Air-conditioning load forecasting is important for building energy scheduling and control. The task is difficult because load data show clear periodic patterns, and sudden changes caused by events such as holidays or heat alerts also appear. These patterns exist together in non-stationary time series, which makes accurate prediction harder. This paper proposes an Event-aware Temporal Fusion Network (ETFN). The model learns long-term periodic patterns and short-term local changes at the same time, and it combines them through an event-aware gating mechanism so that the model can adjust feature importance when conditions change. The model uses air-conditioning operation data, environmental data, time information, and event indicators as inputs. Experiments use hourly VRV air-conditioning data from four enterprises in Nanjing. The results show that ETFN achieves lower prediction errors than common baseline models and keeps stable performance under different operating conditions, including periods with sudden load changes and data disturbances.
Fang et al. (Tue,) studied this question.