| With the continuous development of distributed power generation technology,a large number of distributed energy sources,electric vehicles and energy storage systems and other source-load devices are integrated to the distribution network,which brings great challenges to the stable operation of the distribution network.Distribution network state estimation can help operators to learn about the running state of the distribution network in real time and provide data support for subsequent operations.At present,there are mainly problems of poor system observability and uncertainty in operation in the distribution network.Poor system observability makes it impossible to solve for state estimation.The uncertainty of data and topological uncertainty in the operational uncertainty lead to the poor accuracy and timeliness of the conventional distribution network state estimation methods.Therefore,this thesis addresses the abovementioned problems by conducting research on distribution network state estimation methods,which is divided into three main parts.Firstly,this thesis analyzes the existing traditional physical methods and elaborates.This thesis proposes a method to construct pseudo-measurement,which is different from the traditional load forecasting,using existing data and real-time measurement information as well as historical data as input to avoid the problem of degradation of forecasting accuracy caused by missing data input such as weather and time.The physical model-based state estimation is completed by inputting the constructed pseudomeasurements together with the real-time measurements into the state estimation model,and its drawbacks and limitations are analyzed through different scenario settings.Secondly,during the actual operation of the line,there may be changes in its parameters and measurement data could have uncertainty,which limits the application of traditional physical methods and results in a decrease in accuracy.This thesis proposes iRes Net,a modified version of the traditional residual neural network,and utilizes it to develop a state estimation model.The approach is based on data analysis and has the advantage that it does not require complex line parameters and only requires a small number of real-time measurements and constructed pseudo-measurements to obtain state estimation results It is also robust to large data errors,and its operational efficiency is greatly improved to better meet the system operational timeliness requirements.The simulation is validated on a typical distribution system with IEEE33 node and IEEE118 node respectively,and it is demonstrated that better results can be achieved compared with traditional physical method and other data-driven methods.Finally,topology uncertainty exists for distribution networks whose topology can change frequently due to issues such as system optimization.The unknown nature of the new topology and the small amount of historical data cause the data-driven approach proposed above to be difficult to apply.In this thesis,a time-varying topology state estimation based on migration learning is proposed,which takes the existence of correlation between topologies as the starting point and uses the historical data of the source topology as well as a small amount of data of the new topology to complete the training to improve the effectiveness of the data-driven method.Simulation verification has been performed across diverse topologies and distribution networks of varying scales,providing unequivocal evidence for the efficacy of the innovative approach suggested herein. |
