Deep Multi-Task Learning for Energy Consumption Forecasting of Household Water Heater Usage
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
With electricity market deregulation and the diversification of electricity consumption patterns, precise forecasting of electric demand is crucial for optimal power generation planning and load balancing. However, the increasing complexity of the usage patterns makes the accurate appliance-level prediction more difficult. Water heaters, for example, exhibit distinctive consumption patterns: high electricity usage during active heating periods alternates with complete inactivity (0W), resulting in zero-inflated and high-variance distributions, which are difficult to handle by conventional prediction methods. This study therefore proposes a deep multi-task learning (MTL) framework that simultaneously predicts binary classification (consumed/not consumed) and regression (the amount of consumption) to hierarchically correct the prediction errors for household water heater electricity prediction. We adopt Uncertainty Weighting (UW) which dynamically optimizes heterogeneous loss functions. Using IoT logs with 180,805 samples collected from water heaters in 787 households from January 2022 to May 2024, we demonstrate significant performance improvements over major baselines in RMSE reduction:$-22.492\%$(single-task MLP), 3.962% (Transformer), and$-4.223 \%$(LSTM). In addition, the proposed model showed strong generalizability compared to timeseries Transformer (PatchTST).
Fields of Study & Subjects
Energy Load and Power Forecasting
Smart Grid Energy Management
Building Energy and Comfort Optimization
Publication details
- Journal / source
- International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
- Pages
- 1–6
- Publication date
- 24 February 2026
- Language
- EN