Abstract: To address the dependence of physics-model-driven methods on prior mechanism knowledge, as well as the low training efficiency and high computational burden of typical recurrent neural networks in modeling complex non-stationary degradation signals, this paper proposes a lightweight multi-scale temporal convolutional network model for degradation prediction of switching power supplies. Built upon the temporal convolutional network architecture, the proposed model employs multi-scale dilated convolutions to jointly model short-term fluctuation features and long-term degradation trends in ripple signals. Dropout regularization, Huber loss, and an early stopping mechanism are further introduced to improve the generalization ability and training efficiency of the model under small-sample and strong-noise conditions. Based on the established capacitor degradation experimental platform, ripple data of switching power supplies under multiple stress conditions, including normal, overvoltage, and high-temperature conditions, are collected to validate the proposed LMS-TCN model. Experimental results show that the proposed LMS-TCN model achieves clear advantages in prediction accuracy and convergence efficiency, and can well meet the requirements of online real-time degradation prediction of switching power supplies, providing an efficient and feasible solution for their life assessment and predictive maintenance.