SolarCrossFormer: Improving day-ahead Solar Irradiance Forecasting by Integrating Satellite Imagery and Ground Sensors

TL;DR

This work tackles day-ahead solar irradiance forecasting by fusing satellite imagery with a dense ground-sensor network through a graph-based transformer, enabling probabilistic forecasts at 15-minute resolution for up to $H=96$ steps (24 hours) across multiple sites. The SolarCrossFormer architecture alternates temporal self-attention with cross-modal cross-attention, employing Rotary Positional Encoding and local masked attention to capture spatio-temporal relations, and supports forecasting for unseen locations via dynamic masking. Evaluated on a year-long Swiss dataset with 127 locations, it achieves horizon-averaged improvements in $NRMSE$ of around $0.17\%$ over strong baselines and is competitive with commercial NWP, while offering substantial computational advantages and robustness for real-world deployment. The results show that multimodal fusion yields clear benefits, particularly for day-ahead horizons, and demonstrate the model’s ability to operate without direct ground observations at new sites by leveraging satellite context and neighboring sensors.

Abstract

Accurate day-ahead forecasts of solar irradiance are required for the large-scale integration of solar photovoltaic (PV) systems into the power grid. However, current forecasting solutions lack the temporal and spatial resolution required by system operators. In this paper, we introduce SolarCrossFormer, a novel deep learning model for day-ahead irradiance forecasting, that combines satellite images and time series from a ground-based network of meteorological stations. SolarCrossFormer uses novel graph neural networks to exploit the inter- and intra-modal correlations of the input data and improve the accuracy and resolution of the forecasts. It generates probabilistic forecasts for any location in Switzerland with a 15-minute resolution for horizons up to 24 hours ahead. One of the key advantages of SolarCrossFormer its robustness in real life operations. It can incorporate new time-series data without retraining the model and, additionally, it can produce forecasts for locations without input data by using only their coordinates. Experimental results over a dataset of one year and 127 locations across Switzerland show that SolarCrossFormer yield a normalized mean absolute error of 6.1 % over the forecasting horizon. The results are competitive with those achieved by a commercial numerical weather prediction service.

Figures (18)

• Figure 1: SolarCrossFormer architecture with satellite data inputs. The encoder consists of a temporal transformer to encode the individual nodes' time series, a cross-attention transformer to correlate patch-node information and a second cross-attention transformer to correlate inter-node information. The decoder consists of a temporal transformer followed by a MLP.
• Figure 2: Attention layers. Left: Cross-attention. Right: Temporal attention.
• Figure 3: Local attention heads for the node cross attention mechanism (4 heads). The blue dot is the central node ($i$-th node) and the red dots are the nodes where the mask $\bm{M}_{i.a}$ is set to 1. A similar local attention is used for the cross attention with the satellite images.
• Figure 4: Spatial distribution of the weather stations that conform the network of sensors. The area of the whole image corresponds to the area covered by the cropped satellite images.
• Figure 5: Errors over the horizon: NRMSE for the best models trained under the MSE loss.