Data-driven Surface Solar Irradiance Estimation using Neural Operators at Global Scale

TL;DR

Developed using NVIDIA Modulus, this model represents the first adaptive global framework capable of providing long-term SSI forecasts and can be fine-tuned using satellite data, which significantly enhances its performance in the fine-tuned regions, while maintaining accuracy elsewhere.

Abstract

Accurate surface solar irradiance (SSI) forecasting is essential for optimizing renewable energy systems, particularly in the context of long-term energy planning on a global scale. This paper presents a pioneering approach to solar radiation forecasting that leverages recent advancements in numerical weather prediction (NWP) and data-driven machine learning weather models. These advances facilitate long, stable rollouts and enable large ensemble forecasts, enhancing the reliability of predictions. Our flexible model utilizes variables forecast by these NWP and AI weather models to estimate 6-hourly SSI at global scale. Developed using NVIDIA Modulus, our model represents the first adaptive global framework capable of providing long-term SSI forecasts. Furthermore, it can be fine-tuned using satellite data, which significantly enhances its performance in the fine-tuned regions, while maintaining accuracy elsewhere. The improved accuracy of these forecasts has substantial implications for the integration of solar energy into power grids, enabling more efficient energy management and contributing to the global transition to renewable energy sources.

Figures (8)

• Figure 1: 6-hourly averaged SSI forecasts over a 48-hour period. SFNO SFNOSFNO_checkpoint acts as prognostic model forecasting multiple weather variables at 6-hourly lead times. Our diagnostic model is then applied to retrieve the accumulated 6-hour SSI. The proposed methodology is generic and can be applied to other weather forecasting models.
• Figure 2: On the left: the Pix2Pix UNet architecture with residual blocks (ResBlock) showed in blue. The encoder and decoder are composed by strided convolutions (2D Conv) and strided transposed convolutions (2D TConv). The first encodes the weather forecasted fields to the latent space while the second decodes the latent prediction into a SSI field. On the right: the AFNO architecture with AFNO blocks showed in blue. The encoding in the latent space is obtained by symmetrically padding the input fields and successively applying a strided convolution. Then, the latent fields are processed through AFNO blocks and mapped to SSI by a linear layer.
• Figure 3: Evaluation vs. ERA5 6-hour SSI. Our model ($\text{AFNO}_{\text{ERA5}}$) is evaluated against ERA5 SSI product and compared against two benchmark models (Pix2Pix and MLP), likewise evaluated against the ERA5 SSI product. The validation is performed on 6-hour averaged fields of solar radiation for 2018.
• Figure 4: Evaluation vs SARAH3 6-hour SSI. Three version of our AFNO-based model are validated and compared on SARAH3 6-hour averaged SSI fields at $0.25^\circ$ resolution for 2018. $\text{AFNO}_{\text{ERA5}}$ is the ERA5-trained version, the second model is the same model finetuned on 17 years of SARAH3 SSI fields ($\text{AFNO}_{\text{f}}$), while $\text{AFNO}_{\text{SARAH3}}$ has the same architecture but it is fully trained on SARAH3 without pretraining on ERA5 irradiance product.
• Figure A.2: Evaluation vs BSRN 6-hour SSI. RMSE, MAE and Bias are shown as density plots computed over three different subsets of BSRN stations for all the SSI estimation models. In the first raw, all the stations are used. In the second row, the predictions are compared to the stations locating in the SARAH3 domain. In the third row, the metrics are computed only on the stations outside the SARAH3 domain.