SSMCF: A self-supervised masked contrastive framework for few-shot short term electric load forecasting

Accurate short-term load forecasting (STLF) is crucial for optimizing power supply–demand balance and energy efficiency. However, existing methods rely heavily on massive labeled data, with limited transferability and robustness in load mutations, transient energy and new scenarios, challenging smart grid dispatching. To solve this, a self-supervised masked contrastive framework (SSMCF) is proposed for accurate STLF under few-shot labels. SSMCF is built on three key innovations: (1) A masked time-series modeling (MTM) is proposed to learn local load features, (2) A pre-training framework unifying MTM and contrastive learning (CL) is constructed for self-supervised learning under few-shot, (3) An LSTM–KAN–Transformer hybrid backbone with dynamic fusion module (DFM) is proposed for temporal feature extraction. Experiments on seven public datasets and one utility-grade load forecasting scenario show SSMCF outperforms thirteen methods, with higher accuracy and generalization, proving it a powerful tool for few-shot STLF. The code and dataset is available at https://github.com/wink798/SSMCF.git .