Intelligent Identification of Dominant Error-Causing Parameters for Dynamic Simulation Models of Power Systems Based on Time–Frequency Feature Learning of Response Errors

INTRODUCTION: Dynamic simulation is important for disturbance analysis and security assessment in power systems. Parameter deviations can cause discrepancies between simulated and measured responses, and multiple deviations may produce complex transient response errors. OBJECTIVES: This paper aims to characterize the error-causing effects of candidate parameters from response error patterns and identify priority parameters for subsequent model parameter calibration. It is intended as a pre-calibration parameter screening method rather than direct parameter correction. METHODS: An intelligent identification method for dominant error-causing parameters is proposed based on time–frequency feature learning of response errors. Continuous wavelet transform is employed to extract local time–frequency features from short-term post-fault response errors and construct multi-channel time–frequency feature tensors. Parameter error-causing contribution labels are constructed by jointly considering parameter deviation magnitude and local response sensitivity. A convolutional neural network-based multi-output regression model is then trained using the feature tensors and labels to predict and rank candidate parameters by contribution score. RESULTS: Case studies based on a modified IEEE 39-bus system show that the proposed method can identify a dominant error-causing parameter set more consistent with current response error patterns under multiple-parameter deviation conditions. Compared with conventional sensitivity-based parameter selection, the proposed method achieves smaller post-calibration response errors. When used for subsequent calibration, the selected subset reduced voltage and active-power errors by 85.34% and 90.09%, respectively. CONCLUSION: The proposed method provides a more targeted candidate parameter set for dynamic simulation model parameter calibration and supports more efficient subsequent model refinement.