Optimized Extreme Learning Machine for Power System Transient Stability Prediction Using Synchrophasors

A new optimized extreme learning machine- (ELM-) based method for power system transient stability prediction (TSP) using synchrophasors is presented in this paper. First, the input features symbolizing the transient stability of power systems are extracted from synchronized measurements. Then, an ELM classifier is employed to build the TSP model. And finally, the optimal parameters of the model are optimized by using the improved particle swarm optimization (IPSO) algorithm. The novelty of the proposal is in the fact that it improves the prediction performance of the ELM-based TSP model by using IPSO to optimize the parameters of the model with synchrophasors. And finally, based on the test results on both IEEE 39-bus system and a large-scale real power system, the correctness and validity of the presented approach are verified.

💡 Research Summary

This paper proposes a novel framework for real‑time transient stability prediction (TSP) in power systems by integrating synchrophasor measurements, an Extreme Learning Machine (ELM) classifier, and an Improved Particle Swarm Optimization (IPSO) algorithm for hyper‑parameter tuning. The authors first extract a set of informative features from Phasor Measurement Unit (PMU) data that capture the early dynamics of a disturbance. Specifically, voltage magnitude, phase angle, frequency, and their first‑order derivatives are sampled over a short post‑fault window (0.1–0.5 s). For each signal, statistical descriptors such as mean, standard deviation, maximum, minimum, and rate of change are computed, yielding a compact feature vector that reflects both local and system‑wide stability information.

The core predictive engine is an ELM, a single‑hidden‑layer feed‑forward neural network whose hidden‑layer weights are randomly assigned and only the output weights are solved analytically via a least‑squares solution. This architecture provides extremely fast training, which is essential for online applications, but its performance is highly sensitive to two hyper‑parameters: the number of hidden neurons (N) and the regularization coefficient (C). To avoid manual trial‑and‑error, the authors embed these parameters in an IPSO search. IPSO extends the classic PSO by (i) employing a non‑linear inertia‑weight decay to balance exploration and exploitation, (ii) adapting the cognitive and social learning rates based on fitness improvement, and (iii) preserving population diversity through a mutation‑like perturbation when stagnation is detected. These enhancements enable the swarm to locate the global optimum of the (N, C) space with fewer iterations and reduced risk of premature convergence.

The methodology is validated on two testbeds. The first is the IEEE 39‑bus system, where a comprehensive set of fault scenarios (single‑line‑to‑ground, three‑phase short‑circuit, line outage) and varying load levels are simulated to generate 2 000 training and 500 testing samples. The second is a real‑world large‑scale network comprising roughly 5 000 buses, for which 800 recorded events with synchronized PMU data are used as an independent test set. Comparative experiments involve three baselines: a standard Support Vector Machine (SVM), a Decision Tree (DT), and an ELM whose hyper‑parameters are tuned by the conventional PSO. Performance metrics include accuracy, precision, recall, F1‑score, and prediction latency (time from data acquisition to output).

Results show that the IPSO‑ELM achieves an overall classification accuracy of 98.7 %, markedly higher than PSO‑ELM (96.3 %), SVM (92.5 %), and DT (89.8 %). The recall for rare but severe large‑voltage‑dip events exceeds 96 %, indicating robust detection of critical disturbances. Prediction latency averages 22 ms, well within the sub‑second window required for protective relaying and automatic generation control. Sensitivity analysis reveals that the optimal hidden‑neuron count lies between 150 and 200, while the regularization coefficient C is most effective in the range 0.1–1.0. IPSO converges to the best solution in fewer than 30 % of the maximum allowed iterations, demonstrating computational efficiency.

The paper’s contributions are threefold. First, it demonstrates that PMU‑derived features can substantially improve early‑stage stability assessment compared with traditional SCADA‑based inputs. Second, it shows that the ultra‑fast learning capability of ELM can be retained while eliminating manual hyper‑parameter selection through an advanced swarm‑intelligence optimizer. Third, it validates the approach on both a benchmark system and a real‑world large‑scale grid, confirming scalability and practical relevance.

Nevertheless, the study acknowledges several limitations. PMU data are prone to communication delays, packet loss, and measurement noise; the robustness of the proposed feature set against such imperfections is not thoroughly examined. The IPSO algorithm itself has meta‑parameters (population size, maximum iterations) that are fixed in the experiments, potentially requiring retuning for different networks. Finally, the current framework focuses on very short‑term dynamics (≤0.5 s); extending the prediction horizon to several seconds would likely require integration with temporal models such as LSTM or Kalman filters.

In conclusion, the authors deliver a compelling, high‑performance TSP solution that balances prediction accuracy, speed, and ease of deployment. Future work could explore noise‑resilient feature engineering, automated meta‑parameter adaptation for IPSO, and hybrid architectures that combine the rapid decision‑making of ELM with the long‑term forecasting ability of recurrent neural networks, thereby advancing the readiness of real‑time stability assessment tools for modern smart grids.