A short-term temperature forecaster based on a state space neural network

Abstract A short-term hourly environmental temperature forecaster for using in building electric load forecasting purposes has been designed based on a state space Neural Network ( ssNN ) . The forecaster uses the current coded hour and the temperature as inputs, and predicts the next hour temperature. The training is based on a Random Optimization Method. Because of the non-stationary characteristic of temperature, training is executed daily in order to update the network weights. The dynamic of the outside temperature was satisfactorily captured by the ssNN when real data were used during several experiments. The encouraging results allow to use this predictor as a very good tool in Load Forecasting Systems.