Professional Documents
Culture Documents
1 s2.0 S0306261918315009 Main
1 s2.0 S0306261918315009 Main
1 s2.0 S0306261918315009 Main
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
Center for Computational Engineering & Networking (CEN), Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, India
HIGHLIGHTS
Keywords: The electric load forecasting is extremely important for energy demand management, stability and security of
Short-term electric load forecasting power systems. A sufficiently accurate, robust and fast short-term load forecasting (STLF) model is necessary for
Time-series analysis the day-to-day reliable operation of the grid. The characteristics of load series such as non-stationarity, non-
Dynamic mode decomposition linearity, and multiple-seasonality make such prediction a troublesome task. This difficulty is conventionally
Prediction
tackled with model-driven methodologies that demand domain-specific knowledge. However, the ideal choice is
Smart grid
a data-driven methodology that extracts relevant and meaningful information from available data even when the
physical model of the system is unknown. The present work is focused on developing a data-driven strategy for
short-term load forecasting (STLF) that employs dynamic mode decomposition (DMD). The dynamic mode de-
composition is a matrix decomposition methodology that captures the spatio-temporal dynamics of the under-
lying system. The proposed data-driven model efficiently identifies the characteristics of load data that are
affected by multiple exogenous factors including time, day, weather, seasons, social activities, and economic
aspects. The effectiveness of the proposed DMD based strategy is confirmed by conducting experiments on
energy market data from different smart grid regions. The performance advantage is verified using output
quality measures such as RMSE, MAPE, MAE, and running time. The forecasting results are observed to be
competing with the benchmark methods. The satisfactory performance suggests that the proposed data-driven
model can be used as an effective tool for the real-time STLF task.
⁎
Corresponding author.
E-mail address: neethumohan.ndkm@gmail.com (N. Mohan).
https://doi.org/10.1016/j.apenergy.2018.09.190
Received 11 June 2018; Received in revised form 19 September 2018; Accepted 23 September 2018
Available online 04 October 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
N. Mohan et al. Applied Energy 232 (2018) 229–244
changes, economic aspects, and social activities make the load fore- more accurately. In past years, support vector machines (SVM) [16] and
casting a difficult task [3]. its variants were widely being used for time-series forecast because of
Electricity demand forecasting simply means obtaining an expected its non-linear mapping capability. Mohandes [17] applied SVM with
value for demand in the future and it is mainly categorized as (i) long- sigmoid kernel function for STLF and proved the model to be competing
term forecasting - makes the prediction for 5–20 years ahead. The key with the AR model. The drawback of the SVM model is that the fore-
intention is resource management, long-term planning and investment casting accuracy is proportional to the amount of load data available for
of power systems. (ii) mid-term forecasting - does the prediction training. Yuancheng et al. [18] developed a least-squares SVM (LS-
spanning a few months to certain years ahead (month-5 years). It is SVM) approach by integrating load and temperature values of YanTai
mainly used for power production, financial and operational planning. electric network. Ceperic et al. [19] proposed a strategic-seasonality
(iii) short-term forecasting - makes the prediction spanning from a few adjusted support vector regression (SSA-SVR) technique for STLF. The
hours to weeks ahead and mainly focused on scheduling and analysis of superior performance of SSA-SVR method hinged on the specific se-
the distribution network. lection of input variables and features for SVM. In general, the selection
Short-term load forecasting (STLF) helps to improve the manage- of optimum hyper-parameters and suitable kernel functions are the two
ment efficiency and reduce the grid operating cost. It is highly desirable important liabilities of SVM models. Pai and Hong [20] implemented a
for the secure and reliable operation of electric grid. An accurate STLF genetic algorithm (GA) to select the optimum parameters of SVM to
model should consider the non-linear, time-evolving characteristics of predict the regional electric demand. Hong [21] utilized a chaotic
load series to efficiently forecast the future demand. Additionally, the particle swarm optimization (CPSO) algorithm to tune the hyper-
smart grid infrastructure increases the complexities in modern grid parameters of the SVR model and the results are found to be higher than
management and protection [4]. The hourly load fluctuations generate GA-SVM model. Che and Wang [22] developed a kernel-based SVR
ambiguities in operation and planning phases of smart grid [5]. Dealing model by combining different kernel functions to forecast the electric
with these uncertainties to improve the forecasting performance is load data. The main shortcoming associated with the model is the need
troublesome and cause the STLF task challenging. for the selection of suitable kernel combinations.
Neural network (NN) [23] based models can intrinsically learn the
1.1. Review of related works and motivation non-linear behavior of load series data. Kouhi et al. [24] implemented a
cascaded NN structure by utilizing wavelet transform (WT) for feature
Literature studies on STLF can be broadly divided into four groups selection to avoid irrelevant features. However, the two-stage feature
depending on the techniques used for forecasting or prediction: (i) selection procedure and NN parameter tuning increase the complexity
linear methods, (ii) non-linear methods, (iii) ensemble techniques, (iv) of the method. Amjady et al. [25] integrated a modified harmony
deep learning based models. search technique with NN to efficiently search the solution space for
Linear forecasting methods predict the future values of the electric STLF. This model tries to overcome the over-fitting issue to a certain
load series by using a linear function. The popular examples of linear extent using the search algorithm. Ranaweera et al. [26] developed a
forecasting techniques include autoregressive (AR), moving average radial basis function neural network (RBFN) and proved it to be more
(MA), autoregressive moving average (ARMA) and autoregressive in- efficient than back-propagation (BP) models in terms of accuracy and
tegrated moving average (ARIMA) models [6]. Generally, these are training time. Chang [27] introduced a radial basis function NN for
simple statistical models and easy to implement. The application of forecasting the electric load in Tai power system. Ding et al. [28]
ARIMA model for STLF is initially proposed in [7], in which a third- proposed a predictive NN based STLF model using variable and model
order polynomial function is utilized to efficiently forecast the electric selection to choose the suitable parameters. The proposed model is
loads. Taylor et al. [8] described an STLF model using ARIMA by ex- verified using the distribution network data from France. Khwaja et al.
ploiting time-load relationship. In this work, different ARIMA im- developed two NN models namely bagged NN (BNN) [29] and boosted
plementations are compared using load data from ten European coun- NN (BooNN) [30] to improve the forecasting performance of single
tries. It is found that the Holt-Winters exponential smoothing-AR model artificial neural network (ANN) models. Telbany et al. [31] developed
gives the best performance. Amjady [9] developed a modified ARIMA an NN model trained by particle swarm optimization (PSO) algorithm
by combining human expert knowledge to forecast the daily and sea- to predict the daily peak electric loads of Jordanian electricity market.
sonal load patterns and the model achieved better results than con- The results are found to be better than BP and ARMA models. However,
ventional Box-Jenkins ARIMA. However, the dependency on human the optimal parameter selection for PSO algorithm is a challenging
experts limits its suitability for real-time application. Huang and Shih aspect of this model. Cecati et al. [32] introduced different machine
[10] proposed an ARMA approach including non-Gaussian process for learning techniques namely SVR, extreme learning machine (ELM),
STLF. The average forecast error of the ARMA model for one-week decay RBF, improved second order and error correction to train the RBF
ahead period is 1.57% and one-day ahead period is 1.62%. Taylor [11] network for one-day ahead forecasting. Among these techniques, the
defined a double seasonal load model using multiplicative ARIMA to error correction method received the highest efficiency for STLF. Che
forecast the one-day ahead load. In this model, to accommodate the et al. [33] proposed an adaptive fuzzy combination model for STLF
multiple seasonalities, Holt-Winters exponential smoothing is em- using self-organizing maps (SOM) and SVR. Even though the model
ployed. Mohamed et al. [12] developed a double seasonal ARIMA attained good accuracy, the complexity of the model is high. Cevik et al.
(SARIMA) with kth order polynomial to forecast the daily and weekly [34] developed an adaptive neuro-fuzzy interface system (ANFIS) by
patterns of Malaysian electricity grid. Even though the MAPE obtained incorporating temperature difference and season as additional para-
is less than 1%, its performance is found to be decaying for larger time- meters. Although the average MAPE for ANFIS model is 1.85%, the
step. Tarsitano et al. [13] proposed a two-stage seasonal ARIMA disadvantage is that the real-time implementation is harder due to its
(SARIMAX) model to predict the load demand from one-day to nine-day dependency on additional parameters. Gao et al. [35] proposed dif-
ahead period. Dudek [14] described a linear regression technique for ferent types of multi-block Elman NN (ENN) for STLF. In this complex
pattern-based forecasting. However, the linear models are inadequate model, an initial feature selection stage based on mutual information
to represent the non-linear behavior of load series and hence fails to (MI) is adopted using the relevancy-redundancy criteria. The optimal
predict the accurate future demand values. Thus, the forecasting ac- parameters of ENN are tuned using the enhanced shark smell optimi-
curacy is poor for linear models. zation (ESSO) algorithm. The different types of ENN model considered
The non-linear forecasting models predict the future demands by are (1) series, (2) parallel, (3) parallel-cascade-in-cascade (PCC), (4)
means of a non-linear function with respect to the input values [15]. cascade-parallel-in-cascade (CPC), (5) cascade-parallel-in-parallel
The non-linear functions can learn the complex nature of load series (CPP), (6) parallel-cascade-in-parallel (PCP), (7) hybrid cascade, and
230
N. Mohan et al. Applied Energy 232 (2018) 229–244
(8) hybrid parallel. Bhattacharya et al. [36] forecasted the Australian window EMD method combining with a feature selection procedure
grid data using genetic programming (GP) and found to be competing for small-scale load forecasting in buildings. In this model, an op-
with ANN and fuzzy techniques. The optimization involved in many of timized Elman neural network engine is used to forecast the load
the non-linear models suffer from the disadvantage of converging to the values. Lin et al. [47] developed a hybrid model combining varia-
local minimum. The hyper-parameter tuning and appropriate kernel tional mode decomposition (VMD) with ELM for STLF. In this
selection are additional constraints of non-linear forecasting models. model, the data series is decomposed using VMD into distinct modes
The ensemble or hybrid forecasting methods combine the advantage to remove uncertainties. Further, each mode is forecasted using ELM
of one or more techniques for efficient prediction of future values. optimized with differential evolution (DE) algorithm.
These models overcome the drawbacks of corresponding non-ensemble
structures and achieve better forecasting accuracy. The additional ad- Even though the ensemble models possess novel methodologies and
vantages of ensemble methods are robustness and efficiency. Most of higher accuracy, it suffers from several limitations including the re-
the ensemble models combine linear and non-linear models for ob- quirement of appropriate network structures, convergence to local op-
taining effective prediction. In recent years, various ensemble models timal locations and high complexity.
are developed mainly by combining (1) traditional statistical methods Recently, deep learning based approaches are becoming popular for
[37–39] (2) machine-learning algorithms [40–44], and (3) decom- STLF due to its dominant ability to capture non-linear time series
position techniques [1–3,45–47]. characteristics [48–51]. Deep learning models have the inherent ability
to learn the hidden patterns in the load series data. Hossen et al. [48]
1. Amjady et al. [37] proposed NN/SVR model for price and demand investigated the application of a multi-layer deep neural network
forecasting by integrating an improved mutual information (MI) (DNN) to forecast the Iberian grid data using 90 days of load history.
analysis to remove the irrelevant features of the input data. In this The temperature value, wind speed, and solar irradiance are used as
model, an iterative technique is exploited to predict the inter- additional inputs to train the DNN architecture. The different combi-
dependent variables of the electric grid. Niu et al. [38] developed an nations of activation functions are also studied for accurate load fore-
ANN by integrating Bayesian rules and Monte Carlo algorithm to casting. Zheng et al. [49] explored the possibilities of sequential deep
forecast the Chinese grid data. This model attained good prediction learning models namely long-short term memory (LSTM) and recurrent
accuracy (average MAPE of 0.7831%) by learning the weights of neural network (RNN) and the results are found to be better than
ANN using Monte Carlo algorithm. Fan et al. [39] combined bi- conventional AR models. Narayan et al. [50] developed an LSTM ar-
square kernel (BSK) based on the position similarity of the re- chitecture with an average NRMSE value of 5.9% for STLF. Kong et al.
constructed data for STLF. This model utilized the classical phase [51] proposed an LSTM-RNN based framework to forecast the short-
space reconstruction technique called C-C method for reconstructing term residential loads. The ability of LSTM to learn the long-term
the data. The forecasting accuracy of the method is less than 2.20% temporal characteristics is exploited for the accurate prediction of in-
in terms of MAPE. dividual household load.
2. Singh et al. [40] introduced an ensemble ANN model with the Despite the aforementioned models, several other techniques are
evolutionary algorithm for improved electric demand prediction. also being proposed for STLF. Alvarez et al. [52] developed a pattern
The superior performance of the model is compared to other alter- sequence similarity (PSF) based algorithm that extracts the pattern of
native hybrid approaches. Hu et al. [41] utilized an SVR with filter- the prior day to forecast the next day pattern. In this model, the pat-
wrapper combined feature selection process to eliminate the irre- terns similar to the extracted previous day pattern is identified from the
levant features of load data. However, the improved performance is historical data to predict the future demand. Koprinska et al. [53] in-
reliable with the selection of proper input features. Li et al. [42] tegrated PSF with NN (PSF-NN) to efficiently forecast the hourly load
forecasted the decomposed components of wavelets using ELM series of Australian grid. Jin et al. [54] proposed a hybrid optimization
tuned with modified artificial bee colony (MABC) algorithm. The grey model (HOGM) for STLF and the results are found to be better than
less convergence time is the most significant advantage of the ELM- the conventional grey model. Asrari et al. [55] developed a gray-Fuzzy-
MABC method. Sudheer et al. [43] developed a load model com- Markov model for day-ahead forecasting and achieved superior per-
bining the weighted nearest neighbor (WNN) and triple exponential formance than gray, ARIMA, and wavelet-ARIMA models. Table 1
smoothing (TES) to forecast the wavelet components for day-ahead shows the summery of few state-of-the-art approaches for STLF. The
prediction. Ghadimi et al. [44] introduced a two-stage hybrid NN electricity markets, the horizon of forecasting and the evaluation me-
architecture consisting of ridgelet and Elman neural networks for trics used in each model are described in Table 1. A detailed review of
STLF. The NN parameters are optimized using the chaotic binary several STLF models are available in [56–59].
shark smell optimization (CBSSO) algorithm. Even though many of the existing methods have good prediction
3. Qiu et al. [1] developed a hybrid method by combining restricted accuracy, they fail to provide generality and interpretability for future
Boltzmann machines (RBM) with deep belief networks (DBN) to demand prediction [3]. Single structure models often encounter issues
forecast the individual modes obtained through the empirical mode in dealing with the chaotic behavior of non-linear, timely-emerging
decomposition (EMD). Though the model achieved good forecasting load series. Some of the other methods depend on a feature extraction
performance, the ensemble architecture increases its complexity. and selection stage [4,24,37,41,60]. The hand-crafted feature identifi-
Fan et al. [2] integrated differential-EMD (DEMD) and auto-re- cation requires additional efforts and further increases the complexity
gression (AR) with non-linear SVR to achieve an improved fore- of the method. To achieve acceptable forecasting accuracies, the en-
casting performance. In this model, DEMD is used to decompose the semble models rely on several preprocessing stages and parameter op-
load series into different modes. Further, the high-frequency modes timization based on the individual technique [1,3,39]. The computa-
are forecasted using SVR and the residual is forecasted using AR. tional time required by the ensemble models are higher than its non-
Zhang et al. [3] used non-linear SVM optimized by cuckoo search ensemble counterparts which limits the real-time applications [1,39].
(CS) algorithm and linear ARIMA to forecast the components ob- Although the deep learning models have outstanding capability to learn
tained through singular spectrum analysis (SSA). Finally, the in- the dynamics of the data, its application for STLF is still in a growing
dividual forecasted components are aggregated using the CS algo- stage. Moreover, the deep learning architectures are parameterized
rithm. Fan et al. [45] developed a hybrid model integrating EMD- models, hence their superior predictive performance is proportional to
PSO-GA-SVR for STLF. The high-frequency EMD modes are fore- the suitable hyper-parameter selection including network parameters
casted using SVR tuned by PSO. The low-frequency EMD modes are and network topologies. Thus, there is a requirement for the develop-
forecasted using GA. Mohammadi et al. [46] proposed a sliding ment of a simple and effective model for STLF.
231
N. Mohan et al. Applied Energy 232 (2018) 229–244
Table 1
Summery of various state-of-the-art short-term electric load forecasting models.
Model Technique Electricity market/data Horizon Metrics
Linear models ARIMA [8] Norway, Italy, France, Sweden, Spain, Belgium, Half-an-hr, One-hr MAPE, MAE
Finland, Ireland, Portugal, Britain
Modified ARIMA [9] Iran One day MAPE, PAPE
ARMA [10] Taiwan One-day, One-week Relative error
SARIMA [12] Malaysia One-step, One-week MAPE
Linear regression [14] Poland One-day MAPE
Ensemble models EMD-DBN [1] Australia, China Half-an-hr, One-day, MAPE, RMSE
One-week
DEMD-SVR-SR [2] NSW-Australia, New York One-day, One-week MAPE, RMSE, MAE
SSA-SVM-ARIMA-CS [3] NSW-Australia, Singapore One-day, One-week MAPE, MAE, MSE
PSR-BSK [39] NewYork, NSW-Australia One-day, One-week MAPE, RMSE, MAE
MI feature selection [37] Spain, New York One-month, One-year MAPE, MAE
Bayesian-ANN-Monte Carlo Algo. China One-month MAPE, RMSE
[38]
ANN-Evolutionary Algo. [40] England, Australia, Texas One-day, One-week MAPE, MAE
Filter-Wrapper-SVR [41] Kaggle dataset, North America One-day, One-week MAPE, MASE
WT-ELM-MABC [42] England, North America One-week, One-month MAPE, MAE
WT-TES-WNN [43] California, Spain One-day, One-week MAPE, EV
EMD-PSO-GA-SVR [45] NSW-Australia, New York One-day, One-week MAPE, RMSE, MAE
VMD-ELM [47] NSW & QLD-Australia One-day, One-week MAPE, RMSE, MAE
Other STLF models PSF [52] Spain, Australia, New-York One-week, One-year MAE, MER
PSF-NN [53] NSW-Australia One-day MAPE, MAE
HOGM [54] China One-day Error %
Gray-Fuzzy-Markov [55] Ontario, PJM-US, Iran One-day MAPE, Error %
MAPE: Mean Absolute Percentage Error, MAE: Mean Absolute Error, PAPE: Peak Absolute Percentage Error, CC: Correlation Coefficient, RMSE: Root Mean Square
Error, NMAPE: Normalized MAPE, NRMSE: Normalized RMSE, MER: Mean Error Relative, MSE: Mean Square Error, EV: Error Variance.
1.2. Objective and key contributions initial univariate demand data is Hankelized to obtain a multi-dimen-
sional augmented data matrix with overlapping load series [64]. The
The objective of the present paper is to propose a fast, robust, and Hankel mapping is motivated by the fact that a novel representation is
efficient data-driven strategy for STLF. The data-driven methods take needed to fully learn and identify the spatio-temporal structures from
the advantages of mining the data to identify the latent dynamic the non-stationary, time-evolving electric load data. Further, the ei-
characteristics. It provides several ways to understand the underlying gendecomposition and dynamic mode extraction are done based on the
nature of the system even when the physical model of the system is not Hankel matrix. Finally, a data-rearrangement and averaging operation
known. The complex relational characteristics of the data (temporal are performed on the initial predicted series to get the appropriate fu-
and spatial) are captured by the data-driven methods. It can efficiently ture load series. The key contributions of this paper are,
discover the hidden dynamicities and thus offers an accurate identifi-
cation of the system behavior. • Development of an intelligent methodology for short-term load
Dynamic mode decomposition (DMD) has recently gained huge at- forecasting (STLF) using dynamic mode decomposition (DMD) by
tention due to its dominant ability to extract meaningful information extracting the meaningful, hidden tractable information from load
from available data [61,62]. It has revolutionized the analysis and series data. The proposed method can handle the load series data
modeling of physical systems like turbulent flows, fluid dynamics, that is affected by multiple factors including time, day, seasons,
neuroscience, financial trading markets, etc. [63,61]. This paper ex- climate, and socio-economic activities.
plores the ability of the DMD algorithm to capture the hidden under- • Demonstration of the improvements in forecasting accuracy, low
lying features of load series data. To capture multiple seasonality and complexity and fast running time using the energy market data from
cyclicity of the univariate load data, we have proposed a novel strategy different smart grid regions.
using the DMD algorithm. In the proposed data-driven strategy, the
232
N. Mohan et al. Applied Energy 232 (2018) 229–244
The performance advantage of the proposed data-driven strategy is The major steps involved in DMD algorithm are explained as follows
evaluated through various comparative experiments. The expertise of [61],
the model to forecast demand data for one-day, one-week, and one-
month ahead period is investigated. The rest of this paper is organized 1. Compute singular value decomposition (SVD) of observation matrix
as follows. Section 2 presents the details of the dynamic mode de- X1 as,
composition algorithm followed by the proposed forecasting model in
X1 U VH (7)
Section 3. Section 4 discusses the effectiveness of the proposed model.
Section 5 explains about the datasets, experiments and evaluations where U CN × K ,
V K represents the rank of the
CK × K , CM × K ,
performed. Finally, conclusions are drawn in Section 6. reduced SVD approximation to X1.
2. Compute companion-type matrix Z from,
2. Mathematical background of dynamic mode decomposition X2 X1 Z U V HZ (8)
(DMD)
The dimension of each measurement vector is N and the two observa- W = [ w1 w2 w3 … wM 1] (14)
tion matrices are overlapping over time. By making the assumption that The diagonal matrix of eigenvalues, is visualized as,
the system is evolving slowly, the Mth observation (or snapshot) is
possible to express as a linear combination of previous M 1 observa- 1 0 0 0
tions with residual error (r). 0 2 0 0
=
xM = a1 x1 + a2 x2 + + aM 1 xM 1 +r (3) 0 0 0
M 2
233
N. Mohan et al. Applied Energy 232 (2018) 229–244
Fig. 1. Illustrates the selection of days from load history to create the input load series data.
X1 = [x̄1, x̄2, , x̄L 1] RS × (L 1) (19) Based on the computed dynamic mode matrix , an initial
234
N. Mohan et al. Applied Energy 232 (2018) 229–244
Table 2
Procedure of the proposed data-driven DMD based STLF model
Input vector:
x :- Input load series
Output vector:
x :- Forecasted load series
Steps
1. Normalize x using Eq. (17)
2. Create observation matrices using Eqs. (18)–(20)
3. Compute eigendecomposition using Eqs. (7)–(13)
4. Compute dynamic mode matrix ( ) using Eq. (16)
5. Compute initial prediction (xP ) using Eq. (21)
6. De-normalize xP using Eq. (22)
Fig. 3. Visualization of over the unit circle in complex plane. 7. Perform data-rearrangement and compute XP using Eq. (23)
8. Perform data averaging using Eq. (24)
90 9. Compute final forecasted load series ( x ) using Eq. (25)
120 60
0.8
150 0.6 30
x = [x1, x2, , xF , , xS ] RS . Since the window length for prediction is
0.4
fixed as F, the final forecasted load series from 00 h to next F time-steps
0.2
is given by
180 0
This step achieves the final forecast for 00 h to next F future time-
steps. Here, the previously de-normalized load data vector, xP is rear-
ranged into a matrix XP as shown in Eq. (23) to attain the most suitable
forecast from 00 h to F future values. The XP RG × F with elements
XP (ij) , 1 i G, 1 j F is defined as,
xP (1) xP (2) xP (F )
xP (F + 1) xP (F + 2) xP (2F )
XP =
xP (G) xP (G + 1) xP (G + F 1) (23)
where F denotes the length of the window to be predicted,
G = S F + 1. In order to obtain the final forecasted load series, an
averaging operation is performed on the XP matrix using Eq. (24).
v+1
1
v+1
xP (u, v u + 2) 1 v<G 1
u=1
G
1
xv = G
xP (u, v u + 2) G 1 v<F
u=1
P (S F + 1)
1
S v
x u, v u+2 F v<S
u=v F+2 (24)
The averaging is performed over the diagonals, i + j = v + 2 , and u , v Fig. 5. Schematic illustration of the proposed data-driven strategy for STLF
denotes the locations of elements in XP matrix. The resulting series is using DMD.
235
N. Mohan et al. Applied Energy 232 (2018) 229–244
(a) (b)
6500
8500 Original
Proposed DMD
AR model
8000 6000 ARX model
Demand (MW)
Demand (MW)
ARMAX model
7500
5500
7000 Original
Proposed DMD
AR model
6500 ARX model 5000
ARMAX model
6000
5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 40 45
Time(half-hour) Time(half-hour)
(c) (d)
Demand (MW)
6500 6000 ARX model
ARMAX model
6000
Original
5500
Proposed DMD
5500 AR model
ARX model 5000
ARMAX model
5000
5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 40 45
Time(half-hour) Time(half-hour)
Fig. 6. Illustrates one-day ahead forecasting results using the proposed DMD approach on QLD data. Forecasting results for (a) 10-Feb-2017, (b) 15-Apr-2017, (c) 27-
Jul-2017, (d) 15-Oct-2017.
4. Effectiveness of proposed DMD based forecasting model the actual load demand and forecasted load demand. In each box, the
horizontal central line represents the median, the edges represent 25th
To demonstrate the competence of the proposed DMD based STLF and 75th percentiles and the + sign indicate the outliers. The key
model, data from Queensland (QLD) electric grid is selected [67]. It is advantages of the proposed data-driven DMD based model are (i) the
half-an-hour interval data having 48 observations in a day. The initial model works irrespective of the seasonal variations in the data and thus
load data for generating the observation matrices from Eqs. (18)–(20) is have improved generalization ability for different seasonal patterns, (ii)
created using the four previous days data (ie, two immediate previous computational complexity is less as it does not require any additional
days, same day in the previous week and previous day in the previous feature extraction stage, (iii) the lower MAPE suggests that the model is
week). Further, the eigendecomposition is performed and the dynamic efficient, accurate and robust for STLF, (iv) independent of parameter
modes are estimated using Eq. (16). Finally, the one-day ahead fore- selection, (v) implementation is easy as no additional combination of
casting (00 h to next 48 values) is generated by the DMD based pre- approaches (ensemble) is required.
diction followed by data-rearrangement and averaging operations using
Eqs. (21)–(25). Since the seasonal climate variations have an influence 5. Experiments and discussion
on electric demand data, the days to forecast are chosen from four
different seasons. This will highlight the potentiality of the proposed This section describes the experiments conducted to examine the
model to accommodate the climatic and seasonal characteristics of the performance of the proposed DMD based approach for STLF. To illus-
load data. The results obtained for forecasting are shown in Fig. 6. For trate the advantage of performance, the electric load data from different
the proposed DMD based data-driven methodology, the results illu- energy market operators are considered. This highlights the potentiality
strated in each panel of Fig. 6 is generated by considering the four days and generality of the proposed method for accurate STLF. The fore-
of data from history (as explained in Section 3). The results are com- casting horizon considered in this study are one-day ahead, one-week
pared with autoregressive models such as AR [6], ARX [6], ARMAX [6]. ahead and one-month ahead. Based on the day to be forecasted, the
It can be observed from Fig. 6 that the proposed approach effectively input load series for the experiment is constructed using (i) data from
predict the future load values for a short-term period. Fig. 7 shows an two previous days, (ii) data from the previous day in the previous week,
enlarged view of the marked regions in Fig. 6, where the high peaks are (iii) data from the same day in the previous week. The pictorial re-
generated due to increased load demand. The proposed model exhibit presentation of this selection procedure is shown in Fig. 1 and is dis-
higher flexibility and accuracy at the extreme peak locations as the cussed in Section 3. The similarity in load pattern reinforces the se-
latent dynamics are efficiently captured by the DMD. The average lection of these four distinct days for effective prediction of next day
running time required for the one-day ahead of prediction using the demand. Fig. 9 represents the hourly and half-hourly load data series
proposed DMD model is 0.125 s. The corresponding mean absolute with different cyclic and seasonal patterns. It is clear from Fig. 9 that
percentage error (MAPE) for each model is given in Table 3. The lowest the load series have distinct nature/pattern depending on time, day,
MAPE of the proposed DMD based model highlight its superiority to weather and seasons.
predict the future demand values accurately.
The forecasting performance of the proposed model based on fore- 5.1. Datasets and its characteristics
casting error is visualized through box graph and is shown in Fig. 8. The
forecasting error (FE) is measured as the absolute difference between The electric load data series from two energy market operators
236
N. Mohan et al. Applied Energy 232 (2018) 229–244
(a) (b)
8800
8600 5400
8400
Demand (MW)
Demand (MW)
8200 5200
8000 Original
Original
Proposed DMD
Proposed DMD 5000
7800 AR model
AR model
ARX model
ARX model
7600 ARMAX model
ARMAX model
4800
30 35 40 45 10 12 14 16 18 20 22
Time(half-hour) Time(half-hour)
(c) (d)
6800
Original
6600 Proposed DMD
7000 AR model
ARX model
Demand (MW)
Demand (MW)
6400 ARMAX model
6500
6200
Original
Proposed DMD
6000 AR model
6000
ARX model
ARMAX model 5800
30 35 40 45 35 36 37 38 39 40 41 42 43 44 45 46
Time(half-hour) Time(half-hour)
Fig. 7. Enlarged view of the marked high peak locations in Fig. 6. (a) 10-Feb-2017, (b) 15-Apr-2017, (c) 27-Jul-2017, (d) 15-Oct-2017.
MAPE obtained for four different days for one-day ahead forecasting (The best Daily cycle Weekly cycle
8000
350
6000
300 200 400 600 800 1000 1200 1400
Time
Forecasting error
250
Fig. 9. Example of load data series with cyclic and seasonal patterns. (a) Hourly
200 data, (b) half-hourly data.
150
1
Normalized load-series data
100
0.8
50
0 0.6
Fig. 8. Illustrates forecasting performance of the proposed model for four dif- 0.2
SW
IC
a
ic
A
-S
-V
-Q
-T
er
-N
m
O
EM
O
EM
EM
A
EM
th
A
A
A
or
A
Energy Markets
giving 48 observations per day for five different states including New
Fig. 10. Illustrates the distribution of normalized load series data of a month
South Wales (NSW), Queensland (QLD), Tasmania (TAS), Victoria (VIC) from different energy markets.
and South Australia (SA). Whereas, the North American electricity data
are sampled at every hour giving 24 observations a day. The
237
N. Mohan et al. Applied Energy 232 (2018) 229–244
Table 4
One-day ahead prediction results on Australian grid (AEMO) data.
Data Model RMSE MAPE
NSW Persistence 978.24 729.50 609.82 587.14 8.55 6.71 6.22 5.36
AR [6] 754.02 693.06 466.27 520.35 6.71 6.22 3.51 4.71
ARX [6] 942.61 720.46 514.49 544.76 8.11 6.64 4.08 4.81
ARMAX [6] 940.17 721.54 504.49 544.35 8.25 7.28 3.92 4.73
SVR [16] 703.43 474.38 574.30 393.32 6.23 4.27 5.86 3.74
ANN [23] 750.53 578.05 534.75 345.07 7.20 5.41 5.38 3.48
DBN [69] 639.75 361.63 415.81 350.82 5.95 3.36 4.11 3.41
E-DBN [66] 636.03 551.74 414.90 334.12 5.70 4.78 4.07 3.14
EMD-ANN [70] 748.30 512.59 345.90 299.34 6.66 4.57 3.09 2.90
Proposed DMD 445.79 504.39 329.12 333.23 4.20 5.04 3.13 3.15
TAS Persistence 89.82 157.73 120.47 109.46 7.24 10.22 8.11 7.48
AR [6] 55.40 94.07 80.51 73.86 4.36 6.46 4.86 5.41
ARX [6] 55.66 96.39 83.85 74.65 4.41 6.44 5.25 5.27
ARMAX [6] 56.15 86.81 77.82 77.12 4.36 5.57 4.68 5.43
SVR [16] 60.97 111.89 90.99 79.45 4.81 7.48 5.89 5.55
ANN [23] 69.92 94.40 89.17 72.86 5.42 6.30 6.28 5.24
DBN [69] 63.96 93.81 87.30 75.73 4.98 6.12 6.04 5.15
E-DBN [66] 60.68 109.78 85.19 80.81 4.82 7.28 6.04 5.05
EMD-ANN [70] 63.38 87.41 82.92 80.85 4.87 5.92 5.50 5.63
Proposed DMD 55.38 93.30 77.77 69.06 4.30 7.21 5.16 5.29
QLD Persistence 461.09 489.63 430.46 417.33 5.25 6.25 5.90 5.54
AR [6] 453.43 432.37 305.47 391.03 5.23 6.15 3.71 4.44
ARX [6] 457.26 448.73 325.82 417.55 5.14 6.18 4.04 4.73
ARMAX [6] 418.90 437.95 335.98 413.51 5.07 5.82 4.39 4.70
SVR [16] 282.07 266.39 223.17 298.76 3.65 3.53 3.10 3.93
ANN [23] 299.32 339.93 203.00 263.12 3.61 3.77 3.03 3.46
DBN [69] 228.86 247.56 213.20 251.34 2.78 2.99 2.95 3.40
E-DBN [66] 218.55 259.34 159.45 292.93 2.69 3.33 2.32 3.53
EMD-ANN [70] 273.70 237.58 174.64 248.55 3.28 3.11 2.45 3.27
Proposed DMD 513.10 326.72 176.77 247.36 7.36 3.86 2.42 3.13
VIC Persistence 990.74 669.87 721.85 577.70 9.48 8.40 9.76 8.30
AR [6] 628.93 659.01 557.69 476.06 8.26 8.09 6.38 5.00
ARX [6] 712.61 618.58 583.77 504.52 9.22 8.36 6.83 5.46
ARMAX [6] 784.09 612.64 611.92 538.06 9.40 8.30 7.56 6.52
SVR [16] 587.98 330.93 297.07 391.11 7.16 4.43 4.38 4.50
ANN [23] 811.43 359.03 305.88 347.91 9.32 4.95 4.29 4.79
DBN [69] 915.21 353.02 276.25 389.06 8.79 4.55 3.72 4.85
E-DBN [66] 762.16 343.18 285.14 401.02 9.14 4.49 3.65 4.72
EMD-ANN [70] 781.17 376.12 386.64 332.44 9.07 4.79 5.29 4.15
Proposed DMD 469.67 479.99 401.92 295.90 6.65 5.08 5.38 4.09
distribution of the normalized load series data of different energy forecasting error in percentage.
markets is shown through the box graph in Fig. 10. To visualize the k
discriminating nature of the load data, the normalized data of a specific 100 |Fk Ok |
MAPE =
month, March, are considered in the box graph. The + sign in the k i=1
Ok (26)
graph indicates the outliers.
5.2. Performance evaluation measures • Root mean square error (RMSE): RMSE measures the difference
between the original values to the forecasted values. A lower RMSE
indicates better forecasting result whereas higher RMSE indicates
The performance of the proposed forecasting model is assessed poor forecasting performance.
using three standard metrics. It is defined as,
1 k
238
N. Mohan et al. Applied Energy 232 (2018) 229–244
11
RMSE MAPE
• Mean absolute error (MAE): MAE measures the average difference
10 between the original and forecasted values. It is defined as,
9
Friedman p-value for RMSE: 4.8839e-13 k
Friedman p-value for MAPE: 3.2754e-12 i=1
|Fk Ok |
8 MAE =
k (28)
7
Friedman test ranks
4
5.3. Fixing of window length (L)
3
The window length (L) for Hankelization is fixed based on the
2
characteristics of the load series taken for analysis. For the load data
1 with 48 observations in a day (ie, every half-an-hour observation of
0 load demand), L is fixed at 48. Whereas, for load data with 24 ob-
servations in a day (hourly), the value of L is fixed at 24.
e
AR
AX
BN
ed
nc
AR
SV
AN
DB
AN
os
M
ED
te
D-
AR
op
is
EM
rs
Pr
Pe
Fig. 11. Results for Friedman test for one-day ahead forecasting based on RMSE
5.4. Results and comparisons
and MAPE. The Friedman p-value for RMSE is 4.8839e−13 and MAPE is
3.2754e−12. Case 1: For case 1, the five regions of Australian grid data are
considered and the horizon of forecast is fixed as one-day. To indicate
different seasons, the data from the months January, April, July, and
Table 5 October of 2013 are taken. The forecasting results for one-day ahead
Performance of forecasting models for one-day ahead prediction on May 8, prediction is compared with nine state-of-the-art models including
2007, NSW data. Persistence model, single structure models - AR [6], AR-exogenous
Model RMSE MAPE (ARX) [6], ARMA-exogenous (ARMAX) [6], SVR [16], ANN [23], DBN
[69], ensemble models - ensemble DBN model (EDBN) [66], and EMD
AR [6] 196.78 19.54 combined ANN approach (EMD-ANN) [70]. The average results for the
ARX [6] 190.45 18.87
last one-week prediction in all the four selected months are tabulated in
ARMAX [6] 178.72 18.02
BPNN model [39] 155.71 12.70 Table 4. The best performance values are given in bold. Persistence is a
SVR model [16] 145.87 11.70 basic forecasting model and usually works well for STLF. The persis-
SVR-GA model [45] 150.38 13.52 tence model assumes that the load demand series fluctuates very little
PSO-SVR model [19] 145.69 11.42 during short-interval of prediction [1]. This model is considered as the
SVR-ABC model [39] 145.71 10.70
baseline to evaluate machine learning, ensemble, and data-driven ap-
PSO-BP model [2] 142.26 10.91
Proposed DMD model 127.82 10.18 proaches. The AR and ARX models are estimated using non-iterative,
forward-backward least-squares procedure. The coefficients of linear
ARMAX model is calculated using the iterative prediction-error method.
These regression methods use a polynomial model with identifiable
(a)
10000 Forecast
Original
9500
9000
Demand (MW)
8500
8000
7500
7000
5 10 15 20 25 30 35 40 45
Time (half-hour)
(b)
Forecast
10000
Original
9000
Demand (MW)
8000
7000
Fig. 12. Illustrates forecasting results using the proposed DMD approach. (a) One-day ahead prediction for May 8, 2007 on NSW data, (b) one-week ahead prediction
from May 8, 2007 to May 14, 2007 on NSW data.
239
N. Mohan et al. Applied Energy 232 (2018) 229–244
(a) (b)
300 300
250 250
200 200
RMSE
MAE
150 150
100 100
50 50
0 0
3-Jul-2011 4-Jul-2011 5-Jul-2011 6-Jul-2011 7-Jul-2011 8-Jul-2011 9-Jul-2011 3-Jul-2011 4-Jul-2011 5-Jul-2011 6-Jul-2011 7-Jul-2011 8-Jul-2011 9-Jul-2011
Days Days
(c) (d)
3 0.15
2 0.1
MAPE
1.5
1 0.05
0.5
0 0
3-Jul-2011 4-Jul-2011 5-Jul-2011 6-Jul-2011 7-Jul-2011 8-Jul-2011 9-Jul-2011 3-Jul-2011 4-Jul-2011 5-Jul-2011 6-Jul-2011 7-Jul-2011 8-Jul-2011 9-Jul-2011
Days Days
Fig. 13. Illustrates the performance of the proposed DMD model in terms of forecasting measures. (a) RMSE, (b) MAE, (c) MAPE, (d) running time.
240
N. Mohan et al. Applied Energy 232 (2018) 229–244
Forecast Original
(a) (b)
2400
2000
Demand (MW)
2200
Demand (MW)
1800
2000
1800 1600
Demand (MW)
Demand (MW)
1800 3000
1600
2500
2 4 6 8 10 12 14 16 18 20 22 24 2 4 6 8 10 12 14 16 18 20 22 24
Time (hour) Time (hour)
Fig. 14. Illustrates one-day ahead forecasting results using the proposed DMD approach on North American grid data. Forecasting results for (a) 18-Apr-1990, (b) 25-
Jul-1990, (c) 25-Aug-1990, (d) 14-Dec-1990.
combined ELM model (DE-ELM) [47], basic ARIMA [8], ARMAX [6],
AR [6], ARX [6], WT-MABC-ELM model [42] and EMD-DE-ELM model
150
[47]. The comparison results are tabulated in Table 6 and it can be
observed that the proposed model obtained improved results. This ex-
100
periment highlights the potential of the DMD based STLF model for
50 one-week ahead period.
Experiment 2: Experiment 2 is conducted using North American
0 grid data for December, 1990 to predict the hourly load demand in
January, 1991 (one-month ahead period). The forecasting results and
18-Apr-1990 25-Jul-1990 25-Aug-1990 14-Dec-1990
comparison with SVR [16] and MI [37] model are tabulated in Table 7.
Predicted day This experiment also highlights the computing speed of the proposed
Fig. 15. Visualization of forecasting error of four different days on North DMD model for STLF. It is evident from Table 7 that the DMD model has
American grid data using box graph. effectively predicted the load demand of January 1991. Additionally,
the computation time required for our data-driven method is much
800 10 lesser than the other two referred models. Fig. 14 shows 24-h ahead
forecasting results on four different days, including holidays. It can be
PSF
PSF-NN1 9
700 PSF-NN2 inferred from Fig. 14 that the proposed model predicts the future de-
PSF-NN3 8
600 Proposed DMD mand values regardless of time and day. The graphical visualization of
7 the forecasting error for the selected four days through the box plot is
500 6 shown in Fig. 15.
5
Experiment 3: Experiment 3 demonstrates the potential of the
400
proposed model for forecasting the load demand of one-month period
4
300 using NSW, AEMO market data. The performance advantage is com-
3
pared with PSF [52] and PSF-NN model [53]. PSF-NN is a hybrid model
200
2 which combines the pattern sequence-based forecasting (PSF) with
100 1 neural networks (NN). In [53], three different approaches are devel-
0
oped based on additional inputs. In PSF-NN, the forecasting is done
0
Oct-2011 Dec-2011 Oct-2011 Dec-2011 using the load demand from the previous day, PSF-NN1 predicts the
MAPE
MAE demand using previous week demand data, and PSF-NN2 predicts the
Fig. 16. Illustrates one-month ahead forecasting results on NSW, AEMO data of demand using previous day and week data. The one-month ahead
two different months. forecasting results for two different months are shown in Fig. 16. The
proposed data-driven model has the smallest MAE and MAPE values
than the hybrid and non-hybrid models.
241
N. Mohan et al. Applied Energy 232 (2018) 229–244
Table 8
One-day ahead forecasting results for four different seasons using NSW, AEMO data in year 2010.
Season Day MAPE
ARIMA [8] SVR [16] RBF-NN WT-RBF- Two stage MI- Series ENN Parallel ENN PCC-ENN CPC-ENN CPP-ENN PCP-ENN Proposed
[27] NN WT- BPNN [35] [35] [35] [35] [35] [35] [35] DMD
Spring 18-Mar 17.33 16.87 14.36 12.74 9.05 10.32 2.71 8.32 5.23 20.64 8.32 2.60
23-Mar 32.54 32.54 23.44 15.53 12.65 13.22 2.04 9.65 6.43 34.53 9.65 2.91
Summer 18-Jun 22.43 20.54 17.07 12.33 9.76 11.08 2.65 8.43 5.42 23.65 8.43 2.87
26-Jun 15.43 15.65 15.04 12.84 10.54 10.54 2.34 8.43 5.08 24.33 8.43 2.54
Fall 12-Oct 10.23 19.32 17.87 12.52 9.56 11.54 2.13 8.43 5.24 22.43 8.43 2.15
17-Oct 19.64 19.76 18.04 15.92 13.56 13.22 2.44 11.46 7.64 24.32 11.46 2.23
Winter 19-Dec 23.54 19.65 12.44 11.04 9.56 10.54 2.34 8.54 5.42 25.64 8.54 2.13
25-Dec 37.54 37.54 31.54 16.69 13.87 14.32 2.54 11.23 6.53 38.43 11.23 2.70
Experiment 4: Experiment 4 is conducted to emphasize the com- for STLF task is certified by the lower forecasting errors such as MAPE,
petence of the proposed DMD model to accommodate seasonal and RMSE, MAE, and running time. The competing performance of the
climatic variations in the daily load patterns. Similar to the conditions proposed DMD based model suggests that it can be used as an effective
mentioned in [35], one-day ahead forecasting performance is evaluated tool for forecasting other interdependent variables in power system
by taking two days from all the four seasons of NSW data for 2010. The such as electric price, renewable energy resources (eg; solar, wind), etc.
performance is compared with the results mentioned in [35] and is This is considered as a direction for the future work using the proposed
tabulated in Table 8. The eleven STLF models considered for compar- method.
ison are ARIMA [8], SVR [16], RBF-NN [27], WT combined RBF-NN
(WT-RBF-NN), two-stage MI integrated with WT-BPNN [35], series
ENN [35], parallel ENN [35], PCC-ENN [35], CPC-ENN [35], CPP-ENN 6. Conclusion and future scope
[35], and PCP-ENN [35]. It is clear from Table 8 that the proposed DMD
outperformed ARIMA, SVR, RBF-NN, WT-RBF-NN, two-stage MI-WT- The short-term load forecasting (STLF) plays a key role in the grid
BPNN models in terms of lower MAPE. The proposed DMD model also energy management system. An efficient STLF model is necessary to
obtained comparable results with the parallel ENN model proposed in attain secure supply of reliable electric energy. In this paper, a data-
[35]. driven strategy for short-term load forecasting (STLF) using dynamic
Experiment 5: In experiment 5, the running time required by dif- mode decomposition (DMD) is proposed. The inherent ability of the
ferent models for one-day ahead forecasting is compared. In order to do DMD algorithm to mine the latent spatio-temporal dynamics of the
the real-time prediction in a faster rate, the forecasting model should be time-evolving load data are exploited for accurate future load predic-
computationally effective, less complex and faster. Table 9 shows the tion. The key advantages of the proposed method are:
running time taken by each model for the prediction of load demand on
May 8, 2007. It is clear from Table 9 that the proposed DMD model is • The performance of the method is independent of rigorous training
extremely faster than both single structure and ensemble models. The phases and optimal parameter selection. Also, the method does not
advantage in computing time facilitates the real-time application of the require any hand-crafted feature selection stage.
proposed model for short-term load prediction. Further, the proposed • The method is adaptive to multiple seasonal and cyclic patterns in
data-driven model needs no requirement of parameter settings and load data thus offers an improved generalization ability for load
hence complexity is less unlike other existing models for STLF prediction of different seasons.
[2,39,45]. • The complexity of the method in terms of computation and speed is
The extensive experiments confirms the competitive forecasting much lesser than single structure and ensemble models which fa-
performance of the proposed DMD based model for one-day, one-week, cilitates the real-time application of the proposed model.
and one-month ahead period. The ability of the model to identify the
hidden dynamic characteristics from the load series data leads to an The competence of the proposed DMD based data-driven model is
accurate prediction of the future load. Further, the data-driven model demonstrated through various comparative experiments using the en-
can effectively track the complex load patterns affected by various ergy market data from different geographical locations. The accuracy
exogenous factors. Unlike conventional model-driven methodologies and comprehensibility of the proposed model are satisfactory compared
that assumes a prior system model, the proposed DMD model learns the to other benchmark models for STLF. The significantly lower fore-
spatio-temporal behavior from the available load data. The major ad- casting error measures such as MAPE, RMSE, and MAE confirms the
vantages of the proposed STLF model are, (i) simple, (ii) independent of predictive performance of the proposed model. Therefore it can be
optimum parameter selection, (iii) no requirement of additional feature concluded that the proposed DMD based data-driven strategy is a sui-
extraction stage, (iv) less computation time, and (v) easy to implement. table tool for real-time STLF to build an efficient energy management
The satisfactory performance of the proposed data-driven DMD model system. The application of the proposed DMD model for forecasting the
other interdependent variables in power system such as electric price,
Table 9
Running time required by different models for one-day ahead forecasting.
Models BPNN [39] SVR [16] PSO-SVR PSO-BP SVR-GA SVR-ABC AFCM [71] EMD-SVR- AR DEMD-SVR- AR Proposed DMD
[19] [2] [45] [39] [2] [2]
Running Time 190.4 180.4 165.2 159.9 167.3 181.4 75.3 80.7 76.8 0.13
(sec.)
242
N. Mohan et al. Applied Energy 232 (2018) 229–244
243
N. Mohan et al. Applied Energy 232 (2018) 229–244
regression and time series forecasting. 2014 IEEE symposium on computational Neural Comput 2006;18(7):1527–54.
intelligence in ensemble learning (CIEL). IEEE; 2014. p. 1–6. [70] Liu H, Chen C, Tian H-q, Li Y-f. A hybrid model for wind speed prediction using
[67] AEMO, Australian Energy Market Operator; 2018. URL <http://www.aemo. empirical mode decomposition and artificial neural networks. Renew Energy
com.au>. 2012;48:545–56.
[68] North American Electric Utility Grid); 2018. URL <http://www.ee.washington. [71] An X, Jiang D, Zhao M, Liu C. Short-term prediction of wind power using emd and
edu/class/555/el-sharkawi/datafiles/forecasting.zip>. chaotic theory. Commun Nonlinear Sci Numer Simul 2012;17(2):1036–42.
[69] Hinton GE, Osindero S, Teh Y-W. A fast learning algorithm for deep belief nets.
244