Research Of Data-driven And Physics-based Combined Method For Power Grid Transient Stability Assessment

With expansion of power grid scale,increase of time-varying factors and strong nonlinearity,both data-driven alone and physics-based alone power system transient stability assessment(TSA)methods encounter more challenges than before.Combining these two methods is believed as an effective way to improve application performances considering the natural complementary features of these two methods,which helps deal with those newly emerged challenges.This thesis is founded by National Key Research and Development Program of China(High-performance analysis and situation awareness techniques for interconnected large-scale power grid,2018YFB0904500)and the National Natural Science Foundation(Prediction theories and techniques of transient frequency situation awareness in power systems based on combined physical and data-driven modeling,51877037),focused on how to combine datadriven and physics-based methods organically to solve tough problems in power system TSA,where an in-depth study of combined methods for power system TSA related problems is carried out.The main work content and innovation of this thesis are as follow: typical combination patterns for data-driven and physics-based combined methods are summarized based on review study of data-driven and physics-based combined method applications.Further,data-driven and physics-based combined methods are put forward to solve problems faced with physics-based or data-driven power system TSA methods.The specific works are as follow:1)Research status about data-driven and physics-based combined methods is summarized by investigating researches in various research areas.Four typical combination patterns for data-driven and physics-based combined methods,including feedback,serial,parallel and embedded pattern,are then summarized,which helps further research on applications of data-driven and physics-based combined method in power system TSA problems.2)A deep reinforcement learning(DRL)based feedback parameter correction method for high voltage direct current(HVDC)transient models is put forward.This fast parameter correction method relying on online measurements feedback is put forward to solve the mismatch problem between HVDC model parameters and actual HVDC system.This method is based on deep deterministic policy gradient(DDPG)algorithm,taking HVDC transient measurements including HVDC current,voltages at rectifier and inverter side as feedback inputs and identifying main control parameters of HVDC model.This method enables to guarantee both computation accuracy and efficiency,which is suitable for online parameter correction of HVDC models.3)A power system transient frequency features prediction method based on serial prediction-correction approach is put forward.This frequency features prediction method is used to solve the large error problem of system frequency response(SFR)model caused by model simplification.It is realized by combining SFR model and extreme learning machine(ELM),where SFR model preserves factors strongly related with transient frequency stability.The ELM based error correction model is used to describe the incidence relation between weakly related factors and prediction error.This method is able to realize good performance in both computation accuracy and efficiency.Moreover,the proposed method can take effects when sample quantity is small.4)A parallel large-disturbance rotor angle stability prediction method based on localglobal information is proposed,which aims to solve the hesitation problem of datadriven power system TSA in predicting boundary samples.It relies on combination of local and global information,where the global information is processed by ELM based TSA model.The local information is processed by trajectory fitting(TF)based TSA model and it is activated when result of ELM based TSA model is judged as unauthentic according to ELM based TSA model output value.Once activated,the TF based TSA model is used to check the prediction result of ELM based TSA model with local measurement of generator angle.This method takes full advantage of the computation speed of ELM based TSA model,while applies the TF based TSA model to make verification when hesitated boundary samples are encountered,which thereby further improves the accuracy of the prediction.5)A critical clearing time(CCT)prediction method considering physics knowledge embedding and instance transfer is put forward,to make up the shortages of sample insufficiency and lack of physics knowledge in constructing data-driven models.Firstly,this method combines integrated extended equal area criterion(IEEAC)and ELM in serial prediction-correction pattern.Then the ELM algorithm is transformed to adapt to properties of power system operation,where prediction accuracy under different TSA scenarios is differentiated and samples can be shared among similar power system operation modes.This CCT prediction method is both fast and accurate.Moreover,the performance of CCT prediction accuracy under severe scenarios is improved with the transformed ELM algorithm.