Identifying power load curves is essential to ensure the safety and energy efficiency of the power grid. However, existing algorithms for power load curve identification tasks often suffer from issues such as low recognition accuracy and robustness. To tackle these issues, a multi-model fusion ensemble learning method for power grid load curve recognition is proposed. Temporal convolutional network （TCN）, transformer, and light GBM models are adaptively improved to predict load curve categories, considering three dimensions：local,global and structural features. Then, predictions through stacking ensemble learning（EL）to refine overall accuracy are adaptively fused.Additionally, a truncated Gaussian distribution（TGD）data augmentation strategy is introduced, which models intra-class signal fluctuations to alleviate data category imbalances, thereby enhancing the robustness of the recognition model. Through simulation analysis, compared with methods such as XG Boost, LSTM, and MLP, this approach shows a significant improvement in power load classification accuracy.