The Challenge of Intermittency: Probabilistic Spatiotemporal Forecasting of Photovoltaic Energy Production

A higher renewable energy utilization is essential to tackle climate change. Photovoltaic (PV) energy is an abundant renewable resource but its power grid integration is limited due to its intrinsic volatility, as surface solar radiation (SSR) highly depends on stochastic cloud dynamics. Accurate probabilistic SSR forecasts with lead times from minutes to hours help mitigate the effects of volatility and increase the PV share in the power grid without jeopardizing grid stability.

Most studies to date focus on forecasting PV power production for a few locations. By using satellite-derived SSR maps, we can more effectively capture cloud dynamics and produce forecasts encompassing broader regions. The satellite-based SSR is derived from two components, the clear-sky SSR and the clear-sky index (CSI), a dimensionless cloudiness index measuring the ratio between the actual radiation and the theoretical maximum SSR in a completely cloud-free situation.

Optical flow is the prevalent method for capturing cloud motion from satellite imagery. However, optical flow methods neglect the clouds' temporal evolution, growth and decay, which constitute a major source of SSR uncertainty. To address this research gap, we present the probabilistic SSR forecast model SolarSTEPS (Carpentieri et al., 2023). It exploits the scale dependency of the CSI’s temporal evolution to model the stochastic growth and decay of CSI with different autoregressive (AR) models at multiple spatial scales. A probabilistic forecast ensemble is then generated based on spatially correlated noise fields which are superimposed on the deterministic cloud motion forecasts.

SolarSTEPS shows better performance than state-of-the-art advection-based and analog-based models (Carriere et al., 2021; Ayet et al., 2018) as measured by the continuous ranked probability score (CRPS) at forecast lead times of up to 4 hours. In fact, our model shows an average CRPS that is more than 10% lower than the benchmark models.

Finally, we compare SolarSTEPS with a new probabilistic deep learning model. Our analysis focuses on different approaches of modeling cloudiness dynamics as well as the benefits and drawbacks of each approach.