Objective The high penetration of renewable energy results in power systems with low inertia and weak damping, posing a major challenge to frequency stability following active power disturbances. Existing assessment methods are limited by low accuracy, high computational costs, and insufficient training samples resulting from frequent grid reconfigurations. To address these limitations, this paper proposes a data-driven frequency stability evaluation method based on multi-dimensional stability indicators. The proposed approach improves evaluation accuracy and computational efficiency, alleviates data scarcity, and supports the secure operation of modern power systems. Method The study adopted a multi-step systematic approach. First, a simplified multi-machine frequency response model was established. Under different renewable energy penetration scenarios, the relationships between power imbalance and three key frequency parameters (maximum rate of change of frequency, transient frequency extremum, and quasi-steady-state frequency deviation) were analyzed. Combined with frequency stability criteria, a multi-dimensional power-frequency indicator evaluation system was constructed to assess frequency stability. Second, the input features were preprocessed: CP-SMOTE was used to address sample imbalance, and the LassoNet model was applied for feature selection to reduce the feature dimensionality. A BiGRU-KAN model was then developed by integrating the KAN attention mechanism into the BiGRU module for frequency stability evaluation. Accuracy, precision, recall, and F1-score derived from the confusion matrix were used to evaluate the model performance. Finally, an improved Wasserstein distance generative adversarial network (WGAN) with an optimized gradient penalty mechanism was developed to generate sufficient representative samples for model training under system topology changes. The IEEE 10-machine 39-bus system was used as the test system for simulation-based verification under multiple scenarios. Results and Discussions Comprehensive simulation experiments were conducted on the IEEE 10-machine 39-bus system to evaluate the performance of the proposed method. The key results are summarized as follows. In the sample preprocessing stage, a significant imbalance was observed between stable and unstable samples in the initial dataset. After applying CP-SMOTE, the proportion of stable samples increased from 37.5% to 52.5% of the total samples, effectively balancing the dataset and providing a solid foundation for subsequent model training. For feature selection, the LassoNet model was employed. The feature selection results showed that the model achieved the highest accuracy of 95.14% with 13 selected features, which were subsequently used as the model inputs. In the classification performance evaluation, the BiGRU-KAN model rapidly achieved high accuracy during the early training stage and maintained a stable accuracy of 99.7% thereafter. It significantly outperformed the comparative models (BiLSTM-KAN, GRU-KAN, and KAN), demonstrating superior classification performance. Regarding convergence performance, the cross-entropy loss of the BiGRU-KAN model decreased sharply during the early training stage and approached zero after only a few training epochs, stabilizing at approximately 5 training epochs. Compared with the other models, it exhibited both superior objective function optimization and faster convergence, reaching the optimal solution with fewer training epochs. Furthermore, the BiGRU-KAN model showed smaller loss fluctuations throughout the training process and maintained a consistently low loss after convergence. Performance evaluation based on the confusion matrix showed that the BiGRU-KAN model outperformed the comparative models across all classification metrics, indicating a higher prediction accuracy and a stronger capability to distinguish between stable and unstable system states. In terms of training efficiency, the average processing time per training sample was 5.13 ms for BiGRU-KAN, 6.48 ms for BiLSTM-KAN, 2.55 ms for GRU-KAN, and 2.01 ms for KAN. Although BiGRU-KAN required more processing time per sample than GRU-KAN and KAN, it remained more efficient than BiLSTM-KAN. Considering accuracy, convergence, and computational efficiency together, GRU-KAN and KAN achieved faster processing but at the expense of lower accuracy and inferior convergence. In contrast, the bidirectional gated recurrent unit architecture of BiGRU-KAN more effectively captured sequential dependencies in the data, while the KAN attention mechanism further enhanced key feature extraction, resulting in a better balance between accuracy and computational efficiency. To address sample generation under system topology changes, an improved WGAN with an optimized gradient penalty mechanism was employed. Comparative analysis showed that generating the same number of samples required 9.29 s using the original WGAN and 8.58 s using the improved model. Together with the discriminator and generator loss analysis, these results demonstrate that the improved WGAN provides both higher-quality sample generation and faster generation speed. Furthermore, stability tests under noise interference showed that, although the introduction of noise vectors increased the training loss compared with the noise-free condition, the discriminator could still accurately distinguish generated samples from real samples under high-dimensional noise, while the generator maintained high-quality sample generation. These results verify the robustness of the improved WGAN model. Conclusion This study proposes a data-driven frequency stability assessment method incorporating multi-dimensional stability indicators to address the issues of low accuracy, high computational cost, and insufficient training samples caused by topology changes in frequency stability assessment of modern power systems. The proposed multi-dimensional power-frequency indicator system, CP-SMOTE-based sample balancing, LassoNet-based feature selection, and the BiGRU-KAN model jointly improve assessment accuracy and computational efficiency. In addition, the improved WGAN addresses the shortage of training samples caused by topology changes and enhances the model's generalization capability. Simulation results on the IEEE 10-machine 39-bus system demonstrate that the proposed method achieves significant improvements in both accuracy and computational efficiency compared with conventional methods, providing an effective basis for accurate frequency stability assessment and rapid frequency regulation strategy formulation. Future work will extend the proposed method to more complex power system models and investigate its adaptability under extreme operating conditions to further enhance its practical engineering applicability.
本文模型在Python3.9环境中运行,使用PyCharm Community 2020.1.3进行编译。将仿真样本按照4∶1的比例划分为训练集和验证集。将本文BiGRU‒KAN模型与其他主流组合式模型BiLSTM‒KAN、GRU‒KAN和KAN模型的准确率及交叉熵损失值进行对比,以验证其收敛性和准确性,设置各模型的参数设置情况相同,以保证对比的公平性。
YangLingrui, GuoCheng, GuoZhen,et al.Inertia estimation method of power system based on total variation regularization[J].Electric Power Automation Equipment,2026,46(3):178-185. doi:10.16081/j.epae.202507006
PeiMing, LiYilin, YeLin,et al.Power system frequency response for combined prediction considering spatio-temporal characteristics of inertia[J].Proceedings of the CSEE,2023,43(Supp1):57-69.
JuPing, ZhengYi, JinYuqing,et al.Analytic assessment of the power system frequency security[J].IET Generation,Transmission & Distribution,2021,15(15):2215-2225. doi:10.1049/gtd2.12171
[6]
LiGuoqing, LiuXianchao, XinYechun,et al.Research on frequency stability of power system with high penetration renewable energy:A review[J].High Voltage Engineering,2024,50(3):1165-1181.
PaturetM, MarkovicU, DelikaraoglouS,et al.Stochastic unit commitment in low-inertia grids[J].IEEE Transactions on Power Systems,2020,35(5):3448-3458. doi:10.1109/tpwrs.2020.2987076
[9]
YanZiming, XuYan.Data-driven load frequency control for stochastic power systems:A deep reinforcement learning method with continuous action search[J].IEEE Transactions on Power Systems,2019,34(2):1653-1656. doi:10.1109/tpwrs.2018.2881359
[10]
LuWenan, ZhuQingxiao, LiZhaowei,et al.A prediction me-thod of new power system frequency characteristics based on convolutional neural network[J].Journal of Shanghai Jiao Tong University,2024,58(10):1500-1512.
HuangMingzeng, HuYahan, WenYunfeng,et al.Assessment method of transient stability for power system based on joint mutual information maximization and natural gradient boosting[J].Automation of Electric Power Systems,2021,45(8):155-165. doi:10.7500/AEPS20200812004
HanTiansen, ChenJinfu, LiYinhong,et al.Study on interpretable surrogate model for power system stability evaluation machine learning[J].Proceedings of the CSEE,2020,40(13):4122-4130. doi:10.13334/j.0258-8013.pcsee.190963
ZhaoDongmei, ZhengYarui, XieJiakang,et al.Frequency stability evaluation of power system containing wind power based on light gradient boosting machine and generative adversarial network[J].Power System Technology,2022,46(8):3181-3190.
ZhangYihao, HanSong, RongNa.Frequency stability assessment method of power system based on CNN-Informer and DeepLIFT[J].Electric Power Automation Equipment,2025,45(7):165-171.
XieJian, SunWei.A transfer and deep learning-based method for online frequency stability assessment and control[J].IEEE Access,2021,9:75712-75721. doi:10.1109/access.2021.3082001
[21]
ZhangHan, WangCheng, BiTianshu.On-line fast frequency calculation after power system disturbance based on fusion of physics and data knowledge[J].Power System Technology,2022,46(11):4325-4335.
DengXianzhe, YaoWei, HuangWei,et al.Transient frequency prediction driven by data-model fusion based on adaptive time window[J].Power System Technology,2024,48(4):1551-1562.
HaoLü, ChenJinhui, DuYoutian,et al.Model-data driven intelligent enhanced method for frequency stability discrimination[J].Power System Technology,2025,49(8):3314-3323.
LiuJiahao, WangCheng, BiTianshu.Node equivalent inertia index for temporal-spatial frequency dynamics of renewable energy power system and its applications[J].Proceedings of the CSEE,2023,43(20):7773-7788.
TaoYukun, JiangJiangming, YangFeifei,et al.Active frequency response strategy for wind turbines considering wind-thermal power coordination[J].Automation of Electric Power Systems,2026,50(4):152-162.
LiuZhijian, TangHao, GuoCheng,et al.Step dead zone control strategy and optimization of DC FLC for asynchronous interconnected sending system with high proportion of renewable energy[J].Electric Power Automation Equipment,2024,44(10):116-123.
SheDaoming, JiaMinping.A BiGRU method for remaining useful life prediction of machinery[J].Measurement,2021,167:108277. doi:10.1016/j.measurement.2020.108277
[34]
WangRui, XuXinchao, LuJing.Short-term wind power prediction:Feature selection and ISSA-CNN-BiGRU approach[J].Advanced Engineering Sciences,2024,56(3):228-239.
JingjingLyu, ZhuGuanghui, HeChuan.Data-driven short-term photovoltaic power forecasting under extreme weather conditions using GRU-KAN model[J].Renewable Energy,2026,257:124756. doi:10.1016/j.renene.2025.124756
[37]
WangPenghui, LianZengzeng, Núñez-AndrésM A,et al.A GCN-GRU-KAN-Based Framework for UWB 3D localization in adverse geometric configurations[J].Measurement,2026,258:119066. doi:10.1016/j.measurement.2025.119066
[38]
LiManman.Wind power prediction of BiTCN-BiGRU-KAN model based on attention mechanism[C]//Proceedings of the 2024 IEEE 2nd International Conference on Sensors,Electronics and Computer Engineering(ICSECE).Jinzhou:IEEE,2024:1640-1645. doi:10.1109/icsece61636.2024.10729502
[39]
GongWei, LiuTao, LiXuanyi,et al.Automatic cable defect detection method based on deep learning and time-frequency domain reflectometry[J/OL].Advanced Engineering Sciences,2025[2025-12-01].
LiuZibing, CuiShuangxi, LiuQingqing.New energy scenario generation method based on an improved WGAN-GP integrating multi-scale convolution and multi-head attention mechanisms[J].Power System Technology,2026,50(1):272-283.
PengWeifeng, DongShufeng, QiuJian.A WGAN-GP approach for multi-scenario specific photovoltaic generation with feature-constrained and target-driven optimization[J].Proceedings of the CSEE,2025,45(23):9126-9139.
MiJiaqi, MaCongcong, ZhengLihua,et al.WGAN-CL:A Wasserstein GAN with confidence loss for small-sample augmentation[J].Expert Systems with Applications,2023,233:120943. doi:10.1016/j.eswa.2023.120943
[46]
WangJingwen, WenYunfeng, LiaoBangkun,et al.Rapid evaluation method of regional inertia of power system based on frequency prediction error[J/OL].Advanced Engineering Sciences,2025[2025-12-04].
XuXiaolu, QiZitong, HanXiumei,et al.Combined-task deep network based on LassoNet feature selection for predicting the comorbidities of acute coronary syndrome[J].Computers in Biology and Medicine,2024,170:107992. doi:10.1016/j.compbiomed.2024.107992
[49]
ChenHui, ZhangTian, ChenRunbin.CXR image classification based on residual convolution and multi-headed self-attentions[J].Advanced Engineering Sciences,2024,56(3):219-227.
KomsrimorakotP, SiriborvornratanakulT.Enhancing fraud detection in imbalanced motor insurance datasets using CP-SMOTE and random under-sampling[J].Journal of Big Data,2025,12(1):172. doi:10.1186/s40537-025-01217-3
[52]
ZhouXuan, WuWeili.Transformer fault diagnosis method based on improved SMOTE unbalanced sample processing and IHPO-DBN[J].Power System Protection and Control,2024,52(11):21-30.