Decentralized asynchronous adaptive federated learning algorithm for securely prediction of distributed power data

Introduction: Improving the precision and real-time speed of electricity data prediction while safeguarding data privacy and security holds immense significance for all power system participants' decision-making. To surmount the issues of exorbitant computational expenses and privacy breaches of traditional centralized prediction methods, this paper proposes a decentralized asynchronous adaptive federated learning algorithm for securely prediction of distributed power data, which makes predictions from distributed data more flexible and secure.Methods: First, each regional node trains its own deep neural network model locally. After that, the node model parameters are uploaded to the decentralized federated learning chain for ensuring local data protection. Asynchronous aggregated update of the global prediction model is then achieved via block mining and shared maintenance. The algorithm has been enhanced based on the traditional federated learning algorithm, which introduces an asynchronous mechanism while adaptively adjusting the regional node model weights and local update step size to overcomes the inefficiency of traditional methods.Results and Discussion: The experimental analysis of actual electricity price data is conducted to compare and analyze with the centralized prediction model, study the impact of model adoption and parameter settings on the results, and compare with the prediction performance of other federated learning algorithms. The experimental results show that the method proposed in this paper is highly accurate, efficient, and safe.