Maximum power point tracking using unsupervised learning for photovoltaic power systems

The Maximum Power Point Tracking (MPPT) technique in the solar energy field optimises the performance of solar panels in different atmospheric conditions and variable loads. In this study, we present a new method that uses unsupervised learning (K-means clustering) to identify the atmospheric clusters of solar irradiance and cell temperature (G, T) and delimit homogeneous atmospheric zones or clusters to reduce the search space of the optimal parameters (Vmp, Imp, and Pmp). The data collected for one year is segregated into 12 clusters; in every cluster, 04 regions are defined based on every cluster's centroid (Gc, Tc). A local search of the reference voltage/Duty cycle per cluster region is initiated for every sensed (G, T). Variable atmospheric conditions and resistive loads are tested. The results show that the efficiency of the DC/DC converter is 97.5% with a settling time (4.013 ms/5.577 ms) compared to the Perturb and Observe (P&O) the conventional tracking method and the Particle Swarm Optimisation (PSO), both applied locally inside a cluster and a deviation of 2% from the global maximum.