Decomposition Aided Bidirectional Long-Short-Term Memory Optimized by Hybrid Metaheuristic Applied for Wind Power Forecasting

Increasing global energy demands and environmental concerns have in recent times lead to a shift in energy production towards green and renewable sources. While renewable energy has many advantages, it also highlights certain challenges in storage and reliability. Since many renewable sources heavily rely on weather forecasting the amount of produced energy with a degree of accuracy becomes crucial. Energy production reliant on wind farms requires accurate forecasts in order to make the most of the generated electricity. Artificial intelligence (AI) has previously been used to make tackle many complex tasks. By formulating wind-farm energy production as a time-series forecasting task novel AI techniques may be applied to address this challenge. This work explores the potential of bidirectional long-short-term (BiLSTM) neural networks for wind power production time-series forecasting. Due to the many complexities affecting wind power production data, a signal decomposition technique, variational mode decomposition (VMD), is applied to help BiLSTM networks accommodate data. Furthermore, to optimize the performance of the network an improved version of the reptile search algorithm, which builds on the admirable capabilities of the original, is introduced to optimize hyperparameter selection. The introduced method has been compared to several state-of-the-art technique forecasting wind energy production on real-world data and has demonstrated great potential, outperforming competing approaches.