[Paper Review] Improving day-ahead Solar Irradiance Time Series Forecasting by Leveraging Spatio-Temporal Context

Oussama Boussif, Ghait Boukachab+5 more|arXiv (Cornell University)|Jun 1, 2023

Solar Radiation and Photovoltaics9 citations

TL;DR

Introduces CrossViViT, a multi-modal transformer architecture that fuses satellite spatio-temporal context with historical time series to forecast day-ahead Global Horizontal Irradiance (GHI) and estimates prediction uncertainty via multi-quantile heads, evaluated across multiple stations and out-of-distribution scenarios.

ABSTRACT

Solar power harbors immense potential in mitigating climate change by substantially reducing CO$_{2}$ emissions. Nonetheless, the inherent variability of solar irradiance poses a significant challenge for seamlessly integrating solar power into the electrical grid. While the majority of prior research has centered on employing purely time series-based methodologies for solar forecasting, only a limited number of studies have taken into account factors such as cloud cover or the surrounding physical context. In this paper, we put forth a deep learning architecture designed to harness spatio-temporal context using satellite data, to attain highly accurate extit{day-ahead} time-series forecasting for any given station, with a particular emphasis on forecasting Global Horizontal Irradiance (GHI). We also suggest a methodology to extract a distribution for each time step prediction, which can serve as a very valuable measure of uncertainty attached to the forecast. When evaluating models, we propose a testing scheme in which we separate particularly difficult examples from easy ones, in order to capture the model performances in crucial situations, which in the case of this study are the days suffering from varying cloudy conditions. Furthermore, we present a new multi-modal dataset gathering satellite imagery over a large zone and time series for solar irradiance and other related physical variables from multiple geographically diverse solar stations. Our approach exhibits robust performance in solar irradiance forecasting, including zero-shot generalization tests at unobserved solar stations, and holds great promise in promoting the effective integration of solar power into the grid.

Motivation & Objective

• Motivate accurate day-ahead GHI forecasting by incorporating cloud-driven spatio-temporal context beyond local time series.
• Develop a multi-modal deep learning architecture that fuses satellite imagery with historical solar data.
• Provide a mechanism to extract predictive distributions (uncertainty) for each time step.
• Create a multi-modal, long-span dataset of satellite imagery and solar variables across diverse stations.
• Propose a testing scheme that separates easy and hard cloud-influenced cases to evaluate performance under challenging conditions.

Proposed method

• Propose CrossViViT, a Cross Video Vision Transformer that jointly encodes past time series and satellite context using separate transformers and cross-attention to build a unified latent for future prediction.
• Use ROtary Positional Encoding (ROPE) to encode spatio-temporal coordinates and relative positions of context patches and time-series tokens.
• Mask portions of past context during training as regularization to encourage reliance on context.
• Decode via a Transformer-based decoder to produce future time-series predictions.
• Extend CrossViViT with a Multi-Quantile head to predict multiple quantiles, enabling prediction intervals (e.g., 0.02–0.98) for each time step.
• Assemble a multi-modal dataset combining 15 years of satellite imagery (RSS) with six solar stations and 24h history for 24h ahead GHI forecasts, including derived metrics like DHI, DNI, and clear-sky components.

Experimental results

Research questions

• RQ1 Can satellite-based spatio-temporal context improve day-ahead GHI forecasting at individual stations compared to time-series baselines?
• RQ2 How does CrossViViT perform in zero-shot forecasting at unseen stations and years?
• RQ3 Does the multi-quantile extension provide reliable prediction intervals for cloud-driven variability?
• RQ4 What is the impact of using rotary positional encoding versus learned embeddings on out-of-distribution performance?
• RQ5 How does a dedicated testing scheme that splits easy vs hard cloud scenarios reveal model strengths and weaknesses?

Key findings

• CrossViViT achieves robust day-ahead GHI forecasting performance and shows zero-shot generalization to unseen stations.
• The Multi-Quantile extension provides uncertainty estimates with reasonably high interval coverage.
• Abla tion studies indicate RoPE contributes to better performance on out-of-distribution stations and challenging (cloudy) days.
• A testing scheme that separates easy and hard cloud scenarios reveals model advantages on hard days where cloud movement strongly affects irradiance.
• A new multi-modal dataset combining satellite imagery with solar irradiance measurements over 14 years and six stations enables testing of spatio-temporal forecasting models in diverse conditions.