Post-Processing of ML-Based Weather Prediction for Solar Capacity Factor Forecasting

 1. Introduction

Reliable and computationally affordable forecasts of renewable energy production values are necessary for effective grid management and energy market integration and thus for a fast and sustainable transition of the power sector. State-of-the-art Machine Learning based weather prediction (MLWP) models are getting cheaper and better continuously nowadays, making them the perfect option to provide the needed weather forecasts. On the other hand, they lack the variables for solar power generation (solar capacity factor or irradiance or at least cloud cover). This study aims to answer the question of whether MLWP is suitable for deriving solar energy values from weather forecasts, provide a suitable post-processing pipeline and compare the results for different MLWP models.

2. Methodology

A comprehensive ML-based post-processing technique is developed to predict the solar capacity factor using weather data from forecasts or reanalysis datasets. In addition to basic calculation and data processing steps, the methodology consists of a Convolutional Neural Networks (CNN) trained on ERA5 and the “C3S operational energy dataset”. From ERA5 only the variables wind, humidity, pressure, and temperature were used in the training, making the model suitable for use with MLWP data. From the energy dataset, the solar capacity factor is used as ground truth.

With this architecture, weather forecasts of MLWP models (currently Pangu-Weather, soon NeuralGCM and GraphCast) are used to predict the solar capacity factor for up to 10 days lead time.

3. Current (and Upcoming) Results

Compared to a simple persistence baseline, the CNN applied to Pangu-Weather consistently yields a lower RMSE, with the error reduction ranging from approximately 51% for a one-day lead time to 11% for a lead time of 10 days. Similar results can be achieved by comparing the approach against a climatology baseline.

Future work will include comparing the performance across different MLWP model forecasts to identify the optimal models for energy sector predictions.

4. Conclusions This research demonstrates that sophisticated ML-based post-processing is essential for transforming raw weather model outputs into actionable, reliable capacity factor forecasts for the energy sector. The enhanced skill provides a vital tool for energy trading and grid operators to manage risk, optimize renewable energy resource deployment, and bolster grid resilience in the face of climate variability.