Answers are grounded in this paper and cite their sources.
Partial shading conditions (PSC) produce multiple local maxima in the power–voltage characteristic of photovoltaic (PV) systems, increasing the risk that conventional maximum power point tracking (MPPT) methods converge to suboptimal operating points. This work presents a physics-guided machine-learning (ML)-assisted MPPT framework that combines bypass-diode-aware PV modeling, normalized MPP-voltage prediction, and bounded measurement-based micro-refinement. A five-parameter single-diode model with explicit substring-level bypass-diode activation is used to generate an 8312-sample PSC dataset, and seven supervised regression models — ANN, Random Forest, SVR, GPR, XGBoost, LightGBM, and CatBoost — are evaluated using the normalized target V mpp / V oc . Among the evaluated models, ANN achieved the highest static prediction accuracy, with an MSE of 1.070 V 2 , an MAE of 0.514 V, and an R 2 of 97.40%. Across an independent 30-point PSC tracking benchmark, the ANN achieved a mean tracking factor (TF) of 97.20%, which increased to 99.95% after bounded micro-refinement, with 100% of the evaluated operating points satisfying both TF ≥ 95 % and TF ≥ 98 % . The improvement introduced by refinement relative to ANN alone remained statistically significant after Holm correction ( p Holm = 0.0097 ), whereas the higher mean TF relative to conventional P&O (93.10%) was not statistically significant after correction. Under four dynamic operating scenarios, the refined ANN maintained mean TF values of approximately 99.93%–99.99%, while under the highest evaluated measurement and input uncertainty level, its mean TF remained 98.80%. The selected 10–128–64–1 ANN contains 9729 trainable parameters, requires approximately 9536 MAC operations per inference, and occupies approximately 38.9 kB in FP32 representation. Real-time deployment on a TI C2000 TMS320F28379D with 1 ms ADC acquisition, 20 ms MPPT updates, and 10 kHz pulse-width modulation switching, together with converter-level experiments, demonstrates the practical embedded feasibility of the proposed framework.
No takes yet. Share an insight, caveat, or question.
Khan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: