Uncertainty Quantification and Analysis in Solar Forecasting Using Machine Learning

In the context of climate change, there is an urgent need to develop sustainable and reliable energy systems, particularly renewable energy.
For the planning and operation of such systems, accurate and trustworthy forecasting is essential. This study addresses prediction uncertainty in machine learning-based solar forecasting by quantifying it and systematically analysing its sources.

The uncertainty of a prediction arises from various sources. On the one hand, models may fail to represent the underlying relationship or may not be optimally parametrised. This type of uncertainty is referred to as epistemic uncertainty. On the other hand, uncertainties arise from the data itself, for example due to measurement errors and missing information, or from inherent variability in the physical system. This is also referred to as aleatoric uncertainty. In order to reduce uncertainty, its source must be identified. 

In this study, a solar forecasting model is employed and combined with uncertainty quantification (UQ) methods to assess the reliability of its predictions.
A network of measurement stations across Germany provides irradiance time series data, which are modelled within a graph framework. The model produces short-term forecasts of solar energy, using the data from all stations and potentially additional meteorological input.  It is based on Graph Neural Networks (GNNs) to incorporate both temporal and spatial information in the data.  

Different UQ frameworks are integrated with the forecasting model to assess predictive uncertainty and analyse its underlying sources. Particular emphasis is placed on uncertainty arising from the input data. Classical UQ approaches are compared with perspectives from trustworthy AI, notably adversarial machine learning. In this context, small perturbations of the input are identified that induce significant changes in the model output. This approach is explored as a tool to characterise model sensitivity, and to better understand the influence of input uncertainty on the forecast, thereby complementing established UQ methods.