Electric load forecasting using a structured self-growing neural network model ´CombNET-II´

A neural network approach for electric load forecasting using CombNET-II has been investigated. The records on hourly electric load values from June 1986 to May 1990 (four years) as well as the corresponding maximum temperatures, average temperatures in a day and temperatures in every three hours at Nagoya were used. The networks have been trained to make up the mapping functions between these temperature trends and the electric load trends. The performance of the networks are evaluated by forecasting the records in the years from June 1989 to May 1990. The average errors for all days in a week were 3.18% to 3.01%. Considering that the network utilizes the weather parameters only, these results are quite acceptable. The performance of the load forecasting by CombNET-II is superior to that of the BP network, the average which was 4.72%.