Research On Short-Term Load Forecasting In Energy Internet

Energy is an indispensable element for human economic and social development,and it serves as a fundamental driving force for industrialization and modernization.The Energy Internet is a system based on smart grids,which integrates various energy networks,such as hydro,wind,and solar,utilizing Internet technology to ensure efficient energy consumption.Short-term load forecasting is a crucial component for maintaining a smooth power system operation,particularly in the Energy Internet.However,with the growth of the Energy Internet,the power system has become increasingly intricate,making short-term load forecasting exponentially more challenging.Additionally,no existing short-term load forecasting algorithms offer load data privacy protection.This paper addresses these challenges by exploring ways to improve load forecasting accuracy and safeguard load data privacy.The main contributions of the paper are the following three points:(1)Given that current short-term load forecasting algorithms struggle to meet accuracy demands,this paper delves into attention mechanisms in detail,proposing a GRU-MA load forecasting model based on gated recurrent unit(GRU)and multi-head attention(MA)mechanisms.The model uses GRU to extract features from input vectors,and MA to enhance the weight of key features.Experimental results indicate that the proposed GRU-MA model has substantially improved load prediction accuracy in comparison to the traditional GRU model.(2)To further enhance the accuracy and generalization ability of the load forecasting model,this paper conducts a comprehensive study of Bagging ensemble learning and combines it with the self-attention(SA)mechanism to propose a Bagging-SA short-term load forecasting model.The model uses GRU-MA and temporal convolutional network(TCN)as weak learners,and the structure of Bagging is optimized.The SA mechanism is utilized as a combination strategy for the output results of weak learners,which dynamically adjusts the weight of each weak learner.Experimental results demonstrate that the Bagging-SA model significantly improves the accuracy and generalization ability of the original Bagging model for short-term load forecasting.(3)To address load data privacy concerns,this paper introduces horizontal federated learning(HFL)to the Bagging-SA model and proposes the Bagging-SA-HFL short-term load forecasting model.The HFL mechanism enables joint training of multi-regional load data while protecting the privacy of the data.Experimental results demonstrate that the proposed Bagging-SA-HFL model not only protects the privacy of load data effectively,but also enhances the load forecasting ability of the model.