A Two-Stage Unsupervised–Supervised Machine Learning Framework for Maximum Power Point Tracking of Photovoltaic Systems

Maximum Power Point Tracking (MPPT) is essential for maximizing photovoltaic (PV) energy extraction under changing operating conditions. This paper investigates a two-stage machine-learning framework combining K-means or fuzzy C-means (FCM) clustering of irradiance–temperature conditions with region-specific support vector regression (SVR), Gaussian process regression (GPR), and bagged ensemble trees (BET) to estimate the maximum-power-point reference voltage. The predicted voltage is supplied to a proportional–integral–derivative (PID)-controlled boost converter, and closed-loop performance is evaluated in MATLAB R2024/Simulink under changing irradiance, temperature, and load. Under the original evaluation protocol, two-cluster K-means—BET achieved the lowest MAE among the K-means prediction configurations, whereas integrated-energy rankings did not consistently follow voltage-prediction rankings. An additional leakage-controlled assessment used 70% of data for training and 30% for independent test observations, with clustering, standardization, and five-fold cross-validated Bayesian hyperparameter optimization performed using training observations only. In this stricter comparison, unclustered SVR achieved the lowest overall MAE, while FCM with three clusters improved BET MAE by approximately 9.7%. These results show that clustering benefits depend on the regression model and evaluation protocol. The closed-loop energy findings apply to the original simulated configurations; the new strict hold-out models were not separately tested in closed-loop simulation. Hardware implementation and broader-dataset validation remain future work.