Research On Frequency Stability Prediction Method Of Low Inertia Power System Based On Machine Learnin

With the introduction of the 2030 carbon peak and 2060 carbon neutrality targets,large-scale renewable energy generation units are replacing traditional fossil fuel generation units and integrating into the grid.However,most renewable energy generation units lack the inertial response and primary frequency support capabilities of traditional generation units,leading to a lack of rotating inertia in the grid and the need to maintain frequency stability.Real-time assessment and prediction of grid frequency safety can assist power dispatch and planning personnel in quickly understanding the dynamic frequency response characteristics of the grid and formulating reasonable preventive control measures to ensure that the frequency remains within a specified range after accidents.Traditional time-domain simulation analysis and aggregation equivalent model analysis require the modeling of the physical models of power systems,which is difficult to achieve with renewable energy generation units that are typically integrated into the grid through power electronic interfaces.Additionally,the complexity of calculations increases with the size of the grid.In recent years,with the rapid development of artificial intelligence(AI)technology,building power system frequency stability prediction models using machine learning(ML)algorithms has become an important direction for future digital grid construction.ML algorithms can extract features from operational data to avoid the modeling difficulties associated with complex physical mechanisms in traditional methods.Moreover,trained ML models can rapidly and accurately predict frequency instability events,assisting power dispatch and planning personnel in maintaining frequency stability in low-inertia power systems.This article presents a series of studies on power system frequency stability prediction based on the low-inertia characteristics of power systems and advanced ML algorithms.The main research topics and achievements include:(1)To address the problem of insufficient frequency support capability in low-inertia power systems,we propose a new frequency security index(FSI)as a sample label for the training dataset.This index combines the magnitude of the active power imbalance disturbance to evaluate frequency stability.Specifically,the FSI divides the perturbed system frequency into three states: "absolute security," "relative security," and "insecurity." The relative security reflects the critical stable state of the system in low-inertia conditions,indicating that power dispatchers need to increase inertia.(2)To address the issue that current ML-based frequency stability prediction methods heavily rely on manual experience-based tuning,we propose a method based on automated deep forest for predicting the frequency security index under expected power system faults,which combines the Bayesian optimization algorithm and the deep forest algorithm.We conduct case studies on a New England 10-machine 39-node system and a South Carolina 90-machine 500-node system,and compare three automated ML algorithms and five classical ML algorithms,demonstrating the superiority of the proposed method.(3)To address the issue that current ML-based frequency stability prediction methods have poor prediction accuracy due to the difficulty of extracting global features from power system operational data,we propose a method based on Vision Transformer and copula entropy for predicting the frequency security index under expected power system faults.Due to the O(N2)computational complexity of the Transformer,we use copula entropy for feature selection to reduce the dimensionality of input data and avoid long training times.Finally,we conduct case studies on a New England 10-machine 39-node system and a South Carolina 90-machine500-node system,comparing eight ML methods and verifying the effectiveness and robustness of the proposed method.(4)To further improve the accuracy and efficiency of frequency stability prediction,we propose a method based on Co At Net and SHAP values for predicting frequency stability.Co At Net uses a combination of convolution and attention mechanisms to overcome the limitations of traditional single-model methods in fully extracting data features.Additionally,selecting all features as input to the deep learning model may result in a large computational burden.Therefore,we propose a feature selection method based on SHAP values to select effective features as input for the prediction process,significantly reducing numerical complexity while maintaining high prediction performance.Finally,we conduct case studies on a New England 10-machine 39-node system and a South Carolina 90-machine 500-node system,comparing seven ML methods and verifying the effectiveness and robustness of the proposed method.