A cluster-aware XAI-MCDM framework for province-level solar PV suitability screening in Turkey using real operational PV data

Abstract Identifying suitable regions for solar photovoltaic (PV) deployment is commonly addressed through GIS- and multi-criteria decision-making (MCDM)-based frameworks; however, criterion weights in such studies are often derived from expert judgement or purely statistical assumptions rather than from field-observed PV performance. This study proposes a cluster-aware explainable artificial intelligence–MCDM framework that converts operational PV generation data into interpretable meteorological criterion weights for province-level PV suitability screening. In the first stage, a leakage-controlled artificial neural network (ANN) was trained using 47,185 cleaned hourly observations from four operational PV plants in Turkey, including Adıyaman, Bolu, Çanakkale, and Konya. The target variable was defined as capacity-normalized hourly PV performance, and the predictor space was restricted to five transferable meteorological descriptors: solar-radiation-related information, air temperature, relative humidity, cloud cover, and wind speed. The ANN achieved a test R 2 of 0.782 and an nRMSE of 11.92%, indicating sufficient predictive reliability for explainability-based criterion weighting. Multi-XAI analysis, including SHAP, permutation importance, mutual information, and correlation-based relevance, identified the solar-radiation-related descriptor as the dominant driver of PV performance. The derived ANN-XAI weights were integrated with MEREC objective weights and transferred to a performance-priority MCDM ensemble composed of MARCOS, TOPSIS, and EDAS. In the second stage, the calibrated framework was extended to all 81 Turkish provinces using NASA POWER-based meteorological and solar-radiation proxy descriptors, together with applicability-domain risk assessment and solar–climate clustering. Clustering was used as an interpretation and robustness layer to identify comparable solar–climate regimes and to support within-cluster benchmarking rather than to replace the national MCDM ranking. The results identified Mersin as the most robust national candidate, while Burdur, Isparta, Karaman, Konya, Kayseri, Nevşehir, Niğde, and Aksaray formed a strong high-priority province portfolio. Cluster-wise analysis further showed that Turkey can be interpreted through distinct solar–climate regimes, enabling regional candidates to be evaluated within comparable climatic contexts. The proposed framework provides an explainable and data-driven decision-support tool for strategic province-level PV suitability screening, while its outputs should not be interpreted as final parcel-scale siting recommendations.