Abstract
Gridded global horizontal irradiance (GHI) databases are fundamental for analysing solar energy applications' technical and economic aspects, particularly photovoltaic applications. Today, there exist numerous gridded GHI databases whose quality has been thoroughly validated against ground-based irradiance measurements. Nonetheless, databases that generate data at latitudes above 65$^{\circ}$ are few, and those available gridded irradiance products, which are either reanalysis or based on polar orbiters, such as ERA5, COSMO-REA6, or CM SAF CLARA-A2, generally have lower quality or a coarser time resolution than those gridded irradiance products based on geostationary satellites. Among the high-latitude gridded GHI databases, the STRÅNG model developed by the Swedish Meteorological and Hydrological Institute (SMHI) is likely the most accurate one, providing data across Sweden. To further enhance the product quality, the calibration technique called "site adaptation" is herein used to improve the STRÅNG dataset, which seeks to adjust a long period of low-quality gridded irradiance estimates based on a short period of high-quality irradiance measurements. This study, differing from the conventional statistical approaches, adopts machine learning for site adaptation. Nine machine-learning algorithms have been analysed and compared with conventional statistical ones to identify Sweden's most favourable technique for site adaptation. Three weather stations of SMHI are used for training and validation. The results show that, due to the spatio-temporal heterogeneity in model performance, no universal model can be identified, which suggests that site adaptation is a location-dependent procedure.
Introduction
Accurate solar radiation data is the bedrock of the renewable energy transition. However, for regions near the Arctic Circle, obtaining precise data is challenging. Most high-quality gridded products rely on geostationary satellites, which face a "blind spot" at latitudes above 65° due to their limited field of view. In Sweden, researchers often turn to the STRÅNG model developed by the Swedish Meteorological and Hydrological Institute (SMHI). While STRÅNG is sophisticated—integrating data from polar-orbiting satellites and reanalysis—it still contains inherent biases when compared to ground-level measurements.
This study addresses these biases through "site adaptation," a process that uses short-term, high-quality ground measurements to calibrate long-term gridded datasets. While traditional methods like linear regression or quantile mapping are common, they often struggle with the complex, non-linear atmospheric variables that influence irradiance at high latitudes.
Methodology and Data
The research treats site adaptation as a regression problem, benchmarking nine machine learning (ML) algorithms against traditional statistical baselines. The study utilized two primary data sources: the STRÅNG mesoscale solar radiation product (2.5 km spatial resolution) and ground measurements from three SMHI weather stations representing distinct Swedish climates: Kiruna (high latitude), Visby (high direct irradiance), and Norrköping (representative average climate).
The ML models were categorized into three tiers to evaluate varying levels of complexity:
Predictors included STRÅNG GHI, solar azimuth, solar elevation, and time-based features (month, day, hour). Models were trained on data from 2008–2009 and validated against a multi-year test period from 2010–2015.
Results and Discussion
The analysis revealed several key insights regarding the efficacy of machine learning in solar data calibration:
- Machine Learning Superiority: ML models consistently outperformed traditional linear and quantile mapping methods. For instance, in Kiruna, the raw STRÅNG model's Mean Bias Error (MBE) of 10.32 W/m² was reduced to -0.12 W/m² using the k-NN model.
- Complexity vs. Performance: Interestingly, deep learning models (CNN, LSTM) did not always surpass shallow models. This suggests that standard 1-2 year training datasets might be insufficient for deep architectures to fully realize their potential compared to robust shallow models like SVR or k-NN.
- Spatio-Temporal Heterogeneity: A critical finding was the "No Free Lunch" theorem in practice; no universal model exists. A model optimal for Kiruna was not necessarily the best for Visby, confirming that site adaptation is a highly location-dependent procedure.
- Accuracy Gains: ML models generally improved $R^2$ scores from 0.86 (raw STRÅNG) to as high as 0.92 (LSTM in Visby).
Conclusion
Machine learning offers a powerful toolset for enhancing high-latitude solar radiation products. While ML significantly reduces bias and improves accuracy, the choice of model must be tailored to the specific site due to climatic heterogeneity. Future research will focus on developing monthly trained models to better handle seasonal variations and investigating methods to reduce variance during extreme weather events where underlying gridded models may underperform.
Summary updated