Spatio-temporal characteristics based wind speed predictions

Integration of large-scale wind power plants to the power system is a challenge as the power generation is variable, and energy management systems require accurate prediction of wind power for a stable operation. Frequency control, economic dispatch and unit commitment problems in power system operations depend on forecasted wind power. Due to the dynamic changes in wind patterns, wind speed (and power) is very difficult to predict. In this paper, several computational approaches using neural networks (NN) for wind speed prediction is presented. Cellular Computational Networks (CCNs) are found to be more accurate than Multilayer Perceptrons (MLPs) and Recurrent Neural Networks (RNNs). This is due to capability of CCNs to simultaneously capture spatial-temporal characteristics of wind. The effectiveness of standard backpropagation, Backpropagation Through Time (BPTT) algorithm and Particle Swarm Optimization (PSO) are compared for training the computational networks. Performance of PSO algorithm is comparatively better than that of BPTT for training CCNs with MLPs.