A modular hydrogen UPS intelligent management and fault prediction method suitable for data centers

Abstract

The present application discloses a modular hydrogen-powered UPS intelligent management and fault prediction method suitable for a data center, which relates to the field of intelligent control. The method comprises: obtaining operating parameters of a target data center, wherein the operating parameters include server load, ambient temperature, power demand and network traffic; obtaining a state score of the target data center through a state intelligent identification model based on the above operating parameters; determining the output power of each power supply module through a power supply allocation model based on the above state score and the priority information of each power supply module of the modular hydrogen-powered UPS.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control, and more specifically, to a modular hydrogen UPS intelligent management and fault prediction method suitable for a data center. Background Art

[0002] With the continuous expansion of data center scale and the rapid growth of computing demand, traditional uninterruptible power supply (UPS) systems have been unable to meet the demand for efficient and reliable power supply. When faced with complex and changing load demands and environmental conditions in data centers, traditional UPS systems are often difficult to adjust in real time and dynamically, resulting in low energy efficiency, waste of resources and reduced system reliability. In addition, traditional UPS systems lack intelligent fault prediction and management mechanisms, making it difficult to respond to system failures in a timely manner, which seriously affects the stability of system operation.

[0003] Modular hydrogen UPS systems are gradually gaining attention due to their flexibility and efficient energy management capabilities. However, the management and scheduling of most modular UPS systems currently still rely on preset rules or static configurations, lacking the ability to perceive and dynamically respond to the operating environment in real time. Especially when faced with complex multi-objective optimization requirements, traditional methods are difficult to take into account various indicators at the same time, resulting in difficulty in achieving optimal system performance. In addition, the existing technology lacks comprehensive consideration of the mutual influence between modules, further limiting the overall optimization capabilities of the system. Summary of the invention

[0004] A series of simplified concepts are introduced in the Summary of the Invention section, which will be further described in detail in the Detailed Description of the Invention section. The Summary of the Invention section of this application does not mean to attempt to define the key features and essential technical features of the claimed technical solution, nor does it mean to attempt to determine the scope of protection of the claimed technical solution.

[0005] This application proposes a modular hydrogen UPS intelligent management and fault prediction method suitable for data centers, including:

[0006] Obtaining operating parameters of the target data center, wherein the operating parameters include server load, ambient temperature, power demand, and network traffic;

[0007] Based on the above operating parameters, a state score of the target data center is obtained through a state intelligent identification model;

[0008] Based on the above state scores and the priority information of each energy supply module of the modular hydrogen energy UPS, the output power of each energy supply module is determined through the energy supply allocation model.

[0009] In a feasible implementation, the above-mentioned state intelligent recognition model includes an input layer, a feature extraction layer, a time series processing layer, a multi-dimensional weight calculation layer, a nonlinear transformation and correction layer, a state scoring layer and an output layer;

[0010] The above-mentioned state score of the target data center is obtained through the state intelligent identification model based on the above-mentioned operating parameters, including:

[0011] Input the above operating parameters into the above input layer to obtain standardized operating parameters based on the Z-score method;

[0012] Input the standardized operating parameters into the feature extraction layer to perform principal component analysis to obtain dimension-reduced data, and perform feature extraction on the dimension-reduced data based on an autoencoder to obtain operating feature information;

[0013] The above operation feature information is input into the above time series processing layer to obtain long-term feature sequence and local feature sequence based on LSTM and TCN;

[0014] Input the long-term feature sequence and the local feature sequence into the multi-dimensional weight calculation layer to obtain multi-dimensional weighted features based on a multi-head attention mechanism;

[0015] Input the multi-dimensional weighted features into the non-linear transformation and correction layer for mapping through an activation function to obtain non-linear features;

[0016] The nonlinear features are input into the state scoring layer to obtain the final state score, and the final state score is output through the output layer.

[0017] In a feasible implementation manner, the above-mentioned state scoring layer determines the final state score S by the following formula:

[0018]

[0019] N represents the total number of state categories, α i is the global weight coefficient for each state category, T i represents the time window length of the i-th state category, P i (t) is the probability of the i-th state category at time t, W i represents the weight of the i-th state category, f(x) is the nonlinear transformation function, g(V i (t)) is the correction function, V i (t) is other indicators related to the status of the data center.

[0020] In a feasible implementation manner, the above-mentioned determining the output power of each energy supply module through the energy supply allocation model based on the above-mentioned state score and the priority information of each energy supply module of the modular hydrogen energy UPS includes:

[0021] Determine an objective function, wherein the objective function includes minimizing an energy consumption function, maximizing a reliability function, minimizing a cost function, and a load balancing function;

[0022] Architecture constraints, wherein the constraints include power limit constraints, total power demand constraints, and priority constraints;

[0023] Combine the state scores and priority information to generate a set of possible power allocation schemes

[0024] Evaluate each solution based on the above objective function and find the Pareto optimal solution;

[0025] Repeat the iterations to optimize the power allocation until convergence or reaching a preset number of generations to determine the output power of each functional module.

[0026] In a feasible implementation, the above energy consumption minimization function is:

[0027]

[0028] The above maximized reliability function is:

[0029]

[0030] The above minimization cost function is:

[0031]

[0032] The above load balancing function is:

[0033]

[0034] Where N is the total number of energy supply modules, C i is the energy consumption coefficient of the i-th energy supply module, P i is the output power of the i-th energy supply module, S is the state score of the target data center, R i is the reliability coefficient of the i-th energy supply module, P priority,i is the priority information of the i-th energy supply module, K i is the unit cost coefficient of the i-th energy supply module, P total is the total output power of all energy supply modules.

[0035] In one possible implementation,

[0036] The above power limit constraints are:

[0037]

[0038] The above total power demand constraint is:

[0039]

[0040] The above priority constraints are:

[0041] P i ·P priority，i ≥λ·S

[0042] Among them, P min,i is the minimum output power of the ith energy supply module, P i is the actual output power of the i-th energy supply module, P max,i is the maximum output power of the i-th energy supply module, N is the total number of energy supply modules, P demand is the current total power demand of the data center, P priority,i is the priority information of the i-th energy supply module, S is the state score of the target data center, and λ is an adjustment parameter.

[0043] In a feasible implementation manner, the above adjustment parameters are determined based on the scale, purpose, environmental factors and business criticality of the target data center.

[0044] In a feasible implementation manner, the specific steps of determining the priority information of each energy supply module include:

[0045] Obtain module performance data, environmental data and module location data of each energy supply module;

[0046] A graph neural network is constructed based on the module performance data, environmental data and module location data of each energy supply module obtained as described above, and priority information of each energy supply module is generated based on the graph neural network.

[0047] In a feasible implementation, the above-mentioned construction of a graph neural network based on the module performance data, environmental data and module location data of each energy supply module obtained above, and the generation of priority information of each energy supply module based on the graph neural network includes:

[0048] Each energy supply module is regarded as a node in the graph, and the connection relationship between the energy supply modules is regarded as an edge, wherein the feature vector of the node includes the module performance data, the environment data and the module location data, and the features of the edge include the physical distance between the modules, the heat radiation effect and the shared cooling channel;

[0049] The features of each of the above nodes are convolved based on a multi-layer graph convolutional network, the representation of each node is updated in combination with the information of neighboring nodes, and the priority information of each of the above energy supply modules is output, where the update formula is as follows:

[0050]

[0051] i(k) is the feature representation of node i in the kth layer, N(i) is the set of neighbor nodes of node i, W(k) is the weight matrix of the graph convolutional layer, c ij is the normalization coefficient and σ is the nonlinear activation function.

[0052] In a feasible implementation, the graph neural network is trained based on a hybrid training strategy, which includes a reinforcement learning strategy and a transfer learning strategy. The formula of the hybrid training strategy is:

[0053]

[0054] Among them, θ is the model parameter, η is the learning rate, and J(θ) is the reward function based on reinforcement learning.

[0055] In summary, the method proposed in the present application includes: obtaining the operating parameters of the target data center, wherein the operating parameters include server load, ambient temperature, power demand and network traffic; obtaining the state score of the target data center through a state intelligent identification model based on the above operating parameters; determining the output power of each energy supply module through an energy supply allocation model based on the above state score and the priority information of each energy supply module of the modular hydrogen energy UPS. The method proposed in the embodiment of the present application uses a state intelligent identification model, and the system can intelligently evaluate the overall health status of the data center based on the operating parameters such as server load, ambient temperature, power demand and network traffic obtained in real time, and generate a state score. This score provides an accurate basis for subsequent energy supply optimization, enabling the system to identify and respond to potential risks in a timely manner, thereby improving the safety and stability of the overall operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present specification. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0057] Figure 1 A schematic diagram of the process of intelligent management and fault prediction of a modular hydrogen UPS suitable for a data center provided in an embodiment of the present application. DETAILED DESCRIPTION

[0058] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0059] See also Figure 1 , is a flow chart of a modular hydrogen UPS intelligent management and fault prediction method applicable to a data center provided in an embodiment of the present application. The above method may specifically include:

[0060] S110, obtaining operating parameters of the target data center, wherein the operating parameters include server load, ambient temperature, power demand, and network traffic;

[0061] S120, obtaining a status score of the target data center through a status intelligent identification model based on the above operating parameters;

[0062] S130. Determine the output power of each energy supply module through an energy supply allocation model based on the above state score and the priority information of each energy supply module of the modular hydrogen energy UPS.

[0063] For example, in this embodiment, the operating parameters of the target data center are first obtained. Specifically, the operating parameters include but are not limited to the following key factors:

[0064] Server load: This refers to the current workload of the servers in the target data center, usually measured by indicators such as CPU utilization and memory usage.

[0065] Ambient temperature: This refers to the ambient temperature within the data center, including the temperature in the server room and other critical areas.

[0066] Power demand: This refers to the total amount of electricity currently required by the data center, reflecting the power consumption level of the entire data center.

[0067] Network traffic: This refers to the transmission load of the network in the data center, including data in and out traffic, internal network communication volume, etc.

[0068] These operating parameters are collected in real time through sensors and monitoring systems and provided to subsequent analysis and processing modules.

[0069] After obtaining the above operating parameters, the system enters the step of obtaining the status score of the target data center through the status intelligent identification model based on the above operating parameters. The status intelligent identification model uses the obtained operating parameters to evaluate the current operating status of the target data center.

[0070] After obtaining the status score of the target data center, the system enters the step of determining the output power of each energy supply module through the energy supply allocation model based on the above status score and the priority information of each energy supply module of the modular hydrogen UPS. In this step, the energy supply allocation model will comprehensively consider the status score and the priority information of each energy supply module. Finally, the model determines the specific output power of each energy supply module and issues these power allocation results to each module for execution.

[0071] In summary, the method proposed in the embodiment of the present application uses the state intelligent identification model, and the system can intelligently evaluate the overall health status of the data center based on real-time operating parameters such as server load, ambient temperature, power demand and network traffic, and generate a state score. This score provides an accurate basis for subsequent energy supply optimization, enables the system to identify and respond to potential risks in a timely manner, and improves the safety and stability of the overall operation.

[0072] In some examples, the state intelligent recognition model includes an input layer, a feature extraction layer, a time series processing layer, a multi-dimensional weight calculation layer, a nonlinear transformation and correction layer, a state scoring layer, and an output layer;

[0073] The above-mentioned state score of the target data center is obtained through the state intelligent identification model based on the above-mentioned operating parameters, including:

[0074] Input the above operating parameters into the above input layer to obtain standardized operating parameters based on the Z-score method;

[0075] Input the standardized operating parameters into the feature extraction layer to perform principal component analysis to obtain dimension-reduced data, and perform feature extraction on the dimension-reduced data based on an autoencoder to obtain operating feature information;

[0076] The above operation feature information is input into the above time series processing layer to obtain long-term feature sequence and local feature sequence based on LSTM and TCN;

[0077] Input the long-term feature sequence and the local feature sequence into the multi-dimensional weight calculation layer to obtain multi-dimensional weighted features based on a multi-head attention mechanism;

[0078] Input the multi-dimensional weighted features into the non-linear transformation and correction layer for mapping through an activation function to obtain non-linear features;

[0079] The nonlinear features are input into the state scoring layer to obtain the final state score, and the final state score is output through the output layer.

[0080] Exemplarily, the above operating parameters are input into the above input layer. These operating parameters include server load, ambient temperature, power demand and network traffic. The main function of the input layer is to pre-process these raw data to ensure that they are suitable for subsequent analysis. Specifically, the input layer standardizes the operating parameters based on the Z-score method to obtain standardized operating parameters. The Z-score method sets the mean of the data to 0 and the standard deviation to 1, so that data of different dimensions can be compared and analyzed on the same scale.

[0081] Next, the standardized operating parameters are input into the feature extraction layer. In this layer, principal component analysis (PCA) is first performed to obtain the reduced-dimensional data. The purpose of PCA is to reduce the dimensionality of the data while retaining the most representative features of the data. This step helps to reduce the complexity of the data and eliminate redundant information. Subsequently, the reduced-dimensional data is subjected to feature extraction based on an autoencoder. The autoencoder further extracts deep-level operating feature information from the reduced-dimensional data through unsupervised learning. These feature information provides the basis for subsequent time series processing.

[0082] Then, the above operation feature information is input into the above time series processing layer. In this layer, the time series processing layer obtains long-term feature sequences and local feature sequences based on LSTM (Long Short-Term Memory Network) and TCN (Time Convolution Network) respectively. LSTM is good at capturing long-term dependencies in operation data and can identify trend changes in the operation status of the data center. TCN obtains the local feature sequence of the operation data through time convolution operation, which can capture the changes and fluctuations of the data center in a short period of time. The combination of these two sequences enables the model to fully understand and describe the operation status of the data center.

[0083] Next, the long-term feature sequence and the local feature sequence are input into the multi-dimensional weight calculation layer. The multi-dimensional weight calculation layer calculates the weighted value of each feature dimension based on the multi-head attention mechanism to generate multi-dimensional weighted features. The multi-head attention mechanism assigns different weights to different feature dimensions to ensure that the model can focus on those features that have a greater impact on the operating status of the data center. This step improves the model's sensitivity to key features, making the final status assessment more accurate.

[0084] Then, the multi-dimensional weighted features are input into the non-linear transformation and correction layer. This layer uses an activation function to perform non-linear mapping on the input multi-dimensional weighted features to obtain non-linear features. Activation functions (such as ReLU, LeakyReLU, etc.) can introduce non-linearity, making the model more expressive and able to handle complex non-linear relationships. In addition, the non-linear transformation and correction layer may also include a correction mechanism to adjust the value of the feature to make it more in line with the actual operating environment and business needs.

[0085] Finally, the above nonlinear features are input into the above state scoring layer, where they are integrated and processed to obtain the final state score. The function of the state scoring layer is to convert the multidimensional nonlinear features into a comprehensive state score, which is used to quantify the current operating state of the data center. The above final state score is output through the above output layer for subsequent decision-making and optimization.

[0086] Through the method provided in this embodiment, the various hierarchical structures of the state intelligent identification model are effectively utilized, and the operation status of the target data center is successfully analyzed from multiple dimensions and multiple levels, and a comprehensive status score is generated. This score can accurately reflect the operation health of the data center and provide a scientific basis for subsequent energy supply allocation and system optimization.

[0087] In some examples, the state scoring layer above determines the final state score S as follows:

[0088]

[0089] N represents the total number of state categories, αi is the global weight coefficient of each state category, T i represents the time window length of the i-th state category, P i (t) is the probability of the i-th state category at time t, W i represents the weight of the i-th state category, f(x) is the nonlinear transformation function, g(V i (t)) is the correction function, V i (t) is other indicators related to the status of the data center.

[0090] For example, in this embodiment, N represents the total number of status categories. Status categories refer to the classification of the operating status of the data center into several categories according to different dimensions and indicators in the status intelligent identification model. Each category may represent a different operating status of the data center or a different system component. N reflects the total number of these categories, ensuring that the model can take into account all important status factors in the comprehensive analysis.

[0091] α i is the global weight coefficient for each status category. This coefficient is used to indicate the importance of the i-th status category in the overall status score. The global weight coefficient is pre-set based on the system designer's judgment of the importance of each category. i The system can pay more attention to certain key state categories when calculating the final state score S.

[0092] T i represents the time window length of the i-th state category. The time window length defines the time range considered when evaluating each state category. P is calculated by integration i (t) The average impact within time t, ensuring that the status score reflects the operation of the data center during that time period. The time window can be set based on different analysis requirements, which can be either a short-term instantaneous response or a long-term trend analysis.

[0093] P i (t) is the probability of the i-th state category at time t. This indicates the possibility that the data center is in the i-th state category at a specific time t. The probability P i (t) is obtained by analyzing the operating parameters through the state intelligent identification model, which reflects the proportion of specific state categories in the data center at that point in time.

[0094] W i Represents the weight of the i-th state category. And the global weight coefficient α i Similar, W i It is used to measure the proportion of a certain state category in its own evaluation process. Different from α i The globality of W i More emphasis is placed on the distribution of influence within specific categories, ensuring that the internal characteristics of each category are fully considered in the calculation.

[0095] f(x) is a nonlinear transformation function. When calculating the status score, the nonlinear transformation function f(x) is used to process the original integral value to make it more consistent with the actual system response. This nonlinear transformation can capture the complex relationship in the data center operating status and improve the accuracy of the status score.

[0096] g(Vi (t)) is a correction function. The correction function is used to adjust the influence of the state category when calculating it, so that it can better reflect the actual operation of the data center. The correction function g(V i (t)) Based on other indicators V related to the data center status i (t) Dynamically adjust the status score to ensure that it has strong adaptability and accuracy under different operating conditions.

[0097] V i (t) are other indicators related to the data center status. These indicators may include temperature, humidity, equipment health, etc., which are additional factors that affect the final evaluation of the status category. By combining these indicators, the modified function g(V i (t)) can be used to determine the probability P of the state category i (t) Make finer adjustments to reflect more accurate operating status.

[0098] After the above steps, the state scoring layer comprehensively considers the global weight of each state category, the performance within the time window, nonlinear transformation and correction factors, and finally calculates the final state score S through the above formula. This score provides a quantitative indicator that reflects the current operating status of the target data center and can be used for subsequent energy supply decisions and system optimization.

[0099] In some examples, the above-mentioned determining the output power of each energy supply module through the energy supply allocation model based on the above-mentioned state score and the priority information of each energy supply module of the modular hydrogen energy UPS includes:

[0100] Determine an objective function, wherein the objective function includes minimizing an energy consumption function, maximizing a reliability function, minimizing a cost function, and a load balancing function;

[0101] Architecture constraints, wherein the constraints include power limit constraints, total power demand constraints, and priority constraints;

[0102] Combine the state scores and priority information to generate a set of possible power allocation schemes

[0103] Evaluate each solution based on the above objective function and find the Pareto optimal solution;

[0104] Repeat the iterations to optimize the power allocation until convergence or reaching a preset number of generations to determine the output power of each functional module.

[0105] Exemplarily, first, the system needs to determine the objective function. In this embodiment, the objective function includes the following key components:

[0106] Minimize energy consumption function: This objective function aims to reduce the energy consumption of the entire system and optimize energy efficiency. By minimizing the power consumption of each module during the power allocation process, the overall energy consumption can be minimized.

[0107] Maximizing reliability function: This objective function focuses on the stability and reliability of the system. By giving priority to the use of high-reliability power supply modules, the data center can be ensured to operate stably under various circumstances, thereby maximizing the reliability of the system.

[0108] Minimize cost function: The purpose of this objective function is to minimize operating and maintenance costs while meeting power supply requirements. The system optimizes power distribution and selects the most cost-effective module for energy supply.

[0109] Load balancing function: This objective function aims to achieve load balancing among modules, avoid overloading or underloading of some modules, and thus improve the overall efficiency and stability of the system.

[0110] After determining the objective function, the system then needs to construct constraints. In this embodiment, the above constraints include the following:

[0111] Power limit constraint: The output power of each power supply module must operate within its design range. Specifically, the output power should be between the minimum power and the maximum power.

[0112] Total power demand constraint: The system must meet the current total power demand of the data center equal to the total output power of all power supply modules

[0113] Priority constraint: Based on the priority information of each power supply module and the status score of the target data center, the system needs to ensure that high-priority modules get enough power support when the status is poor. This constraint is adjusted through adjustment parameters to balance the relationship between priority and power allocation.

[0114] After constructing the objective function and constraints, the system will combine the state scores and priority information to generate a set of possible power allocation schemes. These schemes are generated by optimization algorithms under different objective functions and constraints. The system will generate multiple alternatives, each of which represents a possible power allocation strategy.

[0115] After generating the power allocation scheme, the system will evaluate each scheme based on the above objective functions and find the best solution in multi-objective optimization. Specifically, the system will use multi-objective optimization methods to find the Pareto optimal solution. The Pareto optimal solution means that in a certain scheme, the improvement of any objective function will inevitably lead to the degradation of other objective functions. By finding these Pareto optimal solutions, the system can determine the power allocation scheme that best suits the current operating state under various trade-offs.

[0116] After finding the initial Pareto optimal solution, the system will repeat the iteration to optimize the power distribution. In each iteration, the system will adjust the output power of each energy supply module and continue to optimize according to the results of the previous round until the convergence condition is met or the preset number of generations is reached. After the final iteration, the system can determine the specific output power of each energy supply module to ensure that the system operates at the optimal performance in the current state.

[0117] Through the method provided in the above embodiment, the power allocation of the modular hydrogen UPS is successfully optimized based on the state score and module priority information. By considering multiple objective functions and strict constraints, the system can achieve the best balance between energy efficiency, reliability, cost and load balance, ensuring that the data center can maintain efficient and stable energy supply under various operating conditions.

[0118] In some examples, the above energy minimization function is:

[0119]

[0120] The above maximized reliability function is:

[0121]

[0122] The above minimization cost function is:

[0123]

[0124] The above load balancing function is:

[0125]

[0126] Where N is the total number of energy supply modules, C i is the energy consumption coefficient of the i-th energy supply module, P i is the output power of the i-th energy supply module, S is the state score of the target data center, R i is the reliability coefficient of the i-th energy supply module, P priority,i is the priority information of the i-th energy supply module, K i is the unit cost coefficient of the i-th energy supply module, Ptotal is the total output power of all energy supply modules.

[0127] Through the comprehensive optimization of the above four objective functions, the system is able to make trade-offs and decisions in multiple dimensions at the same time, including energy consumption, reliability, cost and load balancing. These objective functions work together to enable the modular hydrogen UPS system to always maintain efficient and stable power supply under different operating conditions while minimizing operating costs and energy consumption.

[0128] In some examples,

[0129] The above power limit constraints are:

[0130]

[0131] The above total power demand constraint is:

[0132]

[0133] The above priority constraints are:

[0134] P i ·P priority，i ≥λ·S

[0135] Among them, P min,i is the minimum output power of the ith energy supply module, P i is the actual output power of the i-th energy supply module, P max,i is the maximum output power of the i-th energy supply module, N is the total number of energy supply modules, P demand is the current total power demand of the data center, P priority,i is the priority information of the i-th energy supply module, S is the state score of the target data center, and λ is an adjustment parameter.

[0136] For example, by setting power limit constraints, total power demand constraints, and priority constraints, it is ensured that the power supply module operates within a reasonable and safe range, meets the power requirements of the data center, and gives priority to the stability of key modules. These constraints work together to ensure that the system always maintains efficient and stable power supply capabilities in a complex and dynamic operating environment.

[0137] In some examples, the aforementioned adjustment parameters are determined based on the size, usage, environmental factors, and business criticality of the target data center.

[0138] Exemplarily, the node parameter λ is determined based on the size, purpose, environmental factors and business criticality of the target data center. This method enables the system to flexibly adjust the priority allocation of energy supply modules under different operating conditions to ensure efficient and stable operation of the system. The following is a detailed explanation and example of this solution.

[0139] 1. Based on the scale of data center:

[0140] Large-scale data centers usually serve a large number of users and devices, with heavy loads and diverse demands. In order to ensure the reliability of the system under high load, the adjustment parameter λ is usually set to a higher value to ensure that high-priority power supply modules get more power allocation.

[0141] Small and medium-sized data centers: The loads of these data centers are relatively balanced and stable. In this case, λ can be set to a medium or low value to strike a balance between reliability and energy efficiency.

[0142] 2. Based on the purpose of data center:

[0143] Data centers with high reliability requirements: such as data centers of financial institutions, government departments or medical systems, any failure may lead to serious consequences. Therefore, the λ value should be high to ensure that key modules can obtain sufficient power support under any circumstances.

[0144] General-purpose data centers: such as those used for backup, development and testing, have lower reliability requirements. Therefore, the λ value can be appropriately reduced, with more attention paid to energy efficiency and cost optimization.

[0145] 3. Based on environmental factors:

[0146] Data centers under harsh environmental conditions: For example, data centers located in areas with high temperature, high humidity, or unstable power supply. In order to adapt to environmental changes, the adjustment parameter λ should be set to a higher value to ensure that high-priority modules can better cope with environmental pressure.

[0147] Conventional data centers: For data centers operating under mild, controlled environmental conditions, λ can be set to a lower value to focus more on cost and energy efficiency optimization.

[0148] 4. Based on business criticality:

[0149] Critical business data center: For data centers that process real-time transactions and critical tasks, λ should be set to a higher value to ensure that the system can prioritize the operation of key modules under high load or abnormal conditions.

[0150] Non-critical data centers: For centers used to store non-critical data or perform background processing, the λ value can be lower to maximize energy efficiency and cost advantages while maintaining necessary stability.

[0151] For example, in a large financial data center, the center needs to process a large number of real-time transactions, has extremely high reliability requirements, and is located in the city center with a relatively stable environment. For such a data center:

[0152] Scale: Large scale, λ=0.9.

[0153] Application: high reliability requirement, λ=0.9.

[0154] Environmental factors: Stable urban environment, λ=0.

[0155] Business criticality: Business critical, λ=1.0.

[0156] Taking all these factors into consideration, the final λ value can be taken as a higher 0.9 to ensure that the system can prioritize the operation of key modules and maintain the continuity and stability of financial transactions under any circumstances.

[0157] For another example, in a medium-sized development and testing data center, which is mainly used for software development and testing, it is located in an area with moderate environmental conditions and has relatively low reliability requirements.

[0158] Scale: Medium, λ=0.5.

[0159] Application: general purpose, λ=0.4.

[0160] Environmental factors: normal environment, λ=0.4.

[0161] Business criticality: non-business critical, λ=0.3.

[0162] The final λ value can be set to 0.4 to ensure that the system pays more attention to cost and energy efficiency optimization while meeting basic operating requirements.

[0163] By determining the adjustment parameter λ based on the size, purpose, environmental factors and business criticality of the data center, the system can flexibly adjust the priority allocation strategy under different operating environments to achieve efficient and reliable power management. This approach ensures that the system can operate in the best way under various conditions, ensuring the stability of critical tasks and maximizing energy saving and cost-effectiveness in appropriate scenarios.

[0164] In some examples, the specific steps of determining the priority information of each energy supply module include:

[0165] Obtain module performance data, environmental data and module location data of each energy supply module;

[0166] A graph neural network is constructed based on the module performance data, environmental data and module location data of each energy supply module obtained as described above, and priority information of each energy supply module is generated based on the graph neural network.

[0167] Exemplarily, the system constructs a graph neural network (GNN) by acquiring module performance data, environmental data, and module location data of each energy supply module, and generates priority information of each energy supply module based on the network.

[0168] After the graph neural network is built, the system generates priority information for each energy supply module based on the graph neural network. Through the multi-layer convolution operation of the graph neural network, the system gradually aggregates the neighbor node information of each node and learns the importance of each energy supply module in the entire network. This learning process can capture the complex relationships between modules and the impact of the overall layout. Finally, the output representation of each node (that is, the feature representation of each energy supply module) is mapped to a priority value. This priority information reflects the importance or urgency of each energy supply module in the current state, and the system will allocate power based on this priority information.

[0169] In some examples, the above-mentioned construction of a graph neural network based on the module performance data, environmental data, and module location data of each energy supply module obtained above, and the generation of priority information of each energy supply module based on the graph neural network includes:

[0170] Each energy supply module is regarded as a node in the graph, and the connection relationship between the energy supply modules is regarded as an edge, wherein the feature vector of the node includes the module performance data, the environment data and the module location data, and the features of the edge include the physical distance between the modules, the heat radiation effect and the shared cooling channel;

[0171] The features of each of the above nodes are convolved based on a multi-layer graph convolutional network, the representation of each node is updated in combination with the information of neighboring nodes, and the priority information of each of the above energy supply modules is output, where the update formula is as follows:

[0172]

[0173] i(k) is the feature representation of node i in the kth layer, N(i) is the set of neighbor nodes of node i, W(k) is the weight matrix of the graph convolutional layer, c ij is the normalization coefficient and σ is the nonlinear activation function.

[0174] Exemplarily, first, the system obtains module performance data, environmental data, and module location data of each energy supply module. These data constitute the input features of the graph neural network:

[0175] Module performance data: including the current output power, energy consumption, historical failure rate, reliability index, etc. of each energy supply module. These data reflect the performance status and operating efficiency of each module.

[0176] Environmental data: including the temperature, humidity, air circulation, etc. in the data center, as well as the local environmental conditions of each module. These data affect the heat dissipation efficiency and stability of the module.

[0177] Module location data: including the physical location of the module in the data center, such as the distance relative to other modules, the distance to the cooling system, etc. These data are used to evaluate the thermal radiation impact between modules, cooling efficiency and the rationality of spatial layout.

[0178] Next, the system builds a graph neural network based on the module performance data, environment data, and module location data obtained above. The steps for building a graph neural network are as follows:

[0179] Node representation: Each energy supply module is represented as a node in the graph neural network, and the node's feature vector includes its module performance data, environmental data, and location data. This feature vector provides a description of the module in various dimensions.

[0180] Edge representation: In the graph neural network, the connection relationship between each energy supply module is represented as an edge. The characteristics of the edge include the physical distance between modules, thermal radiation effects, and shared cooling channels. These edge features reflect the interaction and influence between modules.

[0181] Physical distance: The actual distance between modules affects heat conduction and heat dissipation efficiency.

[0182] Thermal radiation effect: The thermal radiation effect between adjacent modules affects the temperature and operating efficiency of the modules.

[0183] Shared Cooling Channel: Describes whether the modules share the same cooling system or channel, affecting cooling efficiency and temperature control.

[0184] After building the graph neural network, the system performs convolution operations on the features of each of the above nodes based on a multi-layer graph convolution network. Specifically, the system updates the features of each node in the multi-layer graph convolution network, combines the information of its neighboring nodes, and continuously optimizes the representation of the node. The update formula of the convolution operation is as follows:

[0185]

[0186] i(k) is the feature representation of node i in the kth layer, N(i) is the set of neighbor nodes of node i, W(k) is the weight matrix of the graph convolutional layer, c ijis the normalization coefficient and σ is the nonlinear activation function.

[0187] After the multi-layer graph convolution operation is completed, the system outputs the priority information of each of the above energy supply modules. This priority information is based on the final feature representation of each node and reflects the importance of the energy supply module in the current system state. This information is used for subsequent power allocation and management decisions to ensure that the system remains efficient and stable in a complex operating environment.

[0188] In some examples, the graph neural network is trained based on a hybrid training strategy, which includes a reinforcement learning strategy and a transfer learning strategy. The formula of the hybrid training strategy is:

[0189]

[0190] Among them, θ is the model parameter, η is the learning rate, and J(θ) is the reward function based on reinforcement learning.

[0191] Exemplarily, the reinforcement learning strategy is used in this embodiment to guide the graph neural network to gradually optimize its decision-making process in a complex and changing environment. The core of the reinforcement learning strategy is to evaluate the performance of the model in the current state through the reward function J(θ) and adjust the model parameters according to the performance.

[0192] Reward function J(θ): During the training process of the graph neural network, the reward function is used to measure the performance of the model, such as improved energy efficiency, enhanced reliability, reduced failure rate, etc. By setting different reward criteria, the model can be guided to optimize specific goals. For example, if the model successfully reduces the energy consumption of the system, a positive reward is given, thereby encouraging the model to continue to optimize energy-related decisions in future iterations.

[0193] Model parameter update: In each training iteration, the model calculates the gradient of the reward function Get the direction and magnitude of the current parameter adjustment. The learning rate η controls the update step size to ensure that the model can gradually approach the optimal solution.

[0194] Before starting training for a specific data center, the graph neural network can obtain the initial parameters θ from an already trained model (source domain) through transfer learning old These parameters reflect the knowledge and patterns in the source domain and help the model adapt quickly in the target domain (new data center environment). The transfer learning strategy allows the model to be fine-tuned in the new environment and further adjust the model parameters based on the actual situation of the current data center (such as different environmental conditions, load requirements, etc.). This approach effectively reduces the training time and improves the generalization ability of the model.

[0195] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A modular hydrogen UPS intelligent management and fault prediction method suitable for data centers, characterized in that: include: Obtaining operating parameters of the target data center, wherein the operating parameters include server load, ambient temperature, power demand, and network traffic; Acquire a status score of a target data center through a status intelligent identification model based on the operating parameters; The output power of each energy supply module is determined through an energy supply allocation model based on the state score and the priority information of each energy supply module of the modular hydrogen energy UPS.

2. The method according to claim 1, characterized in that The state intelligent recognition model includes an input layer, a feature extraction layer, a time series processing layer, a multi-dimensional weight calculation layer, a nonlinear transformation and correction layer, a state scoring layer and an output layer; The obtaining the status score of the target data center through the status intelligent identification model based on the operating parameters includes: Input the operating parameters into the input layer to obtain standardized operating parameters based on the Z-score method; Inputting the standardized operating parameters into the feature extraction layer for principal component analysis to obtain dimension-reduced data, and performing feature extraction on the dimension-reduced data based on an autoencoder to obtain operating feature information; Input the operation feature information into the time series processing layer to obtain long-term feature sequence and local feature sequence based on LSTM and TCN; Inputting the long-term feature sequence and the local feature sequence into the multi-dimensional weight calculation layer to obtain multi-dimensional weighted features based on a multi-head attention mechanism; Inputting the multi-dimensional weighted features into the non-linear transformation and correction layer for mapping through an activation function to obtain non-linear features; The nonlinear feature is input into the state scoring layer to obtain a final state score, and the final state score is output through the output layer.

3. The method according to claim 1, characterized in that The state scoring layer determines the final state score S as follows: N represents the total number of state categories, α i is the global weight coefficient for each state category, T i represents the time window length of the i-th state category, P i (t) is the probability of the i-th state category at time t, W i represents the weight of the i-th state category, f(x) is the nonlinear transformation function, g(V i (t)) is the correction function, V i (t) is other indicators related to the status of the data center.

4. The method according to claim 1, characterized in that: The step of determining the output power of each energy supply module through an energy supply allocation model based on the state score and the priority information of each energy supply module of the modular hydrogen energy UPS includes: Determining an objective function, wherein the objective function includes minimizing an energy consumption function, maximizing a reliability function, minimizing a cost function, and a load balancing function; Architecture constraints, wherein the constraints include power limit constraints, total power demand constraints, and priority constraints; Combine the state scores and priority information to generate a set of possible power allocation schemes Evaluate each solution based on the objective function to find the Pareto optimal solution; Repeat the iterations to optimize the power allocation until convergence or reaching a preset number of generations to determine the output power of each functional module.

5. The method according to claim 4, characterized in that The minimized energy consumption function is: The maximum reliability function is: The minimization cost function is: The load balancing function is: Where N is the total number of energy supply modules, C i is the energy consumption coefficient of the i-th energy supply module, P i is the output power of the i-th energy supply module, S is the state score of the target data center, R i is the reliability coefficient of the i-th energy supply module, P priority,i is the priority information of the i-th energy supply module, K i is the unit cost coefficient of the i-th energy supply module, P total is the total output power of all energy supply modules.

6. The method according to claim 4, characterized in that The power limit constraint is: The total power demand constraint is: The priority constraints are: P i ·P priority，i ≥λ·S Among them, P min,i is the minimum output power of the ith energy supply module, P i is the actual output power of the i-th energy supply module, P max,i is the maximum output power of the i-th energy supply module, N is the total number of energy supply modules, P demand is the current total power demand of the data center, P priority,i is the priority information of the i-th energy supply module, S is the state score of the target data center, and λ is an adjustment parameter.

7. The method according to claim 6, characterized in that The adjustment parameters are based on the target The data center is determined by its size, purpose, environmental factors and business criticality.

8. The method according to claim 1, characterized in that The specific steps of determining the priority information of each energy supply module include: Obtain module performance data, environmental data and module location data of each energy supply module; A graph neural network is constructed based on the module performance data, environmental data and module location data of each energy supply module, and priority information of each energy supply module is generated based on the graph neural network.

9. The method according to claim 8, characterized in that The step of constructing a graph neural network based on the acquisition of module performance data, environmental data, and module location data of each energy supply module, and generating priority information of each energy supply module based on the graph neural network includes: Each energy supply module is regarded as a node in the graph, and the connection relationship between the energy supply modules is regarded as an edge, wherein the feature vector of the node includes the module performance data, the environment data and the module location data, and the features of the edge include the physical distance between the modules, the thermal radiation effect and the shared cooling channel; The features of each node are convolved based on a multi-layer graph convolutional network, the representation of each node is updated in combination with the information of neighboring nodes, and the priority information of each energy supply module is output, wherein the update formula is as follows: i(k) is the feature representation of node i in the kth layer, N(i) is the set of neighbor nodes of node i, W(k) is the weight matrix of the graph convolutional layer, c ij is the normalization coefficient and σ is the nonlinear activation function.

10. The method according to claim 8, characterized in that The graph neural network is trained based on a hybrid training strategy, which includes a reinforcement learning strategy and a transfer learning strategy. The formula of the hybrid training strategy is: Among them, θ is the model parameter, η is the learning rate, and J(θ) is the reward function based on reinforcement learning.

Images