Intelligent forecast engine for short-term wind speed prediction based on stacked long short-term memory

Abstract

This paper presents machine learning techniques for short-term wind speed prediction by exploiting their computational intelligence capabilities such as random forest regression (RFR), support vector regression (SVR), radial basis function neural networks (RBFNN), and long short-term memory (LSTM) on various wind farm datasets located in Pakistan. Initially, predictions are obtained by employing baseline regressors, RFR and RBFNN in terms of error indices. Later, a stacked LSTM forecast engine (SLFE) has been proposed to improve the efficacy, accuracy, and prediction capability of the forecast engine by using leaky rectified linear units (Leaky ReLU) as kernel function. SLFE has been implemented and tested on the datasets acquired from four different wind farms having a temporal locality of 10 min for short-term wind speed prediction. Furthermore, an ensemble of stacked LSTM with RFR, SVR, and RBFNN has also been developed for the comparison. The strength of the SLFE has been evaluated in terms of various performance measures such as mean absolute error (MAE), root mean squared error (RMSE), R2Score, and explained variance score (EVS). The efficacy of the proposed models is evaluated in terms of performance metrics, MAE, RMSE, R2Score and EVS, 0.043, 0.065, 0.813, and 0.814, respectively, demonstrating the worth of the forecast engine. Additionally, statistical one-way ANOVA is also carried out with multiple independent executions of the proposed algorithm to demonstrate the robustness, efficiency, and reliability of the model.