Short-term load demand forecasting in Smart Grids using support vector regression

In this study, we propose a method based on support vector regression (SVR) to model the nonlinear dynamics of customer load demand given a limited set of previous measurements. Such methodology is used for short-term load forecasting (STLF). SVR model is trained and tested using real-world data from both residential and business load profile types. An important issue in SVR model is addressed: determining a single set of kernel and model parameters suitable for the whole year, regardless of the load profile type. Main advantages of using the proposed methodology are that SVR makes no prior assumptions about the stationarity of the data, the computational complexity of the model does not depend on the dimensionality of the input space and the provided solution is global and unique. Prediction performances of the proposed method are analyzed and compared with those of different modeling approaches recently presented in the literature such as artificial neural networks and time series analysis techniques.