Global population growth and industrial expansion boost electricity demand, prompting power system infrastructure expansion. But this makes power system management and maintenance more complex, highlighting the importance of power load time series prediction and consumption anomaly detection. However, power load time series data have issues like long term dependence, feature nonlinearity, and noise interference. Traditional single deep learning models struggle to capture multi dimensional and long sequence information and extract global and local features. This paper proposes a TCN-Informer combined model for power load forecasting. Which integrates TCN's feature extraction ability and Informer's ability to capture long term dependencies by its attention mechanism, significantly enhancing prediction accuracy. First, two years electricity consumption data which from the Southern Power Grid are collected, with missing values filled by linear interpolation and data normalized by the Min-Max method. Second, the model is constructed, where TCN extracts short term features and Informer captures long term dependencies. Finally, experiments show this model outperforms others in MAE, RMSE, and MAPE, improving prediction accuracy and supporting power system management.