A Hybrid Deep Learning Methodology for Wind Power Forecasting Based on Attention

Abstract

Wind energy, as a sustainable energy source, poses challenges in terms of storage. Therefore, careful planning is crucial to utilize it efficiently. Deep learning algorithms are gaining popularity for analyzing complex time series data. However, as the "no free lunch" theorem suggests, the trade-off is: they need a lot of data to achieve the benefits. This even brings up a severe challenge for time series analysis, as the availability of historical data is often limited. This study aims to address this issue by proposing a novel shallow deep learning approach for wind power forecasting. The proposed model utilizes a fusion of transformers, convolutional and recurrent neural networks to efficiently handle several time series simultaneously. The empirical evidence demonstrates that the suggested innovative method exhibits exceptional forecasting performance, as indicated by a coefficient of determination (R2) of 0.99. When the forecasting horizon reaches 48, the model's performance declines significantly. However, when dealing with long ranges, utilizing the mean as a metric rather than individual point estimates would yield superior results. Even when forecasting up to 96 hrs in advance, obtaining an R2 value of 0.50 is considered a noteworthy accomplishment in the context of average forecasting.