Study On The Generation Method Of Large-scale Power Flow Samples For Data-driven Methods

The complexity and stochastics of the operation of electrical power system are increasing with more permeability of renewable energy,closer interaction with the society as well as larger scale of power electronics applications,while the insecure states of the system become more important and more diversified.It is one of the key problems for future power systems,since the number and dangerousness of insecure states are both accentuated.On the other hand,samples from real power systems' operation is much less than satisfied,since the power systems are operating under normal conditions for most of time.Data-driven method is a suitable method with high-efficiency,in which the most important problem is how to obtain plenty of samples.The most popular methods are continuous power flow(CPF)method and scheduled power flow(SPF)method.The drawback of CPF is the bad ability of covering,while the drawback of SPF is the risk of convergency problems.Furthermore,to implement SPF,the requirement of the quality of the person who are scheduling the power flow is of vital importance.Finally,it is a huge task for manual handling.Both two methods are based on traditional power flow calculations,and the lack of reasonable sample generation speed and risk of non-convergence and ill-condition power flow seriously restrict the real application of the methods.A totally new method for power flow sample generation is proposed in this thesis to solve the related problems.The key idea of the proposed method is to switch the solving target of power flow calculation.Injected powers for all buses are calculated with given bus voltages,instead of the calculation of bus voltage with pre-set injected powers,which can avoid the iteration of nonlinear power flow equations,and only substitutions and simple calculations are required.The efficiency of the proposed method can be faster for more than one order of magnitude comparing with traditional power flow solving method,without non-convergence risk,so it is more suitable to generate large amount of power flow samples under various conditions.Several application scenarios are including: online analysis with AI techniques,which may require tens of millions of samples for offline training;automatic operation mode generation for bulk power systems.The improvement of the speed and the reduce of difficulty for power flow sample generation is validated via multiple scales of sample systems from IEEE14 to IEEE300,in which the power flow for specified interface is predicted by neural networks,and the generated samples are feed into those networks for training.At the same time,the spreadability of the data set generated by the proposed method is better,and it is less inclined to be overfitted.Finally,the generalization performance of the proposed method is much better than traditional CPF.