Parallel Fusion CNN Enabled Data-Driven Nonlinear Model on Time-Frequency Characterization for Power Quality Analysis

Modeling of nonlinear loads is crucial for analyzing and evaluating power quality in modern power system. To further enhance both accuracy and robustness, a novel data-driven nonlinear load model based on time-frequency characterization and the parallel fusion convolutional neural network (PFCNN) is proposed. Specifically, the 1-D voltage signal is fed into the 1-D CNN to extract time-varying feature of nonlinear loads. Also, the fast Fourier transform (FFT) and signal reconstruction method are implemented to transform the 1-D signal into the 2-D color map, which is then processed by the 2-D CNN to extract steady-state frequency feature. The 1-D and 2-D CNNs operate in a fully parallel manner and the above two complementary features are combined by the parallel fusion layer composed of the fully connected layer (FC) and Bi-LSTM. The distinct parameter arrays in the PFCNN are update independently by triple separate optimizers with the modified loss function. By interacting with limited samples, the data-driven model learns the relationship between the input voltage and output current of nonlinear loads, effectively approximating the dynamic admittance with more complex and comprehensive time-frequency features. Compared to other data-driven modeling methods, the test results demonstrate the proposed model offers high accuracy and robust performance.