TurboFed: A Federated Learning Approach to The PHM of Distributed Wind Turbines

This paper proposes one approach with federated learning technique to address practical challenges faced by the emerging green energy industries, i.e., wind turbines in terms of Predictive Health Management (PHM). Not as many federated learning applications being used in the scenarios only for simulation, the application of federated learning in this paper is focused on the real industrial problems with raw data collected from the fields. Huge amount of real data was collected by sensors on more than ten wind turbines across different areas in China and transmitted to the storage for in-time processing. The framework proposed in this paper called TurboFed, can handle the raw data and achieves good prediction performance in the practical wind generated power systems. The framework showed its help on improving the efficiency of the wind turbines. The paper has brought three novel results. First, as far as known, the framework here is the first federated learning framework addressing position adjustment of wind turbines in the real environment. Second, it deploys customized recurrent neural computing models to the wind turbines which are considered the client devices under the federated learning paradigm. Finally, it incorporates new customized aggregation algorithms on the sever side.