A meta-heuristic framework for forecasting household electricity consumption

It may be difficult to model household electricity consumption with conventional methods such as regression due to seasonal and monthly changes. This paper illustrates a flexible integrated meta-heuristic framework based on Artificial Neural Network (ANN) Multi Layer Perceptron (MLP), conventional regression and design of experiment (DOE) for forecasting household electricity consumption. Previous studies base their verification by the difference in error estimation, whereas this study uses various error estimation methods and design of experiment (DOE). Moreover, DOE is based on analysis of variance (ANOVA) and Duncan Multiple Range Test (DMRT). Furthermore, actual data is compared with ANN MLP and conventional regression model through ANOVA. If the null hypothesis is accepted, DMRT is used to select either ANN MLP or conventional regression. However, if the null hypothesis is accepted then the proposed framework selects either the MLP or regression model based on the average of Minimum Absolute Percentage Error (MAPE), Mean Square Error (MSE) and Mean Absolute Error (MAE). The significance of this study is the integration of ANN MLP, conventional regression and DOE for flexible modeling and improved processing, development and testing of household electricity consumption. Some of the previous studies assume that ANN MLP provide better estimation and others estimate electricity consumptions based on the conventional regression approach. However, this study presents a flexible integrated framework to locate the best model based on the actual data. Moreover, it would provide more reliable and precise forecasting for policy makers. To show the applicability and superiority of the integrated approach, annual household electricity consumption in Iran from 1974 to 2003 was collected for processing, training and testing purpose.