Short-term wind speed prediction using Bayesian optimized LSTM network

Abstract

The wind is a highly complex phenomenon that depends on geographical and environmental conditions. Wind speed depends on many variables such as temperature, pressure, humidity, and other lower atmospheric conditions, and thus mathematical modeling is highly complex and requires high computational power and time for wind speed prediction. Over the years, data-driven models for multi-step ahead time-series predictions have been gaining attention and are still in the evolutionary stage. Improvements in prediction models help the wind generation systems to operate efficiently. The accumulation of errors in the multi-step prediction creates a challenge in formulating novel prediction models. A prediction model based on long short-term memory (LSTM) is proposed in this study for short-term wind prediction up to a prediction horizon of 3 hours ahead. Hyperparameters of the LSTM model are tuned by Bayesian optimization. The wind speed data of two different sites are considered for the evaluation of the proposed model. Further, Support vector Regression based on the Multiple Input Multiple Output (MIMO) strategy is used to compare the performance of the proposed model. Bayesian optimized Long Short Term Memory (BO-LSTM) model shows nearly 30% and 16 % improvement in the MSE and RMSE scores, respectively, over the SVR model.