Modelo adaptativo para previsão de carga ativa de curto prazo

The present work proposes, validates and verifies a very short and short time bus active load forecast model that takes into account the regional differences and the dynamics of the load demand in a heterogeneous electric energy market such as the market of the State of Santa Catarina. This work also compares the performance of this model with the Oliveira (1999) model that has proposed for real time systems associated use. Furthermore, the proposed model was used to verify the following hypothesis. Hypothesis I: there is difference among the sets of input variables used in artificial neural networks to minimize the load forecasting error, in a specific time stamp, for regions with different load demand behavior. Hypothesis II: there exists difference among the sets of input variables of the artificial neural networks to minimize the load forecasting error in a specific node of an electrical system at different moments of time. The proposed model can adapt and adjust itself to the dynamic of a market of electrical energy composed of micro regions with heterogeneous characteristics of load demand. The model hybrid and built upon two artificial intelligence techniques: genetic algorithms and artificial neural networks. The artificial neural network uses the radial basis function that is relatively easy to adjust and that is relatively fast to be trained. Such properties are relevant for forecasting systems associated to SCADA. The native modeling capacity of the artificial neural network is complemented with the modeling capacity of the genetic algorithm that set the neural network input layer and some of its internal parameters. When it is compared to the model of Oliveira (1999), the proposed model shows 66% to 73% of precision improvement (reduction on the values of the errors). Hypothesis I and II are both proved. When a specific region of load demand is considered, the model shows 3% to 63% of precision improvement. When a specific season is considered and the model is specialized to the region of load demand, the model shows 7% to 29% of precision improvement. Finally, when the number of climatic variables was intentionally changed, relevant variations were observed, and the precision improvement was up to 25%.