Research And Implementation Of Peak Regulation Method For Data Center Based On Load Forecasting

In the context of promoting the low-carbon power grid,developing virtual power plants is of great significance in facilitating the consumption of renewable energy,the stability of the power supply,and the intelligent energy transformation.Meanwhile,with the booming cloud computing,the rising trend of high energy use of data centers has posed new challenges to the grids.Due to the coupling of energy and information domains,data center loads can be aggregated to virtual power plants to participate in peak regulation in power markets.This exploits the joint scheduling of crossdomain resources and increases grid flexibility.In the above process,accurate load forecasts provide a key basis for data centers and virtual power plants to develop reasonable regulation strategies.Therefore,this paper conducts research on load forecasting and peak regulation for data centers.It also designs and implements a peak regulation system for data center based on load forecasting according to proposed approaches.First,to address the limited prediction performance of existing load forecasting models caused by the underutilization of future information,this paper proposes a data center load prediction method based on feature adaptive exploration.The model uses multi-task learning and long and short-term memory networks to improve prediction accuracy by exploring future features;it also designs a two-stage learning algorithm,including feature iterative learning strategy and adaptive selection mechanism,to enhance both future features through mutual assistance and the stability of feature learning.Experimental results show that the proposed method outperforms all baselines and effectively improves load prediction precision.Then,to address the loss of benefits and lack of willingness to trade due to asymmetric information and lack of interaction between energy and information resources in data center participation in virtual power plant services,this paper proposes a joint optimization method for energy trading and computational scheduling.This approach uses contract theory to establish the joint optimization problem of peak shaving transaction and load scheduling,and solves the asymmetric information through incentivecompatible constraints;considering the huge number of decision variables and constraints,it is solved in parallel using the proximal Jacobi alternating direction multiplier method to reduce time complexity.Empirical experiments show that the proposed scheme surpasses the baselines and effectively improves benefits and cooperation between data centers and virtual power plants.Finally,based on the two approaches presented above,a peak regulation system for data center based on load forecasting is designed and implemented in this paper.The system conducts functional and nonfunctional requirement analysis for cloud operators and algorithm researchers and designs four functional modules:load forecasting,peak regulation,data management,and user management;it also completes the functional and non-functional tests.Test results show that the system is feasible and robust and can verify the effectiveness of the suggested approaches.