Conceptual approach to the application of neural network for short-term load forecasting

The feasibility of using a simple neural network for short-term load forecasting is investigated. A combined linear and nonlinear neural network is developed. The forecasts are computed using weights which are reestimated using only very recent observations. The model operation is tested by using load data obtained from a winter-peaking utility in the Northeastern USA. The results show that the error in most weeks is small, less than 4-5%. This validation test proves that the method is feasible and able to produce accurate forecasts under normal conditions.>