Short Term Load Forecasting Using Gaussian Process Models

The electrical deregulated market increases the need for short-term load forecast algorithms in order to assist electrical utilities in activities such as planning, operating and controlling electric energy systems. Methodologies based on regression methods have been widely used with satisfactory results. However, this type of approach has some shortcomings. This paper proposes a short-term load forecast methodology applied to distribution systems, based on Gaussian Process models. This methodology establishes an interesting and valuable approach to short-term forecasting applied to the electrical sector. The results obtained are in accordance with the best values of expected errors for these types of methodologies. A careful study of the input variables (regressors) was made, from the point of view of contiguous values, in order to include the strictly necessary instances of endogenous variables. Regressors representing the trend of consumption, at homologous time intervals in the past, were also included in the input vector. The proposed approach was tested on real-load from three medium-sized supply electrical distribution substations located in the center of Portugal. To test the performance of the model in different load situations, the case study includes three different electrical distribution substations representative of typical load consuming patterns, namely the residential, the non-residential and the service sector.