Enhancing solar PV suitability mapping in the Middle East using an optimized deep learning framework

The shift toward sustainable energy has underscored the importance of optimizing PhotoVoltaic (PV) site selection through cutting-edge technological approaches. This study introduces an optimized Deep Learning (DL) framework for mapping PV suitability. The framework combines TabNet-an attentive and interpretable DL model-with the Optuna optimizer for efficient hyperparameter tuning and is further enhanced by an eXplainable Artificial Intelligence (XAI) approach. The study evaluates 12 key techno-economic factors along with an inventory of 612 PV solar stations to assess PV site suitability. The performance of the proposed approach was compared with two DL models-Tabular Prior Data Fitted Network (TabPFN) and Feature Tokenizer + Transformer (FT-Transformer)-as well as seven classical Machine Learning models, including Decision Tree, Random Forest, Gradient Boosting, CatBoost, Support Vector Machine, Naïve Bayes and K-Nearest Neighbors. Results demonstrated that the proposed architecture outperformed all other models on both validation and testing datasets, achieving classification accuracies of 0.875 and 0.886, respectively. The spatial suitability map indicated that 17.3 % (∼1231,254 km²) of the Middle East's land area is highly suitable for PV deployment, predominantly along the coasts and in the northern and northwestern regions. XAI, implemented via SHapley Additive exPlanation, revealed that proximity to infrastructure had the most significant impact on predictions.