ACOLM: Adaptive contrastive online learning model for urban extreme weather load forecasting

As global climate change intensifies, the increasing frequency of extreme weather events poses unprecedented challenges for urban power system load forecasting. Traditional forecasting methods and existing deep learning approaches often fail when faced with concept drift caused by extreme weather and are limited by the scarcity of extreme weather data, making it difficult to adapt in real time to drastic changes in load patterns. To address these issues, this paper proposes an adaptive contrastive online learning model (ACOLM) for online load forecasting under extreme weather conditions in urban areas. First, ACOLM employs radian scaling instead of traditional normalization methods, providing the model with unbounded forecasting capabilities and addressing the systematic underestimation caused by concept drift during extreme weather. Second, the model integrates a bidirectional LSTM (BiLSTM) encoder, a multi-head attention mechanism, and a contrastive learning framework. Through a dual-path design with projection heads and prediction heads, it can extract robust temporal feature representations while achieving precise load forecasting. Contrastive learning mitigates the scarcity of extreme weather data by constructing positive and negative sample pairs based on time-series characteristics. Finally, the Z-Score-based concept drift detection mechanism can identify extreme weather patterns in real time and achieve rapid adaptation through dynamic adjustment of learning rates and temperature parameters. Experimental results demonstrate that ACOLM outperforms other comparative models in extreme weather load forecasting.