Dynamic aggregation-based federated learning for fault diagnosis of distributed PV systems

Photovoltaic (PV) power is the primary way to utilize solar energy. Ensuring the safety and reliability of PV systems is essential for a stable and efficient power supply, where intelligent fault diagnosis (FD) is one of the key means. However, in large-scale distributed PV systems, regular intelligent FD methods face practical problems such as response delay, high communication overhead, and storage pressure. Besides, there exist comprehensive requirements for resource utilization and diagnostic reliability. To deal with these problems, this paper proposes a dynamic aggregation-based decentralized federated learning method. Combined with compressed sampling, a diagnostic comprehensive evaluation method is proposed to balance resource utilization and diagnostic reliability. Three strategies of efficient training, effective participation, and dynamic selection are designed to reduce computing and communication costs further and improve model aggregation efficiency. The proposed method is verified based on the grid-connected PV model. Compared with traditional FL, the aggregation interval, aggregation round, and waiting time are reduced by 17.353%, 13.2%, and 7%, respectively. • Accurate and energy-efficient fault diagnosis for photovoltaic systems. • Distributed federated learning with novel dynamic aggregation strategies. • Compressed sampling mechanism with an adjustable compression ratio. • Comprehensive experiments under different meteorological conditions.