Due to non-stationary noise, the low-voltage power line communication (LPCC) encounters significant challenges in smart grid applications. Conventional denoising techniques, such as wavelet thresholding and adaptive filtering, exhibit limited performance in complex industrial environments, while emerging deep learning models often suffer from insufficient real-time capability. In response to the noise characteristics of low-voltage power line channels and the limitations of traditional impedance matching algorithms, we propose a hybrid CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) architecture. A dual-branch feature fusion mechanism is introduced, which employs parallel processing of time-frequency features via STFT+WVD (Short-Time Fourier Transform+Wigner-Ville Distribution) to enhance noise identification accuracy. A dynamic impedance matching module, optimized in real time using a deep reinforcement learning (DRL)-based gradient descent algorithm, is developed to overcome the poor adaptability of conventional fixed-parameter approaches. Furthermore, a joint noise suppression and signal reconstruction framework is designed to effectively preserve useful signal components while suppressing noise. The experimental results verify that the proposed model achieves an SNR (Signal-to-Noise Ratio) improvement of up to 18.2 dB, outperforming the conventional DnCNN (Denoising Convolutional Neural Network) method by 16.7 %. Under harmonic interference conditions (THD = 15 %), the waveform distortion rate is only 2.1 %, with latency optimized to 9.2 ms, meeting real-time requirements. By incorporating multi-scale feature fusion and dynamic gating mechanisms, the model effectively mitigates mixed interference composed of switching impulse noise and additive white Gaussian noise, which will offer a viable solution for enhancing the reliability of LPCC systems.
Zhu et al. (Sat,) studied this question.