Short Term Load Forecasting using Echo State Networks

In this paper a new algorithm is proposed for Short Term Load Forecasting (STLF) using Echo State Networks (ESN). Hourly load data along with only average temperature of each day and day type flag is fed to the ESN and nonlinear mapping is done using training methods. Despite conventional recurrent neural networks, ESN can be trained much easier and with great deal of accuracy. Simulation results show that this method successfully predicts load demands even using limited input data. Using several parallel ESN units with smaller reservoir sizes in which each ESN unit identifies the dynamics of a certain hour of the day throughout the training and testing process results in more efficient use of data. Using this method, there is no need to identify weak correlations between dynamics of certain hours by using bigger neural network.