Existing harmonic source modeling methods struggle to reconcile the frequency-domain steady-state feature and dynamic time-varying feature of harmonic sources when applied to scenarios with unknown internal topologies and unclear mechanisms

hindering the accuracy and robustness of the harmonic models. To address this

a data-driven harmonic source modeling method based on joint time-frequency feature extraction is proposed in this paper. First

the dominant harmonic orders of the harmonic sources are determined based on the frequency domain components of historical voltage and current data. Then

a steady-state voltage time-frequency feature matrix is constructed

and pseudo-color coding is applied to achieve dimensionality expansion. Finally

the dynamic time-varying signals and steady-state feature component color maps are combined and fed into a data-driven harmonic model consisting of a multiple convolutional neural network (MCNN) and a bi-directional long short term memory network (BiLSTM) to establish the V-I mapping that captures both the dynamic and steady-state characteristics of the harmonic source in both time and frequency domains. Simulation results and field data verify that the proposed method not only demonstrates significant superiority in single harmonic source modeling but also achieves high accuracy and sharp robustness under complex multi-harmonic source conditions compared with other data-driven modeling methods.