Method and system for configuring overall electrical scheme of computer room for cabinet layout

Abstract

The present invention provides a method and system for configuring an overall electrical scheme of a computer room for cabinet layout, which relates to the field of electrical design technology, including: establishing a three-dimensional digital model of a computer room and performing functional zoning; optimizing the cabinet layout and determining the maximum capacity; constructing a power load prediction model; using deep reinforcement learning to perform UPS load balancing; designing intelligent bus duct routing and circuit breaker access schemes; constructing a device topology diagram and collecting operating parameters; extracting device association features based on a graph attention network; identifying abnormal power consumption patterns and optimizing control schemes.

Description

Technical Field

[0001] The present invention relates to the field of electrical design technology, and in particular to a method and system for configuring an overall electrical solution of a computer room for cabinet layout. Background Art

[0002] Traditional cabinet layout and space planning methods lack comprehensive consideration of the overall environment of the computer room. They usually adopt empirical layout methods, which make it difficult to make full use of limited space and effectively solve hot spots. This leads to low space utilization in the computer room and poor cooling efficiency. At the same time, it is difficult to meet the special needs of different functional areas.

[0003] Existing power distribution system design methods often adopt static planning, which cannot adapt to the dynamic changes of computer room loads. Traditional load distribution strategies lack flexibility and it is difficult to achieve optimal load balancing of UPS systems. The wiring scheme of the intelligent bus duct system and the access design of the intelligent circuit breaker also lack systematicity and foresight, making it difficult to meet future expansion needs.

[0004] The current computer room monitoring and management system generally has the problem of information islands, lacks in-depth analysis of the relationship between devices, and the existing power consumption pattern recognition method is not accurate enough to effectively identify abnormal power consumption behavior. It lacks a control strategy based on global optimization, resulting in low overall operation efficiency of the computer room and serious energy waste.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for configuring an overall electrical solution of a computer room for cabinet layout, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for configuring an overall electrical solution of a computer room for a cabinet layout, comprising:

[0008] The building area, floor height, load-bearing parameters, depth dimensions and functional zoning requirements of the computer room are collected, and a three-dimensional digital model of the computer room is established. The computer room space is functionally zoned based on the improved particle swarm optimization algorithm to generate multiple computer room areas. For the equipment sub-areas, the layout of hot and cold channels is planned through the computational fluid dynamics model, and a temperature field distribution model considering airflow disturbance is constructed. The cabinet layout scheme is optimized in combination with the improved genetic algorithm to obtain the optimization results and determine the load-bearing constraints based on the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, and a benchmark power density is assigned to each cabinet to generate an initial layout scheme.

[0009] Based on the improved long short-term memory neural network, a cabinet power load prediction model is constructed, the historical data of the power load in the computer room is collected and feature extraction is performed, feature fusion and time series prediction are performed based on the extracted power load features, the power load prediction results are generated and the power distribution system is constructed, the state space of the UPS system group is constructed and dynamic load distribution is performed through a load balancing controller of deep reinforcement learning, and the optimal load distribution strategy is obtained through repeated iterations. Based on the optimal load distribution strategy and the initial layout plan, a routing plan for the intelligent bus duct system is constructed, and the access plan of the intelligent circuit breaker is determined in combination with the hierarchical control architecture to obtain the physical connection plan of the power distribution system;

[0010] Based on the physical connection scheme of the power distribution system, a device topology map is constructed, sensors are set at each node of the device topology map to collect computer room operating parameters, the correlation features between different electrical equipment are extracted based on the improved graph attention network and a device correlation strength matrix is ​​constructed, a power consumption pattern recognition model is constructed based on the device correlation strength matrix and a pre-set multi-layer autoencoder, and the abnormal power consumption pattern is obtained by combining sparse constraints and noise enhancement recognition, the power consumption decision corresponding to the abnormal power consumption pattern is optimized based on the physical constraints and operating status of the power distribution system, and the optimal control plan is generated in combination with the optimal load distribution strategy.

[0011] In an optional embodiment,

[0012] The building area, floor height, load-bearing parameters, depth and functional zoning requirements of the computer room are collected to establish a three-dimensional digital model of the computer room. The computer room space is functionally zoned based on the improved particle swarm optimization algorithm to generate multiple computer room areas including:

[0013] A 3D laser scanner is used to scan the computer room space to obtain point cloud data, wherein the point cloud data corresponds to the building area, floor height, load-bearing parameters, and depth dimensions of the computer room. The scanned point cloud data are fused through point cloud registration technology, and the voxel filtering algorithm is used to reduce noise and sparsely process the point cloud data. The key structural features of the wall and the ground are extracted using a random sampling consistency algorithm, and a 3D digital model of the computer room is established. At the same time, a temperature and humidity sensor network is arranged to collect environmental parameters;

[0014] According to the preset requirements of the functional zoning of the computer room, the computer room space is functionally zoned based on the multi-objective particle swarm algorithm, and an initial particle swarm is generated through logical mapping. For each particle in the initial particle swarm, a real number encoding method is adopted, which includes the boundary coordinates of the functional area and the area ratio information. For each particle, the fitness is calculated and sorted in descending order according to the fitness. The inertia weight is dynamically adjusted according to the sorting result, and the top 20% of the particles are selected for information exchange to generate a new solution space. The iteration is repeated until the preset maximum number of iterations is reached, and multiple computer room areas are obtained by segmentation according to the boundary coordinates of the functional area generated by the last iteration.

[0015] In an optional embodiment,

[0016] For the equipment sub-area, the layout of hot and cold channels is planned by computational fluid dynamics model, and the temperature field distribution model considering airflow disturbance is constructed. The cabinet layout scheme is optimized by combining the improved genetic algorithm, and the optimization results are obtained. The load-bearing constraints are determined according to the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, and the benchmark power density is assigned to each cabinet to generate the initial layout scheme, including:

[0017] For the equipment sub-areas in the computer room area, the computational fluid dynamics model is combined with deep learning to construct an airflow field simulation model, and a convolutional neural network is used to analyze historical operation data to predict load change trends, and the air supply parameters are adaptively adjusted according to the load change trends; guide plates are set at the corners of the cabinet arrangement, intelligent adjustable shutters are installed on the top of the cold channel, and a heat recovery device is set on the top of the hot channel. The cabinet layout is optimized using a multi-objective genetic algorithm, and a simulated annealing mechanism is introduced to set the initial temperature and cooling coefficient to obtain the optimization results;

[0018] Based on the optimization results, the finite element analysis method is used to divide the floor into grid units, the grid size is adaptively encrypted in the high-load area, a nonlinear mechanical model considering the deformation of the floor is established, and an integer programming model for optimizing the installed capacity of the computer room is established. The optimization goal is to maximize the total power of all information technology equipment in the computer room. The static load of the floor does not exceed the static load bearing capacity and the installed density of each functional partition does not exceed the allowable density. The functional relationship between the installed capacity and the number of information technology equipment arranged is established, and the dynamic programming algorithm is used to solve the integer programming model for optimizing the installed capacity. The dynamic programming state is defined to represent the maximum total power when multiple equipment is considered and the total weight does not exceed the given weight, and the transfer equation between adjacent states is established. The initial layout plan that meets the floor bearing capacity constraint is solved.

[0019] In an optional embodiment,

[0020] Based on the improved long short-term memory neural network, a cabinet power load prediction model is constructed, the historical data of the power load in the computer room is collected and feature extraction is performed, feature fusion and time series prediction are performed based on the extracted power load features, the power load prediction results are generated and the power distribution system is constructed, the state space of the UPS system group is constructed and dynamic load distribution is performed through the load balancing controller of deep reinforcement learning, and the optimal load distribution strategy is obtained by repeated iterations. Based on the optimal load distribution strategy and the initial layout plan, a routing plan of the intelligent bus duct system is constructed, and the access plan of the intelligent circuit breaker is determined in combination with the hierarchical control architecture, and the physical connection plan of the distribution system is obtained, including:

[0021] The power monitoring sensor deployed on the cabinet power distribution unit collects the cabinet power load history data. The power monitoring sensor records the cabinet real-time power consumption data every five minutes. The cabinet real-time power consumption data includes voltage, current, active power, and reactive power. The cabinet power load history data is processed for abnormal values ​​and filled with missing values.

[0022] Feature extraction is performed on the historical data of the power load of the cabinet from three dimensions: time domain, frequency domain, and time-frequency domain. In the time domain dimension, the time domain feature is obtained by statistically analyzing the mean, variance, peak value, and waveform factor of the power load. In the frequency domain dimension, the amplitude and phase of different frequency components are obtained by performing fast Fourier transform on the historical data of the power load of the cabinet, and the spectrum characteristics of the power load in the change result are extracted to obtain the frequency domain feature. Wavelet transform is performed on the historical data of the power load of the cabinet to extract the energy distribution characteristics of the power load on different time scales to obtain the time-frequency domain feature.

[0023] An improved long short-term memory neural network model is used to perform feature fusion on the time domain features, frequency domain features, and time-frequency domain features. The improved long short-term memory neural network model comprises an input layer, a long short-term memory layer, a fully connected layer, and an output layer. The long short-term memory layer comprises 128 neurons and adopts a dropout mechanism. The fully connected layer uses a ReLU activation function. The output layer generates a 24-hour electricity load prediction result.

[0024] Based on the power load prediction result, a state space of an uninterruptible power supply system group is constructed, wherein the state space includes the load rate and battery capacity of the uninterruptible power supply module, and a load balancing controller is trained using a deep Q network algorithm, wherein the input of the load balancing controller is the state space, and the output of the load balancing controller is a load distribution action;

[0025] The load balancing controller is iteratively trained to obtain an optimal load distribution strategy, a wiring plan for a three-phase five-wire intelligent bus duct system is constructed based on the optimal load distribution strategy, an intelligent circuit breaker with a hierarchical control architecture is arranged in the three-phase five-wire intelligent bus duct system, and a physical connection plan for a power distribution system is constructed based on the intelligent circuit breaker with the hierarchical control architecture. The physical connection plan for the power distribution system includes a primary main incoming line cabinet, an uninterruptible power supply input and output cabinet, a three-phase five-wire intelligent bus duct system, an inter-row distribution cabinet, and a cabinet power distribution unit. The primary main incoming line cabinet, the uninterruptible power supply input and output cabinet, the three-phase five-wire intelligent bus duct system, the inter-row distribution cabinet, and the cabinet power distribution unit are connected by copper busbars and cables.

[0026] In an optional embodiment,

[0027] The loss function corresponding to the deep Q network algorithm is shown in the following formula:

[0028] ;

[0029] Among them, L(θ) represents the loss value, E represents the expectation, and H δ represents the Huber loss function, r represents the immediate reward, γ represents the discount factor, represents the maximum value of the Q value of all actions a' calculated by the target network with parameters θ' in the next state s', which is used to indicate the expected future reward, It represents the Q value calculated by the target network with parameter θ when taking action a in state s, which is used to indicate the value of the current state and action.

[0030] In an optional embodiment,

[0031] Based on the physical connection scheme of the power distribution system, a device topology map is constructed, sensors are set at each node of the device topology map to collect the operating parameters of the computer room, the correlation features between different electrical devices are extracted based on the improved graph attention network and the device correlation strength matrix is ​​constructed, the power consumption pattern recognition model is constructed based on the device correlation strength matrix and the pre-set multi-layer autoencoder, the abnormal power consumption model is obtained by combining sparse constraints and noise enhancement recognition, the power consumption decision corresponding to the abnormal power consumption pattern is optimized based on the physical constraints and operating status of the power distribution system, and the optimal control plan is generated in combination with the optimal load distribution strategy, including:

[0032] Building a device topology map based on the physical connection scheme of the power distribution system, setting sensors at each node of the device topology map and collecting the operating parameters of the computer room;

[0033] Based on the improved graph attention network, the correlation features between different electrical devices are extracted and the device correlation strength matrix is ​​constructed, wherein the improved graph attention network introduces the attention mechanism and uses multi-head attention to calculate the attention weights between the device node and its neighboring nodes. Specifically, the node features are linearly transformed by the attention function in the form of a scaled dot product, the similarity between the nodes is calculated and the attention coefficient is obtained by normalization through the softmax function, and the features of the neighboring nodes are weighted and aggregated by the attention coefficient. The original features of the nodes are then concatenated with the aggregated features and transformed through a feedforward neural network nonlinearly. After multiple rounds of iterative updates, the node embedding vector containing the device association information is obtained, and finally the embedding vector of each pair of nodes is inner-producted to obtain the device correlation strength matrix;

[0034] A power consumption pattern recognition model is constructed based on the device association strength matrix and a pre-set multi-layer autoencoder, wherein the device association strength matrix is ​​input into a pre-set multi-layer autoencoder network, sparsity is introduced by applying an L1 regularization constraint on the hidden layer neurons, and Gaussian noise with a mean of zero and a variance of a set value is superimposed on the input data to improve noise robustness, and a sparse compressed representation of the power consumption pattern is obtained after training for a set number of iterations;

[0035] The abnormal power consumption model is obtained by combining sparse constraints and noise enhancement recognition, wherein normal power consumption data and known abnormal power consumption data are respectively input into the trained autoencoder, and the respective reconstruction errors are calculated. The reconstruction errors are compared and judged by the preset error threshold. When the reconstruction error exceeds the threshold, it is identified as an abnormal power consumption mode, otherwise it is identified as a normal power consumption mode.

[0036] The power consumption decision corresponding to the abnormal power consumption mode is optimized based on the physical constraints and operating status of the distribution system, wherein the physical constraints and real-time operating status parameters of the distribution system are comprehensively considered, the physical constraints include transformer capacity, line current capacity, voltage qualified range, etc., the real-time operating status parameters include equipment load rate, power supply reliability, power quality, etc. On this basis, a multi-objective optimization model is constructed, and the objective function includes but is not limited to minimizing energy consumption cost, maximizing load balance, etc. The optimal load distribution strategy under abnormal working conditions is obtained by solving the optimization model.

[0037] In an optional embodiment,

[0038] The attention weight between a device node and its neighbor nodes is calculated using multi-head attention as shown in the following formula:

[0039] ;

[0040] in, represents the feature representation of node i after weighting by the multi-head attention mechanism, represents the set of neighbor nodes of node i, K represents the total number of attention heads, and W k represents the linear transformation matrix of the kth attention head, exp represents the exponential function, represents the transpose of the feature vector of node i, h l represents the feature vector of node l, the neighbor node of node i. Indicates concatenating the results of all attention heads.

[0041] A second aspect of an embodiment of the present invention provides a system for configuring an overall electrical solution for a computer room for cabinet layout, including:

[0042] The first unit is used to collect the building area, floor height, load-bearing parameters, depth dimensions and functional zoning requirements of the computer room, establish a three-dimensional digital model of the computer room, perform functional zoning on the computer room space based on the improved particle swarm optimization algorithm, generate multiple computer room areas, and for the equipment sub-areas, plan the layout of hot and cold channels through the computational fluid dynamics model, build a temperature field distribution model considering airflow disturbances, and optimize the cabinet layout plan in combination with the improved genetic algorithm to obtain the optimization results and determine the load-bearing constraints based on the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, assign a benchmark power density to each cabinet, and generate an initial layout plan;

[0043] The second unit is used to build a cabinet power load prediction model based on an improved long short-term memory neural network, collect historical data on the power load of the computer room and perform feature extraction, perform feature fusion and time series prediction based on the extracted power load features, generate power load prediction results and build a power distribution system, build a state space of the UPS system group and perform dynamic load distribution through a load balancing controller of deep reinforcement learning, repeatedly iterate to obtain the optimal load distribution strategy, build a routing plan for the intelligent bus duct system based on the optimal load distribution strategy and the initial layout plan, determine the access plan of the intelligent circuit breaker in combination with the hierarchical control architecture, and obtain the physical connection plan of the power distribution system;

[0044] The third unit is used to construct a device topology map based on the physical connection scheme of the power distribution system, set sensors at each node of the device topology map and collect computer room operating parameters, extract correlation features between different electrical equipment based on the improved graph attention network and construct a device correlation strength matrix, build a power consumption pattern recognition model based on the device correlation strength matrix and a pre-set multi-layer autoencoder, combine sparse constraints and noise enhancement to identify abnormal power consumption patterns, optimize the power consumption decisions corresponding to the abnormal power consumption patterns based on the physical constraints and operating status of the power distribution system, and generate an optimal control plan in combination with the optimal load distribution strategy.

[0045] According to a third aspect of the embodiments of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the embodiments of the present invention is:

[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0052] In the present invention, by establishing a three-dimensional digital model of the computer room and using an improved particle swarm optimization algorithm to perform functional zoning on the computer room space, the architectural parameters, load-bearing constraints and functional requirements of the computer room are taken into consideration, and the reasonable division and utilization of the computer room space are realized. By collecting the historical data of the power load in the computer room and using the improved long short-term memory neural network for feature extraction and time series prediction, an accurate power load prediction model is constructed, which can capture the time dependence and long-term trend of the power load, generate reliable load prediction results, and provide an important basis for the optimization of the power distribution system. Through the load balancing controller of deep reinforcement learning, dynamic load distribution and balancing are realized, and the reliability and energy efficiency of the power distribution system are improved. By setting sensors at each node of the equipment topology map, the operating parameters of the computer room are collected, and the correlation features between devices are extracted using the improved graph attention network, the equipment correlation strength matrix is ​​constructed, which can effectively identify abnormal power consumption behavior and improve the safety and reliability of the power distribution system. In summary, the present invention significantly improves the design and operation efficiency of the data center, reduces energy consumption costs and failure risks, and provides important technical support for building a new generation of highly reliable, energy-efficient and intelligent data centers. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of a flow chart of a method for configuring an overall electrical solution for a computer room according to a cabinet layout according to an embodiment of the present invention;

[0054] Figure 2 The present invention is a schematic diagram of the overall electrical solution configuration system for a computer room with cabinet layout according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0057] Figure 1 FIG. 1 is a flow chart of a method for configuring an overall electrical solution of a computer room for cabinet layout according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] S1. Collect the building area, floor height, load-bearing parameters, depth dimensions and functional zoning requirements of the computer room, establish a three-dimensional digital model of the computer room, perform functional zoning of the computer room space based on the improved particle swarm optimization algorithm, generate multiple computer room areas, plan the layout of hot and cold channels for the equipment sub-areas through the computational fluid dynamics model, build a temperature field distribution model considering airflow disturbances, optimize the cabinet layout plan in combination with the improved genetic algorithm, obtain the optimization results, determine the load-bearing constraints based on the finite element analysis of the computer room structure, use the dynamic programming algorithm to calculate the maximum cabinet capacity of the equipment sub-area, assign a benchmark power density to each cabinet and generate an initial layout plan;

[0059] The fluid dynamics model is a mathematical model used to describe fluid flow and its dynamic behavior, and is usually based on basic equations of fluid mechanics, such as the continuity equation, momentum equation, and energy equation. The temperature field distribution model considering airflow disturbance is a model that incorporates the disturbance factor of airflow into the temperature field calculation, and is usually used for the analysis of phenomena such as heat conduction, convection, and radiation. The improved genetic algorithm is an optimized version of the classical genetic algorithm, and the convergence speed and global search capability of the algorithm are usually improved by adjusting the mechanisms of selection, crossover, and mutation operations. The dynamic programming algorithm is an algorithm for solving optimal decision problems, which improves efficiency by decomposing complex problems into simple sub-problems and avoiding repeated calculations by recording the solutions to the sub-problems. The benchmark power density refers to the power output or energy distribution corresponding to a unit area or unit volume under specific conditions, and is often used to evaluate and compare the energy efficiency of different systems or equipment.

[0060] In an optional embodiment,

[0061] The building area, floor height, load-bearing parameters, depth and functional zoning requirements of the computer room are collected to establish a three-dimensional digital model of the computer room. The computer room space is functionally zoned based on the improved particle swarm optimization algorithm to generate multiple computer room areas including:

[0062] A 3D laser scanner is used to scan the computer room space to obtain point cloud data, wherein the point cloud data corresponds to the building area, floor height, load-bearing parameters, and depth dimensions of the computer room. The scanned point cloud data are fused through point cloud registration technology, and the voxel filtering algorithm is used to reduce noise and sparsely process the point cloud data. The key structural features of the wall and the ground are extracted using a random sampling consistency algorithm, and a 3D digital model of the computer room is established. At the same time, a temperature and humidity sensor network is arranged to collect environmental parameters;

[0063] According to the preset requirements of the functional zoning of the computer room, the computer room space is functionally zoned based on the multi-objective particle swarm algorithm, and an initial particle swarm is generated through logical mapping. For each particle in the initial particle swarm, a real number encoding method is adopted, which includes the boundary coordinates of the functional area and the area ratio information. For each particle, the fitness is calculated and sorted in descending order according to the fitness. The inertia weight is dynamically adjusted according to the sorting result, and the top 20% of the particles are selected for information exchange to generate a new solution space. The iteration is repeated until the preset maximum number of iterations is reached, and multiple computer room areas are obtained by segmentation according to the boundary coordinates of the functional area generated by the last iteration.

[0064] The voxel filtering algorithm is an algorithm for three-dimensional data processing, which improves the quality of three-dimensional reconstruction or image processing by filtering voxel data to remove noise or smooth data. The random sampling consistency algorithm is an algorithm for model estimation and parameter optimization, which is usually used to estimate model parameters by randomly sampling data subsets in the presence of a large amount of noisy data to improve the robustness of the estimation. The logical mapping is a method of converting complex logical relationships into mathematical models, which is usually used in optimization problems to better handle the logical relationships between decision variables.

[0065] A 3D laser scanner is used to scan the computer room space to obtain point cloud data. During the scanning process, the 3D laser scanner is placed at multiple locations in the computer room for multi-site scanning to ensure that complete computer room space information is obtained. The point cloud data obtained includes key information such as the computer room's building area, floor height, load-bearing parameters and depth dimensions.

[0066] The point cloud data obtained by multi-site scanning are fused through point cloud registration technology. Specifically, the iterative closest point algorithm is used for coarse registration, and then the improved non-rigid registration algorithm is used for precise registration. Finally, a complete computer room point cloud model in a unified coordinate system is obtained. The voxel filtering algorithm is used to reduce noise and sparse the fused point cloud data. The point cloud space is divided into several cubic grids, and the center of gravity of the points in each grid is calculated. The center of gravity is used to replace all the points in the grid. By adjusting the grid size, different degrees of noise reduction and sparseness effects can be achieved. For example, for an original point cloud with 10 million points, filtering with a cubic grid with a side length of 5 cm can reduce the number of points to about 1 million.

[0067] The key structural features of the wall and the ground are extracted by using the random sampling consistency algorithm. Several points are randomly selected, the plane equation is fitted, the distance from other points to the plane is calculated, the number of points whose distance is less than the threshold is counted, and the process is repeated many times. The plane with the most internal points is selected as the wall or the ground. The extracted plane features can be used to construct a three-dimensional digital model of the computer room. Based on the extracted structural features, a three-dimensional digital model of the computer room is established. The wall and the ground are fitted and reconstructed using B-spline surfaces to generate a smooth and continuous surface model. At the same time, detailed features such as doors, windows, and columns are added to improve the accuracy and realism of the model. The final three-dimensional digital model can be used for subsequent functional zoning planning.

[0068] A temperature and humidity sensor network is arranged in the computer room to collect environmental parameters. Multiple temperature and humidity sensor nodes are evenly arranged according to the size of the computer room. ZigBee wireless communication technology is used to form a network to achieve low-power, self-organized data transmission. Each sensor node collects temperature and humidity data every 5 minutes and transmits the data to the gateway node for aggregation.

[0069] After obtaining the three-dimensional digital model and environmental parameters of the computer room, the improved multi-objective particle swarm algorithm is used to perform functional zoning of the computer room space according to the pre-set functional zoning requirements of the computer room, and the initial particle swarm is generated through logical mapping. For a computer room with four functional areas, each particle contains 12 variables, which represent the X, Y coordinates and area ratios of the four functional areas. For example, a particle can be represented as (2.5, 3.0, 0.2, 5.0, 6.5, 0.3, 8.0, 4.5, 0.25, 10.5, 7.0, 0.25).

[0070] For each particle in the initial particle group, a real number encoding method is used, which contains the boundary coordinates and area ratio information of the functional area. The fitness value of each particle is calculated, and the factors considered include: whether the area of ​​each functional area meets the requirements, whether the adjacent relationship between functional areas is reasonable, whether the channel width is sufficient, etc. All particles are sorted in descending order according to the fitness value.

[0071] The inertia weight is dynamically adjusted according to the sorting results. For particles with a high ranking, the inertia weight is reduced to enhance the local search capability. For particles with a low ranking, the inertia weight is increased to enhance the global search capability. For example, the inertia weight can be set to 0.4+0.5*(1-current ranking / total number of particles).

[0072] The top 20% particles are selected for information exchange to generate a new solution space. The arithmetic crossover operator is used to perform weighted average of the corresponding variables of the two parent particles to obtain the offspring particles. The mutation operation randomly perturbs certain variables of the offspring particles to increase population diversity. The information exchange and mutation operation process is repeated until the preset maximum number of iterations is reached, such as 1000 times. According to the functional area boundary coordinates generated by the last iteration, the computer room space is divided into multiple sub-areas. The resulting partitioning scheme not only meets the area requirements of each functional area, but also ensures a reasonable layout and channel width.

[0073] In this embodiment, three-dimensional laser scanning technology is used to obtain the spatial information of the computer room, which is more efficient and accurate than traditional manual measurement. Through point cloud processing and feature extraction, a three-dimensional digital model of the computer room can be quickly constructed to provide accurate spatial information for subsequent functional zoning. A temperature and humidity sensor network is arranged to collect environmental parameters, and the temperature and humidity changes in various areas of the computer room can be monitored in real time, which helps to optimize the layout of the air-conditioning system and energy consumption management. By dynamically adjusting the inertia weight and selective information exchange, the convergence speed of the algorithm and the quality of the solution are improved. In summary, this embodiment greatly improves the efficiency and refinement level of computer room planning and design, and lays a solid digital foundation for the intelligent management of the computer room.

[0074] In an optional embodiment,

[0075] For the equipment sub-area, the layout of hot and cold channels is planned by computational fluid dynamics model, and the temperature field distribution model considering airflow disturbance is constructed. The cabinet layout scheme is optimized by combining the improved genetic algorithm, and the optimization results are obtained. The load-bearing constraints are determined according to the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, and the benchmark power density is assigned to each cabinet to generate the initial layout scheme, including:

[0076] For the equipment sub-areas in the computer room area, the computational fluid dynamics model is combined with deep learning to construct an airflow field simulation model, and a convolutional neural network is used to analyze historical operation data to predict load change trends, and the air supply parameters are adaptively adjusted according to the load change trends; guide plates are set at the corners of the cabinet arrangement, intelligent adjustable shutters are installed on the top of the cold channel, and a heat recovery device is set on the top of the hot channel. The cabinet layout is optimized using a multi-objective genetic algorithm, and a simulated annealing mechanism is introduced to set the initial temperature and cooling coefficient to obtain the optimization results;

[0077] Based on the optimization results, the finite element analysis method is used to divide the floor into grid units, the grid size is adaptively encrypted in the high-load area, a nonlinear mechanical model considering the deformation of the floor is established, and an integer programming model for optimizing the installed capacity of the computer room is established. The optimization goal is to maximize the total power of all information technology equipment in the computer room. The static load of the floor does not exceed the static load bearing capacity and the installed density of each functional partition does not exceed the allowable density. The functional relationship between the installed capacity and the number of information technology equipment arranged is established, and the dynamic programming algorithm is used to solve the integer programming model for optimizing the installed capacity. The dynamic programming state is defined to represent the maximum total power when multiple equipment is considered and the total weight does not exceed the given weight, and the transfer equation between adjacent states is established. The initial layout plan that meets the floor bearing capacity constraint is solved.

[0078] The load change trend refers to the change pattern of the load level of a system or equipment within a certain period of time, and is usually used in fields such as load prediction, resource scheduling and equipment management. The finite element analysis method is a numerical technology used to solve the behavior of complex physical systems. It decomposes large problems into small finite elements for accurate calculations. It is widely used in structural analysis, thermal analysis and fluid mechanics, etc. The static load refers to the constant load applied to the structure or material during the analysis process, and is usually used to evaluate the bearing capacity and stability of the material or structure. The dynamic programming state refers to the specific state of the system at a certain moment in the dynamic programming problem. This state affects subsequent decisions and the selection of the optimal path.

[0079] For the equipment sub-area, the layout of the hot and cold channels is planned through the computational fluid dynamics model. Specifically, the ANSYS Fluent software is used to build a three-dimensional model, import the room geometry model and equipment model, set the boundary conditions and meshing parameters, select the k-ε turbulence model, set the inlet wind speed, temperature and other parameters, and perform steady-state solution. The airflow field and temperature field distribution are analyzed to determine the optimal layout of the hot and cold channels.

[0080] A temperature field distribution model considering airflow disturbance is constructed, and the control equation is discretized using the finite volume method to establish a mathematical model including heat transfer mechanisms such as convection and conduction. The influence of factors such as cabinet opening rate and air supply temperature on airflow disturbance is considered, and dynamic grid technology is used to simulate airflow disturbance. The steady-state temperature field distribution is obtained through iterative calculation.

[0081] The cabinet layout scheme is optimized by combining the improved genetic algorithm. First, the cabinet layout is encoded, and the position of each cabinet is represented by a binary string. The fitness function is designed, and indicators such as temperature uniformity and airflow organization efficiency are comprehensively considered. Genetic operations such as tournament selection, single-point crossover and bit reversal mutation are used, and the elite retention strategy is introduced. Through multiple generations of evolutionary iterations, the optimal cabinet layout scheme is obtained.

[0082] To determine the load-bearing constraints, a finite element analysis of the room structure was performed. A finite element model including floor slabs, beams and columns was established, and the mesh units were divided. Material parameters, boundary conditions and load conditions were set. Static analysis was performed using ANSYS software to obtain stress, deformation and other results. The maximum allowable load for each area was determined according to the specification requirements.

[0083] The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, and the problem is decomposed into multiple stages, each corresponding to a type of equipment. State variables are defined to represent the total weight and total power of the installed equipment. A state transition equation is established to describe the change in state after adding a device. Starting from the initial state, the optimal solution of each stage is calculated step by step, and finally the maximum cabinet capacity that meets the load-bearing constraint is obtained.

[0084] Assign a baseline power density to each cabinet. Based on the equipment type and heat dissipation characteristics, the cabinets are divided into three power levels: high, medium, and low. Use cluster analysis to determine the baseline power density value for each level based on historical load data. For example, a high-density cabinet can be assigned 12kW / cabinet, a medium-density cabinet 8kW / cabinet, and a low-density cabinet 4kW / cabinet.

[0085] According to the optimized cabinet layout and the allocated power density, the specific location and power value of each cabinet are determined. The floor plan is drawn using CAD software, and the hot and cold channels, supply and return air vents, etc. are marked. An equipment list is generated, including information such as the number of cabinets and power, to initially form an overall plan for the layout of the computer room and obtain an initial layout plan.

[0086] In this embodiment, the airflow and temperature distribution inside the computer room are accurately simulated through computational fluid dynamics model and temperature field distribution model, and the improved genetic algorithm is combined for optimization to obtain a more reasonable cabinet layout scheme. Finite element analysis is used to determine the load-bearing constraints to ensure the feasibility of the scheme. The simulated annealing mechanism is introduced to set the initial temperature and cooling coefficient. In the optimization process, some poor solutions are accepted to jump out of the local optimum, enhance the global search ability of the algorithm, improve the quality and efficiency of layout optimization, and accurately evaluate the impact of the cabinet layout scheme on the deformation and stress distribution of the floor through the nonlinear mechanical model to ensure the load-bearing safety of the computer room floor. In summary, this embodiment significantly improves the space utilization efficiency, energy efficiency and safety reliability of the computer room, and provides strong technical support for the construction of a new generation of high-density and high-efficiency computer rooms.

[0087] S2. Construct a cabinet power load prediction model based on an improved long short-term memory neural network, collect historical data on power load in the computer room and perform feature extraction, perform feature fusion and time series prediction based on the extracted power load features, generate power load prediction results and construct a power distribution system, construct a state space of a UPS system group and perform dynamic load distribution through a load balancing controller using deep reinforcement learning, repeatedly iterate to obtain an optimal load distribution strategy, construct a wiring plan for the intelligent bus duct system based on the optimal load distribution strategy and the initial layout plan, determine the access plan for the intelligent circuit breaker in combination with the hierarchical control architecture, and obtain a physical connection plan for the power distribution system;

[0088] The UPS system group refers to a component of an uninterruptible power supply system, which is designed to provide backup power to ensure that important equipment can continue to be powered when the main power supply is cut off or fails. The load balancing controller is a device used to optimize the allocation of power resources. It can monitor and adjust the power usage of each load to achieve efficient operation of the overall system. The intelligent bus duct system is a device for power distribution that combines bus ducts and intelligent monitoring technologies. The hierarchical control architecture is a system design method that divides system functions into multiple levels to achieve efficient control and management.

[0089] In an optional embodiment,

[0090] Based on the improved long short-term memory neural network, a cabinet power load prediction model is constructed, the historical data of the power load in the computer room is collected and feature extraction is performed, feature fusion and time series prediction are performed based on the extracted power load features, the power load prediction results are generated and the power distribution system is constructed, the state space of the UPS system group is constructed and dynamic load distribution is performed through the load balancing controller of deep reinforcement learning, and the optimal load distribution strategy is obtained by repeated iterations. Based on the optimal load distribution strategy and the initial layout plan, a routing plan of the intelligent bus duct system is constructed, and the access plan of the intelligent circuit breaker is determined in combination with the hierarchical control architecture, and the physical connection plan of the distribution system is obtained, including:

[0091] The power monitoring sensor deployed on the cabinet power distribution unit collects the cabinet power load history data. The power monitoring sensor records the cabinet real-time power consumption data every five minutes. The cabinet real-time power consumption data includes voltage, current, active power, and reactive power. The cabinet power load history data is processed for abnormal values ​​and filled with missing values.

[0092] Feature extraction is performed on the historical data of the power load of the cabinet from three dimensions: time domain, frequency domain, and time-frequency domain. In the time domain dimension, the time domain feature is obtained by statistically analyzing the mean, variance, peak value, and waveform factor of the power load. In the frequency domain dimension, the amplitude and phase of different frequency components are obtained by performing fast Fourier transform on the historical data of the power load of the cabinet, and the spectrum characteristics of the power load in the change result are extracted to obtain the frequency domain feature. Wavelet transform is performed on the historical data of the power load of the cabinet to extract the energy distribution characteristics of the power load on different time scales to obtain the time-frequency domain feature.

[0093] An improved long short-term memory neural network model is used to perform feature fusion on the time domain features, frequency domain features, and time-frequency domain features. The improved long short-term memory neural network model comprises an input layer, a long short-term memory layer, a fully connected layer, and an output layer. The long short-term memory layer comprises 128 neurons and adopts a dropout mechanism. The fully connected layer uses a ReLU activation function. The output layer generates a 24-hour electricity load prediction result.

[0094] Based on the power load prediction result, a state space of an uninterruptible power supply system group is constructed, wherein the state space includes the load rate and battery capacity of the uninterruptible power supply module, and a load balancing controller is trained using a deep Q network algorithm, wherein the input of the load balancing controller is the state space, and the output of the load balancing controller is a load distribution action;

[0095] The load balancing controller is iteratively trained to obtain an optimal load distribution strategy, a wiring plan for a three-phase five-wire intelligent bus duct system is constructed based on the optimal load distribution strategy, an intelligent circuit breaker with a hierarchical control architecture is arranged in the three-phase five-wire intelligent bus duct system, and a physical connection plan for a power distribution system is constructed based on the intelligent circuit breaker with the hierarchical control architecture. The physical connection plan for the power distribution system includes a primary main incoming line cabinet, an uninterruptible power supply input and output cabinet, a three-phase five-wire intelligent bus duct system, an inter-row distribution cabinet, and a cabinet power distribution unit. The primary main incoming line cabinet, the uninterruptible power supply input and output cabinet, the three-phase five-wire intelligent bus duct system, the inter-row distribution cabinet, and the cabinet power distribution unit are connected by copper busbars and cables.

[0096] The three-phase five-wire intelligent bus duct system is a bus duct system used in three-phase power distribution, and has five lines, generally including three phase lines, a neutral line and a ground line.

[0097] Through the power monitoring sensors deployed on the cabinet power distribution unit, the real-time power consumption data of the cabinet is collected every 5 minutes, including voltage, current, active power and reactive power. The collected cabinet power load historical data is processed for abnormal values, and data points that are obviously deviated from the normal range are removed to ensure the validity of the data. The data with missing values ​​are filled. The missing values ​​can be estimated by the average of the previous and next values, linear interpolation and other methods to ensure the integrity of the data.

[0098] Extract features from the time domain dimension: Calculate the statistical quantities such as the mean, variance, peak value and waveform factor of the power load data to obtain the time domain features that reflect the load change trend and fluctuation characteristics. Extract features from the frequency domain dimension: Perform fast Fourier transform (FFT) on the cabinet power load historical data to obtain the amplitude and phase information of different frequency components and extract the distribution characteristics of the load in the frequency domain. Extract features from the time-frequency domain dimension: Perform wavelet transform on the cabinet power load historical data to extract the energy distribution characteristics of the load on different time scales and capture the time-varying and multi-scale characteristics of the load.

[0099] The improved long short-term memory neural network model is used to fuse the extracted time domain, frequency domain and time-frequency domain features. The improved long short-term memory neural network model includes input layer, LSTM layer, fully connected layer and output layer. Among them, the LSTM layer contains 128 neurons and uses the dropout mechanism to prevent overfitting; the fully connected layer uses the ReLU activation function to introduce nonlinearity. The output of the model is the power load prediction result for the next 24 hours, which can provide the load change trend for the next day;

[0100] Based on the load prediction results, the state space of the uninterruptible power supply system (UPS) group is constructed. The state space contains key parameters such as the load rate and battery capacity of the UPS module. The deep Q network (DQN) algorithm is used to train the load balancing controller, which takes the state space as input and the load distribution action as output. The load balancing strategy is optimized through continuous interaction and learning with the environment. The load balancing controller is iteratively trained to finally obtain the optimal load distribution strategy, achieve balanced load distribution within the UPS group, and improve the reliability and energy efficiency of the system.

[0101] Based on the optimal load distribution strategy, a wiring scheme for the three-phase five-wire intelligent bus duct system is constructed, and the power distribution path and access points are reasonably planned to reduce line loss and voltage drop. An intelligent circuit breaker with a hierarchical control architecture is set up in the intelligent bus duct system to achieve refined protection and management of the power distribution system. Based on the intelligent circuit breaker with a hierarchical control architecture, a physical connection scheme for the power distribution system is constructed, including the primary main incoming line cabinet, UPS input and output cabinet, three-phase five-wire intelligent bus duct system, inter-row power distribution cabinet and cabinet power distribution unit. Each power distribution unit is reliably connected through copper busbars and cables to form a complete power distribution system topology to ensure the safety and stability of power supply.

[0102] Exemplarily, the above-mentioned intelligent power distribution optimization solution is adopted in the computer room of a data center. By deploying power monitoring sensors on the power distribution units of 100 cabinets, power consumption data, including voltage, current, active power and reactive power, are collected every 5 minutes. The collected historical data are processed for outliers, 2% of abnormal data points are eliminated, and the missing values ​​are filled by linear interpolation. The data integrity reaches more than 99%, and load characteristics are extracted from three dimensions: time domain, frequency domain and time-frequency domain. Time domain features include mean, variance, peak value and waveform factor; frequency domain features are obtained by FFT, and the amplitude and phase of the main frequency components are extracted; time-frequency domain features are extracted by wavelet transform, and energy distribution on different time scales is obtained;

[0103] Using the extracted features, a load prediction model based on the improved LSTM model was constructed. The LSTM layer of the model contains 128 neurons, the dropout rate is set to 0.2, and the fully connected layer uses the ReLU activation function. The model training uses historical data from the past year to predict the load changes in the next 24 hours. The average prediction error is controlled within 5%. Based on the load prediction results, a UPS group state space consisting of 5 UPS modules is constructed, and the DQN algorithm is used to train the load balancing controller. After 10,000 iterative training, the load balancing controller converged to the optimal strategy, achieving balanced distribution of load within the UPS group, increasing the average load rate by 15%, and reducing energy consumption by 10%;

[0104] In this embodiment, through high-frequency, multi-parameter real-time sampling, comprehensive and accurate cabinet power load data is obtained, which provides a data basis for subsequent load analysis and optimization. By constructing the UPS group state space and load distribution action space, the controller can autonomously learn and optimize the load balancing strategy according to the current state and predicted load, realize the dynamic balance of the load within the UPS group, and improve the reliability and energy efficiency of the power supply system. Reasonable distribution path planning reduces line loss and voltage drop. The hierarchical control of the intelligent circuit breaker improves the reliability of fault isolation and protection, and enhances the safety and controllability of the distribution system. In summary, this embodiment improves power supply reliability and energy utilization efficiency; the optimized distribution topology and intelligent protection management enhance the safety and maintainability of the system, and lays a solid foundation for the efficient, stable and safe operation of the data center, which is of great significance to improving the overall performance and operational efficiency of the data center.

[0105] In an optional embodiment,

[0106] The loss function corresponding to the deep Q network algorithm is shown in the following formula:

[0107] ;

[0108] Among them, L(θ) represents the loss value, E represents the expectation, and H δ represents the Huber loss function, r represents the immediate reward, γ represents the discount factor, represents the maximum value of the Q value of all actions a' calculated by the target network with parameters θ' in the next state s', which is used to indicate the expected future reward, It represents the Q value calculated by the target network with parameter θ when taking action a in state s, which is used to indicate the value of the current state and action.

[0109] In this embodiment, by using the target network and experience replay, the stability and convergence of the algorithm are improved, and the over-estimation problem in Q-learning is avoided. The Huber loss function can reduce the influence of outliers, making the training process more stable and robust. It can handle high-dimensional state space and approximate complex value functions through deep neural networks. It is suitable for large-scale reinforcement learning problems. In summary, this embodiment can effectively improve the learning efficiency and decision-making quality of the deep Q-network algorithm and promote the performance of intelligent agents in complex environments.

[0110] S3. Construct a device topology map based on the physical connection scheme of the power distribution system, set sensors at each node of the device topology map and collect the operating parameters of the computer room, extract the correlation features between different electrical equipment based on the improved graph attention network and construct a device correlation strength matrix, construct a power consumption pattern recognition model based on the device correlation strength matrix and a pre-set multi-layer autoencoder, combine sparse constraints and noise enhancement to identify abnormal power consumption patterns, optimize the power consumption decisions corresponding to the abnormal power consumption patterns based on the physical constraints and operating status of the power distribution system, and generate an optimal control plan in combination with the optimal load distribution strategy.

[0111] The device topology map is a graphical representation of devices and their interconnections, and is typically used to visualize the structure of a power system or network. The improved graph attention network is a deep learning model that combines graph structure data and an attention mechanism. The association strength matrix is ​​a matrix that represents the strength of relationships between different devices or nodes. By quantifying the degree of association between each pair of nodes, the association strength matrix provides important structural information for the model, which can assist in analyzing the impact and interaction between devices. The power consumption pattern recognition model is a model constructed based on historical power consumption data through machine learning or deep learning algorithms, and is intended to identify and classify users' power consumption behavior. The abnormal power consumption pattern refers to power consumption that is significantly different from normal power consumption behavior, which may indicate equipment failure, power fraud or other abnormal conditions.

[0112] In an optional embodiment,

[0113] Based on the physical connection scheme of the power distribution system, a device topology map is constructed, sensors are set at each node of the device topology map to collect the operating parameters of the computer room, the correlation features between different electrical devices are extracted based on the improved graph attention network and the device correlation strength matrix is ​​constructed, the power consumption pattern recognition model is constructed based on the device correlation strength matrix and the pre-set multi-layer autoencoder, the abnormal power consumption model is obtained by combining sparse constraints and noise enhancement recognition, the power consumption decision corresponding to the abnormal power consumption pattern is optimized based on the physical constraints and operating status of the power distribution system, and the optimal control plan is generated in combination with the optimal load distribution strategy, including:

[0114] Building a device topology map based on the physical connection scheme of the power distribution system, setting sensors at each node of the device topology map and collecting the operating parameters of the computer room;

[0115] Based on the improved graph attention network, the correlation features between different electrical devices are extracted and the device correlation strength matrix is ​​constructed, wherein the improved graph attention network introduces the attention mechanism and uses multi-head attention to calculate the attention weights between the device node and its neighboring nodes. Specifically, the node features are linearly transformed by the attention function in the form of a scaled dot product, the similarity between the nodes is calculated and the attention coefficient is obtained by normalization through the softmax function, and the features of the neighboring nodes are weighted and aggregated by the attention coefficient. The original features of the nodes are then concatenated with the aggregated features and transformed through a feedforward neural network nonlinearly. After multiple rounds of iterative updates, the node embedding vector containing the device association information is obtained, and finally the embedding vector of each pair of nodes is inner-producted to obtain the device correlation strength matrix;

[0116] A power consumption pattern recognition model is constructed based on the device association strength matrix and a pre-set multi-layer autoencoder, wherein the device association strength matrix is ​​input into a pre-set multi-layer autoencoder network, sparsity is introduced by applying an L1 regularization constraint on the hidden layer neurons, and Gaussian noise with a mean of zero and a variance of a set value is superimposed on the input data to improve noise robustness, and a sparse compressed representation of the power consumption pattern is obtained after training for a set number of iterations;

[0117] The abnormal power consumption model is obtained by combining sparse constraints and noise enhancement recognition, wherein normal power consumption data and known abnormal power consumption data are respectively input into the trained autoencoder, and the respective reconstruction errors are calculated. The reconstruction errors are compared and judged by the preset error threshold. When the reconstruction error exceeds the threshold, it is identified as an abnormal power consumption mode, otherwise it is identified as a normal power consumption mode.

[0118] The power consumption decision corresponding to the abnormal power consumption mode is optimized based on the physical constraints and operating status of the distribution system, wherein the physical constraints and real-time operating status parameters of the distribution system are comprehensively considered, the physical constraints include transformer capacity, line current capacity, voltage qualified range, etc., the real-time operating status parameters include equipment load rate, power supply reliability, power quality, etc. On this basis, a multi-objective optimization model is constructed, and the objective function includes but is not limited to minimizing energy consumption cost, maximizing load balance, etc. The optimal load distribution strategy under abnormal working conditions is obtained by solving the optimization model.

[0119] The attention function in the form of scaled dot product is a method to implement the attention mechanism, which generates attention weights by calculating the dot product between the query and the key and adjusting the result according to the scaling factor. The reconstruction error refers to the difference between the original data and the reconstructed data during the data reconstruction process, and is usually used to evaluate the performance of the reconstruction model.

[0120] The equipment topology diagram is constructed based on the physical connection scheme of the power distribution system. Specifically, according to the actual connection relationship between the electrical equipment in the computer room, each device is regarded as a node in the diagram, and the electrical connection relationship between the devices is represented by edges, so as to construct a complete equipment topology diagram. For example, for a small computer room containing a main distribution cabinet, UPS, distribution box and server cabinet, a tree topology structure can be constructed with the main distribution cabinet as the root node, UPS and distribution box as the intermediate nodes, and server cabinet as the leaf node.

[0121] Sensors are installed at each node of the equipment topology diagram to collect the operating parameters of the computer room. For different types of equipment nodes, appropriate sensors are selected for installation. For example, a three-phase energy meter is installed in the main distribution cabinet to collect the total power consumption, a temperature sensor and a load rate sensor are installed in the UPS, and a power meter and a temperature sensor are installed in the server cabinet. These sensors collect the operating parameters of the equipment such as voltage, current, power, temperature, etc. in real time, and transmit the data to the central control system through a wired or wireless network.

[0122] Based on the improved graph attention network, the association features between different electrical devices are extracted and the device association strength matrix is ​​constructed. The collected raw data is preprocessed, including removing outliers and normalizing. The preprocessed data is input into the improved graph attention network as the initial feature vector of the device node. By introducing the multi-head attention mechanism, the attention weight between each node and its neighboring nodes is calculated. Specifically, for each attention head, the node features are mapped to the query vector, key vector and value vector through linear transformation, the dot product of the query vector and the key vector is calculated to obtain the original attention score, and it is scaled by dividing it by the square root of the feature dimension. The attention score is normalized using the softmax function to obtain the attention coefficient, and the value vector is weighted and summed using the attention coefficient to obtain the output of the attention head. The outputs of multiple attention heads are spliced ​​together, and then nonlinearly transformed through a feedforward neural network to obtain the updated node representation. Repeat the above process for multiple rounds to obtain the node embedding vector containing device association information. Finally, the inner product of each pair of node embedding vectors is calculated to construct an n×n device association strength matrix, where n is the number of device nodes.

[0123] A power consumption pattern recognition model is constructed based on the device association strength matrix and a pre-set multi-layer autoencoder. The device association strength matrix obtained in the previous step is used as input and sent to a pre-designed multi-layer autoencoder network. The autoencoder contains multiple encoding layers and decoding layers, with a lower-dimensional bottleneck layer in the middle. During the training process, L1 regularization constraints are applied to the hidden layer neurons to introduce sparsity, and Gaussian noise is superimposed on the input data to improve the robustness of the model. Specifically, the encoder and decoder are set to contain 3 hidden layers each, and the number of neurons is [512, 256, 128] and [128, 256, 512] respectively. ReLU is used as the activation function. The mean square error is used as the reconstruction loss during training. The Adam optimizer is used for parameter update, the batch size is set to 64, and the number of training rounds is 100. After the training is completed, the autoencoder can compress the input power consumption pattern data into a low-dimensional sparse representation;

[0124] Combining sparse constraints and noise enhancement recognition to obtain an abnormal power consumption model, a certain amount of known normal power consumption data and abnormal power consumption data are collected and input into the trained autoencoder model respectively. The reconstruction error of each data is calculated, that is, the Euclidean distance between the input data and the reconstructed data. The distribution of reconstruction errors of normal data and abnormal data is statistically analyzed, and a suitable threshold is set (such as the 99.9th percentile of the reconstruction error of normal data). When new power consumption data is input, if its reconstruction error exceeds the threshold, it is judged as an abnormal power consumption mode, otherwise it is judged as a normal power consumption mode. This method takes advantage of the fact that the autoencoder has a good reconstruction ability for normal modes and a poor reconstruction effect for abnormal modes, and can effectively identify abnormal power consumption behavior.

[0125] Based on the physical constraints and operating status of the distribution system, the power consumption decision corresponding to the abnormal power consumption mode is optimized, and the various physical constraints of the system are clarified, such as the rated capacity of the transformer does not exceed 1000kVA, the maximum allowable current of the distribution line is 400A, and the voltage deviation does not exceed ±7%. The real-time operating status parameters of the system are collected, including the load rate of each device, power supply reliability index, voltage harmonic content, etc., and a multi-objective optimization model is constructed. The objective function includes minimizing the total energy consumption cost, maximizing the load balance, minimizing the power quality degradation, etc. The improved particle swarm algorithm is used to solve the optimization problem. The population size is set to 100 and the maximum number of iterations is 200. The optimal solution obtained is the optimal load distribution strategy under abnormal conditions, including the switch status and operating power of each device.

[0126] In this embodiment, by introducing an improved graph attention network, the topological relationship and mutual influence between devices are fully explored, and more representative correlation features are extracted. By using technologies such as sparse constraints and noise enhancement, the model's resistance to various interferences is enhanced, and abnormal power usage behaviors can be more accurately identified. By solving the model, an operation plan with the lowest energy consumption cost and the most balanced load can be obtained while ensuring power supply safety and quality, effectively responding to abnormal power usage. In summary, this embodiment greatly reduces manual intervention, improves the ability to recognize power usage patterns, and provides strong support for the safe and efficient operation of the computer room.

[0127] In an optional embodiment,

[0128] The attention weight between a device node and its neighbor nodes is calculated using multi-head attention as shown in the following formula:

[0129] ;

[0130] in, represents the feature representation of node i after weighting by the multi-head attention mechanism, represents the set of neighbor nodes of node i, K represents the total number of attention heads, and W k represents the linear transformation matrix of the kth attention head, exp represents the exponential function, represents the transpose of the feature vector of node i, h l represents the feature vector of node l, the neighbor node of node i. Indicates concatenating the results of all attention heads.

[0131] In this embodiment, through the attention mechanism, the model can adaptively learn the correlation and importance between the device node and its neighbor nodes. Through the weighted aggregation of neighbor node features by attention weights, the feature representation of the device node can dynamically integrate the information of the neighbor nodes. The dynamically updated node features can reflect the role and importance of the node in the network topology. The differentiated feature transformation and attention calculation enhance the representation learning ability of heterogeneous networks, so that the model can better characterize the characteristics and interactions of different types of device nodes. The redundant design of the multi-head structure enhances the robustness of the model to noise and missing data, making the model more robust in practical applications. In summary, this embodiment helps to improve the intelligence level and operating efficiency of the data center network, and is of great significance for building an adaptive and highly reliable data center infrastructure.

[0132] Figure 2 FIG. 1 is a schematic diagram of a system for configuring an overall electrical solution for a computer room according to an embodiment of the present invention for a cabinet layout. Figure 2 As shown, the system comprises:

[0133] The first unit is used to collect the building area, floor height, load-bearing parameters, depth dimensions and functional zoning requirements of the computer room, establish a three-dimensional digital model of the computer room, perform functional zoning on the computer room space based on the improved particle swarm optimization algorithm, generate multiple computer room areas, and for the equipment sub-areas, plan the layout of hot and cold channels through the computational fluid dynamics model, build a temperature field distribution model considering airflow disturbances, and optimize the cabinet layout plan in combination with the improved genetic algorithm to obtain the optimization results and determine the load-bearing constraints based on the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, assign a benchmark power density to each cabinet, and generate an initial layout plan;

[0134] The second unit is used to build a cabinet power load prediction model based on an improved long short-term memory neural network, collect historical data on the power load of the computer room and perform feature extraction, perform feature fusion and time series prediction based on the extracted power load features, generate power load prediction results and build a power distribution system, build a state space of the UPS system group and perform dynamic load distribution through a load balancing controller of deep reinforcement learning, repeatedly iterate to obtain the optimal load distribution strategy, build a routing plan for the intelligent bus duct system based on the optimal load distribution strategy and the initial layout plan, determine the access plan of the intelligent circuit breaker in combination with the hierarchical control architecture, and obtain the physical connection plan of the power distribution system;

[0135] The third unit is used to construct a device topology map based on the physical connection scheme of the power distribution system, set sensors at each node of the device topology map and collect computer room operating parameters, extract correlation features between different electrical equipment based on the improved graph attention network and construct a device correlation strength matrix, build a power consumption pattern recognition model based on the device correlation strength matrix and a pre-set multi-layer autoencoder, combine sparse constraints and noise enhancement to identify abnormal power consumption patterns, optimize the power consumption decisions corresponding to the abnormal power consumption patterns based on the physical constraints and operating status of the power distribution system, and generate an optimal control plan in combination with the optimal load distribution strategy.

[0136] According to a third aspect of the embodiments of the present invention,

[0137] An electronic device is provided, comprising:

[0138] processor;

[0139] a memory for storing processor-executable instructions;

[0140] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0141] A fourth aspect of the embodiments of the present invention is:

[0142] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0143] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention 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 replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The overall electrical scheme configuration method for the computer room for cabinet layout is characterized by: include: The building area, floor height, load-bearing parameters, depth dimensions and functional zoning requirements of the computer room are collected, and a three-dimensional digital model of the computer room is established. The computer room space is functionally zoned based on the improved particle swarm optimization algorithm to generate multiple computer room areas. For the equipment sub-areas, the layout of hot and cold channels is planned through the computational fluid dynamics model, and a temperature field distribution model considering airflow disturbance is constructed. The cabinet layout scheme is optimized in combination with the improved genetic algorithm to obtain the optimization results and determine the load-bearing constraints based on the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, and a benchmark power density is assigned to each cabinet to generate an initial layout scheme. Based on the improved long short-term memory neural network, a cabinet power load prediction model is constructed, the historical data of the power load in the computer room is collected and feature extraction is performed, feature fusion and time series prediction are performed based on the extracted power load features, the power load prediction results are generated and the power distribution system is constructed, the state space of the UPS system group is constructed and dynamic load distribution is performed through a load balancing controller of deep reinforcement learning, and the optimal load distribution strategy is obtained through repeated iterations. Based on the optimal load distribution strategy and the initial layout plan, a routing plan for the intelligent bus duct system is constructed, and the access plan of the intelligent circuit breaker is determined in combination with the hierarchical control architecture to obtain the physical connection plan of the power distribution system; Based on the physical connection scheme of the power distribution system, a device topology map is constructed, sensors are set at each node of the device topology map to collect computer room operating parameters, the correlation features between different electrical equipment are extracted based on the improved graph attention network and a device correlation strength matrix is ​​constructed, a power consumption pattern recognition model is constructed based on the device correlation strength matrix and a pre-set multi-layer autoencoder, and the abnormal power consumption pattern is obtained by combining sparse constraints and noise enhancement recognition, the power consumption decision corresponding to the abnormal power consumption pattern is optimized based on the physical constraints and operating status of the power distribution system, and the optimal control plan is generated in combination with the optimal load distribution strategy.

2. The method according to claim 1, characterized in that: The building area, floor height, load-bearing parameters, depth and functional zoning requirements of the computer room are collected to establish a three-dimensional digital model of the computer room. The computer room space is functionally zoned based on the improved particle swarm optimization algorithm to generate multiple computer room areas including: A 3D laser scanner is used to scan the computer room space to obtain point cloud data, wherein the point cloud data corresponds to the building area, floor height, load-bearing parameters, and depth dimensions of the computer room. The scanned point cloud data are fused through point cloud registration technology, and the voxel filtering algorithm is used to reduce noise and sparsely process the point cloud data. The key structural features of the wall and the ground are extracted using a random sampling consistency algorithm, and a 3D digital model of the computer room is established. At the same time, a temperature and humidity sensor network is arranged to collect environmental parameters; According to the preset requirements of the functional zoning of the computer room, the computer room space is functionally zoned based on the multi-objective particle swarm algorithm, and an initial particle swarm is generated through logical mapping. For each particle in the initial particle swarm, a real number encoding method is adopted, which includes the boundary coordinates of the functional area and the area ratio information. For each particle, the fitness is calculated and sorted in descending order according to the fitness. The inertia weight is dynamically adjusted according to the sorting result, and the top 20% of the particles are selected for information exchange to generate a new solution space. The iteration is repeated until the preset maximum number of iterations is reached, and multiple computer room areas are obtained by segmentation according to the boundary coordinates of the functional area generated by the last iteration.

3. The method according to claim 1, characterized in that For the equipment sub-area, the layout of hot and cold channels is planned by computational fluid dynamics model, and the temperature field distribution model considering airflow disturbance is constructed. The cabinet layout scheme is optimized by combining the improved genetic algorithm, and the optimization results are obtained. The load-bearing constraints are determined according to the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, and the benchmark power density is assigned to each cabinet to generate the initial layout scheme, including: For the equipment sub-areas in the computer room area, the computational fluid dynamics model is combined with deep learning to construct an airflow field simulation model, and a convolutional neural network is used to analyze historical operation data to predict load change trends, and the air supply parameters are adaptively adjusted according to the load change trends; guide plates are set at the corners of the cabinet arrangement, intelligent adjustable shutters are installed on the top of the cold channel, and a heat recovery device is set on the top of the hot channel. The cabinet layout is optimized using a multi-objective genetic algorithm, and a simulated annealing mechanism is introduced to set the initial temperature and cooling coefficient to obtain the optimization results; Based on the optimization results, the finite element analysis method is used to divide the floor into grid units, the grid size is adaptively encrypted in the high-load area, a nonlinear mechanical model considering the deformation of the floor is established, and an integer programming model for optimizing the installed capacity of the computer room is established. The optimization goal is to maximize the total power of all information technology equipment in the computer room. The static load of the floor does not exceed the static load bearing capacity and the installed density of each functional partition does not exceed the allowable density. The functional relationship between the installed capacity and the number of information technology equipment arranged is established, and the dynamic programming algorithm is used to solve the integer programming model for optimizing the installed capacity. The dynamic programming state is defined to represent the maximum total power when multiple equipment is considered and the total weight does not exceed the given weight, and the transfer equation between adjacent states is established. The initial layout plan that meets the floor bearing capacity constraint is solved.

4. The method according to claim 1, characterized in that: Based on the improved long short-term memory neural network, a cabinet power load prediction model is constructed, the historical data of the power load in the computer room is collected and feature extraction is performed, feature fusion and time series prediction are performed based on the extracted power load features, the power load prediction results are generated and the power distribution system is constructed, the state space of the UPS system group is constructed and dynamic load distribution is performed through the load balancing controller of deep reinforcement learning, and the optimal load distribution strategy is obtained by repeated iterations. Based on the optimal load distribution strategy and the initial layout plan, a routing plan of the intelligent bus duct system is constructed, and the access plan of the intelligent circuit breaker is determined in combination with the hierarchical control architecture, and the physical connection plan of the distribution system is obtained, including: The power monitoring sensor deployed on the cabinet power distribution unit collects the cabinet power load history data. The power monitoring sensor records the cabinet real-time power consumption data every five minutes. The cabinet real-time power consumption data includes voltage, current, active power, and reactive power. The cabinet power load history data is processed for abnormal values ​​and filled with missing values. Feature extraction is performed on the historical data of the power load of the cabinet from three dimensions: time domain, frequency domain, and time-frequency domain. In the time domain dimension, the time domain feature is obtained by statistically analyzing the mean, variance, peak value, and waveform factor of the power load. In the frequency domain dimension, the amplitude and phase of different frequency components are obtained by performing fast Fourier transform on the historical data of the power load of the cabinet, and the spectrum characteristics of the power load in the change result are extracted to obtain the frequency domain feature. Wavelet transform is performed on the historical data of the power load of the cabinet to extract the energy distribution characteristics of the power load on different time scales to obtain the time-frequency domain feature. An improved long short-term memory neural network model is used to perform feature fusion on the time domain features, frequency domain features, and time-frequency domain features. The improved long short-term memory neural network model comprises an input layer, a long short-term memory layer, a fully connected layer, and an output layer. The long short-term memory layer comprises 128 neurons and adopts a dropout mechanism. The fully connected layer uses a ReLU activation function. The output layer generates a 24-hour electricity load prediction result. Based on the power load prediction result, a state space of an uninterruptible power supply system group is constructed, wherein the state space includes the load rate and battery capacity of the uninterruptible power supply module, and a load balancing controller is trained using a deep Q network algorithm, wherein the input of the load balancing controller is the state space, and the output of the load balancing controller is a load distribution action; The load balancing controller is iteratively trained to obtain an optimal load distribution strategy, a wiring plan for a three-phase five-wire intelligent bus duct system is constructed based on the optimal load distribution strategy, an intelligent circuit breaker with a hierarchical control architecture is arranged in the three-phase five-wire intelligent bus duct system, and a physical connection plan for a power distribution system is constructed based on the intelligent circuit breaker with the hierarchical control architecture. The physical connection plan for the power distribution system includes a primary main incoming line cabinet, an uninterruptible power supply input and output cabinet, a three-phase five-wire intelligent bus duct system, an inter-row distribution cabinet, and a cabinet power distribution unit. The primary main incoming line cabinet, the uninterruptible power supply input and output cabinet, the three-phase five-wire intelligent bus duct system, the inter-row distribution cabinet, and the cabinet power distribution unit are connected by copper busbars and cables.

5. The method according to claim 4, characterized in that The loss function corresponding to the deep Q network algorithm is shown in the following formula: Among them, L(θ) represents the loss value, E represents the expectation, and H δ represents the Huber loss function, r represents the immediate reward, γ represents the discount factor, It represents the maximum Q value of all actions a' calculated by the target network with parameters θ' in the next state s', which is used to indicate the expected future reward. Q(s, a; θ)) represents the Q value calculated by the target network with parameters θ when taking action a in state s, which is used to indicate the value of the current state and action.

6. The method according to claim 1, characterized in that Based on the physical connection scheme of the power distribution system, a device topology map is constructed, sensors are set at each node of the device topology map to collect the operating parameters of the computer room, the correlation features between different electrical devices are extracted based on the improved graph attention network and the device correlation strength matrix is ​​constructed, the power consumption pattern recognition model is constructed based on the device correlation strength matrix and the pre-set multi-layer autoencoder, the abnormal power consumption model is obtained by combining sparse constraints and noise enhancement recognition, the power consumption decision corresponding to the abnormal power consumption pattern is optimized based on the physical constraints and operating status of the power distribution system, and the optimal control plan is generated in combination with the optimal load distribution strategy, including: Building a device topology map based on the physical connection scheme of the power distribution system, setting sensors at each node of the device topology map and collecting the operating parameters of the computer room; Based on the improved graph attention network, the correlation features between different electrical devices are extracted and the device correlation strength matrix is ​​constructed, wherein the improved graph attention network introduces the attention mechanism and uses multi-head attention to calculate the attention weights between the device node and its neighboring nodes. Specifically, the node features are linearly transformed by the attention function in the form of a scaled dot product, the similarity between the nodes is calculated and the attention coefficient is obtained by normalization through the softmax function, and the features of the neighboring nodes are weighted and aggregated by the attention coefficient. The original features of the nodes are then concatenated with the aggregated features and transformed through a feedforward neural network nonlinearly. After multiple rounds of iterative updates, the node embedding vector containing the device association information is obtained, and finally the embedding vector of each pair of nodes is inner-producted to obtain the device correlation strength matrix; A power consumption pattern recognition model is constructed based on the device association strength matrix and a pre-set multi-layer autoencoder, wherein the device association strength matrix is ​​input into a pre-set multi-layer autoencoder network, sparsity is introduced by applying an L1 regularization constraint on the hidden layer neurons, and Gaussian noise with a mean of zero and a variance of a set value is superimposed on the input data to improve noise robustness, and a sparse compressed representation of the power consumption pattern is obtained after training for a set number of iterations; The abnormal power consumption model is obtained by combining sparse constraints and noise enhancement recognition, wherein normal power consumption data and known abnormal power consumption data are respectively input into the trained autoencoder, and the respective reconstruction errors are calculated. The reconstruction errors are compared and judged by the preset error threshold. When the reconstruction error exceeds the threshold, it is identified as an abnormal power consumption mode, otherwise it is identified as a normal power consumption mode. Based on the physical constraints and operating status of the distribution system, the power consumption decision corresponding to the abnormal power consumption mode is optimized, wherein the physical constraints and real-time operating status parameters of the distribution system are comprehensively considered, the physical constraints include transformer capacity, line current capacity, and voltage qualified range, and the real-time operating status parameters include equipment load rate, power supply reliability, and power quality. On this basis, a multi-objective optimization model is constructed, and the objective function includes but is not limited to minimizing energy consumption cost and maximizing load balance. The optimal load distribution strategy under abnormal working conditions is obtained by solving the optimization model.

7. The method according to claim 6, characterized in that The attention weight between a device node and its neighbor nodes is calculated using multi-head attention as shown in the following formula: Among them, h′ i represents the feature representation of node i after weighting by the multi-head attention mechanism, represents the set of neighbor nodes of node i, K represents the total number of attention heads, and W k represents the linear transformation matrix of the kth attention head, exp represents the exponential function, represents the transpose of the feature vector of node i, h l represents the feature vector of node l, the neighbor node of node i. Indicates concatenating the results of all attention heads.

8. A system for configuring the overall electrical scheme of a computer room for cabinet layout, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the building area, floor height, load-bearing parameters, depth dimensions and functional zoning requirements of the computer room, establish a three-dimensional digital model of the computer room, perform functional zoning on the computer room space based on the improved particle swarm optimization algorithm, generate multiple computer room areas, and for the equipment sub-areas, plan the layout of hot and cold channels through the computational fluid dynamics model, build a temperature field distribution model considering airflow disturbances, and optimize the cabinet layout plan in combination with the improved genetic algorithm to obtain the optimization results and determine the load-bearing constraints based on the finite element analysis of the computer room structure. The dynamic programming algorithm is used to calculate the maximum cabinet capacity of the equipment sub-area, assign a benchmark power density to each cabinet, and generate an initial layout plan; The second unit is used to build a cabinet power load prediction model based on an improved long short-term memory neural network, collect historical data on the power load of the computer room and perform feature extraction, perform feature fusion and time series prediction based on the extracted power load features, generate power load prediction results and build a power distribution system, build a state space of the UPS system group and perform dynamic load distribution through a load balancing controller of deep reinforcement learning, repeatedly iterate to obtain the optimal load distribution strategy, build a routing plan for the intelligent bus duct system based on the optimal load distribution strategy and the initial layout plan, determine the access plan of the intelligent circuit breaker in combination with the hierarchical control architecture, and obtain the physical connection plan of the power distribution system; The third unit is used to construct a device topology map based on the physical connection scheme of the power distribution system, set sensors at each node of the device topology map and collect computer room operating parameters, extract correlation features between different electrical equipment based on the improved graph attention network and construct a device correlation strength matrix, build a power consumption pattern recognition model based on the device correlation strength matrix and a pre-set multi-layer autoencoder, combine sparse constraints and noise enhancement to identify abnormal power consumption patterns, optimize the power consumption decisions corresponding to the abnormal power consumption patterns based on the physical constraints and operating status of the power distribution system, and generate an optimal control plan in combination with the optimal load distribution strategy.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images