Accurate 24-hour-ahead Load Forecasting Using Similar Hourly Loads

Most conventional methods require weather conditions for accurate load forecasting. This paper presents the results of 24-hour-ahead load forecasting without involving weather variables. A novel method based on hourly load deviation is proposed to search similar hourly loads as the input of neural network. Levenberg-Marquardt method is used to train a multilayered feed-forward neural network and genetic algorithm is applied to optimize the weights of the trained neural network. The study is performed on the actual electric demands of Ontario, Canada in the year 2005. The 24-hour-ahead forecasting results have high accuracy with the maximum MAPE (mean absolute percentage error) below 1.2% and the prediction errors less than 10%.