Development and validation of a novel model-agnostic machine learning framework for photovoltaic systems performance assessment under complex weather conditions

• Development of a machine learning framework independent of specific PV models, adaptable to different photovoltaic systems and climates. • The model's key strength is its extensive training on 2,187 diverse photovoltaic configurations, making it highly robust against extreme weather variations like partial shading and rapid changes in irradiance and temperature. • Introduces a new Maximum Power Point Tracking (MPT) method using a deeply optimized neural network (9-16-9-1 topology), representing a paradigm shift from conventional approaches. • Benchmarking shows a significant quantitative advantage, achieving asymptotic precision (99%), very fast response dynamics (80ms), and exceptional output stability with minimal fluctuation. • This machine learning-based approach solves the critical issue of weather-dependent recalibration, establishing ML-driven MPT as a robust and universal solution for maximizing solar energy yield in unpredictable environments The optimization of maximum power point tracking represents a fundamental challenge in maximizing the energy yield of photovoltaic systems in the presence of temperature variations or partial shading. While conventional approaches (P&O, INC) exhibit intrinsic limitations under dynamic operating conditions, contemporary hybrid solutions (fuzzy-PID controllers, adaptive control systems) remain constrained by complex parameterization requirements. Our work introduces a paradigm shift through the development of a deeply optimized and universally applicable neural architecture (9-16-9-1 topology) whose performance surpasses current state-of-the-art methods. The distinctive feature of our approach lies in its extensive training on 2,187 heterogeneous photovoltaic configurations, endowing it with unmatched robustness against extreme meteorological varia- tions (200-1600 W/m2, 0-60°C). Furthermore, the framework is truly model-agnostic: the same pre-trained network can be deployed on diverse commercial PV technologies without fine-tuning by simply inputting the target module's specifications, eliminating the need for module-specific retraining. Comprehensive benchmarking reveals marked quantitative superiority: a tracking efficiency of 99.99%, representing a 2.15-8.16% absolute improvement over conventional methods (P&O, INC), and a 0.2–4.9% improvement over Fuzzy Logic Control, a 108% reduction in power oscillations compared to P&O, and sub-100ms response times. The framework’s computational efficiency ensures real-time deployability on commercial PV modules, as verified via MATLAB/Simulink simulations. By addressing the critical limitations of weather-dependent performance recalibration, this work establishes ML-driven MPPT as a robust, universal solution for energy yield maximization in stochastic environments, with implications for scalable renewable energy integration.