A novel data-driven approach for transient stability prediction of power systems considering the operational variability

Data driven methods are playing an increasingly important role in transient stability assessment, primarily because of the availability of large annotated datasets. Nevertheless, training data cannot cover all the possible operating conditions of a modem power system with variable power generations and loads. The classifier should adjust to the near-future operation condition in limited time, and this adjustment may be hindered by the computational time of the simulations and classifier training. To dramatically reduce the computational cost, this paper presents a systematic approach for building and updating an accurate transient stability classifier. First, the time-series trajectories of generators after disturbance are used as the inputs, and then a convolutional neural network (CNN) ensemble method is proposed to generate the transient stability predictor using these multi-dimensional data. To reduce the misclassification of instability, different cost weights are considered for the stable and unstable instances in the loss function. When the operating condition changes substantially and makes the pre-trained classifier unavailable, the active learning and fine-tuning techniques are integrated to update the classifier with good performance using fewer labelled instances and short computational time. The simulation results of two power systems illustrate the effectiveness of the proposed approach.