Objective In response to the problem of low prediction accuracy of traditional power load forecasting models for long sequences a novel temporal convolutional neural network TCN model called STE-TCN combining skip-level convolutional connections and time encoding networks was proposed. Methods Firstly skip-convolution channels across periods were added to the TCN model to extract cycle information from power data. The Skip-TCN network was obtained by fusing the features so that the network captures the cycle pattern and increases the length of information utilization. Finally the time encoding network was designed to capture the life cycle and seasonal features and the STE-TCN model was obtained by fusing the features with the Skip-TCN. The long sequence prediction of power load data was realized. Results Experimental results show that compared with the TCN model and traditional sequential networks Skip-TCN exhibited improved prediction accuracy especially in longer prediction tests. Conclusion Experimental results validate that by capturing longer-spanned temporal relationships the improved method of STE-TCN network effectively enhances the prediction accuracy of long sequence power load data.