Leveraging state-of-the-art AI models to forecast wind power generation using deep learning

In this paper, we present a novel approach for forecasting weather variables that are not currently available in many state-of-the-art AI models. A variable not found in most models is the 100-m wind speed, which is commonly used in the energy sector to predict the power generated by wind turbines. We trained a convolutional neural network model on 12 years of ERA5 data to instantaneously predict the 100-m wind speed based on a subset of variables found in the ECMWF-AIFS forecast. We evaluated our model with 2020 ERA5 data and achieved an average 100-m wind speed RMSE of 0.18 m/s, outperforming the wind profile power law method with an RMSE of 0.63 m/s. Using the AIFS output as input to our trained model, we generated 10-day 100-m wind speed forecasts without requiring autoregressive steps, significantly reducing computational costs. We compared our predictions with the ECMWF-IFS forecast using the ECMWF analysis as 'ground truth' and showed greater accuracy at longer lead times. Additionally, we produced power generation forecasts for onshore and offshore wind farms across the United Kingdom, with improvements over the IFS after a lead time of 3 days. We also showed that our model exhibits spatial and temporal coherence between local predictions and discussed the common limitation of over-smoothing in the predictions of AI models.