Short-term load forecasting based on ANN applied to electrical distribution substations

The short-term load forecasting (STLF) algorithms belong to the set of methodologies which aim to furnish more effectiveness in planning, operation and conduction in electric energy systems (EES). The presence of a de-regulated environment reinforces the need of forecast, particularly in distribution networks. Actions like network management, load dispatch and network reconfiguration, under quality of service constraints, require reliable short-term (next hour) load forecasts. Artificial neural networks (ANN) are widely used in this horizon of prevision, with satisfactory results. The construction of an "efficient" ANN goes through, among, other factors, the construction of an "efficient" input vector, in order to avoid overfitting problems and keeping the global simplicity of the model. This paper deals with a methodological approach, in order to provide a more solid basis decision regarding the composition of the input vector, namely, in the choice of the number of the contiguous values of the principal variable (active power). In a first approach we established a search for any "chains with complete connections", in the active power signal, based on Gibbs measure, and a relative entropy analysis. We introduced the concept of "consumption tendency". We also analyzed the correlation between the consumption and the climatic data, having been established a nonweather sensitive model. The methodological approach is discussed and compared with another input vector. The model was tested in a real life case study for illustration of the defined steps.