Hybrid learning algorithm based neural networks for short-term load forecasting

This paper proposes a hybrid algorithm to improve the accuracy of short-term load forecasting (STLF). In the hybrid algorithm, first, support vector regression (SVR) is used to determine the initial structure of RBFNNs (SVR-RBFNNs); then, an annealing robust concept with time-varying learning algorithm (ARTVLA) is then applied to train the SVR-RBFNNs (ARTVLA-SVR-RBFNNs). In ARTVLA, we adopt a particle swarm optimization (PSO) method to find a set of promising rates to overcome the problem for the trade-off between stability and speed of convergence in training procedure of RBFNNs. Finally, the optimal RBFNNs are applied to predict short-term load demands. The performance of the proposed approach is evaluated on the hourly empirical load data of the Taiwan power Company (TPC) in the case for 24-hour-ahead prediction. Simulation results show that the proposed ARTVLA-SVRRBFNNs yield more accurate load forecasting than the SVRRBFNNs based on annealing robust learning algorithm (ARLA-SVR-RBFNNs) with fixed learning rates.