EVLA-Energy: A Physics-Grounded Vision–Language–Action Dataset for Residential Energy Control

EVLA-Energy is a physics-grounded multimodal dataset designed to support research in vision–language–action (VLA) models for residential energy management and demand response. The dataset integrates visual load representations, natural language instructions, physical state variables, and Pareto-consistent control trajectories generated by a model predictive control (MPC) oracle. EVLA-Energy comprises 439,961 samples spanning 19 residential buildings, with strict household-level train–test separation to support zero-shot generalization studies. Each sample includes:(i) a 224×224 log-scaled power field encoding aggregated electrical demand,(ii) a natural language control instruction expressing operational intent (e.g., cost prioritization or comfort preservation),(iii) a four-dimensional physical state vector (price, mean power, battery state-of-charge, indoor temperature), and(iv) a short-horizon (16-step) action sequence corresponding to battery and HVAC control. The dataset is derived from real-world residential electricity measurements collected in the UK REFIT Smart Meter Electrical Load Dataset, which provides high-frequency (8-second) household load data over multi-year periods. EVLA-Energy does not redistribute raw REFIT measurements. Instead, REFIT data are transformed through a physics-consistent pipeline that includes log-power field construction, state extraction, and MPC-based trajectory generation. The resulting representations are task-specific, temporally aggregated, and transformed through nonlinear log-scaling and MPC supervision, and therefore cannot be inverted to recover the original smart meter signals.Original data source: Murray et al., “An electrical load measurements dataset of United Kingdom households,” Scientific Data (2015). Price variability and demand fluctuations in EVLA-Energy act as operational proxies for renewable intermittency (e.g., solar and wind variability), enabling evaluation of control policies under conditions characteristic of high-renewable-penetration energy systems without explicitly modeling generation assets at the household level. EVLA-Energy is intended for research in physics-aware multimodal learning, demand response, human-aligned energy control, and intelligent energy systems. Access is restricted during peer review and will be made publicly available upon acceptance of the associated manuscript.