To fully explore the complex temporal relationships among the inherent multi-scale features (MSF) of electrical load data and further improve the performance of electricity load forecasting models, especially their accuracy during holidays, a short-term electricity load forecasting model based on multi-scale secondary feature extraction is proposed. First, the Prophet algorithm is used to decompose and fit the load data, extracting components at different scales, which are then combined with correlated weather data to construct a multivariant dataset. Then, an improved feature pyramid network (IFPN) is employed to match the multi-scale characteristics of load data. A convolutional feature enhancement module is designed to strengthen the model’s ability to express holiday-specific features, achieving the first extraction of MSF. Leveraging the advantages of temporal convolutional neural networks, the model deeply mines the temporal dependencies among the primary features. Squeeze-and-excitation networks (SENet) is introduced to adaptively assign weights to features, completing the secondary extraction of MSF. Finally, load forecasting is performed using a Transformer model optimized by the Osprey algorithm. Validation on two domestic and international load datasets shows that the proposed model outperforms comparison models, particularly in improving prediction accuracy during holidays.