A preprint proposes a global climate map for photovoltaic modules that combines two modeled outcomes: annual energy yield and module lifetime. Its selected system has six main climate clusters and 15 subclusters, named Tropical, Desert, Continental, Temperate, Boreal and Polar.
The source dataset combines hourly weather-station data with 12 environmental features and two performance targets. Those targets are annual energy yield and module lifetime, simulated for crystalline-silicon (c-Si) modules installed worldwide with the PVMD Toolbox.
The authors then applied nearest-neighbor interpolation to all features and targets, using three neighbors to make the global dataset more spatially uniform. The resulting map is therefore a model-based comparison shaped by both the interpolation and the physical simulation.
That pairing is the point of the exercise: the proposed classification is built around how a photovoltaic module is modeled to produce electricity and retain performance, rather than around environmental variables in isolation.
The model's clearest signals
Gaussian process regression produced the best reported accuracy, with an RMSE of 0.007 MW h for energy yield and 1.5 years for lifetime. RMSE is the study's measure of prediction error. Each regression model was trained on 85% of the data and tested on the remaining 15%.
Adding more inputs did not keep improving the predictions. RMSE performance leveled off at around six features and rose again at 12 in some models, a pattern that suggests overfitting. To rank the inputs, the study used a brute-force feature-importance procedure across five regression methods.
The leading feature depended on the target. GHIann ranked highest for energy yield, while Tmean ranked highest for lifetime. In plain language, annual global horizontal irradiation was the top-ranked input for predicted output, and mean ambient temperature was the top-ranked input for predicted lifetime.
From ranked inputs to climate zones
The climate groups were assembled with hierarchical clustering. Ward linkage was selected because it produced the most balanced structure among the tested linkage methods. The final classification contains six main clusters and 15 subclusters. Because the system combines two performance targets, it is designed as a PV-specific classification rather than an energy-yield-only map.
One result may surprise readers: the main-text results place the highest discounted lifetime energy yield in Temperate-LT. Low-temperature, or LT, subclusters outperform their high-temperature, or HT, counterparts. This is a comparison of modeled lifetime-energy output, not a measured report from deployed systems.
The labels are useful, but not fixed
The classification's stability is a central caveat. Small perturbations to the feature-importance scores can change the dendrogram and the selected number of clusters, even when the top-three ranking remains unchanged. The six-zone result should therefore be read as one selected representation of the modeled data, not as a uniquely settled climate standard.
The study presents the map as a framework for organizing model-based comparisons. Its outputs are predictions from physical simulations for c-Si modules, so the supplied analysis does not establish whether the same classification transfers to other module designs or PV technologies.
The document is arXiv:2608.25448v1, dated 26 Aug 2026. It states that the dataset and methods software are available through a 4TU dataset. No funding statement appears in the supplied text; the explicit disclosure concerns generative-AI use in writing.
Paper data and sources
Original title: Energy Yield and Lifetime Climate Classification via Machine Learning for Optimizing Photovoltaic Module Design and Materials
Authors: Youri Blom, Sofia Dutto, Alexandru Costache et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text