Data Driven-based Dynamic Optimal Power Flow Algorithm For Distribution Network

With the development of advanced metering infrastructure and artificial intelligence technology in distribution network,it is possible to mine historical data based on data driven technology and find optimal control strategy quickly based on prior knowledge,instead of relying on physical model.To this end,this thesis studies the data driven-based dynamic optimal power flow algorithm for distribution network.It focuses on the study of power curve generation method for loads,solar plants,and wind farms based on deep neural network and the study of dynamic optimal power flow framework based on data driven technology,then verification is carried out.The main work of this thesis is as follows.The methods for generating curves of loads,solar plants,and wind farms based on deep neural network are proposed under the condition of insufficient data.Three different deep generative models are constructed and compared.Based on the concept of variational lower bound,the variational function is used to replace the posterior probability,and the similarity is measured by information divergence.A variational automatic encoder model is proposed.For the problems of low generated data quality and unstable training process of traditional generative adversarial network,an improved generative adversarial network is constructed.The parameters of the discriminator and the generator are updated by back propagation algorithm,and the power curves are generated by the trained generator.The non-linear independent components estimation model is proposed.The simple distribution is transformed into complex distribution by a series of reversible transformation functions to obtain the probability distribution of the real load curve.The effectiveness of the proposed method is verified on the London smart meter dataset.Several control devices such as on-load tap changer,shunting capacitors,controllable distributed generators and intelligent soft open point are considered.The method for constructing the database with optimal power flow solution is proposed.The optimal power flow model of the distribution network is built with the minimum power loss as the objective function.The daily load curve is divided into several time intervals by enumeration method,and each time interval is solved by using multiple heuristic algorithms 50 times,and the best result is used as the solution of static optimal power flow.The effectiveness of the proposed method is verified by the simulation on the IEEE 33-bus system.A data driven-based dynamic optimal power flow method for distribution network is proposed.The numerical features of the daily load curve of the distribution network are extracted by the piecewise aggregate approximation.A classification tree is constructed based on the symbol features,and the historical daily load curves are initially screened using the classification tree.The threshold is tentatively set by dichotomy,and the historical daily load curves are quadratically screened by using the extracted numerical features to obtain the candidate set.The distance between each historical daily load curve in the candidate set and the current daily load curve is calculated,and the historical daily load curve with the smallest distance is found.The segmentation strategy of the historical daily load curve is used to divide the current daily load curve into n time intervals.In each interval,the active power and reactive power of each node are taken as the features of static optimal power flow.Principal component analysis is used to reduce the dimension of the feature and the entropy weight method is used to determine the weight of the feature.After the rough matching and the fine matching,the historical load series with the smallest distance from the current load series is obtained,and the optimal power flow control scheme of the historical load series is applied to this interval.The simulation is carried out on IEEE 33-bus system and IEEE 69-bus system,and the effectiveness and superiority of the proposed method are verified by comparing with traditional methods.