Deep Multi-Modal Fusion and Kalman Online Adaptation for Electric Load Forecasting Against Extreme Rainfall Events

Extreme rainfall (ERF), one of the most frequent and widespread severe weather hazards, has significant impacts on power systems. The sudden and unpredictable nature of precipitation can cause abrupt changes in power consumption behavior, posing significant challenges for accurate electric load forecasting. Satellite images provide fine-grained spatial characteristics of weather phenomena, which can be leveraged to capture rainfall patterns and trajectories that influence power demand. To this end, this paper proposes a multimodal learning and online adaptation-based load forecasting method for ERF disasters by taking both satellite images and time series data as input. Instead of relying on convolutional neural networks or Transformers, a novel lightweight MLP-Mixer architecture is devised. It leverages three types of Mixer operations—time-mixing, token-mixing, and channel-mixing—to process sequential satellite images, which allows the proposed model to capture both local rainfall details and broader weather dynamics while ensuring lower computational costs. Following this, a gated multimodal unit is devised to adaptively fuse multimodal representations of images and numerical features by learning to decide how modalities influence the activation of the unit. To enable the model to rapidly adapt to load pattern changes caused by ERF events, a modified extended Kalman filter-based online adaptation algorithm is further developed to dynamically update model parameters using newly arrived load data. Extensive experiments on distribution network load data from ERF events show that the proposed method achieves superior performance compared to state-of-the-art load forecasting methods.