Optimized recurrent neural networks with attention for wind farm energy generation forecasting

The shift towards renewable sources of power has brought several associated challenges. The sporadic nature of power generation, high reliance on weather, and difficulties with long-term storage are just some of the challenges facing large-scale adoption. Storage is especially challenging due to the large associated upfront costs of batteries coupled with limited lifetimes further shortened by uneven charging. One possible step for improving the adoption of renewable sources is to develop robust techniques for forecasting power generation to better plan power demands. This work proposed an approach based on time series forecasting through the use of recurrent neural networks augmented with an attention mechanism. To further optimize the performance of this approach a modified version of a well-known metaheuristic algorithm is also introduced. The proposed approach is evaluated on a real-world dataset consisting of weather and wind farm power generation data from a wind farm located in mainland China. Several metaheuristics have been applied to optimize the proposed approach and the outcomes have been subjected to a comparative analysis to determine the best approach.