Monthly electric demand forecasting with neural filters

Neural networks have proved to be a very efficient tool for time series forecasting. Furthermore, the structure of the neural model known as Multilayer Perceptron is well suited to behave as a digital filter. These two neural properties have been used to forecast the monthly electric demand. The corresponding time series has been split into two new series: one representing its trend and the other describing a fluctuation around that trend. Trend has been forecasted with a neural network, while fluctuation has been predicted by splitting its time series into six series associated to each of the six peak frequencies of the fluctuation spectrum, so that a filtering-forecasting process will be carried out by six neural networks to obtain six predictions. Then all the predictions have been added to obtain the monthly demand forecasting. It has been proved that a Multilayer Perceptron is able to perform both filtering and forecasting at once if properly trained.

► Trend and fluctuations of a time series are independently forecasted.
► The fluctuation series is filtered to split it into frequency bands.
► Each band is independently forecasted.
► An only neural network carries out both filtering and forecasting.
► Bands and trend predictions are added to obtain the series forecasting.