Transient Stability Assessment Method for Power System Based on SVM with Adaptive Parameters Adjustment

Nowadays machine learning and deep learning methods have been widely used in transient stability assessment for power system. Some simple machine learning models have strong applicability for transient stability evaluation due to their high efficiency, but these methods often suffer from insufficient accuracy. Although the deep learning models can achieve higher precision, the offline training usually costs a lot of time. In order to improve the evaluation accuracy of the model and shorten its time of searching parameters. This paper proposes a parameter searching method based on Bayesian optimization that integrates verification curves and RFE (recursive feature elimination) to optimize the SVM (support vector machine). The global search is replaced by the directed local search strategy, which can reduce the model training time and achieve the global optimum. On the one hand, we combine the verification curve with the RFE based on random forest. By this method, we can determine the Bayesian optimized area. Then we can achieve the optimal parameters. On the other hand, the prior function in Bayesian optimization is used to simulate the distribution of hyperparameters. At the same time, the acquisition function is used to determine the best search point, thereby greatly reducing the time cost of offline training. The simulation results on New England 10-machine 39-bus system show that the offline training time of the SVM based on the improved Bayesian optimization is significantly shortened and its evaluation accuracy can be upgraded to a large extend.