A zero-shot load forecasting method for extreme weather integrating causal learning and meta-learning

The absence of historical load data for extreme weather events makes forecasting future extreme weather loads quite challenging. Additionally, extreme weather often leads to significant peak loads and increased peak-valley differences, further complicating accurate forecasting. To address these issues, this paper proposes a zero-shot load forecasting method for extreme weather that integrates causal learning and meta-learning. Firstly, the causal graph between meteorological features and loads is established through the structural causal model (SCM). Based on this, causal interventions are made on key variables based on meteorological domain knowledge to generate reasonable virtual extreme weather scenarios, thus solving the problem of absent extreme weather training data. Secondly, a paradigm shift in the learning paradigm is achieved by employing model-agnostic meta-learning (MAML) with a two-layer optimization mechanism. Instead of focusing on learning specific extreme weather patterns, the proposed model acquires meta-knowledge, enabling it to rapidly adapt to novel distributions. Finally, the Transformer is introduced to capture global dependence under extreme weather through its self-attention mechanism, which ensures extreme scenario forecast robustness in extreme weather scenarios. The proposed method breaks through the reliance of traditional methods on historical data and realizes accurate zero-shot extreme weather load forecasting. Compared with other advanced forecasting methods, the proposed method reduces the MAE and RMSE by 34.03 %–85.17 % and 33.55 %–84.28 %, demonstrating its excellent performance under zero-shot extreme weather load forecasting.