Research On Methods For Transient Stability And Frequency Stability Assessment Based On Ensemble Learning

In recent years,the increasing load is approaching the operating limit of power system,and the integration of new energy resources has significantly reduced the inertia level.Both factors increase complexity of operation mode and stability form of the power grid,imposing unprecedented challenges on the power system stability analysis and control.Currently,it still mainly relies on the time-domain simulation to simulate the dynamic response process of the system after fault.However,the difficulty of modeling and intensively computational burden make it only suitable for offline simulation analysis.Therefore,this paper introduces advanced artificial intelligence technology such as ensemble learning,focusing on transient stability and frequency stability assessment,and aims to improve the computational efficiency,accuracy and adaptability of the evaluation method.In general,the main works and contributions are as follow:(1)In order to realize the on-line rapid assessment and credibility evaluation of power system transient stability,a transient stability assessment method combining joint mutual information maximization(JMIM)and natural gradient boosting(NGBoost)is proposed.Based on JMIM,joint mutual information and“maximum of the minimum” principle are adopted to mine data correlation,so as to screen out the key characteristics of the power grid and reduce the computational complexity.Besides,a transient stability assessment method driven by NGBoost is constructed,which can predict parameters of the conditional probability distribution of the model in the form of a function to achieve probability prediction,thereby quantifying the credibility.Meanwhile,an adaptive credibility threshold correction method is designed to assist NGBoost to complete the transient stability assessment with high-confidence in a very short assessment period.(2)To improve the practicability of data-driven methods,a CatBoost-based transient stability assessment and its interpretable analysis method are proposed.A hyperparameter optimization model based on Bayesian optimization is constructed,which can avoid the non-optimality and inefficiency of heuristic parameter tuning.Meanwhile,the objective function of multi-weighted evaluation index is designed to ensure that the performance of CatBoost is improved comprehensively during the hyperparameter optimization process.Furthermore,SHapley Additive ex Planation(SHAP)is introduced to understand why a stable/unstable status prediction is made by CatBoost.Therefore,the trust of operators in the predictions can be enhanced,and more reference information for the formulation of emergency control measures can be provided.(3)Aiming at the problem of the current frequency stability assessment methods ignore the effect of uneven sample distribution,this paper proposes a frequency stability assessment method considering the frequency deviation distribution and penalty cost.To avoid the dimensional explosion problem,the multi-source information fusion method is used to extract key input feature subsets from power system operation information.Then,a cascaded light gradient boosting machine(CasLightGBM)is constructed,and a punishment sensitivity mechanism is embedded into its loss function,which helps CasLightGBM automatically correct the sample loss value in training according to the probability distribution of frequency deviation samples and the penalty cost of conservative prediction.Leveraging CasLightGBM and the punishment sensitivity mechanism,the prediction accuracy of maximum frequency deviation prediction can be improved,and the misjudgment rate of frequency instability can be reduced.