Transient Voltage Stability Assessment of Power System Based on Improved Deep Residual Network

In traditional model to assess power system transient voltage stability there are defects in two aspects, i.e., difficult to capture the key information during the failure process and the imbalance between transient stability samples and unstability samples leads to the tendentiousness of the model for majority class samples. For this reason, a voltage stability forewarning model based on improved deep residual network was proposed. Firstly, to capture the key information during the failure process, a convolutional attention module was embedded in the residual network, and through the dual attention of the time channel and the space channel the potential spatiotemporal relationship in the dynamic trajectory of the power system was dug. Secondly, in allusion to problem that during the training process the model was tended to majority class samples, the gradient harmonizing mechanism-based loss function was led in to reduce the influence of imbalance samples on assessment results. Thirdly, to intensify the model’s ability of extracting data features, traditional convolution kernel was replaced with asymmetric convolution block. Finally, by means of connecting two different wind power ratios to IEEE 39-bus system, the performance of the proposed method in the assessment on transient voltage stability was verified.