To fully mine the multi-scale time series information contained in power load data and improve short-term power load prediction accuracy, an improved temporal convolutional neural network with multi-scale feature enhancement (ECA-MS-DHTCN) model is proposed. First, the load data features are extracted using causal convolutions with four convolution kernels of different sizes, and an efficient channel attention (ECA) module is embedded in the feature extraction layer to achieve local cross-channel interactions without dimensional reduction. It obtains multi-scale loading features with channel attention. Then, the basic TCN residual block structure is improved using double-hybrid dilated convolutional layers to overcome the problems of information omission and long-distance information irrelevance in the dilated convolution structure of the TCN model. It also takes into account the shallow details and deep connections of the load characteristics. Finally, the ECA-MS-DHTCN load forecasting framework is built by combining the ECA-optimized multi-scale feature extraction module with the improved TCN model to complete the short-term load forecasting task. Through the simulation of actual power grid load data, the results show that the ECA-MS-DHTCN model proposed can effectively improve prediction accuracy while maintaining fast training.