Neural Networks in Forecasting Electrical Energy Consumption

This paper presents an artificial neural network (ANN) approach to electric energy consumption (EEC) forecasting. In order provide the forecasted energy consumption, the ANN interpolates between the EEC and its determinants in a training data set. In this study, two ANN models are presented and implemented on real EEC data. The first model is a univariate model based on past consumption values. The second model is a multivariate model based on EEC time series and a weather dependent variable, namely, degree days (DD). Forecasting performance measures such as mean square errors (MSE), mean absolute deviations (MAD), mean percentage square errors (MPSE) mean absolute percentage errors (MAPE) are presented for both models. Moreover, autoregressive and ARIMA modeling was applied (Saab, Badr and Nasr 2001). In this paper, two models were built to forecast electrical energy consumption (EEC) in Lebanon using artificial neural networks (ANN). The first model is univariate and fully connected model based on past EEC values. The second model is a multivariate not fully connected model based on past degree days (DD) and EEC time series. The monthly DD data is calculated from the daily mean temperatures obtained from climatological monthly bulletins (Ghaddar, 1995-1999). DD is used to indicate the days requiring energy usage for comfortable indoor living. Introduction Electric energy consumption forecasting is a crucial component of any energy management system. Traditional forecasting models can be classified as time series or regression models. Various techniques for forecasting energy consumption have been proposed in the last decade. Specifically, Multivariate modeling along with cointegration techniques (Dincer and Dost 1997; Eltony and Hosque 1997; Eltony and AI-Mutairi 1995; Eltony 1996; Ranjan and Jain 1999; Nasr, Badr and Dibeh 2000) are used to study the impact of different determinants on energy demand in different countries. Also, univariate modeling such as the AutoRegressive Moving Average (ARMA) modeling technique has been successfully used for forecasting (Abdel-Aal and A1 Garni 1997). Neural and abductive network models have also been successfully used for energy forecasting (AI-Shehri 1999; AbdeI-Aal, AI-Garni and AI-Nassar 1997; Park et al. 1991). The determinants of electricity consumption have also been studied using econometric models (Nasr, Badr and Dibeh 2000). Copyright ©2001, American Association for Artificial Intelligence (www.aaai.org) All rights reserved. ANN Implementation The study period spans the time period from 1995 to 1999. This period is used to train, test and evaluate the ANN models. The training of the models is based on a three year training set, January 1995 to December 1997 while the testing stage covers the period from January 1998 to February 1999. The evaluation stage covers the period between January 1998 and December 1999. Since the purpose of this paper is to forecast future data, the backpropagation algorithm is used. This method is proven to be highly successful in training of multilayered neural nets. The consists training of the network by backpropagation of three stages: Thefeedforward of the input training pattern. The calculation and backpropagation of the associated error. The adjustment of the weights Before applying the BPN algorithm, two steps are required: NEURAL NETWORK I FUZZY 489 From: FLAIRS-01 Proceedings. Copyright © 2001, AAAI (www.aaai.org). All rights reserved. 1. Initialize Weights: the weights from the input layer to the hidden layer and the weights from the hidden layer to the output layer are randomly chosen to have values between -0.5 and 0.5. 2. Normalize data: The monthly data are normalized in order to have values between 0.1 and 0.9. The formula used is the following: DataMin x (Hi Lo ) + MaxMin Where: Min = Monthly minimum data. Max = Monthly maximum data. Hi = 0.9 = Maximum value for the normalized data. Lo = 0.1 = Minimum value for the normalized data. The ANN models are implemented in a C program using Microsoft Visual C++. This program is used to train, test and evaluate the net. The C program involves many steps: I. The weights are chosen randomly. 2. The learning rate, the momentum parameter and the slope parameter are initialized. 3. The minimum test error is initialized to the maximum real value. 4. The data are normalized. 5. The training data set is used more than once. 6. The net is tested using the testing data set and the test error is computed. 7. If the test error is less than the minimum test error, the weights are saved and the test error will be the minimum test error. 8. Otherwise, if the net is tested for more than 100 times, the weights are restored. 9. Otherwise, step 4 to 7 are repeated. 10. The net is evaluated. The mean square error (MSE), the mean absolute deviation (MAD), the mean percentage square error (MPSE), and the mean absolute percentage error (MAPE) are calculated. ANN Models Model I Since present and future electric energy demand depends on previous electric energy consumptions, a univariate ANN model is implemented. The ANN model I (EEC model) requires real monthly EEC data as input parameters and has the structure as shown in Figure 1. The network architecture selected consists of an input, a hidden layer and an output layer. This model is a fully 49O FLAIRS-2001 connected model since each input unit broadcasts its signal to each hidden unit. The parameters are selected following extensive testing by varying the values of the learning rate, the momentum parameter, the slope parameter and the number of input units. Parameter values yielding lowest error figures are given in Table 1. The error measures used in this study are the mean square error (MSE), the mean absolute deviation (MAD), the mean percentage square error (MPSE) the mean absolute percentage error (MAPE).