In this section, the applied neural network for power load forecasting, either ANN alone or with other pre-processing techniques, is discussed. In 2009, Xiao et al. used a back propagation neural network with rough sets for power demand forecasting. The system was compared with standard BP and in general the performance of BP with rough sets was better than standard BP [
5]. Din and Marnerides applied the Feed Forward Neural Network (FFN) and Recurrent Neural Network (RNN) with deep learning for short period power load forecasting, using a dataset collected from New England for the period from 2007 to 2012. The model was tested for two cases: the first case time domain features were used, while in the second case both features from time and frequency domains were used. The evaluated system using MAPE, RMSE and MAE errors, which rendered lower rates in the second case than in the first one, and the accuracy of the model were improved in the second case [
6]. An applied ANN with wavelet decomposition was designed by Reddy and Jung; the experiment results showed the efficiency performance of the proposed system, which exceeded ANN [
7]. Electrical load forecasting using advanced wavelets with neural networks was proposed by Rana and Koprinska. The proposed system consists of four steps: load data decomposed into high and low frequencies using wavelet transform, feature chosen based mutual information, training NN for each component and testing the trained model. The model was evaluated for two data sets from Australia and Spain. The mean absolute percentage errors were 0.268% and 1.716% for the Australian and Spanish data sets respectively. In addition, the articles conclude that the system out-performed the other existing models [
8]. Mordjaoui et al. proposed using a dynamic Neural network to forecast the electricity load. The proposed system was designed and tested using a dataset of the French Transmission System Operator. The simulation results proved the validation of the designed method [
9]. The idea for using loads of identical days as the input variable of the combination from wavelet transform with a neural network to predict future values of the load was proposed by Chen et al. [
10]. Zheng et al. designed an intelligent model for demand power forecasting, k-means for cluster data wavelet transform to decompose the data and finally NN to forecast the final value of the power load [
11]. Niu et al. used a Hybrid Monte Carlo technique for training a Bayesian neural network (BNN) for the purpose of designing a power load forecasting model. The designed system was compared with BNN trained using a La-place algorithm and ANN trained using a Backpropagation technique using MAPE and RMSE criteria. The experiment’s result proved the validity of the designed method for load forecasting [
12]. Combining the K-means clustering with ANN for load forecasting was made by Jahan et al. The experiments used k-means and k-medoids for clustering the original data into groups then measuring the distance between each sample and each cluster as new features which were fed into ANN. The results of ANN proved better than a decision tree when comparing the results using the MAPE criterion [
13]. Khwaja et al. proposed using a bagged neural network (BNN) for load forecasting. The idea of BNN is dividing the data set into random parts, then training the neural network for each part, the average of the outputs representing the output of the model. The outcome of the proposed idea reduced the forecasting error when compared with standard ANN and other existing approaches [
14]. Wang et al. proposed BPNN for power load forecasting. The idea was to optimize the network weights using a genetic algorithm faster than standard BPNN. The results showed that the proposed optimization algorithm improved the learning speed and the accuracy of the learning process [
15]. Ekici applied an extreme learning machine (ELM), regularized ELM (RELM) and ANN for electrical load forecasting, and comparing their performance. The outcome of the experiments confirmed that the RELM learned much faster than ANN and the forecasting accuracy of RELM was better than standard ELM [
16]. A hybrid model for short load forecasting was studied by Zhang et al. [
17]. The model has been constructed from improved empirical mode decomposition, an autoregressive integrated moving average (ARIMA) and wavelet neural network optimized by the fruit fly optimization algorithm. The MAPE of the model’s forecasting results was improved and is about 0.82% higher than other compared systems. In the USA, Ashfaq et al. designed a one day-ahead system for power load forecasting. The designed system constructed in three stages: pre-processing, in this stage removing unwanted samples, the forecasting stage using ANN, and the optimization stage for minimizing the forecasting errors. The forecasting accuracy of the system improved in comparison with other models [
18]. Yi Liang et al. introduced a hybrid system for electricity forecasting. The model is constructed as follows: empirical mode decomposition, minimal redundancy maximal relevance, neural network for regression with the fruit fly optimization algorithm. The simulation results of the model proved the validity of the system in STLF [
19]. In Spain, a short-term load forecasting model was designed using three stages: SOM maps used for pattern recognition, k-means for clustering the patterns, and ANN to predict the power load. The methodology has been trained and tested using a data set from the Iberdrola company. The system has a small error when compared with others [
20]. Short-term weekday power load forecasting was proposed in [
21]. The paper compared ANN with different learning algorithms. The best results were obtained when using a Generalized Neural Network with wavelet transform that was trained using an adaptive genetic algorithm and fuzzy system. In Canada, El-Hendawi and Wang designed a method for short-term demand power forecasting. The method combined the full wavelet packet transform with neural networks. The designed system decreases the forecasting error by 20% when compared with standard neural networks [
22].