SPIRIT: Short-term Prediction of solar IRradIance for zero-shot Transfer learning using Foundation Models

TL;DR

SPIRIT tackles the challenge of forecasting solar irradiance at new installations lacking historical data by leveraging foundation-model vision embeddings, physics-informed features, and future covariates for zero-shot transfer. The approach combines a ViT-based nowcasting pipeline with an autoregressive Transformer forecast module, achieving substantial gains over a state-of-the-art baseline, including about $70\%$ improvement in zero-shot nowcasting and $45\%$ in cross-location forecasting as measured by $nMAP$. It supports rapid deployment and adaptive fine-tuning with limited data, maintaining robustness across diverse camera setups and locations, and demonstrates statistically significant improvements over prior methods. This work enables scalable, data-efficient solar forecasting crucial for grid reliability and accelerated renewable integration in emerging and remote markets.

Abstract

Traditional solar forecasting models are based on several years of site-specific historical irradiance data, often spanning five or more years, which are unavailable for newer photovoltaic farms. As renewable energy is highly intermittent, building accurate solar irradiance forecasting systems is essential for efficient grid management and enabling the ongoing proliferation of solar energy, which is crucial to achieve the United Nations' net zero goals. In this work, we propose SPIRIT, a novel approach leveraging foundation models for solar irradiance forecasting, making it applicable to newer solar installations. Our approach outperforms state-of-the-art models in zero-shot transfer learning by about 70%, enabling effective performance at new locations without relying on any historical data. Further improvements in performance are achieved through fine-tuning, as more location-specific data becomes available. These findings are supported by statistical significance, further validating our approach. SPIRIT represents a pivotal step towards rapid, scalable, and adaptable solar forecasting solutions, advancing the integration of renewable energy into global power systems.

Figures (4)

• Figure 1: Illustration of our system: A vision encoder (top-left) extracts embeddings from a sky camera image sampled from a diverse set spanning multiple locations and setups. Physics-inspired features are derived and integrated with auxiliary values, then merged with the image embedding (top-middle) into a unified representation. For nowcasting (right), a regressor predicts Global Horizontal Irradiance from this feature vector. For forecasting (bottom), a time-series model processes past feature vectors to create a context embedding, which is concatenated with a future covariate vector constructed from known future values to form the final latent representation. A regressor then maps this representation to future GHI values (bottom-right).
• Figure 2: Mean nMAP performance of SPIRIT and wacv2022 across varying fine-tuning data sizes in weeks. Solid lines denote average accuracy and shaded regions show 95% confidence intervals across multiple runs, including cross-setup (TSI to ASI and vice versa) and cross-location (TSI or ASI to SKIPP’D) transfer scenarios.
• Figure 3: Forecasting performance of SPIRIT and wacv2022 using nMAP error across different forecast intervals. Subfigures (a), (b), (c), and (d) correspond to 1-hour, 2-hour, 3-hour, and 4-hour forecasting, respectively. The solid lines represent the average performance, with varying fine-tuning data sizes (in weeks). The shaded regions denote the 95% confidence interval, illustrating the variability across multiple runs, including training on one dataset and finetuning and testing on another, as well as utilizing randomized seasonal sampling. SPIRIT exhibits consistently better performance and low variance compared to the baseline, particularly with severely limited data, demonstrating its ability to maintain stability over the baseline's.
• Figure 4: Examples of sunny, partly cloudy, and overcast conditions, captured by different sky cameras, are shown from left to right, across the three datasets: TSI, ASI, and SKIPP'D, displayed from top to bottom.