Federated Learning-Based MPPT for Distributed Rooftop PV Systems: A Multi-Client Simulation Study
DOI:
https://doi.org/10.4314/drjeit.v14i3.7Keywords:
Federated learning; maximum power point tracking; rooftop photovoltaic systems; neural-network controller; non-IID dataAbstract
Rooftop photovoltaic (PV) systems operate under heterogeneous irradiance, temperature, orientation and shading conditions, making accurate maximum power point tracking (MPPT) essential for effective energy extraction. This study evaluates a federated-learning MPPT (FL-MPPT) framework for distributed rooftop PV systems using 15 heterogeneous PV clients and a 24-h environmental dataset sampled at one-minute intervals. Each client employs a compact 4–32–16–1 neural-network controller, while federated averaging (FedAvg) combines locally learned model parameters without requiring exchange of the underlying sensor data. Across the 15 clients, mean daily energy yield was 4.230 kWh for P&O, 4.325 kWh for incremental conductance, 4.366 kWh for local neural MPPT, 4.620 kWh for centralized neural MPPT and 4.486 kWh for FL-MPPT. Relative to the distributed/local baselines, FL-MPPT yielded gains of 6.06%, 3.73% and 2.75%, respectively, while centralized NN-MPPT was 2.99% higher than FL-MPPT. The centralized and FL results were available as matched client-level observations; the Wilcoxon signed-rank comparison of centralized NN-MPPT with FL-MPPT gave W = 0 and exact two-sided p = 6.10×10⁻⁵. Mean steady-state power ripple across six operating conditions was 1.08% for FL-MPPT, compared with 3.23%, 2.22% and 1.75% for P&O, incremental conductance and local neural MPPT. FL-MPPT also recorded a 0.55 s rise time, 0.90 s settling time, 1.9% overshoot and 3.5 W oscillation amplitude. Ten independent FL-MPPT runs under the study's 40-round training configuration produced 96.946 ± 0.268% MPPT efficiency, 4.4355 ± 0.0133 kWh harvested energy and 47.995 ± 0.055 V mean DC-bus voltage. With 705 trainable parameters, the controller requires 2,820 bytes per float32 model update, giving 3.384 MB of bidirectional model traffic over 40 federated rounds for 15 clients. These results show that centralized learning provides the highest recorded energy yield in the available matched comparison, while FL-MPPT retains strong controller performance with the distributed-learning and communication characteristics of federated training.
