Data center water loop heat pump air conditioning system and control method

Abstract

According to the data center water loop heat pump air-conditioning system and the control method, the temperature control response speed and precision are remarkably improved by constructing a feature matrix fused with waterway delay compensation factors and adopting a load-sensitive double-engine prediction algorithm and a digital twinning pre-verification mechanism; the original delay compensation model brings the waterway transmission delay into a prediction feedforward link, so that the response delay is shortened from 12 minutes to less than 3 minutes; the dynamic weight fusion algorithm adaptively adjusts the long and short time prediction weight according to the load sudden change intensity, and the temperature overshoot probability is obviously reduced; the digital twinborn verification module avoids an actual system execution risk through a hybrid model rehearsal control strategy, and the number of times of daily start and stop of a compressor is obviously reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data center air conditioning, in particular to a data center water ring heat pump air conditioning system and a control method. BACKGROUND

[0002] With the rapid development of cloud computing and artificial intelligence technology, the energy consumption problem of data centers is increasingly prominent, and the energy consumption of air conditioning systems accounts for as high as 30%-40%. In order to reduce the power usage effectiveness index, the water ring heat pump air conditioning system has been widely used in the field of data center cooling due to its advantage of energy recovery by utilizing the heat storage characteristics of water. The existing technology generally adjusts the compressor frequency in response to temperature changes or uses a valve control method based on PID feedback, both of which try to balance energy saving and temperature control stability. The current mainstream commercial systems (such as Schneider EcoStruxure and Emerson Liebert series) generally use a sensor real-time monitoring + feedback control architecture to implement closed-loop regulation by collecting parameters such as cabinet inlet temperature and cooling water flow.

[0003] However, due to the inherent thermal inertia characteristics of the water ring heat pump system, there is a significant delay in the transmission of cooling water in the pipeline (the average delay is 8-12 minutes), which causes serious hysteresis in traditional feedback control. The specific performance is as follows:

[0004] Response lag: when the IT load changes suddenly, the existing system needs to wait for the temperature sensor feedback before starting the adjustment, at which time the computer room temperature has exceeded the safety threshold, forcing the server to trigger a protective frequency reduction;

[0005] Control oscillation: the phase lag caused by the delay makes the PID controller produce continuous oscillation, and the compressor is frequently started and stopped up to 12 times within 1 hour, accelerating the equipment aging;

[0006] Energy efficiency loss: to compensate for the hysteresis effect, the temperature set value is often reduced by 2-3℃ as a safety margin by the operation and maintenance personnel, resulting in an increase in annual PUE to above 1.45;

[0007] Prediction loss: the existing scheme lacks deep mining of historical operation data and cannot predict the load change trend, and the existing technology only uses linear extrapolation method to predict temperature, with high error rate.

[0008] Therefore, it is urgent to develop a new type of water ring heat pump control system that can effectively solve the temperature control hysteresis problem caused by water delay, realize the advanced prediction of computer room temperature changes, the collaborative optimization control of multiple actuators, and the safety verification mechanism of control strategy, to meet the stringent requirements of high-end data centers for temperature control stability and system reliability. SUMMARY

[0009] The purpose of this invention is to provide a data center water-loop heat pump air conditioning system and control method to solve the problems existing in the prior art.

[0010] To achieve the above objectives, the present invention provides the following solution:

[0011] This invention provides a data center water-loop heat pump air conditioning system, comprising:

[0012] The multi-source heterogeneous data acquisition module is used to connect with the sensor network deployed in the data center through industrial communication protocols and acquire multi-source heterogeneous operating data in real time.

[0013] The feature fusion preprocessing module is connected to the multi-source heterogeneous data acquisition module and is used to receive the multi-source heterogeneous running data and construct a spatiotemporal feature matrix.

[0014] A multi-scale prediction engine module, connected to the feature fusion preprocessing module, is used to receive the spatiotemporal feature matrix and output a final predicted temperature sequence.

[0015] The dynamic optimization decision module, connected to the multi-scale prediction engine module, is used to solve a constrained optimization problem within a preset rolling time domain, with the objectives of minimizing the deviation between the final predicted temperature sequence and the preset target temperature and minimizing the total energy consumption of the system, thereby generating a set of optimal control commands.

[0016] The collaborative execution control module, connected to the dynamic optimization decision module, is used to receive the optimal control command and distribute it to the physical device.

[0017] Preferably, the multi-source heterogeneous operating data includes at least: the cabinet air inlet temperature sequence, the heat pump outlet water temperature sequence, the real-time cooling water flow rate, the water pump operating power, the water valve opening degree, the heat pump compressor operating frequency, and the cabinet real-time power consumption.

[0018] Preferably, the feature fusion preprocessing module includes:

[0019] The delay compensation factor calculation unit is used to calculate the water system delay compensation factor based on the received real-time flow rate of the cooling water and the pre-stored cooling water pipeline system length parameters.

[0020] The feature matrix construction unit is used to apply the water system delay compensation factor to correct the time dimension information of the real-time cooling water flow data, and integrate the corrected real-time cooling water flow data with the cabinet air inlet temperature sequence and the cabinet real-time power consumption data into a multi-dimensional feature vector sequence including multiple historical time steps, thus forming the spatiotemporal feature matrix.

[0021] Preferably, the multi-scale prediction engine module includes:

[0022] The short-term forecasting submodule includes a first forecasting model for generating a first forecasting temperature sequence in response to short-term fluctuations in IT load.

[0023] The long-term forecasting submodule includes a second forecasting model for generating a second predicted temperature sequence that reflects the long-term trend and cycle of temperature.

[0024] The dynamic fusion unit is used to first calculate the time change rate of the real-time power consumption data of the cabinet, then determine a dynamic confidence weight based on the time change rate, and use the dynamic confidence weight to perform a weighted summation operation on the first predicted temperature sequence and the second predicted temperature sequence to obtain the final predicted temperature sequence.

[0025] Preferably, the collaborative execution control module adopts a valve-compressor decoupling control strategy, which includes: for water valves, using a PID controller based on the current temperature deviation for independent basic adjustment to maintain stable water flow; for heat pump compressors, using a control law based on feedforward compensation, which takes the final predicted temperature sequence output by the multi-scale prediction engine module and the IT load change rate calculated by the feature fusion preprocessing module as inputs to directly calculate the forward-looking target compressor frequency to proactively respond to upcoming heat load changes.

[0026] The present invention also provides a control method for a data center water-loop heat pump air conditioning system, comprising the following steps:

[0027] S1. Data Acquisition: Real-time acquisition of multi-source heterogeneous operational data within the data center via a preset communication protocol. This multi-source heterogeneous operational data includes at least: time-series temperature data characterizing the thermal environment of the server racks, including the temperature sequence at the server rack air inlets; water system parameter data characterizing the cooling water circuit status, including real-time cooling water flow rate and water pump operating power; equipment status data characterizing the status of the cooling equipment, including water circuit valve opening and heat pump compressor operating frequency; and IT load data characterizing the computing task intensity of the data center, including the real-time power consumption of the server racks.

[0028] S2. Feature preprocessing: Processing multi-source heterogeneous operational data to construct a spatiotemporal feature matrix for temperature prediction. The feature preprocessing steps include: calculating a water system delay compensation factor to compensate for physical transport delays in the water system based on the real-time cooling water flow rate and the preset cooling water pipeline system length; and performing timestamp alignment processing on the water system parameter data based on the water system delay compensation factor, and combining the aligned water system parameter data with temperature time series data and IT load data to generate a spatiotemporal feature matrix with multiple time steps and temporal and spatial coupling characteristics.

[0029] S3. Multi-scale fusion prediction: Based on the spatiotemporal feature matrix, the final predicted temperature sequence within the future control time domain is generated. The multi-scale fusion prediction steps include: inputting the spatiotemporal feature matrix in parallel to the short-term temperature prediction submodule and the long-term temperature prediction submodule; the short-term temperature prediction submodule uses a first prediction model to generate a first predicted temperature sequence to respond to short-term drastic fluctuations in IT load data; the long-term temperature prediction submodule uses a second prediction model to generate a second predicted temperature sequence to capture the long-term trend and periodicity of temperature time series data; calculating the load change rate, which characterizes the drastic fluctuation of IT load data; determining the dynamic confidence weights for fusing the first and second predicted temperature sequences based on the load change rate; and weighting and fusing the first and second predicted temperature sequences according to the dynamic confidence weights to obtain the final predicted temperature sequence.

[0030] S4. Optimize decision-making: Based on the final predicted temperature sequence and the preset target temperature, solve the multi-objective optimization problem under the preset equipment operation constraints to generate a set of optimal control instructions. The optimal control instruction set includes at least the target compressor frequency and the target valve opening.

[0031] S5. Digital Twin Pre-simulation Verification: The optimal control command set generated by the optimization decision-making steps is first input into a pre-constructed digital twin model. The digital twin model is a virtual simulation environment that integrates a device thermodynamic model based on physical laws and a behavioral model trained based on historical data. The optimal control command set is simulated and executed in the digital twin model, and its possible virtual temperature response and virtual energy consumption results are predicted. It is determined whether the virtual temperature response meets the preset safe temperature threshold. Only when the virtual temperature response meets the safe temperature threshold is the execution of the command authorized, and the optimal control command set is distributed to the physical device.

[0032] S6. Command execution: Distribute the optimal control command set to the heat pump compressor and water valves to execute the control.

[0033] Preferably, in step S2, the method for calculating the water system delay compensation factor is as follows:

[0034] τ=f(G w ,L pipe );

[0035] Among them, L pipe G is the total length of the cooling water piping system. w This represents the real-time flow rate of the cooling water collected at the current moment.

[0036] The construction of the spatiotemporal feature matrix specifically includes: concatenating the cabinet air inlet temperature sequence and cabinet real-time power consumption sequence, which include the current time and historical time, with water system parameter data with timestamp k-τ after delay compensation factor compensation, to form a multi-dimensional input vector.

[0037] Preferably, in step S3, the dynamic confidence weight is calculated using the Sigmoid function, with the following formula:

[0038]

[0039] in, γ is the absolute value of the rate of change of real-time power consumption of IT load with respect to time, and γ is a preset coefficient used to adjust the weight sensitivity.

[0040] Preferably, in step S3, the first prediction model used in the short-term temperature prediction submodule is a long short-term memory network model, which is used to receive the spatiotemporal feature matrix as input and output the point-by-point temperature prediction value within the future preset short-term time window.

[0041] The second prediction model used in the long-term temperature prediction submodule is the Prophet Algorithm Model, which is used to fit historical hourly average temperature data into trend, periodic and holiday effect terms to output the predicted temperature trend value within a preset long-term time window.

[0042] Preferably, in step S4, the multi-objective optimization problem is solved within a rolling time-domain optimization framework, which defines a finite future time domain at the beginning of each control cycle; the solver used to solve the multi-objective optimization problem is a particle swarm optimization algorithm.

[0043] The objective function of the optimization problem specifically includes a temperature deviation penalty term and an energy consumption penalty term. The temperature deviation penalty term is proportional to the square integral of the difference between the final predicted temperature sequence and the target temperature, and the energy consumption penalty term is the weighted sum of the operating power of the heat pump compressor and the operating power of the water pump.

[0044] The preset equipment operating constraints include: the operating frequency of the heat pump compressor must be between the preset minimum frequency and the preset maximum frequency, and the absolute value of its frequency change rate must not exceed the preset maximum frequency conversion rate.

[0045] The present invention achieves the following beneficial technical effects compared to the prior art:

[0046] This invention provides a data center water-loop heat pump air conditioning system and control method. By constructing a feature matrix that integrates water path delay compensation factors, employing a load-sensitive dual-engine prediction algorithm, and a digital twin pre-verification mechanism, it significantly improves the temperature control response speed and accuracy. The unique delay compensation model incorporates water path transmission hysteresis into the prediction feedforward stage, reducing the response delay from 12 minutes to less than 3 minutes. The dynamic weight fusion algorithm adaptively adjusts the long and short-time prediction weights according to the intensity of load mutations, significantly reducing the probability of temperature overshoot. The digital twin verification module uses a hybrid model to pre-simulate the control strategy, avoiding actual system execution risks and significantly reducing the number of compressor start-ups and shutdowns per day. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart of the control method for a data center water-loop heat pump air conditioning system provided by the present invention. Detailed Implementation

[0049] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages) and should not be construed as limiting the invention.

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] The purpose of this invention is to provide a data center water-loop heat pump air conditioning system and control method to solve the problems existing in the prior art.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] This invention provides a data center water-loop heat pump air conditioning system and control method. The system consists of six main parts: a multi-source heterogeneous data acquisition module, a feature fusion preprocessing module, a multi-scale prediction engine module, a dynamic optimization decision-making module, a collaborative execution control module, and a digital twin verification module. The method operates according to a closed-loop process of "data acquisition, feature preprocessing, multi-scale fusion prediction, optimization decision-making, digital twin pre-simulation verification, and command execution." Through deep hardware and software collaboration, the system and method achieve millisecond-level precise control and minute-level rolling optimization of the data center water-loop heat pump air conditioning equipment, ultimately achieving the technical effect of "high temperature control accuracy and low PUE throughout the year."

[0054] Example 1:

[0055] The data center water-loop heat pump air conditioning system in this embodiment includes:

[0056] 1. Multi-source heterogeneous data acquisition module

[0057] On the hardware side: Three types of sensor networks are deployed in the data center server room, hydraulic pipeline network, and heat pump unit;

[0058] Thermal environment sensors: One Pt1000 temperature probe is installed on each of the front and rear doors of each cabinet, with a sampling period of 1 second;

[0059] Water circuit sensors: Electromagnetic flow meters (accuracy 0.2%R) and pressure transmitters (0.1%FS) are installed on the main cooling water pipe and branch pipes respectively;

[0060] Equipment status sensors: The heat pump compressor, water pump, and electric regulating valve are all equipped with power / frequency / opening degree transmitters with Modbus-RTU interface.

[0061] On the software side: The data acquisition module runs embedded Linux and has built-in three industrial communication protocol stacks: OPC-UA, Modbus-TCP, and Bacnet / IP. It polls the above devices in parallel through a thread pool, and writes the raw data with a 64-bit nanosecond-level timestamp to a shared memory circular buffer.

[0062] 2. Feature fusion preprocessing module, which includes:

[0063] The delay compensation factor calculation unit uses the following formula to calculate the delay compensation factor τ:

[0064] τ=f(G w ,L pipe );

[0065] Among them, L pipe G is the total length of the cooling water piping system. wThe current real-time cooling water flow rate is τ. This unit updates τ in real time at a frequency of 1Hz and rounds τ down to the nearest integer second for subsequent timestamp alignment.

[0066] The feature matrix construction unit maintains a fixed-length time window of T=120 (two minutes of historical data). Using the current time k as a reference, it concatenates the rack inlet temperature sequence, the rack real-time power consumption sequence, and the delay-compensated water system parameters into a 120×7 spatiotemporal feature matrix X. k , serving as the unified input for the multi-scale prediction engine.

[0067] 3. Multi-scale prediction engine module, which includes:

[0068] The short-term prediction submodule employs a two-layer LSTM network (7 input dimensions, 128 hidden layers, 1 output dimension), trained on 1-second real-time data from the past 30 days. The loss function is HuberLoss (δ = 0.5). This submodule outputs a second-by-second temperature prediction sequence Y for the next 30 seconds. short .

[0069] The long-term forecasting submodule uses the Facebook Prophet algorithm. The input is the hourly average temperature over the past 60 days, which is decomposed into trend, daily cycle, weekly cycle, and holiday effect terms. The output is a minute-level temperature trend sequence Y for the next 30 minutes. long .

[0070] The dynamic fusion unit first calculates the time change rate of the real-time power consumption data of the cabinet, then determines a dynamic confidence weight based on the time change rate, and performs a weighted summation operation on the first predicted temperature sequence and the second predicted temperature sequence using the dynamic confidence weight to obtain the final predicted temperature sequence. The load change rate ΔP is defined as: ΔP = |P IT (k)–P IT (k-1)| / P IT (k-1), the dynamic confidence weight β uses the Sigmoid function. The dynamic confidence weight is calculated using the Sigmoid function, and the formula is:

[0071]

[0072] in, γ represents the absolute value of the rate of change of real-time power consumption of the IT load with respect to time, and γ is a preset coefficient used to adjust the weight sensitivity; the final predicted temperature sequence Y final =β·Y short +(1-β)·Y long .

[0073] 4. The dynamic optimization decision-making module adopts a rolling time-domain MPC framework for its optimization model, with a prediction domain N. p =30, control domain N c =10. Decision variables include compressor frequency, water pump frequency, and valve opening. The solver uses particle swarm optimization (PSO) with 30 particles, 50 iterations, and an inertia weight of 0.7.

[0074] 5. The collaborative execution control module adopts a "valve-compressor decoupled control strategy." On the valve side, a PID controller is used, taking the current temperature deviation as input and outputting the valve opening degree. On the compressor side, a feedforward compensator is used based on the Y... final The forward target frequency is calculated with ΔP, and the control command is sent to the frequency converter via the EtherCAT bus at a cycle of 4kHz.

[0075] 6. The digital twin pre-simulation and verification module deploys a Modelica-based thermodynamic twin model on an edge server, coupling a CFD data center airflow model with a pump-pipeline hydraulic model. The twin model simulates the data center temperature field and energy consumption for the next 30 seconds in real time with a 1-second step. If the twin results show that any rack intake air temperature exceeds 28℃ or energy consumption exceeds the threshold, the optimization decision is immediately rolled back, and instructions are prohibited from being issued; otherwise, security authentication is passed and execution is carried out.

[0076] Example 2:

[0077] The control method flow of the above system is as follows: Figure 1 As shown, it includes:

[0078] Step S1: Data Acquisition

[0079] After the system starts up, the acquisition module continuously acquires seven types of data at a frequency of 1Hz, including rack air inlet temperature, cooling water flow rate, water pump power, valve opening, compressor frequency, and rack power consumption. The data is cached in shared memory for subsequent modules to read without copying.

[0080] Step S2: Feature Preprocessing

[0081] The delay compensation factor τ is calculated and updated in real time; then, the water system data is timestamped with τ and concatenated with temperature and power consumption data to form a 120×7 spatiotemporal feature matrix.

[0082] Step S3: Multi-scale fusion prediction

[0083] The feature matrix is ​​fed into an LSTM to obtain a 30-second short-term sequence; simultaneously, Prophet outputs a 30-minute long-term trend; β is calculated based on ΔP and weighted fusion is performed to obtain the final predicted temperature sequence Y. final .

[0084] Step S4: Optimize Decisions

[0085] At the start of the current control cycle, the PSO solver searches for the optimal control scheme within the 30s prediction domain.

[0086] Step S5: Digital Twin Pre-visualization Verification

[0087] Inject the optimal control command obtained in step S4 into the digital twin model; if the twin verification passes, proceed to step S6; if it fails, trigger re-optimization, retrying up to 3 times; if it still fails, execute a conservative strategy (compressor frequency +5Hz, valve opening +10%).

[0088] Step S6: Instruction Execution

[0089] The final command is sent to the compressor inverter, water pump inverter, and electric regulating valve via the EtherCAT bus to complete a closed-loop control.

[0090] Experimental verification

[0091] This system was deployed in a Tier III data center (120 racks, 600kW IT load) and ran continuously for 30 days. Test results showed: the standard deviation of rack inlet air temperature decreased from 1.2℃ to 0.25℃; the average operating frequency of the heat pump compressor decreased by 11%, and water pump power consumption decreased by 6%; the annual PUE decreased from 1.48 to 1.36, resulting in energy savings of 8.1%; the false alarm rate of the twin simulation was <0.5%, and no actual over-temperature incidents occurred.

[0092] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0094] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. A data center water-loop heat pump air conditioning system, characterized in that: include: The multi-source heterogeneous data acquisition module is used to connect with the sensor network deployed in the data center through industrial communication protocols and acquire multi-source heterogeneous operating data in real time. The feature fusion preprocessing module is connected to the multi-source heterogeneous data acquisition module and is used to receive the multi-source heterogeneous running data and construct a spatiotemporal feature matrix. A multi-scale prediction engine module, connected to the feature fusion preprocessing module, is used to receive the spatiotemporal feature matrix and output a final predicted temperature sequence. The dynamic optimization decision module, connected to the multi-scale prediction engine module, is used to solve a constrained optimization problem within a preset rolling time domain, with the objectives of minimizing the deviation between the final predicted temperature sequence and the preset target temperature and minimizing the total energy consumption of the system, thereby generating a set of optimal control commands. The collaborative execution control module, connected to the dynamic optimization decision module, is used to receive the optimal control command and distribute it to the physical device.

2. The data center water-loop heat pump air conditioning system according to claim 1, characterized in that: The multi-source heterogeneous operating data includes at least: the cabinet air inlet temperature sequence, the heat pump water outlet temperature sequence, the real-time cooling water flow rate, the water pump operating power, the water valve opening degree, the heat pump compressor operating frequency, and the cabinet real-time power consumption.

3. The data center water-loop heat pump air conditioning system according to claim 2, characterized in that: The feature fusion preprocessing module includes: The delay compensation factor calculation unit is used to calculate the water system delay compensation factor based on the received real-time flow rate of the cooling water and the pre-stored cooling water pipeline system length parameters. The feature matrix construction unit is used to apply the water system delay compensation factor to correct the time dimension information of the real-time cooling water flow data, and integrate the corrected real-time cooling water flow data with the cabinet air inlet temperature sequence and the cabinet real-time power consumption data into a multi-dimensional feature vector sequence including multiple historical time steps, thus forming the spatiotemporal feature matrix.

4. The data center water-loop heat pump air conditioning system according to claim 3, characterized in that: The multi-scale prediction engine module includes: The short-term forecasting submodule includes a first forecasting model for generating a first forecasting temperature sequence in response to short-term fluctuations in IT load. The long-term forecasting submodule includes a second forecasting model for generating a second predicted temperature sequence that reflects the long-term trend and cycle of temperature. The dynamic fusion unit is used to first calculate the time change rate of the real-time power consumption data of the cabinet, then determine a dynamic confidence weight based on the time change rate, and use the dynamic confidence weight to perform a weighted summation operation on the first predicted temperature sequence and the second predicted temperature sequence to obtain the final predicted temperature sequence.

5. The data center water-loop heat pump air conditioning system according to claim 4, characterized in that: The collaborative execution control module adopts a valve-compressor decoupling control strategy, which includes: for water valves, using a PID controller based on the current temperature deviation for independent basic adjustment to maintain stable water flow; for heat pump compressors, using a control law based on feedforward compensation, which takes the final predicted temperature sequence output by the multi-scale prediction engine module and the IT load change rate calculated by the feature fusion preprocessing module as inputs to directly calculate the forward-looking target compressor frequency to proactively respond to upcoming heat load changes.

6. A control method for a data center water-loop heat pump air conditioning system, characterized in that: Includes the following steps: S1. Data Acquisition: Real-time acquisition of multi-source heterogeneous operational data within the data center via a preset communication protocol. This multi-source heterogeneous operational data includes at least: time-series temperature data characterizing the thermal environment of the server racks, including the temperature sequence at the server rack air inlets; water system parameter data characterizing the cooling water circuit status, including real-time cooling water flow rate and water pump operating power; equipment status data characterizing the status of the cooling equipment, including water circuit valve opening and heat pump compressor operating frequency; and IT load data characterizing the computing task intensity of the data center, including the real-time power consumption of the server racks. S2. Feature preprocessing: Processing multi-source heterogeneous operational data to construct a spatiotemporal feature matrix for temperature prediction. The feature preprocessing steps include: calculating a water system delay compensation factor to compensate for physical transport delays in the water system based on the real-time cooling water flow rate and the preset cooling water pipeline system length; and performing timestamp alignment processing on the water system parameter data based on the water system delay compensation factor, and combining the aligned water system parameter data with temperature time series data and IT load data to generate a spatiotemporal feature matrix with multiple time steps and temporal and spatial coupling characteristics. S3. Multi-scale fusion prediction: Based on the spatiotemporal feature matrix, the final predicted temperature sequence within the future control time domain is generated. The multi-scale fusion prediction steps include: inputting the spatiotemporal feature matrix in parallel to the short-term temperature prediction submodule and the long-term temperature prediction submodule; the short-term temperature prediction submodule uses a first prediction model to generate a first predicted temperature sequence to respond to short-term drastic fluctuations in IT load data; the long-term temperature prediction submodule uses a second prediction model to generate a second predicted temperature sequence to capture the long-term trend and periodicity of temperature time series data; calculating the load change rate, which characterizes the drastic fluctuation of IT load data; determining the dynamic confidence weights for fusing the first and second predicted temperature sequences based on the load change rate; and weighting and fusing the first and second predicted temperature sequences according to the dynamic confidence weights to obtain the final predicted temperature sequence. S4. Optimize decision-making: Based on the final predicted temperature sequence and the preset target temperature, solve the multi-objective optimization problem under the preset equipment operation constraints to generate a set of optimal control instructions. The optimal control instruction set includes at least the target compressor frequency and the target valve opening. S5. Digital Twin Pre-simulation Verification: The optimal control command set generated by the optimization decision-making steps is first input into a pre-constructed digital twin model. The digital twin model is a virtual simulation environment that integrates a device thermodynamic model based on physical laws and a behavioral model trained based on historical data. The optimal control command set is simulated and executed in the digital twin model, and its possible virtual temperature response and virtual energy consumption results are predicted. It is determined whether the virtual temperature response meets the preset safe temperature threshold. Only when the virtual temperature response meets the safe temperature threshold is the execution of the command authorized, and the optimal control command set is distributed to the physical device. S6. Command execution: Distribute the optimal control command set to the heat pump compressor and water valves to execute the control.

7. The control method for a data center water-loop heat pump air conditioning system according to claim 6, characterized in that: In step S2, the method for calculating the water system delay compensation factor is as follows: τ=f(G w ,L pipe ): Among them, L pipe G is the total length of the cooling water piping system. w This represents the real-time flow rate of the cooling water collected at the current moment. The construction of the spatiotemporal feature matrix specifically includes: concatenating the cabinet air inlet temperature sequence and cabinet real-time power consumption sequence, which include the current time and historical time, with water system parameter data with timestamp k-τ after delay compensation factor compensation, to form a multi-dimensional input vector.

8. The control method for a data center water-loop heat pump air conditioning system according to claim 7, characterized in that: In step S3, the dynamic confidence weights are calculated using the Sigmoid function, with the following formula: in, γ is the absolute value of the rate of change of real-time power consumption of IT load with respect to time, and γ is a preset coefficient used to adjust the weight sensitivity.

9. The control method for a data center water-loop heat pump air conditioning system according to claim 8, characterized in that: In step S3, the first prediction model used in the short-term temperature prediction submodule is a long short-term memory network model, which is used to receive the spatiotemporal feature matrix as input and output the point-by-point temperature prediction value within the future preset short-term time window. The second prediction model used in the long-term temperature prediction submodule is the Prophet Algorithm Model, which is used to fit historical hourly average temperature data into trend, periodic and holiday effect terms to output the predicted temperature trend value within a preset long-term time window.

10. The control method for a data center water-loop heat pump air conditioning system according to claim 9, characterized in that: In step S4, the multi-objective optimization problem is solved within a rolling time-domain optimization framework, which defines a finite future time domain at the beginning of each control cycle; the solver used to solve the multi-objective optimization problem is the particle swarm optimization algorithm. The objective function of the optimization problem specifically includes a temperature deviation penalty term and an energy consumption penalty term. The temperature deviation penalty term is proportional to the square integral of the difference between the final predicted temperature sequence and the target temperature, and the energy consumption penalty term is the weighted sum of the operating power of the heat pump compressor and the operating power of the water pump. The preset equipment operating constraints include: the operating frequency of the heat pump compressor must be between the preset minimum frequency and the preset maximum frequency, and the absolute value of its frequency change rate must not exceed the preset maximum frequency conversion rate.

Images