To improve the accuracy of short-term power load forecasting, this study proposes a model for short-term power load forecasting using multi-scale features. First, feature selection is performed on the power load data using the Pearson correlation coefficient and the maximal information coefficient, followed by dynamic spectral filtering to remove high-noise components. Second, the data are fed into a multi-scale temporal convolution module for multi-scale temporal feature extraction. Third, correlations in the data are mined, and features are encoded using a variable-dimensional multi-head attention mechanism combined with feed-forward neural networks. Finally, a linear hierarchical fusion module is developed to process the encoded features in parallel with a projection layer. The weighted combination of these two components yields prediction results through the deep fusion of data features. Experimental results on Dataset 1 show that the proposed model achieves good results in both comparative and ablation experiments, outperforming the next-best model by 5.2% and 6.8% in MAE and RMSE, respectively. On Dataset 2, the proposed model surpasses baseline models across all evaluation metrics, achieving an MAPE of 0.467 and an R² of 0.949. The proposed model enhances short-term power load forecasting accuracy and exhibits strong generalization capabilities.