Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Mar 22, 2019

Content may be subject to copyright.

Electricity Load Forecasting for Each Day

of Week Using Deep CNN

Sajjad Khan1, Nadeem Javaid1(B

), Annas Chand2,

Abdul Basit Majeed Khan3, Fahad Rashid4, and Imran Uddin Afridi1

1COMSATS University Islamabad, Islamabad 44000, Pakistan

nadeemjavaidqau@gmail.com

2COMSATS University Islamabad, Abbottabad Campus,

Abbottabad 22010, Pakistan

3Abasyin University Islamabad, Islamabad 44000, Pakistan

4Bahria University Islamabad, Islamabad 44000, Pakistan

http://www.njavaid.com

Abstract. In smart grid, precise and accurate electricity load forecast-

ing is one of the most challenging tasks. It is due to the high volatile, non-

stationary and non-linear behavior of electricity load data. In this paper,

a Deep Convolution Neural Network (DCNN) model is proposed to fore-

cast the electricity load for each day of the week of Victoria (Australia).

To forecast the electricity load for one day of the week, we analyzed the

electricity load data consumed on the same day for the previous three

months. To show the usefulness of our proposed scheme, comparison is

made with the state of the art forecasting models namely recurrent neural

network, extreme learning machine, CNN and auto regressive integrated

moving average. Results show that the proposed DCNN has the low-

est mean absolute percentage error, mean absolute error and root mean

square error of 2.1%, 138.771 and 116.417.

Keywords: Smart grid ·For e casting ·Energy management ·

Neural Network ·Deep learning

1 Background

Forecasting the accurate electricity load demand plays a key role in demand

side management. It helps the generation companies to develop a tolerable pro-

duction plan and avoid energy wastage. For electricity load forecasting many

techniques have been used in literature. These prediction techniques are mainly

classiﬁed into three main groups, i.e., data driven techniques, classical techniques

and Artiﬁcial Intelligence (AI) techniques. Data driven techniques predict the

desired outcomes on the basis of analysing historical data. Classical techniques

comprises of the statistical and mathematical methods like Autoregressive Inte-

grated Moving Average (ARIMA), Seasonal ARIMA and random forest etc.

c

Springer Nature Switzerland AG 2019

L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1107–1119, 2019.

https://doi.org/10.1007/978-3-030-15035-8_107

1108 S. Khan et al.

Such methods are eﬀective for one step forecast. AI methods inhibit the behavior

of biological neurons e.g., Feed Forward Neural Network (FFNN), Convolution

Neural Network (CNN) and Long Short Term Memory (LSTM) etc.

In [2], Bayesian network and Multi Layer Perceptron (MLP) are used for

short and long term load forecasting. Bayesian network performs behavioral

analytics using the data obtained from the smart meters. Number of experi-

ments were performed on the obtained data. For Short Term Load Forecasting

(STLF), multiple linear regression is used in [3]. However, it has the limitation

that it can not be used for long term prediction. The authors in [4] performed

electricity load forecasting on the basis of weather data obtained from the mete-

orological department. In this work, load prediction is performed using residual

Neural Network (NN). A framework of diﬀerent models is used for prediction

in [5]. This framework consists of Gated Recurrent Unit (GRU), Recurrent Neu-

ral Network (RNN) and LSTM. GRU is a variant of the RNN. It outperformed

LSTM and many other statistical models in terms of prediction accuracy. Based

on the simulation results, it is proved that deep learning models are better than

the statistical models. The authors in [6] used Restricted Boltzmann machine

and Rectiﬁed Linear Unit (ReLU) for training the data and electricity load

prediction respectively. Discrete wavelet transform and inconsistency rate meth-

ods are used for optimal feature selection from the features set in [7]. It helps

in dimensionality reduction. The parameter tuning of Support Vector Machine

(SVM) is performed using Sperm whale algorithm. Authors in [8]proposeda

model for STLF. In this work, for feature selection, Mutual Information (MI) is

used whereas, better prediction results are achieved by modifying the Artiﬁcial

Neural Network (ANN). Extreme Learning Machine (ELM) is used in [9]for

eﬃcient load prediction.

The authors in [10] proposed an intelligent approach to forecast electricity

load using MI and ANN. In [11] the authors forecast electricity load and price

using enhanced CNN and enhanced SVM. ELM is a generalized single hidden

layer FFNN learning algorithm. It is proved to be eﬀective for regression and

classiﬁcation purposes. Authors used the Neural Network (NN) for achieving

better load prediction in [12]. Input weights and biases are randomly assigned

in ELM learning process, whereas output weights are calculated using Moore-

Penrose generalized inverse technique. Eﬃcient load forecasting is performed

using sparse Bayesian ELM in [13]. This method is used for multi-classiﬁcation

purpose. Authors in [14] used Particle Swarm Optimization (PSO) and discrete

PSO techniques for predicting day ahead electricity load. The Grey wolf opti-

mizer is used in combination with NN to optimize weights and biases in [15].

The simulation results proved that weights and biases optimization technique

improved the prediction accuracy. Back propagation technique is used for train-

ing ELM in [16]. Training is done using context neurons as input to hidden and

input layer of ELM. Authors in [17] enhanced RNN to achieve better prediction

accuracy by adjusting weights using previous iteration errors. Moreover, adjust-

ing biases and neurons selection also aﬀects the prediction accuracy. The authors

in [18] developed a knowledge based system for STLF. In [19], 24 h ahead predic-

tion of cooling load of building is performed using deep learning. The simulation

Electricity Load Forecasting for Each Day of Week Using Deep CNN 1109

results showed that deep learning techniques improved the cooling load predic-

tion. Similarly in [20], RNN is used for electricity load prediction. In this work,

RNN groups the consumers into pool of inputs. It then handles the consumers

data as aggregated data, which reduces the computational time and resource

utilization.

1.1 Motivation and Problem Statement

The authors in [2] and [3] forecast electricity load using shallow NN by analysing

the historic load data. In [4], the authors used weather data to forecast the

electricity load demand using RNN. ELM is used to forecast the electricity load

demandin[16]. In [19], the cooling load of a residential building is forecasted

using a deep NN. The aforementioned work motivates us to forecast electricity

load using deep NN.

MLP is used in [2] and [3] for short term load forecasting. Similarity a single

layerFFNNisusedin[4] to forecast load. In [16], a back propagation technique

is used. Similarly, the authors in [19] used a deep RNN to forecast the load

demand of residential buildings. None of the above mentioned work used Deep

CNN (DCNN) to forecast electricity load demand for all days of the week.

1.2 Contribution

The main contribution of this paper are following:

•An DCNN is proposed to forecast the electricity load of Victoria.

•A performance comparison of the benchmark schemes with the proposed

scheme is performed.

The rest of this paper is organized as follows. Section 2is the proposed sys-

tem model. The proposed and existing schemes used to forecast electricity load

are discussed in Sect. 3. A detailed discussion on the simulation results of each

scheme is presented in Sect. 4. Section 5concludes the paper along with the

future directions of our work.

2SystemModel

Figure 3presents a complete picture of our proposed system model. Forecasting

electricity load belongs to the paradigm of Time Series (TS) forecasting. Accu-

rate load forecasting using TS data is not an easy task. Therefore, in this paper,

we have ﬁrst converted the TS electricity load data into a supervised learning

problem. Afterwards, training and testing is performed to forecast the electricity

load using the proposed model.

To convert the TS electricity load data into a supervised learning problem, we

have ﬁrst extracted all days of the week separately from the historical load data.

All Mondays are extracted to form a Monday group. Similarly, all Tuesdays are

extracted to form a Tuesday group etc. This processes is repeated for all days.

1110 S. Khan et al.

After the formation of seven groups, we further divide each group into a train

and test dataset sub-group. Every train data sub-group contains load data of

eight days. Whereas, the test dataset sub-group contains the load data of the

ninth day. After the formation of train and test data sub-groups, these groups

are appended one after the other in such a way that all Mondays are followed by

all Tuesdays, Similarly all Tuesdays are followed by Wednesdays etc. Once all

the days are grouped into one set of train and test dataset, we create the input

and output variables such that output to Xiis Xi+1 and output of Xnis X1.

The process of input and output variable formation is shown in Fig.1.

Monday

Tuesday t1 t2 ... n-1 n t1

Sunday t1 t2 ... n-1 n t1

t1 t2 ... n-1 n t1 Input variab les

Output

variables

Fig. 1. Input and output variables formation

2.1 Dataset Description

In this paper, the dataset released by the Australian energy market operator is

used to forecast the electricity load of Victoria. The dataset is publicly available

at [1]. In this paper, the dataset used for forecasting contains load data from the

last week of April 2018 till the second week of July 2018. The load values are

recorded after every thirty minutes in a day and there are forty eight observations

in one day. Figure 2shows the electricity load data used in this paper.

0500 1000 1500 2000 2500 3000

3500

5500

7500

Time (30 min interval)

Electricity load (MW)

Fig. 2. Electricity load of Victroia

Electricity Load Forecasting for Each Day of Week Using Deep CNN 1111

2.2 Performance Metrics

In forecasting the electricity load, researchers have used a number of performance

metric parameters. However, none of these performance evaluation metrics have

been marked as the standard parameters for evaluating the performance of a fore-

casting model. In order to evaluate, the eﬀectiveness of our proposed scheme, we

used three performance metric parameters. These parameters are Mean Abso-

lute Percentage Error (MAPE), Mean Absolute Error (MAE) and Root Mean

Square Error (RMSE). Equations 1,2and 3presents the mathematical formula

of MAPE, MAE and RMSE respectively. Furthermore, to assess the performance

of each forecasting model, the lower the value of each performance metric metric

indicator, the higher the forecasting accuracy will be.

MAP E =(1

N

N

t=1

|Xi−Yi

Yi

|)×100 (1)

MAE =1

N

N

t=1

|Xi−Yi|2(2)

RM SE =

1

N

N

t=1

|Xi−Yi|2(3)

Here Xipresents the actual load value, whereas, Yipresents the load forecast

values. To evaluate the forecasting results of a model, the lower the performance

metric parameter values, the better the results would be.

3 Existing and Proposed Forecasting Schemes

In this section we discussed the proposed and existing benchmark schemes used

to forecast the electricity load.

3.1 RNN

Among all the NN models, LSTM [21] is the most common type of RNN due

to its ability of memorizing information for a longer period of time. In LSTM,

every node act as a memory cell to store information. Furthermore, each node

has its own activation function. In order to process the input sequence, RNN

use the internal states. LSTM also have internal states, however, the diﬀerence

between other NN and LSTM is that in LSTM the old state becomes the input

to the next state along with the current input.

1112 S. Khan et al.

Start

Pre processing

Days extraction and dataset formation

Pre processing

Deep CNN structure

Selection of loss function

Selection of an activation function

Add hidden layers

Training and testing

Electricity load forecasting

Performance metrics

MAPE

MAE

RMSE

End

Parameter

tuni ng

Yes/No

Fig. 3. Proposed system model

3.2 CNN

The most successful type of neural network to solve various problems in medical

imaging and computer vision is CNN [25]. It is due to the fact that CNN can

handle large amount of input sequences. Furthermore, it has the ability to extract

the hidden features in the input data. The architecture of CNN comprises of

convolution, fully connected, dense, ﬂatten and max pooling layers. Commonly,

the deeper the CNN model, the higher the accuracy will be. In this paper, CNN

has one convolution, one dropout and two dense layers.

Electricity Load Forecasting for Each Day of Week Using Deep CNN 1113

3.3 ARIMA

It is one of the most commonly used statistical technique for TS forecasting.

ARIMA as a TS forecasting model is proposed by Box and Jenkins [22]asa

combination of Auto Regression (AR) and Moving Average (MA). The key idea

to combine AR and MA is to approximate the real world representation in a

more accurate and precise way. The main advantage of using ARIMA to forecast

electricity load is that it can accurately forecast the future load values if the TS

is a stationary series. However, if the TS is not stationary, then the input series

is converted to a stationary TS by taking the ﬁrst or second diﬀerence.

3.4 ELM

Huang et al. in [23] and [24] proposed ELM, which a type of feed forward neu-

ral network. Due to its generalization and learning abilities, ELM as forecast-

ing model has gained popularity in several ﬁelds. The core advantage of ELM

includes: (i) parameters of the hidden layers do not require ﬁne tunning: (ii)

biasness and the input weights to ELM are randomly assigned.

3.5 DCNN

We enhance the forecasting performance of CNN by adding extra layers. Increas-

ing the number of layers in CNN scale up the size and complexity of the proposed

DCNN. However, in this work, our focus is to minimize the diﬀerence between

the actual and forecast values. The proposed DCNN comprises of convolution

and dense layers. A dropout layer at 20% along with ReLU activation function.

4 Simulation Results and Discussion

This section discusses the forecasting results. To aﬃrm the usefulness of our

scheme, the proposed DCNN is compared with four benchmark schemes. To

better understand the forecasting behavior of each model, this section is further

divided in to ﬁve subsections.

4.1 Forecasting Results of RNN

The electricity load forecasted by RNN for Victoria is shown in Fig. 7. In this

ﬁgure, it can be clearly seen that there is a huge diﬀerence between the actual

and forecasted electricity load using RNN. From the performance metric graphs

showninFigs.4,5and 6it is concluded that RNN has the worst performance

in forecasting electricity load as compared to all the benchmark schemes used in

this paper. According to the performance metric values of RNN, Thursday has

the lowest error values. The minimum MAPE, MAE and RMSE values using

RNN to forecast electricity load of Victoria are 13.439%, 642.835 and 620.471

respectively. Furthermore, Monday has the worst performance metric values as

compare to all days of the week. To further study the usefulness of RNN in

forecasting electricity load of Victoria, the performance metric values for all

days of the week are displayed in Table 1.

1114 S. Khan et al.

ELM RNN DCNN CNN ARIMA

0

2

4

6

8

10

12

14

MAPE Scores

Mon

Tue

Wed

Thu

Fri

Sat

Sun

Fig. 4. MAPE of the ﬁve schemes using Victoria dataset

4.2 Forecasting Results of CNN

Figure 7shows the electricity load forecast of Victoria for all days using CNN.

As shown in this ﬁgure, the performance of CNN is slightly better as compare to

RNN. However, the diﬀerence in the forecasted and actual electricity is still very

large. The bar plots in Figs. 4,5and 6presents the performance errors metrics

of CNN. According to Table 1, CNN has the lowest error values in forecasting

the load of Thursday and Saturday. The MAPE, MAE and RMSE values for

CNN are 7.20%, 385.869 and 359.788 respectively. Whereas, in forecasting the

electricity load of Monday and Friday CNN has the worst performance metric

values.

ELM RNN DCNN CNN ARIMA

0

100

200

300

400

500

600

700

MAE Scores

Mon

Tue

Wed

Thu

Fri

Sat

Sun

Fig. 5. MAE of the ﬁve schemes using Victoria dataset

Electricity Load Forecasting for Each Day of Week Using Deep CNN 1115

4.3 Forecasting Results of ARIMA

According to Fig. 7, ARIMA achieved average results in forecasting electricity

load for Victoria as compare to CNN and RNN. Figure 7shows that at some time

slots the load forecasted by ARIMA is exactly the same as actual load. However,

at some time slots the load forecasted by ARIMA is totally unacceptable e.g., the

ﬁrst time slot of Friday and Tuesday. According to the performance metric graphs

in Figs. 4,5and 6the performance of ARIMA in forecasting load for Victoria

is better on Wednesday as compare to other days of the week. The minimum

MAPE, MAE and RMSE achieved in forecasting electricity load by ARIMA for

Victoria is 3.58%, 238.154 and 194.830 respectively. Similarly, ARIMA has the

worst performance metric values in forecasting load of Tuesday.

ELM RNN DCNN CNN ARIMA

0

100

200

300

400

500

600

700

RMSE Sco res

Mon

Tue

Wed

Thu

Fri

Sat

Sun

Fig. 6. RMSE of the ﬁve schemes using Victoria dataset

4.4 Forecasting Results of ELM

Table 1presents the performance metric values of ELM for all days of the week.

AsshowninFig.7, ELM is the second best choice to forecast electricity load

for Victoria. From the performance metric graphs in Figs. 4,5and 6Wednesday

has the least error values. The MAPE, MAE and RMSE values for Wednesday

are 2.170%, 342.677 and 259.769 respectively. Similarly, ELM has the worst

performance metric values in forecasting electricity load of Monday.

4.5 Forecasting Results of DCNN

The performance metric graphs in Figs. 4,5and 6shows that the proposed

DCNN has the lowest performance metric values for all days of the week as

compare to all the benchmark schemes. According to Fig. 7, the load forecasted

by the proposed DCNN for Victoria almost overlaps the actual values for all

days of the week. Table 1shows the performance metric values for all days of the

week using Deep CNN. From this table, it can be clearly seen that the minimum

1116 S. Khan et al.

3500

5500

7500

Monda y

3500

5500

7500

Tues day

3500

5500

7500

Wednesday

3500

5500

7500

Thu rsd ay

3500

5500

7500

Fri d ay

3500

5500

7500

Sat u r da y

0 4 8 12 16 20 24

3500

5500

7500

Sun d ay

Hour (s)

Ele c t ri c it y l oa d ( MW )

DCNN

ARIMA

CNN

RNN

ELM

Act ual

Fig. 7. Forecasting results of all schemes for Victoria

Electricity Load Forecasting for Each Day of Week Using Deep CNN 1117

Table 1. Performance metric values

ELM RNN DCNN CNN ARIMA

MAPE

Monday 4.303 13.969 2.893 7.515 4.629

Tuesday 3.577 13.708 2.805 7.263 5.047

Wednesday 2.170 13.781 2.152 7.379 3.583

Thursday 2.606 13.439 2.489 7.370 3.984

Frid a y 3.846 13.937 2.705 7.609 4.633

Saturday 4.270 13.637 2.924 7.200 4.538

Sunday 4.017 13.787 2.927 7.370 4.578

MAE

Monday 342.677 767.183 218.879 469.614 347.052

Tuesday 271.484 732.568 201.115 441.339 352.144

Wednesday 152.037 683.748 138.771 401.976 238.154

Thursday 194.463 642.835 184.972 385.869 265.644

Frid a y 288.611 738.617 192.779 449.356 343.053

Saturday 336.182 752.452 218.885 455.808 351.269

Sunday 306.203 745.615 210.575 453.192 332.788

RMSE

Monday 259.769 732.158 171.732 415.240 273.986

Tuesday 209.630 701.686 162.286 392.409 293.096

Wednesday 117.390 666.721 116.417 376.971 194.830

Thursday 139.007 620.471 132.354 359.788 212.860

Frid a y 220.306 705.959 148.614 404.669 256.654

Saturday 256.228 716.093 172.620 399.010 266.153

Sunday 235.286 711.188 167.861 400.558 261.256

MAPE, MAE and RMSE in forecasting electricity load for Victoria are 2.15%,

138.771 and 116.417 respectively. Moreover, the lowest error values in forecasting

load using DCNN are achieved for Thursday. Similarly, the highest MAPE, MAE

and RMSE are 2.927%, 218.885 and 172.620. To sum up, by comparing the

usefulness of our proposed DCNN with the benchmark schemes, it can be said

that the proposed scheme has the best result in forecasting the electricity load

for all days.

5 Conclusion and Future Work

For utilities and generation companies, forecasting the actual demand of electric-

ity minimizes the power risk and production cost of electricity. NN are gaining

popularity in forecasting the electricity load demand because of their ability

1118 S. Khan et al.

to handle non linear and non stationary structure of the electricity load data.

In this paper, a DCNN is proposed to forecast the electricity load demand for

each day of week. Our experimental results illustrate that the proposed scheme

has the lowest MAPE, MAE and RMSE scores as compare to the benchmark

schemes. The proposed DCNN exceeds the existing benchmark schemes by scor-

ing the lowest MAPE, MAE and RMSE of 2.152%, 138.771 116.417. ELM has

the second best performance metric scores of 2.17, 152.037 and 117.390. ARIMA

achieved average results whereas, CNN and RNN has the worst performance

metric values. In future, we plan to investigate the impact of feature engineering

in forecasting the electricity load and price. Furthermore, to validate the eﬀec-

tiveness of our proposed scheme, we plan to forecast the electricity load of other

states of Australia such as New South Wales and Queensland etc.

References

1. https://www.aemo.com.au/. Accessed 01 Jan 2019

2. Singh, S., Yassine, A.: Big data mining of energy time series for behavioral analytics

and energy consumption forecasting. Energies 11(2), 452 (2018)

3. Ahmad, T., Chen, H.: Short and medium-term forecasting of cooling and heating

load demand in building environment with data-mining based approaches. Energy

Build. 166, 460–476 (2018)

4. Chen, K., Chen, K., Wang, Q., He, Z., Hu, J., He, J.: Short-term load forecasting

with deep residual networks. IEEE Trans. Smart Grid (2018)

5. Jesus, L., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: Deep

learning approaches and empirical comparison of traditional algorithms. Appl.

Energy 221, 386–405 (2018)

6. Ryu, S., Noh, J., Kim, H.: Deep neural network based demand side short term load

forecasting. Energies 10(1), 3 (2016)

7. Liu, J., Li, C.: The short-term power load forecasting based on sperm whale algo-

rithm and wavelet least square support vector machine with DWT-IR for feature

selection. Sustainability 9(7), 1188 (2017)

8. Ashfaq, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and fast

converging short-term load forecasting model for industrial applications in a smart

grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)

9. Huang, G.-B., Zhou, H., Ding, X., Zhang, R.: Extreme learning machine for regres-

sion and multiclass classiﬁcation. IEEE Trans. Syst. Man Cybern. Part B (Cybern.)

42(2), 513–529 (2012)

10. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.: Short-term load forecast-

ing in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019)

11. Zahid, M., Ahmed, F., Javaid, N., Abbasi, R.A., Kazmi, H.S.Z., Javaid, A., Bilal,

M., Akbar, M., Ilahi, M.: Electricity price and load forecasting using enhanced

convolutional neural network and enhanced support vector regression in smart

grids. Electronics 8(2), 122 (2019)

12. Fallah, S., Deo, R., Shojafar, M., Conti, M., Shamshirband, S.: Computational

intelligence approaches for energy load forecasting in smart energy management

grids: state of the art, future challenges, and research directions. Energies 11(3),

596 (2018)

13. Luo, J., Vong, C.-M., Wong, P.-K.: Sparse Bayesian extreme learning machine for

multi-classiﬁcation. IEEE Trans. Neural Netw. Learn. Syst. 25(4), 836–843 (2014)

Electricity Load Forecasting for Each Day of Week Using Deep CNN 1119

14. Yu, J., Wang, S., Xi, L.: Evolving artiﬁcial neural networks using an improved PSO

and DPSO. Neurocomputing 71(4–6), 1054–1060 (2008)

15. Zeng, Y.-R., Zeng, Y., Choi, B., Wang, L.: Multifactor-inﬂuenced energy consump-

tion forecasting using enhanced back-propagation neural network. Energy 127,

381–396 (2017)

16. Saremi, S., Mirjalili, S.Z., Mirjalili, S.M.: Evolutionary population dynamics and

grey wolf optimizer. Neural Comput. Appl. 26(5), 1257–1263 (2015)

17. Ertugrul, ¨

O.F.: Forecasting electricity load by a novel recurrent extreme learning

machines approach. Int. J. Electr. Power Energy Syst. 78, 429–435 (2016)

18. Khan, M., Javaid, N., Naseem, A., Ahmed, S., Riaz, M., Akbar, M., Ilahi, M.:

Game theoretical demand response management and short-term load forecasting

by knowledge based systems on the basis of priority index. Electronics 7(12), 431

(2018)

19. Fan, C., Xiao, F., Zhao, Y.: A short-term building cooling load prediction method

using deep learning algorithms. Appl. Energy 195, 222–233 (2017)

20. Shi, H., Minghao, X., Li, R.: Deep learning for household load forecasting-a novel

pooling deep RNN. IEEE Trans. Smart Grid 9(5), 5271–5280 (2018)

21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),

1735–1780 (1997)

22. Box, G.E.P., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time Series Analysis:

Forecasting and Control. Wiley, Hoboken (2015)

23. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K.: Extreme learning machine: a new learning

scheme of feedforward neural networks. In: 2004 Proceedings of the IEEE Interna-

tional Joint Conference on Neural Networks, vol. 2, pp. 985–990. IEEE (2004)

24. Huang, G.-B., Chen, L., Siew, C.K.: Universal approximation using incremental

constructive feedforward networks with random hidden nodes. IEEE Trans. Neural

Netw. 17(4), 879–892 (2006)

25. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classiﬁcation with deep con-

volutional neural networks. In: Advances in Neural Information Processing Sys-

tems, pp. 1097–1105 (2012)