Short-term power load forecasting by non-fixed neural network model with fuzzy BP learning algorithm

The authors develop a non-fixed neural network model with fuzzy back-propagation (BP) learning algorithm for short-term power load forecasting of Taipower system. The nonfixed structure used in our model is achieved by changeable input variables in the input layer of the neural network. Next-day peak load forecasting and one-to-two-day-ahead hourly load forecasting are investigated in this study. Gray relational analysis is employed for proper selection of input variables to improve the learning efficiency of the network. Furthermore, the dynamic fuzzy learning rate and momentum are also adopted in BP learning algorithm to improve the network's ill-learning phenomenon. Hourly loads and relevant temperature data from 1992 to 1996 provided by Taipower Utility and the Central Weather Bureau are implemented. As a comparison, the same experiments are performed using traditional neural network with fixed size and constant learning rate and momentum.