Spatiotemporal CNN LSTM Based Multi Class Stability Assessment of Renewable Integrated Power Systems Using PMU Measurements
Abstract The increasing penetration of renewable energy resources and the resulting variability in power-system operating conditions have created significant challenges for reliable real-time stability monitoring and corrective control. Conventional stability assessment methods often require extensive computation and may not satisfy the low-latency requirements of Wide Area Monitoring Systems (WAMS). To address these challenges, this paper proposes a spatiotemporal deep learning framework that integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks for multi-class stability assessment using synchronized Phasor Measurement Unit (PMU) data streams. The CNN layers learn spatial correlations among measurements from different buses, while the LSTM layers capture the temporal evolution of disturbance dynamics. The proposed framework simultaneously classifies four stability conditions: Stable, Transient Instability, Voltage Instability, and Frequency Instability. The methodology is evaluated on the IEEE 9-bus, IEEE 39-bus, and IEEE 118-bus benchmark systems under diverse disturbances, including three-phase faults, generator disconnections, line contingencies, and sudden load changes. A total of 5,150 operating and disturbance cases are considered. The proposed CNN-LSTM model achieves an overall classification accuracy of 97.4%, with an average model inference latency of 17.3 ms for each 500 ms PMU observation window, outperforming standalone SVM, ANN, CNN, and LSTM models. Because the classifier requires a complete observation window, a stability decision becomes available 519.2 ms after disturbance onset, and the corrective action is applied at 0.649 s for STATCOM, 0.769 s for SVC, and 1.019 s for battery dispatch, excluding communication delay. All closed-loop results reported in this paper use these actuation instants. For transient instability on the IEEE 39-bus system, STATCOM support limits the critical generator to a peak rotor angle of 75.0 degrees, against loss of synchronism at 1.33 s without control, and synchronism is retained for communication delays up to approximately 334 ms. For voltage instability on the IEEE 118-bus system, SVC support limits the weak-bus voltage to 0.894 p.u. and the excursion below the 0.90 p.u. ANSI limit to 0.18 s, against collapse to 0.621 p.u. without control. For frequency instability following a 650 MW generator trip, the decision delay places the battery dispatch after the rate-of-change-of-frequency peak, so the maximum ROCOF is unchanged at 1.48 Hz/s and the nadir improves only from 48.79 Hz to 48.89 Hz; the measurable benefit is a shorter recovery, 4.3 s against 7.5 s to return above 49.8 Hz. The results therefore establish a mechanism-dependent latency limit: measurement-window-based deep-learning control is effective for rotor-angle and voltage mechanisms, which evolve over the first one to two seconds, and cannot act on the inertial phase of a frequency event. The proposed framework provides a scalable approach for data-driven stability monitoring and coordinated corrective control in renewable-integrated power systems, with the latency constraint quantified for each mechanism.
Authors
- Shashank Kumar (ORCID: https://orcid.org/0000-0003-0841-987X)
- Mangal Singh
- Mahesh Singh
Institutions
- Symbiosis International University (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11572281/v1
- Primary Topic
- Power System Optimization and Stability
- Type
- preprint