The paper reports a new methodology for the medium term load forecasting providing monthly energy consumption and monthly maximum demand for a municipal utility. To this aim a modular procedure, based on an artificial neural network (ANN), which is a multi-layer perceptron using a back-propagation feed-forward algorithm, is implemented. The monthly forecasts are obtained through some knowledge based activities from the output of stage providing annual energy forecast. The choice of the prediction stage is reported by illustrating the results of a comparison with canonical statistical methods, such as exponential smoothing and ARIMA. The whole knowledge based procedure is illustrated in due detail and some best forecasting performances are reported thus demonstrating validity of the proposed approach