Short-term wind speed forecasting based on recurrent neural networks and Levy crystal structure algorithm

Abstract

Short-term wind speed forecasting accuracy is essential for power generating operations planning, energy market scheduling, wind power equipment maintenance and safety management. Nevertheless, predicting wind speed is highly challenging since wind speed is inherently non-stationary and nonlinear. This research addresses this issue through the development of an innovative paradigm for wind speed forecasting. Initially, the benchmark wind speed forecasting model is constructed by integrating the long short-term memory network, gate recurrent unit network, bi-directional long short-term memory network, and auto regressive integrated moving average model to capture sequential dependencies, seasonality and trends. Subsequently, for identifying the ideal parameters for both the neural network and statistical models, a grid search technique is applied. To further improve the accuracy of wind speed forecasting, Levy CryStal structure algorithm is employed to optimize the weights of the four models to ensure faster and more effectively reach optimal configuration. The optimization process starts with the GoodPoint set method for initial population optimization, adds the levy operator to avoid local optima, and the models expounded within this study exhibit a high degree of statistical confidence, registering at a 99% level of certainty when compared to established benchmark models by Diebold-Mariano test.