Towards online operation of a RBF neural network model to forecast the Portuguese electricity consumption

In previous work the authors successfully identified a radial basis function neural network to forecast the Portuguese electricity consumption profile within a 48 hour predictive horizon. As the model is a static mapping employing external dynamics and the electricity consumption trends and dynamics are varying with time, its predictive performance degrades after a certain period. One of the simpler ways to counteract this effect is by retraining the model at certain time intervals. In this paper this methodology is investigated considering regular and irregular retraining periods. For the latter, a criterion is defined in order to trigger the retraining procedure. The results obtained are compared to a nearest-neighbour predictive approach that achieves acceptable predictive performance and operates on a sliding window of data, therefore providing some level of adaptation. Also an analysis is made in order to find the time of day where the prediction error is smaller. Globally the retraining technique provides satisfactory maintenance of predictive performance although exhibiting alternating levels.