Multi-channel Deep Convolutional Networks With Shared Weight Layers For Power System Transient Stability Ssessment

With the gradual popularization of renewable energy supply, the further increase of power load and the extensive application of power electronic equipment, the network structure and operating mechanism of power system has become more and more complex, and its safe and stable operation is facing severe challenges. In recent years, there have been many large-scale blackouts at home and abroad, resulting in serious economic and social impact, but also exposed the various deficiencies of the existing grid stability analysis system in the operating mechanism and timeliness. There is still a huge room for improvement.In this paper, an architecture of deep convolutional network for power system transient stability assessment is proposed, which is inspired by deep convolutional neural network used in the feature learning studies.The method can automatically obtain the characteristic information of different abstraction levels from the voltage and phase angle raw data collected by the PMU, and therefore does not need to deal with the complex manual feature extraction process. In particular, we propose an efficient deep convolution network model for multidimensional multichannel time series. Each channel receives a two-dimensional time series, which represents the complex voltage timing information of a single bus in the power network (from the bus voltage and the phase angle data are transformed by simple coordinates). The data from each channel are then processed by the weight sharing deep convolutional networks, and the hidden feature information is obtained. The model then inputs this information into a multi-layer perceptron (MLP) to perform a stability prediction. We use the model-based gradient method to determine the network parameters, and then based on IEEE39 system simulation scenarios for performance verification. The results indicate that our model has a good effect in both predicting speed and precision.In addition, in order to take advantage of the large amount of untagged real data existing in the real power grid system, this paper proposes an unsupervised pre-training method to improve the accuracy of the model. This paper also provides a visual representation of the learned features by visualizing the convolution kernels.