A method and system for pre-fabricated data center energy efficiency adaptive and robust control

Abstract

The application relates to the field of distributed control, and discloses a prefabricated data center energy efficiency adaptive and robust control method and system. A system-level controller registers prefabricated parameters and positions, and solves an initial resource scheduling strategy set according to real-time task requirements. A unit controller performs forward-looking state calculation based on local historical operation data and generates a forward-looking state abstract, and simultaneously calculates a local state change amount. When preset sending conditions are met, the abstract is sent to the system-level controller through flexible data exchange middleware. The system-level controller combines multiple-end abstracts, the initial strategy set, task requirements and prefabricated parameters, performs forward-looking resource collaborative scheduling in a rolling window to generate a global collaborative scheduling strategy, determines an operation mode and issues a control instruction, further establishes a distributed takeover mechanism through a double-redundancy hard-wired ring network, and realizes non-disturbance switching back through a shadow operation mode. While reducing communication overhead, the application realizes system-level high-availability autonomy.

Description

Technical Field

[0001] This invention relates to the field of distributed control, and in particular to a prefabricated data center energy efficiency adaptive and robust control method and system. Background Technology

[0002] Prefabricated data centers are typically formed by assembling computing, power distribution, cooling, integrated management, and interconnection interface units on-site after they are prefabricated in a factory. Compared with traditional civil engineering data centers, prefabricated data centers have the advantages of shorter deployment cycles, flexible expansion, and replaceable units. However, different prefabricated units vary in terms of capacity, installation location, thermal inertia, power distribution efficiency, and the impact of adjacent units.

[0003] Existing modular, containerized, or prefabricated data center technologies are typically designed around cold aisle systems, high-efficiency air conditioning, air-liquid co-cooling, liquid cooling units, power supply redundancy, or monitoring and alarm systems. Some solutions also utilize artificial intelligence or predictive models for cooling regulation. These technologies indicate that energy efficiency control in prefabricated data centers needs to simultaneously consider heat load, cooling capacity, power supply capacity, and monitoring link reliability.

[0004] However, existing energy-saving scheduling technologies are mostly based on centralized data acquisition and centralized solution. In prefabricated data centers, if each prefabricated unit continuously uploads full status data, it will increase communication and control overhead, affecting plug-and-play expansion. Existing cooling or power distribution structures focus on single devices or local channels, failing to incorporate prefabricated parameters, installation location, adjacent unit types, and real-time load into the operating mode selection. Although existing high-availability control can perform link switching or redundant power supply, unit local autonomy cannot maintain global load balancing when the system-level controller fails or the network is isolated. Therefore, a control method for prefabricated data centers is needed to form a continuous closed loop, enabling unit local look-ahead computation, event-driven low-overhead communication, system-level rolling collaborative scheduling, prefabricated parameter-driven operating mode selection, and control level migration in case of failure. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in related technologies. To this end, the present invention provides a prefabricated data center energy efficiency adaptive and robust control method and system.

[0006] The first technical solution provided by this invention is: An adaptive and robust energy efficiency control method for prefabricated data centers includes the following steps: Initialization phase: The system-level controller registers the prefabrication parameter set, installation location, and adjacent unit types for each prefabricated unit; Initial phase: The system-level controller receives real-time task requirements and solves them based on the registered set of pre-configured parameters and the installation location to obtain an initial resource scheduling strategy set; Forward-looking stage: The unit controller built into each prefabricated unit performs forward-looking state calculations based on local historical operating data, generates a forward-looking state summary, and calculates the local state change amount; Data exchange phase: A flexible data exchange middleware runs between the unit controller and the system-level controller. When the local state change meets the preset sending conditions, the look-ahead state summary is sent to the system-level controller. Collaborative optimization phase: The system-level controller receives multiple forward-looking state summaries and merges them with the initial resource scheduling strategy set, the real-time task requirements, and the registered prefabricated parameter set. It then performs forward-looking resource collaborative scheduling within a rolling time window to generate a global collaborative scheduling strategy, thereby determining the operating mode of each prefabricated unit and issuing physical control commands.

[0007] Furthermore, the pre-set parameter set includes a mapping table of compressor speed and cooling capacity of the refrigeration unit, a mapping table of fan speed and air volume, voltage and current efficiency curves, rated capacity, thermal inertia parameters, and equipment lifespan influencing factors corresponding to each candidate operating parameter of the power distribution unit. The installation location is represented by row and column coordinates, container number, or cabinet group number; If the task constraints or capacity constraints in the real-time task requirements are not feasible, the system-level controller freezes the mode switching of the current control cycle and outputs a conservative resource allocation strategy.

[0008] Furthermore, in the prospective phase, generating a prospective state summary specifically includes: The unit controller inputs the historical computing load, historical heat load, historical power distribution load, task queue length, ambient temperature, current operating mode identifier, and fault alarm record of the prefabricated unit into the long short-term memory network as the local historical operating data. The Long Short-Term Memory network is used to output the probability distribution of resource demand and the expected failure probability at each predicted moment within the future rolling time window; The unit controller performs heat load prediction and energy consumption prediction calculations based on the local historical operating data, and compresses the resource demand probability distribution, the expected failure probability, the heat load prediction value, and the energy consumption prediction value into the prospective state summary.

[0009] Furthermore, in the look-ahead phase, the local state change is calculated in the following manner: The unit controller collects the real-time computing load, real-time thermal load, and real-time expected failure probability at the current moment, and obtains the historical load value, thermal load prediction value, and expected failure probability corresponding to the last time the prospective state summary was sent. Calculate the most recent effective window difference between the current value and the corresponding historical value at the last transmission time, and divide each of the most recent effective window differences by the calibrated maximum fluctuation value to obtain the dimensionless calculated load change component, heat load change component, and fault probability change component. The maximum value among the calculated load change component, the heat load change component, and the fault probability change component is taken as the local state change quantity. Its mathematical expression is defined as: in, Let be the local state change of the i-th prefabricated unit at time t; To calculate the load variation components; This represents the component of heat load variation; This represents the component representing the change in failure probability.

[0010] Furthermore, during the data exchange phase: Sending conditions are If true, in, Whether the i-th prefabricated unit sends a look-ahead state summary at time t; Let be the local state change of the i-th prefabricated unit at time t; The unit dynamic threshold of the i-th prefabricated unit; This is a system-level query event flag; I(·) is a mathematical indicator function; When the value of the indicator function is 1, the flexible data exchange middleware performs digest packaging and publishes the prospective state digest to the system-level controller; When the value of the indicator function is 0, the unit controller only sends a heartbeat signal to the system-level controller.

[0011] Furthermore, in the aforementioned collaborative optimization phase: Within the rolling time window, the system-level controller converts energy consumption, task migration cost, fault risk, and mode switching cost into dimensionless indices and sums them up. The goal is to minimize the rolling collaborative scheduling target value obtained from the summation and to perform forward-looking resource collaborative scheduling to generate the global collaborative scheduling strategy. The global collaborative scheduling strategy includes task migration instructions, computing capacity start / stop instructions, cooling capacity targets, and power distribution targets.

[0012] Furthermore, in the collaborative optimization phase, determining the operating mode of each prefabricated unit according to the global collaborative scheduling strategy includes the following steps: The system-level controller reads the candidate operating mode set of the prefabricated unit; Based on the registered prefabricated parameter set, the installation location, the registered adjacent unit type, and the real-time load, a mode score is calculated for each candidate operating mode. The model score is obtained by weighted summation of normalized predicted energy consumption, temperature deviation, insufficient capacity margin, and equipment life loss. The system-level controller sorts the mode scores of each candidate operating mode and selects the candidate operating mode that ranks first in the score ranking and whose score difference with the current operating mode reaches the hysteresis switching threshold as the target operating mode of each prefabricated unit. The system-level controller sends physical control commands corresponding to the target operating mode to the corresponding refrigeration unit and power distribution unit. The physical control commands include combinations of operating parameters.

[0013] Furthermore, it also includes the takeover phase: The integrated management unit monitors the heartbeat of the system-level controller, the system-level controller's computational response, and the connectivity of the control network; When a system-level controller failure, control network isolation, or heartbeat timeout is detected, the distributed collaborative control module built into the integrated management unit enters the takeover state; In the takeover state, the distributed collaborative control module obtains the prospective state summary sent by each unit controller in real time through the backup link established by the dual redundant hard-wired ring network. The distributed collaborative control module calls the historical policy database, the most recent global collaborative scheduling policy, and the currently obtained forward state summary to calculate the migration optimization target, and publishes the takeover identifier and takeover policy version number to each unit controller to implement alternative control.

[0014] Furthermore, it also includes a switchback phase: When the system-level controller is reset and brought back online, the integrated management unit instructs the system-level controller to enter shadow operation mode; In the shadow operation mode, the system-level controller operates in parallel as a cold standby node without execution rights. It reads system parameters, topology and the latest prospective state summary in real time through the dual-redundant hard-wired ring network, and performs rolling scheduling and mode scoring strategy calculation in the background. The distributed collaborative control module performs consistency verification between the strategy calculated by the system-level controller in the background and the strategy currently executed by the integrated management unit, comparing the differences in task allocation, cooling parameters, and power distribution parameters between the two. When the differences of the three types of parameters are consistently less than the corresponding consistency hard threshold at each sampling moment within the preset continuous verification period, the integrated management unit issues a control migration instruction to each of the unit controllers to officially close the shadow operation mode and seamlessly switch the hierarchical control function back to the system-level controller.

[0015] Based on the first technical solution, the second technical solution proposed is as follows: An adaptive and robust energy efficiency control system for prefabricated data centers is provided for implementing the aforementioned adaptive and robust energy efficiency control method for prefabricated data centers. This includes: a system-level controller, multiple unit controllers, a comprehensive management unit, and flexible data exchange middleware; The system-level controller is used to: register the prefabricated parameter set, installation location, and adjacent unit type of each prefabricated unit during the initialization phase; solve the initial resource scheduling strategy set according to real-time task requirements, the prefabricated parameter set, and the installation location during the initial allocation phase; and merge the received multiple forward state summaries, the initial resource scheduling strategy set, the real-time task requirements, and the registered prefabricated parameter set during the collaborative optimization phase, and perform collaborative scheduling within a rolling time window to generate a global collaborative scheduling strategy to determine the operating mode of each prefabricated unit and issue physical control commands. Each of the aforementioned unit controllers is built into the corresponding prefabricated unit and is used to perform forward state calculations based on local historical operating data during the forward look-ahead phase to generate a forward state summary and to calculate the local state change amount. The flexible data exchange middleware runs between the unit controller and the system-level controller, and is used to send the look-ahead state summary to the system-level controller when the local state change meets the preset sending conditions during the data exchange phase. The integrated management unit has a built-in distributed collaborative control module, which is used to enable the distributed collaborative control module to enter the takeover state when the system-level controller is detected to be faulty, the control network is isolated, or the heartbeat timeout occurs during the takeover phase. The module then obtains the prospective state summary through a dual-redundant hardwired ring network to execute alternative control.

[0016] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: 1. This solution performs look-ahead state calculations locally on the unit controller, utilizing a Long Short-Term Memory (LSTM) network to compress multiple results such as resource requirements and failure probabilities within future rolling time windows into a lightweight look-ahead state summary. By running an event-driven state summary protocol in the middleware, only low-overhead heartbeat packets are sent when the local state change amount does not reach a dynamic threshold. This completely changes the traditional centralized mode of continuously uploading all data, which generates high communication overhead, providing solid technical support for dynamic access of prefabricated units, low-overhead collaboration, and seamless capacity expansion.

[0017] 2. The rolling collaborative scheduling mechanism converts energy consumption, task migration costs, fault risk, and mode switching costs into dimensionless indicators, performing dynamic rolling global optimization across physical fields within the prediction time slot. Subsequently, a refined operation mode scoring mechanism incorporating predicted energy consumption, thermal load deviation, insufficient capacity margin, and core equipment lifespan degradation factors, combined with a hysteresis switching threshold calibrated based on equipment response time, is used to issue decisions for the target mode. This effectively prevents repeated ping-pong switching of operation modes within adjacent control cycles, substantially protecting and extending the service life of high-value physical actuators within the module.

[0018] 3. The solution innovatively designs a dual-redundant hardwired ring network that runs through all physical units, the system-level controller, and the integrated management unit in terms of physical communication paths. In extreme isolation scenarios such as core controller failure or severe network disruption, the distributed collaborative control module built into the integrated management unit will forcibly switch to takeover mode, completely bypassing the potentially failed conventional routing network. It will continue to ensure forward summary transmission using bidirectional hardwired backup links and call upon the historical policy database to maintain global load balancing, thus ensuring the system's continuous control loop and high availability survivability.

[0019] 4. After the core controller fails and resets, the solution introduces a "shadow operation mode" without execution rights, which operates as a cold standby node in parallel. During shadow operation, the system-level controller only reads real-time hardware parameters and summaries and performs calculations in the background. The distributed coordination module performs cross-domain (task allocation, cooling, power distribution) multi-dimensional consistency difference verification between its background strategy and the current takeover strategy. Only after several consecutive rolling control cycles are stable below the consistency hard threshold, the algorithm state machine is confirmed to have smoothly warmed up, and the parameters are aligned, can the formal, non-disruptive switchback be performed. This completely eliminates the physical transient step oscillations that are easily caused by the hard switch of control rights, greatly maintaining the continuous stability of the precision environment of the precision data center.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the topology of the prefabricated data center physical framework and control system of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0024] This invention describes and illustrates unit look-ahead computing, event-driven communication, rolling collaborative scheduling, pre-defined parameter-driven operation mode selection, and fault takeover migration within a continuous control chain.

[0025] Figure 1 The topology of the prefabricated data center physical framework and control system of the present invention is shown.

[0026] The prefabricated data center includes computing units 201, cooling units 202, power distribution units 203, integrated management units 103, and interconnection interface units 204. The computing unit 201 is a prefabricated cabinet module or computing cabin; the cooling unit 202 is an air conditioning module, liquid-cooled distribution module, or hybrid cooling module; the power distribution unit 203 is a low-voltage power distribution module, bus module, or energy storage access module; and the interconnection interface unit 204 provides network and control bus connections between units. Each type of prefabricated unit has a built-in unit controller 102, a system-level controller 101 responsible for global coordination, and the integrated management unit 103 has a built-in distributed collaborative control module 104.

[0027] During the initialization phase, the system-level controller 101 registers the prefabricated parameter set, installation location, adjacent unit type, rated capacity, selectable operating mode, and communication address for each prefabricated unit. The prefabricated parameter set includes a compressor speed-to-cooling capacity mapping table and a fan speed-to-airflow mapping table for the refrigeration unit, voltage-current efficiency curves, rated capacity, thermal inertia parameters, and equipment lifespan impact factors corresponding to each candidate operating parameter for the power distribution unit. The installation location is represented by row and column coordinates, container number, or cabinet group number. The adjacent unit type describes whether there are high heat density computing units, power distribution bottleneck units, or spare capacity units around the target unit.

[0028] At the start of each control cycle, the system-level controller 101 receives real-time task requirements. These requirements include the computational resource requirements of the task to be scheduled, task completion deadlines, power limit constraints, fault tolerance levels, and portability flags. Based on the real-time task requirements, pre-defined parameter sets, installation locations, and adjacent unit types, the system-level controller 101 constructs a mixed-integer linear programming model and solves for the initial resource scheduling strategy set according to equation (1). The output of equation (1) is used as the initial solution in subsequent rolling optimizations. If task constraints or capacity constraints are not feasible, the system-level controller 101 freezes the current cycle mode switching and outputs a conservative resource allocation strategy.

[0029] The initial optimization objective is calculated according to equation (1): (1) in, This is the initial resource scheduling target value; For prefabricated unit index; This represents the total number of prefabricated units; Index for candidate running modes; The total number of candidate operating modes; The normalized energy consumption cost for the m-th candidate operating mode of the i-th prefabricated unit; The decision variable is whether the i-th prefabricated unit should select the m-th candidate operating mode. For refrigeration resource indexing; This represents the total number of refrigeration resources. The cost of normalizing the c-th refrigeration resource; Let c be the decision variable for whether or not the c-th cooling resource is activated; For power distribution resource index; The total number of power distribution resources; The normalization activation cost for the d-th power distribution resource; Let d be the decision variable for whether the d-th power distribution resource is activated.

[0030] Decision variables , , The range of values ​​is limited. .

[0031] In Equation (1), the energy consumption cost, cooling capacity activation cost, and power distribution capacity activation cost are normalized according to rated power, rated cooling capacity, and rated power distribution capacity, respectively, before entering the objective function. Therefore, both the left and right sides of Equation (1) are dimensionless scheduling costs. The weights or activation costs are derived from the pre-set parameter set, equipment rated parameters, and on-site commissioning records. The values ​​are written during initialization or operation and maintenance calibration. The previous valid value is frozen when communication anomalies, missing capacity curves, or calibration errors exceed the limit.

[0032] Each unit controller 102 runs a look-ahead state calculation module locally. The look-ahead state calculation module receives historical computed load, historical cooling load, historical power distribution load, task queue length, ambient temperature, current operating mode identifier, and fault alarm records for its unit, and inputs them into a Long Short-Term Memory (LSTM) network. The LSM network outputs the resource demand probability distribution and expected fault probability for each prediction time within the future rolling time window. The unit controller 102 also calculates predicted heat load and energy consumption, and compresses these results into a look-ahead state summary.

[0033] The training samples for the local load prediction model consist of historical data from completed operating cycles, with training labels representing the actual resource demand, heat load, energy consumption, and fault status for that cycle. During online operation, the unit controller 102 only uses the model output to generate a prospective state summary; when measured records are not retrieved, fault labels are not confirmed, or the operating mode undergoes an abnormal switch, the window prediction output is not used to update the model parameters, thereby avoiding self-training with unverified prediction results.

[0034] The look-ahead state summary includes unit identifier, timestamp, resource demand quantile for the future rolling time window, peak heat load forecast, energy consumption forecast, expected failure probability, current operating mode identifier, available computing capacity, available cooling capacity, and available power distribution capacity. Compared to the full original sampled data, the look-ahead state summary retains only the compressed fields required for scheduling, making it suitable for low-overhead coordination after plug-and-play expansion of prefabricated units.

[0035] A flexible data exchange middleware 105 driven by events runs between the unit controller 102 and the system-level controller 101. The flexible data exchange middleware 105 adopts an on-demand state summary protocol. The local state change is calculated according to equation (2), and the conditions for sending the prospective state summary are determined according to equation (3). The load change component, the heat load change component, and the fault probability change component are calculated by dividing the difference of the most recent effective window by the corresponding historical fluctuation upper limit or calibration threshold, and are all dimensionless components.

[0036] The change in local state is calculated according to formula (2): (2) in, Let be the local state change of the i-th prefabricated unit at time t; To calculate the load variation components; This represents the component of heat load variation; This represents the component representing the change in failure probability.

[0037] Among them, the above three components at the sampling time The corresponding dimensionless mathematical expressions are defined as follows: In the formula, , , They represent the first Each prefabricated unit at the current moment Real-time calculation of load, real-time thermal load, and real-time expected failure probability; , , These respectively represent the time when the unit last sent the look-ahead state summary. The corresponding historical load values, heat load forecast values, and expected failure probabilities; , , The maximum value of the fluctuation, either historically statistically or artificially defined, is used as the denominator to eliminate the physical dimensions of each component, transforming it into a dimensionless scalar ranging from 0 to 1.

[0038] The conditions for sending the look-ahead state summary are determined according to equation (3): (3) in, Whether the i-th prefabricated unit sends a look-ahead state summary at time t; Let be the local state change of the i-th prefabricated unit at time t; The unit dynamic threshold of the i-th prefabricated unit; This is a system-level query event flag.

[0039] in, The middleware performs summary packaging and deployment when... It sends heartbeat packets with minimal overhead. For mathematical indicator functions: Used to explicitly define the true or false output of the sending condition.

[0040] The output of equation (2) serves as the input of equation (3). When equation (3) determines that the data should be sent, the flexible data exchange middleware 105 sends a look-ahead state summary to the system-level controller 101. When equation (3) determines that the data should not be sent, the unit controller 102 only sends a heartbeat signal. The unit dynamic threshold is determined based on the unit type, historical fluctuation level, control cycle length, and artificial safety margin. The threshold update is only performed when multiple consecutive windows of data are valid and there are no communication anomalies. If the summary timestamp expires, the data goes out of bounds, or the heartbeat is lost, the system-level controller 101 does not use the window to update the scheduling model parameters.

[0041] The system-level controller 101 receives forward-looking state summaries sent by multiple unit controllers 102, merges them with the initial resource scheduling strategy set, real-time task requirements, and pre-set parameter set, and performs forward-looking resource collaborative scheduling according to equation (4) within a rolling time window. The global collaborative scheduling strategy includes task migration instructions, computing capacity start / stop instructions, cooling capacity targets, and power distribution targets.

[0042] The rolling collaborative scheduling target is calculated according to formula (4): (4) in, The target value for rolling collaborative scheduling; For the index of the predicted time within the scrolling window; This is the current control moment; The length of the scrolling time window; As the weight of the energy consumption item; For a moment The normalized energy consumption term; Weights for task migration or load imbalance items; For a moment Normalized task migration cost term; Weights for fault risk items; For a moment Normalized fault risk item; Weights for mode switching cost items; For a moment The normalization mode switching cost term.

[0043] The energy consumption, task migration cost, fault risk, and mode switching cost in equation (4) are all converted into dimensionless indices and then summed.

[0044] Weight , , and The weights are provided by the operation and maintenance strategy table or the on-site debugging process, and range from 0 to 1, with a total value of 1. When the task completion deadline is tighter, the task migration cost weight is increased; when the expected failure probability is higher, the failure risk weight is increased; and when the mode switching is frequent, the mode switching cost weight is increased. If any weight source is missing, the system controller 101 uses the most recently verified weight set and prohibits online self-learning updates.

[0045] The system-level controller 101 further selects the operating mode of each prefabricated unit according to the global collaborative scheduling strategy. For each prefabricated unit, the system-level controller 101 reads its candidate operating mode set and calculates the mode score according to formula (5) based on the prefabricated parameter set, installation location, adjacent unit type and real-time load.

[0046] The model score of the candidate operating mode is calculated according to formula (5): (5) in, The mode score is given to the k-th candidate operating mode of the i-th prefabricated unit; Weighting for predicted energy consumption; Normalized predicted energy consumption for the k-th candidate mode of the i-th unit; Weighted by temperature deviation; The normalized temperature deviation of the k-th candidate mode in the i-th unit; Weights for items with insufficient capacity margin; The normalized capacity margin is insufficient for the k-th candidate pattern of the i-th unit; Weighting of equipment lifespan depreciation items; This represents the normalized device life loss of the i-th unit and the k-th candidate mode.

[0047] To ensure the scoring function has a clear physical meaning and eliminates differences in physical dimensions, the input components of the scoring function are specifically calculated using the following normalization formula: In the formula, For pattern Next unit Predicted energy consumption Its rated power; To predict temperature, To set the temperature, This is the maximum permissible deviation; For rated capacity, For pattern Next unit Estimated available capacity; This refers to the equipment lifespan degradation factor.

[0048] The scoring weight coefficients satisfy the following constraints: .

[0049] The system-level controller 101 selects the candidate operating mode whose score ranking is first and whose score difference with the current operating mode reaches the hysteresis switching threshold. The hysteresis switching threshold is calibrated based on the mode switching cost, control cycle length, and equipment response time, and is used to avoid repeated switching of operating modes within adjacent control cycles. The system-level controller 101 issues operating parameter combinations to the refrigeration unit 202 and the power distribution unit 203. The refrigeration unit 202 adjusts the compressor speed, fan speed, coolant flow rate, or supply air temperature; the power distribution unit 203 adjusts the upper limit of power distribution branch, voltage operating range, or energy storage access ratio; and the computing unit 201 adjusts the task load according to the task migration command.

[0050] In scenarios involving controller failure or network isolation, the integrated management unit 103 monitors the system-level controller's heartbeat, computational response, and control network connectivity. When a system-level controller 101 failure, network isolation between the system-level controller 101 and prefabricated units, or a system-level controller heartbeat timeout is detected, the distributed collaborative control module 104 built into the integrated management unit 103 enters takeover mode. On the physical communication path, the computing unit 201, cooling unit 202, power distribution unit 203, system-level controller 101, and integrated management unit 103 establish a communication network using a dual-redundant hardwired ring network. This network ensures continued transmission of forward-looking state summaries and issuance of control commands through bidirectional backup hardwired links when the main control network becomes isolated due to routing failures or network fragmentation. The dual-redundant hardwired ring network consists of two physically independent sets of ring hardwired networks, providing backup communication channels in the event of a control network failure.

[0051] During takeover, the distributed collaborative control module 104 calls the historical policy database of the system-level controller, the most recent global collaborative scheduling policy, and the currently available forward state summary, and calculates the migration optimization objective according to equation (6).

[0052] The migration optimization objective after takeover is calculated according to equation (6): (6) in, Optimize target values ​​for migration during takeover; For prefabricated unit index; This represents the total number of prefabricated units; This is the normalized energy consumption term during the takeover of the i-th prefabricated unit; For the normalized load imbalance term during the takeover of the i-th prefabricated unit; This is the normalized fault risk term during the takeover of the i-th prefabricated unit.

[0053] To prevent semantic overlap or logical confusion with the weight symbol in equation (4) under normal scheduling, the weight symbol in the takeover state is explicitly changed to... , , Furthermore, all weighting coefficients are real numbers between 0 and 1; The following normalization constraints are satisfied under different fault levels: .

[0054] In equation (6), the energy consumption, load imbalance, and failure risk terms are all dimensionless indicators. (Weights) , , The value ranges from 0 to 1 and the sum is 1. When network isolation leads to insufficient digests, the weight of the load imbalance item is increased and the access of new tasks is reduced. When the probability of failure increases, the weight of the failure risk item is increased. If there is no historical policy in the historical policy database that meets the nearest neighbor condition, the integrated management unit 103 maintains the current task load, maintains the cooling safety margin, limits the power distribution load ramp-up rate, and requests the system-level controller 101 to recover and recalculate.

[0055] In takeover mode, the integrated management unit 103 issues a takeover identifier and a takeover policy version number to each unit controller 102. Upon receiving the takeover identifier, the unit controller 102 suspends accepting new policies from the system-level controller 101, accepting only policies issued by the integrated management unit 103, or executes a local backup policy when communication is unavailable. The local backup policy includes maintaining the current task load, restricting new task access, maintaining cooling safety margin, and limiting the power distribution load ramp-up rate.

[0056] After the system-level controller 101 recovers, the integrated management unit 103 performs a strategy consistency check. To eliminate the transient step oscillation (bumpless transfer) that may be caused by hard switching of control, this system is designed with a shadow mode and a disturbance-free switchback mechanism. When the system-level controller 101 is reset and comes back online, the integrated management unit 103 instructs it to enter the "shadow mode". In the shadow mode, the system-level controller 101 is not allowed to issue specific control commands to physical actuators (such as compressor speed controllers or circuit breaker controllers), but instead operates in parallel as a cold standby node without execution rights. The system-level controller 101 reads the current system's hardware parameters, topology, and the latest look-ahead state summary in real time through the dual-redundant hard-wired ring network, and runs the rolling scheduling of formula (4) and the mode scoring calculation of formula (5) in the background.

[0057] During this period, the distributed collaborative control module 104 performs real-time consistency verification on the strategy output by the system-level controller 101 in the background. The consistency verification includes comparing the task allocation differences, cooling parameter differences, and power distribution parameter differences between the global collaborative scheduling strategy calculated by the system-level controller 101 after recovery and the current execution strategy of the integrated management unit 103. When the differences of the three types of parameters are consistently less than the corresponding consistency hard threshold at each sampling moment within the preset continuous verification period (set to 5 consecutive rolling control cycles in this embodiment), it indicates that the algorithm state machine of the system-level controller 101 has completed smooth warm-up and parameter alignment in shadow mode. At this time, the integrated management unit 103 issues control migration instructions to each unit controller 102 to officially close the shadow operation mode and seamlessly switch the hierarchical control function back to the system-level controller 101; if any difference does not meet the condition, it is considered that the system-level controller 101 has not been aligned during cold start, and the integrated management unit 103 continues to take over and requests the system-level controller 101 to re-align and calculate in shadow mode.

[0058] The system in this embodiment includes a system-level controller 101, multiple unit controllers 102, an integrated management unit 103, a flexible data exchange middleware 105, a dual-redundant hardwired ring network, and physical module assemblies with fast liquid-cooled blind connectors and high-current electrical contact sockets. The system-level controller 101 performs initial resource scheduling, forward-looking resource coordination scheduling, and operating mode selection; the unit controllers 102 calculate forward-looking state summaries and publish them on demand through the flexible data exchange middleware 105; the integrated management unit 103 takes over global load balancing and energy efficiency optimization calculations in the event of system-level controller failure, network isolation, or heartbeat timeout. The electronic devices include memory, a processor, unit communication interfaces, and policy output interfaces; the processor executes the computer program to complete the aforementioned continuous control chain.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A prefabricated data center energy efficiency adaptive and robust control method, characterized in that, Includes the following steps: Initialization phase: The system-level controller registers the prefabrication parameter set, installation location, and adjacent unit types for each prefabricated unit; Initial phase: The system-level controller receives real-time task requirements and solves them based on the registered set of pre-configured parameters and the installation location to obtain an initial resource scheduling strategy set; Forward-looking stage: The unit controller built into each prefabricated unit performs forward-looking state calculations based on local historical operating data, generates a forward-looking state summary, and calculates the local state change amount; Data exchange phase: A flexible data exchange middleware runs between the unit controller and the system-level controller. When the local state change meets the preset sending conditions, the look-ahead state summary is sent to the system-level controller. Collaborative optimization phase: The system-level controller receives multiple forward-looking state summaries and merges them with the initial resource scheduling strategy set, the real-time task requirements, and the registered prefabricated parameter set. It then performs forward-looking resource collaborative scheduling within a rolling time window to generate a global collaborative scheduling strategy, thereby determining the operating mode of each prefabricated unit and issuing physical control commands.

2. The prefabricated data center energy efficiency adaptive and robust control method according to claim 1, characterized in that, The pre-set parameter set includes a mapping table of compressor speed and cooling capacity for the refrigeration unit, a mapping table of fan speed and air volume for the fan, voltage and current efficiency curves, rated capacity, thermal inertia parameters, and equipment lifespan impact factors corresponding to each candidate operating parameter for the power distribution unit. The installation location is represented by row and column coordinates, container number, or cabinet group number; If the task constraints or capacity constraints in the real-time task requirements are not feasible, the system-level controller freezes the mode switching of the current control cycle and outputs a conservative resource allocation strategy.

3. The prefabricated data center energy efficiency adaptive and robust control method according to claim 1, characterized in that, In the aforementioned look-ahead phase, generating the look-ahead state summary specifically includes: The unit controller inputs the historical computing load, historical heat load, historical power distribution load, task queue length, ambient temperature, current operating mode identifier, and fault alarm record of the prefabricated unit into the long short-term memory network as the local historical operating data. The Long Short-Term Memory network is used to output the probability distribution of resource demand and the expected failure probability at each predicted moment within the future rolling time window; The unit controller performs heat load prediction and energy consumption prediction calculations based on the local historical operating data, and compresses the resource demand probability distribution, the expected failure probability, the heat load prediction value, and the energy consumption prediction value into the prospective state summary.

4. The prefabricated data center energy efficiency adaptive and robust control method according to claim 3, characterized in that, In the look-ahead phase, the local state change is calculated in the following manner: The unit controller collects the real-time computing load, real-time thermal load, and real-time expected failure probability at the current moment, and obtains the historical load value, thermal load prediction value, and expected failure probability corresponding to the last time the prospective state summary was sent. Calculate the most recent effective window difference between the current value and the corresponding historical value at the last transmission time, and divide each of the most recent effective window differences by the calibrated maximum fluctuation value to obtain the dimensionless calculated load change component, heat load change component, and fault probability change component. The maximum value among the calculated load change component, the heat load change component, and the fault probability change component is taken as the local state change quantity. Its mathematical expression is defined as: in, Let be the local state change of the i-th prefabricated unit at time t; To calculate the load variation components; This represents the component of heat load variation; This represents the component representing the change in failure probability.

5. The prefabricated data center energy efficiency adaptive and robust control method according to claim 4, characterized in that, During the data exchange phase: Sending conditions are If true, in, Whether the i-th prefabricated unit sends a look-ahead state summary at time t; Let be the local state change of the i-th prefabricated unit at time t; The unit dynamic threshold of the i-th prefabricated unit; This is a system-level query event flag; I(·) is a mathematical indicator function; When the value of the indicator function is 1, the flexible data exchange middleware performs digest packaging and publishes the prospective state digest to the system-level controller; When the value of the indicator function is 0, the unit controller only sends a heartbeat signal to the system-level controller.

6. The prefabricated data center energy efficiency adaptive and robust control method according to claim 1, characterized in that, In the aforementioned collaborative optimization phase: Within the rolling time window, the system-level controller converts energy consumption, task migration cost, fault risk, and mode switching cost into dimensionless indices and sums them up. The goal is to minimize the rolling collaborative scheduling target value obtained from the summation and to perform forward-looking resource collaborative scheduling to generate the global collaborative scheduling strategy. The global collaborative scheduling strategy includes task migration instructions, computing capacity start / stop instructions, cooling capacity targets, and power distribution targets.

7. The prefabricated data center energy efficiency adaptive and robust control method according to claim 6, characterized in that, In the collaborative optimization phase, determining the operating mode of each prefabricated unit according to the global collaborative scheduling strategy includes the following steps: The system-level controller reads the candidate operating mode set of the prefabricated unit; Based on the registered prefabricated parameter set, the installation location, the registered adjacent unit type, and the real-time load, a mode score is calculated for each candidate operating mode. The model score is obtained by weighted summation of normalized predicted energy consumption, temperature deviation, insufficient capacity margin, and equipment life loss. The system-level controller sorts the mode scores of each candidate operating mode and selects the candidate operating mode that ranks first in the score ranking and whose score difference with the current operating mode reaches the hysteresis switching threshold as the target operating mode of each prefabricated unit. The system-level controller sends physical control commands corresponding to the target operating mode to the corresponding refrigeration unit and power distribution unit. The physical control commands include combinations of operating parameters.

8. The prefabricated data center energy efficiency adaptive and robust control method according to claim 1, characterized in that, It also includes the takeover phase: The integrated management unit monitors the heartbeat of the system-level controller, the system-level controller's computational response, and the connectivity of the control network; When a system-level controller failure, control network isolation, or heartbeat timeout is detected, the distributed collaborative control module built into the integrated management unit enters the takeover state; In the takeover state, the distributed collaborative control module obtains the prospective state summary sent by each unit controller in real time through the backup link established by the dual redundant hard-wired ring network. The distributed collaborative control module calls the historical policy database, the most recent global collaborative scheduling policy, and the currently obtained forward state summary to calculate the migration optimization target, and publishes the takeover identifier and takeover policy version number to each unit controller to implement alternative control.

9. The prefabricated data center energy efficiency adaptive and robust control method according to claim 1, characterized in that, It also includes the switchback phase: When the system-level controller is reset and brought back online, the integrated management unit instructs the system-level controller to enter shadow operation mode; In the shadow operation mode, the system-level controller operates in parallel as a cold standby node without execution rights. It reads system parameters, topology and the latest prospective state summary in real time through the dual-redundant hard-wired ring network, and performs rolling scheduling and mode scoring strategy calculation in the background. The distributed collaborative control module performs consistency verification between the strategy calculated by the system-level controller in the background and the strategy currently executed by the integrated management unit, comparing the differences in task allocation, cooling parameters, and power distribution parameters between the two. When the differences of the three types of parameters are consistently less than the corresponding consistency hard threshold at each sampling moment within the preset continuous verification period, the integrated management unit issues a control migration instruction to each of the unit controllers to officially close the shadow operation mode and seamlessly switch the hierarchical control function back to the system-level controller.

10. A prefabricated data center energy efficiency adaptive and robust control system, used to execute the prefabricated data center energy efficiency adaptive and robust control method as described in any one of claims 1-9. Its features are, This includes: a system-level controller, multiple unit controllers, a comprehensive management unit, and flexible data exchange middleware; The system-level controller is used to: register the prefabricated parameter set, installation location, and adjacent unit type of each prefabricated unit during the initialization phase; solve the initial resource scheduling strategy set according to real-time task requirements, the prefabricated parameter set, and the installation location during the initial allocation phase; and merge the received multiple forward state summaries, the initial resource scheduling strategy set, the real-time task requirements, and the registered prefabricated parameter set during the collaborative optimization phase, and perform collaborative scheduling within a rolling time window to generate a global collaborative scheduling strategy to determine the operating mode of each prefabricated unit and issue physical control commands. Each of the aforementioned unit controllers is built into the corresponding prefabricated unit and is used to perform forward state calculations based on local historical operating data during the forward look-ahead phase to generate a forward state summary and to calculate the local state change amount. The flexible data exchange middleware runs between the unit controller and the system-level controller, and is used to send the look-ahead state summary to the system-level controller when the local state change meets the preset sending conditions during the data exchange phase. The integrated management unit has a built-in distributed collaborative control module, which is used to enable the distributed collaborative control module to enter the takeover state when the system-level controller is detected to be faulty, the control network is isolated, or the heartbeat timeout occurs during the takeover phase. The module then obtains the prospective state summary through a dual-redundant hardwired ring network to execute alternative control.

Images