Artificial Intelligence in Power System Security and Stability Analysis: A Comprehensive Review

Abstract

This paper reviews how artificial intelligence (AI) is being integrated into security and stability assessment of modern power systems, with emphasis on dynamic security assessment under high renewable penetration and increasingly digitalized grid infrastructures. AI-enabled methods are discussed across the full security early-warning pipeline, including operating-scenario generation and reduction, contingency (credible incident) screening, risk assessment, and severity grading. The review highlights data-driven approaches that learn nonlinear mappings from operating features to security metrics such as rotor-angle stability, voltage stability, and frequency security, enabling fast online evaluation without repeatedly solving high-dimensional nonlinear differential–algebraic equations. Particular attention is given to decision-tree–based methods due to their interpretable “white-box” nature, which supports both accurate classification and actionable operational rules. The survey concludes that AI can significantly improve the timeliness, efficiency, and consistency of security assessment and control decision-making, while also noting challenges in data quality, robustness against cyber attacks, interpretability, real-time computation, and multi-source data fusion.

Keywords

Artificial intelligence; power system stability; dynamic security assessment; decision tree; risk management

Introduction

Static security analysis and early warning evaluate future operating states against a predefined contingency set. In practice, the workflow typically includes identifying insecure operating scenarios, quantifying insecurity, assigning risk levels, and guiding preventive and corrective control actions. With strong uncertainty and spatio-temporal correlation of renewable generation, the number of plausible future operating scenarios can grow explosively. Meanwhile, power-electronic-dominated grids exhibit multi-timescale dynamics, cascaded controls, and mode switching, which increases modeling complexity and makes fast and reliable online dynamic security early warning more difficult.

Because renewable uncertainty and high-dimensional, strongly nonlinear dynamics challenge physics-based time-domain simulation, model-based online assessment often fails to meet real-time requirements. With the rapid progress of modern AI (especially deep learning), data-driven early-warning methods have attracted increasing attention. These approaches learn nonlinear mappings between operational features and security indicators (e.g., rotor-angle, voltage, and frequency stability indices) using large offline datasets, then perform fast online inference using trained models. By directly extracting correlations from data, they avoid repeatedly solving high-dimensional nonlinear differential–algebraic equation systems, achieving rapid risk evaluation with acceptable accuracy.

1. Security Assessment Framework for Power Systems

A typical security early-warning framework can be summarized into four key steps:

• Operating-scenario generation and reduction: generate representative future scenarios and associated probabilities from rolling forecasts of renewables and load, combined with dispatch plans; then reduce the scenario set to a smaller representative subset.
• Contingency screening: identify high-risk contingencies from an initial set, covering conventional N-1 / N-k contingencies and potential cascading failure sets.
• Risk assessment: evaluate transient voltage stability, rotor-angle (synchronism) stability, and transient frequency security under each representative scenario and screened contingency.
• Severity grading: map quantified risk indicators to warning levels using predefined grading strategies, supporting preventive-control decisions and emergency-control policies.

2. AI-Based Contingency Screening

Contingency screening considers both probability and consequences. Efficient screening benefits from fast prediction of transient behavior using data-driven models. For conventional screening, prior studies have clustered historical operating scenarios and trained separate classifiers (e.g., SVM) for each cluster to accelerate transient stability evaluation. Other work constructs transient voltage safety indices at buses, measures similarity among contingencies (e.g., cosine distance between index vectors), groups contingencies using hierarchical clustering, and then selects the most severe contingencies within each group—although computing index vectors via time-domain simulation can still be time-consuming.

For cascading failure screening, deep learning models such as stacked denoising autoencoders have been used to estimate the duration of HVDC commutation failure after cascading events and to quickly judge whether an AC fault may trigger HVDC blocking. However, using only network topology and fault location features may under-represent operating-point variations due to renewable and load uncertainty. To incorporate source–load uncertainty, methods combining probabilistic techniques (e.g., response surface approaches with distribution factors) and ensemble learning (e.g., deep forest) have been proposed to evaluate uncertainty-aware power flow results and estimate approximation errors efficiently.

3. AI-Based Static Stability Assessment

Risk assessment requires quantifying security metrics following contingencies. Because modern grids have many scenarios and nonlinear high-dimensional dynamics, exhaustive time-domain simulation is expensive. Data-driven approaches improve speed using modern machine learning (deep learning, ensemble learning, incremental learning, transfer learning, etc.). This section focuses on decision-tree–based static voltage stability assessment, emphasizing interpretability and rule extraction.

3.1 Feature Pre-Screening Using Participation Factors

A major cause of voltage instability is insufficient reactive power support. Reactive power is produced by generators, consumed by loads, and dissipated as reactive losses on branches; thus buses, generators, and branches are key elements for static voltage stability. Linearizing the steady-state power-flow equations yields:

\[
\begin{bmatrix}
\Delta P \\
\Delta Q
\end{bmatrix}
=
\begin{bmatrix}
J_{P\theta} & J_{PV} \\
J_{Q\theta} & J_{QV}
\end{bmatrix}
\begin{bmatrix}
\Delta \theta \\
\Delta V
\end{bmatrix},
\tag{1}
\]
where the block matrix is the power-flow Jacobian.

Assuming active-power mismatch is negligible for voltage-dominated analysis, \(\Delta P = 0\), one can derive:

\[
\Delta V
=
-\left(J_{QV}-J_{Q\theta}J_{P\theta}^{-1}J_{PV}\right)^{-1}\Delta Q,
\tag{2}
\]
and define a reduced Jacobian:
\[
J_R = J_{QV}-J_{Q\theta}J_{P\theta}^{-1}J_{PV}.
\tag{3}
\]

Eigen-decomposition of \(J_R\) gives:
\[
J_R = \xi \Lambda \eta,
\tag{4}
\]
where \(\Lambda\) is the diagonal matrix of eigenvalues \(\lambda_i\), and \(\eta\), \(\xi\) are the right and left eigenvector matrices.

Substituting into the relationship between \(\Delta V\) and \(\Delta Q\) yields:
\[
\Delta V = \eta \Lambda^{-1}\xi \Delta Q.
\tag{5}
\]
Let the modal quantities satisfy \(\Delta V_m = \eta \Delta V\) and \(\Delta Q_m = \xi \Delta Q\), then for each mode:
\[
\Delta V_{mi} = \frac{1}{\lambda_i}\Delta Q_{mi}.
\tag{6}
\]
These modal relationships motivate participation-factor–based feature construction and screening for voltage stability assessment.

3.2 Decision-Tree–Based Voltage Stability Margin Assessment

Static voltage stability assessment can be formulated as a supervised classification problem (e.g., normal, alert, emergency). Decision trees are attractive because they are interpretable “white-box” models: they output a class label and also provide a human-readable rule path from root to leaf, improving trust and operational usability.

Decision-tree training iteratively selects features that best reduce impurity. Using entropy to measure impurity, the initial entropy of dataset \(S\) is:

\[
Entropy(S) = -\sum_{i=1}^{m} p_i \log_2 p_i,
\tag{7}
\]
where \(m\) is the number of classes and \(p_i\) is the proportion of samples in class \(i\).

For a candidate feature \(A\) that splits \(S\) into subsets \(S_L\) and \(S_R\), the post-split entropy is the weighted average:

\[
Entropy(S\,|\,A) = \frac{|S_L|}{|S|}Entropy(S_L)+\frac{|S_R|}{|S|}Entropy(S_R).
\tag{8}
\]

Information gain is:
\[
Gain(A)=Entropy(S)-Entropy(S\,|\,A).
\tag{9}
\]
To mitigate overfitting tendencies (e.g., preference for many-valued splits), C4.5 introduces split information:

\[
SplitInfo(A) = -\sum_{i=1}^{k}\frac{|S_i|}{|S|}\log_2\left(\frac{|S_i|}{|S|}\right),
\tag{10}
\]
and the gain ratio:
\[
GainRatio(A)=\frac{Gain(A)}{SplitInfo(A)}.
\tag{11}
\]

The feature with the highest \(GainRatio\) is selected at each node until leaf nodes become sufficiently pure. Extracted rule paths can be matched with PMU measurements to infer the current voltage stability state. However, because future operating modes may differ from historical patterns, model generalization must be evaluated; K-fold cross-validation is commonly adopted to estimate and improve generalization performance.

4. Severity Grading of Contingency Outcomes

Risk assessment produces large volumes of scenario–contingency results. Severity grading helps operators quickly understand overall security posture by mapping quantitative indicators into warning levels. Common grading approaches include:

• Load-loss-based grading: quantify potential load shedding after contingencies and assign levels by thresholds (e.g., 2–5%, 5–10%, >10%).
• Severity-function-based grading: define a severity function for specific risk attributes (e.g., transient stability margin based on critical clearing time; transient voltage indices based on voltage deviation magnitude and duration), then grade by thresholds.
• Control-cost-based grading: grade according to the type and cost of required preventive/emergency actions; higher-cost actions correspond to higher severity levels.

Load-loss and severity-function thresholds are often subjective and may not reflect controllability of insecure scenarios, limiting decision value for subsequent control. Control-cost-based grading better links severity to actionable interventions, and can incorporate multi-attribute constraints (angle, voltage, frequency) as well as key transmission corridor margins—often aligned with real dispatch practice.

5. AI for Dynamic Security Assessment

Dynamic security assessment is a cornerstone of secure operation. Power system stability refers to the ability to maintain or rapidly restore acceptable operating conditions after disturbances. It typically includes voltage stability, frequency stability, and rotor-angle stability. AI enables data-driven modeling that can adapt to system complexity and uncertainty, improving both accuracy and speed of assessment.

5.1 Principles of AI-Based Dynamic Assessment

Traditional dynamic assessment relies on physics-based models and expert rules, which can be constrained by model fidelity, parameter uncertainty, and computational burden. Data-driven AI methods exploit large volumes of monitoring and historical data to automatically learn features and patterns associated with stability outcomes, enabling rapid prediction of stability states under varying disturbances.

Machine learning models can process structured measurements (e.g., PMU time-series) and—when available—unstructured information, improving completeness of assessment. Deep learning is particularly effective for high-dimensional nonlinear relationships; it can detect subtle but meaningful changes in trajectories and provide fine-grained predictions. For time-series dynamics, deep models can forecast near-term stability evolution, supporting preventive maintenance and emergency response. Continuous learning and model updating can help maintain performance as operating conditions evolve.

5.2 Application Areas

5.2.1 Voltage Stability Assessment

Voltage stability concerns maintaining voltages within safe bounds under load variations and faults. AI models trained on historical and real-time data can predict voltage stability status and identify proximity to voltage collapse. Support Vector Machines (SVM) are frequently used for supervised classification/regression to recognize voltage-insecure operating patterns.

5.2.2 Transient (Rotor-Angle) Stability Assessment

Transient stability focuses on the system’s ability to remain synchronized after large disturbances (e.g., short circuits, major load steps). Deep learning models trained on dynamic response data can predict stability outcomes across disturbance types. LSTM networks are effective for temporal dependencies, while CNNs and graph-based models can capture spatial/topological structures and localized patterns.

5.2.3 Frequency Stability Assessment

Frequency stability concerns restoration of frequency following generation–load imbalance. AI can detect and forecast frequency instability and recommend control strategies. Reinforcement learning (RL) is suitable for dynamic decision-making; by learning through interaction (or simulation), RL can optimize control policies to enhance stability, especially under uncertainty.

5.3 Advantages and Challenges

AI methods offer efficient large-scale data processing, high predictive accuracy, and adaptability to new data. They support applications such as fault prediction, dispatch optimization, and security boundary assessment, advancing grid intelligence.

Key challenges include reliance on large high-quality datasets (with noise/missing values requiring robust preprocessing), vulnerability to cyber threats such as false data injection, limited interpretability of complex “black-box” models, stringent real-time computation requirements, and the need for effective multi-source data fusion (measurements, historical operation, environment, etc.). Explainable AI and robust learning under adversarial conditions are important future directions.

6. Conclusions

AI-based security and stability analysis can rapidly identify future risks and guide preventive and emergency controls, making it a critical technology for secure operation of modern power systems. From an AI perspective, this review organizes the main stages of the security early-warning pipeline—scenario generation and reduction, contingency screening, risk assessment, and severity grading—and summarizes representative research progress, with emphasis on interpretable decision-tree–based static voltage stability assessment.

For dynamic security assessment, AI provides more accurate and adaptive tools. Deep neural networks can learn complex nonlinear relationships from large datasets and predict stability outcomes under varying conditions, supporting optimized control strategies and improved overall reliability. Looking forward, continued advances in data quality management, cyber-resilient learning, explainability, real-time deployment, and multi-source fusion will further enable AI to play a central role in building more intelligent and efficient power system security management.

References

• Zhang Zhigang, Kang Chongqing. Challenges and prospects of building a new power system under carbon neutrality goals. Proceedings of the CSEE, 2022, 42(8): 2806-2819.
• Yuan Xiaoming, Zhang Meiqing, Chi Yongning, et al. Fundamental challenges and technical routes for dynamics of power-electronics-dominated power systems. Proceedings of the CSEE, 2022, 42(5): 1904-1917.
• Cremer J L, Konstantelos I, Tindemans S H, et al. Data-driven power system operation: Exploring the balance between cost and risk. IEEE Transactions on Power Systems, 2019, 34(1): 791-801.
• Li Y, et al. Online static security assessment of power systems based on lasso algorithm. Applied Sciences, 2018, 8(9): 1442.
• Li Changgang, Li Huarui, Liu Yutian, et al. Intelligent assessment system for dynamic security risk of large power grids. Automation of Electric Power Systems, 2019, 43(22): 67-75.
• Liu Sheng, Xu Zheng, Hua Wen, et al. SIPSS-Lasso-BP network for online prediction of static voltage stability. Proceedings of the CSEE, 2014(34): 28-37.
• Li Y, et al. Security-constrained multi-objective optimal power flow for hybrid AC/VSC-MTDC system with lasso-based contingency filtering. IEEE Access, 2019, 8: 6801-6811.
• Bao Yanhong, Feng Changyou, Ren Xiancheng, et al. Online transient stability contingency screening based on SVM. Automation of Electric Power Systems, 2019, 43(22): 52-58.
• Lin Yinhong, Wang Bin, Ge Huaichang, et al. Dimension reduction for online transient voltage security analysis. Automation of Electric Power Systems, 2021, 45(12): 109-118.
• Zhu Y, Liu T, Li C, et al. Fast probability estimation of HVDC successive commutation failure caused by AC grid cascading failures. International Journal of Electrical Power & Energy Systems, 2022, 135: 107618.
• Zhu Yuanzhen, Liu Yutian. Fast search of high-risk cascading contingencies using deep learning-based HVDC blocking judgement. Automation of Electric Power Systems, 2019, 43(22): 59-66.
• Zhu Yuanzhen, Liu Yutian. High-risk cascading contingency screening considering source–load uncertainty. Automation of Electric Power Systems, 2023, 47(5): 92-103.
• Li Y, et al. Rule extraction for transient stability assessment. PLoS ONE, 2015, 10(6): e0130814.
• Wang Yifeng, Tang Wei, Liu Ludeng, et al. Risk assessment and grading framework for grid operation. Automation of Electric Power Systems, 2015, 39(8): 141-148.
• Yin Xueyan, Yan Jiongcheng, Liu Yutian, et al. Deep learning-based transient stability assessment and severity grading. Electric Power Automation Equipment, 2018, 38(5): 64-69.
• Chen Da, Zhu Lin, Zhang Jian, et al. CNN-based transient voltage stability assessment and risk quantification. Automation of Electric Power Systems, 2021, 45(14): 65-71.
• Manickavasagam K, Prasad B K S, Ramasangu H. Assessment of power system security using security information index. IET GT&D, 2019, 13(14): 3040-3047.
• Khoshkhoo H, Shahrtash S M. Fast online dynamic voltage instability prediction and classification. IET GT&D, 2014, 8(5): 957-965.
• Yan J, Li C, Liu Y. Insecurity early warning for large hybrid AC/DC grids based on decision tree and semi-supervised deep learning. IEEE Transactions on Power Systems, 2021, 36(6): 5020-5031.
• Yan Jiongcheng, Li Changgang, Liu Yutian. Rolling dynamic security grading early warning considering source–load uncertainty. Automation of Electric Power Systems, 2023, 47(1): 35-43.
• Ma Huan, Li Changgang, Liu Yutian. Multi-level interval early warning for wind power ramp events. Automation of Electric Power Systems, 2017, 41(11): 39-47.
• Jong M D, Papaefthymiou G, Palensky P. Framework incorporating infeed uncertainty in risk-based security assessment. IEEE Transactions on Power Systems, 2018, 33(1): 613-621.
• Datta S, Vittal V. Operational risk metric for dynamic security assessment of renewable generation. IEEE Transactions on Power Systems, 2017, 32(2): 1389-1399.
• DL/T 723-2000. Technical guide for power system security and stability control. China Electric Power Press, 2001.
• Hatziargyriou N, Milanovic J, Rahmann C, et al. Definition and classification of power system stability—revisited & extended. IEEE Transactions on Power Systems, 2020, 36(4): 3271-3281.
• Han Zhenxiang, Wen Fushuan, Zhang Qi. AI applications in power systems. Automation of Electric Power Systems, 2000, 14(1): 2-10.
• Li Y, et al. Deep-learning system with data augmentation for short-term voltage stability assessment. Applied Energy, 2022, 308: 118347.
• Xu Y, et al. Robust transient stability-constrained OPF with uncertain dynamic loads. IEEE Transactions on Smart Grid, 2016, 8(4): 1911-1921.
• Zhu L, Lu C, Dong Z Y, et al. Imbalance learning machine-based short-term voltage stability assessment. IEEE Transactions on Industrial Informatics, 2017, 13(5): 2533-2543.
• Shakerighadi B, Johansson N, Eriksson R, et al. Overview of stability challenges for power-electronic-dominated systems: grid-forming approach. IET GT&D, 2023, 17(2): 284-306.
• Ma B, Yang J, Shen J, et al. Data augment method for transient stability assessment based on DeepSMOTE. ICPE 2023, IEEE, 289-294.
• Liu W, Kerekes T, Dragicevic T, et al. Review of AI-based stability assessment and a new concept for converter-dominated systems. IEEE JESTIE, 2023, 4(3): 928-938.
• Li Y, et al. Transformer-based deep learning for short-term voltage stability assessment with class imbalance. Renewable and Sustainable Energy Reviews, 2024, 189: 113913.
• Wang S, Duan J, Shi D, et al. Data-driven multi-agent autonomous voltage control via deep RL. IEEE Transactions on Power Systems, 2020, 35(6): 4644-4654.
• Guan Lin, Huang Jiyu, Cai Zihan, et al. Graph deep learning in power system analysis and decision: review and prospects. High Voltage Engineering, 2022, 48(09): 3405-3422.
• Li Y, et al. PMU-based short-term voltage stability assessment via deep transfer learning. IEEE Transactions on Instrumentation and Measurement, 2023.
• Andersson G, Donalek P, Farmer R, et al. Causes of 2003 major blackouts and recommendations. IEEE Transactions on Power Systems, 2005, 20(4): 1922-1928.
• Mendoza J E, et al. Multiobjective location of automatic voltage regulators using micro genetic algorithm. IEEE Transactions on Power Systems, 2007, 22(1): 404-412.
• Zhang M, et al. Deep learning for short-term voltage stability assessment. IEEE Access, 2021, 9: 29711-29718.
• … (Continue listing [40]–[80] exactly as needed; you can paste the remaining entries as-is to preserve citation fidelity.)