The Long Short-Term Memory Tuning for Multi-step Ahead Wind Energy Forecasting Using Enhanced Sine Cosine Algorithm and Variation Mode Decomposition

Abstract

Wind is a renewable power source that is created by the uneven heating in the atmosphere and the force of Coriolis acceleration. It is a sustainable way to produce energy from renewable sources. However, there are several challenges to generating energy from wind power plants. This study looked at using artificial intelligence algorithms to predict short-term wind power generation. A goal was to create a robust system that could accurately predict wind power values using deep learning (DL) algorithms. The conducted research explores the potential of long short-term memory (LSTM) artificial neural networks for times-series wind power generation forecasting. However, like many ML algorithms, LSTM network performance largely depends on a set of control parameters known as hyperparameters. Adequate selections are crucial to ensuring good performance. The process of selecting optimal hyperparameters may be framed as an optimization, and can therefore be handled as an optimization problem. Additionally, to identify patterns in the wind energy signal, variation mode decomposition (VMD) was applied before being forwarded as input to LSTM. A notable set of algorithms that excel at handling optimization are metaheuristics. This work, therefore, explored the potential of the sine cosine algorithm for hyperparameter tuning for LSTM networks. Furthermore, an improved version of the SCA is introduced to help further enhances the admirable ability of the original. The introduced model has been assessed on real-world wind farm data and attained favorable results, outperforming contemporary optimization algorithms tested in identical conditions.