Research map / Probabilistic and scientific modeling / Dynamical systems

NeuralBES: A Differentiable, Control-Aware Emulator for Scalable Building Energy Modeling

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent-space control and planning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NeuralBES, a differentiable two-node RC building-energy emulator whose coefficients come from a shared metadata encoder, solved via a log-space parallel scan with a predictor-corrector thermostat loop, evaluated on ResStock across CA/TX/NY.

Problem

Scalable, physically valid prediction of residential building thermal loads and indoor temperature across heterogeneous buildings, where classical simulators need per-building calibration and black-box sequence models produce physically inconsistent trajectories.

Approach

A shared neural encoder maps static building metadata to bounded RC parameters that define the decay coefficients of an exact discrete linear recurrence solved by a log-space parallel scan, with a short predictor-corrector loop closing the thermostat saturation nonlinearity.

Key results

• NeuralBES achieves 100% mode exclusivity and 99.7% bounds adherence on the annual rollout, while LSTM (36.6M params) has lower raw HVAC MAPE (38.5%/37.6%) but only 36.9% mode exclusivity.
• At parameter parity, NeuralBES reaches 41.5% and 41.3% HVAC MAPE versus 86.5% and 92.5% for the grey-box RC baseline.
• NeuralBES uses about 1.0M parameters, roughly 36x fewer than LSTM and about 20x fewer than Roformer.

Surprising findings

• PatchTST returns a flat zero response to a uniform outdoor-temperature perturbation, i.e. its indoor-temperature prediction is insensitive to outdoor temperature.
• PI-Roformer's mode-exclusivity ranking inverts between annual (31.0%) and weekly (2.75%) rollouts, suggesting physics heads on black-box trunks do not help short-window switching.

Limitations

• No ~1M-parameter LSTM/Roformer baseline row, so capacity versus inductive bias for the 3-4 pp MAPE gap is untested.
• Sensible-only two-node model omits latent loads and domestic hot water, which matters for humid-climate cooling.
• Trained on 672-step windows yet evaluated on 35,040-step annual rollouts (out-of-distribution horizon).

Experimental setup

Datasets: ResStock

Benchmarks: ResStock test set (annual 35,040-step rollout) ResStock weekly 672-step rollout

Metrics: MAPE RMSE mode exclusivity bounds adherence

Baselines: DLinear LSTM PatchTST Roformer PI-Roformer RC Grey-Box

Builds on: S4 Mamba linear recurrent units DOE-2 EnergyPlus

Closest papers by meaning

• Building Timeseries Dataset: Empowering Large-Scale Building Analytics (NeurIPS 2024 · similarity 0.82)
• SysCaps: Language Interfaces for Simulation Surrogates of Complex Systems (ICLR 2025 · similarity 0.82)
• The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization (NeurIPS 2025 · similarity 0.81)
• Active Reinforcement Learning for Robust Building Control (AAAI 2024 · similarity 0.81)
• CE-NAS: An End-to-End Carbon-Efficient Neural Architecture Search Framework (NeurIPS 2024 · similarity 0.80)
• Data Center Cooling System Optimization Using Offline Reinforcement Learning (ICLR 2025 · similarity 0.80)