Leveraging Spatially Explicit Data for Accurate Renewable Energy Forecasting in France

To meet France’s CO2 emission reduction of 33 % by 2030 compared to 1990 and reach greenhouse gas neutrality in 2050, sustainable energy sources are key to clean power production and reduced emissions from the energy sector. However, non-dispatchable renewables such as wind and solar photovoltaic (PV) power require accurate forecasts to improve their grid stability, reliability, and penetration level not to mention supply-demand matching. Indeed, those sources are dependent on weather conditions such as solar radiation or wind speeds, making their load highly variable and challenging to balance for grid operators.
Despite the increase of data availability from both weather and energy fields, regional wind and PV supply forecasts are usually indirect. Either a bottom-up approach of plant-level forecasts or a time series prediction incorporating lagged values is used. The potential of spatially explicit data for direct prediction is still underestimated. In this work, we present a methodology for predicting solar and wind power production at the country scale in France using machine learning models trained with spatially resolved weather data combined with geospatial information about production sites’ capacity.

A dataset spanning from 2012 to 2023 is built, using daily power production data from the national grid operator as the target variable, with daily weather data from ERA5, the capacity and location of the production sites, and electricity prices as input features. Three modeling approaches are explored to handle spatially resolved weather data: spatial averaging over the country, dimension reduction through principal component analysis, and a convolutional neural network (CNN) architecture to exploit complex spatial patterns. We benchmarked state-of-the-art machine learning models such as tree-based architectures, additive models, and neural networks on daily power supply for the midterm horizon. Hyperparameter tuning procedures based on different cross-validation methods were also investigated to reach the lowest generalization error possible.

Despite the variance introduced by the model and the data, our cross-validation experiments showed that while using one-to-one models on the spatial average of weather data, the time-series dedicated procedures tend to estimate the generalization error better than standard methods like K-Fold. This allowed us to push the model calibration to reach the best performance on unseen test data. However, they fall short of the CNN ingesting entire weather maps which predicts twice as good. Indeed, CNN is the best model for both PV and wind, achieving errors of around 5 %. This is mainly due to its ability to exploit spatial weather patterns on production site locations to extrapolate the trend in renewable power supply as underlined by an interpretability method. In fact, one-to-one models utilized on both spatial average and principal components extracted from weather maps are struggling to grasp the increase in power supply due to the growth in installed capacity.

Our study highlighted the potential of spatially explicit data and dedicated models to improve the accuracy of direct regional renewable power supply. Such enhancements will lead to a better supply-demand balance while incorporating a growing part of sustainable energy into our electricity mix.