Hybrid Post-Processing for Solar Power: Bridging Nowcasting to Short-Range

Accurate forecasting of solar power generation is crucial for grid operators, as location-dependent photovoltaic (PV) installations exhibit diverse production patterns. The need for high temporal and spatial resolution, combined with the inherent variability of PV outputs, presents significant challenges for forecasting and post-processing across different time horizons. This study addresses these challenges in post-processing optimal point forecasts for PV sites across multiple forecasting ranges, with the aim of providing seamless output for end-users in the energy sector. Specifically, we focus on two-day-ahead PV site forecasts, with an emphasis on a highly resolved nowcasting range (from minutes to hours ahead) and a smooth transition to short-range forecasts. Advanced machine learning techniques, gridded meteorological models, and a variety of location-specific data sources are employed to enhance our post-processing approach for optimal site forecasts.

Focusing on an Austrian case study, we develop a post-processing framework based on machine learning approaches for time-series forecasting, with particular emphasis on Long Short-Term Memory (LSTM) models compared to more classical methods such as Random Forest (RF) and Multiple Linear Regression (MLR). Our primary objective is to smoothly post-process and identify transitions among a set of range-specific, mostly gridded background models spanning various spatial and temporal resolutions. The post-processed models used as input primarily represent irradiance and related parameters. Our work integrates IrradPhyD-Net, a high-resolution AI-based nowcasting model, with AROME, a limited-area Numerical Weather Prediction (NWP) model for the alpine region, providing valuable physical information extending into the short- and medium-range. To exploit the location-specific characteristics of the site, we incorporate additional time-series models that capture the climatology and trends of PV, irradiance, and strongly correlated parameters identified during pre-processing. Given the substantial and growing input data needs of AI and machine learning, we build on our previous contributions by integrating semi-synthetic data to address challenges posed by limited or inconsistent historical PV data, thereby improving model stability. In this context, additional data sources, such as satellite-based CAMS radiation time-series and ERA-5 reanalysis, are essential.

By leveraging skillful input models, supported by synthetic data, our post-processing framework demonstrates strong forecast skill across the studied ranges. Thus, sourcing and transforming data from multiple inputs proves to be an effective way to achieve seamless, high-skill forecasts while maintaining high temporal resolution for nowcasting.