"DC-HVAC-DT23: Data Center Cooling Telemetry and Probabilistic Surrogate Rollout Dataset"

"A real dataset collected from an experimental HVAC set-up of an industrial data center, recorded under normal (fault-free) operation and intended for the development and benchmarking of data-driven digital twins of data center cooling systems (by \"DATARK\" LLC and \"DATCHECK\" LLC).The experimental testbed reproduces a single server-rack footprint inside a thermally contained hot\/cold-aisle enclosure. Cooling is provided by a commercial in-row direct-expansion (DX) precision air conditioner operating on R410A refrigerant, comprising a compressor, an outdoor air condenser, an evaporator, a receiver, a filter-drier, a hot-gas bypass (HGBP) control valve, a thermostatic expansion valve, and a forced ventilation system. The IT heat load is emulated by industrial heat guns providing approximately 5 kW of sensible load in the hot aisle; a built-in PID controller maintains a 22 \u00b0C return-air setpoint with a 1 \u00b0C deadband. Although the testbed uses a single DX circuit and a single rack footprint, it captures the core thermodynamic couplings \u2014 compressor cycling, condenser-side ambient dependence, superheat control, and air-side recirculation \u2014 that govern energy behaviour in production-scale installations, providing a physically representative basis for evaluating surrogate models of cooling energy.The dataset is released in two parts.Part 1 \u2014 Measured testbed telemetry. Telemetry was acquired continuously through a PLC and a Zabbix monitoring server at 1-minute resolution and resampled to 15-minute averages, which matches the thermal inertia of the enclosure (\u03c4 > 10 min) and ASHRAE PUE measurement practice. Data were recorded over six multi-day sessions spanning April 2025 to January 2026 at a site in Yekaterinburg, Russia, whose sharply continental climate exposes the cooling system to an ambient range from \u221222.8 \u00b0C to +36.1 \u00b0C across the recording period; only steady-state, fault-free periods were retained. The release contains 9,570 samples (approximately 100 days of cumulative operation) split chronologically into training, validation, and test partitions. The partitioning is deliberately non-random: the test partition covers a deep-winter period with ambient temperatures down to \u221222.8 \u00b0C that does not overlap the training temperature range, providing a genuine out-of-distribution regime for evaluating extrapolation.Each sample contains 2 exogenous drivers (outdoor ambient temperature and total IT load) and 23 system-response variables describing the refrigeration circuit and the air distribution system, ordered below along the physical flow path:- Air side.C_Unit_Inlet_Air_Temp \u2014 air temperature entering the unit at the evaporator inlet. *Hot_Aisle_Temp* \u2014 return air temperature in the contained hot aisle. *Evaporator_Fan_Speed_Avg* \u2014 average operating speed of the indoor evaporator fans. *Air_Flow_Rate* \u2014 volumetric air flow through the indoor unit fans. *Air_Filter_Differential_Pressure* \u2014 pressure drop across the evaporator air filter. *AC_Unit_Outlet_Air_Temp* \u2014 supply air temperature leaving the unit after the evaporator. *Rack_Inlet_Air_Temp* \u2014 air temperature at the rack inlet in the cold aisle.- Refrigerant circuit. Suction_Pressure* \u2014 refrigerant pressure on the compressor suction side. *Suction_Line_Refrigerant_Temp* \u2014 refrigerant vapour temperature in the suction line. *Compressor_Active_Power* \u2014 active electrical power drawn by the compressor. *Discharge_Pressure* \u2014 refrigerant pressure on the discharge side. *Discharge_Line_Refrigerant_Temp* \u2014 refrigerant temperature in the discharge gas line. *HGBP_Valve_Position* \u2014 relative opening of the hot gas bypass valve. *Outdoor_Unit_Active_Power* \u2014 active electrical power drawn by the outdoor condenser unit. *Refrigerant_Temp_After_Receiver* \u2014 refrigerant temperature immediately downstream of the receiver. *Filter_Drier_Inlet_Refrigerant_Temp* and *Filter_Drier_Outlet_Refrigerant_Temp* \u2014 refrigerant temperatures on either side of the filter-drier. *Evaporator_Superheat_Temp* \u2014 superheat of the refrigerant leaving the evaporator.- **Derived energy indicators.** *AC_Cooling_Capacity* \u2014 instantaneous cooling capacity delivered by the unit. *Cooling_Demand* \u2014 cooling capacity currently required to hold the temperature setpoint. *Cooling_Power* \u2014 total electrical power of the cooling system (compressor, outdoor unit, and indoor fans). *PUE* \u2014 Power Usage Effectiveness. *EER* \u2014 Energy Efficiency Ratio.Part 2 \u2014 Probabilistic surrogate rollout. To support research on long-horizon digital twin simulation, the release also contains a one-year open-loop rollout produced by six probabilistic surrogate models (MLP, LSTM, Parallel MLP-LSTM, TCN, Transformer encoder, and a CatBoost ensemble) trained on Part 1. The models were driven over a full calendar year of exogenous boundary conditions at 15-minute resolution \u2014 real historical 2024 ambient temperature together with a synthesized IT load profile \u2014 and each emits, for every one of the 23 system-response variables, both a point estimate and an input-dependent uncertainty scale. This part enables the study of calibration, long-horizon thermodynamic consistency, and seasonal energy behaviour without access to the original models or GPU resources. The rollout year is independent of, and does not overlap with, the measurement period of Part 1; it is a counterfactual scenario rather than a held-out measurement."