| Resumen |
This study explores the effectiveness of an integrated heuristic-machine learning approach in forecasting solar radiation in various K & ouml;ppen climate zones. Our objective was to refine the model selection process for solar zoning, which involves characterizing solar radiation patterns in various geographic regions. We evaluated 107 heuristic models in 1216 automatic weather stations spread across tropical, arid, and temperate zones. Model performance was assessed based on criteria such as similarity of the data and distance to determine the most effective models. Of the 107 models tested, 98 proved optimal for at least one weather station. However, only 35.4% of all weather stations met the minimum performance benchmark, with 38.7% in Climate A, 31.5% in Climate B, and 26.3% in Climate C. Kmeans++, Hierarchical Clustering and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) techniques were used to classify regions based on solar irradiance. By integrating climatic data, including K & ouml;ppen climate types and meteorological variables, with geographical location. Key findings reveal significant clustering patterns related to specific climate types , demonstrating how solar radiation correlates with data dispersion and climatic subtypes. For Climate A, clusters primarily formed around a tropical subtype with irregular precipitation patterns, where climate data yielded welldefined clusters, and data dispersion showed more irregular grouping, often favoring multivariate linear and polynomial models. Climate B clusters around hot desert climates with rainfall in winter, showing a preference for daily multivariate linear models, while Climate C identifies optimal models only for weather stations with mild and low data dispersion. Differences between climate-based and data dispersion-based clustering were evident: climate information aligned closely with climatic subtypes, whereas data dispersion captured geographic dependencies. DBSCAN was particularly effective at identifying patterns in arid and temperate climates. This research not only clarifies the relationship between climate types and model performance, but also improves the methodology for selecting appropriate models in regions with scarce data, contributing significant advancements in the development of gray-box models for solar radiation forecasting. |
