Cooling tower group energy efficiency collaborative control method and system based on edge computing

Abstract

The present application belongs to the technical field of intelligent control, and particularly relates to a cooling tower group energy efficiency collaborative control method and system based on edge computing, which comprises the following steps: constructing a topological graph with the heat plume recirculation index and the liquid force coupling coefficient as the edge features; using a graph attention network coding embedded with a heat and mass transfer efficiency prior bias to predict the control sequence and the accidental uncertainty variance; constructing a combined loss function model combining the node centrality and the uncertainty variance penalty to train the model; and when the outlet water temperature deviation exceeds the confidence boundary, a correction control sequence is regenerated and executed. The present application focuses on high energy efficiency nodes by fusing physical topology and efficiency prior guide model, and timely responds to working condition deviation relying on dynamic confidence boundary, thereby improving the system comprehensive energy efficiency and operation reliability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology. More specifically, this invention relates to a method and system for coordinated energy efficiency control of cooling tower groups based on edge computing. Background Technology

[0002] There is some thermal interference between the cooling towers. For example, the recirculation of hot, humid air discharged from adjacent equipment can easily form a hot plume recirculation phenomenon, thereby reducing the intake air heat exchange efficiency. The hydraulic coupling caused by the pipeline connection may also cause uneven flow distribution. Existing control methods treat each cooling tower as an independent control unit, or simply operate them simultaneously based on rule logic, which can easily cause the equipment to operate under conditions that deviate from the high-efficiency range for a long time. At the same time, changes in external meteorological conditions and equipment aging and degradation can introduce random uncertainties, making it difficult for static control strategies to respond to real-time deviations in a timely manner, and also making it difficult to achieve adaptive adjustment of the cooperative control sequence. To solve this problem, intelligent cooperative control technology based on graph neural networks and deep learning can map the cooling tower group into a topological graph. With the equipment as nodes, a graph attention network encoder deployed on the edge side extracts the spatial correlation features of the group, and uses data-driven models to output cooperative control strategies.

[0003] Currently, adaptive control is achieved by collecting multi-dimensional operational data for state modeling, extracting features using intelligent algorithms, and outputting control strategies. Chinese patent application CN121386400A discloses an adaptive energy-saving control method for cooling tower groups based on reinforcement learning. This method utilizes feature mapping to construct independent subspaces for energy efficiency, stability, and safety, and uses a lightweight attention network to calculate the importance weights of each sub-objective to generate a joint state representation vector, which is then input into the reinforcement learning policy network. Based on these techniques, multi-objective dynamic optimization and collaborative control of cooling tower groups can be achieved.

[0004] However, the aforementioned technical solutions can, to some extent, achieve adaptive adjustment of control strategies through data-driven models. However, existing technologies typically rely solely on spatial distance or logical connections when constructing topology graphs, resulting in insufficient representation of the true intensity of disturbances within the group. Conventional graph attention networks primarily depend on data-driven features during feature encoding, lacking prior knowledge constraints such as cooling tower heat and mass transfer efficiency, leaving room for improvement in model generalization ability and consistency. Regarding model training, the centrality differences of different cooling towers within the group topology are often ignored, lacking representation of the variance of accidental uncertainty in prediction results, leading to insufficient optimization of key nodes. When actual operating conditions deviate from predictions, such training and control mechanisms struggle to promptly trigger corrections and regenerate control sequences based on confidence boundaries, thus limiting the overall energy efficiency improvement of the cooling tower group collaborative control system. Summary of the Invention

[0005] To address the technical challenges of limited overall energy efficiency and operational reliability in a coordinated control system for cooling tower groups, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for coordinated energy efficiency control of cooling tower groups based on edge computing, comprising: S1, acquiring real-time operating condition data and historical performance data of the cooling tower group, constructing a topology map of the cooling tower group based on the real-time operating condition data and historical performance data, wherein the nodes of the cooling tower group topology map are each cooling tower, calculating the thermal plume recirculation index and the hydraulic coupling coefficient, and the edge features are weighted by the thermal plume recirculation index and the hydraulic coupling coefficient; S2, using a graph attention network with embedded priors to perform feature encoding on the cooling tower group topology map, wherein in the self-attention calculation, the dot product result of the query vector and the key vector is adjusted by a prior bias function and then subjected to Softmax normalization processing, the prior bias function being based on the transmission of each cooling tower. The heat transfer efficiency parameter is set based on the encoded group features. The decoder predicts the collaborative control sequence, future outlet water temperature prediction sequence, and random uncertainty variance of each cooling tower in parallel. S3, a combined loss function training model is constructed. A loss component is set for each cooling tower. The loss component includes a weighted basic loss and a logarithmic penalty term for the random uncertainty variance. The weight of the basic loss is proportional to the centrality of the node in the cooling tower group topology graph and inversely proportional to the random uncertainty variance. The outlet water temperature of each cooling tower is continuously monitored. When the deviation between the measured and predicted outlet water temperature exceeds the confidence boundary set by the random uncertainty variance, the latest data containing the deviation is used as input to regenerate the correction control sequence and execute it.

[0007] This invention constructs a topology graph by expressing physical fluid interference and incorporates an efficiency prior mechanism, which can accurately characterize the real interference intensity and guide the model feature aggregation to focus on high-efficiency nodes. By combining a loss function of node centrality and uncertainty penalty to implement differentiated training, it effectively improves the optimization strength of key nodes and the model's generalization ability. Relying on the closed-loop correction mechanism of uncertainty dynamic confidence boundary, it can respond to deviations from actual operating conditions in a timely manner and adaptively reconstruct the sequence, thereby improving the overall energy efficiency and operational reliability of the cooling tower group collaborative control system.

[0008] Preferably, the acquisition of real-time operating condition data and historical performance data of the cooling tower group includes: real-time acquisition of inlet water temperature, outlet water temperature, water flow rate and fan operating frequency of each cooling tower as real-time operating condition data through sensors installed on the inlet and outlet water pipes and fans of each cooling tower; extraction of operating condition data sequence within a set period from the historical database of the control system, and alignment of the operating condition data sequence with the real-time operating condition data to construct an input time series feature matrix for network training and inference.

[0009] This invention can eliminate the time delay and timing deviation of multi-source asynchronous sampling devices, providing a precise alignment of temporal multidimensional physical quantities for subsequent network training and inference, and ensuring a high degree of consistency in the feature input time step.

[0010] Preferably, the step of constructing a cooling tower group topology map based on real-time operating data and historical performance data includes: setting each cooling tower in the cooling tower group as a node in the cooling tower group topology map; calculating the thermal plume recirculation index and hydraulic coupling coefficient between any two nodes; performing dimensionless processing on the thermal plume recirculation index and hydraulic coupling coefficient respectively; assigning preset weight constants to the dimensionless thermal plume recirculation index and hydraulic coupling coefficient respectively; summing the weighted result of the thermal plume recirculation index and the weighted result of the hydraulic coupling coefficient; and using the summed result as the edge feature between the corresponding two nodes.

[0011] This invention can transform the two heterogeneous physical properties of complex three-dimensional airflow thermal interference and uneven hydraulic distribution in parallel pipe networks into dimensionless weights of a unified scale, thereby realizing a standardized topological structural characterization of the physical coupling strength within a cooling tower group.

[0012] Preferably, the calculation of the thermal plume recirculation index includes: obtaining the three-dimensional spatial coordinates of any two cooling towers and calculating the spatial Euclidean distance between the two cooling towers; obtaining the real-time wind direction and calculating the cosine of the angle between the airflow propagation direction vector and the direction vector of the line connecting the two cooling towers, and taking the larger value between the cosine of the angle and 0 as the effective wind direction factor; calculating the sum of the operating frequencies of the fans of the two cooling towers; and multiplying the ratio of the preset reference distance to the spatial Euclidean distance, the effective wind direction factor, and the sum of the fan operating frequencies to the preset reference frequency to obtain the thermal plume recirculation index between the two cooling towers.

[0013] Preferably, calculating the hydraulic coupling coefficient includes: measuring the pressure drop of the supply and return water mains of the cooling tower group; obtaining the real-time water flow of the branches where the two target cooling towers are located, and calculating the square of the water flow of each branch; calculating the ratio of the sum of the squares of the water flow of the two target cooling tower branches to the sum of the squares of the preset reference flow to obtain the flow coupling component; calculating the ratio of the pressure drop of the supply and return water mains to the preset reference pressure drop to obtain the pressure drop coupling component; and performing a weighted summation of the flow coupling component and the pressure drop coupling component to obtain the hydraulic coupling coefficient between the two cooling towers.

[0014] Preferably, in the self-attention calculation, the dot product of the query vector and the key vector is adjusted by a prior bias function and then normalized using Softmax. This includes: multiplying the input features of each node by the query weight matrix and the key weight matrix respectively to obtain the corresponding query vector and key vector; calculating the inner product of the query vector and the key vector of any two nodes to obtain the basic attention score; extracting the ratio of the actual heat exchange to the theoretical maximum heat exchange of each cooling tower from the real-time operating data as the heat transfer and mass transfer efficiency parameter; adding the logarithm of the heat transfer and mass transfer efficiency parameter of the corresponding target cooling tower node as a bias term to the basic attention score pointing from the source cooling tower node to the target cooling tower node to obtain the adjusted attention score; and applying the Softmax function to normalize the adjusted attention score.

[0015] Preferably, the construction of the combined loss function training model includes: calculating the feature vector centrality of each cooling tower node based on the constructed cooling tower group topology graph and edge features; obtaining the collaborative control sequence of the decoder output for each cooling tower and the corresponding random uncertainty variance; calculating the basic loss value between the predicted output of each cooling tower and the true label; multiplying the basic loss value of each cooling tower by the corresponding feature vector centrality, dividing the product by the corresponding random uncertainty variance, adding a smoothing term to prevent the denominator from being zero, and then adding it to the logarithmic penalty term of the random uncertainty variance to obtain the loss component of each cooling tower; summing the loss components of all cooling towers to obtain the total combined loss function.

[0016] This invention can force the network to be adaptively sensitive to core hub device nodes during gradient backpropagation optimization by using the mathematical logic of centrality weighting and variance penalty, while smoothing and isolating the inherent system noise of edge nodes or sensors, thus ensuring the convergence stability of end-to-end forward prediction.

[0017] Preferably, before using a graph attention network with embedded priors to encode the features of the cooling tower group topology, the method further includes: inputting the input temporal feature matrix into a temporal feature encoding module, extracting temporal patterns through recurrent neural units to generate a temporal hidden feature matrix; the graph attention network includes a graph attention interaction layer, which uses the temporal hidden feature matrix as the feature of each node, and interacts with the cooling tower group topology containing edge features. Figure 1 The input is fed to the graph attention interaction layer, where self-attention information is aggregated in the spatial dimension to generate node spatial feature representations; the decoder is a dual-branch prediction decoding module.

[0018] Preferably, when the deviation between the measured value and the predicted value of the outlet water temperature exceeds the confidence boundary set by the random uncertainty variance, the corrected control sequence is regenerated and executed using the latest data containing the deviation as input, including: calculating the absolute deviation between the measured value and the predicted future outlet water temperature; triggering a recalculation signal when the square of the absolute deviation is greater than 9 times the random uncertainty variance; and transmitting the latest state features to the decoder through a unified architecture protocol to infer the corrected fan and water valve control sequence through forward propagation.

[0019] This invention can strictly and dynamically define the error tolerance by utilizing the probability statistical standard deviation of network self-evaluation, avoiding false triggering or missed triggering caused by conventional static single threshold setting, and ensuring that the closed-loop correction control mechanism has a stable and reliable response capability under severe external environmental disturbances.

[0020] Secondly, the present invention provides an edge computing-based cooling tower group energy efficiency collaborative control system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned edge computing-based cooling tower group energy efficiency collaborative control method is implemented.

[0021] By adopting the above technical solution, the above-mentioned edge computing-based cooling tower group energy efficiency collaborative control method is generated into a computer program and stored in a memory for loading and execution by a processor. This allows for the creation of terminal devices based on the memory and processor, making them convenient to use.

[0022] The beneficial effects of this invention are as follows: This invention represents the thermodynamic and hydraulic interactions between cooling towers by constructing a group topology graph containing the thermal plume recirculation index and the hydraulic coupling coefficient. During the feature encoding stage, a graph attention network with embedded priors is used to integrate heat and mass transfer efficiency parameters into the attention calculation process, achieving a fusion of data-driven approaches and physical laws. This allows for the parallel output of collaborative control sequences, future outlet water temperature prediction sequences, and random uncertainty variance. By constructing a combined loss function that integrates node centrality and uncertainty variance, differentiated training is performed on each cooling tower, enabling the model to accurately focus on core nodes of the group and high-confidence predictions. By continuously monitoring the outlet water temperature and regenerating the correction control sequence when the deviation exceeds the confidence boundary, a rigorous closed-loop correction mechanism is constructed, which helps improve the control accuracy in actual operation and enhances the overall energy efficiency optimization level and operational reliability of the cooling tower group. Attached Figure Description

[0023] Figure 1 This is a flowchart of the cooling tower group energy efficiency collaborative control method based on edge computing in this invention; Figure 2This is a comparison chart of the attention scores of cooling tower nodes before and after the addition of heat and mass transfer efficiency bias in this invention. Figure 3 This is a comparison chart of the ablation test performance of the cooling tower group control model in this invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] This invention discloses a method for coordinated energy efficiency control of cooling tower groups based on edge computing, referring to... Figure 1 This includes steps S1-S3: S1. Construct the topology diagram of the cooling tower group.

[0026] In an optional embodiment, real-time operating data and historical performance data of the cooling tower group are obtained, and a topology map of the cooling tower group is constructed based on the real-time operating data and historical performance data. The nodes of the cooling tower group topology map are each cooling tower. The thermal plume recirculation index and hydraulic coupling coefficient are calculated, and the edge features are composed of the weighted sum of the thermal plume recirculation index and the hydraulic coupling coefficient.

[0027] A connection is established with the data acquisition and monitoring control system, and a standard query language is executed to extract the inlet water temperature, outlet water temperature, water flow rate, and fan operating frequency of each cooling tower as real-time operating data and historical performance data for the nodes. Simultaneously, ambient temperature and humidity, real-time wind speed, and real-time wind direction are extracted as auxiliary meteorological data. This auxiliary meteorological data is used to calculate the thermal plume recirculation index, theoretical maximum heat transfer, and heat and mass transfer efficiency parameters. This auxiliary meteorological data is not incorporated into the four-dimensional input time-series feature matrix described later. Each cooling tower is mapped to a node of a directed graph object by instantiating a directed graph object. The thermal plume recirculation index is constructed based on the spatial distance between the cooling towers, the effective angle between the real-time wind direction and the node connection direction, and the sum of the operating frequencies of the two cooling tower fans. The hydraulic coupling coefficient is constructed based on the normalized component of the main pipe pressure drop and the normalized component of the sum of squares of the branch flow rates. The thermal plume recirculation index and the hydraulic coupling coefficient are added using a multiply-add operation according to a preset ratio to generate a tensor containing comprehensive features. This tensor containing comprehensive features is then assigned to the edge attribute field of the data object defined in the PyTorch geometry learning library.

[0028] In an optional embodiment, real-time operating data and historical performance data of the cooling tower group are acquired. Specifically, sensors installed on the inlet and outlet water pipes and fans of each cooling tower are used to collect real-time data on the inlet water temperature, outlet water temperature, water flow rate, and fan operating frequency of each cooling tower. The operating data sequence within a set period in the past is extracted from the historical database of the control system, and the operating data sequence is aligned with the real-time operating data to construct an input time-series feature matrix for network training and inference.

[0029] At the hardware acquisition layer, each cooling tower is equipped with a high-precision industrial-grade sensor network. Inlet and outlet water temperatures are measured using PT100 armored resistance temperature detectors embedded in the pipes, covering a temperature range of 0℃ to 80℃ with an accuracy controlled within ±0.1℃. Water flow rate is monitored non-contactly using electromagnetic flowmeters, with a range designed from 0 m³ / h to 1000 m³ / h and an accuracy error within ±0.5%. The fan operating frequency is synchronously read from the inverter register controlling the fan at a refresh rate of 1Hz via RS485 communication protocol, with a value range between 0Hz and 50Hz. Ambient temperature and humidity, real-time wind speed, and real-time wind direction are collected by a micro-weather station, serving as auxiliary meteorological variables for edge feature and heat and mass transfer efficiency calculations. These auxiliary meteorological variables are not used as node operating condition dimensions in the input time-series feature matrix. The data collection polling cycle is set to once per minute, which helps to detect the hydraulic and thermal changes in the system using real-time operating data.

[0030] At the data processing layer, the control center pulls historical operating data from the historical time series database built into the industrial control system, which has been used for the past 72 hours as a set period, and extracts 4,320 consecutive historical data points with a time step of one minute.

[0031] A timestamp hard alignment and first-order linear interpolation algorithm are used to align historical sequences with current real-time sampled values. After interpolation preprocessing, the four dimensions of operating features of a single cooling tower—inlet water temperature, outlet water temperature, water flow rate, and fan frequency—are concatenated along the time axis to form a three-dimensional dataset of N cooling towers in the cluster. This constructs an input time-series feature matrix with dimensions (N, 4320, 4). This input time-series feature matrix is ​​then used as a structured input tensor to the encoder of the network model for batch parallel training and inference. Dimension 4 corresponds to inlet water temperature, outlet water temperature, water flow rate, and fan operating frequency, respectively, to avoid confusion with auxiliary meteorological variables.

[0032] The network model is a spatiotemporal graph neural network used for cooling tower group control and prediction. The network model structure includes a sequentially connected temporal feature encoding module, a graph attention interaction layer, and a two-branch prediction decoding module. The input to the network model includes an input temporal feature matrix and a cooling tower group topology graph, where the cooling tower group topology graph contains the nodes corresponding to each cooling tower and the edge features between any two cooling tower nodes.

[0033] The network model outputs a cooperative control sequence for each cooling tower and the corresponding random uncertainty variance. To support subsequent correction judgments, the main regression branch of the dual-branch prediction decoding module also simultaneously outputs a future outlet water temperature prediction sequence corresponding to the cooperative control sequence. This future outlet water temperature prediction sequence is used to compare the deviation with the measured outlet water temperature. Specifically, the input time-series feature matrix is ​​fed into the time-series feature encoding module, which extracts time-series patterns through recurrent neural units to generate a time-series hidden feature matrix.

[0034] The temporal hidden feature matrix serves as the feature of each node, and is used in conjunction with the topology of the cooling tower group containing edge features. Figure 1 The input is fed into the graph attention interaction layer, where a query weight matrix and a key weight matrix are generated through linear transformation. The self-attention information is then aggregated in the spatial dimension by combining edge features to generate a node spatial feature representation that integrates multimodal interaction relationships.

[0035] The node spatial feature representation is input to the dual-branch prediction decoding module. The main regression branch of the dual-branch prediction decoding module maps and outputs the collaborative control sequence of future time steps through a fully connected layer, and simultaneously outputs the future effluent temperature prediction sequence. The parallel uncertainty branch outputs the corresponding random uncertainty variance number through a fully connected layer to characterize the system noise.

[0036] The mathematical expression for the network model is as follows: Let the input temporal feature matrix be X, and the parameters of the recurrent neural unit in the temporal feature encoding module be... The input node features at time t are The previous hidden state was The temporal hidden state is then obtained through the recurrent neural unit. And record the last time step or the hidden state after time pooling as .thus, Preserve the temporal dependencies of historical operating conditions, rather than obtaining them solely from a single-layer linear mapping.

[0037] In the graph attention interaction layer, construct the self-attention mechanism: ;

[0038] in, This is the attention coefficient matrix; This is the normalization function; It is in a hidden state; To query the weight mapping matrix; This is the key weight mapping matrix; Value mapping matrix; For query vector dimensions; This is the edge feature adjustment coefficient; The side feature matrix is ​​composed of the thermal plume recirculation index and the hydraulic coupling coefficient. This is used to directly incorporate the thermal and hydraulic coupling strength between cooling towers into the attention score; For input prior bias terms; This represents the spatial features of the nodes.

[0039] In the dual-branch prediction decoding module, the dual-branch prediction decoding module is constructed as follows: ; ;

[0040] in, For coordinated control sequences; The weight matrix for the main regression branch; Representation of node spatial features; The bias term of the main regression branch; For future effluent temperature prediction sequences; The weight matrix for the auxiliary prediction head of effluent temperature; The bias term for the auxiliary prediction head of the outlet water temperature; The variance represents random uncertainty. For activation functions; The weight matrix for the uncertain branch; This is the bias term for the uncertain branch.

[0041] This completes the end-to-end forward prediction computation process.

[0042] In an optional embodiment, a cooling tower group topology map is constructed based on real-time operating data and historical performance data. Specifically, each cooling tower in the cooling tower group is set as a node in the cooling tower group topology map; the thermal plume recirculation index and hydraulic coupling coefficient between any two nodes are calculated; the thermal plume recirculation index and hydraulic coupling coefficient are respectively dimensionlessized, and preset weight constants are assigned to the dimensionless thermal plume recirculation index and hydraulic coupling coefficient respectively; the weighted result of the thermal plume recirculation index and the weighted result of the hydraulic coupling coefficient are summed, and the summed result is used as the edge feature between the corresponding two nodes, thus constructing a cooling tower group topology map containing node and edge features.

[0043] In the initialization and construction phase of the topological graph structure at the bottom layer of the graph neural network, the N cooling towers operating in parallel within the water system are abstractly mapped as N discrete nodes with spatial characteristics in the graph model, and a weighted adjacency matrix structure of dimension N×N is pre-allocated in memory. Through the computing engine, the thermal plume recirculation index, which represents the degree of short-circuit interference of air humidity and heat around the tower group, and the hydraulic coupling coefficient, which represents the degree of flow competition between the supply and return water sharing the main pipe, are calculated respectively.

[0044] The calculated thermal plume recirculation index and hydraulic coupling coefficient are each subjected to independent Min-Max normalization dimensionless operations. After scale unification, based on prior knowledge of whether meteorological intervention or pipeline resistance dominates in the environment, fusion weight constants are assigned, for example, the thermal plume constant is assigned 0.65 and the hydraulic constant is assigned 0.35 respectively. For source node A and target node B, if the normalized thermal plume index and hydraulic coefficient are 0.8 and 0.4 respectively, a weighted summation is performed on the thermal plume index and hydraulic coefficient, and the composite edge feature weight value is calculated to be 0.66. All edge weights representing the intensity of fluid dynamics and thermodynamics interference are filled into the adjacency matrix, and together with the input temporal feature matrix, they are encapsulated into a group topological directed graph containing multi-dimensional connectivity attributes.

[0045] In an optional embodiment, the thermal plume recirculation index is calculated. Specifically, the three-dimensional spatial coordinates of any two cooling towers are obtained, and the spatial Euclidean distance between the two cooling towers is calculated; the real-time wind direction is obtained, and the cosine of the angle between the airflow propagation direction vector and the direction vector of the line connecting the two cooling towers is calculated. The larger value between the cosine of the angle and 0 is taken as the effective wind direction factor; the sum of the operating frequencies of the fans of the two cooling towers is calculated; the ratio of the preset reference distance to the spatial Euclidean distance, the effective wind direction factor, and the ratio of the sum of the fan operating frequencies to the preset reference frequency are multiplied to obtain the thermal plume recirculation index between the two cooling towers.

[0046] By connecting to the three-dimensional Cartesian coordinate system established from the engineering drawings, the geometric center coordinates of the air outlet of the target cooling tower A are extracted. , , The coordinates of the geometric center of the air outlet of the target cooling tower B are ( ) , , The Euclidean distance between target cooling tower A and target cooling tower B in three-dimensional space is obtained by applying the standard three-dimensional Euclidean distance formula. The distance value is between 10m and 30m.

[0047] Real-time absolute wind direction angles with a measurement accuracy of 1° are obtained from a micro-weather station and converted into airflow propagation direction vectors. When the weather station output is the incoming wind direction, the incoming wind direction is reversed to obtain the airflow propagation direction vector. The cosine value of the angle between the three-dimensional airflow propagation direction vector and the spatial direction vector of the line connecting point A to point B is calculated based on the spatial geometric dot multiplication method. When the spatial angle is 60°, the calculated cosine value of the spatial angle is 0.5. For crosswinds or leeward winds that bring negative cosine values, the activation function Max(0, cosine value) is used to truncate the value to 0, and the positive real number output of the directed cyclic action is taken as the wind direction factor.

[0048] The sum of the actual operating frequencies of the two cooling tower fans involved in the calculation is extracted. For example, if the frequency of tower A is 35Hz and the frequency of tower B is 45Hz, the cumulative sum is 80Hz. During the product operation, the ratio of the preset reference distance of 15m to the real-time measured spatial Euclidean distance of 20m is calculated to obtain a ratio of 0.75. The previously obtained effective wind direction factor of 0.5 is extracted. The ratio of the sum of the combined frequencies to the sum of the preset reference rated frequencies when the two cooling towers are running at full load is calculated to be 0.8. The product of the three factors, including the ratio of 0.75, the effective wind direction factor of 0.5, and the ratio of 0.8, yields a calculated value of 0.3. This continuous parameter is used as the thermal plume recirculation index. It should be noted that the thermal plume recirculation index is calculated using an empirical product model. The computational fluid dynamics model can be used as an offline verification tool during the engineering calibration phase, but it is not a necessary step in the online calculation of the thermal plume recirculation index in this embodiment.

[0049] In an optional embodiment, the hydraulic coupling coefficient is calculated. Specifically, the pressure drop of the supply and return water main of the cooling tower group is measured; the real-time water flow rate of the branch where the two target cooling towers are located is obtained, and the square of the water flow rate of each branch is calculated; the ratio of the sum of the squares of the water flow rates of the two target cooling tower branches to the sum of the squares of the preset reference flow rates is calculated to obtain the flow coupling component; the ratio of the pressure drop of the supply and return water main to the preset reference pressure drop is calculated to obtain the pressure drop coupling component; the flow coupling component and the pressure drop coupling component are weighted and summed to obtain the hydraulic coupling coefficient between the two cooling towers.

[0050] Differential pressure transmitters pre-installed at the main circulation return water pipe and the inlet water collection pipe of the water system are used to obtain the real-time pressure difference of the cooling water circuit as the pressure drop value of the main pipe, for example, the obtained measurement data is a stable 120 kPa. Simultaneously, ultrasonic flow sensors deployed at the end of the branch network are used to obtain the instantaneous water flow data on the branch pipes where cooling tower A and cooling tower B are located, and convert them into m³ / h units; following the principle that the resistance is proportional to the square of the flow velocity in pipe network fluid mechanics, the two water flow values ​​are squared.

[0051] For example, the flow sensor reports an instantaneous branch flow rate of 150 m³ / h for tower A, which squares to 22500; the corresponding flow rate for tower B is 200 m³ / h, which squares to 40000. The sum of the squares of the two flow rates is 62500. If the preset baseline flow rate sum of squares is 50000, the calculated flow coupling component is 1.25; if the preset baseline pressure drop is 100 kPa, the calculated pressure drop coupling component is 1.2. The sum of the squares of the two flow rates is 62500. If the preset baseline flow rate sum of squares is 50000, the calculated flow coupling component is 62500 / 50000 = 1.25; if the preset baseline pressure drop is 100 kPa, the calculated pressure drop coupling component is 120 / 100 = 1.2. With the flow coupling component weight set to 0.6 and the pressure drop coupling component weight set to 0.4, the hydraulic coupling coefficient is 0.6 × 1.25 + 0.4 × 1.20 = 1.23. Therefore, the larger the branch flow or the greater the main pipe pressure drop, the larger the hydraulic coupling coefficient, which characterizes the increased competition for flow between the supply and return water systems connected in parallel through the shared main pipe.

[0052] S2. Predict control sequences using network models.

[0053] In an optional embodiment, a graph attention network with embedded priors is used to encode the topology of the cooling tower group. In the self-attention calculation, the dot product of the query vector and the key vector is adjusted by a prior bias function and then normalized by a Softmax function. The prior bias function is set based on the heat and mass transfer efficiency parameters of each cooling tower. Based on the encoded group features, the decoder predicts in parallel the cooperative control sequence of each cooling tower, the future outlet water temperature prediction sequence, and the random uncertainty variance.

[0054] A graph attention network model is constructed, and three fully connected Linear layers are initialized to extract the query vector, key vector, and value vector of each node. The inner product matrix of the query vector and key vector is calculated, and the edge feature matrix obtained by weighting the thermal plume recirculation index and the hydraulic coupling coefficient is used as the edge feature bias and incorporated into the inner product matrix. At the same time, the ratio of the actual heat transfer to the theoretical maximum heat transfer of each cooling tower is calculated from real-time operating data, and the natural logarithm of this ratio is used as the prior bias term for heat and mass transfer efficiency.

[0055] The summation result is normalized along the adjacent node dimension using the Softmax function and multiplied with the value vector to obtain the feature encoding sequence. This feature encoding sequence is then input into a dual-branch prediction decoding module consisting of a multilayer perceptron and a fully connected mapping layer. A linear mapping layer is set in parallel at the end of the decoder to output a cooperative control sequence containing the fan frequency and the water pump valve opening, a future outlet water temperature prediction sequence corresponding to the cooperative control sequence, and a random uncertainty variance prediction value representing the Gaussian distribution variance.

[0056] In an optional embodiment, during self-attention computation, the dot product of the query vector and key vector is adjusted by a prior bias function and then normalized using the Softmax function. Specifically, the input features of each node are multiplied by the query weight matrix and the key weight matrix, respectively, to obtain the corresponding query vector and key vector; the inner product of the query vector and key vector of any two nodes is calculated to obtain the basic attention score; the ratio of the actual heat exchange to the theoretical maximum heat exchange of each cooling tower is extracted from the real-time operating data as the heat and mass transfer efficiency parameter; the logarithm of the heat and mass transfer efficiency parameter of the corresponding target cooling tower node or the aggregated adjacent cooling tower node is used as a bias term and added to the basic attention score pointing from the source cooling tower node to the target cooling tower node to obtain the adjusted attention score; the adjusted attention score is normalized using the Softmax function, and the attention weight matrix is ​​output.

[0057] In the forward propagation layer of the graph attention network, the network configures and maintains two sets of linear transformation weight matrices that are optimized with batch updates. When a 64-dimensional temporal representation of node feature sequences is processed, a query weight matrix with a dimension specification of [64×128] is applied. and the key weight matrix of the same dimension During the dot product operation, a 128-dimensional query feature vector and a 128-dimensional key feature vector are generated.

[0058] Based on this, a basic attention score measuring sequence density is calculated by performing an inner product operation on the paired query vectors and key vectors formed by graph nodes. Subsequently, the composite edge feature weights between corresponding node pairs are multiplied by the edge feature adjustment coefficient and then added to the basic attention score, so that thermal plume recirculation and hydraulic coupling strength participate in adjusting the neighborhood information aggregation weights.

[0059] Simultaneously, the control system extracts the inlet and outlet water temperature differences captured by each set of sensors and calculates the actual heat exchange based on the medium flow rate, comparing it with the theoretical maximum heat exchange established by the limiting wet-bulb temperature. For example, if the actual heat exchange of cooling tower A is determined to be 800kW and the theoretical maximum heat exchange is 1000kW, then its parameter ratio is set to 0.8. The natural logarithm of the sum of the network calculation ratio and the smoothing term is used as the heat and mass transfer efficiency bias term corresponding to the aggregated adjacent cooling tower nodes. For the attention score of the source cooling tower node pointing to the target cooling tower node, the heat and mass transfer efficiency bias term of the target cooling tower node is superimposed on the basic attention score for correction. The Softmax function is applied along the adjacent node dimension for normalization processing to ensure that the contribution coefficients of different adjacent nodes to the generation of a single source node are summed to 1, so that adjacent nodes with better energy efficiency performance receive higher weights in information aggregation, resulting in an attention weight allocation matrix.

[0060] Reference Figure 2 After adjusting by introducing a prior bias function, the adjusted scores of different nodes showed differentiated increases and decreases compared to the base scores. The attention scores of some nodes were significantly improved, while the attention scores of others were suppressed. The data trend reflects that the mechanism of incorporating the heat and mass transfer efficiency parameter as a bias term into the self-attention calculation can intervene in the feature encoding process, enabling the network to give greater weight to nodes with high energy efficiency when aggregating neighborhood information. This proves that the feature encoding mechanism can guide the model to focus on high energy efficiency equipment.

[0061] S3. Build the model and generate control sequences.

[0062] In an optional embodiment, a combined loss function training model is constructed, and a loss component is set for each cooling tower. The loss component includes a weighted base loss and a logarithmic penalty term for the random uncertainty variance. The weight of the base loss is proportional to the centrality of the cooling tower in the cooling tower group topology graph and inversely proportional to the random uncertainty variance. The outlet water temperature of each cooling tower is continuously monitored. When the deviation between the measured outlet water temperature and the corresponding predicted outlet water temperature exceeds the confidence boundary set by the random uncertainty variance, the latest data including the deviation is used as input to regenerate the correction control sequence and execute it.

[0063] Before calculating the loss, the cooperative control sequence and the future outlet water temperature prediction sequence are normalized, or errors of different dimensions are converted into dimensionless loss terms using preset weighting coefficients. The centrality score of each cooling tower node in the topology diagram is measured. The mean square error (MSELoss) algorithm is used to calculate the control loss between the predicted value of the control sequence and the corresponding optimized control label, and the temperature prediction loss between the future outlet water temperature prediction sequence and the measured outlet water temperature label at the corresponding time. The control loss and temperature prediction loss are then weighted and summed according to preset weights to obtain the basic loss. The optimized control labels are selected from samples with low energy consumption per unit cooling capacity in historical operating data, or obtained offline from a preset energy consumption constraint optimization model. The basic loss is multiplied by the centrality score and divided by the sum of the variance of random uncertainty and the smoothing term to obtain the weighted basic loss. The logarithmic value of the sum of the random uncertainty variance and the smoothing term is then calculated using the log function. This logarithmic value is used as the log-regularization term. The weighted basic loss is added to the log-regularization term to construct the loss components for each cooling tower node. The Adam optimizer is then used to perform backpropagation on the combined loss function to update the model parameters. The random uncertainty variance is used to characterize the degree of uncertainty in the prediction error of future effluent temperature.

[0064] Establish a monitoring service to continuously read the measured outlet water temperature values ​​from the programmable logic controller (PLC). Calculate the absolute deviation between the measured outlet water temperature value and the predicted future outlet water temperature value output by the model. If the square of this absolute deviation is greater than 9 times the variance of the random uncertainty, the deviation is considered a positive result. If this happens, a recalculation signal is triggered. The latest state characteristics are transmitted to the decoder via the OPC unified architecture protocol, the corrected fan and water valve control sequences are inferred through forward propagation, and the corrected fan and water valve control sequences are then sent down to the underlying controller for execution.

[0065] In an optional embodiment, a combined loss function training model is constructed, including: calculating the feature vector centrality of each cooling tower node based on the constructed cooling tower group topology graph and edge features; obtaining the collaborative control sequence of the decoder output for each cooling tower and the corresponding random uncertainty variance; calculating the basic loss value between the predicted output of each cooling tower and the true label; multiplying the basic loss value of each cooling tower by the corresponding feature vector centrality, dividing the product by the corresponding random uncertainty variance, adding a smoothing term to prevent the denominator from being zero, and then adding it to the logarithmic penalty term of the random uncertainty variance to obtain the loss component of each cooling tower; summing the loss components of all cooling towers to obtain the total combined loss function used for model backpropagation optimization.

[0066] The eigenvector centrality is calculated based on the topological graph with multimodal edge characteristics. The principal eigenvalues ​​of the N×N adjacency matrix are solved using the matrix eigenvalue decomposition operator, and the corresponding normalized eigenvector components are derived. The numerical range of these normalized eigenvector components is mapped to [0, 1] as a weight index representing the hydraulic and thermal status level of the center; for example, a score of 0.8 is obtained for a cooling tower. During the end-to-end model training phase, the control action sequence for a specified future T steps, the predicted future outlet water temperature sequence, and the random uncertainty variance of the uncertain branch output are obtained from the model module.

[0067] When calculating the loss value for a single cooling tower device, the error between the model's predicted result and the baseline measured response is calculated using the mean square error algorithm as the base loss value. The base loss value (0.1) is multiplied by the eigenvector centrality (0.8) to obtain a weighted value. This weighted value is then divided by the sum of the random uncertainty variance (0.04) and a smoothing factor set to 1e-6. A logarithmic penalty term for the sum of the random uncertainty variance and the smoothing factor is then added to obtain the corresponding loss component. This strategy enables noisy predictions to have the ability to smoothly optimize the path. In each cycle, the loss components of all cooling tower nodes are summed to generate the overall combined loss function used for iterative optimization of the driving parameters.

[0068] This ablation experiment was conducted based on three months of historical operating data from six parallel cooling towers in a large data center, with a data sampling step of one minute. The experimental model was configured to input the historical operating conditions and topology graph into the network to predict the cooperative control sequence. The evaluation metrics for the model performance were the root mean square error (RMSE), mean absolute error (MAE), and the overall system energy efficiency ratio after implementing the control strategy. Three comparative networks were set up in the experiment. The basic network adopted a spatiotemporal graph neural network with conventional edge features and conventional attention, and the loss function was the traditional MSE. Variant Network 1 removed the heat and mass transfer efficiency bias term in the forward propagation of the graph attention network based on the complete scheme. Variant Network 2 removed the combined loss function combining eigenvector centrality and random uncertainty variance based on the complete scheme.

[0069] Reference Figure 3 Compared with the basic network, variant network 1, and variant network 2, the complete scheme has the lowest values ​​for both root mean square error and mean absolute error, while achieving the highest value for the overall system energy efficiency ratio. The data trend reflects the mechanism of integrating the heat and mass transfer efficiency bias term, eigenvector centrality, and random uncertainty variance into the model construction and loss function calculation. This mechanism can effectively reduce the prediction error of the cooperative control sequence and significantly improve the overall system energy efficiency, proving that the combined loss function training model and feature encoding optimization strategy can improve the control accuracy and overall system energy efficiency in actual operation.

[0070] The basic network's predictive control sequence has a root mean square error (RMSE) of 0.85 and a mean absolute error (MAO) of 0.62, resulting in a system overall energy efficiency ratio (EER) of 4.2. Variant network 1 reduces the RMS to 0.68 and the MAO to 0.51, improving the EER to 4.6. Variant network 2 achieves an RMS of 0.72 and an MAO of 0.54, reaching an EER of 4.5. The complete scheme employing all innovative modules reduces the RMS to 0.52 and the MAO to 0.4, while achieving the optimal EER of 5.1.

[0071] During the attention calculation phase, the bias of the ratio of actual heat exchange to theoretical heat exchange enables the network model to lock in and assign greater weight to nodes with high energy efficiency when aggregating neighborhood information, thereby reducing prediction bias under operating conditions. Simultaneously, the edge features formed by the thermal plume recirculation index and the hydraulic coupling coefficient are incorporated into the attention score, allowing the network aggregation process to directly reflect the thermal and hydraulic coupling strength between cooling towers. The combined loss function, which integrates topological centrality and prediction random uncertainty variance, not only makes the network optimization process sensitive to high-position thermal nodes but also assesses and isolates inherent system noise interference, ensuring the stability of end-to-end forward prediction and generating a coordinated control sequence for the cooling tower group, which helps improve the overall system energy efficiency.

[0072] This invention also discloses an edge computing-based cooling tower group energy efficiency collaborative control system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the edge computing-based cooling tower group energy efficiency collaborative control method according to the present invention is implemented.

[0073] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0074] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for coordinated energy efficiency control of cooling tower groups based on edge computing, characterized in that, include: S1. Obtain real-time operating data and historical performance data of the cooling tower group. Construct a topology map of the cooling tower group based on the real-time operating data and historical performance data. The nodes of the topology map are each cooling tower. Calculate the heat plume recirculation index and the hydraulic coupling coefficient. The edge features are weighted by the heat plume recirculation index and the hydraulic coupling coefficient. S2. Encode the features of the cooling tower group topology map using a graph attention network with embedded priors. In the self-attention calculation, the dot product of the query vector and the key vector is adjusted by a prior bias function and then normalized using Softmax. The prior bias function is set based on the heat and mass transfer efficiency parameters of each cooling tower. Based on the encoded group features... The system first predicts the coordinated control sequence, future outlet water temperature prediction sequence, and random uncertainty variance of each cooling tower in parallel using a decoder. Then, it constructs a combined loss function training model, setting a loss component for each cooling tower. This loss component includes a weighted base loss and a logarithmic penalty term for the random uncertainty variance. The weight of the base loss is proportional to the centrality of the node in the cooling tower group topology and inversely proportional to the random uncertainty variance. The system continuously monitors the outlet water temperature of each cooling tower. When the deviation between the measured and predicted outlet water temperature exceeds the confidence boundary set for the random uncertainty variance, it regenerates and executes a correction control sequence using the latest data containing the deviation as input.

2. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 1, characterized in that, The acquisition of real-time operating condition data and historical performance data of the cooling tower group includes: real-time collection of inlet water temperature, outlet water temperature, water flow rate and fan operating frequency of each cooling tower through sensors installed on the inlet and outlet water pipes and fans of each cooling tower; extraction of operating condition data sequence within a set period from the historical database of the control system, and alignment of the operating condition data sequence with the real-time operating condition data to construct an input time series feature matrix for network training and inference.

3. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 1, characterized in that, The step of constructing a cooling tower group topology map based on real-time operating data and historical performance data includes: setting each cooling tower in the cooling tower group as a node in the cooling tower group topology map; calculating the thermal plume recirculation index and hydraulic coupling coefficient between any two nodes; performing dimensionless processing on the thermal plume recirculation index and hydraulic coupling coefficient respectively; assigning preset weight constants to the dimensionless thermal plume recirculation index and hydraulic coupling coefficient respectively; summing the weighted result of the thermal plume recirculation index and the weighted result of the hydraulic coupling coefficient; and using the summed result as the edge feature between the corresponding two nodes.

4. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 3, characterized in that, The calculation of the thermal plume recirculation index includes: obtaining the three-dimensional spatial coordinates of any two cooling towers and calculating the spatial Euclidean distance between the two cooling towers; obtaining the real-time wind direction and calculating the cosine of the angle between the airflow propagation direction vector and the direction vector of the line connecting the two cooling towers, and taking the larger value between the cosine of the angle and 0 as the effective wind direction factor; calculating the sum of the operating frequencies of the fans of the two cooling towers; and multiplying the ratio of the preset reference distance to the spatial Euclidean distance, the effective wind direction factor, and the sum of the fan operating frequencies to the preset reference frequency to obtain the thermal plume recirculation index between the two cooling towers.

5. The edge computing-based energy efficiency collaborative control method for cooling tower groups according to claim 1 or 3, characterized in that, The calculation of the hydraulic coupling coefficient includes: measuring the pressure drop of the supply and return water main of the cooling tower group; obtaining the real-time water flow of the branches where the two target cooling towers are located, and calculating the square of the water flow of each branch; calculating the ratio of the sum of the squares of the water flow of the branches of the two target cooling towers to the sum of the squares of the preset reference flow to obtain the flow coupling component; calculating the ratio of the pressure drop of the supply and return water main to the preset reference pressure drop to obtain the pressure drop coupling component; and performing a weighted summation of the flow coupling component and the pressure drop coupling component to obtain the hydraulic coupling coefficient between the two cooling towers.

6. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 1, characterized in that, In the self-attention calculation, the dot product of the query vector and the key vector is adjusted by a prior bias function and then normalized using Softmax. This includes: multiplying the input features of each node by the query weight matrix and the key weight matrix respectively to obtain the corresponding query vector and key vector; calculating the inner product of the query vector and key vector of any two nodes to obtain the basic attention score; extracting the ratio of the actual heat exchange to the theoretical maximum heat exchange of each cooling tower from the real-time operating data as the heat transfer and mass transfer efficiency parameter; adding the logarithm of the heat transfer and mass transfer efficiency parameter of the corresponding target cooling tower node as a bias term to the basic attention score pointing from the source cooling tower node to the target cooling tower node to obtain the adjusted attention score; and applying the Softmax function to normalize the adjusted attention score.

7. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 1, characterized in that, The construction of the combined loss function training model includes: calculating the feature vector centrality of each cooling tower node based on the constructed cooling tower group topology graph and edge features; obtaining the collaborative control sequence of the decoder output for each cooling tower and the corresponding random uncertainty variance; calculating the basic loss value between the predicted output of each cooling tower and the true label; multiplying the basic loss value of each cooling tower by the corresponding feature vector centrality, dividing the product by the corresponding random uncertainty variance, adding the sum of the smoothing terms to prevent the denominator from being zero, and then adding it to the logarithmic penalty term of the random uncertainty variance to obtain the loss components of each cooling tower; summing the loss components of all cooling towers to obtain the total combined loss function.

8. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 2, characterized in that, Before using a graph attention network with embedded priors to encode the features of the cooling tower group topology map, the process includes: inputting the input temporal feature matrix into the temporal feature encoding module, extracting temporal patterns through recurrent neural units to generate a temporal hidden feature matrix; the graph attention network includes a graph attention interaction layer, which uses the temporal hidden feature matrix as the feature of each node, and inputs it together with the cooling tower group topology map containing edge features into the graph attention interaction layer to perform self-attention information aggregation in the spatial dimension and generate a node spatial feature representation; the decoder is a dual-branch prediction decoding module.

9. The energy efficiency collaborative control method for cooling tower groups based on edge computing according to claim 1, characterized in that, When the deviation between the measured and predicted outlet water temperature exceeds the confidence boundary set by the random uncertainty variance, the corrected control sequence is regenerated and executed using the latest data containing the deviation as input. This includes: calculating the absolute deviation between the measured outlet water temperature and the predicted future outlet water temperature; triggering a recalculation signal when the square of the absolute deviation is greater than 9 times the random uncertainty variance; and transmitting the latest state features to the decoder through a unified architecture protocol to infer the corrected fan and water valve control sequence through forward propagation.

10. A cooling tower group energy efficiency collaborative control system based on edge computing, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the edge computing-based energy efficiency collaborative control method for cooling tower groups according to any one of claims 1-9.

Images