Adaptive operation strategy optimization method and system for refrigeration system

Abstract

The application provides a refrigeration system adaptive operation strategy optimization method and system, relates to the technical field of automatic control, and comprises the following steps: obtaining temperature and humidity data and fluid boundary parameters, reconstructing a temperature and humidity distribution field by using a physical constraint neural network, extracting enthalpy difference characteristics, constructing a topological coupling network, calculating a corrected refrigeration load, and optimizing refrigeration resource allocation parameters based on energy consumption and enthalpy value uniformity, so that precise regulation and control of a dynamic load area is realized, the energy efficiency of the system is improved, the enthalpy value uniformity requirement of each area is met, and the ability of the system to adapt to environmental changes is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic control, in particular to a self-adaptive operation strategy optimization method and system for a refrigeration system. BACKGROUND

[0002] A refrigeration system is an air conditioning system widely used in data centers, clean workshops, precision machining workshops and other places. This kind of system maintains the enthalpy value of the environment in an appropriate range by regulating the temperature and humidity in the space, ensuring the normal operation of the equipment and the stability of the process. At present, the operation strategy of this kind of refrigeration system is mainly based on the traditional temperature and humidity control method, which adjusts by setting fixed temperature and humidity thresholds. However, there are the following defects and deficiencies:

[0003] This kind of refrigeration system operation strategy lacks accurate perception ability of the enthalpy value distribution in the space, and only relies on limited temperature and humidity sampling points for control, which cannot accurately reflect the dynamic change characteristics of the enthalpy value of the whole space, resulting in unreasonable allocation of refrigeration resources and serious energy waste.

[0004] The prior art does not fully consider the thermal-hygroscopic coupling relationship between different areas in the refrigeration space, the mutual influence between areas is ignored, and the overall collaborative optimization control cannot be realized, resulting in the possibility of overcooling or overheating in local areas, affecting the stability and uniformity of the system.

[0005] The control strategy lacks self-adaptive adjustment mechanism and cannot flexibly adjust the control parameters according to the dynamic load change. The system response is lagging, and it is difficult to respond to sudden load changes, especially in the scene of high-density equipment distribution, the balance between refrigeration effect and energy consumption is difficult to optimize. SUMMARY

[0006] The embodiments of the present application provide a self-adaptive operation strategy optimization method and system for a refrigeration system, which can solve the problems in the prior art.

[0007] In a first aspect of the embodiments of the present application, a self-adaptive operation strategy optimization method for a refrigeration system is provided, comprising:

[0008] Obtaining temperature and humidity data and fluid boundary parameters in the space acted on by the refrigeration system;

[0009] Reconstructing a continuous temperature and humidity distribution field based on a physical constraint neural network, calculating an enthalpy distribution field and extracting a gradient change trajectory, dividing dynamic load areas according to gradient direction turning points and gradient periodic fluctuation points, and extracting load fluctuation intensity of each area as a local enthalpy difference feature;

[0010] The mutual correlation coefficient and phase difference of the time series of the load fluctuation intensity of adjacent areas are calculated to construct a topological coupling network; the enthalpy value decay rate of air supply arriving at each area is calculated according to the fluid boundary parameter; for each area, the local enthalpy difference feature, the coupling weight of the topological coupling network and the enthalpy value decay rate are calculated to obtain a modified refrigeration load through energy balance calculation;

[0011] The system energy consumption and the enthalpy value uniformity of each area are calculated based on the modified refrigeration load as an optimization target, the target weight coefficient is dynamically adjusted according to the load fluctuation intensity, and the refrigeration resource allocation parameter is solved;

[0012] Actual temperature and humidity data after the refrigeration resource allocation parameter is executed are collected, and the dynamic load area boundary is adjusted according to the deviation type of the actual enthalpy value distribution and the expected enthalpy value distribution.

[0013] The continuous temperature and humidity distribution field is obtained by field reconstruction of the temperature and humidity data based on the physical constraint neural network, and the steps of calculating the enthalpy value distribution field and extracting the gradient change trajectory include:

[0014] The physical constraint neural network includes a temperature prediction branch and a humidity prediction branch, the temperature prediction branch receives spatial position and time information as input, and the humidity prediction branch receives spatial position, time information and hidden layer features of the temperature prediction branch as input;

[0015] The loss function of the physical constraint neural network includes a data fitting term and a physical constraint term, the physical constraint term includes a thermodynamic conservation constraint and a time causality constraint; the time causality constraint calculates the enthalpy value according to the output of the temperature prediction branch and the humidity prediction branch, and constrains the time derivative sign of the enthalpy value to be consistent with the direction of energy inflow and outflow at the position; the weight proportion of the data fitting term and the physical constraint term is dynamically adjusted based on the density of temperature and humidity measurement points;

[0016] The continuous temperature and humidity distribution field is obtained by field reconstruction of the temperature and humidity data based on the physical constraint neural network, the enthalpy value distribution field is calculated, and the time gradient and spatial gradient of the enthalpy value distribution field are calculated, and the evolution process of the time gradient and spatial gradient in the time-space domain is taken as the gradient change trajectory.

[0017] The dynamic load area is divided according to the gradient direction turning point and the gradient periodic fluctuation point, and the load fluctuation intensity of each area is extracted as the local enthalpy difference feature, and the steps include:

[0018] The direction of the enthalpy value spatial gradient in the gradient change trajectory is tracked, and the spatial position where the gradient direction vector angle change rate exceeds a first set threshold is identified as the gradient direction turning point;

[0019] performing time domain analysis on the enthalpy value time gradient in the gradient change trajectory, and identifying a spatial position where a gradient amplitude change rate at adjacent time points exceeds a second set threshold and a change direction is reversed as a gradient periodic fluctuation point;

[0020] performing clustering analysis on the gradient direction turning points and the gradient periodic fluctuation points in spatial distribution to obtain a plurality of spatial clustering clusters, and taking a boundary of each spatial clustering cluster as a boundary of a dynamic load area;

[0021] calculating fluctuation characteristics of the enthalpy value time gradient in each dynamic load area as a load fluctuation intensity, and combining regional position information to form a local enthalpy difference feature.

[0022] The steps of calculating a cross-correlation coefficient and a phase difference of a load fluctuation intensity time sequence of adjacent areas to construct a topological coupling network include:

[0023] forming a load fluctuation intensity time sequence of each dynamic load area based on the load fluctuation intensity of each dynamic load area in the local enthalpy difference feature;

[0024] determining a spatial adjacent relationship between dynamic load areas, for two spatially adjacent dynamic load areas, calculating a cross-correlation function of their load fluctuation intensity time sequences, extracting a maximum value of the cross-correlation function as a cross-correlation coefficient, and extracting a time offset corresponding to the maximum value as a phase difference;

[0025] taking each dynamic load area as a network node, taking a cross-correlation coefficient between spatially adjacent dynamic load areas as a weight of a connection edge, and taking the phase difference as a direction attribute of the connection edge to construct a topological coupling network.

[0026] The steps of performing energy balance calculation on the local enthalpy difference feature, a coupling weight of the topological coupling network, and an enthalpy value decay rate to obtain a corrected refrigeration load include:

[0027] For each dynamic load area, extracting the load fluctuation intensity and regional boundary information of the area from the local enthalpy difference feature, and extracting the coupling weight and phase difference between the area and adjacent areas from the topological coupling network;

[0028] calculating a basic refrigeration load based on the load fluctuation intensity and regional boundary information of the area;

[0029] According to the coupling weight and phase difference between the dynamic load area and each adjacent area, calculating a heat and moisture transfer influence amount of the adjacent area on the area, which is equal to a product of the load fluctuation intensity and the coupling weight of the adjacent area and a time offset corrected cumulative value according to the phase difference;

[0030] According to the enthalpy attenuation rate of the dynamic load area, the effective refrigeration capacity of the supply air reaching the area is calculated, which is equal to the energy value corresponding to the difference between the nominal refrigeration capacity of the supply air and the enthalpy attenuation rate;

[0031] The base refrigeration load of the dynamic load area is added to the heat and moisture transfer impact quantity to obtain the total heat and moisture load of the area, and energy balance calculation is performed on the total heat and moisture load of the area and the effective refrigeration capacity to obtain the corrected refrigeration load of the dynamic load area.

[0032] Based on the corrected refrigeration load, system energy consumption and enthalpy uniformity of each area are calculated as optimization targets, target weight coefficients are dynamically adjusted according to load fluctuation intensity, and the steps of solving refrigeration resource allocation parameters include:

[0033] A state vector containing load fluctuation intensity and area coupling characteristics is constructed, and a target weight coefficient of each dynamic load area is generated based on a reinforcement learning strategy network according to the state vector, and the generation of the target weight coefficient is based on a collaborative adjustment strategy adopted by the topological coupling network;

[0034] Based on the coupling weight and phase difference in the topological coupling network, an influence matrix representing the degree of influence and time delay of heat and moisture transfer between areas is constructed;

[0035] Taking the supply air parameters of each dynamic load area as decision variables, taking the sum of the system energy consumption and the enthalpy uniformity index of each area weighted by the target weight coefficient as the optimization target, and taking the heat and moisture transfer constraint between areas defined by the influence matrix and the physical capacity constraint of the refrigeration equipment as the constraint condition, a multi-objective optimization function is constructed, and the refrigeration resource allocation parameters of each dynamic load area are obtained by solving the multi-objective optimization function;

[0036] The refrigeration resource allocation parameters are executed, the actual running effect is monitored, the parameters of the reinforcement learning strategy network are updated according to the monitoring results, and online learning and continuous optimization of the target weight coefficient adjustment strategy are realized.

[0037] The steps of generating the target weight coefficient based on the collaborative adjustment strategy adopted by the topological coupling network include:

[0038] Based on the distance relationship and connection mode between areas in the topological coupling network, a collaborative adjustment matrix is constructed combined with the enthalpy gradient change trajectory of each area, and the elements in the collaborative adjustment matrix represent the influence relationship of the adjustment of the target weight coefficient between areas;

[0039] According to the element values of the collaborative adjustment matrix, the areas are divided into groups, the areas with an influence relationship exceeding a preset influence threshold are determined as a collaborative area group, the master-slave relationship of the areas in the group is determined based on the size of the influence relationship, and an initial adjustment scheme of the target weight coefficient of the collaborative areas is generated;

[0040] According to the cooperative adjustment matrix, a change trend of the enthalpy gradient in the cooperative region group is calculated, when the change trend of the enthalpy gradient exceeds a preset range, a target weight coefficient of a corresponding region in the initial adjustment scheme is adjusted until the change trend of the enthalpy gradient meets a preset range requirement;

[0041] An adjustment scheme meeting the preset range requirement is taken as a final target weight coefficient, and is used for solving the multi-objective optimization function.

[0042] A second aspect of the embodiment of the present application provides a refrigeration system adaptive operation strategy optimization system, comprising:

[0043] A first unit is configured to acquire temperature and humidity data and fluid boundary parameters in a space acted on by a refrigeration system;

[0044] A second unit is configured to perform field reconstruction on the temperature and humidity data based on a physical constraint neural network to obtain a continuous temperature and humidity distribution field, calculate an enthalpy distribution field and extract a gradient change trajectory, divide dynamic load regions according to gradient direction turning points and gradient periodic fluctuation points, and extract load fluctuation intensities of the regions as local enthalpy difference features;

[0045] A third unit is configured to calculate mutual correlation coefficients and phase differences of time series of the load fluctuation intensities of adjacent regions to construct a topological coupling network, calculate enthalpy attenuation rates of air supply when reaching the regions according to the fluid boundary parameters, and perform energy balance calculation on the local enthalpy difference features, coupling weights of the topological coupling network, and the enthalpy attenuation rates to obtain a corrected refrigeration load for each region;

[0046] A fourth unit is configured to calculate system energy consumption and enthalpy uniformity of the regions as optimization objectives based on the corrected refrigeration load, dynamically adjust target weight coefficients according to the load fluctuation intensities, and solve to obtain refrigeration resource allocation parameters;

[0047] A fifth unit is configured to collect actual temperature and humidity data after the refrigeration resource allocation parameters are executed, and adjust boundaries of the dynamic load regions according to a deviation type of actual enthalpy distribution and expected enthalpy distribution.

[0048] A third aspect of the embodiment of the present application,

[0049] An electronic device is provided, comprising:

[0050] a processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0053] A fourth aspect of the embodiment of the present application,

[0054] A computer readable storage medium is provided, and computer program instructions are stored on the computer readable storage medium, and the computer program instructions are executed by a processor to implement the method.

[0055] The present application realizes accurate reconstruction of the temperature and humidity field by physically constraining the neural network, scientifically divides the dynamic load area based on the gradient change trajectory, accurately captures the complex heat flow dynamic characteristics, and greatly improves the operation state perception accuracy.

[0056] The topology coupling network constructed by the present application can represent the transmission relationship of load fluctuations between regions, and combined with enthalpy decay rate for energy balance calculation, a modified refrigeration load more consistent with the actual physical process is obtained, avoiding the load estimation deviation caused by ignoring the heat exchange between regions in the traditional method, and making the refrigeration resource distribution more reasonable.

[0057] The present application dynamically adjusts the optimization target weight according to the load fluctuation intensity, realizes adaptive balance between system energy consumption and enthalpy uniformity, and adjusts the dynamic load area boundary through the deviation feedback of actual enthalpy distribution and expected enthalpy distribution, forming a closed-loop adaptive control mechanism, improving the adaptability of the system to environmental changes, and significantly reducing the system energy consumption while ensuring the comfort of each region. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 A flowchart of the refrigeration system adaptive operation strategy optimization method of the embodiment of the present application is shown in

[0059] Figure 2 A flowchart of the energy balance modified refrigeration load calculation based on local features and topology coupling is shown in DETAILED DESCRIPTION

[0060] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0062] Figure 1 A flowchart of the refrigeration system adaptive operation strategy optimization method of the embodiment of the present application is shown in Figure 1 As shown in the figure, the method comprises:

[0063] acquire temperature and humidity data and fluid boundary parameters in a space of a refrigeration system;

[0064] reconstruct a continuous temperature and humidity distribution field based on a physically constrained neural network, calculate an enthalpy distribution field and extract a gradient change trajectory, divide dynamic load areas according to gradient direction turning points and gradient periodic fluctuation points, and extract load fluctuation intensity of each area as a local enthalpy difference feature;

[0065] calculate a cross-correlation coefficient and a phase difference of a time series of load fluctuation intensity of adjacent areas to construct a topological coupling network, calculate an enthalpy decay rate of air supply when reaching each area according to the fluid boundary parameters, and perform energy balance calculation on the local enthalpy difference feature, coupling weight of the topological coupling network, and the enthalpy decay rate to obtain a corrected refrigeration load for each area.

[0066] calculate system energy consumption and enthalpy uniformity of each area based on the corrected refrigeration load as an optimization target, dynamically adjust a target weight coefficient according to the load fluctuation intensity, and solve to obtain refrigeration resource allocation parameters;

[0067] collect actual temperature and humidity data after the refrigeration resource allocation parameters are executed, and adjust the boundary of the dynamic load area according to the deviation type of the actual enthalpy distribution and the expected enthalpy distribution.

[0068] In an optional implementation, the step of reconstructing a continuous temperature and humidity distribution field based on a physically constrained neural network, calculating an enthalpy distribution field, and extracting a gradient change trajectory includes:

[0069] The physically constrained neural network includes a temperature prediction branch and a humidity prediction branch, the temperature prediction branch receives spatial position and time information as input, and the humidity prediction branch receives spatial position, time information, and hidden layer features of the temperature prediction branch as input.

[0070] The loss function of the physically constrained neural network includes a data fitting term and a physical constraint term, the physical constraint term includes a thermodynamic conservation constraint and a time causality constraint, the time causality constraint calculates an enthalpy value according to the output of the temperature prediction branch and the humidity prediction branch, and constrains the time derivative sign of the enthalpy value to be consistent with the direction of energy inflow and outflow at the position, and dynamically adjusts the weight proportion of the data fitting term and the physical constraint term based on the density of temperature and humidity measurement points.

[0071] reconstruct a continuous temperature and humidity distribution field based on the physically constrained neural network, calculate an enthalpy distribution field, and calculate a time gradient and a spatial gradient of the enthalpy distribution field, and take the evolution process of the time gradient and the spatial gradient in the time-space domain as a gradient change trajectory.

[0072] For example, the physically constrained neural network employs a dual-branch architecture to jointly reconstruct the temperature and humidity fields. The temperature prediction branch receives three-dimensional spatial coordinates and a timestamp as input. The spatial coordinates are normalized to the 0-1 interval in meters, and the timestamp is normalized to the 0-1 interval in seconds for the current control cycle. The temperature prediction branch contains four fully connected layers, with the number of neurons in each layer being 128, 256, 256, and 128 respectively. The activation function is a hyperbolic tangent function, and the last layer outputs a single scalar representing the predicted temperature at that spatiotemporal location. The humidity prediction branch, while receiving the spatial coordinates and timestamp, also receives the 256-dimensional hidden layer feature vector from the third fully connected layer of the temperature prediction branch. This feature vector is concatenated with the input of the humidity prediction branch and then fed into the fully connected layer of the humidity prediction branch. The humidity prediction branch also consists of four fully connected layers with 384, 512, 256, and 128 neurons. The first layer has an input dimension of 387, including 3D spatial coordinates, 1D time, 256-dimensional temperature hidden features, and 127-dimensional self-hidden features. The activation function is consistent with that of the temperature branch. The last layer outputs a single scalar representing the predicted humidity. The interlayer connection weights between the temperature and humidity branches are initialized using the Xavier method, with the bias term initialized to 0.01.

[0073] The loss function of the physically constrained neural network consists of a data fitting term and a physical constraint term. The data fitting term calculates the mean square error between the neural network's predicted values ​​and the actual measured data. For temperature measurement points, it calculates the square of the difference between the predicted and measured temperatures; for humidity measurement points, it calculates the square of the difference between the predicted and measured humidity. The sum of the squared errors for all measurement points is divided by the total number of measurement points to obtain the value of the data fitting term. The physical constraint term includes two parts: thermodynamic conservation constraints and time causality constraints. The thermodynamic conservation constraints require that the temperature and humidity fields satisfy the energy conservation equation. Specifically, 1000 verification points are uniformly sampled within the field. For each verification point, the partial derivatives of temperature with respect to time, the second-order partial derivatives of temperature with respect to the three spatial directions, the partial derivatives of humidity with respect to time, and the second-order partial derivatives of humidity with respect to the three spatial directions are calculated. An automatic differentiation mechanism is used to obtain the derivatives of each order from the neural network output. The time derivative of temperature and the second-order spatial derivative of temperature are linearly combined using a thermal diffusivity of 0.0023 m² / s, and the squared residual is calculated. The time derivative and spatial second derivative of humidity are linearly combined using a moisture diffusion coefficient of 0.0019 m² / s, and the squared residuals are calculated. The sum of the squared residuals of temperature and humidity conservation at all verification points is divided by the number of verification points to obtain the value of the thermodynamic conservation constraint term.

[0074] The time causality constraint requires that the direction of enthalpy change over time be consistent with the direction of energy inflow and outflow at that location. For each verification point, the enthalpy is calculated based on the outputs of the temperature and humidity prediction branches. The enthalpy calculation involves adding two terms: the first is the temperature multiplied by the specific heat of air at constant pressure (1.005 kJ / (kg·℃), and the second is the humidity multiplied by the latent heat of vaporization of water vapor (2501 kJ / kg). The sum of these two terms is the enthalpy at that point. The first partial derivative of the enthalpy with respect to time is used to obtain the time derivative of the enthalpy. The first partial derivatives of the temperature with respect to the three spatial directions are used to obtain the temperature gradient vector, and the first partial derivatives of the humidity with respect to the three spatial directions are used to obtain the humidity gradient vector. The energy flux density vector is then calculated based on the temperature and humidity gradient vectors. Each component of the energy flux density vector is calculated in two parts: the first part is the product of the negative thermal diffusivity and the corresponding component of the temperature gradient; the second part is the product of the negative moisture diffusivity and the corresponding component of the humidity gradient, multiplied by the latent heat of vaporization. A given component of the energy flux density vector is equal to the first part minus the second. The divergence of the energy flux density vector is calculated; the divergence value is the sum of the partial derivatives of the three components of the energy flux density vector with respect to their respective spatial directions. The time causality constraint requires that the product of the enthalpy time derivative and the energy flux density divergence be greater than or equal to zero. When the product is less than zero, a penalty term is generated, the value of which is equal to the square of the absolute value of the product. The sum of the time causality penalty terms for all verification points, divided by the number of verification points, yields the value of the time causality constraint term.

[0075] The measurement point density is defined as the number of measurement points divided by the field volume, where the field volume is in cubic meters. When the measurement point density is less than 0.1 points / m³, the weight of the data fitting term is set to 0.3, the weight of the thermodynamic conservation constraint in the physical constraint term is set to 0.5, and the weight of the time causality constraint is set to 0.2. When the measurement point density is between 0.1 and 0.5 points / m³, the weight of the data fitting term increases linearly to 0.6, the weight of the thermodynamic conservation constraint in the physical constraint term decreases linearly to 0.3, and the weight of the time causality constraint remains unchanged at 0.1. When the measurement point density exceeds 0.5 points / m³, the weight of the data fitting term is fixed at 0.8, the weight of the thermodynamic conservation constraint in the physical constraint term is fixed at 0.15, and the weight of the time causality constraint is fixed at 0.05. The calculation of the total loss function is divided into three parts, which are added together: the first part is the data fitting term multiplied by its weight, the second part is the thermodynamic conservation constraint multiplied by its weight, and the third part is the time causality constraint multiplied by its weight.

[0076] The neural network was trained using the Adam optimizer with an initial learning rate of 0.001, decaying by 10% every 500 iterations. The batch size was set to 256, containing randomly sampled test and validation data. Training lasted 5000 iterations, with the loss function evaluated on the validation set every 100 iterations. Training was terminated when the validation loss failed to decrease after three consecutive evaluations. The trained neural network model retained its weight parameters and network structure configuration for subsequent field reconstruction inference.

[0077] In the field reconstruction inference stage, a three-dimensional spatial grid is established within the operating space of the refrigeration system. This grid is divided into 50 nodes along the length, 30 nodes along the width, and 20 nodes along the height, totaling 30,000 spatial grid points. For a given moment, the coordinates of these 30,000 spatial grid points and their timestamps are input into the trained physical constraint neural network. The temperature prediction branch and humidity prediction branch output 30,000 temperature prediction values ​​and 30,000 humidity prediction values, respectively, forming a continuous temperature and humidity distribution field. Based on the temperature and humidity distribution fields, the enthalpy distribution field is calculated. The enthalpy calculation for each grid point involves the sum of two terms: the first is the temperature multiplied by the specific heat of air at constant pressure, and the second is the humidity multiplied by the latent heat of vaporization of water vapor. The enthalpy distribution field is stored as a three-dimensional array with dimensions of 50×30×20.

[0078] The temporal and spatial gradients of the enthalpy distribution field are calculated. The temporal gradient is calculated using the central difference method. For the current moment, the enthalpy distribution fields of the previous and next moments are obtained. The enthalpy temporal gradient of each grid point at the current moment is calculated by subtracting the enthalpy of that point from the enthalpy of the previous moment from the enthalpy of that point at the next moment, and then dividing the difference by the time interval. The time interval is 1 / 10 of the control period, typically 6 seconds. The spatial gradient is also calculated using the central difference method. For each grid point, the enthalpy spatial gradient along the length direction is calculated by subtracting the enthalpy of the adjacent point to the right from the enthalpy of the adjacent point to the left, and then dividing the difference by the grid spacing along the length direction. The grid spacing is obtained by dividing the length of the refrigeration system's operating space by 49. The enthalpy spatial gradients along the width and height directions are calculated similarly, by dividing by the corresponding grid spacing. Each grid point has one temporal gradient scalar and three spatial gradient components, forming a four-dimensional gradient vector.

[0079] The gradient change trajectory records the evolution of the temporal and spatial gradients of enthalpy in the spatiotemporal domain. For a control period, it includes snapshots of the enthalpy distribution field at 10 time points, with 30,000 grid points at each time point and a four-dimensional gradient vector at each grid point, totaling 300,000 gradient vectors constituting the gradient change trajectory for that control period. The gradient change trajectory is stored in a time-series data structure, supporting multi-dimensional queries by time, spatial location, and gradient component indexing. For scenarios requiring extraction of gradient evolution in a specific spatial region, a spatial coordinate range can be specified to extract the gradient vector sequence for all time points within that range. For scenarios requiring analysis of gradient abrupt changes, a gradient change rate threshold can be set. When the difference in gradient vectors at the same grid point between adjacent time points exceeds the threshold, it is marked as a mutation point, and the spatiotemporal location and gradient value of the mutation point are recorded in the mutation event list.

[0080] This invention achieves high-precision enthalpy distribution field construction. Through the collaborative design of temperature and humidity prediction branches, and by integrating physical principles with data fitting, it ensures that the field reconstruction results simultaneously satisfy data consistency and physical rationality. In particular, the introduction of time causality constraints guarantees that enthalpy changes conform to energy flow laws, overcoming the limitations of traditional interpolation methods in terms of physical consistency. By dynamically adjusting weights based on measurement point density, this method maintains good reconstruction accuracy even in sparse data regions.

[0081] In one optional implementation, the step of dividing the dynamic load region according to the gradient direction inflection point and the gradient periodic fluctuation point, and extracting the load fluctuation intensity of each region as a local enthalpy difference characteristic includes:

[0082] The enthalpy spatial gradient in the gradient change trajectory is oriented, and the spatial position where the rate of change of the gradient direction vector angle exceeds a first set threshold is identified as the gradient direction inflection point.

[0083] A time-domain analysis is performed on the enthalpy time gradient in the gradient change trajectory to identify spatial locations where the rate of change of the gradient amplitude at adjacent time points exceeds a second set threshold and the direction of change reverses as gradient periodic fluctuation points.

[0084] Cluster analysis is performed on the spatial distribution of the gradient direction inflection points and the gradient periodic fluctuation points to obtain multiple spatial clusters, and the boundary of each spatial cluster is used as the boundary of the dynamic load region.

[0085] The fluctuation characteristics of the enthalpy time gradient in each dynamic load region are calculated as the load fluctuation intensity, and combined with the regional location information to form the local enthalpy difference characteristics.

[0086] For example, the gradient change trajectory includes the evolution data of the temporal and spatial gradients of enthalpy throughout the control period. Direction tracking is performed on the spatial gradient to identify gradient direction inflection points. The enthalpy spatial gradient has three components at each grid point, corresponding to the gradient values ​​in the length, width, and height spatial directions, respectively. The gradient direction vector is calculated by constructing a three-dimensional vector from the three components, and the direction angle of the vector is determined by the ratio of the three components. For the spatial gradient vectors of a grid point at two adjacent time points, the angle between the two vectors is calculated. The angle calculation uses the ratio of the vector dot product to the vector magnitude; specifically, it is the sum of the products of the corresponding components of the two gradient vectors divided by the product of the magnitudes of the two vectors, yielding the cosine value of the angle. The inverse cosine function is then used to obtain the angle in radians. The rate of change of the gradient direction vector angle is defined as the angle between the gradient vectors at adjacent time points divided by the time interval, where the time interval is the time difference between adjacent snapshots within the control period. When the rate of change of the gradient direction vector angle at a grid point exceeds a first set threshold, that grid point is marked as a gradient direction inflection point. The first threshold value is determined based on the spatial scale of the cooling system's operating space and the rate of load change, typically ranging from 0.5 to 2.0 radians per second, with a default value of 1.0 radians per second. The distribution of gradient direction inflection points in space is recorded as a coordinate list, with each element containing three-dimensional spatial coordinates and the moment the inflection was detected.

[0087] Enthalpy time gradient time-domain analysis is used to identify points of periodic gradient fluctuations. Each grid point has multiple enthalpy time gradient values ​​within the control period, forming a time series. This time series is analyzed point-by-point, calculating the rate of change of the enthalpy time gradient between adjacent time points. The rate of change is defined as the difference between the gradient value at the next time point and the gradient value at the previous time point, divided by the time interval. The method for determining whether the direction of change has reversed is to check the sign of the gradient rate of change at three consecutive time points. A reversal of the direction of change is considered to have occurred when the first rate of change is positive and the second is negative, or vice versa. Simultaneously, the absolute value of the rate of change must exceed a second set threshold, which is determined based on the fluctuation range of the enthalpy time gradient, typically ranging from 0.01 to 0.1 kJ / (kg·s²), with a default value of 0.05 kJ / (kg·s²). Grid points that meet both conditions—a rate of change exceeding the threshold and a reversal of the direction of change—are marked as points of periodic gradient fluctuations. The gradient periodic fluctuation points are also recorded as a list of coordinates, with each element containing three-dimensional spatial coordinates and the time when the fluctuation was detected.

[0088] Gradient direction inflection points and gradient periodic fluctuation points are merged into a unified feature point set, where each element represents a three-dimensional spatial coordinate. Clustering analysis employs a density-based spatial clustering algorithm, which automatically clusters points by defining a neighborhood radius and a minimum point count threshold. The neighborhood radius is defined as the maximum spatial distance considered as neighbors around a given feature point, typically ranging from 0.5 to 2.0 m, adjusted according to the size of the cooling system's operating space and the desired region granularity; the default value is set to 1.0 m. The minimum point count threshold is defined as the minimum number of feature points required to form a cluster, typically ranging from 5 to 20; the default value is set to 10. The clustering algorithm iterates through all feature points. For each feature point, it calculates the number of other feature points within its neighborhood radius. When the number of feature points in its neighborhood reaches or exceeds the minimum point count threshold, the feature point is marked as a core point. Starting from any core point, all points within its neighborhood are added to the same cluster, and the neighborhood of newly added points is recursively checked, continuously adding points with achievable density to the cluster. Feature points not belonging to any cluster are marked as noise points. After clustering is completed, multiple spatial clusters are obtained, and each cluster corresponds to a dynamic load region.

[0089] The boundary of the dynamic load region is determined by the outer points of the clusters. For each cluster, the minimum and maximum values ​​of all its feature points in three spatial directions are calculated to form a cuboid bounding box surrounding the cluster. The six faces of the bounding box are determined by the minimum and maximum values ​​in the length, width, and height directions, respectively. Each face of the bounding box is extended outward by a boundary buffer distance, which is set to half the neighborhood radius to ensure that the bounding box fully covers the influence range of load changes within the region. The extended bounding box serves as the final boundary of the dynamic load region. When the bounding boxes of multiple clusters overlap, the overlapping areas are assigned according to the feature point density. The number of feature points in each cluster within the overlapping area is calculated, and the overlapping area is assigned to the cluster with the more feature points. The number of dynamic load regions is equal to the number of clusters, and each region has a unique region number, incrementing from 1. The region boundaries are stored in the form of six boundary values, including the minimum and maximum values ​​in the length, width, and height directions.

[0090] The load fluctuation intensity of each dynamic load region is obtained by calculating the fluctuation characteristics of the enthalpy time gradient within the region. For a given dynamic load region, the enthalpy time gradient values ​​of all grid points within the region's boundary are extracted over the entire control period. The standard deviation of these gradient values ​​is calculated as the basic indicator of fluctuation characteristics. The standard deviation is calculated as the square root of the sum of the squares of the differences between all gradient values ​​and the average gradient value, divided by the total number of gradient values. The standard deviation reflects the dispersion of the enthalpy time gradient; a larger standard deviation indicates more severe load fluctuations. In addition to the standard deviation, the peak-to-peak value of the gradient values ​​is also calculated, defined as the maximum gradient value minus the minimum gradient value. The peak-to-peak value reflects the extreme range of the enthalpy time gradient; a larger peak-to-peak value indicates a larger load fluctuation amplitude. The load fluctuation intensity is obtained by weighting the standard deviation and peak-to-peak value, with weighting coefficients of 0.6 and 0.4, respectively. The load fluctuation intensity equals the standard deviation multiplied by 0.6 plus the peak-to-peak value multiplied by 0.4. The unit for load fluctuation intensity is kJ / (kg·s), and the numerical range depends on the load characteristics of the space in which the refrigeration system operates, with a typical range of 0.05 to 0.5 kJ / (kg·s).

[0091] The local enthalpy difference characteristic is formed by combining load fluctuation intensity and regional location information. Regional location information includes the geometric center coordinates of the region's boundary box. The three components of the geometric center coordinates are the average of the minimum and maximum values ​​in the length direction, the width direction, and the height direction. Regional location information also includes the region volume, which is equal to the length boundary difference multiplied by the width boundary difference multiplied by the height boundary difference. The local enthalpy difference characteristic is stored in structured data format, with one record corresponding to each dynamic load region. Record fields include region number, geometric center length coordinates, geometric center width coordinates, geometric center height coordinates, region volume, load fluctuation intensity, minimum boundary in the length direction, maximum boundary in the length direction, minimum boundary in the width direction, maximum boundary in the width direction, minimum boundary in the height direction, and maximum boundary in the height direction. The local enthalpy difference characteristic set contains records for all dynamic load regions and is used for subsequent control strategy optimization.

[0092] This invention innovatively extends the enthalpy difference principle from a testing and evaluation tool to a fundamental theory for the operation and control of refrigeration systems. Through precise perception of the enthalpy distribution field within space, dynamic extraction of regional enthalpy difference characteristics, and energy balance calculation based on enthalpy difference, it achieves adaptive optimization of the refrigeration system's operating strategy, transforming the enthalpy difference principle from a static testing and evaluation tool into a dynamic operation and control method. This invention overcomes the limitations of traditional fixed-region division. By identifying the directional inflection points and periodic fluctuation points of the enthalpy gradient, it can adaptively capture the dynamic changes in the actual load distribution. This method considers both spatial and temporal gradients, achieving automatic identification and boundary division of load regions. This dynamic division method can respond promptly to changes in load distribution, avoiding the problem of uneven load distribution within regions caused by fixed-region division.

[0093] In one optional implementation, the steps of calculating the cross-correlation coefficient and phase difference of the load fluctuation intensity time series of adjacent regions to construct the topologically coupled network include:

[0094] Based on the load fluctuation intensity of each dynamic load region in the local enthalpy difference characteristics, a time series of load fluctuation intensity for each dynamic load region is formed;

[0095] To determine the spatial adjacency between dynamic load areas, for two spatially adjacent dynamic load areas, calculate the cross-correlation function of their load fluctuation intensity time series, extract the maximum value of the cross-correlation function as the cross-correlation coefficient, and extract the time offset corresponding to the maximum value of the cross-correlation function as the phase difference;

[0096] A topologically coupled network is constructed by using each dynamic load region as a network node, the cross-correlation coefficient between spatially adjacent dynamic load regions as the weight of the connecting edge, and the phase difference as the directional attribute of the connecting edge.

[0097] For example, the local enthalpy difference characteristics of each dynamic load region include load fluctuation intensity values, which represent the load fluctuation characteristics of that region at a specific moment. Forming a time series of load fluctuation intensity requires repeatedly calculating the load fluctuation intensity for each dynamic load region over multiple consecutive control cycles. The time interval of the control cycle is set to 60 seconds, and the continuous acquisition time window is set to 30 control cycles, corresponding to an observation duration of 1800 seconds. For a given dynamic load region, at the end of each control cycle, the enthalpy time gradient values ​​of all grid points within the region's boundary are extracted, and the standard deviation and peak-to-peak value are calculated. The load fluctuation intensity for that control cycle is obtained by weighting the standard deviation by 0.6 and the peak-to-peak value by 0.4. The load fluctuation intensities of the 30 control cycles are arranged chronologically to form a time series of length 30. Each element in the time series corresponds to a load fluctuation intensity value at a specific moment, with element indices ranging from 0 to 29. Index 0 corresponds to the first control cycle, and index 29 corresponds to the thirtieth control cycle. The load fluctuation intensity time series is stored in the form of a one-dimensional array, where the array elements are floating-point numbers with a value range determined according to the regional load characteristics, typically ranging from 0.05 to 0.5 kJ / (kg·s). A load fluctuation intensity time series is generated for each dynamic load region. If there are 6 dynamic load regions, 6 time series are generated, each with a length of 30.

[0098] Spatial adjacency between dynamic load regions is determined by the geometric position of the region boundaries. For any two dynamic load regions, six boundary values ​​of their bounding boxes are obtained, including minimum and maximum length values, minimum and maximum width values, minimum and maximum height values, and maximum height values. The method for determining whether two regions are spatially adjacent is to check the shortest distance between their bounding boxes. The coordinates of the eight vertices of the first region's bounding box are calculated. Each vertex coordinate is obtained by combining the boundary values ​​in the length, width, and height directions, resulting in eight possible combinations for each vertex. The coordinates of the eight vertices of the second region's bounding box are calculated similarly. The shortest distance from each vertex of the first region to each face of the second region is calculated. Similarly, the shortest distance from each vertex of the second region to each face of the first region is calculated. The minimum value of all vertex-to-face distances is defined as the shortest distance between the two bounding boxes. When the shortest distance is less than or equal to the adjacency threshold, the two regions are considered spatially adjacent. The adjacency threshold is determined based on the spatial scale and granularity of the cooling system's operating space, typically ranging from 0.5 to 2.0 m, with a default value of 1.5 m. Spatial adjacency relationships are stored in the form of an adjacency matrix. The row and column indices of the matrix correspond to the dynamic load region numbers. The matrix elements are Boolean values; the corresponding element is true when two regions are spatially adjacent, and false otherwise. The adjacency matrix is ​​a symmetric matrix, with diagonal elements being false, indicating that a region is not adjacent to itself.

[0099] For two spatially adjacent dynamic load regions, calculate the cross-correlation function of their load fluctuation intensity time series. The cross-correlation function describes the similarity between the two time series at different time offsets. Let the time series of the first region be series A, and the time series of the second region be series B. Let the time offset be parameter tau, where tau ranges from negative sequence length plus one to positive sequence length minus one. For a given time offset tau, the cross-correlation function value is calculated by multiplying series A and the offset sequence B point by point and summing the results. When tau is positive, sequence B is offset backward by tau time steps, and the product of the first few elements of sequence A and the corresponding elements of sequence B after the offset is calculated. When tau is negative, sequence B is offset forward by the absolute value tau time steps, and the product of the last few elements of sequence A and the corresponding elements of sequence B after the offset is calculated. The effective calculation range is the overlapping part of the two sequences, where the length of the overlapping part is equal to the sequence length minus the absolute value of the offset. Summing the point-by-point products of the overlapping part yields the original cross-correlation value at that offset. To eliminate the influence of sequence length and amplitude on the cross-correlation function, the original values ​​are normalized. The normalization method involves dividing the original value by the product of the standard deviations of the two sequences in the overlapping region, and then dividing by the overlap length. The normalized cross-correlation function values ​​range from -1 to +1, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation; the absolute value represents the correlation strength. All time offsets (tau) are iterated, and the cross-correlation function value corresponding to each tau is calculated, forming a cross-correlation function curve. The cross-correlation function curve is stored in a one-dimensional array, with the array index corresponding to the time offset and the array element being the normalized cross-correlation function value.

[0100] The maximum value of the cross-correlation function is used as the cross-correlation coefficient, reflecting the strongest correlation between the time series of load fluctuations in two regions. By iterating through all elements of the cross-correlation function curve, the element with the largest absolute value is found; this value is the cross-correlation coefficient. The cross-correlation coefficient ranges from -1 to +1. Values ​​close to +1 indicate highly synchronized load fluctuations in the two regions, values ​​close to -1 indicate highly inverse load fluctuations, and values ​​close to zero indicate no significant correlation between load fluctuations. The time offset corresponding to the maximum value of the cross-correlation function is used as the phase difference, reflecting the temporal relationship between load fluctuations in the two regions. The array index corresponding to the cross-correlation coefficient is extracted, and the index value is converted into a time offset, with the unit being the number of control cycles. A positive phase difference indicates that the load fluctuation in the second region lags behind the first region, while a negative phase difference indicates that the load fluctuation in the second region leads the first region. The absolute value of the phase difference reflects the degree of time delay; multiplying the phase difference by the control cycle time interval yields the actual time delay, in seconds. The cross-correlation coefficient and phase difference are stored as key-value pairs, where the key is a combination of the two region numbers, and the value is a structure containing the cross-correlation coefficient and phase difference.

[0101] The topologically coupled network is represented by a graph data structure, which contains a set of nodes and a set of edges. Each dynamic load region serves as a network node, and the number of nodes equals the number of dynamic load regions. Each node has a unique node identifier, which is consistent with the region number. Node attributes include the geometric center coordinates of the region, the region volume, and the region boundary extent; these attributes are extracted from local enthalpy difference characteristics. Connection edges are established between spatially adjacent dynamic load regions, and the existence of an edge is determined by the adjacency matrix. For elements in the adjacency matrix with a true value, an edge is created between the corresponding two nodes. The weight of the edge is the cross-correlation coefficient, with weight values ​​ranging from -1 to +1; a larger absolute weight value indicates a stronger coupling between the two regions. The direction attribute of the edge is the phase difference, with phase difference values ​​being integers, in units of the number of control cycles, ranging from negative sequence length plus one to positive sequence length minus one. The edge data structure contains four fields: start node identifier, end node identifier, weight, and phase difference. Due to the symmetry of the cross-correlation, the edge from node i to node j has the same weight as the edge from node j to node i, but the phase difference sign is opposite. When a topologically coupled network is represented as an undirected weighted graph, edges are recorded only once, and the sign of the phase difference is defined as the phase difference when pointing from a node with a smaller number to a node with a larger number. Topologically coupled networks are stored in the form of adjacency lists, with each node maintaining a list of neighbors. Each list element contains the neighbor node identifier, edge weight, and phase difference. Global network attributes include the total number of nodes, the total number of edges, the average weight, and the standard deviation of the weights. These attributes are used for network topology analysis and control strategy optimization.

[0102] The implementation method for calculating the enthalpy decay rate of air supply reaching each region based on fluid boundary parameters is as follows: Fluid boundary parameters refer to the set of key parameters describing the physical characteristics of the air supply system, including air outlet temperature, humidity, wind speed, air supply duct length, duct material thermal conductivity, duct cross-sectional dimensions, number of bends, and branch structure. These parameters directly affect the energy loss of air during transmission. When calculating the enthalpy decay rate, the path from the air outlet to each dynamic load region is decomposed into three basic units: straight pipe sections, bends, and branches. For each straight pipe section, the heat loss coefficient α is calculated based on the pipe length L (meters), pipe diameter D (meters), wall thermal conductivity K (W / m·K), and pipe wall thickness δ (meters); for each bend, the local resistance coefficient β is calculated based on the bending angle and cross-sectional change; for each branch, the flow distribution influence coefficient γ is calculated based on the flow distribution ratio. These coefficients are combined with the supply air temperature difference ΔT (the difference between supply air temperature and ambient temperature) and the humidity difference ΔW, and the total energy loss is obtained using a cumulative calculation method. The energy loss is then divided by the initial enthalpy value to obtain the enthalpy decay rate. For example, for a path containing a 15-meter straight pipe (α=0.02 / meter), two 90° bends (β=0.15 / benchmark), and one three-way branch (γ=0.1), with a supply air temperature of 15℃ and an ambient temperature of 25℃, the calculated enthalpy decay rate is 15×0.02×10 + 2×0.15×10 + 0.1×10 = 5.0%. This simplifies the complex heat transfer process into an engineering calculation model, balancing accuracy and practicality.

[0103] This invention reveals the temporal correlation and impact propagation characteristics of load changes between regions by calculating the cross-correlation of time series of load fluctuation intensities in adjacent regions. By introducing cross-correlation coefficients as edge weights and phase difference as a directional attribute, the network can simultaneously characterize spatial adjacency and temporal influence relationships. This characterization method overcomes the limitations of traditional methods that only consider spatial adjacency, and can capture the dynamic coupling characteristics between regions.

[0104] In one optional implementation, the step of calculating the corrected cooling load by performing energy balance calculations using local enthalpy difference characteristics, coupling weights of the topological coupled network, and enthalpy decay rate includes:

[0105] For each dynamic load region, the load fluctuation intensity and region boundary information of the region are extracted from the local enthalpy difference characteristics, and the coupling weight and phase difference between the region and its neighboring regions are extracted from the topological coupling network.

[0106] The base cooling load is calculated based on the intensity of load fluctuations in the region and the region boundary information.

[0107] Based on the coupling weight and phase difference between the dynamic load area and each adjacent area, the heat and moisture transfer influence of the adjacent area on the area is calculated. The heat and moisture transfer influence is equal to the product of the load fluctuation intensity of the adjacent area and the coupling weight, and the accumulated value after time offset correction according to the phase difference.

[0108] The effective cooling capacity when the supply air reaches the area is calculated based on the enthalpy decay rate of the dynamic load area. The effective cooling capacity is equal to the energy value corresponding to the difference between the nominal cooling capacity of the supply air and the enthalpy decay rate.

[0109] The total heat and humidity load of the region is obtained by adding the basic cooling load of the dynamic load region to the heat and humidity transfer influence. The energy balance calculation is performed on the total heat and humidity load of the region and the effective cooling capacity to obtain the corrected cooling load of the dynamic load region.

[0110] Combination Figure 2 The flowchart for calculating the energy balance correction cooling load based on local features and topological coupling is illustrated below. For example, when calculating the correction cooling load for each dynamic load region, relevant parameters need to be extracted from the local enthalpy difference characteristics and the topological coupling network. The local enthalpy difference characteristics are stored in a structured table, with each row corresponding to a dynamic load region. For a given dynamic load region, the load fluctuation intensity field and the region boundary information field are read from the table. The load fluctuation intensity is a floating-point number in kJ / (kg·s), reflecting the degree of fluctuation in the time gradient of the enthalpy value of the region. The region boundary information includes six boundary values: minimum boundary in the length direction, maximum boundary in the length direction, minimum boundary in the width direction, maximum boundary in the width direction, minimum boundary in the height direction, and maximum boundary in the height direction, in meters. The topological coupling network is stored in an adjacency list. For a node corresponding to a given dynamic load region, the node identifiers, coupling weights, and phase differences of all adjacent regions are extracted from its neighbor list. The coupling weights are floating-point numbers, ranging from -1 to +1, reflecting the correlation strength between the load fluctuations of this region and its adjacent regions. The phase difference is an integer, measured in control cycles, reflecting the time offset of load fluctuations in adjacent areas relative to the current area. The extraction process is completed using area number indexes, which serve as table row indexes and network node identifiers, ensuring consistent data access.

[0111] The basic cooling load is calculated based on the load fluctuation intensity and boundary information of the region. The boundary information is used to calculate the region volume, which is equal to the length boundary difference multiplied by the width boundary difference multiplied by the height boundary difference. The length boundary difference is the maximum length boundary minus the minimum length boundary; the width and height boundary differences are calculated similarly. The region volume is measured in cubic meters, reflecting the spatial scale of the region. The basic cooling load is defined as the load fluctuation intensity multiplied by the region volume and then by the air density. The air density is determined based on the temperature and humidity conditions of the space where the cooling system operates, typically taking a value of 1.2 kg / m³. This value is applicable under standard atmospheric pressure and room temperature conditions, but needs to be corrected when the experimental chamber environment deviates from standard conditions. The basic cooling load is measured in kilowatts (kW). The calculation process involves multiplying the load fluctuation intensity by the region volume and then by the air density. Since the load fluctuation intensity is measured in kJ / (kg·s), multiplying it by the region volume (cubic meters) and air density (kg / m³) converts the unit to kJ / s, i.e., kilowatts. The basic cooling load reflects the heat and humidity load generated by the area itself, without considering the influence of adjacent areas.

[0112] The calculation of the impact of heat and moisture transfer considers the coupling relationship between the dynamic load area and its adjacent areas. For each adjacent area, the load fluctuation intensity of the adjacent area, the coupling weight between the current area and the adjacent area, and the phase difference are obtained. The load fluctuation intensity of the adjacent area is extracted from the local enthalpy difference characteristic table, and the load fluctuation intensity field of the corresponding row is indexed according to the adjacent area number. The instantaneous impact of the adjacent area on the current area is calculated by multiplying the load fluctuation intensity of the adjacent area by the coupling weight. The absolute value of the coupling weight reflects the intensity of the impact, and the sign reflects the direction of the impact. A positive value indicates that an increase in the load fluctuation of the adjacent area will lead to an increase in the load of the current area, and a negative value indicates a reverse impact. The phase difference is used for time offset correction. When the phase difference is positive, it indicates that the load fluctuation of the adjacent area lags behind the current area. When calculating the impact of the current area, the load fluctuation intensity of the adjacent area at a past time should be used. Specifically, the time offset is the current time index minus the phase difference value to obtain the index of the adjacent area at a past time. The value at that time is extracted from the time series of the load fluctuation intensity of the adjacent area. When the phase difference is negative, it indicates that the load fluctuation of the adjacent area leads the current area. The load fluctuation intensity of the adjacent area at a future time should be used. If the index after time offset exceeds the time series range, the influence of the adjacent region on the region at the current time is set to zero. The total influence of heat and moisture transfer is obtained by summing the influence values ​​of all adjacent regions after time offset correction. The summation process involves traversing all adjacent regions, calculating the influence value for each, and then summing them. The unit of heat and moisture transfer influence is kJ / (kg·s), consistent with the unit of load fluctuation intensity.

[0113] Enthalpy decay rate describes the enthalpy change characteristics of supplied air as it travels from the air outlet to the dynamic load area. It is calculated by analyzing the temperature and humidity field distribution along the air supply path. For a given dynamic load area, the spatial distance between this area and the nearest air outlet is determined. This spatial distance is calculated as the Euclidean distance between the geometric center coordinates of the area and the coordinates of the air outlet, using the square root of the sum of the squares of the differences in the three coordinate axes. The air outlet coordinates are determined by the layout of the cooling system's operating space; a typical configuration includes both location coordinates and air supply direction. Based on the spatial distance and enthalpy distribution field, an enthalpy sequence is extracted along the air supply path, containing both the enthalpy at the air outlet and the enthalpy at the geometric center of the area. The enthalpy decay rate is defined as the enthalpy at the air outlet minus the enthalpy at the area center, divided by the spatial distance, with units of kJ / (kg·m). A positive enthalpy decay rate indicates that the enthalpy of the supplied air decreases during transmission, resulting in reduced cooling capacity. A negative enthalpy decay rate indicates that the enthalpy of the air supply increases during transmission, influenced by the heat source. The typical range for the enthalpy decay rate is 0 to 2.0 kJ / (kg·m), with a default value of 0.8 kJ / (kg·m).

[0114] The nominal cooling capacity of the air supply is defined as the difference between the enthalpy of the air supply at the air outlet and the target enthalpy of the indoor air, multiplied by the air supply mass flow rate. The enthalpy of the air supply is extracted from the enthalpy distribution field at the air outlet location. The target enthalpy of the indoor air is set according to the control objective, with a typical value of 48 kJ / kg. The air supply mass flow rate is determined based on the operating parameters of the air supply fan, typically ranging from 0.5 to 5.0 kg / s. The nominal cooling capacity of the air supply is expressed in kilowatts (kW), calculated as the enthalpy difference multiplied by the air supply mass flow rate. The unit conversion is kJ / kg multiplied by kg / s to obtain kJ / s, which is equivalent to kW. The calculation of the effective cooling capacity requires deducting the capacity loss caused by enthalpy decay. The loss is equal to the enthalpy decay rate multiplied by the spatial distance multiplied by the air supply mass flow rate. The enthalpy decay rate is expressed in kJ / (kg·m), multiplied by the spatial distance in meters to obtain kJ / kg, and then multiplied by the air supply mass flow rate in kg / s to obtain kilowatts. Effective cooling capacity equals the nominal cooling capacity of the air supply minus the loss, and is measured in kilowatts. When the enthalpy decay rate is large or the spatial distance is far, the effective cooling capacity is significantly lower than the nominal value, requiring improvement by increasing the air supply volume or optimizing the air supply layout.

[0115] The total regional heat and humidity load is calculated by adding the base cooling load to the heat and humidity transfer effect. The base cooling load is in kilowatts (kW), and the heat and humidity transfer effect is in kJ / (kg·s). To ensure unit consistency, the heat and humidity transfer effect needs to be multiplied by the regional volume and air density to convert it to kilowatts. The conversion formula is: heat and humidity transfer effect multiplied by regional volume multiplied by air density, yielding the transfer effect power in kilowatts. The total regional heat and humidity load equals the base cooling load plus the transfer effect power, in kilowatts, reflecting the actual heat and humidity load level of the region after considering the coupling effects of adjacent regions.

[0116] Energy balance calculations compare the total heat and humidity load of a region with its effective cooling capacity to obtain a corrected cooling load. The corrected cooling load is defined as the difference between the total heat and humidity load and the effective cooling capacity. A positive difference indicates that the heat and humidity load of the region exceeds the effective cooling capacity of the existing air supply, requiring increased cooling capacity or adjustment of air supply parameters to achieve energy balance. A positive corrected cooling load represents the additional cooling capacity required, measured in kilowatts (kW). A negative difference indicates that the effective cooling capacity of the region exceeds the heat and humidity load, posing a risk of overcooling; a negative corrected cooling load represents the amount of cooling capacity that can be reduced. A zero corrected cooling load indicates that the region has reached energy balance and requires no adjustment. The absolute value of the corrected cooling load reflects the degree of deviation from the region's energy balance; a larger value indicates a more severe deviation and a higher control priority.

[0117] This invention provides a modified cooling load calculation method based on energy balance, which comprehensively considers the basic cooling load, the influence of heat and moisture transfer between regions, and the air supply attenuation factor. It achieves an accurate estimate of the actual cooling demand, makes up for the shortcomings of traditional load calculation that ignores transmission losses, ensures that the allocated cooling resources can meet the actual needs of the region, and avoids the problems of local overcooling or insufficient cooling.

[0118] In one optional implementation, the steps for obtaining the cooling resource allocation parameters, based on the energy consumption of the corrected cooling load calculation system and the enthalpy uniformity of each region as optimization objectives, and dynamically adjusting the objective weight coefficients according to the intensity of load fluctuations, include:

[0119] A state vector containing load fluctuation intensity and regional coupling characteristics is constructed. Based on the reinforcement learning policy network, target weight coefficients for each dynamic load region are generated according to the state vector. The generation of the target weight coefficients is based on the topological coupling network using a cooperative adjustment strategy.

[0120] Based on the coupling weights and phase differences in the topologically coupled network, an influence matrix is ​​constructed to characterize the degree of influence and time delay of heat and moisture transfer between regions.

[0121] Using the air supply parameters of each dynamic load region as decision variables, the sum of the system energy consumption after weighting by the target weight coefficient and the enthalpy uniformity index of each region as the optimization objective, and the inter-regional heat and moisture transfer constraints and the physical capacity constraints of the refrigeration equipment defined by the influence matrix as constraints, a multi-objective optimization function is constructed. Solving the multi-objective optimization function yields the refrigeration resource allocation parameters of each dynamic load region.

[0122] The cooling resource allocation parameters are executed, the actual operating effect is monitored, and the parameters of the reinforcement learning strategy network are updated based on the monitoring results, so as to realize the online learning and continuous optimization of the target weight coefficient adjustment strategy.

[0123] For example, the construction of the state vector includes two parts: load fluctuation intensity characteristics and regional coupling characteristics. Load fluctuation intensity characteristics are extracted from local enthalpy difference characteristics. For the six dynamic load regions, the load fluctuation intensity values ​​for each region are extracted, forming a load fluctuation intensity vector of length 6. Regional coupling characteristics are extracted from the topological coupling network, including the number of neighbors, average coupling weight, maximum coupling weight, and average absolute phase difference for each region. The number of neighbors is defined as the total number of adjacent regions connected to a given region in the topological coupling network, obtained from the length of the neighbor list of the region's nodes. The average coupling weight is calculated as the arithmetic mean of the coupling weights between the region and all its adjacent regions; when a region has no adjacent regions, the average coupling weight is set to zero. The maximum coupling weight is the maximum absolute value of the coupling weights between the region and all its adjacent regions. The average absolute phase difference is calculated as the arithmetic mean of the absolute phase differences between the region and all its adjacent regions. For the six dynamic load regions, the number of neighbors, average coupling weight, maximum coupling weight, and average absolute phase difference for each region are each formed into a vector of length 6. The state vector is obtained by concatenating the load fluctuation intensity vector with the coupling feature vectors of the four regions, resulting in a one-dimensional vector of length 30. The elements of this vector are floating-point numbers, normalized to the interval between zero and one. The normalization method is to divide the load fluctuation intensity by the maximum load fluctuation intensity of 0.5 kJ / (kg·s), divide the coupling weight by 1.0, and divide the absolute value of the phase difference by the time series length of 30. The state vector is stored as a one-dimensional array. Array indices 0 to 5 correspond to the load fluctuation intensity of each region, indices 6 to 11 correspond to the number of neighbors, indices 12 to 17 correspond to the average coupling weight, indices 18 to 23 correspond to the maximum coupling weight, and indices 24 to 29 correspond to the average absolute value of the phase difference.

[0124] The reinforcement learning policy network is implemented using a deep neural network, receiving a state vector as input and outputting target weight coefficients for each dynamic load region. The policy network consists of three fully connected layers. The first layer has an input dimension of 30 and 64 neurons, using a modified linear unit (MRU) activation function. The second layer has 32 neurons and also uses an MRU activation function. The third layer has an output dimension of 12, containing system energy consumption weights and enthalpy uniformity weights for six regions. The third layer uses a Softmax function to ensure that the sum of the two weight coefficients for each region is 1.0, and that all weight coefficients are positive. The Softmax function groups and normalizes the 12 outputs of the third layer. For region i, the system energy consumption weight coefficient and enthalpy uniformity weight coefficient are calculated by exponentially taking the corresponding output value and then dividing by the sum of the two exponents. The weight parameters of the policy network are initialized using Xavier, and the bias term is initialized to 0.01. The cooperative adjustment policy is implemented based on a topologically coupled network. For adjacent region pairs whose absolute coupling weights are greater than the cooperative threshold, their target weight coefficients are adjusted to be closer to each other. The coordination threshold is set to 0.7. When the absolute value of the coupling weight between two regions exceeds 0.7, the average value of the system energy consumption weight coefficient and the enthalpy uniformity weight coefficient for each region is calculated. This average value is used as the adjusted weight coefficient, covering the original output of the policy network. Coordination adjustment ensures that strongly coupled regions adopt similar control strategies, avoiding oscillations caused by local optimization.

[0125] The influence matrix characterizes the degree of influence and time delay of heat and moisture transfer between regions. The influence matrix is ​​a square matrix, with both the number of rows and columns equal to the number of dynamic load regions. For six regions, the influence matrix dimension is 6×6. Matrix element (i,j) represents the influence coefficient of heat and moisture transfer from region j to region i. The influence coefficient is calculated based on the coupling weights and phase differences in the topologically coupled network. When region j is an adjacent region of region i, the influence coefficient equals the absolute value of the coupling weight multiplied by the time delay attenuation factor. The time delay attenuation factor is determined based on the absolute value of the phase difference, calculated as 1.0 minus the absolute value of the phase difference divided by the time series length of 30, then multiplied by the attenuation factor of 0.5. The attenuation factor reflects the degree to which the time delay weakens the influence intensity, typically ranging from 0.3 to 0.7, with a default value of 0.5. When region j is not an adjacent region of region i, the influence coefficient is set to zero. The diagonal elements of the influence matrix are set to zero, indicating that the region is not affected by itself. The influence matrix is ​​stored as a two-dimensional array, with array elements being floating-point numbers ranging from zero to one. The influence matrix is ​​used to construct constraints, ensuring that the control strategy takes into account the interactions between regions.

[0126] The decision variables for the multi-objective optimization function are the supply air parameters for each dynamic load zone, including supply air temperature, supply air humidity, and supply air mass flow rate. For the six zones, the total number of decision variables is 18, with three supply air parameters for each zone. The supply air temperature ranges from 12 to 18 degrees Celsius, the supply air humidity ranges from 0.006 to 0.010 kg / kg dry air, and the supply air mass flow rate ranges from 0.5 to 5.0 kg / s. The decision variables are represented as a one-dimensional vector with a length of 18. Indices 0 to 2 correspond to the supply air temperature, supply air humidity, and supply air mass flow rate for zone 1; indices 3 to 5 correspond to zone 2, and so on. The optimization objectives include system energy consumption and enthalpy uniformity. System energy consumption is defined as the sum of the energy consumption of the supply fans and the chiller units across all zones. The supply fan energy consumption is proportional to the cube of the supply air mass flow rate, with a proportionality coefficient of 0.0015 kW / (kg / s) cubed. The energy consumption of the refrigeration unit is directly proportional to the total cooling capacity, with the proportionality coefficient being the reciprocal of the energy efficiency ratio (EER). A typical EER value is 3.5, and the proportionality coefficient is 0.286. The total cooling capacity equals the sum of the corrected cooling loads for all regions, which are recalculated based on the current air supply parameters. The enthalpy uniformity index is defined as the square root of the sum of the squares of the deviations between the actual and target enthalpy values ​​for each region. The actual enthalpy value for a region is calculated based on the temperature and humidity distribution field of that region, taking the arithmetic mean of the enthalpy values ​​at all grid points within the region's boundary. The target enthalpy value is set at 48.0 kJ / kg. The enthalpy uniformity index is expressed in kJ / kg; a smaller value indicates that the enthalpy value of each region is closer to the target value and the distribution is more uniform. The optimization objective function is the system energy consumption multiplied by its target weight coefficient, plus the enthalpy uniformity index multiplied by its target weight coefficient. The weighted sum of the objectives for the six regions yields the overall objective function value. The objective function value is in kilowatts. Since the enthalpy uniformity index has different dimensions from the system energy consumption, it needs to be normalized. The enthalpy uniformity index is multiplied by the normalization coefficient of 10.0 so that its numerical range is comparable to the system energy consumption.

[0127] The constraints include inter-regional heat and moisture transfer constraints and refrigeration equipment physical capacity constraints. Inter-regional heat and moisture transfer constraints are based on the influence matrix definition, requiring that the actual cooling effect of a region equals the corrected cooling load of that region minus the heat and moisture transfer influence of all adjacent regions on that region. The heat and moisture transfer influence is calculated as the sum of the products of the elements in the corresponding row of the influence matrix and the corrected cooling load of each region. This constraint ensures that the control strategy considers the coupling relationship between regions when allocating cooling resources, avoiding insufficient actual effect in a region despite sufficient cooling capacity due to heat and moisture transfer from adjacent regions. Refrigeration equipment physical capacity constraints include upper and lower limits for supply air temperature, upper and lower limits for supply air humidity, upper and lower limits for supply air mass flow rate, upper limit for total supply air volume, and upper limit for refrigeration unit capacity. The upper and lower limits for supply air temperature are 12 to 18 degrees Celsius, the upper and lower limits for supply air humidity are 0.006 to 0.010 kg / kg dry air, and the upper and lower limits for supply air mass flow rate are 0.5 to 5.0 kg / s. The upper limit for total supply air volume is that the sum of the supply air mass flow rates of all regions does not exceed 20.0 kg / s, a limitation stemming from the total power constraint of the supply fans. The maximum capacity of the refrigeration unit is 50.0 kW for total cooling capacity, a limitation derived from the unit's rated power. Constraints are expressed as inequalities. The optimization solver checks whether the decision variables satisfy all constraints; solutions that do not satisfy the constraints are rejected.

[0128] The multi-objective optimization function is solved using a particle swarm optimization algorithm. The particle swarm contains 50 particles, each representing a set of air supply parameter configurations. The particle position vector has a dimension of 18, corresponding to 18 decision variables. The particle velocity vector also has a dimension of 18, representing the rate of change of position. During particle swarm initialization, the position of each particle is randomly sampled within the range of decision variable values, and the velocity is initialized to zero. During particle swarm iteration, each particle updates its velocity and position based on its own historical best position and the global best position. The historical best position is the position where the particle achieves the minimum objective function value in each iteration, and the global best position is the position where all particles achieve the minimum objective function value in each iteration. The velocity update formula comprises three parts: an inertial term, a cognitive term, and a social term. The inertial term is the current velocity multiplied by an inertial weight of 0.7. The cognitive term is the difference between the current position and the previous historical best position, multiplied by a cognitive coefficient of 1.5, and then multiplied by a random number between zero and one. The social term is the difference between the current position and the global best position, multiplied by a social coefficient of 1.5, and then multiplied by a random number between zero and one. The sum of these three terms is the updated velocity. Position update is the current position plus the updated velocity. The velocity is limited to between the negative and positive maximum velocities, with the maximum velocity set at 20% of the decision variable's range. After position update, constraints are checked; if constraints are violated, the position is adjusted to the constraint boundary. The particle swarm optimization is performed 100 times. In each iteration, the objective function value of all particles is calculated, and the previous historical best and global best are updated. After the iteration terminates, the global best position is used as the optimization solution result, i.e., the cooling resource allocation parameters for each dynamic load region.

[0129] The cooling resource allocation parameters include the supply air temperature, supply air humidity, and supply air mass flow rate for each zone, output in vector form. When executing the cooling resource allocation parameters, the supply air parameters are sent to the air conditioning controller via the communication interface. The communication interface uses the Modbus protocol, with the controller address and register mapping pre-configured. The supply air temperature is written to the temperature setpoint register, the supply air humidity is written to the humidity setpoint register, and the supply air mass flow rate is achieved by adjusting the damper opening. The mapping relationship between damper opening and flow rate is determined through calibration. After the parameters are sent, the air conditioning system adjusts the supply fan speed, chiller load, and damper opening according to the new setpoint to achieve the target supply air parameters. The parameter execution delay is a control cycle time interval of 60 seconds, and the actual operating effect is monitored after execution.

[0130] Monitoring actual operational performance is achieved by collecting temperature and humidity sensor data from various areas. The monitoring cycle is 60 seconds, corresponding to one control cycle. Monitoring indicators include average temperature, average humidity, average enthalpy, and total system energy consumption for each area. Average temperature and average humidity for each area are calculated as the arithmetic mean of all sensor readings within that area, while average enthalpy is calculated based on average temperature and average humidity. Total system energy consumption is obtained by summing the readings from the blower power meter and the chiller power meter. The monitoring indicators are compared with the optimization target to calculate the deviation between actual and predicted system energy consumption, and the deviation between actual and predicted enthalpy uniformity indicators. This deviation serves as a reward signal for reinforcement learning. The reward value is defined as a negative weighted sum of deviations: energy consumption deviation multiplied by an energy consumption penalty coefficient of 0.5, and enthalpy uniformity deviation multiplied by a uniformity penalty coefficient of 5.0. The negative sum of these two values ​​yields the reward. Smaller deviations result in larger rewards; positive rewards occur when actual performance is better than prediction, and negative rewards occur when actual performance is worse than prediction.

[0131] The reinforcement learning policy network parameter update employs the policy gradient method combined with an advantage function to improve training stability. The system stores the current control cycle's state vector, target weight coefficients, allocation parameters, and reward value as empirical samples in a 1000-sample replay buffer. When 32 samples have accumulated, they are randomly selected for batch training. The loss function comprises two parts: policy loss and value loss. The policy network and value network share the first two layers of the network structure. The policy loss is calculated as the log probability of the weight coefficients multiplied by the negative mean of the advantage value. The value loss is the mean squared error between the value estimate and the actual reward. The advantage value equals the actual reward minus the value estimate; a positive value indicates performance better than the average. The total loss function combines both parts, with the value loss weight coefficient set to 0.5. Network parameter updates use the Adam optimizer with a learning rate of 0.0001. The optimizer calculates gradients through backpropagation, maintaining estimates of the first and second moments for each parameter. The parameter update step size comprehensively considers both the learning rate and the moment estimates. The system saves a checkpoint after every 10 training batches, including network weights and optimizer state. A rollback mechanism is triggered when a new version performs poorly on the validation set. The online learning process forms a complete closed loop: the policy network continuously receives state vectors, generates weight coefficients, performs optimization, monitors the effect, and updates parameters, achieving continuous optimization of the weight adjustment strategy.

[0132] The reinforcement learning-based dynamic adjustment method for target weight coefficients in this invention achieves an adaptive balance between system energy consumption and enthalpy uniformity. The introduced influence matrix incorporates inter-regional heat and moisture transfer constraints into the optimization framework, ensuring the physical feasibility of the optimization results. The closed-loop feedback mechanism and online learning capability enable the system to continuously accumulate operational experience and optimize control strategies, achieving adaptive optimization of refrigeration resource allocation and long-term performance improvement.

[0133] In one optional implementation, the generation of the target weight coefficients based on the cooperative adjustment strategy employed by the topologically coupled network includes the following steps:

[0134] Based on the distance relationships and connection methods between regions in the topologically coupled network, and combined with the enthalpy gradient change trajectory of each region, a collaborative adjustment matrix is ​​constructed. The elements in the collaborative adjustment matrix represent the influence relationship of the adjustment of the target weight coefficients between regions.

[0135] The region groups are divided according to the element values ​​of the collaborative adjustment matrix. Regions whose influence relationship exceeds the preset influence threshold are identified as collaborative region groups. The master-slave relationship of the regions within the group is determined based on the magnitude of the influence relationship, and an initial adjustment scheme for the target weight coefficient of the collaborative region is generated.

[0136] The enthalpy gradient change trend within the collaborative region group is calculated based on the collaborative adjustment matrix. When the enthalpy gradient change trend exceeds the preset range, the target weight coefficient of the corresponding region in the initial adjustment scheme is adjusted until the enthalpy gradient change trend meets the preset range requirement.

[0137] The adjustment scheme that meets the preset range requirements is used as the final target weight coefficient for solving the multi-objective optimization function.

[0138] For example, the construction of the coordinated adjustment matrix is ​​based on the distance relationships, connection methods, and enthalpy gradient change trajectories between regions in the topologically coupled network. The distance relationships between regions are obtained by calculating the Euclidean distance between the geometric center coordinates of each dynamic load region. For region i and region j, the distance is calculated as the square root of the sum of the squares of the differences in the coordinates of the geometric centers of the two regions in the length, width, and height directions. The distance matrix is ​​a 6×6 symmetric matrix, where the matrix element (i,j) represents the spatial distance between region i and region j in meters, with diagonal elements being zero. The connection method is extracted from the adjacency matrix of the topologically coupled network. The adjacency matrix elements are Boolean values; the element is true when two regions are spatially adjacent, and false otherwise. The enthalpy gradient change trajectory contains the evolution data of the temporal and spatial gradients of enthalpy for each region within the control period. For a given region, the average temporal gradient of enthalpy at 30 time points is extracted, forming a time series of length 30, which reflects the trend of enthalpy change over time in that region. The difference between the enthalpy time gradients at two adjacent moments is calculated to obtain the enthalpy gradient change rate sequence, which has a length of 29. The standard deviation of the enthalpy gradient change rate is used as an indicator of the activity of the enthalpy gradient change in this region; the larger the standard deviation, the more drastic the enthalpy gradient change in this region.

[0139] The calculation of the collaborative adjustment matrix element (i,j) integrates three factors: distance relationship, connection mode, and enthalpy gradient change activity. When region i and region j are adjacent in the adjacency matrix, the basic influence coefficient is set to the absolute value of the corresponding coupling weight in the topological coupled network. When region i and region j are not adjacent, the basic influence coefficient is set to zero. The distance correction factor is calculated based on the spatial distance between the two regions, using the method of 1.0 divided by 1.0 plus the distance value divided by the distance attenuation constant, which is set to 5.0m. This value reflects the degree to which distance weakens the influence relationship. The distance correction factor ranges from zero to one; the larger the distance, the smaller the correction factor. The activity correction factor is calculated based on the geometric mean of the enthalpy gradient change activity of the two regions. The geometric mean is the square root of the product of the standard deviations of the activity of the two regions. The activity correction factor is obtained by normalizing the geometric mean to the interval between zero and one, with the normalization divisor being the maximum value of the standard deviations of the activity of all regions. The collaborative adjustment matrix element equals the basic influence coefficient multiplied by the distance correction factor, and then multiplied by the activity correction factor. The collaborative adjustment matrix is ​​a 6×6 asymmetric matrix, with elements ranging from zero to one. Larger element values ​​indicate a stronger influence of region j on the target weight coefficient adjustment of region i. The collaborative adjustment matrix is ​​stored as a two-dimensional array, with array elements being floating-point numbers.

[0140] The division of region groups is based on the comparison between the element values ​​of the collaborative adjustment matrix and a preset influence threshold. The preset influence threshold is set to 0.5, which is determined according to the regional coupling strength distribution of the cooling system's operating space, typically ranging from 0.4 to 0.7. All elements of the collaborative adjustment matrix are traversed. When the value of element (i,j) exceeds the preset influence threshold, region i and region j are identified as a region pair with a collaborative relationship. A disjoint-set data structure is used to manage the collaborative relationships. During disjoint-set initialization, each region is an independent set. When a collaborative relationship is found between region i and region j, the sets containing the two regions are merged. After traversal, each connected component in the disjoint-set corresponds to a collaborative region group. The number of collaborative region groups is equal to the number of root nodes in the disjoint-set. Each collaborative region group contains one or more dynamic load regions. For a collaborative region group containing a single region, the target weight coefficient is adjusted independently for that region, unaffected by other regions. For a collaborative region group containing multiple regions, the target weight coefficients of the regions within the group need to be adjusted collaboratively.

[0141] The master-slave relationship within a collaborative region group is determined based on the magnitude of their influence. For a given collaborative region group, the relevant rows and columns of all regions within that group are extracted from the collaborative adjustment matrix to form a submatrix. The sum of the elements in each row of the submatrix is ​​calculated; this row sum represents the total strength of the influence of other regions within the group on that region. The sum of the elements in each column of the submatrix is ​​also calculated; this column sum represents the total strength of the influence of that region on other regions within the group. Dominance is defined as the column sum minus the row sum. A positive dominance value indicates that the region's influence on other regions is greater than the influence of other regions on that region, and that region is dominant in the collaborative region group. A negative dominance value indicates that the region is subordinate in the collaborative region group. All regions within the collaborative region group are sorted from largest to smallest dominance, with the region with the highest dominance determined as the master region, and the others as slave regions. The target weight coefficient adjustment of the master region has the highest priority, while the target weight coefficient adjustment of the slave regions must refer to the adjustment direction of the master region.

[0142] The initial adjustment scheme for the target weight coefficients of the collaborative regions is generated based on the output of the reinforcement learning policy network. For a group of collaborative regions, the initial target weight coefficients for each region within the group are obtained. These initial target weight coefficients are the system energy consumption weight coefficient and enthalpy uniformity weight coefficient directly output by the policy network. For the main region, the initial adjustment scheme keeps the target weight coefficients output by the policy network unchanged. For the secondary region, the initial adjustment scheme moves the target weight coefficients of the secondary region closer to those of the main region. The convergence coefficient is determined based on the influence coefficient from the secondary region to the main region in the collaborative adjustment matrix. The convergence coefficient equals the influence coefficient multiplied by the collaborative strength parameter, which is set to 0.6, with a typical value range of 0.4 to 0.8. The adjusted system energy consumption weight coefficient of the secondary region is calculated by multiplying the initial system energy consumption weight coefficient of the secondary region by 1.0, subtracting the convergence coefficient, and adding the system energy consumption weight coefficient of the main region multiplied by the convergence coefficient. The adjusted enthalpy uniformity weight coefficient of the secondary region is calculated by subtracting the adjusted system energy consumption weight coefficient from 1.0, ensuring that the sum of the two weight coefficients is 1.0. The initial adjustment scheme is stored in dictionary form, with the key being the region number and the value being a tuple containing the system energy consumption weight coefficient and the enthalpy uniformity weight coefficient.

[0143] For a given coordinated region group, the average enthalpy time gradient of each region within the group over the most recent 10 control cycles is extracted, forming a time series of length 10. The slope of the linear fit of this time series is calculated as the trend of enthalpy gradient change. The linear fit uses the least squares method, and the slope is calculated as the covariance of the time series value and the time index divided by the variance of the time index. The time index is an integer sequence from 0 to 9. The covariance is calculated as the difference between the time series value and its average value multiplied by the difference between the time index and its average value, summed over 10 time points, and then divided by 10. The variance of the time index is the sum of the squares of the differences between the index value and the average index value divided by 10. The unit of the enthalpy gradient trend slope is kJ / (kg·s²), with a positive value indicating an upward trend in the enthalpy time gradient and an increase in load fluctuation intensity, and a negative value indicating a downward trend in the enthalpy time gradient and a decrease in load fluctuation intensity. The enthalpy gradient trend slope of all regions within the collaborative region group is calculated by weighting the regional volume of each region. The slope of the weighted average is used as the overall enthalpy gradient trend of the collaborative region group.

[0144] The preset range is defined as the allowable interval of the enthalpy gradient trend slope, with a lower limit of -0.005 kJ / (kg·s²) and an upper limit of +0.005 kJ / (kg·s²). This range reflects the system's desired stable load fluctuation state. When the enthalpy gradient trend slope falls within this range, the load fluctuation is considered controllable. When the overall enthalpy gradient trend slope of the coordinated region group exceeds the preset range, the target weight coefficient of the corresponding region in the initial adjustment scheme needs to be adjusted. The adjustment strategy is determined based on the positive or negative direction of the enthalpy gradient trend slope. When the slope is positive and exceeds the upper limit, it indicates that the load fluctuation intensity is increasing too rapidly, requiring an increase in the enthalpy uniformity weight coefficient to strengthen uniformity control, while simultaneously reducing the system energy consumption weight coefficient. When the slope is negative and below the lower limit, it indicates that the load fluctuation intensity is decreasing too rapidly, leading to overcooling, requiring an increase in the system energy consumption weight coefficient to reduce energy consumption, while simultaneously reducing the enthalpy uniformity weight coefficient.

[0145] The adjustment amount of the target weight coefficient is calculated based on the deviation of the slope of the enthalpy gradient trend from the preset range boundary. The deviation is defined as the absolute value of the difference between the actual slope and the preset range boundary. The adjustment step size is equal to the deviation multiplied by the adjustment sensitivity coefficient, which is set to 20.0, with a typical range of 10.0 to 30.0. This coefficient reflects the response speed of the weight coefficient to the enthalpy gradient trend. The adjustment step size is limited to between 0.01 and 0.15 to avoid oscillations caused by excessively large single adjustments. When it is necessary to increase the enthalpy uniformity weight coefficient, the adjustment step size is increased for all regions within the collaborative region group in the initial adjustment scheme, while the adjustment step size is decreased by the same amount for the system energy consumption weight coefficient. When it is necessary to increase the system energy consumption weight coefficient, the adjustment direction is reversed. The adjusted target weight coefficients must satisfy the non-negativity constraint and the constraint that the sum is 1.0. If one weight coefficient is less than zero after adjustment, it is set to zero, and the other weight coefficient is set to 1.0. If one of the weight coefficients is greater than 1.0 after adjustment, it is set to 1.0, and the other weight coefficient is set to zero.

[0146] After adjustment, the overall enthalpy gradient trend of the coordinated region group is recalculated. Based on the adjusted target weight coefficients, a multi-objective optimization function is simulated to obtain new cooling resource allocation parameters. The enthalpy time gradient of each region is predicted based on these new parameters, and the slope of the predicted enthalpy gradient trend is calculated. The prediction method uses a physically constrained neural network inference. The input is the new air supply parameters and the current time, and the output is the temperature and humidity field distribution for the next 10 control cycles. The enthalpy time gradient of each region is calculated, and the slope of the trend is extracted. The slope of the predicted enthalpy gradient trend is checked to see if it meets the preset range requirement. If it does, the current adjustment scheme is used as the final target weight coefficient; otherwise, adjustment continues. The maximum number of iterations is set to 5. If the preset range requirement is still not met after 5 iterations, the adjustment scheme with the enthalpy gradient trend slope closest to the preset range center value of zero in the 5 iterations is selected as the final target weight coefficient.

[0147] The final target weight coefficients are used to replace the target weight coefficients directly output by the reinforcement learning policy network when solving the multi-objective optimization function. For regions belonging to cooperative region groups, the cooperatively adjusted target weight coefficients are used. For independent regions not belonging to any cooperative region group, the target weight coefficients directly output by the policy network are used. The target weight coefficients are passed to the multi-objective optimization function in dictionary form, with the region number as the key and tuples containing the system energy consumption weight coefficient and the enthalpy uniformity weight coefficient as the value. The multi-objective optimization function constructs a weighted objective function based on the target weight coefficients and executes the particle swarm optimization algorithm to solve for the cooling resource allocation parameters.

[0148] This invention quantitatively characterizes the mutual influence of weight adjustments between regions by constructing a collaborative adjustment matrix. The collaborative region groups and master-slave relationship determination mechanism, based on influence relationships, enable weight adjustments to be performed while considering global impacts. The introduction of a prediction and verification mechanism for enthalpy gradient trends ensures that weight adjustments do not lead to imbalances in enthalpy distribution between regions. This collaborative adjustment method avoids the optimization conflicts caused by traditional independent adjustments, improving the overall coordination and operational stability of multi-region refrigeration systems.

[0149] The implementation method for adjusting the dynamic load area boundary based on the deviation type between the actual and expected enthalpy distributions after collecting actual temperature and humidity data following the execution of refrigeration resource allocation parameters is as follows: A network of temperature and humidity sensors distributed within the refrigeration space collects actual temperature and humidity data after the execution of refrigeration resource allocation parameters at 60-second intervals. Based on the collected temperature and humidity data, the actual enthalpy value at each point is calculated using the enthalpy calculation formula. The enthalpy value is equal to the sum of the enthalpy value of dry air and the enthalpy value of water vapor. The enthalpy value of dry air is temperature multiplied by the specific heat capacity coefficient 1.005, and the enthalpy value of water vapor is humidity multiplied by the latent heat of vaporization 2501 kJ / kg. The calculated actual enthalpy value is compared with the expected enthalpy value predicted by the physical constraint neural network to generate an enthalpy deviation field. The system identifies three typical deviation types in the deviation field: boundary drift type, characterized by a steep deviation gradient at the region boundary; area change type, characterized by the overall region shrinking or expanding; and splitting and merging type, characterized by single-region splitting or multi-region merging. For boundary drift type deviations, correction is achieved by moving the boundary coordinates, with the movement proportional to the gradient direction. For area change type deviations, this is achieved by proportionally adjusting the six boundary values ​​of the region bounding box. For splitting and merging type deviations, the clustering-based region partitioning algorithm is re-executed. After boundary adjustment, the region boundary information database is updated and used as the initial boundary for the next control cycle. The boundary adjustment frequency is dynamically adjusted based on system stability: once every 10 control cycles in a stable state, and once every 2-3 cycles in an unstable state, ensuring that the boundary reflects the actual load distribution.

[0150] A second aspect of the present invention provides an adaptive operation strategy optimization system for a refrigeration system, comprising:

[0151] The first unit is used to acquire temperature and humidity data and fluid boundary parameters within the operating space of the refrigeration system.

[0152] The second unit is used to reconstruct the temperature and humidity data based on the physical constraint neural network to obtain a continuous temperature and humidity distribution field, calculate the enthalpy distribution field and extract the gradient change trajectory, divide the dynamic load region according to the gradient direction inflection point and the gradient periodic fluctuation point, and extract the load fluctuation intensity of each region as the local enthalpy difference feature.

[0153] The third unit is used to calculate the cross-correlation coefficient and phase difference of the load fluctuation intensity time series of adjacent areas to construct a topological coupling network; calculate the enthalpy decay rate when the air supply reaches each area based on the fluid boundary parameters; and perform energy balance calculations on the local enthalpy difference characteristics, the coupling weight of the topological coupling network, and the enthalpy decay rate for each area to obtain the corrected cooling load.

[0154] The fourth unit is used to calculate the system energy consumption and the enthalpy uniformity of each region based on the corrected cooling load as optimization objectives. The objective weight coefficients are dynamically adjusted according to the intensity of load fluctuations to obtain the cooling resource allocation parameters.

[0155] The fifth unit is used to collect actual temperature and humidity data after executing the refrigeration resource allocation parameters, and adjust the dynamic load zone boundary according to the deviation type between the actual enthalpy distribution and the expected enthalpy distribution.

[0156] A third aspect of the present invention provides an electronic device, comprising:

[0157] processor;

[0158] Memory used to store processor-executable instructions;

[0159] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0160] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0161] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

Claims

1. A method for optimizing the adaptive operation strategy of a refrigeration system, characterized in that, include: Acquire temperature and humidity data, and fluid boundary parameters within the operating space of the refrigeration system; The temperature and humidity data are reconstructed using a physically constrained neural network to obtain a continuous temperature and humidity distribution field. The enthalpy distribution field is calculated, and the gradient change trajectory is extracted. Dynamic load regions are divided based on gradient direction inflection points and gradient periodic fluctuation points, and the load fluctuation intensity of each region is extracted as a local enthalpy difference feature. Specifically, the physically constrained neural network includes a temperature prediction branch and a humidity prediction branch. The temperature prediction branch receives spatial location and temporal information as input, while the humidity prediction branch receives spatial location, temporal information, and the hidden layer features of the temperature prediction branch as input. The loss function of the physically constrained neural network includes a data fitting term. The system includes physical constraints, which include thermodynamic conservation constraints and temporal causality constraints. The temporal causality constraints calculate enthalpy based on the outputs of the temperature prediction branch and the humidity prediction branch, and constrain the sign of the time derivative of the enthalpy to be consistent with the direction of energy inflow and outflow. The weight ratio of the data fitting term and the physical constraint term is dynamically adjusted based on the density of temperature and humidity measuring points. The temperature and humidity data are reconstructed using the physical constraint neural network to obtain a continuous temperature and humidity distribution field, and the enthalpy distribution field is calculated. The temporal gradient and spatial gradient of the enthalpy distribution field are obtained, and the evolution process of the temporal gradient and spatial gradient in the spatiotemporal domain is taken as the gradient change trajectory. Calculate the cross-correlation coefficient and phase difference of the load fluctuation intensity time series of adjacent areas to construct a topological coupling network; calculate the enthalpy decay rate when the supply air reaches each area based on the fluid boundary parameters; for each area, perform energy balance calculations based on the local enthalpy difference characteristics, the coupling weight of the topological coupling network, and the enthalpy decay rate to obtain the corrected cooling load; Based on the energy consumption of the modified cooling load calculation system and the uniformity of enthalpy values ​​in each region as optimization objectives, the target weight coefficients are dynamically adjusted according to the intensity of load fluctuations to obtain the cooling resource allocation parameters. Collect actual temperature and humidity data after executing cooling resource allocation parameters, and adjust the dynamic load zone boundary according to the deviation type between the actual enthalpy distribution and the expected enthalpy distribution.

2. The method according to claim 1, characterized in that, The steps for dividing dynamic load regions based on gradient direction inflection points and gradient periodic fluctuation points, and extracting the load fluctuation intensity of each region as a local enthalpy difference characteristic, include: The enthalpy spatial gradient in the gradient change trajectory is oriented, and the spatial position where the rate of change of the gradient direction vector angle exceeds a first set threshold is identified as the gradient direction inflection point. A time-domain analysis is performed on the enthalpy time gradient in the gradient change trajectory to identify spatial locations where the rate of change of the gradient amplitude at adjacent time points exceeds a second set threshold and the direction of change reverses as gradient periodic fluctuation points. Cluster analysis is performed on the spatial distribution of the gradient direction inflection points and the gradient periodic fluctuation points to obtain multiple spatial clusters, and the boundary of each spatial cluster is used as the boundary of the dynamic load region. The fluctuation characteristics of the enthalpy time gradient within each dynamic load region are calculated as the load fluctuation intensity, and combined with regional location information to form local enthalpy difference characteristics.

3. The method according to claim 1, characterized in that, The steps for calculating the cross-correlation coefficient and phase difference of the load fluctuation intensity time series of adjacent regions and constructing the topologically coupled network include: Based on the load fluctuation intensity of each dynamic load region in the local enthalpy difference characteristics, a time series of load fluctuation intensity for each dynamic load region is formed; To determine the spatial adjacency between dynamic load areas, for two spatially adjacent dynamic load areas, calculate the cross-correlation function of their load fluctuation intensity time series, extract the maximum value of the cross-correlation function as the cross-correlation coefficient, and extract the time offset corresponding to the maximum value of the cross-correlation function as the phase difference; A topologically coupled network is constructed by using each dynamic load region as a network node, the cross-correlation coefficient between spatially adjacent dynamic load regions as the weight of the connecting edge, and the phase difference as the directional attribute of the connecting edge.

4. The method according to claim 1, characterized in that, The steps for calculating the corrected cooling load by using local enthalpy difference characteristics, coupling weights of the topologically coupled network, and enthalpy decay rate to perform energy balance calculations include: For each dynamic load region, the load fluctuation intensity and region boundary information of the region are extracted from the local enthalpy difference characteristics, and the coupling weight and phase difference between the region and its neighboring regions are extracted from the topological coupling network. The base cooling load is calculated based on the intensity of load fluctuations in the region and the region boundary information. Based on the coupling weight and phase difference between the dynamic load area and each adjacent area, the heat and moisture transfer influence of the adjacent area on the area is calculated. The heat and moisture transfer influence is equal to the product of the load fluctuation intensity of the adjacent area and the coupling weight, and the accumulated value after time offset correction according to the phase difference. The effective cooling capacity when the supply air reaches the area is calculated based on the enthalpy decay rate of the dynamic load area. The effective cooling capacity is equal to the energy value corresponding to the difference between the nominal cooling capacity of the supply air and the enthalpy decay rate. The total heat and humidity load of the region is obtained by adding the basic cooling load of the dynamic load region to the heat and humidity transfer influence. The energy balance calculation is performed on the total heat and humidity load of the region and the effective cooling capacity to obtain the corrected cooling load of the dynamic load region.

5. The method according to claim 1, characterized in that, The steps for obtaining the cooling resource allocation parameters, based on the energy consumption of the corrected cooling load calculation system and the enthalpy uniformity of each region as optimization objectives, and dynamically adjusting the objective weight coefficients according to the intensity of load fluctuations, include: A state vector containing load fluctuation intensity and regional coupling characteristics is constructed. Based on the reinforcement learning policy network, target weight coefficients for each dynamic load region are generated according to the state vector. The generation of the target weight coefficients is based on the topological coupling network using a cooperative adjustment strategy. Based on the coupling weights and phase differences in the topologically coupled network, an influence matrix is ​​constructed to characterize the degree of influence and time delay of heat and moisture transfer between regions. Using the air supply parameters of each dynamic load region as decision variables, the sum of the system energy consumption after weighting by the target weight coefficient and the enthalpy uniformity index of each region as the optimization objective, and the inter-regional heat and moisture transfer constraints and the physical capacity constraints of the refrigeration equipment defined by the influence matrix as constraints, a multi-objective optimization function is constructed. Solving the multi-objective optimization function yields the refrigeration resource allocation parameters of each dynamic load region. The cooling resource allocation parameters are executed, the actual operating effect is monitored, and the parameters of the reinforcement learning strategy network are updated based on the monitoring results, so as to realize the online learning and continuous optimization of the target weight coefficient adjustment strategy.

6. The method according to claim 5, characterized in that, The steps for generating the target weight coefficients based on the cooperative adjustment strategy of the topologically coupled network include: Based on the distance relationships and connection methods between regions in the topologically coupled network, and combined with the enthalpy gradient change trajectory of each region, a collaborative adjustment matrix is ​​constructed. The elements in the collaborative adjustment matrix represent the influence relationship of the adjustment of the target weight coefficients between regions. The region groups are divided according to the element values ​​of the collaborative adjustment matrix. Regions whose influence relationship exceeds the preset influence threshold are identified as collaborative region groups. The master-slave relationship of the regions within the group is determined based on the magnitude of the influence relationship, and an initial adjustment scheme for the target weight coefficient of the collaborative region is generated. The enthalpy gradient change trend within the collaborative region group is calculated based on the collaborative adjustment matrix. When the enthalpy gradient change trend exceeds the preset range, the target weight coefficient of the corresponding region in the initial adjustment scheme is adjusted until the enthalpy gradient change trend meets the preset range requirement. The adjustment scheme that meets the preset range requirements is used as the final target weight coefficient for solving the multi-objective optimization function.

7. A refrigeration system adaptive operation strategy optimization system, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire temperature and humidity data and fluid boundary parameters within the operating space of the refrigeration system. The second unit is used to reconstruct the temperature and humidity data based on the physical constraint neural network to obtain a continuous temperature and humidity distribution field, calculate the enthalpy distribution field and extract the gradient change trajectory, divide the dynamic load region according to the gradient direction inflection point and the gradient periodic fluctuation point, and extract the load fluctuation intensity of each region as the local enthalpy difference feature. The third unit is used to calculate the cross-correlation coefficient and phase difference of the load fluctuation intensity time series of adjacent areas to construct a topological coupling network; and to calculate the enthalpy decay rate when the air supply reaches each area based on the fluid boundary parameters. For each region, the corrected cooling load is obtained by energy balance calculation based on local enthalpy difference characteristics, coupling weights of the topological coupling network, and enthalpy decay rate. The fourth unit is used to calculate the system energy consumption and the enthalpy uniformity of each region based on the corrected cooling load as optimization objectives. The objective weight coefficients are dynamically adjusted according to the intensity of load fluctuations to obtain the cooling resource allocation parameters. The fifth unit is used to collect actual temperature and humidity data after executing the refrigeration resource allocation parameters, and adjust the dynamic load zone boundary according to the deviation type between the actual enthalpy distribution and the expected enthalpy distribution.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images