A method and system for intelligent optimization control of data center cooling systems

Abstract

This invention belongs to the field of cooling system optimization control technology, and particularly relates to an intelligent optimization control method and system for data center cooling systems. By setting a spatiotemporal prediction model for cooling demand, the method extracts the demand fluctuation rate of each cooling zone, completing a dynamic assessment of the thermal stability of each zone. By setting an online calculation mechanism for thermal coupling strength based on the correlation between physical location and historical temperature response, a zone coupling group and thermal coupling topology are constructed, completing the quantitative identification of the thermal influence relationship between zones. By setting an upper-level collaborative control decision module, stability conditions are dynamically judged based on the thermal stability assessment results and the thermal coupling topology, completing the generation of global collaborative control commands or hierarchical decoupling control commands. By setting a hierarchical control strategy with inertial time decoupling, the upper-level decoupling controller tracks low-frequency trend loads, while the lower-level decoupling controller responds quickly to real-time demand, achieving an adaptive balance between optimal global energy efficiency and rapid local response.

Description

Technical Field

[0001] This invention belongs to the field of cooling system optimization and control technology, specifically relating to an intelligent optimization and control method and system for a data center cooling system. Background Technology

[0002] Currently, data center cooling system control mainly adopts the following methods: First, PID control based on fixed setpoints relies on manual experience to set cooling parameters, which cannot be dynamically adjusted according to real-time load, resulting in a large amount of energy waste; Second, rule-based group control strategies, such as changing the temperature setpoints of each air conditioner in turn to obtain sensor responses and build temperature change correlation models, but the control granularity is relatively coarse and it is difficult to cope with complex and ever-changing thermal environments; Third, data-driven methods use machine learning or neural networks to build temperature prediction models to guide cooling regulation.

[0003] In recent years, with the rapid development of artificial intelligence technology, some research has begun to attempt to apply intelligent algorithms such as reinforcement learning to data center cooling control. However, existing technologies generally have the following shortcomings:

[0004] First, existing AI optimization methods typically adopt a unified global optimization architecture, which does not take into account the spatiotemporal differences in the thermal characteristics of each cooling zone in the data center. In actual operation, the server load fluctuation characteristics in different areas are significantly different. Some zones have stable loads, while others fluctuate frequently due to changes in business characteristics. When the load of some zones fluctuates drastically, global optimization will force the cooling source side equipment to frequently adjust its operating parameters to respond to the fluctuations in end-user demand, which will not only exacerbate the risk of equipment oscillation but also reduce the overall energy efficiency of the system.

[0005] Secondly, existing methods generally neglect the assessment of thermal coupling relationships between cooling zones, especially the potential risk of coupling resonance between fluctuating zones. In high-density data centers, the thermal environments of adjacent zones influence each other. If there is strong thermal coupling between two fluctuating zones, their temperature responses may overlap or even resonate, leading to local thermal environment runaway. Existing technologies do not yet address the identification and prevention of this risk.

[0006] Third, existing methods have a single control mode and lack the ability to dynamically switch control strategies based on the actual thermal state of the system. Whether it is global optimization or independent zone control, it is impossible to achieve an adaptive balance between global collaborative energy efficiency and local hierarchical rapid response. The root cause is the lack of online quantitative evaluation of the thermal stability of each zone and the strength of mutual thermal coupling, which cannot provide a decision basis for the dynamic selection of control strategies.

[0007] To address the aforementioned issues, this application presents an intelligent optimization control method and system for data center cooling systems. Summary of the Invention

[0008] To address the shortcomings of existing technologies mentioned in the background section, this application proposes an intelligent optimization control method and system for data center cooling systems. This invention establishes a spatiotemporal prediction model for cooling demand, extracts the demand volatility of each cooling zone, and dynamically assesses the thermal stability of each zone. It constructs a zone coupling group and thermal coupling topology map by setting an online calculation mechanism based on the correlation between physical location and historical temperature response, thus quantitatively identifying the thermal impact relationships between zones. An upper-level collaborative control decision module dynamically judges stability conditions based on the thermal stability assessment results and the thermal coupling topology, generating global collaborative control commands or hierarchical decoupling control commands. A hierarchical control strategy with inertial time decoupling is implemented, where the upper-level decoupling controller tracks low-frequency trend loads, and the lower-level decoupling controller responds quickly to real-time demand, achieving an adaptive balance between optimal global energy efficiency and rapid local response. Finally, a graded safety protection and smooth transition mechanism is established to achieve seamless switching between optimized control and safety protection, thereby solving the problems in the background section.

[0009] Firstly, to achieve the above objectives, this application provides an intelligent optimization control method for a data center cooling system, which includes the following specific steps:

[0010] Step S1: Collect real-time operational data related to the data center cooling system. The operational data related to the system includes server load power, air inlet and outlet temperatures of each rack, outdoor ambient temperature and humidity data, and operating parameters of each precision air conditioner and cooling room equipment. Clean, align, and normalize the real-time collected data from each dimension to construct a standardized time-series operational dataset.

[0011] Step S2: Construct a spatiotemporal prediction model for cooling demand that integrates convolutional neural networks and long short-term memory networks based on the time-series running dataset. Extract and output the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone within a preset future time period, as well as the mean, peak, valley and demand volatility of cooling demand for each cooling zone determined based on the prediction sequence.

[0012] Step S3: Based on the demand volatility output in step S2, perform thermal stability assessment on each cooling zone. Cooling zones with demand volatility less than a preset stability threshold are classified as first-class stable zones, and cooling zones with demand volatility greater than or equal to the preset stability threshold are classified as second-class volatile zones. At the same time, based on the correlation between the physical location of each cooling zone and its historical temperature response, calculate the thermal coupling strength value between any two cooling zones. Cooling zones with thermal coupling strength values ​​greater than a preset coupling threshold are assigned to the same zone coupling group, and a thermal coupling topology diagram describing the thermal coupling relationship between each cooling zone and between zone coupling groups is constructed.

[0013] Step S4: Input the thermal stability assessment results and thermal coupling topology diagram generated in step S3 into the upper-level collaborative control decision module. The upper-level collaborative control decision module traverses all second-type fluctuation partitions and performs the following judgment for each second-type fluctuation partition:

[0014] First, check whether the partition coupling group containing the second type of fluctuation partition contains a first type of stable partition;

[0015] If not, it is further determined whether there is at least one other second-type fluctuation partition in the partition coupling group, and the thermal coupling strength value between the second-type fluctuation partition and at least one other second-type fluctuation partition is greater than the preset resonance trigger threshold. If so, it is determined that the second-type fluctuation partition does not meet the stability condition, and it and the corresponding other second-type fluctuation partition are marked as a coupled resonance risk partition pair. If not, it is determined that it meets the stability condition.

[0016] If a first-class stable partition is included, the maximum value of each thermal coupling strength between the second-class fluctuating partition and all first-class stable partitions in the group is obtained. If the maximum value is greater than the preset decoupling trigger threshold, it is determined that it does not meet the stability condition; otherwise, it is determined that it meets the stability condition.

[0017] When there is no second type of fluctuation partition, a global collaborative control command is generated directly. When there is a second type of fluctuation partition and all second type of fluctuation partitions meet the stability conditions, a global collaborative control command is generated. Otherwise, a hierarchical decoupling control command is generated and a linkage control identifier is attached.

[0018] Step S5: When it is a global collaborative control command, the deep reinforcement learning agent performs global optimization with the global cooling energy efficiency ratio as the optimization target, and outputs the set values ​​of the operating parameters of each cooling device.

[0019] When the hierarchical decoupling control command is executed, the hierarchical control strategy of inertial time decoupling is implemented, and the upper-level decoupling controller and the lower-level decoupling controller work together to output the set values ​​of the operating parameters of each refrigeration device.

[0020] Step S6: Send the set values ​​of the operating parameters of each refrigeration device generated in step S5 to the bottom controller of the refrigeration system for execution, and execute steps S1 to S5 in a cycle with a preset control period to achieve closed-loop optimization control.

[0021] Step S7: Continuously monitor the air intake temperature and its rate of change of each cabinet. When the temperature or rate of temperature rise exceeds the limit, trigger the graded protection, pause the current optimized control and switch to conservative control mode. Optimized control will be resumed after the temperature returns to normal and the system status is verified.

[0022] Based on the above scheme, the preferred option is that the demand volatility rate in step S3 is the ratio of the standard deviation of the predicted cooling demand for each refrigeration zone in the future preset period to the mean cooling demand in that period.

[0023] The thermal coupling strength value in step S3 is calculated by combining the linear correlation coefficient and the nonlinear coupling index. The thermal coupling topology graph uses each cooling zone as a node and the thermal coupling strength value as the weight of the edge connecting the corresponding node.

[0024] Based on the above scheme, the preferred offline optimization method for the preset coupling threshold, preset resonance trigger threshold and preset decoupling trigger threshold is as follows: obtain the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone and the corresponding actual control effect data of the cooling system in the historical operation data, take the minimum weighted sum of energy consumption and temperature over-limit times of the cold source side equipment as the optimization objective, and use the grid search method or Bayesian optimization method to determine the preset coupling threshold, preset resonance trigger threshold and preset decoupling trigger threshold offline;

[0025] After offline optimization is determined, it is automatically updated online every preset period based on actual operating data. During online update, the thermal coupling strength value between each cooling zone is recalculated based on the newly accumulated operating data in the current period, and the thresholds are updated accordingly.

[0026] When a sudden change in system load is detected that exceeds the preset load change threshold or when system equipment is replaced, an emergency update of the threshold is triggered.

[0027] Based on the above scheme, step S5 specifically includes the following steps:

[0028] Step S51: When a global collaborative control command is received, the setpoints of the outlet air temperature and supply air volume of each precision air conditioner, the setpoint of the chilled water outlet temperature, the frequency of the cooling tower fan, the frequency of the chilled water pump, and the frequency of the cooling water pump are used as the joint action vector. The deep reinforcement learning agent performs global optimization with the global cooling energy efficiency ratio as the optimization target, and outputs the setpoints of the operating parameters of each refrigeration device. The deep reinforcement learning agent adopts a three-stage deployment mode of offline pre-training, simulation verification, and online low exploration rate fine-tuning. The online fine-tuning stage sets up a manual intervention interface.

[0029] Step S52: When a hierarchical decoupling control command is received, execute the hierarchical control strategy of inertial-time decoupling:

[0030] The upper-level decoupling controller sets the corresponding low-pass filter cutoff frequency according to the demand volatility of each second-type fluctuation zone. The larger the demand volatility of the zone, the lower the cutoff frequency. The cooling demand prediction sequence of each zone is filtered to extract the cooling trend component. The sum of the cooling trend components of all second-type fluctuation zones is taken as the total low-frequency trend load. The set value of chilled water outlet temperature, cooling tower fan frequency, chilled water pump frequency and cooling water pump frequency are used as action variables. The goal is to minimize the total energy consumption of the cold source side equipment in the control time domain. Under the constraints of satisfying the supply of total low-frequency trend load and the air inlet temperature of all cabinets not exceeding the first safe temperature threshold, the upper-level optimal control command is solved. The control cycle is the first preset cycle.

[0031] The lower-level decoupling controller uses the precision air conditioning outlet temperature setpoint and supply air volume setpoint corresponding to each second-type fluctuation zone as action variables, and uses the predicted value of the current control moment in the cooling demand prediction sequence of the zone output in step S2 as the tracking target. Under the constraint that the air inlet temperature of all cabinets in the zone does not exceed the first safe temperature threshold, the controller solves the lower-level optimal control command. The control cycle is a second preset cycle shorter than the first preset cycle. When a linkage control flag is received, the controller performs linkage optimization on the coupled resonance risk zone pair, and performs coordinated control with the minimum comprehensive temperature deviation and the minimum total energy consumption of the two zones according to preset weights as the joint objective.

[0032] For all Class I stable zones, the average cooling demand of that zone is used as the feedforward input. The deviation between the actual inlet air temperature and the preset target temperature is combined to perform proportional-integral feedback correction and lag compensation for the feedback deviation, and the set value of the corresponding precision air conditioner is calculated.

[0033] Among them, both the first safe temperature threshold and the preset target temperature are lower than the temperature threshold for triggering graded protection in step S7.

[0034] Based on the above scheme, the preferred embodiment of the deep reinforcement learning agent in step S51 employs a constrained reinforcement learning algorithm, with the reward function being:

[0035]

[0036] in, It is the reciprocal of the global cooling energy efficiency ratio. Let be the air intake temperature of the i-th rack. The soft-constraint reference temperature is less than or equal to the first safe temperature threshold of sub-step S52. These are the combined action vectors formed by the setpoints of the outlet air temperature, supply air volume, chilled water outlet temperature, cooling tower fan frequency, chilled water pump frequency, and cooling water pump frequency for the current and previous control cycles, respectively. The L2 norm square of the change in action between adjacent control cycles, α, β, and γ are preset non-negative weighting coefficients, and M is the total number of cabinets. Actions that violate the minimum start-stop interval, maximum number of start-stop cycles, or upper limit of pump frequency during training are judged as invalid and discarded.

[0037] In a preferred embodiment based on the above scheme, both the upper-level decoupling controller and the lower-level decoupling controller in step S52 adopt a model predictive control framework, and the constraints also include the adjustment range constraints of each action variable.

[0038] Step S52 also includes a fault redundancy strategy: when a fault or communication interruption is detected in the upper decoupling controller, the lower decoupling controller automatically switches to independent control mode and independently solves the precision air conditioning control command based on the real-time cooling demand forecast value and temperature constraints of this zone. After the upper decoupling controller recovers, the hierarchical collaborative control is restored.

[0039] Based on the above-mentioned preferred scheme, step S7 specifically includes the following steps:

[0040] Step S71: When the rate of increase of the air intake temperature of any cabinet exceeds the preset rate threshold, pre-protection is triggered, and the air volume of the precision air conditioner in the zone where the cabinet is located is increased to the preset value.

[0041] Step S72: A second safe temperature threshold is preset, and the first safe temperature threshold is lower than the preset second safe temperature threshold preset level difference. When the air inlet temperature of any cabinet exceeds the second safe temperature threshold, the strategy output of the currently active controller in step S5 is immediately paused, and all precision air conditioner outlet air temperature setting values ​​and supply air volume setting values ​​are adjusted from the current optimized values ​​to preset conservative values ​​along a preset smooth transition curve. The preset conservative value is the preset ratio of the outlet air temperature setting to the preset low temperature value and the supply air volume setting to the maximum rated value.

[0042] Step S73: After the air intake temperature of all cabinets returns to the normal range, verify the operating status of each actuator and the effectiveness of the temperature control logic. After confirming that it is normal, maintain the conservative control stability for a preset time, and then transition the control parameters from the conservative value back to the optimized value along the preset smooth recovery curve, and restore the optimized control of step S5.

[0043] Secondly, this application provides an intelligent optimization control system for a data center cooling system, which specifically includes: a data acquisition and preprocessing module, a cooling demand spatiotemporal prediction module, a spatiotemporal feature decoupling analysis module, an upper-level collaborative control decision module, a hierarchical collaborative optimization execution module, a collaborative regulation execution module, and a security protection module.

[0044] The data acquisition and preprocessing module is used to collect server load power, air inlet and outlet temperature of each rack, outdoor ambient temperature and humidity data, and operating parameters of each precision air conditioner and refrigeration room equipment in real time, and to clean, align and normalize the collected data to build a standardized time-series running dataset.

[0045] The input end of the cooling demand spatiotemporal prediction module is connected to the output end of the data acquisition and preprocessing module. It has a built-in cooling demand spatiotemporal prediction model that integrates convolutional neural networks and long short-term memory networks. It is used to extract and output the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone within a future preset time period, as well as the mean, peak, valley and demand volatility of cooling demand for each cooling zone determined based on the prediction sequence.

[0046] The spatiotemporal feature decoupling analysis module has its input connected to the output of the cooling demand spatiotemporal prediction module, and is used to perform the following operations:

[0047] Based on the comparison between demand volatility and preset stability threshold, each cooling zone is divided into either a first-class stable zone or a second-class volatile zone.

[0048] Based on the correlation between the physical location of each cooling zone and its historical temperature response, the thermal coupling strength value between any two cooling zones is calculated, and cooling zones with thermal coupling strength values ​​greater than a preset coupling threshold are grouped into the same zone coupling group to construct a thermal coupling topology map.

[0049] The upper-level collaborative control decision module has its input end connected to the output end of the spatiotemporal feature decoupling analysis module. It is used to receive the partition type division and thermal coupling topology map, and according to the logic described in step S4, traverse the second type of fluctuation partition, determine whether it meets the stability conditions, and generate a global collaborative control command or a hierarchical decoupling control command carrying a linkage control identifier.

[0050] The input end of the hierarchical collaborative optimization execution module is connected to the output end of the upper-level collaborative control decision module. When receiving a global collaborative control command, it calls the built-in deep reinforcement learning agent to perform optimization with the global cooling energy efficiency ratio as the optimization target and outputs the set values ​​of the operating parameters of each cooling device.

[0051] When a hierarchical decoupling control command is received, the built-in upper-level decoupling controller is invoked. Based on the demand volatility of the second type of fluctuation zone, the low-pass filter cutoff frequency is set and the cooling capacity trend component is extracted. The upper-level optimal command is solved with the goal of minimizing the total energy consumption of the cold source side equipment. At the same time, the lower-level decoupling controller is invoked. The lower-level optimal command is solved with the real-time value of the cooling capacity demand prediction sequence of the zone as the tracking target. When carrying the linkage control flag, the linkage optimization is performed on the coupled resonance risk zone pair to output the set values ​​of the operating parameters of each refrigeration equipment.

[0052] The input end of the collaborative control execution module is connected to the output end of the hierarchical collaborative optimization execution module, and the output end is connected to the bottom controller of the refrigeration system. It is used to send down the received operating parameter set values ​​of each refrigeration device for execution.

[0053] The safety protection module is used to continuously monitor the air intake temperature of each cabinet and its rate of change. Its output is connected to the enable terminal of the collaborative control execution module. When the temperature or temperature rise rate exceeds the limit, it triggers an enable signal to pause the current optimized control and switch to the conservative control mode. After the status verification is passed, the optimized control is restored.

[0054] Thirdly, this application provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0055] The processor executes the aforementioned intelligent optimization control method for a data center cooling system by calling a computer program stored in the memory.

[0056] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned intelligent optimization control method for a data center cooling system.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] First, this invention constructs a progressive intelligent control architecture that combines spatiotemporal prediction of cooling demand, decoupling analysis of spatiotemporal features, and hierarchical collaborative optimization. Based on the differences in thermal stability and thermal coupling relationships of each cooling zone, it dynamically decides on a hierarchical control strategy of global collaborative control or inertial time decoupling. In the hierarchical control mode, the upper-level decoupling controller tracks the sum of the cooling trend components of each fluctuating zone over a longer period, while the lower-level decoupling controller responds to the real-time cooling demand fluctuations of each zone over a shorter period. This enables the fast-responding terminal precision air conditioner to handle high-frequency fluctuations and the cold source system with large inertia to handle low-frequency trends, effectively avoiding frequent oscillations of cold source-side equipment caused by global unified optimization, and improving the overall operating energy efficiency and stability of the refrigeration system.

[0059] Secondly, by constructing a thermally coupled topology diagram, this invention not only evaluates the thermal coupling relationship between fluctuating and stable partitions, but also introduces a coupling resonance risk identification mechanism between fluctuating partitions. When the thermal coupling strength value between two fluctuating partitions exceeds the preset resonance trigger threshold, they are marked as a pair of coupling resonance risk partitions. The joint optimization of the lower-level decoupling controller is triggered by the linkage control identifier. The joint objective is to minimize the weighted comprehensive temperature deviation of the two partitions and the total energy consumption, thereby effectively preventing local thermal runaway caused by coupling resonance and improving the comprehensiveness and safety of control decisions.

[0060] Third, in the upper-level decoupling controller of this invention, the low-pass filter cutoff frequency is adaptively set according to the magnitude of the demand volatility of each fluctuation zone. The larger the demand volatility of the zone, the lower the cutoff frequency is, which makes the cooling trend component smoother and reduces the adjustment requirements of the cold source side equipment. This adaptive mechanism enables the time scale division of the hierarchical control to be dynamically matched with the actual thermal characteristics, avoiding the problem of over-decoupling or under-decoupling caused by a fixed cutoff frequency.

[0061] Fourth, this invention uses an upper-level collaborative control decision module to sequentially check whether each second-type fluctuation partition contains a first-type stable partition or other strongly coupled fluctuation partitions within its partition coupling group. Based on the judgment results, it dynamically generates global collaborative control instructions or hierarchical decoupling control instructions. When there are no second-type fluctuation partitions in the system, global collaborative control instructions are directly generated, ensuring that the decision logic is complete even under extreme conditions and without control blind spots. Furthermore, this invention sets up a two-level temperature protection system with a first safety temperature threshold and a second safety temperature threshold. A preset difference between the two thresholds forms a safety buffer, and a temperature rise rate pre-protection mechanism is introduced to intervene in advance in rapid temperature rise scenarios. When safety protection is triggered, a smooth transition curve is used to adjust the control parameters to avoid the impact of sudden changes in control parameters on the refrigeration system. Before restoring optimized control, the operating status of each actuator and the effectiveness of the temperature control logic must be verified, which can effectively prevent secondary over-temperature. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a schematic diagram of the overall process of an intelligent optimization control method for a data center cooling system according to the present invention;

[0064] Figure 2 This is a flowchart of step S5 in the intelligent optimization control method for a data center cooling system of the present invention;

[0065] Figure 3 This is a schematic diagram of the framework of an intelligent optimization control system for a data center cooling system according to the present invention. Detailed Implementation

[0066] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0067] Example 1

[0068] To address the technical problems raised in the background art, this application provides a preferred embodiment: such as Figures 1-3 As shown, an intelligent optimization control method for a data center cooling system includes the following specific steps:

[0069] Step S1: Collect real-time operational data related to the data center cooling system. The operational data related to the system includes server load power, air inlet and outlet temperatures of each rack, outdoor ambient temperature and humidity data, and operating parameters of each precision air conditioner and cooling room equipment. Clean, align, and normalize the real-time collected data from each dimension to construct a standardized time-series operational dataset.

[0070] Step S2: Construct a spatiotemporal prediction model for cooling demand that integrates convolutional neural networks and long short-term memory networks based on the time-series running dataset. Extract and output the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone within a preset future time period, as well as the mean, peak, valley and demand volatility of cooling demand for each cooling zone determined based on the prediction sequence.

[0071] Step S3: Based on the demand volatility output in step S2, perform thermal stability assessment on each cooling zone. Cooling zones with demand volatility less than a preset stability threshold are classified as first-class stable zones, and cooling zones with demand volatility greater than or equal to the preset stability threshold are classified as second-class volatile zones. At the same time, based on the correlation between the physical location of each cooling zone and its historical temperature response, calculate the thermal coupling strength value between any two cooling zones. Cooling zones with thermal coupling strength values ​​greater than a preset coupling threshold are assigned to the same zone coupling group, and a thermal coupling topology diagram describing the thermal coupling relationship between each cooling zone and between zone coupling groups is constructed.

[0072] Step S4: Input the thermal stability assessment results and thermal coupling topology diagram generated in step S3 into the upper-level collaborative control decision module. The upper-level collaborative control decision module traverses all second-type fluctuation partitions and performs the following judgment for each second-type fluctuation partition:

[0073] First, check whether the partition coupling group containing the second type of fluctuation partition contains a first type of stable partition;

[0074] If not, it is further determined whether there is at least one other second-type fluctuation partition in the partition coupling group, and the thermal coupling strength value between the second-type fluctuation partition and at least one other second-type fluctuation partition is greater than the preset resonance trigger threshold. If so, it is determined that the second-type fluctuation partition does not meet the stability condition, and it and the corresponding other second-type fluctuation partition are marked as a coupled resonance risk partition pair. If not, it is determined that it meets the stability condition.

[0075] If a first-class stable partition is included, the maximum value of each thermal coupling strength between the second-class fluctuating partition and all first-class stable partitions in the group is obtained. If the maximum value is greater than the preset decoupling trigger threshold, it is determined that it does not meet the stability condition; otherwise, it is determined that it meets the stability condition.

[0076] When there is no second type of fluctuation partition, a global collaborative control command is generated directly. When there is a second type of fluctuation partition and all second type of fluctuation partitions meet the stability conditions, a global collaborative control command is generated. Otherwise, a hierarchical decoupling control command is generated and a linkage control identifier is attached.

[0077] Step S5: When it is a global collaborative control command, the deep reinforcement learning agent performs global optimization with the global cooling energy efficiency ratio as the optimization target, and outputs the set values ​​of the operating parameters of each cooling device.

[0078] When the hierarchical decoupling control command is executed, the hierarchical control strategy of inertial time decoupling is implemented, and the upper-level decoupling controller and the lower-level decoupling controller work together to output the set values ​​of the operating parameters of each refrigeration device.

[0079] Step S6: Send the set values ​​of the operating parameters of each refrigeration device generated in step S5 to the bottom controller of the refrigeration system for execution, and execute steps S1 to S5 in a cycle with a preset control period to achieve closed-loop optimization control.

[0080] It should be noted that step S6 also includes a closed-loop robustness enhancement mechanism: when the sensor fails or data acquisition is interrupted, alternative estimated data is generated based on the spatiotemporal prediction model of cooling demand in step S2 to maintain the control closed loop. Every preset evaluation cycle, the deviation between the actual cooling energy efficiency ratio of the system and the expected value is compared. If the deviation exceeds the preset threshold, the control parameters are automatically corrected.

[0081] Step S7: Continuously monitor the air intake temperature and its rate of change of each cabinet. When the temperature or rate of temperature rise exceeds the limit, trigger the graded protection, pause the current optimized control and switch to conservative control mode. Optimized control will be resumed after the temperature returns to normal and the system status is verified.

[0082] The advantages of this embodiment compared to the prior art are as follows: This invention constructs a progressive intelligent control architecture of spatiotemporal prediction of cooling demand, spatiotemporal feature decoupling analysis, and hierarchical collaborative optimization execution. Based on the differences in thermal stability and thermal coupling relationships of each cooling zone, it dynamically decides on a hierarchical control strategy of global collaborative control or inertial time decoupling. In the hierarchical control mode, the upper-level decoupling controller tracks the sum of cooling trend components of each fluctuating zone with a longer period, while the lower-level decoupling controller responds to the real-time cooling demand fluctuations of each zone with a shorter period. This allows the fast-responding terminal precision air conditioner to handle high-frequency fluctuations, and the cold source system with large inertia to handle low-frequency trends, effectively avoiding the cold source issues caused by global unified optimization. The frequent oscillations of the side equipment improve the overall operating efficiency and stability of the refrigeration system. Furthermore, by constructing a thermal coupling topology diagram, this invention not only evaluates the thermal coupling relationship between fluctuating and stable zones, but also introduces a coupling resonance risk identification mechanism between fluctuating zones. When the thermal coupling strength value between two fluctuating zones exceeds the preset resonance trigger threshold, they are marked as a coupling resonance risk zone pair. The joint optimization of the lower-level decoupling controller is triggered by the linkage control identifier. The joint objective of minimizing the weighted comprehensive temperature deviation of the two zones and minimizing the total energy consumption is used for coordinated control, which can effectively prevent local thermal runaway caused by coupling resonance and improve the comprehensiveness and safety of control decisions.

[0083] Furthermore:

[0084] In an optional embodiment, the demand volatility rate in step S3 is the ratio of the standard deviation of the predicted cooling demand for each refrigeration zone over a future preset period to the mean cooling demand for that period.

[0085] The thermal coupling strength value in step S3 is calculated by combining the linear correlation coefficient and the nonlinear coupling index. The thermal coupling topology graph uses each cooling zone as a node and the thermal coupling strength value as the weight of the edge connecting the corresponding node.

[0086] In an optional embodiment, the offline optimization method for the preset coupling threshold, preset resonance trigger threshold, and preset decoupling trigger threshold is as follows: obtain the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone and the corresponding actual control effect data of the cooling system in the historical operation data, take the minimum weighted sum of energy consumption and temperature over-limit times of the cold source side equipment as the optimization objective, and use the grid search method or Bayesian optimization method to offline optimize and determine the preset coupling threshold, preset resonance trigger threshold, and preset decoupling trigger threshold.

[0087] After offline optimization is determined, it is automatically updated online every preset period based on actual operating data. During online update, the thermal coupling strength value between each cooling zone is recalculated based on the newly accumulated operating data in the current period, and the thresholds are updated accordingly.

[0088] When a sudden change in system load is detected that exceeds the preset load change threshold or when system equipment is replaced, an emergency update of the threshold is triggered.

[0089] In an optional embodiment, step S5 specifically includes the following steps:

[0090] Step S51: When a global collaborative control command is received, the setpoints of the outlet air temperature and supply air volume of each precision air conditioner, the setpoint of the chilled water outlet temperature, the frequency of the cooling tower fan, the frequency of the chilled water pump, and the frequency of the cooling water pump are used as the joint action vector. The deep reinforcement learning agent performs global optimization with the global cooling energy efficiency ratio as the optimization target, and outputs the setpoints of the operating parameters of each refrigeration device. The deep reinforcement learning agent adopts a three-stage deployment mode of offline pre-training, simulation verification, and online low exploration rate fine-tuning. The online fine-tuning stage sets up a manual intervention interface.

[0091] Step S52: When a hierarchical decoupling control command is received, execute the hierarchical control strategy of inertial-time decoupling:

[0092] The upper-level decoupling controller sets the corresponding low-pass filter cutoff frequency according to the demand volatility of each second-type fluctuation zone. The larger the demand volatility of the zone, the lower the cutoff frequency. The cooling demand prediction sequence of each zone is filtered to extract the cooling trend component. The sum of the cooling trend components of all second-type fluctuation zones is taken as the total low-frequency trend load. The set value of chilled water outlet temperature, cooling tower fan frequency, chilled water pump frequency and cooling water pump frequency are used as action variables. The goal is to minimize the total energy consumption of the cold source side equipment in the control time domain. Under the constraints of satisfying the supply of total low-frequency trend load and the air inlet temperature of all cabinets not exceeding the first safe temperature threshold, the upper-level optimal control command is solved. The control cycle is the first preset cycle.

[0093] The lower-level decoupling controller uses the precision air conditioning outlet temperature setpoint and supply air volume setpoint corresponding to each second-type fluctuation zone as action variables, and uses the predicted value of the current control moment in the cooling demand prediction sequence of the zone output in step S2 as the tracking target. Under the constraint that the air inlet temperature of all cabinets in the zone does not exceed the first safe temperature threshold, the controller solves the lower-level optimal control command. The control cycle is a second preset cycle shorter than the first preset cycle. When a linkage control flag is received, the controller performs linkage optimization on the coupled resonance risk zone pair, and performs coordinated control with the minimum comprehensive temperature deviation and the minimum total energy consumption of the two zones according to preset weights as the joint objective.

[0094] For all Class I stable zones, the average cooling demand of that zone is used as the feedforward input. The deviation between the actual inlet air temperature and the preset target temperature is combined to perform proportional-integral feedback correction and lag compensation for the feedback deviation, and the set value of the corresponding precision air conditioner is calculated.

[0095] Among them, both the first safe temperature threshold and the preset target temperature are lower than the temperature threshold for triggering graded protection in step S7.

[0096] In an optional embodiment, the deep reinforcement learning agent in step S51 employs a constrained reinforcement learning algorithm, with the reward function being:

[0097]

[0098] in, It is the reciprocal of the global cooling energy efficiency ratio. Let be the air intake temperature of the i-th rack. The soft-constraint reference temperature is less than or equal to the first safe temperature threshold of sub-step S52. These are the combined action vectors formed by the setpoints of the outlet air temperature, supply air volume, chilled water outlet temperature, cooling tower fan frequency, chilled water pump frequency, and cooling water pump frequency for the current and previous control cycles, respectively. The L2 norm square of the change in action between adjacent control cycles, α, β, and γ are preset non-negative weighting coefficients, and M is the total number of cabinets. Actions that violate the minimum start-stop interval, maximum number of start-stop cycles, or upper limit of pump frequency during training are judged as invalid and discarded.

[0099] It should be noted that the reward function consists of three weighted terms, each corresponding to a different control objective: The term-driven agent seeks the optimal combination of operating parameters for energy efficiency, directly minimizing power consumption per unit of cooling capacity; a temperature limit violation penalty term is also included. To ensure that intelligent agents pursue energy conservation without sacrificing data center temperature safety, any action that causes the rack intake air temperature to exceed the soft constraint reference temperature will be penalized; action smoothing penalty item. It suppresses large fluctuations in equipment operating parameters during adjacent control cycles, avoids mechanical stress and wear caused by frequent adjustments, and the three factors are flexibly prioritized through weighting coefficients α, β, and γ, so that the three objectives of energy saving, safety, and equipment protection are synergistically optimized under a unified mathematical framework.

[0100] In the reward function The soft constraint reference temperature is set at a value less than or equal to the first safe temperature threshold in the hierarchical control. This design provides a safety buffer for the deep reinforcement learning agent during global optimization. It generates a penalty signal before the cabinet temperature reaches the hard constraint boundary, guiding the strategy to avoid risks in advance, rather than passively responding after the temperature exceeds the limit. This is complemented by the hierarchical control using the first safe temperature threshold as a hard constraint for the optimization problem, and the second safe temperature threshold in the safety protection mechanism, forming a three-level progressive safety defense line of soft constraint early warning, hard constraint limitation, and emergency protection as a safety net.

[0101] The action smoothing penalty term uses the square of the L2 norm of the difference between the joint action vectors of adjacent control cycles to apply a greater penalty to large-amplitude adjustments, prompting the agent to generate smooth and gradual control command sequences. This design not only reduces equipment wear and tear, but also makes the control commands more feasible at the engineering execution level, avoiding the problem that the underlying actuator cannot track or overshoots due to excessive command jumps.

[0102] In addition to the soft constraint penalty in the reward function, this invention clearly defines actions that violate the minimum start-stop interval, maximum number of start-stops, or upper limit of pump frequency constraints during training as invalid and discarded. The reward function guides the agent to learn safety strategies autonomously, while hard constraints prevent dangerous actions from being output. The combination of the two ensures from the algorithm level that the control commands output by the agent are always within the safe operating range of the equipment.

[0103] Furthermore:

[0104] In an optional embodiment, both the upper-level decoupling controller and the lower-level decoupling controller in step S52 adopt a model predictive control framework, and the constraints also include the adjustment range constraints of each action variable.

[0105] Step S52 also includes a fault redundancy strategy: when a fault or communication interruption is detected in the upper decoupling controller, the lower decoupling controller automatically switches to independent control mode and independently solves the precision air conditioning control command based on the real-time cooling demand forecast value and temperature constraints of this zone. After the upper decoupling controller recovers, the hierarchical collaborative control is restored.

[0106] The advantages of this embodiment are as follows: In the upper-level decoupling controller, the low-pass filter cutoff frequency is adaptively set according to the magnitude of the demand volatility of each fluctuation zone. The larger the demand volatility of the zone, the lower the cutoff frequency is, which makes the cooling trend component smoother and reduces the adjustment requirements of the cold source side equipment. This adaptive mechanism enables the time scale division of the hierarchical control to be dynamically matched with the actual thermal characteristics, avoiding the problems of over-decoupling or under-decoupling caused by a fixed cutoff frequency.

[0107] In an optional embodiment, step S7 specifically includes the following steps:

[0108] Step S71: When the rate of increase of the air intake temperature of any cabinet exceeds the preset rate threshold, pre-protection is triggered, and the air volume of the precision air conditioner in the zone where the cabinet is located is increased to the preset value.

[0109] Step S72: A second safe temperature threshold is preset, and the first safe temperature threshold is lower than the preset second safe temperature threshold preset level difference. When the air inlet temperature of any cabinet exceeds the second safe temperature threshold, the strategy output of the currently active controller in step S5 is immediately paused, and all precision air conditioner outlet air temperature setting values ​​and supply air volume setting values ​​are adjusted from the current optimized values ​​to preset conservative values ​​along a preset smooth transition curve. The preset conservative value is the preset ratio of the outlet air temperature setting to the preset low temperature value and the supply air volume setting to the maximum rated value.

[0110] Step S73: After the air intake temperature of all cabinets returns to the normal range, verify the operating status of each actuator and the effectiveness of the temperature control logic. After confirming that it is normal, maintain the conservative control stability for a preset time, and then transition the control parameters from the conservative value back to the optimized value along the preset smooth recovery curve, and restore the optimized control of step S5.

[0111] The advantages of the above embodiments are as follows: by setting a two-level temperature protection system with a first safety temperature threshold and a second safety temperature threshold, a preset difference is set between the two thresholds to form a safety buffer, and a temperature rise rate pre-protection mechanism is introduced to intervene in advance in the case of rapid temperature rise. When the safety protection is triggered, a smooth transition curve is used to adjust the control parameters to avoid the impact of sudden changes in control parameters on the refrigeration system. Before restoring the optimized control, it is necessary to verify the operating status of each actuator and the effectiveness of the temperature control logic, which can effectively prevent secondary over-temperature.

[0112] Example 2

[0113] Based on the same inventive concept as in Embodiment 1, such as Figure 3 As shown, this embodiment provides an intelligent optimization control system for a data center cooling system, which specifically includes: a data acquisition and preprocessing module, a cooling demand spatiotemporal prediction module, a spatiotemporal feature decoupling analysis module, an upper-level collaborative control decision module, a hierarchical collaborative optimization execution module, a collaborative regulation execution module, and a security protection module;

[0114] The data acquisition and preprocessing module is used to collect server load power, air inlet and outlet temperature of each rack, outdoor ambient temperature and humidity data, and operating parameters of each precision air conditioner and refrigeration room equipment in real time. It also cleans, aligns and normalizes the collected data to build a standardized time-series running dataset.

[0115] The input end of the cooling demand spatiotemporal prediction module is connected to the output end of the data acquisition and preprocessing module. It has a built-in spatiotemporal prediction model for cooling demand that integrates convolutional neural networks and long short-term memory networks. It is used to extract and output the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone within a preset future time period, as well as the mean, peak, valley and demand volatility of cooling demand for each cooling zone determined based on the prediction sequence.

[0116] The spatiotemporal feature decoupling analysis module has its input connected to the output of the cooling demand spatiotemporal prediction module, and is used to perform the following operations:

[0117] Based on the comparison between demand volatility and preset stability threshold, each cooling zone is divided into either a first-class stable zone or a second-class volatile zone.

[0118] Based on the correlation between the physical location of each cooling zone and its historical temperature response, the thermal coupling strength value between any two cooling zones is calculated, and cooling zones with thermal coupling strength values ​​greater than a preset coupling threshold are grouped into the same zone coupling group to construct a thermal coupling topology map.

[0119] The upper-level collaborative control decision module has its input end connected to the output end of the spatiotemporal feature decoupling analysis module. It is used to receive the partition type division and thermal coupling topology diagram, and according to the logic of step S4, it traverses the second type of fluctuation partition, determines whether it meets the stability conditions, and generates global collaborative control instructions or hierarchical decoupling control instructions carrying linkage control identifiers.

[0120] The hierarchical collaborative optimization execution module has its input end connected to the output end of the upper-level collaborative control decision module. When it receives a global collaborative control command, it calls the built-in deep reinforcement learning agent to perform optimization with the global cooling energy efficiency ratio as the optimization target and outputs the set values ​​of the operating parameters of each cooling device.

[0121] When a hierarchical decoupling control command is received, the built-in upper-level decoupling controller is invoked. Based on the demand volatility of the second type of fluctuation zone, the low-pass filter cutoff frequency is set and the cooling capacity trend component is extracted. The upper-level optimal command is solved with the goal of minimizing the total energy consumption of the cold source side equipment. At the same time, the lower-level decoupling controller is invoked. The lower-level optimal command is solved with the real-time value of the cooling capacity demand prediction sequence of the zone as the tracking target. When carrying the linkage control flag, the linkage optimization is performed on the coupled resonance risk zone pair to output the set values ​​of the operating parameters of each refrigeration equipment.

[0122] The input end of the collaborative control execution module is connected to the output end of the hierarchical collaborative optimization execution module, and the output end is connected to the bottom controller of the refrigeration system. It is used to send down the received operating parameter set values ​​of each refrigeration device for execution.

[0123] The safety protection module is used to continuously monitor the air intake temperature and its rate of change of each cabinet. Its output is connected to the enable terminal of the collaborative control execution module. When the temperature or temperature rise rate exceeds the limit, it triggers an enable signal to pause the current optimized control and switch to the conservative control mode. After the status verification is passed, the optimized control is restored.

[0124] The steps for implementing the corresponding functions of each parameter and each unit module in the intelligent optimization control system for a data center cooling system of the present invention described above can be referred to the parameters and steps in the embodiments of the intelligent optimization control method for a data center cooling system described above, and will not be repeated here.

[0125] Specific implementation method: First, the data acquisition and preprocessing module is started. This module collects data in real time, including the load power of each server, the air inlet and outlet temperature of each rack, the outdoor temperature and humidity, the outlet air temperature, return air temperature, air volume, compressor operating frequency, and operating parameters of each precision air conditioner and the equipment in each refrigeration room, through the building automation system and the data center infrastructure management system. All the collected data is processed by removing outliers, filling missing values, aligning timestamps, and normalizing to build a standardized time-series running dataset for subsequent modules to call.

[0126] Then, the spatiotemporal prediction module for cooling demand is connected to the output of the data acquisition and preprocessing module. This module has a built-in spatiotemporal prediction model for cooling demand that integrates convolutional neural networks and long short-term memory networks. After reading in the time-series running dataset, the convolutional neural network first extracts the spatial correlation features between each cooling zone, and then the long short-term memory network captures the temporal dependence of each feature dimension. Finally, it outputs the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone in the future preset time period. At the same time, based on the prediction sequence, the mean, peak, valley and demand volatility of cooling demand for each cooling zone are calculated and determined.

[0127] Next, the spatiotemporal feature decoupling analysis module is connected to the output of the cooling demand spatiotemporal prediction module. After reading the demand volatility and cooling demand prediction sequence of each cooling zone, this module performs two operations: First, it compares the demand volatility of each cooling zone with a preset stability threshold one by one. Zones with volatility less than the threshold are classified as first-class stable zones, and those with volatility greater than or equal to the threshold are classified as second-class volatile zones. Second, based on the physical location and historical temperature response correlation of each cooling zone, it calculates the thermal coupling strength value between any two cooling zones. This thermal coupling strength value is obtained by combining the Pearson linear correlation coefficient and the mutual information nonlinear coupling index. Two cooling zones with thermal coupling strength values ​​greater than the preset coupling threshold are grouped into the same zone coupling group, and a thermal coupling topology graph is constructed with each cooling zone as a node and the thermal coupling strength value as the edge weight.

[0128] Secondly, the upper-level collaborative control decision module is connected to the output of the spatiotemporal feature decoupling analysis module. After reading the partition type division results and the thermal coupling topology, this module traverses all second-type fluctuation partitions and performs stability condition judgments one by one: If the partition coupling group of a certain second-type fluctuation partition does not contain a first-type stable partition, it checks whether there are other second-type fluctuation partitions in the group and whether the thermal coupling strength value between them is greater than the preset resonance trigger threshold. If they exist, it is determined that it does not meet the stability condition, and these two partitions are marked as a coupled resonance risk partition pair. If they do not exist, it is determined that it meets the stability condition. If the partition coupling group of the partition contains a first-type stable partition, it obtains the maximum value of each thermal coupling strength value between the partition and all first-type stable partitions in the group. If the maximum value is greater than the preset decoupling trigger threshold, it is determined that it does not meet the stability condition. Otherwise, it is determined that it meets the stability condition. After all second-type fluctuation partitions have been judged, if there are no second-type fluctuation partitions in the system, or if all second-type fluctuation partitions meet the stability condition, a global collaborative control instruction is generated. Otherwise, a hierarchical decoupling control instruction is generated, and the linkage control identifier of the coupled resonance risk partition pair is issued with the instruction.

[0129] Next, the hierarchical collaborative optimization execution module is connected to the output of the upper-level collaborative control decision module. This module reads the instruction type and processes it in two ways:

[0130] If the received command is a global collaborative control command, the built-in deep reinforcement learning agent is invoked. The joint action vector is formed by the setpoints of the outlet air temperature of each precision air conditioner, the setpoint of the supply air volume, the setpoint of the chilled water outlet temperature, the frequency of the cooling tower fan, the frequency of the chilled water pump, and the frequency of the cooling water pump. The global cooling energy efficiency ratio is used as the optimization objective to find the optimal solution and output the setpoints of the operating parameters of each refrigeration equipment.

[0131] If the received command is a hierarchical decoupling control command, the built-in upper-level decoupling controller and lower-level decoupling controller are invoked to work together: The upper-level decoupling controller first sets the low-pass filter cutoff frequency according to the demand volatility of each second-type fluctuation zone. The higher the demand volatility of the zone, the lower the cutoff frequency. Then, the cooling demand prediction sequence of each zone is filtered to extract the cooling trend component. The sum of the cooling trend components of all fluctuation zones is taken as the total low-frequency trend load. Then, with the optimization objective of minimizing the total energy consumption of the cold source side equipment in the control time domain, the setpoint of the chilled water outlet temperature and the frequency of the cooling tower fan are obtained under the constraints of satisfying the total low-frequency trend load supply and ensuring that the air inlet temperature of all cabinets does not exceed the first safe temperature threshold. The control cycle for the water pump frequency setpoint is the first preset cycle. The lower-level decoupling controller uses the precision air conditioner outlet temperature setpoint and supply air volume setpoint corresponding to each second-type fluctuation zone as action variables, and the predicted value at the current control moment in the cooling demand prediction sequence of that zone as the tracking target. Under the constraint that the inlet air temperature of all cabinets in the zone does not exceed the first safe temperature threshold, the setpoint of the precision air conditioner in that zone is solved, and the control cycle is the second preset cycle, which is shorter than the first preset cycle. When a linkage control flag is received, the two zones marked as coupled resonance risk zones are linked for optimization according to preset weights, and the joint goal is to solve the problem collaboratively with the minimum comprehensive temperature deviation and the minimum total energy consumption of the two zones as the joint objective. For all first-type stable zones, no optimization processing is performed under the hierarchical decoupling control command. Instead, the average cooling demand of that zone is used as the feedforward input, and proportional-integral feedback correction is performed in combination with the deviation between the actual inlet air temperature and the preset target temperature. After lag compensation for the feedback deviation, the outlet temperature setpoint and supply air volume setpoint of the corresponding precision air conditioner are directly calculated.

[0132] Finally, the input of the coordinated control execution module is connected to the output of the hierarchical coordinated optimization execution module, and the output is connected to the underlying controllers of each precision air conditioner, chiller unit, chilled water pump, cooling water pump, and cooling tower. This module sends the setpoints of the operating parameters of each refrigeration device output by the hierarchical coordinated optimization execution module to the underlying controllers of the corresponding devices for execution via a communication protocol, and receives execution status feedback. After completing the execution of each control cycle, it returns to the first step to start the data acquisition and processing for the next cycle, and the above steps are executed cyclically to form a closed-loop control.

[0133] During the aforementioned cyclic operation, the safety protection module operates independently and continuously, monitoring the inlet air temperature and its rate of rise of each cabinet in real time. When the rate of rise of the inlet air temperature of any cabinet exceeds a preset threshold, the safety protection module triggers a pre-protection action in advance, increasing the airflow of the precision air conditioner in the zone where that cabinet is located. When the inlet air temperature of any cabinet exceeds a preset second safety temperature threshold, the safety protection module sends an enable signal to the collaborative control execution module, pausing the optimization instructions output by the currently active deep reinforcement learning agent or decoupled controller, and adjusting the outlet air temperature setpoint and airflow setpoint of all precision air conditioners to the preset conservative value along a smooth transition curve. After the inlet air temperature of all cabinets returns to normal, the safety protection module verifies the operating status of each actuator and the effectiveness of the temperature control logic. After confirming that everything is normal, it maintains the conservative control state for a preset period of time, and then transitions the control parameters back to the optimized value along the smooth recovery curve, restoring normal optimized control.

[0134] Example 3

[0135] Based on the same inventive concept as Embodiment 1, this embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0136] The processor executes the aforementioned intelligent optimization control method for a data center cooling system by calling the computer program stored in the memory.

[0137] It should be noted that all computer programs for the intelligent optimization control method of a data center cooling system are implemented using the C language.

[0138] Example 4

[0139] Based on the same inventive concept as in Embodiment 1, this embodiment proposes a computer-readable storage medium having an erasable and rewritable computer program stored thereon.

[0140] When a computer program runs on a computer device, it causes the computer device to execute the aforementioned intelligent optimization control method for a data center cooling system.

[0141] For example, computer-readable storage media can be read-only memory, random access memory, read-only optical disc, magnetic tape, floppy disk, and optical data storage devices.

[0142] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the embodiments for IoT devices and media are relatively simple in description because they are fundamentally similar to the method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0143] The systems, media, and methods provided in the embodiments of the present invention are in one-to-one correspondence. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0149] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0150] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0151] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for intelligent optimization control of a data center cooling system, characterized in that, Includes the following steps: Step S1: Collect real-time operational data related to the data center cooling system. The operational data related to the system includes server load power, air inlet and outlet temperatures of each rack, outdoor ambient temperature and humidity data, and operating parameters of each precision air conditioner and cooling room equipment. Clean, align, and normalize the real-time collected data from each dimension to construct a standardized time-series operational dataset. Step S2: Construct a spatiotemporal prediction model for cooling demand that integrates convolutional neural networks and long short-term memory networks based on the time-series running dataset. Extract and output the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone within a preset future time period, as well as the mean, peak, valley and demand volatility of cooling demand for each cooling zone determined based on the prediction sequence. Step S3: Based on the demand volatility output in step S2, perform thermal stability assessment on each cooling zone. Cooling zones with demand volatility less than a preset stability threshold are classified as first-class stable zones, and cooling zones with demand volatility greater than or equal to the preset stability threshold are classified as second-class volatile zones. At the same time, based on the correlation between the physical location of each cooling zone and its historical temperature response, calculate the thermal coupling strength value between any two cooling zones. Cooling zones with thermal coupling strength values ​​greater than a preset coupling threshold are assigned to the same zone coupling group, and a thermal coupling topology diagram describing the thermal coupling relationship between each cooling zone and between zone coupling groups is constructed. Step S4: Input the thermal stability assessment results and thermal coupling topology diagram generated in step S3 into the upper-level collaborative control decision module. The upper-level collaborative control decision module traverses all second-type fluctuation partitions and performs the following judgment for each second-type fluctuation partition: First, check whether the partition coupling group containing the second type of fluctuation partition contains a first type of stable partition; If not, it is further determined whether there is at least one other second-type fluctuation partition in the partition coupling group, and the thermal coupling strength value between the second-type fluctuation partition and at least one other second-type fluctuation partition is greater than the preset resonance trigger threshold. If so, it is determined that the second-type fluctuation partition does not meet the stability condition, and it and the corresponding other second-type fluctuation partition are marked as a coupled resonance risk partition pair. If not, it is determined that it meets the stability condition. If a first-class stable partition is included, the maximum value of each thermal coupling strength between the second-class fluctuating partition and all first-class stable partitions in the group is obtained. If the maximum value is greater than the preset decoupling trigger threshold, it is determined that it does not meet the stability condition; otherwise, it is determined that it meets the stability condition. When there is no second type of fluctuation partition, a global collaborative control command is generated directly. When there is a second type of fluctuation partition and all second type of fluctuation partitions meet the stability conditions, a global collaborative control command is generated. Otherwise, a hierarchical decoupling control command is generated and a linkage control identifier is attached. Step S5: When it is a global collaborative control command, the deep reinforcement learning agent performs global optimization with the global cooling energy efficiency ratio as the optimization target, and outputs the set values ​​of the operating parameters of each cooling device. When the hierarchical decoupling control command is executed, the hierarchical control strategy of inertial time decoupling is implemented, and the upper-level decoupling controller and the lower-level decoupling controller work together to output the set values ​​of the operating parameters of each refrigeration device. Step S6: Send the set values ​​of the operating parameters of each refrigeration device generated in step S5 to the bottom controller of the refrigeration system for execution, and execute steps S1 to S5 in a cycle with a preset control period to achieve closed-loop optimization control. Step S7: Continuously monitor the air intake temperature and its rate of change of each cabinet. When the temperature or rate of temperature rise exceeds the limit, trigger the graded protection, pause the current optimized control and switch to conservative control mode. Optimized control will be resumed after the temperature returns to normal and the system status is verified.

2. The intelligent optimization control method for a data center cooling system according to claim 1, characterized in that: The demand volatility rate in step S3 is the ratio of the standard deviation of the predicted cooling demand for each refrigeration zone in the future preset period to the mean cooling demand in that period. The thermal coupling strength value in step S3 is calculated by combining the linear correlation coefficient and the nonlinear coupling index. The thermal coupling topology graph uses each cooling zone as a node and the thermal coupling strength value as the weight of the edge connecting the corresponding node.

3. The intelligent optimization control method for a data center cooling system according to claim 2, characterized in that: The offline optimization method for the preset coupling threshold, preset resonance trigger threshold, and preset decoupling trigger threshold is as follows: obtain the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone and the corresponding actual control effect data of the cooling system in the historical operation data, take the minimum weighted sum of energy consumption and temperature over-limit times of the cold source side equipment as the optimization objective, and use the grid search method or Bayesian optimization method to determine the preset coupling threshold, preset resonance trigger threshold, and preset decoupling trigger threshold offline. After offline optimization is determined, it is automatically updated online every preset period based on actual operating data. During online update, the thermal coupling strength value between each cooling zone is recalculated based on the newly accumulated operating data in the current period, and the thresholds are updated accordingly. When a sudden change in system load is detected that exceeds the preset load change threshold or when system equipment is replaced, an emergency update of the threshold is triggered.

4. The intelligent optimization control method for a data center cooling system according to claim 3, characterized in that: Step S5 specifically includes the following steps: Step S51: When a global collaborative control command is received, the setpoints of the outlet air temperature and supply air volume of each precision air conditioner, the setpoint of the chilled water outlet temperature, the frequency of the cooling tower fan, the frequency of the chilled water pump, and the frequency of the cooling water pump are used as the joint action vector. The deep reinforcement learning agent performs global optimization with the global cooling energy efficiency ratio as the optimization target, and outputs the setpoints of the operating parameters of each refrigeration device. The deep reinforcement learning agent adopts a three-stage deployment mode of offline pre-training, simulation verification, and online low exploration rate fine-tuning. The online fine-tuning stage sets up a manual intervention interface. Step S52: When a hierarchical decoupling control command is received, execute the hierarchical control strategy of inertial-time decoupling: The upper-level decoupling controller sets the corresponding low-pass filter cutoff frequency according to the demand volatility of each second-type fluctuation zone. The larger the demand volatility of the zone, the lower the cutoff frequency. The cooling demand prediction sequence of each zone is filtered to extract the cooling trend component. The sum of the cooling trend components of all second-type fluctuation zones is taken as the total low-frequency trend load. The set value of chilled water outlet temperature, cooling tower fan frequency, chilled water pump frequency and cooling water pump frequency are used as action variables. The goal is to minimize the total energy consumption of the cold source side equipment in the control time domain. Under the constraints of satisfying the supply of total low-frequency trend load and the air inlet temperature of all cabinets not exceeding the first safe temperature threshold, the upper-level optimal control command is solved. The control cycle is the first preset cycle. The lower-level decoupling controller uses the precision air conditioning outlet temperature setpoint and supply air volume setpoint corresponding to each second-type fluctuation zone as action variables, and uses the predicted value of the current control moment in the cooling demand prediction sequence of the zone output in step S2 as the tracking target. Under the constraint that the air inlet temperature of all cabinets in the zone does not exceed the first safe temperature threshold, the controller solves the lower-level optimal control command. The control cycle is a second preset cycle shorter than the first preset cycle. When a linkage control flag is received, the controller performs linkage optimization on the coupled resonance risk zone pair, and performs coordinated control with the minimum comprehensive temperature deviation and the minimum total energy consumption of the two zones according to preset weights as the joint objective. For all Class I stable zones, the average cooling demand of that zone is used as the feedforward input. The deviation between the actual inlet air temperature and the preset target temperature is combined to perform proportional-integral feedback correction and lag compensation for the feedback deviation, and the set value of the corresponding precision air conditioner is calculated. Among them, both the first safe temperature threshold and the preset target temperature are lower than the temperature threshold for triggering graded protection in step S7.

5. The intelligent optimization control method for a data center cooling system according to claim 4, characterized in that: The deep reinforcement learning agent in step S51 employs a constrained reinforcement learning algorithm, with the reward function being: in, It is the reciprocal of the global cooling energy efficiency ratio. Let be the air intake temperature of the i-th rack. The soft-constraint reference temperature is less than or equal to the first safe temperature threshold of sub-step S52. These are the combined action vectors formed by the setpoints of the outlet air temperature, supply air volume, chilled water outlet temperature, cooling tower fan frequency, chilled water pump frequency, and cooling water pump frequency for the current and previous control cycles, respectively. The L2 norm square of the change in action between adjacent control cycles, α, β, and γ are preset non-negative weighting coefficients, and M is the total number of cabinets. Actions that violate the minimum start-stop interval, maximum number of start-stop cycles, or upper limit of pump frequency during training are judged as invalid and discarded.

6. The intelligent optimization control method for a data center cooling system according to claim 5, characterized in that: In step S52, both the upper-level decoupling controller and the lower-level decoupling controller adopt a model predictive control framework, and the constraints also include the adjustment range constraints of each action variable. Step S52 also includes a fault redundancy strategy: when a fault or communication interruption is detected in the upper decoupling controller, the lower decoupling controller automatically switches to independent control mode and independently solves the precision air conditioning control command based on the real-time cooling demand forecast value and temperature constraints of this zone. After the upper decoupling controller recovers, the hierarchical collaborative control is restored.

7. The intelligent optimization control method for a data center cooling system according to claim 6, characterized in that: Step S7 specifically includes the following steps: Step S71: When the rate of increase of the air intake temperature of any cabinet exceeds the preset rate threshold, pre-protection is triggered, and the air volume of the precision air conditioner in the zone where the cabinet is located is increased to the preset value. Step S72: A second safe temperature threshold is preset, and the first safe temperature threshold is lower than the preset second safe temperature threshold preset level difference. When the air inlet temperature of any cabinet exceeds the second safe temperature threshold, the strategy output of the currently active controller in step S5 is immediately paused, and all precision air conditioner outlet air temperature setting values ​​and supply air volume setting values ​​are adjusted from the current optimized values ​​to preset conservative values ​​along a preset smooth transition curve. The preset conservative value is the preset ratio of the outlet air temperature setting to the preset low temperature value and the supply air volume setting to the maximum rated value. Step S73: After the air intake temperature of all cabinets returns to the normal range, verify the operating status of each actuator and the effectiveness of the temperature control logic. After confirming that it is normal, maintain the conservative control stability for a preset time, and then transition the control parameters from the conservative value back to the optimized value along the preset smooth recovery curve, and restore the optimized control of step S5.

8. A smart optimization control system for a data center cooling system, implemented based on the smart optimization control method for a data center cooling system as described in any one of claims 1-7, characterized in that, Specifically, it includes: a data acquisition and preprocessing module, a cooling demand spatiotemporal prediction module, a spatiotemporal feature decoupling analysis module, an upper-level collaborative control decision-making module, a hierarchical collaborative optimization execution module, a collaborative regulation execution module, and a safety protection module; The data acquisition and preprocessing module is used to collect server load power, air inlet and outlet temperature of each rack, outdoor ambient temperature and humidity data, and operating parameters of each precision air conditioner and refrigeration room equipment in real time, and to clean, align and normalize the collected data to build a standardized time-series running dataset. The input end of the cooling demand spatiotemporal prediction module is connected to the output end of the data acquisition and preprocessing module. It has a built-in cooling demand spatiotemporal prediction model that integrates convolutional neural networks and long short-term memory networks. It is used to extract and output the spatiotemporal distribution prediction sequence of cooling demand for each cooling zone within a future preset time period, as well as the mean, peak, valley and demand volatility of cooling demand for each cooling zone determined based on the prediction sequence. The spatiotemporal feature decoupling analysis module has its input connected to the output of the cooling demand spatiotemporal prediction module, and is used to perform the following operations: Based on the comparison between demand volatility and preset stability threshold, each cooling zone is divided into either a first-class stable zone or a second-class volatile zone. Based on the correlation between the physical location of each cooling zone and its historical temperature response, the thermal coupling strength value between any two cooling zones is calculated, and cooling zones with thermal coupling strength values ​​greater than a preset coupling threshold are grouped into the same zone coupling group to construct a thermal coupling topology map. The upper-level collaborative control decision module has its input end connected to the output end of the spatiotemporal feature decoupling analysis module. It is used to receive the partition type division and thermal coupling topology map, and according to the logic described in step S4, traverse the second type of fluctuation partition, determine whether it meets the stability conditions, and generate a global collaborative control command or a hierarchical decoupling control command carrying a linkage control identifier. The input end of the hierarchical collaborative optimization execution module is connected to the output end of the upper-level collaborative control decision module. When receiving a global collaborative control command, it calls the built-in deep reinforcement learning agent to perform optimization with the global cooling energy efficiency ratio as the optimization target and outputs the set values ​​of the operating parameters of each cooling device. When a hierarchical decoupling control command is received, the built-in upper-level decoupling controller is invoked. Based on the demand volatility of the second type of fluctuation zone, the low-pass filter cutoff frequency is set and the cooling capacity trend component is extracted. The upper-level optimal command is solved with the goal of minimizing the total energy consumption of the cold source side equipment. At the same time, the lower-level decoupling controller is invoked. The lower-level optimal command is solved with the real-time value of the cooling capacity demand prediction sequence of the zone as the tracking target. When carrying the linkage control flag, the linkage optimization is performed on the coupled resonance risk zone pair to output the set values ​​of the operating parameters of each refrigeration equipment. The input end of the collaborative control execution module is connected to the output end of the hierarchical collaborative optimization execution module, and the output end is connected to the bottom controller of the refrigeration system. It is used to send down the received operating parameter set values ​​of each refrigeration device for execution. The safety protection module is used to continuously monitor the air intake temperature of each cabinet and its rate of change. Its output is connected to the enable terminal of the collaborative control execution module. When the temperature or temperature rise rate exceeds the limit, it triggers an enable signal to pause the current optimized control and switch to the conservative control mode. After the status verification is passed, the optimized control is restored.

9. An electronic device, comprising: A processor and a memory, wherein the memory stores a computer program that can be called by the processor, characterized in that: the processor executes a data center cooling system intelligent optimization control method as described in any one of claims 1-7 by calling the computer program stored in the memory.

10. A computer-readable storage medium, characterized in that: The system stores instructions that, when executed on a computer, cause the computer to perform a smart optimization control method for a data center cooling system as described in any one of claims 1-7.

Images