面向异常事件的在线时空集成学习负荷预测方法

This letter proposes a novel online spatiotemporal ensemble learning framework that can rapidly adapt to load pattern changes caused by abnormal events. Unlike existing online learning approaches that focus solely on temporal dependencies, the proposed method also exploits spatial correlations across different regions to achieve fast convergence. An online complementary learning network that can instantly adapt to new patterns while recalling similar historical knowledge is first built as the basic forecast expert to extract spatial and temporal features. The two information streams are then combined using an online convex programming framework, which is further solved by exponentiated gradient descent and reinforcement learning methods. Experiments on real-world electricity load datasets from the COVID-19 period demonstrate the proposed method's effectiveness.