Data-driven Reactive Power Optimization Strategy of Distribution Network Based on Autoencoder Constrained Temporal Convolutional Networks
The increasing penetration of renewables in the distribution networks causes frequent voltage violations.As a widely studied method,reactive power optimization has been successfully applied to distribution networks to reduce power losses and stabilize voltages.By coordinating multiple multi-timescale regulation devices like capacitor banks and smart inverters through three stages,this paper proposes an innovative data-driven reactive power optimization strategy by using autoencoder and constrained temporal convolutional networks.First,the optimization problem is formulated as a mixed-integer second-order cone programming to solve for the historical optimal reactive power dispatching strategies.Then,the proposed model is trained with the historical operational data and the optimal strategies,and the unreasonable results are circumvented by a corrective layer.In actual applications,the well-trained model gives optimization solutions based on real-time system measurements to deal with fluctuations in the distribution networks.Finally,as verified by a modified IEEE 33-bus system,the proposed method can achieve 98.80%accuracy of the mixed-integer second-order cone model with consuming only 7.14%of its time.Compared to other popular deep learning methods,the proposed method has the best performance and better practicality.
reactive power optimizationconstrained temporal convolutional networksdata-drivensecond-order cone programmingautoencoder
万方数据
维普
