Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Feb 17, 2020

Content may be subject to copyright.

Short-Term Load Forecasting using

EEMD-DAE with Enhanced CNN in Smart Grid

Afrah Naeem, Hira Gul, Arooj Arif, Sahiba Fareed, Mubbashra Anwar, and

Nadeem Javaid

Abstract Traditional grid moves toward Smart Grid (SG). In traditional grids, elec-

tricity was wasted in generation-transmission-distribution. SG is introduced to solve

prior issues. In smart grids, how to utilize massive smart meter’s data in order to im-

prove and promote the efﬁciency and viability of both generation and demand side

is a compelling issue. A good forecasting model makes an acceptable use of all char-

acteristics of the electric loads’ data and also reduces dimensionality of that data.

Many data-driven methods have been proposed in the literature for load forecasting.

In this paper, EEMD based ECNN model is proposed to forecast load of electic-

ity using AEMO data. From the results, ECNN outperforms benchmark methods

especially by applying EEMD for decomposition and DAE for feature extraction.

1 Introduction

A huge amount of electricity consumption data has been collected due to preva-

lent popularity of smart grids. Smart grid combines information and communication

technology in the existing energy grids to provide optimum power with remarkable

efﬁciency by interchanging the real time data between consumers and suppliers. For

the time being, the liberalism of power industry on the consumer side, has vastly

grown worldwide. How to utilize massive data of smart meters in order to improve

and promote the efﬁciency and viability of both generation and demand side is a

compelling issue [1]. Short-term load forecasting plays a vital role in energy distri-

bution and management of modern power system, as it is necessary to ensure the

balance between generation and demand. Also, accurate load forecasting becomes

an important issue with the increased complexity of smart grids. Due to imprecise

forecasting, its effect in the planning and scheduling of demand are drastic. For

Afrah Naeem, Hira Gul, Arooj Arif, Sahiba Fareed, Mubbashra Anwar, and Nadeem Javaid (Cor-

responding Author)

COMSATS University Islamabad, email: nadeemjavaidqau@gmail.com

1

2 Afrah et al.

example, an underestimation of electricity results in power outage whereas overes-

timation drives the units to be wasted causing excessive energy supply [2]. A good

forecasting model makes an acceptable use of all characteristics of the electric loads

data and also reduces dimensionality of that data. In other words, the goal of electric

load forecasting is to ensure stable use of supply while keeping its operational cost

as low as possible [3].

Electricity load forecasting can be categorized into four main types summarized

as follows: Very Short-Term Load Forecasting (VSTLF) ranges from few minutes

to hours, Short-Term Load Forecasting (STLF) performs hours or days prediction,

Medium-Term Load Forecasting (MTLF) falls between weeks or months whereas

Long-Term Load Forecasting (LTLF) lies between years. STLF aims to assure relia-

bility of the electric power system and adjust it to prevent from loss of power failure

and overloading. In the last decades, STLF algorithms have been widely studied [4].

In this paper, we focus on day-ahead load forecasting i.e. STLF.

Since 1940s, numerous statistical based linear time series forecasting approaches

have been proposed whose shared goal is to practice time series analysis for deduc-

ing the future energy requirement. The most acknowledged methods are based on

Auto Regressive Integrated Moving Average (ARIMA) [5], linear regression [6]

and Kalman ﬁltering [7] etc. These statistical models acts as a baseline for electric

load forecasting nowadays [8]. At present, traditional methods are developed, the-

ories become mature and approaches are simple, which have numerous composite

advantages. These traditional methods do not perform well when dealing with non-

linear data as they carry out linear based analysis and establish quantitative relation

between data.

The non-linearity of time series data can be handled with the decomposition tech-

niques. EMD is frequently applied to remove noisy part of load signals in the pre-

processing phase of short-term load forecasting. However, important issues with

EMD are end data points extending and mode mixing problem that needs to be

considered [9]. In other words, EMD lacks its mathematical foundation. In [10],

characteristic waves were used to solve this problem. Besides this, it may devi-

ate from actual output. In [11, 12], authors proposed mirror extending method to

help resolve end data points extending difﬁculty. End mirror extending for high fre-

quency and least square polynomial for low frequency were used. In [13], authors

combined EMD with Multiplex Overlap Discrete Wavelet Transform (MODWT)

that tried to overcome this problem. MODWT applied low and high pass ﬁlters to

the input which is then transferred to EMD to further denoise data. Although, this

creates computational complexity. So, there is room for improving solutions for end

point problem of EMD method.

Some artiﬁcial intelligence based methods like ANN, deep neural network and

SVR have been proposed with improved efﬁciency by utilizing more reﬁned fea-

tures. However, these features are directly given into model for training without

separate pre-processing. Thus, we expect to build a model that separately performs

pre-processing and feature engineering in order to ensure performance with well-

deﬁned features. Extreme Learning Machine (ELM) performs forecasting in one

go, though, it is highly sensitive to initial parameters. It reduces computational time

Short-Term Load Forecasting using EEMD-DAE with Enhanced CNN in Smart Grid 3

however, it cannot update biases, weights and parameters. Radial Basis Function

(RBF) kernel used in ELM also assigns same weights to all attributes because they

are treated equally in distance computation. Also, EMD based ELM gives less ac-

curate results for the large dataset [14].

This paper intents to propose a model for electricity forecasting using ECNN:

DAE is ﬁrst used to perform feature extraction then ECNN is used to predict load.

The proposed model is then compared with the performance of two different meth-

ods to see which technique gives best results. The limitations and future challenges

are also identiﬁed for further research.

The remaining paper is structured as follows: in Section II, the problems of each

paper are discussed critically. Section III introduces a model proposed. Section IV

describes the data and forecasting results. Limitations and future research challenges

are discussed in Section V.

2 Literature survey

In paper [14], authors considered a problem that it is tough for an individual model

to perform well in all cases; most probably when dealing with non-linear data. One

kernel function may not capture all properties of data. When dealing with Extreme

Learning Machine (ELM), the choice of kernel is most serious problem. They pro-

posed a model in which to reduce noise, EMD is used that diminish noise inference

of original load data. That denoised data is then decomposed into training and test-

ing set. The processed data is added to mixed kernel-based ELM model to perform

forecasting. Mixed kernel is composed of RBF kernel and UKF kernel function. If

forecasting results are greater than accuracy limitation then it gets ﬁnal results, oth-

erwise goes back to EMD and repeats the whole process. The proposed model is ﬁrst

compared with three methods: RBF-ELM, UKF-ELM and Mixed-ELM. Then com-

pared with three benchmark methods: MFES, ESPLLSVM and combined method.

First three performed well at some speciﬁc time than proposed method. Whereas, on

comparison with benchmark methods, the proposed method has improved electricity

load forecasting accuracy undoubtedly. Four performance metrics are chosen to per-

form simulation, which are: RMSE, MAPE, MAE and AE. Each of them evaluate

the methods from different facets. Whereas, AE is neglected as it contains negative

values. ELM performs forecasting in one go, though, it is highly sensitive to initial

parameters. It reduces computational time however, it cannot update biases, weights

and parameters. EMD based ELM gives less accurate results for large dataset.

In paper [15], authors examined a problem that an important challenge with EMD

is end data points extending problem that needs to be considered. In [10], character-

istic waves were used to solve this problem. Beside this, it may deviate from actual

output. Also, the weights in single feedforward neural network are conventionally

tuned by BP algorithm. BP has gradient descent problem i.e. trapped in local min-

ima. It is also time consuming due to large number of iterations. The proposed model

is comprised of two phases: pre-training phase and incremental learning phase. In

4 Afrah et al.

pre-training, MODWT is applied to decompose signal into several IMFs and one

residue. The training vector is then passed to RVFL network and trained for each

obtained sub-series. A new input matrix is constructed that combines three aspects:

original time series data, temperature data and prediction results of all sub-series. In

incremental learning phase, this matrix is passed so new weight matrix can be con-

structed. The testing data is used to check the error level. If error decreases than cur-

rent error level, updates remain kept otherwise weight matrix is not changed. Two

error measures are considered to validate accuracy of forecasting model: RMSE

and MAPE. The proposed method is compared with eight benchmark learning algo-

rithms: GLMLF-B, SLFN, RVFL, RF, EMD-RF, EMD-SLFN, persistence method

and EMD-RVFL. The critical distance calculated is 3.1. Despite of RVFL has great

approximation ability and efﬁcient generalization performance although, when it

comes to deep learning, there is tradeoff between functional links and number of

hidden neurons in RVFL. The advantages of functional links decrease as number of

hidden neurons increase in order to train model sufﬁciently [15, 16]. DWT is used

to analyze non-linear and non-stationary data. However, this technique is based on

assumption that processed data is linear and stationary. It is highly ﬂexible; wavelet

functions can be openly chosen. It also allows good localization in time domain.

Despite this, it takes longer compression time. In paper [16, 17], EMD is used to

denoise data. Also, one classiﬁer is used to deal with linear and non-linear series i.e.

IMFs and residue.

In paper [18], authors addressed a problem that original non-linear time series

data often lost some useful information due to dissonance; it contains some noisy

part which affects the forecasting results. ‘A divide and conquer’ approach i.e. EMD

is used that decomposes load demand data into several IMFs and one residue. One

training matrix is then constructed for IMFs and residue which is then passed as

input to DBN. DBN is trained to obtain predicted results. By using summation or

linear neural network, prediction results of DBN are then combined to devise an

ensemble output for time series. Performance comparison for day-ahead and half-

an-hour load forecasting is conducted. Two error measures are considered to check

precision of forecasting model: RMSE and MAPE. The proposed method is com-

pared with nine benchmark algorithms: persistence method, ANN, SVR, RF, DBN,

EDBN, EMD-RF, EMD-SVR and EMD-ANN. The critical distance calculated is

3.0.

In paper [19], problem was addressed i.e. the neural network-related algorithms

are mostly practiced to express non-linear forecasting relationship. They have solved

underﬁtting problem successfully, although, overﬁtting is a major disadvantage of

neural networks when it comes to large number of hidden nodes. Box-Cox nor-

malization and copula analysis were applied in data preprocessing. A forecasting

based on DBN was constructed to predict the power load. A comparison analysis

between SVR, NN, classical DBN, R-DNN, SOM-SVM and ELM was conducted.

Performance of proposed model was evaluated using day-ahead and week-ahead

peak load forecasting using three performance metrics namely MAPE, RMSE and

Hit Rate (HR). Backpropagation was used for ﬁne tuning of parameters. Although,

it may trap in local optima while search which then increase computational time

Short-Term Load Forecasting using EEMD-DAE with Enhanced CNN in Smart Grid 5

and affect performance accuracy of the model. In both papers, DBN was used for

forecasting electricity load. However, major drawback of DBN is that it may lead to

overﬁtting problem when performing feature engineering step.

Another paper [20], addressed a problem that due to increase in market compe-

tition, more active and unpredictable electricity market, renewable requirements of

integration, shareholders are more interested in some alternative to point forecasting

techniques. Also, most common technique used in area of daily load forecasting is

deep learning theory. Still, there are two common issues about the theory; overﬁt-

ting and more computation time consumption. To solve this, the improved WNN

and generalized ELM are used in probabilistic load forecasting problem. Uncertain-

ties of noise in data are considered which is a consequence of non-iterative training

machine. A combined GELM, wavelet preprocessing and bootstrapping model is

proposed. Two tests; Friedman and post-hoc tests are used to validate the approach.

Four performance indicators are used for error measures, Prediction Intervals’ Nom-

inal Conﬁdence (PINC), Prediction Intervals Coverage Probability (PICP), ECR and

Average Coverage Error (ACE). The proposed model is then compared with four

methods: RELM/Bootstrapping, BP/Bootstrapping, Quantile Regression and Dol-

phin/Bootstrapping. GELM was applied as SLFN, so it cannot encode more than

one layer of abstraction, it also has overﬁtting troubles. It has fast training time but

slow evaluation speed.

In paper [7], authors addressed pain in the neck that by using only individual

time series model, we cannot predict good results. A combination or hybrid of two

models is more effective than individual model. In addition. EMD has an end effect.

An improved EMD (IEMD) was used for reducing information by decomposing sig-

nals into residual term and non-linear series with calendar effects. ARIMA model

was used for the ﬁrst segment as it performs well with the linear data. WNN was

then used to capture calendar effects i.e. non-linear series. At last, Fruit Fly Opti-

mization Algorithm (FOA) was used for optimization. To make a fair comparison,

four performance metrics were used as measurement index; MAPE, MAE, MPE

and RMSE. Load data was collected from AEMO and New York market. For com-

parison, schemed model was compared with two methods: WTNNEA, WGMIPSO.

Paper [21] addressed a problem that, most researches only focused on one-step load

forecasting and also ignore time feature that is crucial for performing accurate fore-

casting. For this, a Multi Scale Convolutional Network (MS-CNN) was proposed

to extract high-level features. For strengthning periodical description and encode

load data for improving prediction, the periodical coding strategy was designed. In

addition, the deﬁciencies of traditional time coding approaches were evaluated. Au-

thors also proposed TCMS-CNN which optimizes the structure of CNN and raise

the accuracy of multi-step load forecasting by extracting more relationships with

periodical characters. Three error measures were chosen to evaluate performance:

RMSE, MAE and MAPE. To verify MS-CNN, it was compared with some popular

models; LSTM, SVR and ResNet on one step prediction. From the results, it was

concluded that TCMS-CNN was improved by 3.54. For direct multi step load fore-

casting, the performance of TCMS-CNN was not satisﬁed. Its network structure is

also complex that cause longer execution time for training and inference process.

6 Afrah et al.

In paper [22], authors focused on issue that single artiﬁcial intelligence mod-

els that deals with non-linear patterns may stuck in local optima and also failed to

achieve satisfactory performance. Authors proposed hybrid model, that ﬁrst decom-

pose raw wind speed series and send it to LSTM which acts as predictor. RELM

improves performance of model signiﬁcantly. At last, IEWT is employed for outlier

detection and correction of the forecasting series. Four performance evaluation in-

dexes are chosen: the MAE, MAPE, RMSE and SDE. The proposed hybrid model

has high computational time. Even if LSTM is an improved gate version RNN, it has

still some limitations. It has probability of loosing important information in training

process. Also, it does not possess the ability to perform parallel processing, which

results in more time consumption and wastage of resources. The problem addressed

in this [23] is that, in order to save power energy and logically schedule electric gen-

erator units, it is necessary to forecast day-ahead overall load of electricity. Three

kinds of data were collected for solve problem encountered: historical load data,

weather and season parameters. SDA was used for feature extraction because of its

ability to adapt numerical inputs and impressiveness. SDA was composed of sub-

networks which receives data collected and extract some high-level features. The

reﬁned one were then fed into SVR to forecast electricity load. This paper used

MAPE as performance indicator to evaluate error measures. Proposed model was

compared with SVR and ANN method. SVR has good regularization capability and

can easily handle non-linear data. It also deals with classiﬁcation and regression

problem instead feature scaling must be done before applying SVM, which is not

carried out. SVR has long training time. Also, algorithmic complexity and memory

requirements of SVM are very high. One key issue addressed in this [24] paper is

that it is important for stable power supply to predict electric load accurately. By

using conventional statistical methods, it is not easy to predict exact power con-

sumption for short-term load forecasting for university buildings due to their differ-

ent usage patterns. For this, the data collected serve as input variable for forecasting

model. In pre-processing, temperature data is adjusted. The processed data is passed

to simple moving average method which is used to smoothen short-term ﬂuctuations

and highlight long-term trends for time-series data. RF is used to perform predic-

tion which yields improved performance than SVR and ANN because of conducting

Time Series Cross Validation (TSCV). To evaluate performance of forecast model,

two performance metrics were chosen; RMSE and MAE. Finally, performance is

evaluated using TSCV. The moving average method used predicts future demand

based on assumption of past ﬂuctuations or trends. However, in case if assumption

is not valid, this method gives poor result.

In [25], authors highlighted a problem that Demand Response (DR) aims at im-

proving the efﬁciency of smart grids. One key issue of DR is the calculation of load

baselines. The aim of this paper was to propose a generic method to forecast a fair

baseline in DR programs. A historical data of ofﬁce buildings was used to determine

optimal functional relationship between inputs and train SVR model to predict out-

come. Dry bulb temperature was used as input to model in order to improve accuracy

of SVR. The baseline was calculated up to 8 hours. The gap between baseline and

actual load was calculated during DR event. AE, Mean Error (ME) and MAE were

Short-Term Load Forecasting using EEMD-DAE with Enhanced CNN in Smart Grid 7

taken as evaluation standards during a response time. Comparison was done with

eight existing methods for four ofﬁce buildings. SVR model is computationally de-

manding and it requires a lot of training data. In addition, it also suffers from the

curse of dimensionality.

Paper [26] focuses on problem that due to volatility and stochasticity of loads

data, load forecasting for an individual household is much more serious problem.

The sparse code features are used to improve the accuracy of individual meter load

forecasting. Error analysis is done on MAPE error distribution of different meters.

The next-day and next-week total load forecasting is done and evaluated using er-

ror measures i.e. MAE, RMSE and MAPE. In paper [27], addressed a problem

that some of the traditional forecasting models ignores the necessity of data pre-

processing. Also, limitations of individual model are often ignored, which leads to

highly inaccurate prediction. Data was ﬁrst de-noised using decomposition and en-

semble technique. A novel deciding weight method based on a swarm intelligence

and the leave-one-out strategy was developed to combine the individual models. The

accuracy of short-term wind speed was improved based on combined novel model.

Three performance metrics were used to evaluate combined novel model: RMSE,

MAE and Sum of Error Square (SSE). The Diebold–Mariano (DM) test was used

to evaluate and verify the performance of proposed combined model. Although the

proposed model had high computational time. Some more work using artiﬁcial in-

telligence based methods like ANN and deep neural network have been done in

literature [28]-[34].

3 Methodology

In this paper, EEMD based ECNN is proposed to forecast electricity load as shown

in Figure 1. The applied approach consists of the following steps:

1. The time series data is ﬁrst normalized using min-max normalization and stan-

dardization (or Z-score normalization).

2. After normalization, input data is then transferred to EEMD. EEMD decom-

poses data into several IMFs and one residue.

3. For further denoising that data and performing feature extraction, we construct

a training matrix for each IMFs and one residue and pass it to DAE.

4. This denoised data is then transferred to ECNN to formulate ﬁnal load fore-

casting results.

3.1 Dataset

The electricity load data sets from Australian Energy Market Operator (AEMO)

[35] are collected for comparison. AEMO is an electric power and gas industry. It

generates, transmits and distributes electricity to all Australians. It generates a large

8 Afrah et al.

Table 1: Overview of Related Work

Techniques Objectives Data source/ achievements Limitations

EMD mixed kernel-based ELM [1] Electric load forecasting

Improved accuracyby using EMD, RBF and

UKF kernel function. AEMO data collected

EMD may cause mode mixing problem and elm

cannot tune parameters, bias and weights

EMD based Ensemble DBN [2] Short-term and very short-term forecasting

Efﬁciently extractfeatures in an

unsupervised way and better performance for

half-an-hour ahead forecasting. AEMO data collected

Time consuming, EMD has end points extendingproblem

MODWT-EMD-RVFL[3] Short-term load forecasting Betterperformance in terms of

accuracy,efﬁciency. AEMO data collected

RVFLtradeoffs between functional links and number of hidden neurons

SDA-SVR[4] Day-ahead load forecasting Mainlyfocus on feature

extraction to extract well reﬁned features

DWT takeslonger compression time

SMA-RF-CV [6] Short-term load forecasting Loadprediction for university building with acceptable accuracy long training time and feature scaling not done

IWNN-GELM [7] Probabilistic load forecasting Better accuracy and reliability,beat benchmark methods Simple moving averagegives poor result in case of false assumption

TCMS-CNN [8] Multi-step short-term load forecasting Time coding wasconsidered GELM cannot encode more than one layer of abstraction, overﬁtting troubles

CM-DBN [21] Short-term power load

Predicts the power load using

day-ahead, week-ahead forecasting

and week-ahead peak load forecasting

the performance of TCMS-CNN was not satisﬁed,

network structure is complex and time consuming

IEMD-ARIMA-WNN-FOA [22] Short-term load Better accuracy BP may trap in local optima

IEMD-SVR [23] Short-term load Calculate load demand baselines,

take ambient temperature of two hours as input

High computational time

ICEEMDAN-GWOcombined model [24] Short-term wind speed Improve the accuracyof individual meter load forecasting Computationally demanding, requires a lot of training data

EWT-LSTM-RELM-IEWT[25] Short-termwind speed Improved performance More computation time, lose important information in the training process

amount of data about electricity load, its price, demand, generation and distribution,

etc. Load data of 2018 is taken from AEMO, which is used to train and test the

proposed method.

3.2 Pre-processing

Load data collected from AEMO is ﬁrst normalized using min-max normalization

and then standardization normalization. Min-max rescales the range of data [0,1].

Standardization sets mean of data to 0 and standard deviation to 1. It gives equal con-

sideration for each feature. For regression data, standardization improves numerical

stability of model and may speed up training process [36].

3.3 Noise reduction

EEMD is used to reduce noise inference of original load signal and decompose data

into several IMFs and one residue. A practical approach to obtain swift frequency

data from non-linear and non-stationary datasets. IMF function has mean value of

Short-Term Load Forecasting using EEMD-DAE with Enhanced CNN in Smart Grid 9

zero, along with one extreme between two modes [14]. After EEMD, decomposed

data is then passed to feature engineering phase.

3.4 Feature Extraction

Feature extraction is used to extract some features; it creates a new, smaller set of

features that still captures useful information in order to give more accurate results

than original data. In literature, different techniques are used for feature extraction.

DAE is used for feature extraction in the proposed model. During training, loss

function is deﬁned similar to Root Mean Square Error (RMSE) that was deﬁned

previously in convolutional autoencoder. At every iteration of training, network will

compute a loss between noisy data output by decoder and ground truth (de-noisy

data). It also tried to minimize loss or difference between reconstructed data and

original noise-free data.

3.5 ECNN

We propose Enhanced version of Convolutional Neural Network (ECNN), in which

additional hidden layers are added. In order to avoid overﬁtting problem and in-

crease performance of ECNN, parameters are adjusted. ECNN is composed of two

parts; in ﬁrst part, ECNN uses hidden layers to learn the given input features with

weight sharing property. Whereas in the second part, ECNN ﬂattens the output of

above layers and perform forecasting. In ECNN, one convolution layer with sixty-

Fig. 1: Proposed Model

four ﬁlters of two by two matrix is added using relu as activation function. After

convolution layer, two dense layers are added. Dense layer represents a multiplica-

tion vector matrix and the values in the matrix are trainable parameters which get

updated during backpropagation. After dense layer, a dropout layer appear which is

used for regularization to prevent overﬁtting problem. During each iteration of train-

ing process, a neuron is temporarily discarded with some certain probability. The

10 Afrah et al.

reason behind dropout is to prevent the network being dependent on a small num-

ber of neurons and force every neuron to operate independently which increases

the accuracy of ECNN. Lastly, two Max Pooling layers are added which reduce

the number of parameters, computation time, dimensionality, and controlling over-

ﬁtting problem by reducing the dimentional size of the network. In Max Pooling

layer, at each iteration, maximum number of parameters are taken out and the rest

are dropped, which both compresses the training time as well as combats overﬁtting.

In ﬂatten layer, output of layers is ﬂattened as discussed above to deliver into the

next fully connected layers to perform forecasting.

4 Simulation results

For all-time series data sets, data is linearly scaled between [0,1]. To perform simu-

Fig. 2: Results of ECNN compared with CNN

lation, toolbox named as deep learning was used for neural network. The implemen-

tation of DBN, EMD mixed kernel-based ELM and proposed EMD based ECNN is

performed in Python. There are two types of performance matrix that are used for

error measurement of proposed model i.e. RMSE and MAPE.

Short-Term Load Forecasting using EEMD-DAE with Enhanced CNN in Smart Grid 11

The values on x-axis of EMD based DBN and ELM are the number of observa-

tions performed which are then compared with actual dataset. On one hand, DBN

was applied on subset of actual dataset. Activation function chosen for DBN is lin-

ear. On y-axis electricity power (MW) is plotted. MAPE value of DBN computed

Fig. 3: Results of ECNN compared with DBN

is 2.67 and RMSE is 1.626. Also, MAPE value of EMD mixed kernel-based ELM

is 3.87 whereas RMSE is 0.4.ELM on the other hand is applied for comparison with

proposed model. ELM is a single feedforward neural network.ELM does not fully

capture values from dataset; it does not cover up whole search space for forecasting

on basis of features extracted. From the results, it is clear that ECNN outperforms

benchmark methods.

12 Afrah et al.

Fig. 4: Results of MLP

5 Conclusion

Accurate electricity forecasting is essential for the stability of grid. ML is a well-

known technique that is used for electricity forecasting. This paper proposed a

model for electricity load forecasting using EEMD-DAE-ECNN. The performance

is then compared with three methods: DBN, ELM and original CNN. From the re-

sults, it has been seen that ECNN outperforms benchmark methods especially by

applying EEMD for decomposition and DAE for feature extraction.

References

1. Wang, Y., Chen, Q., Hong, T. and Kang, C., 2018. Review of smart meter data analytics: Ap-

plications, methodologies, and challenges. IEEE Transactions on Smart Grid, 10(3), pp.3125-

3148.

2. Fan, S. and Hyndman, R.J., 2011. Short-term load forecasting based on a semi-parametric

additive model. IEEE Transactions on Power Systems, 27(1), pp.134-141.

3. Rahman, S. and Bhatnagar, R., 1988. An expert system based algorithm for short term load

forecast. IEEE Transactions on Power Systems, 3(2), pp.392-399.

4. Hong, T. and Fan, S., 2016. Probabilistic electric load forecasting: A tutorial review. Interna-

tional Journal of Forecasting, 32(3), pp.914-938.

Short-Term Load Forecasting using EEMD-DAE with Enhanced CNN in Smart Grid 13

5. Lee, C.M. and Ko, C.N., 2011. Short-term load forecasting using lifting scheme and ARIMA

models. Expert Systems with Applications, 38(5), pp.5902-5911.

6. T. Hong , 2010. Short Term Electric Load Forecasting, Ph.D. thesis, Graduate Program of

Operation Research and Dept. of Electrical and Computer Engineering, North Carolina State

University.

7. Zhang, J., Wei, Y.M., Li, D., Tan, Z. and Zhou, J., 2018. Short term electricity load forecasting

using a hybrid model. Energy, 158, pp.774-781.

8. Yu, C.N., Mirowski, P. and Ho, T.K., 2016. A sparse coding approach to household electricity

demand forecasting in smart grids. IEEE Transactions on Smart Grid, 8(2), pp.738-748.

9. Gao, Y., Ge, G., Sheng, Z. and Sang, E., 2008, May. Analysis and solution to the mode mixing

phenomenon in EMD. In 2008 Congress on Image and Signal Processing (Vol. 5, pp. 223-

227). IEEE.

10. Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q., Yen, N.C., Tung, C.C.

and Liu, H.H., 1998. The empirical mode decomposition and the Hilbert spectrum for non-

linear and non-stationary time series analysis. Proceedings of the Royal Society of London.

Series A: Mathematical, Physical and Engineering Sciences, 454(1971), pp.903-995.

11. Rilling, G., Flandrin, P. and Goncalves, P., 2003, June. On empirical mode decomposition and

its algorithms. In IEEE-EURASIP workshop on nonlinear signal and image processing (Vol.

3, No. 3, pp. 8-11). NSIP-03, Grado (I).

12. Qingjie, Z., Huayong, Z. and Lincheng, S., 2010, March. A new method for mitigation of

end effect in empirical mode decomposition. In 2010 2nd International Asia Conference on

Informatics in Control, Automation and Robotics (CAR 2010) (Vol. 1, pp. 400-403). IEEE.

13. Qiu, X., Suganthan, P.N. and Amaratunga, G.A., 2018. Ensemble incremental learning ran-

dom vector functional link network for short-term electric load forecasting. Knowledge-

Based Systems, 145, pp.182-196.

14. Chen, Y., Kloft, M., Yang, Y., Li, C. and Li, L., 2018. Mixed kernel based extreme learning

machine for electric load forecasting. Neurocomputing, 312, pp.90-106.

15. Qiu, X., Suganthan, P.N. and Amaratunga, G.A., 2018. Ensemble incremental learning ran-

dom vector functional link network for short-term electric load forecasting. Knowledge-

Based Systems, 145, pp.182-196.

16. Zhang, L. and Suganthan, P.N., 2016. A comprehensive evaluation of random vector func-

tional link networks. Information sciences, 367, pp.1094-1105.

17. Zhang, P.B. and Yang, Z.X., 2017. A new learning paradigm for random vector functional-

link network: RVFL+. arXiv preprint arXiv:1708.08282.

18. Qiu, X., Ren, Y., Suganthan, P.N. and Amaratunga, G.A., 2017. Empirical mode decompo-

sition based ensemble deep learning for load demand time series forecasting. Applied Soft

Computing, 54, pp.246-255.

19. Ouyang, T., He, Y., Li, H., Sun, Z. and Baek, S., 2019. Modeling and Forecasting Short-Term

Power Load With Copula Model and Deep Belief Network. IEEE Transactions on Emerging

Topics in Computational Intelligence, 3(2), pp.127-136.

20. Raﬁei, M., Niknam, T., Aghaei, J., Shaﬁe-Khah, M. and Catalao, J.P., 2018. Probabilistic

load forecasting using an improved wavelet neural network trained by generalized extreme

learning machine. IEEE Transactions on Smart Grid, 9(6), pp.6961-6971.

21. Deng, Z., Wang, B., Xu, Y., Xu, T., Liu, C. and Zhu, Z., 2019. Multi-Scale Convolutional

Neural Network With Time-Cognition for Multi-Step Short-Term Load Forecasting. IEEE

Access, 7, pp.88058-88071.

22. Li, Y., Wu, H. and Liu, H., 2018. Multi-step wind speed forecasting using EWT decompo-

sition, LSTM principal computing, RELM subordinate computing and IEWT reconstruction.

Energy Conversion and Management, 167, pp.203-219.

23. Tong, C., Li, J., Lang, C., Kong, F., Niu, J. and Rodrigues, J.J., 2018. An efﬁcient deep model

for day-ahead electricity load forecasting with stacked denoising auto-encoders. Journal of

Parallel and Distributed Computing, 117, pp.267-273.

24. Moon, J., Kim, K.H., Kim, Y. and Hwang, E., 2018, January. A short-term electric load fore-

casting scheme using 2-stage predictive analytics. In 2018 IEEE International Conference on

Big Data and Smart Computing (BigComp) (pp. 219-226). IEEE.

14 Afrah et al.

25. Chen, Y., Xu, P., Chu, Y., Li, W., Wu, Y., Ni, L., Bao, Y. and Wang, K., 2017. Short-term

electrical load forecasting using the Support Vector Regression (SVR) model to calculate the

demand response baseline for ofﬁce buildings. Applied Energy, 195, pp.659-670.

26. Yu, C.N., Mirowski, P. and Ho, T.K., 2016. A sparse coding approach to household electricity

demand forecasting in smart grids. IEEE Transactions on Smart Grid, 8(2), pp.738-748.

27. Song, J., Wang, J. and Lu, H., 2018. A novel combined model based on advanced optimization

algorithm for short-term wind speed forecasting. Applied energy, 215, pp.643-658.

28. Samuel, O., Alzahrani, F.A., Hussen Khan, R.J.U., Farooq, H., Shaﬁq, M., Afzal, M.K. and

Javaid, N., 2020. Towards Modiﬁed Entropy Mutual Information Feature Selection to Fore-

cast Medium-Term Load Using a Deep Learning Model in Smart Homes. Entropy, 22(1),

p.68.

29. Khalid, R., Javaid, N., Al-zahrani, F.A., Aurangzeb, K., Qazi, E.U.H. and Ashfaq, T., 2020.

Electricity Load and Price Forecasting Using Jaya-Long Short Term Memory (JLSTM) in

Smart Grids. Entropy, 22(1), p.10.

30. Mujeeb, S. and Javaid, N., 2019. ESAENARX and DE-RELM: Novel schemes for big data

predictive analytics of electricity load and price. Sustainable Cities and Society, 51, p.101642.

31. Sana Mujeeb, Turki Ali Alghamdi, Sameeh Ullah, Ayesha Fatima, Nadeem Javaid and Tanzila

Saba, ”Exploiting Deep Learning for Wind Power Forecasting based on Big Data Analytics”,

Applied Sciences, Volume: 9, Issue: 20, Article Number: 4417, Pages: 1-19, Published: Oc-

tober 2019, ISSN: 2076-3417, DOI: https://doi.org/10.3390/app9204417.

32. Aqdas Naz, Nadeem Javaid, Muhammad Babar Rasheed, Abdul Haseeb, Musaed Alhussein

and Khursheed Aurangzeb, ”Game Theoretical Energy Management with Storage Capacity

Optimization and Photo-Voltaic Cell Generated Power Forecasting in Micro Grid”, Sustain-

ability, 2019, Volume: 11, Issue: 10, Article Number: 2763, Pages: 1-22, Published: May

2019, ISSN: 2071-1050. DOI:https://doi.org/10.3390/su11102763.

33. Aqdas Naz, Muhammad Umar Javed, Nadeem Javaid, Tanzila Saba, Musaed Alhussein

and Khursheed Aurangzeb, ”Short-Term Electric Load and Price Forecasting Using En-

hanced Extreme Learning Machine Optimization in Smart Grids”, Energies, Volume: 12,

Issue: 5, Article Number: 866, Pages: 1-30, Published: March 2019, EISSN 1996-1073.

DOI:https://doi.org/10.3390/en12050866.

34. Sana Mujeeb, Nadeem Javaid, Manzoor Ilahi, Zahid Wadud, Farruh Ishmanov and Muham-

mad Khalil Afzal, ”Deep Long Short-Term Memory: A New Price and Load Fore-

casting Scheme for Big Data in Smart Cities”, Sustainability, Volume: 11, Issue: 4,

Article Number: 987, Pages: 1-29, Published: February 2019, ISSN: 2071-1050. DOI:

https://doi.org/10.3390/su11040987.

35. Australian energy market operator, 2016. URL http://www.aemo.com.au/.

36. Z. Zixuan, 2019. URL https://towardsdatascience.com/understand-data-normalization-in-

machine-learning-8ff3062101f0.