province of Ontario, Canada.
IV. E XPERIMENTS

LSTM-RNN for short term electric load forecasting is used

as an autoregressive model where it can only access input from
the current time step. Various competitors to LSTM such as
multi-layer perceptron see several consecutive inputs in a given
time window when trained by back-propagation.

Fig. 6: Hourly power demand for 1-year in Ontario, Canada

Fig. 4: LSTM as time series predictor It is important to capture both the seasonal, daily, and
hourly variations in the power demand. Most of these features
need to be modeled separately when using statistical time
In Figure 4, X(t) denotes the input, S is a scaling factor, series modeling approaches or even standard feed forward
and X(t + p) is the prediction p time steps ahead. neural networks. In LSTM, these features are learned by the
network automatically as the data is presented to the network.
Network Topology
Data
The input units are fully connected to a hidden layer con-
The data for electrical load are usually maintained on sisting of memory block with 1 cell each. The cell outputs are
an hourly basis. Electrical load has complex and non-linear fully connected to the cell inputs, to all gates, and to the output
relationships. The electrical load data for Ontario is obtained units. All the gates, the cell, and the output units are biased.
from Independent Electricity System Operator (IESO) [25]. It Bias weights are initialized in steps of -0.5 starting from -0.5
is observed in Figure 5 that maximum demand occurs between for each block with forget gates having symmetric biases on
4:00 pm and 8:00pm whereas the minimum demand happens the positive side. All other weights are initialized in the range
between 3:00 am and 5:00 am. We used 10 years of (from [-0.1,+0.1]. The cell input function g is sigmoid in the range [-
2006 to 2016) historical data for training and validation of our 1,+1] and the output function h is tangent hyperbolic. We used
LSTM network in the ratio of 70% for training and 30% for a constant learning rate of α = 10−3 . We used mean squared
testing and validation. error as our loss function for optimizing the parameters of our

2576

Authorized licensed use limited to: The University of British Columbia Library. Downloaded on January 14,2022 at 21:27:43 UTC from IEEE Xplore. Restrictions apply.
 Month ANN ARIMA LSTM-RNN
January 4.6 5.7 4.4
May 6.3 8.2 5.9
September 3.8 3.9 3.8
TABLE I: NRMSE in % for ANN, ARIMA and LSTM-RNN
for different months in a year
LSTM-RNN. We used Adam as the optimization algorithm for
our loss function. It is an algorithm for first-order gradient-
based optimization of stochastic objective functions, based on
adaptive estimates of lower-order moments [26].
Results
The overall goal of the network N is to predict the future Fig. 8: Hourly power demand and forecast using LSTM for a
load at time instance t + T , where t is the current time. The day in the Fall of 2015
target for the network N is the difference between the values
x(t + p) of the electrical load time series p time steps ahead
and the current value, with a scaling factor S.
Therefore N (t) = S.(x(t + p) − x(t)) = S.∆x(t). The
scaling factor is used to bring all the values in the range
[-1,+1]. THe same scaling factors are used for training and
testing purposes. We do the reverse when evaluating the
predicted values.
The square root of the mean/average of the square of all of
the error is a good measure of performance of the algorithm.
The use of root mean square error (RMSE) is very common
and it makes an excellent general purpose error metric for
numerical predictions.
Compared to the similar Mean Absolute Error, RMSE
amplifies and severely punishes large errors. If the RMSE Fig. 9: Hourly power demand and forecast using LSTM for
is normalized then it is referred to as Normalized RMSE or day in the Winter of 2015
NMRSE and is given by Equation 8.
v
u n  2 !
u1 X yi − yˆi
N M RSE = t (8)
n i=1
ymax − ymin
In Figure 7 we show the original values with the forecast values
obtained using the training data for the past 10 years.
We present our results for three different months in differ-
ent seasons (i.e. Fall, Winter and Spring) and for a year.
In Figure 7, the annual hourly power demands and forecasts
are presented.

Fig. 10: Hourly power demand and forecast using LSTM for
day in the Spring of 2015

In Figures 8, 9, and 10 the actual and predicted power

demand for a day in three different seasons is presented. It
is clear from the predictions and NRMSE values as presented
in Table I, that LSTM-RNN peforms better or equivalent to
existing techniques. It is also important to note that we do not
require to segregate the data based on seasonality, time of day,
and other factors for training the network. The network trains
Fig. 7: Hourly power demand and forecast using LSTM for and learns for itself the underlying features from the data.
the year 2015
2577

Authorized licensed use limited to: The University of British Columbia Library. Downloaded on January 14,2022 at 21:27:43 UTC from IEEE Xplore. Restrictions apply.
 [5] S. Li, P. Wang, and L. Goel, “Short-term load forecasting by wavelet
transform and evolutionary extreme learning machine,” Electric Power
Systems Research, vol. 122, pp. 96–103, 2015.
[6] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.
[7] U. Mukherjee, A. Maroufmashat, A. Narayan, A. Elkamel, and
M. Fowler, “A stochastic programming approach for the planning and
operation of a power to gas energy hub with multiple energy recovery
pathways,” Energies, vol. 10, no. 7, p. 868, 2017.
[8] A. Narayan and K. Ponnambalam, “Risk-averse stochastic programming
approach for microgrid planning under uncertainty,” Renewable Energy,
vol. 101, pp. 399–408, 2017.
[9] G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time series
analysis: forecasting and control. John Wiley & Sons, 2015.
[10] S.-J. Huang and K.-R. Shih, “Short-term load forecasting via arma
model identification including non-gaussian process considerations,”
Fig. 11: Decay of loss function value with epochs IEEE Transactions on Power Systems, vol. 18, no. 2, pp. 673–679,
2003.
In Figure 11, the decay of the loss function with epochs is [11] H. Hahn, S. Meyer-Nieberg, and S. Pickl, “Electric load forecasting
shown. The loss function used is the mean squared error which methods: Tools for decision making,” European Journal of Operational
decays quite rapidly to a low value given the large training set. Research, vol. 199, no. 3, pp. 902–907, 2009.
[12] Z. Aung, M. Toukhy, J. Williams, A. Sanchez, and S. Herrero, “Towards
V. C ONCLUSIONS accurate electricity load forecasting in smart grids,” in The Fourth
International Conference on Advances in Databases, Knowledge, and
In our work we implemented LSTM deep recurrent neural Data Applications, DBKDA, 2012.
network for modeling short-term electrical load time series for [13] E. Ceperic, V. Ceperic, and A. Baric, “A strategy for short-term load
the Province of Ontario, Canada. Short-term electrical load forecasting by support vector regression machines,” IEEE Transactions
on Power Systems, vol. 28, no. 4, pp. 4356–4364, 2013.
modeling has been a challenging task and is important for
reliable and profitable operation of electricity markets and [14] C. Cecati, J. Kolbusz, P. Różycki, P. Siano, and B. M. Wilamowski,
“A novel rbf training algorithm for short-term electric load forecasting
utilities . It is well established that electrical load is dependent and comparative studies,” IEEE Transactions on Industrial Electronics,
on a large number of underlying factors which are hard to vol. 62, no. 10, pp. 6519–6529, 2015.
identify and model using currently established techniques. [15] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
Deep learning has become a very promising approach for deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554,
pattern recognition for complex and high dimensional data sets. 2006.
We used the deep learning approach to model sequential short- [16] T. Kuremoto, S. Kimura, K. Kobayashi, and M. Obayashi, “Time
term electrical demand time series. The results from LSTM- series forecasting using restricted boltzmann machine,” in International
Conference on Intelligent Computing. Springer, 2012, pp. 17–22.
RNN are reliable, robust, and comparable with other state of
[17] E. Busseti, I. Osband, and S. Wong, “Deep learning for time series
the art techniques. modeling,” Technical report, Stanford University, Tech. Rep., 2012.
The power of LSTM can be utilized for even more complex [18] N. K. Ahmed, A. F. Atiya, N. E. Gayar, and H. El-Shishiny, “An
empirical comparison of machine learning models for time series
tasks given the ability of these networks to mine complex forecasting,” Econometric Reviews, vol. 29, no. 5-6, pp. 594–621, 2010.
hidden patterns in unlabeled data which is present in large
[19] G. Bontempi, S. B. Taieb, and Y.-A. Le Borgne, “Machine learn-
quantities. We wish to extend the LSTM model for forecasting ing strategies for time series forecasting,” in Business Intelligence.
not only short term loads to even finer levels of granularity Springer, 2013, pp. 62–77.
such as 5 min and 10 min. This would enable ease in [20] L. Beriman, “Bias, variance, and arching classifiers,” Technical Report,
integration of the renewable source of energy with the existing Tech. Rep., 1996.
power systems. [21] S. Chatterjee, A. Dash, and S. Bandopadhyay, “Ensemble support vector
machine algorithm for reliability estimation of a mining machine,”
R EFERENCES Quality and Reliability Engineering International, vol. 31, no. 8, pp.
1503–1516, 2015.
[1] P. Geurts, “Pattern extraction for time series classification,” in European
[22] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
Conference on Principles of Data Mining and Knowledge Discovery.
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
Springer, 2001, pp. 115–127.
[23] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependen-
[2] A. Mardani, A. Jusoh, and E. K. Zavadskas, “Fuzzy multiple criteria
cies with gradient descent is difficult,” IEEE transactions on neural
decision-making techniques and applications–two decades review from
networks, vol. 5, no. 2, pp. 157–166, 1994.
1994 to 2014,” Expert Systems with Applications, vol. 42, no. 8, pp.
4126–4148, 2015. [24] K. Greff, R. K. Srivastava, J. Koutnı́k, B. R. Steunebrink, and
J. Schmidhuber, “Lstm: A search space odyssey,” arXiv preprint
[3] S. I. Vagropoulos, E. G. Kardakos, C. K. Simoglou, A. G. Bakirtzis, and
arXiv:1503.04069, 2015.
J. P. Catalão, “Artificial neural network-based methodology for short-
term electric load scenario generation,” in Intelligent System Application [25] IESO, “Independent electricity system operator,” http://www.ieso.ca/
to Power Systems (ISAP), 2015 18th International Conference on. Pages/Power-Data/Demand.aspx.
IEEE, 2015, pp. 1–6. [26] D. P. Kingma and J. Ba, “Adam: A method for stochastic
[4] A. Narayan, K. W. Hipel, K. Ponnambalam, and S. Paul, “Neuro- optimization,” CoRR, vol. abs/1412.6980, 2014. [Online]. Available:
fuzzy inference system (asupfunis) model for intervention time series http://arxiv.org/abs/1412.6980
prediction of electricity prices,” in Systems, Man, and Cybernetics
(SMC), 2011 IEEE International Conference on. IEEE, 2011, pp.
2121–2126.

2578

Authorized licensed use limited to: The University of British Columbia Library. Downloaded on January 14,2022 at 21:27:43 UTC from IEEE Xplore. Restrictions apply.