A Multi-Objective Cooperative Optimization Control Method and System for Data Center Computing and Cooling

Abstract

This invention provides a multi-objective collaborative optimization control method and system for computing, power, and cooling in data centers. The method includes collecting information technology load, server temperature, power supply and distribution losses, and cooling and energy storage status; constructing a joint optimization model with task allocation, server start / stop, air conditioning supply temperature, chilled water temperature difference, and energy storage power as decision variables, and computing power deviation, power supply power consumption, and cooling power consumption as objectives, and setting boundary constraints; generating an initial candidate solution population; calculating the objective value and constraint violation degree; performing non-dominated sorting; iterative evolution through selection, crossover, and mutation operations; constructing a stagnation criterion based on the objective improvement magnitude, feasible solution ratio, and variable dispersion of continuous multi-generation solutions; dividing the decision variables into strong and weak coupling groups; implementing constraint neighborhood perturbation and diversified expansion search respectively; updating the external retained solution set; selecting the optimal feasible solution through secondary sorting; generating control commands; and realizing collaborative optimization control.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to a multi-objective collaborative optimization control method and system for computing, power and cooling in data centers. Background Technology

[0002] The total energy consumption of a data center mainly consists of three parts: the computing power consumption of IT equipment, the transmission loss of the power supply and distribution system, and the power consumption of the cooling system to maintain the thermal environment. In traditional data center operation and maintenance management, computing power, power supply and distribution, and cooling systems are usually in a state of isolation and independent control. When allocating computing tasks, IT scheduling systems often only aim to meet latency and throughput requirements, completely ignoring the direct impact of uneven spatial distribution of computing load on the thermal environment of the computer room, which can easily lead to local hotspots within the computer room. At the same time, in order to ensure the safe operation of servers, cooling systems generally adopt conservative and fixed low air supply temperature and high air volume strategies, resulting in serious over-cooling and energy waste. Traditional control models cannot achieve overall energy coordination, resulting in persistently high overall power utilization efficiency indicators for data centers, which not only increases the operating costs of enterprises but also becomes a core pain point restricting the development of green and low-carbon data centers. By using task allocation, server start-up and shutdown, air conditioning supply temperature, and energy storage power as joint decision variables, a computing-power-cooling collaborative optimization model is constructed to minimize overall energy consumption while ensuring computing performance.

[0003] However, the data center computing, power, and cooling collaborative model contains a large number of discrete and continuous variables, and there are complex nonlinear coupling relationships between these variables. Conventional evolutionary algorithms are prone to getting stuck in local optima and experiencing search stagnation during multi-objective optimization iterations. Because existing methods lack the ability to identify and differentiate the coupling strength of variables, indiscriminate crossover and mutation operations cannot accurately break the search stagnation, resulting in low optimization accuracy and an inability to approximate the true Pareto front. When selecting an execution scheme from a multi-objective solution set, current technologies typically rely solely on the power consumption target or the optimal value of computing power deviation calculated by the theoretical model as the sole criterion, neglecting the switching costs incurred by the data center transitioning from its current physical state to that target solution state. In real-world operating environments, drastic adjustments that deviate from the current system state not only increase the mechanical wear and lifespan of hardware equipment but also easily trigger oscillations and instability in the cooling and power distribution systems. This static evaluation mechanism, which ignores switching costs, often selects theoretically optimal solutions that lack engineering feasibility and cannot be safely and smoothly translated into underlying actual control commands. Summary of the Invention

[0004] To address the problem that existing technologies suffer from local stagnation in global energy efficiency optimization and difficulty in smoothly executing theoretically optimal solutions due to the isolated control of computing power, power supply and distribution and cooling systems, and the failure to consider the costs of variable coupling and state switching in the optimization process.

[0005] In the first aspect, this disclosure provides a multi-objective collaborative optimization control method for computing power and cooling in data centers, including: Data on information technology load, server temperature, power supply and distribution losses, and cooling and energy storage status are collected. A joint optimization model is constructed with task allocation, server start-up and shutdown, air conditioning supply temperature, chilled water temperature difference, and energy storage power as decision variables, and computing power deviation, power consumption, and cooling power consumption as objectives. Boundary constraints are set to generate an initial candidate solution population, calculate the objective value and constraint violation degree, and perform non-dominated sorting. The population is iteratively evolved through selection, crossover, and mutation operations. Stagnation criteria are constructed based on the target improvement magnitude, feasible solution ratio and variable dispersion of continuous multi-generation solutions. When the stagnation criteria are met, the decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength. A constrained neighborhood perturbation search is performed on the strongly coupled variable group, and a diversified extended search is performed on the weakly coupled variable group. Non-dominated sorting and crowding evaluation are performed on the candidate solutions after the search is updated, and the external reserved solution set is updated. The solutions in the externally retained solution set are sorted a second time based on their target improvement degree, constraint satisfaction, and the switching cost from the current system state to the target solution state. The feasible solution with the best weighted comprehensive evaluation of the normalized target improvement degree and the forward-processed system switching cost is selected as the global optimal solution. Control commands for server equipment, power distribution system, cooling system, and energy storage system are then generated to complete the collaborative optimization control operation of computing power and cooling power consumption in the data center.

[0006] Optionally, the joint optimization model targeting computing power deviation, power consumption, and cooling power consumption, setting boundary constraints, generating an initial candidate solution population, calculating the target value and constraint violation degree, and performing non-dominated sorting, includes: Obtain the upper and lower limits of control for each decision variable, and apply the Latin hypercube sampling algorithm to generate an initial candidate solution population of a specified size in the decision variable space; Substitute the initial candidate solution population into the joint optimization model, and calculate the computing power deviation, power consumption and cooling power consumption values ​​corresponding to each candidate solution, which are used as the independent target values ​​of each candidate solution in the multi-objective dimension. The degree to which each candidate solution deviates from the boundary constraints is detected, the magnitude of the deviation from the boundary is converted into a penalty value and accumulated to generate the corresponding constraint violation degree; Based on the independent objective value and constraint violation of each candidate solution, a non-dominated sorting algorithm is applied to assign the highest Pareto level to the solution that is not dominated by other candidate solutions, and the population is divided into multiple Pareto level levels according to the dominance relationship.

[0007] Optionally, the construction of the stagnation criterion based on the target improvement magnitude, feasible solution ratio, and variable dispersion of continuous multi-generation solutions includes: Extract the target values ​​of the first-level Pareto non-dominated solution set within a continuous preset algebra, normalize and weighted aggregate them to obtain the comprehensive target value, and calculate the ratio of the target reduction difference between the current optimal comprehensive target value solution and the historical corresponding optimal comprehensive target value solution to the initial value, as the target improvement margin; Count the number of candidate solutions in the current population that do not violate any boundary constraints, calculate the percentage of the total population size, and use this as the proportion of feasible solutions; After normalizing each decision variable, the mean Euclidean distance of each candidate solution in the current population to the geometric center of the population in the multidimensional decision space is calculated as the variable dispersion. If the target improvement magnitude and variable dispersion are both lower than their respective lower limit judgment thresholds within a consecutive preset algebra, and the proportion of feasible solutions is higher than the corresponding upper limit judgment threshold, then a stagnation judgment condition is triggered to perform a group search operation.

[0008] Optionally, the division of decision variables into strongly coupled variable groups and weakly coupled variable groups based on coupling strength includes: Obtain historical operation data logs within a preset time period, and extract the time series of decision variable feedback from the computing power system, power supply and distribution system, and refrigeration system under different adjustment conditions; The Spearman rank correlation coefficient algorithm is applied to calculate the correlation coefficient between different pairs of decision variables in the system and generate the decision variable association weight matrix. Extract decision variable combinations whose absolute values ​​of correlation coefficients are greater than preset limits, record positive and negative correlations, and cluster the variable combinations into strongly coupled variable groups. The remaining decision variables that are not assigned to the strongly coupled variable group are assigned to the weakly coupled variable group.

[0009] Optionally, the constrained neighborhood perturbation search of the strongly coupled variable set includes: Extract the solution with the highest Pareto level and the lowest constraint violation from the current population as the benchmark guiding solution; For the combination of variables in the strongly coupled variable group in the benchmark guided solution, select the active variable and superimpose random disturbances that follow a normal distribution and whose amplitude is within a preset step size. For the passive variable in the continuous variable, determine the disturbance direction based on the current variable value and the disturbance direction and positive / negative correlation of the active variable. Superimpose the absolute value of the random disturbances whose amplitude is within a preset step size. Invert the state of the discrete variable to generate multiple neighborhood search solutions. The boundary constraints are used to filter the neighborhood search solutions, eliminating solutions that exceed the upper and lower limits of the decision variables, and retaining feasible solutions that do not violate the constraints to complete the perturbation update of the strongly coupled variable set.

[0010] Optionally, the diversified extended search of the weakly coupled variable set includes: Obtain the upper and lower control limits for each decision variable in a loosely coupled variable group; Based on the reverse learning strategy, the symmetric reverse solution of the weakly coupled continuous variable in the current candidate solution with respect to the center of the interval in the corresponding upper and lower bound interval is calculated, and the state is reversed for the weakly coupled discrete variable to generate the corresponding extended search solution. The original candidate solutions of the current population are merged with the generated extended search solutions to form a mixed set. Inferior solutions are eliminated according to the non-dominated sorting and fitness selection rules, and a specified number of solutions are selected to form the updated weakly coupled variable set solution set.

[0011] Optionally, the step of performing non-dominated ranking and crowding evaluation on the updated candidate solutions and updating the externally retained solution set includes: The candidate solutions generated after the search update are merged with the previous generation population to form a comprehensive population. Non-dominated sorting is performed on the comprehensive population, and the non-dominated solution set with Pareto level 1 is extracted as the initial external retained solution set. For each solution in the initial external retained solution set, calculate the crowding distance parameter between each solution and its neighboring solutions at the same level in the multi-objective dimension space; When the size of the initial external retained solution set exceeds the preset maximum capacity, solutions with spatially crowded distributions are eliminated one by one in order of increasing crowding distance parameter until the size of the solution set is reduced to within the preset capacity limit, thus obtaining the updated external retained solution set.

[0012] In a second aspect, this disclosure provides a multi-objective collaborative optimization control system for data center computing power cooling, comprising: The computing module is used to collect information technology load, server temperature, power supply and distribution loss and cooling energy storage status, construct a joint optimization model with task allocation, server start and stop, air conditioning supply temperature, chilled water temperature difference and energy storage power as decision variables, and computing power deviation, power supply power consumption and cooling power consumption as objectives, and set boundary constraints, generate an initial candidate solution population, calculate the objective value and constraint violation degree and perform non-dominated sorting, and iteratively evolve the population through selection operation, crossover operation and mutation operation; The update module is used to construct stagnation criteria based on the target improvement magnitude, feasible solution ratio and variable dispersion of continuous multi-generation solutions. When the stagnation criteria are met, the decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength. A constrained neighborhood perturbation search is performed on the strongly coupled variable group, and a diversified extended search is performed on the weakly coupled variable group. The candidate solutions after the search update are subjected to non-dominated ranking and crowding evaluation, and the external reserved solution set is updated. The generation module is used to perform secondary sorting based on the target improvement degree, constraint satisfaction, and switching cost of the current system state to the target solution state in the externally retained solution set. The feasible solution with the best weighted comprehensive evaluation of the normalized target improvement degree and the forward-processed system switching cost is selected as the global optimal solution. The module generates control commands for server equipment, power distribution system, cooling system, and energy storage system to complete the collaborative optimization control operation of computing power and cooling power consumption in the data center.

[0013] Furthermore, the joint optimization model targeting computing power deviation, power consumption, and cooling power consumption, and setting boundary constraints, generates an initial candidate solution population, calculates the target value and constraint violation degree, and performs non-dominated sorting, including: Obtain the upper and lower limits of control for each decision variable, and apply the Latin hypercube sampling algorithm to generate an initial candidate solution population of a specified size in the decision variable space; Substitute the initial candidate solution population into the joint optimization model, and calculate the computing power deviation, power consumption and cooling power consumption values ​​corresponding to each candidate solution, which are used as the independent target values ​​of each candidate solution in the multi-objective dimension. The degree to which each candidate solution deviates from the boundary constraints is detected, the magnitude of the deviation from the boundary is converted into a penalty value and accumulated to generate the corresponding constraint violation degree; Based on the independent objective value and constraint violation of each candidate solution, a non-dominated sorting algorithm is applied to assign the highest Pareto level to the solution that is not dominated by other candidate solutions, and the population is divided into multiple Pareto level levels according to the dominance relationship.

[0014] Furthermore, the construction of the stagnation criterion based on the target improvement magnitude, feasible solution ratio, and variable dispersion of continuous multi-generation solutions includes: Extract the target values ​​of the first-level Pareto non-dominated solution set within a continuous preset algebra, normalize and weighted aggregate them to obtain the comprehensive target value, and calculate the ratio of the target reduction difference between the current optimal comprehensive target value solution and the historical corresponding optimal comprehensive target value solution to the initial value, as the target improvement margin; Count the number of candidate solutions in the current population that do not violate any boundary constraints, calculate the percentage of the total population size, and use this as the proportion of feasible solutions; After normalizing each decision variable, the mean Euclidean distance of each candidate solution in the current population to the geometric center of the population in the multidimensional decision space is calculated as the variable dispersion. If the target improvement magnitude and variable dispersion are both lower than their respective lower limit judgment thresholds within a consecutive preset algebra, and the proportion of feasible solutions is higher than the corresponding upper limit judgment threshold, then a stagnation judgment condition is triggered to perform a group search operation.

[0015] Furthermore, the division of decision variables into strongly coupled and weakly coupled variable groups based on coupling strength includes: Obtain historical operation data logs within a preset time period, and extract the time series of decision variable feedback from the computing power system, power supply and distribution system, and refrigeration system under different adjustment conditions; The Spearman rank correlation coefficient algorithm is applied to calculate the correlation coefficient between different pairs of decision variables in the system and generate the decision variable association weight matrix. Extract decision variable combinations whose absolute values ​​of correlation coefficients are greater than preset limits, record positive and negative correlations, and cluster the variable combinations into strongly coupled variable groups. The remaining decision variables that are not assigned to the strongly coupled variable group are assigned to the weakly coupled variable group.

[0016] Furthermore, the constrained neighborhood perturbation search of the strongly coupled variable set includes: Extract the solution with the highest Pareto level and the lowest constraint violation from the current population as the benchmark guiding solution; For the combination of variables in the strongly coupled variable group in the benchmark guided solution, select the active variable and superimpose random disturbances that follow a normal distribution and whose amplitude is within a preset step size. For the passive variable in the continuous variable, determine the disturbance direction based on the current variable value and the disturbance direction and positive / negative correlation of the active variable. Superimpose the absolute value of the random disturbances whose amplitude is within a preset step size. Invert the state of the discrete variable to generate multiple neighborhood search solutions. The boundary constraints are used to filter the neighborhood search solutions, eliminating solutions that exceed the upper and lower limits of the decision variables, and retaining feasible solutions that do not violate the constraints to complete the perturbation update of the strongly coupled variable set.

[0017] Furthermore, the diversified extended search of the weakly coupled variable set includes: Obtain the upper and lower control limits for each decision variable in a loosely coupled variable group; Based on the reverse learning strategy, the symmetric reverse solution of the weakly coupled continuous variable in the current candidate solution with respect to the center of the interval in the corresponding upper and lower bound interval is calculated, and the state is reversed for the weakly coupled discrete variable to generate the corresponding extended search solution. The original candidate solutions of the current population are merged with the generated extended search solutions to form a mixed set. Inferior solutions are eliminated according to the non-dominated sorting and fitness selection rules, and a specified number of solutions are selected to form the updated weakly coupled variable set solution set.

[0018] Furthermore, the step of performing non-dominated ranking and crowding evaluation on the updated candidate solutions and updating the externally retained solution set includes: The candidate solutions generated after the search update are merged with the previous generation population to form a comprehensive population. Non-dominated sorting is performed on the comprehensive population, and the non-dominated solution set with Pareto level 1 is extracted as the initial external retained solution set. For each solution in the initial external retained solution set, calculate the crowding distance parameter between each solution and its neighboring solutions at the same level in the multi-objective dimension space; When the size of the initial external retained solution set exceeds the preset maximum capacity, solutions with spatially crowded distributions are eliminated one by one in order of increasing crowding distance parameter until the size of the solution set is reduced to within the preset capacity limit, thus obtaining the updated external retained solution set.

[0019] This invention breaks through the energy efficiency barriers caused by the independent operation of each subsystem in traditional data centers by collecting multi-dimensional operating status and constructing a joint optimization model. It achieves deep collaborative optimization of computing power allocation, power supply and distribution, and cooling systems, reducing the overall power consumption and computing power deviation of the data center. During the population iteration process of model solving, a variable grouping search strategy based on coupling strength is introduced. For evolutionary stagnation phenomena, constrained neighborhood perturbations and diversified extended searches are applied to strongly and weakly coupled variables respectively, improving solution accuracy and optimization efficiency. When determining the optimal solution, the target improvement degree and system state switching cost are combined for comprehensive evaluation and secondary ranking, fully balancing power reduction effects and equipment operational stability. This avoids frequent equipment switching and excessive wear caused by blindly pursuing extreme indicators, ensuring the smooth and secure implementation of multi-system collaborative control commands in real business scenarios. Attached Figure Description

[0020] Figure 1 A flowchart of a multi-objective collaborative optimization control method for computing power cooling in a data center; Figure 2 This is a schematic diagram of the monitoring of the stagnation criterion. Figure 3 This is a schematic diagram illustrating the optimization effect of computing power deviation. Figure 4 A schematic diagram of the Pareto front distribution of external archives. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Firstly, this invention proposes a multi-objective collaborative optimization control method for computing power and cooling in data centers, such as... Figure 1 As shown, it includes: S1 collects information technology load, server temperature, power supply and distribution losses, and cooling and energy storage status. It constructs a joint optimization model with task allocation, server start-up and shutdown, air conditioning supply temperature, chilled water temperature difference, and energy storage power as decision variables, and computing power deviation, power supply power consumption, and cooling power consumption as objectives. It sets boundary constraints, generates an initial candidate solution population, calculates the objective value and constraint violation degree, performs non-dominated sorting, and iteratively evolves the population through selection, crossover, and mutation operations.

[0023] A monitoring system, combined with network management protocols and industrial Ethernet protocols, is used to collect real-time time-series data on information technology load, server temperature, power supply and distribution losses, and cooling and energy storage status. A joint optimization model is constructed using mixed-integer nonlinear programming, with task allocation, air conditioning supply temperature, chilled water temperature difference, and energy storage power as continuous decision variables, and server start / stop as discrete decision variables. Boundary constraints are set as the physical capacity of computing nodes, the maximum heat dissipation of cooling towers, and the safe range of battery state of charge. The Latin hypercube sampling algorithm from the statistics module of the scientific computing library is used to generate an initial candidate solution population. By writing a target fitness calculation function and using a constraint violation calculation function with a static penalty function, the computing power deviation, power consumption, and cooling power consumption of each individual are calculated. The fast non-dominated sorting operator of the fast non-dominated sorting genetic algorithm with an elite retention strategy is used to divide the population into different levels of non-dominated layers based on Pareto dominance. The binary tournament selection operator, simulated binary crossover operator, and polynomial mutation operator based on a multi-objective optimization library are used to generate offspring populations and merge parents to complete iterative evolution.

[0024] In some embodiments, the joint optimization model can be constructed using a task allocation amount x based on the load rate of each server, a server start / stop state vector s, and an air conditioning supply temperature. Chilled water temperature difference and energy storage charging and discharging power As a decision variable; Define the target computing power deviation as Where c is the server's computing power capacity, d is the computing power required for the task, and ⊙ represents element-wise multiplication; the power consumption target is defined as... ,in and Let be the idle and full-load power consumption of the i-th server, respectively. For power conversion efficiency, For power supply and distribution fixed losses; define the cooling power consumption target as: in The total heat generated by the servers is represented by CoP(·), which is the coefficient of performance regression function fitted based on historical operating data. Boundary constraints include the load rate of each server. ∈[0,1], air supply temperature ∈[18℃, 26℃], chilled water temperature difference ∈[3℃,7℃], energy storage power ∈[-1000kW, 1000kW] and satisfy the state of charge constraints and total computing power supply constraints ,in This represents the minimum computing power required for the task.

[0025] In some embodiments, the joint optimization model targeting computing power deviation, power consumption, and cooling power consumption, setting boundary constraints, generating an initial candidate solution population, calculating the target value and constraint violation degree, and performing non-dominated sorting includes: Obtain the upper and lower limits of control for each decision variable, and apply the Latin hypercube sampling algorithm to generate an initial candidate solution population of a specified size in the decision variable space; Substitute the initial candidate solution population into the joint optimization model, and calculate the computing power deviation, power consumption and cooling power consumption values ​​corresponding to each candidate solution, which are used as the independent target values ​​of each candidate solution in the multi-objective dimension. The degree to which each candidate solution deviates from the boundary constraints is detected, the magnitude of the deviation from the boundary is converted into a penalty value and accumulated to generate the corresponding constraint violation degree; Based on the independent objective value and constraint violation of each candidate solution, a non-dominated sorting algorithm is applied to assign the highest Pareto level to the solution that is not dominated by other candidate solutions, and the population is divided into multiple Pareto level levels according to the dominance relationship.

[0026] When implementing the Latin hypercube sampling algorithm, the pre-defined hard constraint boundary ranges of the multidimensional decision variables are read. For example, the allocation ratio of task allocation must be constrained to [0, 100%], the server start / stop status is limited to the discrete integer set {0, 1}, the air conditioning supply temperature range is set to [18℃, 26℃], the allowable temperature difference range of chilled water is set to [3℃, 7℃], and the charging / discharging power of the energy storage system is set to [-1000kW, 1000kW]. The value range of each decision variable is divided into a specified number of non-overlapping intervals according to the cumulative probability distribution. Sample values ​​are randomly drawn from each interval and randomly combined to generate an initial candidate solution population of size N=200, ensuring that the samples are uniformly distributed in the solution space. The 200 initial candidate solutions are substituted into the constructed joint optimization model in turn to obtain the three-dimensional independent objective value of each candidate solution in the current state. Specifically, the objective value includes the computing power deviation value, which represents the degree of matching between computing power supply and demand; the actual active power loss of the power supply and distribution network, which represents the power consumption of the power supply; and the overall power consumption of the precision air conditioner and chiller, which represents the cooling power consumption. Among them, the computing power deviation value is positively correlated with the task response delay. The smaller the computing power deviation, the lower the task response delay.

[0027] In the multi-objective feasibility assessment stage, it is checked whether the decision variables within each candidate solution exceed the preset boundary, and a constant coefficient is used as the penalty weight for violation. For example, if the energy storage power optimization output value of a candidate solution is 1050kW, which deviates from the upper limit boundary by 50kW, then the deviation of 50 is multiplied by the penalty coefficient M, and the multi-dimensional limit violations are accumulated to form a single constraint violation degree. In the non-dominated sorting stage, it is agreed that feasible solutions with a constraint violation degree of 0 strictly dominate any infeasible solutions; among infeasible solutions, those with smaller constraint violation degrees dominate those with larger constraint violation degrees. In the non-dominated sorting stage, a fast non-dominated sorting strategy similar to that in the NSGA-II framework is adopted to identify all candidate solutions in the current population that have a constraint violation degree of 0 and are not simultaneously superior to any other feasible solutions in the three objective dimensions of computing power deviation, power supply power consumption, and cooling power consumption. These candidate solutions are classified as the first Pareto front and assigned the highest Pareto level 1. The solution set of level 1 is temporarily removed, and the current non-dominated solution is continued to be searched in the remaining solution set and assigned level 2, and so on, to complete the hierarchical division of the entire population based on the dominance relationship.

[0028] S2. Based on the target improvement magnitude, feasible solution ratio and variable dispersion of continuous multi-generation solutions, a stagnation criterion is constructed. When the stagnation criterion is met, the decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength. A constrained neighborhood perturbation search is performed on the strongly coupled variable group, and a diversified extended search is performed on the weakly coupled variable group. The candidate solutions after the search are updated are subjected to non-dominated ranking and crowding evaluation, and the external reserved solution set is updated.

[0029] A monitoring window is set up with 20 consecutive generations. The multi-objective optimization evaluation index library is used to calculate the improvement in hypervolume index, the proportion of feasible solutions satisfying all boundary constraints, and the dispersion of variables (Euclidean distance of all decision variables) for each generation within the window. The weighted sum of these three factors, if below a set threshold, is used as the stagnation criterion. When stagnation is triggered, the mutual information regression algorithm from the feature selection module of the machine learning library is used to calculate the information gain between each pair of decision variables. A hierarchical clustering algorithm is then used to divide the variables into strongly coupled and weakly coupled variable groups based on the mutual information matrix. For the strongly coupled variable group, a truncated Gaussian random number generator is used to perform a perturbation search with a limited step size in the neighborhood of the current value to deeply explore local extrema. For the weakly coupled variable group, the reverse learning strategy is invoked to implement a diversified expansion search with large steps using symmetric mapping features to escape local optima; the new solutions generated by the search are merged with the original population, and the non-dominated sorting operator of the fast non-dominated sorting genetic algorithm with elite retention strategy is invoked again to hierarchically divide the population, and the density of solutions in the same layer is calculated by calling the crowding distance function. According to the rule of non-dominated layer priority and priority of the larger crowding distance in the same layer, redundant individuals are truncated and eliminated, and the retained excellent individuals are stored in the external archive set.

[0030] In some embodiments, the construction of the stagnation criterion based on the target improvement magnitude, feasible solution ratio, and variable dispersion of consecutive multi-generation solutions includes: Extract the target values ​​of the first-level Pareto non-dominated solution set within a continuous preset algebra, normalize and weighted aggregate them to obtain the comprehensive target value, and calculate the ratio of the target reduction difference between the current optimal comprehensive target value solution and the historical corresponding optimal comprehensive target value solution to the initial value, as the target improvement margin; Count the number of candidate solutions in the current population that do not violate any boundary constraints, calculate the percentage of the total population size, and use this as the proportion of feasible solutions; After normalizing each decision variable, the mean Euclidean distance of each candidate solution in the current population to the geometric center of the population in the multidimensional decision space is calculated as the variable dispersion. If the target improvement magnitude and variable dispersion are both lower than their respective lower limit judgment thresholds within a consecutive preset algebra, and the proportion of feasible solutions is higher than the corresponding upper limit judgment threshold, then a stagnation judgment condition is triggered to perform a group search operation.

[0031] During the evolutionary iteration monitoring process, the observation window is set to a continuous preset number of generations, and the target data of the first-level Pareto non-dominated solution set in the above 20 generations is extracted in real time. To eliminate the differences in the dimensions of each dimension, the minimax normalization method is used to map the computing power deviation, power consumption, and cooling power consumption to the [0,1] interval respectively, and weights are set according to the needs of the scenario, such as computing power weight. =0.4, power supply weight =0.3, cooling weight =0.3 is used for linear weighted aggregation to obtain the comprehensive objective value. The reduction of the optimal comprehensive objective value of generation k relative to the optimal value of generation k-20 is calculated using a formula, and divided by the comprehensive objective value of the initial generation to obtain the percentage of objective improvement within the current evolution window. The number of solutions with constraint violation of 0 is selected in the population, and this number is divided by the total population size to calculate the percentage of feasible solutions in the current generation; and the mean Euclidean distance from the current population to the centroid of the multidimensional geometric space of the current generation is calculated using the normalized coordinate vectors of each solution to represent the variable dispersion.

[0032] The stagnation criterion is triggered by the combined application of the three sets of strictly quantitative thresholds mentioned above. For example, when the lower limit threshold for the target improvement is set to 0.1%, the lower limit threshold for variable dispersion is set to 0.05, and the upper limit threshold for the proportion of feasible solutions is set to 90%, the current state will be compared generation by generation. If, within a window period of 20 consecutive generations, the target improvement remains below 0.1%, the variable dispersion remains below 0.05, and the proportion of feasible solutions remains above 90%, it indicates that the algorithm population has highly converged to a certain local space under the premise of high probability of satisfying the constraints, losing the motivation to explore better solutions and the diversity of the population. Once this combination of conditions is met, the engine automatically determines that the population has fallen into an evolutionary stagnation or local optimum state, immediately triggering the subsequent grouping and isolation of decision variables and the jump search operation mechanism, such as... Figure 2 As shown, after about 150 iterations, the target improvement magnitude and variable dispersion both dropped to a low level, while the proportion of feasible solutions remained high, thus satisfying the stagnation criterion triggering condition.

[0033] In some embodiments, dividing the decision variables into strongly coupled variable groups and weakly coupled variable groups based on coupling strength includes: Obtain historical operation data logs within a preset time period, and extract the time series of decision variable feedback from the computing power system, power supply and distribution system, and refrigeration system under different adjustment conditions; The Spearman rank correlation coefficient algorithm is applied to calculate the correlation coefficient between different pairs of decision variables in the system and generate the decision variable association weight matrix. Extract decision variable combinations whose absolute values ​​of correlation coefficients are greater than preset limits, record positive and negative correlations, and cluster the variable combinations into strongly coupled variable groups. The remaining decision variables that are not assigned to the strongly coupled variable group are assigned to the weakly coupled variable group.

[0034] Before implementing variable decoupling, historical operational data logs from the data center's environmental monitoring and IT scheduling platform were retrieved over a preset time period. Long-series feedback operational data, including all decision variables such as computing system load rate, number of servers started and stopped, output current and voltage of power distribution nodes, supply air temperature of precision air conditioners, and temperature difference between chilled water supply and return water of chiller units, were extracted at a fixed sampling frequency. For the highly nonlinear response relationships within the data center caused by thermodynamics or electrical logic, a nonparametric Spearman rank correlation coefficient algorithm was used to rank any two time series pairs of decision variables and calculate the degree of monotonic correlation, thereby constructing a D×D symmetric decision variable association weight matrix.

[0035] Based on the generated correlation weight matrix, a preset limit is set for the absolute value of the correlation coefficient. The matrix elements are traversed, and all pairs of variables with an absolute correlation coefficient greater than or equal to 0.65 are identified. For example, if the Spearman coefficient for task allocation and server start / stop count is 0.82, it is recorded as a pair of variables with a strong positive correlation; similarly, if it is -0.82, it is recorded as a pair of variables with a strong negative correlation. Variables exceeding the limit and their correlation directions are clustered together, forming a strongly coupled variable group with extremely high local sensitivity. For instance, a combination of independent variables, such as the correlation coefficient of only 0.12 between energy storage system power and air conditioning supply temperature, fails the strong coupling test limit and is directly classified into a weakly coupled variable group with less independent moderating influence, thus preparing parameters for the subsequent differentiated targeted neighborhood search.

[0036] In some embodiments, performing a constrained neighborhood perturbation search on the strongly coupled variable set includes: Extract the solution with the highest Pareto level and the lowest constraint violation from the current population as the benchmark guiding solution; For the combination of variables in the strongly coupled variable group in the benchmark guided solution, select the active variable and superimpose random disturbances that follow a normal distribution and whose amplitude is within a preset step size. For the passive variable in the continuous variable, determine the disturbance direction based on the current variable value and the disturbance direction and positive / negative correlation of the active variable. Superimpose the absolute value of the random disturbances whose amplitude is within a preset step size. Invert the state of the discrete variable to generate multiple neighborhood search solutions. The boundary constraints are used to filter the neighborhood search solutions, eliminating solutions that exceed the upper and lower limits of the decision variables, and retaining feasible solutions that do not violate the constraints to complete the perturbation update of the strongly coupled variable set.

[0037] In the currently stagnant population, elite solutions with a Pareto level of 1 and an overall constraint violation of 0 are selected and used as the benchmark guiding solution for the strong coupling perturbation mechanism. Centered on the benchmark guiding solution, for the positive and negative correlation combinations identified within the strongly coupled variable group, a high-impact factor variable is randomly selected from the variable cluster as the active variable. The current value of this active variable is then superimposed with a normally distributed... Furthermore, the random Gaussian perturbation is limited to a preset step size. If the active variable generates an incremental perturbation of +0.5℃, and the determination matrix records a strong negative correlation between it and the passive variable, then the passive variable is required to synchronously adopt the opposite perturbation direction, that is, to superimpose an absolute value of a random perturbation with an amplitude within the corresponding preset step size and a negative sign onto the passive variable; at the same time, for strongly coupled discrete variables, a binary XOR inversion state flip is performed.

[0038] By implementing a composite perturbation with directional linkage between active and passive variables, a specified number of new neighborhood search solutions can be generated in parallel at once. Since the superposition of random perturbation step sizes may cause some solutions to deviate from the equipment's safety control limits, a rigid boundary constraint filtering mechanism will be used to perform numerical truncation and rationality scanning on each generated neighborhood solution. All non-compliant solutions exceeding the limits, such as those with supply air temperatures above 26℃ or temperature differences below 3℃, will be immediately eliminated. Only feasible solutions that strictly do not violate all constraint boundary conditions will be retained to overwrite and update the corresponding old data in the current solution set, thereby completing the local in-depth exploration and perturbation escape of strongly coupled variable groups.

[0039] In some embodiments, the diversified extended search of the weakly coupled variable set includes: Obtain the upper and lower control limits for each decision variable in a loosely coupled variable group; Based on the reverse learning strategy, the symmetric reverse solution of the weakly coupled continuous variable in the current candidate solution with respect to the center of the interval in the corresponding upper and lower bound interval is calculated, and the state is reversed for the weakly coupled discrete variable to generate the corresponding extended search solution. The original candidate solutions of the current population are merged with the generated extended search solutions to form a mixed set. Inferior solutions are eliminated according to the non-dominated sorting and fitness selection rules, and a specified number of solutions are selected to form the updated weakly coupled variable set solution set.

[0040] The system retrieves the preset upper and lower control limits for all continuous and discrete decision variables belonging to the weakly coupled variable group, and uses a global back-learning strategy to calculate the mapping solution. For weakly coupled continuous variables, such as energy storage charging and discharging power range [-1000kW, 1000kW], assuming a candidate solution is currently set to 200kW, the energy storage charging and discharging power value is added to the upper and lower limits and then subtracted from the current value, i.e., -200kW, to obtain the absolutely symmetric back-inverse solution about the geometric midpoint of the control range. For weakly coupled discrete variables, such as the start and stop of some backup computing power nodes, the same logical inversion process is performed, and based on the above operations, a mirror-expanded search solution with a scale equal to the current population size is generated in batches.

[0041] After generating the reverse expansion solution, the current population set containing N solutions is merged with the newly generated N expanded search solutions to form a massive hybrid set of size 2N. Within this vast solution space pool, a fast non-dominated sorting is performed to establish new Pareto superiority levels, and fitness parameters, including crowding distance, are calculated within each level. Following the rule of prioritizing lower-level solutions and then prioritizing higher-crowding distance within the same level, the worst-performing N solutions are filtered out, extracting only the top N high-quality individuals with the highest overall performance and most dispersed spatial distribution. This constitutes a new, weakly coupled variable set of solutions after iterative evolution.

[0042] In some embodiments, performing non-dominated ranking and crowding evaluation on the updated candidate solutions and updating the externally retained solution set includes: The candidate solutions generated after the search update are merged with the previous generation population to form a comprehensive population. Non-dominated sorting is performed on the comprehensive population, and the non-dominated solution set with Pareto level 1 is extracted as the initial external retained solution set. For each solution in the initial external retained solution set, calculate the crowding distance parameter between each solution and its neighboring solutions at the same level in the multi-objective dimension space; When the size of the initial external retained solution set exceeds the preset maximum capacity, solutions with spatially crowded distributions are eliminated one by one in order of increasing crowding distance parameter until the size of the solution set is reduced to within the preset capacity limit, thus obtaining the updated external retained solution set.

[0043] After processing the grouped searches of strongly and weakly coupled variables, the candidate solution set derived from the next-generation perturbation search is directly merged with the parent elite population preserved from previous generations to form a comprehensive evolutionary population that doubles in size. During multi-objective optimization filtering, a comprehensive and rapid non-dominated ranking evaluation is performed on this comprehensive population. The set of top-tier elites that are not overwhelmed by other solutions in any computational power consumption or cooling power consumption optimization dimension is extracted and marked as Pareto first level. This batch of high-quality frontier solutions is unconditionally input into a pre-constructed empty external cache to form the initial external reserved solution set. To prevent the obtained decision combinations from being too concentrated in a certain preference region, for each candidate individual in the initial external reserved solution set, the crowding distance parameter between the solution and its neighboring solutions within the hypercube shell is calculated based on the sum of the normalized side length differences of the maximum and minimum values ​​of each objective within the three-dimensional mapping space of computational power deviation, power consumption, and cooling power consumption formed by the surrounding solutions of the same level.

[0044] During the long-term accumulation process of multidimensional optimization, when the number of externally archived Pareto non-dominated solutions exceeds the system-level preset maximum memory capacity limit set by the administrator, an internal congestion pruning mechanism is triggered. Non-dominated solutions that have crossed the capacity warning line are sorted in ascending order from smallest to largest according to the calculated congestion distance parameter. The cluster point with the smallest parameter, i.e., the most densely and redundantly distributed multi-objective mapping, is located, and these non-dominated solutions are removed sequentially. This pruning process iteratively updates the distance values ​​of neighboring points until the total number of retained solutions drops back to within the 100 capacity limit threshold, thus obtaining an external set of retained solutions that is representative and evenly distributed across the entire system-level global Pareto front of optimization.

[0045] S3. Based on the target improvement degree, constraint satisfaction, and switching cost of the current system state to the target solution state in the externally retained solution set, a second sort is performed. The feasible solution with the best weighted comprehensive evaluation of the normalized target improvement degree and the forward-processed system switching cost is selected as the global optimal solution. Control commands for server equipment, power distribution system, cooling system, and energy storage system are generated to complete the collaborative optimization control operation of computing power and cooling power consumption of the data center.

[0046] The system reads the target vectors of each feasible solution from the external archive set, calculates the absolute value of the target improvement relative to the current operating baseline state, and uses the norm distance function to calculate the system switching cost between the setpoints of the decision variables corresponding to each solution and the actual operating parameters of the current equipment, representing the mechanical losses of equipment adjustment and start-up / shutdown. The system preprocessing library's maximum-minimum normalization function maps the target improvement and system switching cost to the same dimension interval of 0 to 1; and transforms the system switching cost into a positive index by subtracting the maximum value, so that a larger index value indicates a smaller switching cost; the analytic hierarchy process (AHP) is used to determine the static weight vectors of the target improvement and the positively transformed system switching cost, calculates the weighted comprehensive evaluation value of the two, performs a descending secondary sort based on this evaluation value, and extracts the feasible solution ranked first as the global optimal solution. The command parsing script maps discrete decision variables in the global optimal solution to power on / off and wake-up network data packets of the server backplane management controller, and maps continuous decision variables to descriptive state transition interface network request commands and register write commands based on the industrial Ethernet protocol of the data center infrastructure management system. These commands are then sent to the uninterruptible power supply of the power distribution system, the chiller and pump frequency converter of the refrigeration system, and the bidirectional inverter controller of the energy storage system to perform specific operations.

[0047] This experiment used nearly 30 days of actual operational data logs from the data center as test cases. Three groups were set up for comparison: the first group was a standard multi-objective genetic algorithm baseline group without any decoupling or stall handling mechanisms; the second group was a partial ablation group that removed Spearman variable grouping search and retained only Latin hypercube sampling and global random mutation; and the third group was the complete application group of this invention, including variable grouping based on coupling strength and a strong-weak joint search mechanism. The initial population size for all groups was set to 200, the maximum number of evolutionary iterations was limited to 500 generations, and each group was run independently 30 times to calculate the mean.

[0048] After multiple rounds of iterative testing, the first benchmark group achieved an average computing power delay deviation of 12.5%, a power supply and distribution network active power loss of 5800kW, and an overall cooling power consumption of 7600kW, and encountered local convergence around generation 350. The second ablation group reduced its computing power delay deviation to 9.2%, with power supply loss and cooling power consumption decreasing to 5200kW and 6900kW respectively, and the convergence generation being advanced to 280 generations. The complete group of this invention optimized the computing power delay deviation to 4.5%, reduced power supply loss to 4100kW, and reduced cooling power consumption to 5300kW. Furthermore, it could accurately identify stagnant states and trigger jump search within generation 150, generating 100% feasible solutions with the most uniform distribution of the externally retained solution set. Figure 3 As shown, this demonstrates the advantages of the present invention in improving computing power response efficiency and reducing latency deviation.

[0049] It can be seen that the complete set of this invention improves the computational response efficiency by 64% compared with the traditional benchmark set, reduces the overall power supply and cooling energy consumption by about 30%, and improves the global convergence efficiency. This proves that the multidimensional comprehensive stagnation criterion can prevent premature convergence of the algorithm population and avoid meaningless computational waste in inefficient regions. Meanwhile, the strategy of separating variables based on coupling strength plays a crucial role. The linked directional perturbation of strongly coupled variables perfectly matches the thermodynamic and electrical laws of the physical device itself, eliminating solutions that exceed limits. The reverse learning of weakly coupled variables broadens the search range of the global space, enabling the population to successfully escape local optimum traps in high-dimensional and complex joint optimization, such as... Figure 4 As shown, the externally retained solution set generated by this invention achieves optimization in both power supply and cooling objectives, and its distribution uniformity is better than that of traditional methods.

[0050] The second aspect is a multi-objective collaborative optimization control system for data center computing power cooling, comprising: The computing module is used to collect information technology load, server temperature, power supply and distribution loss and cooling energy storage status, construct a joint optimization model with task allocation, server start and stop, air conditioning supply temperature, chilled water temperature difference and energy storage power as decision variables, and computing power deviation, power supply power consumption and cooling power consumption as objectives, and set boundary constraints, generate an initial candidate solution population, calculate the objective value and constraint violation degree and perform non-dominated sorting, and iteratively evolve the population through selection operation, crossover operation and mutation operation; The update module is used to construct stagnation criteria based on the target improvement magnitude, feasible solution ratio and variable dispersion of continuous multi-generation solutions. When the stagnation criteria are met, the decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength. A constrained neighborhood perturbation search is performed on the strongly coupled variable group, and a diversified extended search is performed on the weakly coupled variable group. The candidate solutions after the search update are subjected to non-dominated ranking and crowding evaluation, and the external reserved solution set is updated. The generation module is used to perform secondary sorting based on the target improvement degree, constraint satisfaction, and switching cost of the current system state to the target solution state in the externally retained solution set. The feasible solution with the best weighted comprehensive evaluation of the normalized target improvement degree and the forward-processed system switching cost is selected as the global optimal solution. The module generates control commands for server equipment, power distribution system, cooling system, and energy storage system to complete the collaborative optimization control operation of computing power and cooling power consumption in the data center.

[0051] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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. Unless otherwise limited, 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 said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0052] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0053] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-objective collaborative optimization control method for computing power and cooling in a data center, characterized in that, include: Data on information technology load, server temperature, power supply and distribution losses, and cooling and energy storage status are collected. A joint optimization model is constructed with task allocation, server start-up and shutdown, air conditioning supply temperature, chilled water temperature difference, and energy storage power as decision variables, and computing power deviation, power consumption, and cooling power consumption as objectives. Boundary constraints are set to generate an initial candidate solution population, calculate the objective value and constraint violation degree, and perform non-dominated sorting. The population is iteratively evolved through selection, crossover, and mutation operations. Stagnation criteria are constructed based on the target improvement magnitude, feasible solution ratio and variable dispersion of continuous multi-generation solutions. When the stagnation criteria are met, the decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength. A constrained neighborhood perturbation search is performed on the strongly coupled variable group, and a diversified extended search is performed on the weakly coupled variable group. Non-dominated sorting and crowding evaluation are performed on the candidate solutions after the search is updated, and the external reserved solution set is updated. The solutions in the externally retained solution set are sorted a second time based on their target improvement degree, constraint satisfaction, and the switching cost from the current system state to the target solution state. The feasible solution with the best weighted comprehensive evaluation of the normalized target improvement degree and the forward-processed system switching cost is selected as the global optimal solution. Control commands for server equipment, power distribution system, cooling system, and energy storage system are then generated to complete the collaborative optimization control operation of computing power and cooling power consumption in the data center.

2. The method according to claim 1, characterized in that, The joint optimization model, which targets computing power deviation, power consumption, and cooling power consumption, sets boundary constraints, generates an initial candidate solution population, calculates the target value and constraint violation degree, and performs non-dominated sorting, including: Obtain the upper and lower limits of control for each decision variable, and apply the Latin hypercube sampling algorithm to generate an initial candidate solution population of a specified size in the decision variable space; Substitute the initial candidate solution population into the joint optimization model, and calculate the computing power deviation, power consumption and cooling power consumption values ​​corresponding to each candidate solution, which are used as the independent target values ​​of each candidate solution in the multi-objective dimension. The degree to which each candidate solution deviates from the boundary constraints is detected, the magnitude of the deviation from the boundary is converted into a penalty value and accumulated to generate the corresponding constraint violation degree; Based on the independent objective value and constraint violation of each candidate solution, a non-dominated sorting algorithm is applied to assign the highest Pareto level to the solution that is not dominated by other candidate solutions, and the population is divided into multiple Pareto level levels according to the dominance relationship.

3. The method according to claim 1, characterized in that, The stagnation criterion, constructed based on the target improvement magnitude, feasible solution ratio, and variable dispersion of continuous multi-generation solutions, includes: Extract the target values ​​of the first-level Pareto non-dominated solution set within a continuous preset algebra, normalize and weighted aggregate them to obtain the comprehensive target value, and calculate the ratio of the target reduction difference between the current optimal comprehensive target value solution and the historical corresponding optimal comprehensive target value solution to the initial value, as the target improvement margin; Count the number of candidate solutions in the current population that do not violate any boundary constraints, calculate the percentage of the total population size, and use this as the proportion of feasible solutions; After normalizing each decision variable, the mean Euclidean distance of each candidate solution in the current population to the geometric center of the population in the multidimensional decision space is calculated as the variable dispersion. If the target improvement magnitude and variable dispersion are both lower than their respective lower limit judgment thresholds within a consecutive preset algebra, and the proportion of feasible solutions is higher than the corresponding upper limit judgment threshold, then a stagnation judgment condition is triggered to perform a group search operation.

4. The method according to claim 3, characterized in that, The decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength, including: Obtain historical operation data logs within a preset time period, and extract the time series of decision variable feedback from the computing power system, power supply and distribution system, and refrigeration system under different adjustment conditions; The Spearman rank correlation coefficient algorithm is applied to calculate the correlation coefficient between different pairs of decision variables in the system and generate the decision variable association weight matrix. Extract decision variable combinations whose absolute values ​​of correlation coefficients are greater than preset limits, record positive and negative correlations, and cluster the variable combinations into strongly coupled variable groups. The remaining decision variables that are not assigned to the strongly coupled variable group are assigned to the weakly coupled variable group.

5. The method according to claim 1, characterized in that, The constrained neighborhood perturbation search for the strongly coupled variable set includes: Extract the solution with the highest Pareto level and the lowest constraint violation from the current population as the benchmark guiding solution; For the combination of variables in the strongly coupled variable group in the benchmark guided solution, select the active variable and superimpose random disturbances that follow a normal distribution and whose amplitude is within a preset step size. For the passive variable in the continuous variable, determine the disturbance direction based on the current variable value and the disturbance direction and positive / negative correlation of the active variable. Superimpose the absolute value of the random disturbances whose amplitude is within a preset step size. Invert the state of the discrete variable to generate multiple neighborhood search solutions. The boundary constraints are used to filter the neighborhood search solutions, eliminating solutions that exceed the upper and lower limits of the decision variables, and retaining feasible solutions that do not violate the constraints to complete the perturbation update of the strongly coupled variable set.

6. The method according to claim 1, characterized in that, The diversified extended search for the weakly coupled variable set includes: Obtain the upper and lower control limits for each decision variable in a loosely coupled variable group; Based on the reverse learning strategy, the symmetric reverse solution of the weakly coupled continuous variable in the current candidate solution with respect to the center of the interval in the corresponding upper and lower bound interval is calculated, and the state is reversed for the weakly coupled discrete variable to generate the corresponding extended search solution. The original candidate solutions of the current population are merged with the generated extended search solutions to form a mixed set. Inferior solutions are eliminated according to the non-dominated sorting and fitness selection rules, and a specified number of solutions are selected to form the updated weakly coupled variable set solution set.

7. The method according to any one of claims 1-6, characterized in that, The process of performing non-dominated ranking and crowding evaluation on the updated candidate solutions and updating the externally retained solution set includes: The candidate solutions generated after the search update are merged with the previous generation population to form a comprehensive population. Non-dominated sorting is performed on the comprehensive population, and the non-dominated solution set with Pareto level 1 is extracted as the initial external retained solution set. For each solution in the initial external retained solution set, calculate the crowding distance parameter between each solution and its neighboring solutions at the same level in the multi-objective dimension space; When the size of the initial external retained solution set exceeds the preset maximum capacity, solutions with spatially crowded distributions are eliminated one by one in order of increasing crowding distance parameter until the size of the solution set is reduced to within the preset capacity limit, thus obtaining the updated external retained solution set.

8. A multi-objective collaborative optimization control system for computing power cooling in a data center, characterized in that, include: The computing module is used to collect information technology load, server temperature, power supply and distribution loss and cooling energy storage status, construct a joint optimization model with task allocation, server start and stop, air conditioning supply temperature, chilled water temperature difference and energy storage power as decision variables, and computing power deviation, power supply power consumption and cooling power consumption as objectives, and set boundary constraints, generate an initial candidate solution population, calculate the objective value and constraint violation degree and perform non-dominated sorting, and iteratively evolve the population through selection operation, crossover operation and mutation operation; The update module is used to construct stagnation criteria based on the target improvement magnitude, feasible solution ratio and variable dispersion of continuous multi-generation solutions. When the stagnation criteria are met, the decision variables are divided into strongly coupled variable groups and weakly coupled variable groups according to the coupling strength. A constrained neighborhood perturbation search is performed on the strongly coupled variable group, and a diversified extended search is performed on the weakly coupled variable group. The candidate solutions after the search update are subjected to non-dominated ranking and crowding evaluation, and the external reserved solution set is updated. The generation module is used to perform secondary sorting based on the target improvement degree, constraint satisfaction, and switching cost of the current system state to the target solution state in the externally retained solution set. The feasible solution with the best weighted comprehensive evaluation of the normalized target improvement degree and the forward-processed system switching cost is selected as the global optimal solution. The module generates control commands for server equipment, power distribution system, cooling system, and energy storage system to complete the collaborative optimization control operation of computing power and cooling power consumption in the data center.

9. The system according to claim 8, characterized in that, The joint optimization model, which targets computing power deviation, power consumption, and cooling power consumption, sets boundary constraints, generates an initial candidate solution population, calculates the target value and constraint violation degree, and performs non-dominated sorting, including: Obtain the upper and lower limits of control for each decision variable, and apply the Latin hypercube sampling algorithm to generate an initial candidate solution population of a specified size in the decision variable space; Substitute the initial candidate solution population into the joint optimization model, and calculate the computing power deviation, power consumption and cooling power consumption values ​​corresponding to each candidate solution, which are used as the independent target values ​​of each candidate solution in the multi-objective dimension. The degree to which each candidate solution deviates from the boundary constraints is detected, the magnitude of the deviation from the boundary is converted into a penalty value and accumulated to generate the corresponding constraint violation degree; Based on the independent objective value and constraint violation of each candidate solution, a non-dominated sorting algorithm is applied to assign the highest Pareto level to the solution that is not dominated by other candidate solutions, and the population is divided into multiple Pareto level levels according to the dominance relationship.

10. The system according to claim 8, characterized in that, The stagnation criterion, constructed based on the target improvement magnitude, feasible solution ratio, and variable dispersion of continuous multi-generation solutions, includes: Extract the target values ​​of the first-level Pareto non-dominated solution set within a continuous preset algebra, normalize and weighted aggregate them to obtain the comprehensive target value, and calculate the ratio of the target reduction difference between the current optimal comprehensive target value solution and the historical corresponding optimal comprehensive target value solution to the initial value, as the target improvement margin; Count the number of candidate solutions in the current population that do not violate any boundary constraints, calculate the percentage of the total population size, and use this as the proportion of feasible solutions; After normalizing each decision variable, the mean Euclidean distance of each candidate solution in the current population to the geometric center of the population in the multidimensional decision space is calculated as the variable dispersion. If the target improvement magnitude and variable dispersion are both lower than their respective lower limit judgment thresholds within a consecutive preset algebra, and the proportion of feasible solutions is higher than the corresponding upper limit judgment threshold, then a stagnation judgment condition is triggered to perform a group search operation.

Images