Stochastic Heuristic Optimization of Machine Learning Estimators for Short-Term Wind Power Forecasting

Abstract

The continuous fluctuation of wind speed, wind direction and other climatic variables affects the power produced by wind turbines. Accurate short-term wind power prediction models are vital for the power industry to evaluate future energy extraction, increase wind energy penetration and develop cost-effective operations. This research examines short-term wind power forecasting and investigates the effect of sharp, smooth and slow temperature reduction functions on the Simulated Annealing (SA) optimization technique for several prominent prediction models. The regressors under investigation include a Support Vector Machine, a Multi-Layer Perceptron and a Long-Short Term Memory neural network. Their optimization is based on the SA, which is used to specify the hyperparameters of each model in order to enhance the prediction accuracy. The results for each model based on the data of the Greek island of Skyros denote the superiority of the slow temperature reduction function in terms of error metrics and observe that the optimized Multi-Layer Perceptron is the most suitable model for this forecasting task when slow temperature reduction is implemented.