The Short-term Load Forecasting Based On Neural Networks

The models for the short-term load forecasting are proposed by combining the artificial neural networks and electric load characteristics on JiLin and ShenYang Power Markets. Based on the models, the load forecasting software has programmed by Object Oriented method.Because the loads vary periodically and are sensitive with some factors, such as the near loads, temperatures, precipitations, the load data are divided by four seasons and three day-types, and neural networks models for weekday, weekend and holiday establish independently. In order to enhance the training speed and the precision of neural networks, the daily load curve is divided and proper input and output variables are chose in each model.The application of Scaled Conjugate Gradient Algorithm makes the neural networks converge more quickly and accurately. During forecasting, the models implement active selection of training data with approximate clime factors to the forecasting day, employing the K-nearest neighbors concept.The forecasting accuracy of neural networks models including climate factors and no those factors is compared by the load data from JiLin and ShenYang. The result presents that the neural networks models can consider relative factors of load easily, forecast accurately and can be used on-line.