Recurrent neural network for forecasting next 10 years loads of nine Japanese utilities

In this paper, two artificial neural networks, a recurrent neural network (RNN) and a three-layer feed-forward back-propagation (BP), are applied for long-term load forecasting. The RNN is designed to forecast the loads of 1 yr ahead and the BP is used to forecast the next 5 and 10 years loads. The proposed networks are trained with the past 20 years (1975-1994) of actual data and are tested for target years 1995-1997, 2000, and 2005. In addition to the target year load forecasting, a sliding window training method is proposed for the continuous retraining of the RNN. The actual data is used for forecasting the loads of 1975-1994. However, forecasted data is applied for forecasting the loads beyond 1994. Since the weather condition data is not available for longer than two weeks ahead, a sensitivity program is developed to produce the future temperature from the present one. Very reasonable results have been obtained for the Seen (inner sample) and unseen (out-of-sample or loads of target years) data. In this study, total system load forecast reflecting current and future trends, tempered with good judgement which is the key to all planning and indeed financial success is carried out for nine utilities in Japan. (C) 1998 Elsevier Science B.V. All rights reserved.