Dynamic event monitoring using unsupervised feature learning towards smart grid big data

In this paper, a novel framework for dynamic event monitoring in smart grid with synchrophasor data is proposed. The fundamental principle is that the nonstationary signatures of any dynamic event in the system will be captured by the data set collected from the phase measurement units (PMUs). The framework is performed in three steps: 1) Energy function components construction by using PMU data. Each of the component in the energy function, such as the load, transmission line, and generation, are formulated explicitly; 2) Unsupervised feature learning and fusion by using the energy function components. In the feature learning stage, stacked autoencoders (SAE) is used to learn representative features from each component. Then in the feature fusion stage, a shared representation between modalities is further learned; and 3) Supervised classifier training and online application. Classifier, such as a simple neural network, is trained for monitoring the system dynamics to detect and classify events. Compared with purely data-driven methods, the proposed framework utilizes the physical basis to understand and correlate the features with different events. Meanwhile, because of learning features automatically and adaptively, the proposed method reduces the need of human labor and can be more easy to process big data in smart grid. Simulation is carried out on IEEE 39-bus system, and the results show promising effectiveness on dynamic event detection and classification.