Federal learning load prediction method based on dynamic weighted aggregation

Abstract

The invention relates to a short-term load prediction technology of a power system, and aims to provide a federated learning load prediction method based on dynamic weighted aggregation. According to the method, an edge computing device adopts local data to carry out neural network training, network parameter change vectors are obtained and uploaded to a cloud server to carry out pairwise similarity calculation to generate a similarity matrix, and a consistency vector between local models in the current round of training is calculated. And then weighting network parameter changes of different local models based on the consistency between the accuracy of each round of local model for the server verification set and the local models to realize a local model cleaning effect, issuing the updated global model to the edge computing device by the cloud server, and repeating the above steps to realize the cleaning effect of the local models. And the server reaches the preset training round number. The problem that a dirty model possibly trained by a local model and data isomerism in federal learning cannot be effectively avoided in the prior art is solved, and convergence of global model training is accelerated.

Description

technical field [0001] The invention relates to a short-term load forecasting method of a power system and the fields of distributed and machine learning, belonging to the optimization technology and distributed load forecasting technology based on federated learning, and in particular to a federated learning load forecasting method based on dynamic weighted aggregation. Background technique [0002] With the development of digitization of the power system, the amount of data to be processed by the power system increases exponentially. Accurate load forecasting can assist in power dispatching and help maintain power stability in residential areas. The traditional centralized power load forecasting method puts a heavy burden on the server due to the increase in the amount of data, and also has the problem of data privacy leakage. In order to solve the problems of data privacy in distributed training, some researchers have proposed a federated learning framework. [0003] Fed...

AI Technical Summary

Problems solved by technology

[0004] However, due to the volatility of the power load data and the different data quality of each collection point, the training effect of the local model is also uneven. Directly aggregating the local models on the server may have problems such as a decrease in the accuracy of the aggregation model.

In order to solve this problem, some researchers have proposed a solution for model cleaning for specific devices, but these methods cannot be applied to other edge devices with similar characteristics.

Images