Data-Driven Small-Signal Stability Assessment Of Power Systems Based On Virtual Regional Eigenvalue

With the expansion of grid scale and the increase of the proportion of renewable energy generation connected to the grid,the power system faces the issues of high dynamic complexity and many uncertainties,which pose challenges to the safety and stability analysis of the power system.In order to remove the reliance on the detailed time-varying models and improve the accuracy,real-time and engineering practicality of the small-signal stability assessment,this thesis focuses on the power system small-signal stability assessment with the support of massive operational data,use of machine learning technologies,core of the virtual regional eigenvalue,and aims to improve the model accuracy and computational efficiency.The following four studies were conducted in the aspects of offline training and online updating of the small-signal stability assessment: rapid generation of effective data for data-driven small-signal stability model,small-signal stability assessment model based on virtual regional eigenvalue,probabilistic assessment of small-signal stability taking into account uncertainty factors,and online smallsignal stability assessment method based on adaptive partial updating strategy.Firstly,an efficient sampling generation method suitable for data-driven small-signal stability assessment models was investigated to provide data support for the subsequent assessment model construction.This method generates a sufficient number of high-entropy operating points using damping ratio sensitivity as the guiding information and adaptive search step size as the sampling approach,thereby reducing the sample size and information redundancy required for training the assessment model.To address the issue of inconsistent search step sizes,which leads to over-sampling or under-sampling during the sampling process,an adaptive step size determination method based on sensitivity magnitude changes was proposed to improve the sampling efficiency.To tackle the challenges of having too many scenarios to consider in the training samples and the high-dimensional operating space in large systems,scenario screening methods based on key transmission lines and parameter screening methods based on sensitivity were proposed,further reducing the number of generated samples and the complexity of the proposed method.Using the generated effective data as training samples,this thesis proposes a fast assessment method for small-signal stability based on virtual regional eigenvalues.In time-varying power system operation,understanding the exact location of each eigenvalue is unnecessary and may lead to computational waste;thus,intervals are typically used as control units.This method abandons the focus on individual critical eigenvalues and instead focuses on their distribution regions.By dividing the weak damping ratio range into critical regions,the concept of virtual regional eigenvalues is proposed,which realizes the identification of dominant eigenvalues within each region with minimal computational cost to meet real-time engineering requirements.Considering that the performance of this method is sensitive to the region partitioning method,a region sparsity index based on the Mean-shift algorithm is introduced to verify the rationality of the current region partitioning method.Finally,to accommodate the uncertainty of the number of critical eigenvalues at different operating points in the model,a composite long short-term memory network is designed to improve the accuracy of the small-signal stability assessment.Furthermore,to address the issue that deterministic assessment cannot reflect the stochastic nature of the system due to the increased uncertainty factors in the power system,a probabilistic small-signal stability assessment model based on data-driven models is proposed.This method employs the concept of virtual regional eigenvalues,which mitigates the problem of reduced accuracy in data-driven models caused by large differences in eigenvalue values at adjacent operating points due to uncertainty.To tackle the complex modeling issue of dual uncertainty from wind power and load,a quantile regression model based on deep neural networks is constructed,which achieves accurate prediction of the confidence intervals of virtual area eigenvalues.Finally,the probability distribution function is obtained by kernel density estimation based on the aforementioned results,and a stability probability index is further proposed to provide more guidance information for preventive control.Finally,since the operating scenarios are contantly changing,the performance of the datadriven models may struggle to adapt to the time-varying power system in the long term,potentially leading to a decrease in the assessment accuracy at the current operating point.Therefore,to enhance the adaptability of the data-driven small-signal stability assessment model,a timely updating online assessment method is proposed.This method proposes an adaptive partial update strategy by measuring the similarity between the old and new data in adjacent update cycles,thereby reducing redundant computational overhead and time consumption during online updates.Moreover,within a single update cycle,to address the issue of insufficient learning of some key characteristics by the data-driven model,the concept of the base state point is introduced,further improving the accuracy of the small-signal stability assessment.