Develop of a Context-Aware Deep Learning Model for Demand Response Reduction Prediction Based on AMI Data

Accurate forecasting of load reductions during Demand Response (DR) events is essential for effective grid operation and settlement. This paper presents a context-aware deep learning model for predicting DR reductions using Advanced Metering Infrastructure (AMI) data, enhanced with contextual features such as weather, time, customer attributes, and DR history. The proposed model integrates a time-series encoder with multi-attention mechanisms, enabling dynamic weighting of input variables based on event-specific context. Experiments conducted on real-world AMI datasets demonstrate that the model outperforms baseline methods including linear regression, LSTM, and XGBoost in terms of MAE and RMSE. The results show improved prediction accuracy across customer types and DR scenarios. This approach enhances the interpretability and reliability of DR forecasting and provides a foundation for adaptive, context-sensitive DR management strategies. The model also supports applications in pre-settlement estimation and real-time DR operations.