Improving recurrent network load forecasting

In this article, we present a not fully connected recurrent network applied to the problem of load forecasting. Although many authors have pointed out that Recurrent Networks were able to modelize NARMAX process (Non linear Auto Regressive Moving Average with eXogeneous variables), we present a constructing scheme for the MA part. In addition we present a modification of the learning step which improves learning convergence and the accuracy of the forecast. At last, the use of a continuous learning scheme and a robust learning scheme, which appeared to be necessary when using a MA part, enables us to reach a good precision of the forecast, compared to the accuracy of the model in use at the utility at present.