Fast Power Flow Analysis Based On Deep Neural Network And Its Applications In Power Systems

In order to build a clean,low-carbon and sustainable energy system,developing the power systems intergrated with high-proportion renewable energy resources has become an inevitable trend in our country and even the world.Due to the random fluctuations of wind,light,etc.,the uncertainty in power systems has increased dramatically,and the operation states of the power grid are more complex and changeable,which seriously threatens the reliable and economic operation of the power grid,and even causes widespread blackouts.Power system power flow analysis(including power flow calculation and optimal power flow calculation)is a basic analysis tool to ensure reliable and economic operation of the power grid.However,in order to deal with the important impact of strong uncertainty in a timely and effective manner,the calculation frequency of power flow analysis need be increased sharply: on the one hand,in order to effectively consider various uncertainties,it is necessary to introduce a large number of scenarios and perform the power flow analysis,which is computationally burdensome;on the other hand,the calculation time for the above-mentioned massive power flow analyses should be reduced from days to hours even minutes.The contradiction between high-frequency calculation requirements and heavy calculation burden poses a new challenge to the existing power flow analysis methods.The data-driven deep neural network has the characteristics of strong approximation ability of high-dimensional complex nonlinear features and fast online calculation speed(input is directly mapped to output),which provides new ideas for solving the above-mentioned contradiction.In this regard,this article focuses on the "Fast Algorithm of Power System Power Flow Analysis Based on Deep Neural Networks" and explores the application in the reliability assessment of bulk power grids,including:1.A data-driven power flow model suitable for the changes of source/network/load in power systems is propose,which lays a model foundation for the deep neural networks to effectively mine the characteristics of the power system power flow analysis.First,low-dimensional feature vectors that can effectively characterize the changes of source/network/load are proposed.The dimension of the feature vectors only increases linearly with the scale of power systems,and the discrete variables of topology changes can be converted into continuous variables,which can effectively represent the disconnection state of the branch line and the important influence on power flow analysis.Then,according to the learning process of deep neural networks,considering the numerical characteristics of power system power flow analysis such as the properties and varying range of input and output feature vectors,a deep neural network power flow analysis model is built to realize the rapid calculation of basic power flow analysis in power systems.2.In order to obtain faster training efficiency and higher power flow calculation accuracy,this paper further proposes an efficient training method of deep neural network guided by the power flow model.First,based on the transfer principle of information in deep neural networks,that is,the information is not enlarged or reduced exponentially,an initialization method of deep neural network parameters for power flow analysis is derived.Besides,based on the branch power flow equation,a comprehensive loss function is proposed by considering the physical relationship among the training data and the approximation degree of the training data.A deep learning algorithm that considers the physical characteristics of the transmission network is further proposed.The proposed method can improve the training efficiency of deep neural networks,alleviate the problem of overfitting,and improve the accuracy of data-driven power flow calculations in power systems.3.Aiming at the problem that the optimal power flow has a large scale of constraints,which leads to heavy computational burdens,a deep learning method of neural networks oriented to the safety-constrained optimal power flow is further proposed.A calculation framework of security-constrained optimal power flow based on deep neural networks is proposed.The stacked denoising auto encoders is used to intelligently identify the active constraints that account for a small proportion.The framework can greatly reduce the scale of model,and effectively improve the calculation efficiency without precision loss.The generalization ability of deep neural networks is difficult to effectively deal with the problems of new scenarios such as topology change.A rapid updating strategy for deep neural networks based on transfer learning is proposed to provide technical support for the popularization and application of data-driven optimal power flow calculations.4.Bulk power system reliability assessment suffer from heavey computational burden.Therefore,this paper proposes a high-efficiency assessment method to resolve above problem based on data-driven power flow analysis from the following two aspects:complex feature extraction and calculation accuracy guarantee strategy.First,based on data-driven power flow analysis,the calculation framework of bulk power system reliability assessment is constructed,and the complex reliability feature extraction process is divided into several sub-tasks to reduce the learning complexity and alleviate the imbalance problem of load shedding samples,and this paper further proposes load shedding sensitive loss functions to improve the sensitivity of deep neural networks to load shedding samples.Secondly,based on the sample similarity,the adaptability criteria for deep neural network are proposed to determine whether to use the deep neural network to analyze for new scenarios and a single sample.Besides,a correction strategy for the output of the deep neural network is proposed based on theoretical derivation and rule design.Then,the sample with unreasonable calculation results can be effectively corrected and provide calculation accuracy guarantee for the industry application of data-driven bulk power system reliability assessment.In summary,this paper proposes a fast algorithm for power system power flow analysis based on deep neural networks,and studies the application in the bulk power system reliability assessment.The following three difficulties are overcome:high-dimensional complex nonlinear feature extraction,update strategies for power grid evolution scenarios,and accuracy guarantee strategy for data-driven calculation results.This paper adopts data-driven methods to solve the calculation bottleneck of power system power flow analysis,which has a good industrial application prospect.