Cooling uses a large share of data center energy, and learned controllers now save real energy over conventional control: 9% of cooling cost at Google, and 14–21% in a production deployment in China. Each of these controllers, however, rests on a dynamics model fitted to one room. The model has to be rebuilt, or re-excited with deliberate perturbations, every time it moves to a new building. This note argues that the move to a new site bundles two different kinds of change. The first is a change in observed conditions: outdoor weather, grid carbon intensity, and IT workload. A sequence model already receives these as inputs, and in principle it can generalize over them. The second is a change in unobserved physics: thermal mass, airflow recirculation, cooling-plant time constants, and room layout. No input tells the model about these. Current cross-domain methods either treat "site" as a known label with its own weights, or ignore the distinction altogether.
We propose a recurrent dynamics model that handles the second kind of change by inference rather than retraining. A shared gated recurrent network — a sequence model whose internal state is updated step by step and acts as a running memory of the past — predicts rack temperatures and cooling power. It is conditioned on a small facility code. A second recurrent encoder reads the code off a few hours of ordinary operating logs from the new site, with no gradient updates and no exploration on the live floor. The core claim is falsifiable. On sites that differ only in climate, this model should be no better than a pooled model given weather as input. On sites that differ in physical configuration, it should approach a per-site model trained on months of data, and it should beat per-site-weight methods and brief fine-tuning.
Why this gap, and why this design
The data center literature shows that site-specificity is the practical bottleneck. Data center cooling using model-predictive control fits a linear autoregressive model with exogenous inputs (ARX) and spatial sparsity, and it controls a live floor every 30 seconds. It needs about three hours of safe random-walk perturbation per floor. It also reports that a week of steady PID logs was worse than those three hours, because steady operation carries too little variation to reveal cause and effect. This result constrains our design: any method that adapts from passive logs must be tested on exactly the low-variation data that defeated the Google model. Data Center Cooling System Optimization Using Offline Reinforcement Learning uses a graph neural network over rack–cooling-unit topology and a time-symmetry-regularized dynamics model. It is evaluated on two rooms in one region, with no mechanism for moving between them. SustainDC trains agents from scratch in each of eight US states and finds that no single algorithm wins everywhere. This finding motivates the proposal, but it mixes the climate and facility effects we want to separate.
The recurrent literature supplies the mechanism. NeuralHydrology shows that the memory cells of a long short-term memory network (LSTM) track unobserved physical storage terms (soil water, snowpack) that were never supervised. This suggests that a recurrent state can carry slow thermal storage, and that a conditioning code could carry the parameters governing it. PhyLSTM adds a first-order heat and mass balance to the LSTM training loss for a cooling tower. It reports under 2% estimation error and inference under 15 ms, on a single tower with Belgian weather. We reuse its physics residual but apply it across a population of facilities. SegRNN shows that a gated recurrent unit (GRU) run over segments rather than single points stays competitive on long horizons: it is best in 50 of 56 forecasting settings against PatchTST, with over 78% less training time. We use its segment encoder for the facility encoder, which must read hours of 1-minute data cheaply.
The cross-domain literature supplies the contrasts. STAR multiplies shared weights by per-domain weights and keeps per-domain normalization statistics. It improves multi-domain AUC to 0.6506 from 0.6364 for a pooled model, but it needs every domain to be seen in training, so it has nothing to offer a new building. Selective Cross-Domain Consistency Regularization shows on time-series classification that forcing all domains into one representation causes negative transfer, while aligning only similar domains helps (Bearings accuracy 89.1% against 82.2% for standard training). We take its lesson but not its method. It targets classification and relies on known domain labels. Our encoder instead measures similarity between facilities in a learned code space, and it works for continuous multi-step regression.
Core Claim
Let $x_t$ be the observed exogenous inputs at step $t$ (weather, IT load, carbon intensity), $u_t$ the cooling actions (setpoints, fan speeds), $y_t$ the measured rack temperatures and cooling power, and $h_t$ the recurrent state. The model is
$$z = E_\phi(\{x_s,u_s,y_s\}_{s \in \text{probe}}), \qquad h_t = \mathrm{GRU}_\theta(h_{t-1}, [x_t,u_t]; z), \qquad \hat y_t = g_\theta(h_t),$$
Here $E_\phi$ is a segment-recurrent encoder that maps a probe window of ordinary logs (default three hours, matching the Google exploration budget) to a code $z$ of 4–16 dimensions. The code enters the GRU through feature-wise scaling of its gates. Training uses a population of facilities and minimizes
$$\mathcal{L} = \sum_{k=1}^{H} \lVert \hat y_{t+k} - y_{t+k} \rVert^2 + \lambda\, \mathcal{L}_{\text{energy}},$$
where $H$ is the forecast horizon and $\mathcal{L}_{\text{energy}}$ is a PhyLSTM-style residual. That residual penalizes any mismatch between predicted cooling power and IT heat plus the change in stored heat. The probe window and the forecast window are drawn from different periods of the same facility. This forces $z$ to encode persistent physics rather than the recent trajectory.
Research Program
We first need a population of facilities with known ground-truth physics. The DCRL-Green digital twin from Sustainability of Data Center Digital Twins with Reinforcement Learning exposes cabinet geometry, server power specifications, and chiller, pump, and cooling-tower parameters. We sample several hundred configurations over thermal mass, rack density, recirculation fraction, and cooling-plant sizing. We cross them with the eight SustainDC weather and carbon-intensity locations and the Alibaba and Google workload traces. Training data come from logs generated under ASHRAE Guideline 36 control with small setpoint jitter, so the adaptation data are as uninformative as real PID logs.
Second, we define the split that makes the claim testable. The split has three held-out cells: new facility with a seen climate, seen facility with a new climate, and both new. The prediction is asymmetric. Where only the climate is new, the facility code should add nothing. Where the facility is new, it should add a lot.
Third, we check what the code captures. Following NeuralHydrology, we regress $z$ onto the simulator's true parameters. If $z$ is reliably informative about thermal mass and recirculation, the model is identifying physics. If $z$ mostly tracks climate, it is not.
Fourth, we close the loop. The adapted model drives a model-predictive controller in the style of the Google system: at every step it plans the cooling actions that minimize predicted energy, subject to a rack-temperature limit. We run this on held-out facilities, and on the LC-Opt Frontier liquid-cooling twin as a structurally different cooling topology that never appears in training.
Baselines
The naive reference is ASHRAE Guideline 36 rule-based control, together with a persistence forecaster for the modeling task. Every energy number is reported relative to Guideline 36. The file gives two reference points for learned control against it: Sustainability of Data Center Digital Twins with Reinforcement Learning reports a 13% carbon reduction for multi-agent reinforcement learning (MADDPG) and 7% for single-agent PPO. Those policies are trained per configuration, so they serve as a per-site upper reference rather than a transfer competitor.
The strongest competitor on the modeling question is a pooled GRU of identical size and segment structure, trained on all facilities, with weather, workload, and carbon intensity as inputs but no facility code. This is the baseline most likely to be good enough. It gets the same data, the same hyperparameter search budget (equal trials on the same validation facilities), and the same physics residual.
A second competitor is a STAR-style recurrent model: shared GRU weights multiplied by per-facility weights, plus per-facility normalization, as in STAR. For an unseen facility it receives new facility weights fitted on the same three-hour probe.
The third competitor is per-site ARX identification on the same probe, following Data center cooling using model-predictive control. It is run twice: once with passive jittered logs, and once with the Google-style random-walk excitation, which our method is not allowed to use. The latter is the paper's own recipe, which achieved a 9% cooling-cost reduction against PID in production.
Fourth, the pooled GRU is fine-tuned on the probe, with early stopping on a held-out slice of the probe itself.
An oracle GRU trained on months of data from the target facility bounds achievable accuracy from above.
The file contains no reported result on Lacuna for cross-facility forecasting of data center thermal dynamics. All modeling comparisons are therefore run fresh under matched budgets.
Ablations isolate the key ingredient. The first removes the code (which reduces to the pooled baseline). The second replaces the inferred code with hand-written facility metadata (rack count, rated cooling capacity). The third drops the physics residual. The fourth shrinks the probe from three hours to thirty minutes.
Beating the field means three things on new facilities. The model must close at least half the gap in 1–6 hour forecast error between the pooled GRU and the oracle. It must match or beat excited-ARX without excitation. In closed loop, it must deliver cooling energy savings against Guideline 36 within a few points of the oracle-model controller, with no increase in temperature violations.
Why not just…?
Why not just pool all sites and give the model weather as input? This is the real threat, and the design is built around it. For sites that differ only in climate we expect the pooled model to be as good as ours, because weather is an observed input and a recurrent model can generalize over it. That is why the held-out-climate cell is included. The proposal survives only if a clear gap opens on held-out facilities, where the relevant differences (thermal mass, recirculation) are not in any input. If the pooled model closes that gap too, the conclusion is that data center transfer is mostly a covariate problem, and the facility code should be dropped.
Why not just fine-tune a pooled model on the new site's logs? With months of target data, fine-tuning will probably match us, so the comparison is made at equal and small target-data budgets. The Google result that a week of steady PID logs underperformed three hours of excitation suggests that gradient updates on low-variation logs mostly overfit the operating point. Our encoder instead matches the probe against a code space learned across hundreds of facilities. The probe-length ablation shows where, if anywhere, fine-tuning catches up.
Why not just identify a per-site linear model, as Google did? Inside its excited range, ARX may be competitive, and the excited-ARX baseline measures exactly that. Our claim is narrower. We expect to match it without live perturbation, which operators of risk-averse facilities often cannot allow, and to degrade less when control drives the room outside the probe's operating range. PhyLSTM's evidence of nonlinear humidity and mass coupling in evaporative cooling is the reason to expect linear extrapolation to fail. The LC-Opt transfer test checks this directly.
Why not just attach metadata descriptors to each facility? Descriptors such as rack count and rated capacity are cheap and often available. The metadata ablation tests whether they carry the same information as the inferred code. Selective Cross-Domain Consistency Regularization found learned similarity beat metadata similarity (89.1% against 87.9% on Bearings). That result is modest, and in our setting metadata could plausibly win. If it does, the simpler method should be adopted.
Decisive Checks
The experiment that settles the idea is the two-by-two held-out split. We need the code to be useless under climate shift and valuable under facility shift. Both halves must hold. A code that helps under climate shift is probably memorizing weather, and a code that does not help under facility shift has failed at its purpose. The second check is whether $z$ predicts the simulator's hidden thermal parameters on unseen facilities. The third is closed-loop performance on the LC-Opt liquid-cooling twin, whose topology never appears in training. There we expect graceful degradation rather than success, and we report honestly where the facility code stops transferring. If the real-telemetry rooms from the offline reinforcement learning study become accessible, the final test is to adapt from Room A to Room B using only a three-hour log window.
References
- Data center cooling using model-predictive control
- Data Center Cooling System Optimization Using Offline Reinforcement Learning
- SustainDC: Benchmarking for Sustainable Data Center Control
- NeuralHydrology - Interpreting LSTMs in Hydrology
- Physics-Informed LSTM Network for Flexibility Identification in Evaporative Cooling System
- SegRNN: Segment Recurrent Neural Network for Long-Term Time-Series Forecasting
- One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction
- Selective Cross-Domain Consistency Regularization for Time Series Domain Generalization
- Sustainability of Data Center Digital Twins with Reinforcement Learning
- LC-Opt: Benchmarking Reinforcement Learning and Agentic AI for End-to-End Liquid Cooling Optimization in Data Centers