PaperlayerView exact revision on arXiv

Source-linked AI summary

Very Exciting: Zero-Shot Model Predictive Control of Buildings via Excitation-Based Generalized Transfer Learning Models

Fabian Raisch, Felix Koch, Zack Xuereb Conti, Christoph Goebel, Benjamin Tischler

arXiv:2609.12853v2eess.SYcs.LGcs.SY

TL;DR

Building MPC is difficult to scale when each building requires substantial data collection and model training, and prior transfer-learning evaluations often measure prediction rather than control. This paper evaluates generalized models with excitation-based source pretraining in zero-shot target MPC, finding lower control costs than the stated benchmarks.

  • Problem

    Building-specific data collection and model training impede scalable MPC, while transfer-learning studies typically assess target prediction without establishing downstream control performance.

    Source: §1 · pp. 1–2

  • Method

    The paper pretrains generalized models on multiple source buildings, including excitation-based data, and applies them zero-shot in target-building MPC.

    Source: §2.3 · p. 5, §5 · p. 13

  • Results

    6.4% average MPC-cost reduction versus online linear model-based MPC and 36.9% versus a PI controller were reported for excitation-based pretraining.

    Source: §5 · p. 13

  • Takeaways & Limitations

    Excitation-based generalized models transfer across buildings without target-specific data collection or modeling effort, contributing to reduced MPC setup cost.

    Source: §5 · p. 13

    Evaluation is limited to simulated target buildings; the authors identify validation on a real-world building as future work.

    Source: §4.2 · p. 12

Abstract

from arXiv · show

The widespread adoption of data-driven, energy-efficient model predictive control (MPC) in buildings remains hindered by substantial effort to collect data and train models for individual buildings. Transfer learning (TL) has consequently gained increasing attention for target building modeling, as it reduces data requirements and modeling effort by reusing pretrained source models. However, these TL models are typically evaluated only on prediction accuracy in the target, without testing downstream control performance. To address this gap, we apply a state-of-the-art TL approach - pretraining a generalized model on multiple source buildings using standard operational data - within an MPC setup in a target building. We show that this approach is insufficient to achieve satisfactory control performance. As a solution, we introduce generalized models pretrained on excitation-based operational source data - purposefully probed inputs that explore the building's state-action space. For evaluation, we apply the generalized models via zero-shot (i.e., without fine-tuning) to 32 simulated target buildings and assess MPC performance. Our results show that excitation-based generalized models achieve the strongest control performance among all benchmarks, outperforming an online linear model-based MPC and a PI controller by 6.4% and 36.9%, respectively. By combining strong control performance with the ability to generalize across multiple buildings, without requiring any target-specific data, our approach reduces MPC setup cost and simplifies its widespread deployment in the building sector.

1. Introduction

Building MPC needs thermal models, but building-specific data collection and training impede scalable deployment; prior transfer-learning studies generally assess prediction rather than downstream control. This study investigates generalized transfer-learning models in closed-loop MPC and whether excitation-based source pretraining improves zero-shot target control.

1. Introduction

  • Prior transfer-learning studies focused on target prediction, while higher neural-network predictive accuracy does not necessarily translate into better building-control performance.The paper identifies this prediction–control performance gap as a motivation for evaluating generalized models in downstream control.Source: §1 · pp. 1–2
  • Standard operational data cover a narrow region near the temperature setpoint, whereas designed excitation signals probe a broader state-action space.Excitation in occupied buildings is constrained by comfort, energy use, and planning effort; source-building pretraining offers a way to apply it without disrupting target occupants.Source: §1 · pp. 1–2

1.1. Literature Review

  • In building-control studies, MLPs achieved the highest prediction accuracy, but this did not yield the best control performance; linear models performed better and were recommended in studies.Source: §1.1 · p. 2, §1 · p. 2
  • Earlier transfer-learning studies addressed building thermal modeling but did not examine dedicated excitation during pretraining or downstream MPC across independent source and target buildings.Source: §1.1 · p. 3

1.2. Contribution

  • To the authors’ knowledge, this is the first study to investigate generalized transfer-learning models in a closed-loop MPC setup.The contribution also identifies excitation-based pretraining as a route to effective zero-shot control and narrowing the prediction–control gap.Source: §1.2 · p. 3
  • 6.4% average reduction in MPC cost versus online linear model-based MPC, with no target-specific data, is the stated contribution.The contribution concerns excitation-based generalized models.Source: §1.2 · p. 3

2. Method

The method trains generalized thermal-dynamics models on multiple source buildings, then applies them without target fine-tuning in MPC for unseen buildings. Source data can use excitation schemes designed to probe building responses.

2.1. Building-Simulation Environment

  • The study uses 64 simulated source buildings and 32 targets representing varied Central European single-family houses; obtaining data from 64 real sources would be challenging.Source data generation is intended to take place in simulation, with target control prospectively intended for a real building.Source: §2.1 · pp. 3–4

2.2. General MPC Formulation

  • MPC minimizes heating activity and comfort-band violations while smoothing control changes, then applies the first action and repeats over a receding horizon.The objective omits electricity-price costs to preserve direct comparison with the primary benchmark.Source: §2.2 · pp. 4–5

2.3. Generalized Transfer Learning Models

  • Generalized transfer learning pretrains a model across multiple source buildings, with each excitation scheme applied to the same 64 buildings for later unseen-target MPC.The approach builds on prior findings that generalized models outperform single-source transfer learning, models trained from scratch, and time-series foundation models [13].Source: §2.3 · p. 5

2.3.1. Excitation Schemes

  • Source excitation ranges from setpoint-tracking standard operation to PRBS, ramps, and bounded random walks; PRBS combines short- and long-time constants.Source data are simulated over two winter periods, each spanning January 1 to March 30.Source: §2.3.1 · pp. 5–6

2.3.2. Model Pretraining

  • The neural models predict future indoor temperatures from past measurements, weather forecasts, and planned heating controls, using a 96-step lookback and 48-step horizon.MLP, LSTM, and Transformer architectures are trained on each source dataset, yielding 12 generalized models.Source: §2.3.2 · p. 6

2.3.3. Target Application of Generalized Models

  • In zero-shot application, each pretrained model predicts target temperatures directly from past measurements, forecasts, and future controls, without fine-tuning or target-specific data collection.For neural models, MPC optimizes the control sequence by gradient descent while keeping pretrained weights frozen.Source: §2.3.3 · pp. 6–7

2.4. Evaluation Setup and Baselines

  • Evaluation tests 32 target buildings over one January month for control performance and compares prediction accuracy on standard-operation and MPC-operation test sets.Selected experiments also report control results over ten weeks.Source: §2.4 · p. 7
  • Benchmarks include standard-operation pretraining, an online linear model retrained daily on 12 days of past data, and a model-free PI controller.The linear formulation yields a convex quadratic program, while the PI controller reacts to measured temperature error.Source: §2.4 · pp. 7–8

3. Experiments

The experiments evaluate generalized models across 32 target buildings, first comparing architectures and excitation schemes and then benchmarking the best generalized model against other controllers. Excitation-pretrained models achieve lower MPC costs than models pretrained on standard operational data, and the ramp-excitation MLP performs best against the controller benchmarks.

3.1. Generalized Models

  • Prediction gaps are 1.07–3.50 K for standard-data models versus 0.0–0.037 K for excitation-based models, while good standard-test accuracy alone does not indicate good control.The MLP trained on ramp-excitation data achieves the best comfort and total cost among the generalized models.Source: §3.1 · p. 8, Figure 5 · p. 10
  • The ramp-excitation MLP has the shortest MPC optimization time, while update time depends on architecture rather than excitation strategy.Average time per optimization step is 0.82 s for MLP, 1.54 s for Transformer, and 9.48 s for LSTM.Source: §3.1 · pp. 8–9
  • Across 12 generalized models, excitation-pretrained models substantially reduce target MPC costs; ramp excitation has the lowest median, variance, and mean cost.With excitation-based pretraining, MLP, LSTM, and Transformer costs are comparable; PRBS has greater variance and slightly higher median costs than ramp and random walk.Source: §3.1 · p. 8, Figure 4 · p. 9

3.2. Comparison to Benchmarks

  • In the illustrated building, the generalized MLP keeps temperature within comfort bounds and adjusts control to ambient temperature and solar irradiation.The linear model occasionally deviates from comfort bounds; the PI controller reacts after temperature changes rather than anticipating weather or setpoint changes.Source: §3.2 · p. 10
  • 6.4% over linear-model MPC and 36.9% over the PI controller: the ramp-excitation MLP achieves the lowest average total cost across the 32 targets.It outperforms the linear model on 25 of 32 buildings and both the PI controller and standard-data LSTM on all 32.Source: Table 4 · p. 10, §3.2 · p. 10, Figure 6 · p. 11
  • Linear models generalize in prediction better than the standard-data LSTM but worse than the ramp-excitation MLP; the MLP also performs consistently better week by week over ten weeks.The linear-model prediction comparison covers both standard-operation and MPC-based test sets.Source: §3.2 · pp. 10–11

4. Discussion

The discussion finds that excitation-based pretraining supports effective zero-shot MPC, with the ramp-trained MLP outperforming benchmark controllers. It also identifies evaluation scope and model-selection choices as limitations or open directions.

4.1. Principal Findings

  • Excitation-based pretraining reduces control cost by around 86% while generally keeping prediction error below 0.3 K; standard-data models have poor MPC performance.Standard-data models nevertheless achieve good prediction accuracy on the standard-operation test set.Source: §4.1 · p. 11
  • The prediction gap is 1.07–3.50 K for standard-data models versus 0.0–0.037 K for excitation-based models across the two test sets.For the ramp-trained MLP, the gap is 0.016 K, compared with 0.56 K for the linear benchmark.Source: §4.1 · pp. 11–12
  • The ramp-pretrained MLP improves control performance over online linear model-based MPC by 6.4% and over the PI controller by 36.9%.The linear model requires daily retraining and ongoing collection and storage of building-specific data.Source: §4.1 · p. 12
  • Ramp excitation performs best, random walk follows, and PRBS performs weakest among the excitation schemes, though its MPC results remain satisfactory.The authors suggest that ramp and random-walk inputs perform better because they cover intermediate control values.Source: §4.1 · p. 12
  • With excitation-based pretraining, the MLP performs slightly better than the LSTM and Transformer, and all three architectures show small performance differences.The MLP takes 17 minutes to pretrain and 0.8 seconds per MPC update, about 2 times faster than the Transformer and 11 times faster than the LSTM.Source: §4.1 · p. 12
  • Excitation-based generalized models can be applied without target-specific data, model training, or fine-tuning, addressing repeated per-building setup effort.The discussion cautions that standard-test prediction accuracy alone does not indicate good control performance.Source: §4.1 · p. 12

4.2. Limitations and Future Work

  • Evaluation is limited to simulated, single-zone residential buildings in Central Europe under heating-dominated operation; real-building validation and broader building domains remain future work.The authors identify multi-zone and non-residential buildings and combined heating and cooling as directions for extension.Source: §4.2 · pp. 12–13
  • Fine-tuning could improve on zero-shot results, but its closed-loop control effects remain open; future work could also tune hyperparameters directly for control performance.The study avoids standard tuning because source-validation performance is not indicative of downstream MPC performance on a target building.Source: §4.2 · p. 13

5. Conclusion

The conclusion reports that excitation-based generalized models improve control across unseen buildings without target-specific data collection or modeling. They outperform both benchmark controllers while standard-operational-data pretraining yields high control costs.

5. Conclusion

  • Excitation-based pretraining reduces control costs by an average of 86.8%, outperforming online linear model-based MPC by 6.4% and a PI controller by 36.9%.The generalized models are applied zero-shot to unseen target buildings.Source: §5 · p. 13
  • Standard-operational-data pretraining results in high control costs across target buildings.Source: §5 · p. 13
  • The approach transfers across buildings without target-specific data collection or modeling effort, contributing to reduced MPC setup cost.Source: §5 · p. 13

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

The authors disclose using Claude during manuscript preparation to improve language, clarity, and sentence structure. They report reviewing and editing the text and taking responsibility for the published content.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

A. Hyperparameter Selection

The study adopts most architecture hyperparameters from prior work, with separate preliminary choices for the MLP. It avoids standard hyperparameter tuning because source validation performance does not indicate target-building MPC performance.

A. Hyperparameter Selection

  • The Transformer uses three encoder blocks, two dense decoder layers, and cosine positional embeddings to address temporal invariance.Source: A. Hyperparameter Selection · p. 13
  • The study adopts hyperparameters from for the architectures, while selecting MLP hyperparameters based on preliminary experiments.Source: A. Hyperparameter Selection · p. 13
  • The authors avoid standard hyperparameter tuning because source validation performance is not indicative of downstream MPC performance on unseen target buildings.Source: A. Hyperparameter Selection · p. 13

B. Cost Function Weights

The excitation-pretrained MLP has the lowest cost across the tested cost-weight configurations. The section also explains why the study uses equal comfort and heating weights for its comparison with the linear benchmark.

B. Cost Function Weights

  • Normalized cost is lowest at w0 = 100 and w1 = 1, while the linear model has a large upper quartile at that weighting.The linear model's normalized average cost at that weighting is 87785.Source: B. Cost Function Weights · pp. 13–14, Table B.1 · p. 14
  • The excitation-pretrained MLP achieves the lowest cost across all tested cost-weight configurations.Table B.1 compares the linear MPC benchmark with the MLP pretrained on ramp excitation across 32 target buildings.Source: B. Cost Function Weights · pp. 13–14, Table B.1 · p. 14
  • The study uses w0 = w1 = 1 for a fair comparison with the linear benchmark, since practice rarely weights heating more than comfort.The passage says doing so would undermine the heating system's primary purpose of maintaining thermal comfort.Source: B. Cost Function Weights · p. 14

C. Building-Level Control Performance

The ramp-excitation-pretrained MLP has the lowest total MPC cost for most target buildings, while the online linear model performs best for the rest.

C. Building-Level Control Performance

D. Prediction Performance of Linear Model

The linear model's prediction error is close to the standard-data LSTM's on the standard-operation test set but increases substantially on the MPC-based test set. Its test-set prediction gap is between those of the LSTM and ramp-excitation MLP.

D. Prediction Performance of Linear Model

  • The linear model is trained using December data from its first online cycle and tested on January data.Source: D. Prediction Performance of Linear… · p. 14
  • On the standard-operation test set, the linear model's mean MAE is 0.10 K, compared with 0.08 K for the standard-data LSTM.The linear model shows higher variability across buildings.Source: D. Prediction Performance of Linear… · p. 14
  • On the MPC-based test set, the linear model's mean prediction error is 0.66 K; its 0.56 K prediction gap is smaller than the LSTM's 1.07 K gap but larger than the ramp-excitation MLP's 0.016 K gap.Most buildings remain below an MAE of 1 K on the MPC-based test set.Source: D. Prediction Performance of Linear… · p. 14

E. Ten-Week Control Evaluation

Across the ten-week evaluation, the ramp-excitation-pretrained MLP has the best MPC performance, closely followed by the linear model. Weekly costs show a similar temporal pattern across the three controllers.

E. Ten-Week Control Evaluation

Related papers

Loading 2609.12853v2…