Representational learning approach for power system transient stability assessment based on convolutional neural network

The transient stability assessment (TSA) problem can be mapped into a two-class classification problem in machine learning, which estimates the dynamic security boundary of the power system by learning from large amount samples. A representational learning approach is proposed to solve the problem based on big data collected from Phasor Measurement Units (PMUs), which includes four stages: (i) Construct original input features by using PMUs data to describe the dynamic characteristics of the power system. (ii) Unsupervised representational feature learning by using the original features. Stacked autoencoders (SAEs) perform representational learning for crucial features. (iii) Supervised classifier training. A powerful deep learning model, convolutional neural network, which is added to SAE, is trained and tested with the learned representation. (iv) Online application, the trained model is applied to the online evaluation for TSA. Simulation on the New England 39-bus test system shows that the proposed approach has high accuracy, rare misclassification of the unstable sample and excellent robustness with noise in PMUs for TSA.