Emilly Pereira Alves 
João Fausto Lorenzato de Oliveira Francisco Madeiro & Manoel Henrique da Nóbrega Marinho
Abstract: Smart grids are an alternative to minimize environmental impacts, such as CO2 emissions, and improve the efficiency of electricity consumption in buildings. Power grids enable adequate management and monitoring of consumption because of the periodic storage of measurements and easy access to them. In this scenario, an accurate prediction is a challenging task. Forecasting of consumption series is a defiant problem because data present linear and nonlinear patterns, and a dependence on external variables may be observed. Hybrid models are an alternative to mapping both patterns, which have been widely used to forecast load time series. Autoregressive Integrated Moving Average (ARIMA) and Support Vector Regression (SVR) models are used for this purpose, to map the linear and nonlinear patterns of the series, respectively. In this paper, a nonlinear optimized hybrid system based on ARIMA, SVR, and Particle Swarm Optimization (PSO) is proposed. The system can be divided into three steps. First, the linear patterns are predicted by the statistical model ARIMA. Then, the residual series is modeled using an optimized SVR, in which the parameters are selected from the PSO. One particularity from the proposal is to incorporate the choice of the topology and the inertia coefficient into the system. Lastly, the predictions are combined using the SVR. The simulations were conducted using a real database from smart meters of a building in Taiwan. To evaluate the performance of the proposed method, four related approaches were implemented and compared: a single ARIMA, two linear combination systems, and one non-linear combination system. The results show a superiority of the proposed method in terms of the metrics Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).
Keywords: Load forecasting, time series, smart grid, hybrid models, SVR, ARIMA.
DOI code: 10.21528/lnlm-vol20-no1-art2
PDF file: vol20-no1-art2.pdf
BibTex file: vol20-no1-art2.bib