Does AI-based building energy optimisation deliver net energy savings? Code, hardware energy-measurement logs and results

Abstract

Code, per-job hardware energy-measurement logs, classical-controller results, derived tables and figure scripts for the study "Does AI-based building energy optimisation deliver net energy savings? A measured life-cycle assessment of computational energy against operational benefits across 2,000+ building–climate–algorithm configurations" (research article, 2026; the reference will be added once published). The archive contains:(1) the energy-measurement module (Intel RAPL via Windows performance counters, NVIDIA NVML power integration, CPU-time-share apportioning of package energy to processes, calibrated against isolated runs);(2) a vectorised 2R2C gray-box building testbed with six archetypes and 23 climates, rule-based, model-predictive (cvxpy/Clarabel) and deep reinforcement learning controllers (PPO, SAC, DQN; Stable-Baselines3), and the experiment grids: 1,525 measured DRL training runs and 1,380 classical building-years;(3) battery/PV dispatch experiments on the 17 real homes of the CityLearn Challenge 2022 (LP, receding-horizon LP, DRL, shared fleet policy);(4) a day-ahead load-forecasting zoo on 1,246 real meters of the Building Data Genome Project 2 (persistence, ridge, LightGBM, per-building and global MLP/GRU/Transformer) with forecast-driven dispatch;(5) EnergyPlus 26.2 simulation-energy measurements and EnergyPlus-in-the-loop RL training;(6) a Monte-Carlo (20,000 draws) and Sobol life-cycle model for five deployment architectures, fleet sizes and shared policies, with break-even savings by building archetype;(7) every derived table and figure, one JSON record per training run (arguments, hardware, raw counter readings, apportioned energy, in-sample and out-of-sample evaluation, inference energy, thermostat baselines). Reproduction: see README.md (pip install -r requirements.txt; the analysis and figure scripts regenerate every table and figure from the included records in minutes; the full experiment pipeline takes about three days on a 24-thread laptop). Hardware and measurement details are in ENVIRONMENT.md. Licences: code MIT; data, logs, tables and figures CC BY 4.0. Third-party inputs keep their original terms (climate.onebuilding.org weather files, CityLearn 2022 dataset [MIT], Building Data Genome 2 [MIT], Ember/Our World in Data carbon intensities [CC BY 4.0]); see LICENSE-DATA.md. Not included (regenerable with the scripts): trained model checkpoints (1.0 GB), Tier-3 forecast arrays (0.66 GB), EnergyPlus binary outputs (0.25 GB), the EnergyPlus installer and the 167 MB BDG2 raw meter file (download links in README.md).