Evaluating the significance of samples in deep learning-based transient stability assessment

Deep learning-based transient stability assessment has achieved big success in power system analyses. However, it is still unclear how much of the data superfluous and which samples are important for training. In this work, we introduce latest technique from artificial intelligence community to evaluate significance used deep learning model assessment. From empirical experiments, found that nearly 50% low-significance can be pruned without affecting testing performance at early training stages, thus saving computational time effort. We also observe with fault-clearing close critical clearing often have higher indexes, indicating decision boundary learned by network highly related boundary. This intuitive, but best our knowledge, work first analyze connection sample aspects. addition, combine scores index provide an auxiliary criterion degree stability, distance between a The ultimate goal study create tool generate some benchmark datasets analysis, so various algorithms tested unified standard platform like computer vision or natural language-processing fields.