Fault-tolerant water-cooled refrigeration station system optimization control method based on large model

Abstract

The application relates to the technical field of industrial control and discloses a fault-tolerant water-cooled refrigeration station system optimization control method based on a large model, which comprises the following steps: firstly, a pipe network hydraulic resistance topology model is established, all single-pipe failure working conditions are traversed to calculate equivalent hydraulic impedance from a cold source to an end, and impedance values under the most unfavorable working condition are extracted to construct a fault-tolerant hydraulic bottleneck parameter set; subsequently, the parameter set is mapped into a whole-network topology constraint vector, and combined with real-time flow data to be input into a pre-trained time-series attention neural network model; the model uses an attention mechanism to fuse the topology constraint into time-series deduction, and directly predicts a supply and return water pressure difference set value meeting a potential fault safety boundary. The method effectively solves the problem that a system is paralyzed due to insufficient safety margin when a pipe network suddenly fails in a traditional control strategy, and guarantees the continuity of key load cooling and the economy of system operation.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and more specifically, to an optimized control method for fault-tolerant water-cooled refrigeration station systems based on large models. Background Technology

[0002] In modern large public buildings and industrial parks, district cooling systems typically employ complex hydraulic configurations involving multi-source ring networks, multi-stage pressure zones, and differential pressure control. These systems utilize multiple chiller units working in conjunction with a ring-shaped distribution network, employing variable differential pressure control strategies to regulate chilled water flow to adapt to dynamic changes in terminal loads. Traditional control logic generally relies on real-time feedback regulation, dynamically adjusting the supply and return water differential pressure setpoints of the chiller station's manifolds by monitoring the real-time status of the most unfavorable thermal loop in the network (the area where terminal differential pressure is most difficult to meet). Under ideal stable operating conditions, this approach can meet terminal cooling demands while minimizing pump energy consumption, making it the mainstream method in the current HVAC automatic control field.

[0003] However, with increasing demands for system reliability, especially in scenarios like data centers where cooling continuity is paramount, traditional control methods have gradually revealed serious safety vulnerabilities. Existing technologies, when calculating the optimal differential pressure setpoint, often optimize based solely on the hydraulic characteristics of the pipe network under the current moment and normal connection conditions. Most rule-based or data-driven optimization algorithms (such as conventional neural networks and reinforcement learning) are limited to historical data and the current topology, tending to maximize differential pressure under normal operating conditions to achieve energy savings. However, complex ring networks contain numerous bypass or backup pipe sections with extremely low or even zero flow rates during normal operation. While seemingly insignificant under normal conditions, these sections are the only lifeline for maintaining system connectivity during sudden network failures (such as critical pipe ruptures or valve jamming).

[0004] Because existing control models cannot detect minimum cut sets or potential bottlenecks in a graph theory sense, the calculated differential pressure setpoints are often close to the lower limit of the current operating conditions. Once a single point of failure occurs in the system (N-1 condition), the hydraulic flow direction is forced to be reconfigured, and the resistance to flow through the backup path will surge instantaneously. At this time, if the system still maintains the original low differential pressure setpoint, it will be unable to overcome the sudden increase in pipeline impedance, leading to instantaneous loss of cooling at critical terminals and causing serious safety accidents. Summary of the Invention

[0005] This invention provides an optimized control method for a fault-tolerant water-cooled refrigeration station system based on a large model, which solves the technical problems mentioned in the background art.

[0006] This invention provides an optimized control method for a fault-tolerant water-cooled chiller system based on a large model, comprising:

[0007] Establish a topological model of hydraulic resistance of the pipeline network that maps the connection status of the water-cooled refrigeration station and the transmission and distribution pipeline network, and determine the linearized resistance coefficient of each pipeline segment under operating conditions.

[0008] Based on the aforementioned hydraulic resistance topology model of the pipeline network, the single-point failure conditions of each pipe section in the transmission and distribution pipeline network are traversed, and the equivalent hydraulic impedance from the cold source side to the terminal critical load side is calculated. The maximum equivalent hydraulic impedance of each critical load side under all the aforementioned single-point failure conditions is extracted as the most unfavorable failure condition impedance value, and a fault-tolerant hydraulic bottleneck parameter set is constructed based on the most unfavorable failure condition impedance values ​​of all the critical load sides.

[0009] The system collects real-time flow data, calculates the fault protection differential pressure benchmark index using the fault-tolerant hydraulic bottleneck parameter set, generates an input sequence containing time-series operating condition characteristics, and maps the fault-tolerant hydraulic bottleneck parameter set into a network-wide topology constraint feature vector.

[0010] The input sequence and the network topology constraint feature vector are input into a pre-trained temporal attention neural network model. The attention mechanism module in the temporal attention neural network model is used to fuse the network topology constraint feature vector into the time step of the input sequence. The optimized set value of the supply and return water pressure difference is output after forward calculation.

[0011] The optimized setpoint for the supply and return water pressure difference is predicted by the temporal attention neural network model under the hydraulic safety boundary conditions defined by the fault-tolerant hydraulic bottleneck parameter set.

[0012] The beneficial effects of this invention are as follows: by introducing hydraulic skeleton parameters that contain information on potential faults across the entire network, the control model is endowed with the ability to predict and defend against unknown risks; the optimized supply and return water pressure difference setpoint not only remains in the energy-saving range during normal operation, but also meets the hydraulic transport requirements under the worst operating conditions in the event of any single-point fault, without relying on delayed fault feedback adjustment to maintain system stability, thereby avoiding critical load cooling loss due to lack of safety margin, and achieving a balance between extreme energy efficiency and high reliability fault tolerance. Attached Figure Description

[0013] Figure 1 This is a flowchart of the fault-tolerant water-cooled refrigeration station system optimization control method based on a large model according to the present invention;

[0014] Figure 2 This is the overall system architecture and closed-loop control logic diagram of the refrigeration station and distribution pipeline network of the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0016] like Figure 1 As shown, the fault-tolerant water-cooled chiller system optimization control method based on a large model includes:

[0017] Establish a topological model of hydraulic resistance of the pipeline network that maps the connection status of the water-cooled refrigeration station and the transmission and distribution pipeline network, and determine the linearized resistance coefficient of each pipeline segment under operating conditions.

[0018] Based on the aforementioned hydraulic resistance topology model of the pipeline network, the single-point failure conditions of each pipe section in the transmission and distribution pipeline network are traversed, and the equivalent hydraulic impedance from the cold source side to the terminal critical load side is calculated. The maximum equivalent hydraulic impedance of each critical load side under all the aforementioned single-point failure conditions is extracted as the most unfavorable failure condition impedance value, and a fault-tolerant hydraulic bottleneck parameter set is constructed based on the most unfavorable failure condition impedance values ​​of all the critical load sides.

[0019] The system collects real-time flow data, calculates the fault protection differential pressure benchmark index using the fault-tolerant hydraulic bottleneck parameter set, generates an input sequence containing time-series operating condition characteristics, and maps the fault-tolerant hydraulic bottleneck parameter set into a network-wide topology constraint feature vector.

[0020] The input sequence and the network topology constraint feature vector are input into a pre-trained temporal attention neural network model. The attention mechanism module in the temporal attention neural network model is used to fuse the network topology constraint feature vector into the time step of the input sequence. The optimized set value of the supply and return water pressure difference is output after forward calculation.

[0021] The optimized setpoint for the supply and return water pressure difference is predicted by the temporal attention neural network model under the hydraulic safety boundary conditions defined by the fault-tolerant hydraulic bottleneck parameter set.

[0022] In a preferred embodiment, establishing a pipeline hydraulic resistance topology model that maps the connection state of the water-cooled chiller station and the distribution network includes: abstracting the water-cooled chiller station and the distribution network into a graph model. ;in, The set of nodes includes the set of cold source-side nodes representing the main chiller unit of the refrigeration plant. The set of key load-side nodes representing end users And the set of intermediate nodes representing the branch points of the pipeline network; The set of edges includes fluid connection channels representing water supply and distribution pipes, regulating valves, and heat exchange devices; an association matrix describing the connection relationships between nodes and edges is constructed based on the connection relationships. .

[0023] Since the physical connections of complex water-cooled refrigeration plants and distribution networks directly determine the hydraulic transmission paths, it is impossible to systematically analyze the impact of faults without structural abstraction. Therefore, the refrigeration plant main unit, critical terminal loads, network branch points, and diameter change points are first categorized into node sets, and fluid channels such as water distribution pipes, regulating valves, and heat exchangers are categorized into edge sets, forming a graph model that accurately reflects the connection relationships. To make this graph model computable, an association matrix is ​​further constructed. The connection direction between each edge and each node is clearly defined through the values ​​of the matrix elements, transforming the network topology into a quantifiable computational carrier.

[0024] In a preferred embodiment, determining the linearized resistance coefficient of each pipe segment under operating conditions includes: for the set of edges... Each pipe section The nonlinear hydraulic characteristics under the current operating conditions are linearized using the following formula to determine the linearized drag coefficient:

[0025] ;

[0026] in, Indicates pipe section The pressure difference between the supply and return water at both ends, Indicates the pipe section through which the flow passes Traffic, Indicates pipe section The linearized drag coefficient at the current operating point; calculate the linearized drag coefficient. The reciprocal of the value is used to obtain the side admittance weight of the pipe segment. :

[0027] ;

[0028] And construct a diagonal matrix of edge weights using the edge admittance weights of all pipe segments. .

[0029] The relationship between pressure drop and flow rate in a pipe segment affects hydraulic transmission, but this relationship is nonlinear. Directly using it for network-wide fault traversal calculations would result in extremely high complexity and difficulty in convergence. Therefore, a linear approximation is needed at the operating point of the current operating condition to obtain a linearized resistance coefficient that characterizes the proportional relationship between pressure difference and flow rate at both ends of the pipe segment. Considering that subsequent analysis of network connectivity will be conducted through matrix operations, the reciprocal of the linearized resistance coefficient (side admittance weight) more directly reflects the flow capacity of the pipe segment. Based on the side admittance weights of all pipe segments, a side-weighted diagonal matrix is ​​constructed, which integrates the hydraulic characteristics of each pipe segment into a unified matrix form, facilitating global hydraulic analysis in conjunction with the correlation matrix.

[0030] In a preferred embodiment, based on the network hydraulic resistance topology model, the equivalent hydraulic impedance from the cold source side to the terminal critical load side is calculated by traversing the single-point failure conditions of each pipe segment in the transmission and distribution network, including: using the correlation matrix And edge weight diagonal matrix Construct the weighted Laplace matrix of the pipeline network. :

[0031] ;

[0032] For edge set Each pipe section Construct a single-point-of-failure topology after removing the pipe segment. And calculate the corresponding weighted Laplacian matrix. and its Moore-Penrose pseudoinverse matrix The following formula is used to calculate the single-point fault topology. Next, from the source node To critical load nodes The equivalent hydraulic impedance :

[0033] ;

[0034] in, and These respectively represent the source node and critical load nodes Standard basis vectors This indicates the matrix transpose.

[0035] To comprehensively cover single-point failure scenarios, it is necessary to simulate the network state after the failure of each pipe segment. Therefore, a single-point failure topology is constructed for each pipe segment after its removal. The weighted Laplace matrix can integrate the topological structure reflected by the correlation matrix and the flow capacity reflected by the edge-weighted diagonal matrix, thereby quantifying the connectivity characteristics of the entire network. The weighted Laplace matrix corresponding to the failure topology will exhibit singular characteristics due to the removal of the pipe segment. The Moore-Penrose pseudo-inverse matrix can effectively solve this problem. Equivalent hydraulic impedance is the core quantitative indicator of the difficulty of hydraulic transmission from the cold source side to the critical load side at the end. The standard basis vector can accurately locate the positions of the nodes on the cold source side and the critical load side. By combining the quadratic form of the difference between the two with the pseudo-inverse matrix, the hydraulic resistance between the two points under the failure condition can be accurately calculated.

[0036] In a preferred embodiment, the maximum equivalent hydraulic impedance of each critical load side under all single-point failure conditions is extracted as the most unfavorable failure condition impedance value, and a fault-tolerant hydraulic bottleneck parameter set is constructed based on the most unfavorable failure condition impedance values ​​of all critical load sides, including: for each critical load node Traverse all possible faulty pipe sections The impedance value of the node under the most unfavorable fault condition is calculated using the following formula. :

[0037] ;

[0038] in, For the set of nodes on the cold source side, For the pipe section Single point of failure from the source node To critical load nodes The equivalent hydraulic impedance; all critical load nodes Calculated Combine and construct the fault-tolerant hydraulic bottleneck parameter set. :

[0039] ;

[0040] in, This is the set of critical load-side nodes.

[0041] When a system has multiple cooling source-side nodes, after a single point of failure, the terminal critical load-side nodes will automatically select the cooling source with the least resistance to obtain cooling capacity. Therefore, it is necessary to select the minimum equivalent hydraulic impedance from each cooling source to the target load as the effective supply impedance under each failure condition to ensure that the optimal cooling path resistance is reflected during the failure. Based on this, for each terminal critical load-side node, all single-point failure conditions are traversed, and the maximum value of the corresponding equivalent hydraulic impedance or effective supply impedance is selected as the most unfavorable failure condition impedance value. This value directly corresponds to the maximum hydraulic resistance when the node encounters the worst failure. The most unfavorable failure condition impedance values ​​of all terminal critical load-side nodes are combined to form a fault-tolerant hydraulic bottleneck parameter set, realizing a centralized characterization of the fault resistance bottleneck of the entire network's critical loads.

[0042] In a preferred embodiment, real-time flow data of the system is collected, and the fault-tolerant hydraulic bottleneck parameter set is used to calculate the fault protection differential pressure benchmark index, including: for each critical load-side node Calculate its current time using the following formula. node-level structural pressure difference requirements :

[0043] ;

[0044] in, The fault-tolerant hydraulic bottleneck parameter set corresponds to the node. The impedance value under the most unfavorable fault condition. For nodes At any moment Real-time traffic data; calculate the current time using the following formula. System-level structural pressure envelope value :

[0045] ;

[0046] in, For the set of critical load-side nodes, This indicates the operation of finding the maximum value; by and Together, they constitute the fault protection differential pressure benchmark index.

[0047] The current moment serves as the time base for dynamic time-series adjustments, ensuring that differential pressure demand matches real-time operating conditions. Real-time flow data at each node reflects the current cooling demand of critical loads. By multiplying this value by the impedance value corresponding to the most unfavorable fault condition in the fault-tolerant hydraulic bottleneck parameter set, the node-level structural differential pressure demand is obtained. This value quantifies the minimum differential pressure required for a single node to maintain cooling supply when encountering the most unfavorable fault under the current load. Since the system must simultaneously guarantee the fault tolerance capability of all critical loads, the maximum value of the current node-level structural differential pressure demand is selected as the system-level structural differential pressure envelope value, clearly defining the overall system differential pressure baseline. The vector formed by all node-level structural differential pressure demands is combined with the system-level structural differential pressure envelope value to form a fault-tolerant differential pressure benchmark index. This preserves the individual differential pressure demands of each node while establishing a unified safety boundary for the system, providing a clear constraint for the model to predict the supply and return water differential pressure setpoints.

[0048] In a preferred embodiment, generating an input sequence containing time-series operating condition characteristics and mapping the fault-tolerant hydraulic bottleneck parameter set to a network-wide topology constraint feature vector includes: constructing the time sequence using the following formula. Structural Enhancement Feature Vector :

[0049] ;

[0050] in, For real-time traffic data vectors, This is the water pump operating frequency data. External environmental parameters; utilizing a preset time window length. Construct the input sequence containing time-series operating condition features. :

[0051] ;

[0052] The fault-tolerant hydraulic bottleneck parameter set is obtained using the following formula. Mapped to the network-wide topology constraint feature vector :

[0053] ;

[0054] in, and These are the learnable weight matrix and bias vector, respectively. It is a non-linear activation function.

[0055] Real-time flow data vectors reflect the current cooling demand distribution of all critical loads, pump operating frequency data characterizes the current pump distribution capacity, and external environmental parameters relate to external factors causing end-load fluctuations. These three, combined with the system-level structural differential pressure envelope and node-level structural differential pressure demand vectors in the fault-tolerant differential pressure benchmark, form a structural enhancement feature vector. This comprehensively integrates dynamic operating information and fault safety constraints under the current operating conditions. Considering the temporal inertia of system operation, structural enhancement feature vectors containing the current moment and historical moments of a preset length are selected and arranged chronologically to form a temporal operating condition feature input sequence, helping the model capture the evolution of operating conditions. The fault-tolerant hydraulic bottleneck parameter set, as the core of static topology constraints, undergoes feature extraction and dimensional adjustment through a linear transformation layer with learnable weights. A complex mapping relationship is then introduced through a nonlinear activation function to generate a network-wide topology constraint feature vector that adapts to the dimensions of the hidden layer of the temporal attention neural network model. This allows the model to simultaneously utilize temporal dynamic information and topology safety constraints to predict differential pressure setpoints.

[0056] In a preferred embodiment, the pre-training process of the temporal attention neural network model includes: generating a simulation dataset. And calculate the corresponding theoretical optimal supply and return water pressure difference label. The label is determined by the following formula:

[0057] ;

[0058] in, Let the pump power consumption objective function be the constraint condition, which means that any pipe segment in the network must satisfy the constraint condition. Hydraulic safety requirements in case of failure; constructing a pre-trained loss function This includes the main regression loss term and the structural safety margin loss term:

[0059] ;

[0060] in, These are the model's predicted values. These are the weighting coefficients; The implicit safety margin of the model is defined as follows: , The system-level structural differential pressure envelope value in the fault protection differential pressure benchmark index. The model parameters are updated by minimizing the pre-trained loss function using the gradient descent algorithm to achieve a preset target safety margin.

[0061] Since it is difficult to obtain a large amount of real-world fault data in actual scenarios, which cannot be directly used for model training, it is necessary to construct a simulation dataset containing various random load scenarios and environmental parameters. Simultaneously, a fault-tolerant hydraulic bottleneck parameter set is used to calculate the fault protection pressure differential benchmark for each simulation scenario, ensuring that the dataset covers different operating conditions and fault constraints. To provide the model with a clear learning objective, for each simulation scenario, a numerical optimization solver is used to solve for the theoretically optimal supply and return water pressure differential that minimizes system pump power consumption, under the premise of satisfying all single-point fault hydraulic constraints. This solution serves as the training label, allowing the model to clearly define the balance between safety and energy saving. The design of the pre-training loss function must simultaneously consider two core requirements: the main regression loss term guides the model to fit the theoretically optimal supply and return water pressure differential, ensuring energy saving during normal operation; the structural safety margin loss term constrains the model output to exceed the system-level structural pressure differential envelope value in the fault protection pressure differential benchmark, ensuring fault tolerance capability. The weighting coefficients are used to balance the influence of the two loss terms, preventing either objective from being neglected. The implicit safety margin of the model directly quantifies the extent to which the model output exceeds the safety threshold. The preset target safety margin provides a clear standard for this indicator. The gradient descent algorithm continuously updates the model parameters by minimizing the pre-training loss function, enabling the model to gradually master the mapping relationship between the fault-tolerant hydraulic bottleneck parameter set and the optimized set value of the supply and return water pressure difference, thus possessing accurate prediction capabilities.

[0062] In a preferred embodiment, the input sequence and the network-wide topology constraint feature vector are input into a pre-trained temporal attention neural network model. The attention mechanism module in the temporal attention neural network model is used to fuse the network-wide topology constraint feature vector into the time step of the input sequence. After forward calculation, the optimized setpoint for the supply and return water pressure difference is output, including: using the input weight matrix. The time sequence in the input sequence Structural Enhancement Feature Vector Mapped to time step embedding vector The following formula is used to generate a hybrid feature sequence that incorporates physical topological constraints. :

[0063] ;

[0064] in, The network topology constraint feature vector is... As a position encoding vector, this step injects static structural constraints into each time step of the dynamic timing sequence; Inputting the Transformer encoder yields the context state representation. And input it into the LSTM decoder to update the hidden state. ;Utilizing output weights and bias Calculate the optimized setpoint for the supply and return water pressure difference. :

[0065] .

[0066] A temporal attention neural network is used to achieve deep fusion reasoning of dynamic operating conditions and static topological constraints. The time step serves as the basic unit of the input sequence. The time step structure-enhanced feature vector corresponding to each time step integrates the operating condition information and safety constraints at that moment, providing the basic input for the model. The input weight matrix maps the low-dimensional time step structure-enhanced feature vector to a high-dimensional time step embedding vector, improving the model's ability to process features. The position encoding vector supplements the model with the sequential information of the time series, preventing the model from confusing the operating condition logic of different time steps. The hybrid feature sequence is generated by superimposing the time step embedding vector, the network-wide topological constraint feature vector, and the position encoding vector, achieving an organic combination of dynamic temporal information and static topological constraints. This allows the model to consider both the evolution of operating conditions and fault-safe boundaries during reasoning. The Transformer encoder, with its multi-head self-attention mechanism, captures long-distance temporal dependencies in the hybrid feature sequence, explores the correlations between operating conditions at different time steps, and outputs a context state representation rich in temporal information. The LSTM decoder possesses long short-term memory capabilities, enabling temporal evolution calculations on the context state representation sequence, smoothing output fluctuations, and generating stable LSTM decoder hidden states. The output weight vector, in conjunction with the output bias scalar, maps the hidden state of the LSTM decoder to the final optimized setpoint of the supply and return water pressure difference, ensuring that the output results meet both fault tolerance constraints and energy-saving requirements of the current operating conditions.

[0067] A graph model is a structured model that maps the connection state of a water-cooled refrigeration station and its distribution network. It consists of a set of nodes and a set of edges. The model is constructed based on the network construction drawings, equipment layout diagrams, and on-site measured network connection relationships. It is derived by abstracting and transforming the equipment and channels of the actual physical network. It is used for subsequent hydraulic characteristic analysis and fault condition simulation. The construction tools can be Python's NetworkX library or MATLAB's graph theory toolbox.

[0068] The node set is the core component of the graph model, containing three types of specific nodes: the set of nodes representing the cold source side of the refrigeration station host, the set of nodes representing the critical load side of the end users (such as data center server rooms and hospital operating rooms), and the set of intermediate nodes representing the branch points of the pipeline network or the points where the pipe diameter changes. The number and location of the nodes are determined based on the pipeline construction drawings, equipment installation locations, and load distribution. Each node corresponds to an actual physical location or equipment interface.

[0069] The set of nodes on the cold source side consists of nodes corresponding to all the refrigeration units in the refrigeration station. The number of nodes is exactly the same as the number of refrigeration units actually deployed in the refrigeration station. For example, if there are 4 refrigeration units in the refrigeration station, then the set contains 4 nodes. Each node uniquely corresponds to the outlet or inlet of a refrigeration unit (determined according to the flow direction of the pipeline). The node numbers are arranged sequentially from 1 to N according to the installation order of the refrigeration units (N is the number of refrigeration units).

[0070] The critical load-side node set consists of all nodes corresponding to end users that require continuous cooling. The number of nodes is the same as the number of critical loads. Each node corresponds to the chilled water inlet of a critical load. For example, if a data center has 3 critical server rooms, then the set contains 3 nodes. The node numbers are arranged from high to low according to the importance level of the load, and the node locations are determined based on the pipeline access points of the end loads.

[0071] An edge set is a collection of channels connecting various nodes in a graph model. Each edge corresponds to a fluid transmission unit in the actual pipe network, including water distribution pipes, regulating valves, plate heat exchangers, etc. The number of edges is consistent with the number of the above-mentioned fluid transmission units in the actual pipe network. The length, pipe diameter, material, and other attributes of each edge are entered according to the pipe network construction drawings and equipment technical parameters. The edge numbers are arranged sequentially from the cold source side to the load side according to the water flow direction of the pipe network.

[0072] An incidence matrix is ​​a two-dimensional matrix that describes the connection relationship between a set of nodes and a set of edges. The number of rows in the matrix equals the number of edges in the edge set, and the number of columns equals the number of nodes in the node set. The rules for the values ​​of the matrix elements are as follows: if the water flow direction of an edge is from node A to node B, then the column corresponding to node A in the row of that edge has a value of 1, and the column corresponding to node B has a value of -1; if an edge has no direct connection to a node, then the column corresponding to that node in the row of that edge has a value of 0. The matrix is ​​constructed based on the actual connection relationship between nodes and edges, and is generated either manually or through automated recognition tools based on pipeline network drawings.

[0073] A pipe segment is any element in the edge set, corresponding to a specific fluid transmission unit in the actual pipe network. It can be divided into three categories: water distribution pipes (made of stainless steel or seamless steel pipes, with a pipe diameter range of DN50 to DN300), regulating valves (such as electric regulating valves, with a nominal pressure of 1.6MPa), and heat exchange devices (such as the flow channels of plate heat exchangers). The unique identifier of each pipe segment is consistent with the edge number in the edge set. The physical parameters of the pipe segment (length, inner diameter, roughness) are all from the pipe network construction drawings and equipment technical specifications.

[0074] The pressure difference between the supply and return water at both ends of a pipe section refers to the pressure difference between the inlet and outlet ends of a certain pipe section during operation. The measurement location is the pressure gauge interface at both ends of the pipe section. The measuring equipment adopts a high-precision differential pressure sensor (measurement range 0 to 1MPa, accuracy ±0.5% of full scale). The measurement frequency is consistent with the system data acquisition frequency (preferably 1Hz). The measurement data is transmitted to the control system in real time through an industrial bus (such as ModbusTCP).

[0075] Pipe flow rate refers to the mass flow rate of chilled water flowing through a certain pipe section per unit time. The measurement location is the middle of the pipe section (a straight pipe section away from bends and valves, with a length of not less than 5 times the pipe diameter). The measuring equipment is an electromagnetic flow meter (accuracy class 0.5, measurement range 0 to 10 kg / s). During measurement, it is necessary to ensure that the flow meter is installed coaxially with the pipe section, and that the fluid fills the pipe section. The measurement data is output as mass flow rate after temperature compensation (the compensation is based on the change curve of chilled water density with temperature).

[0076] The linearized resistance coefficient is a proportionality coefficient at the current operating point, representing the direct relationship between the pressure difference across the pipe segment and the flow rate. The calculation steps are as follows: First, obtain the pressure drop versus flow rate curve of the pipe segment under the current operating condition through experiments or simulations (the flow rate range is ±20% of the current operating point flow rate). Then, draw a tangent line at the current operating point of this curve. The slope of the tangent line is the linearized resistance coefficient. The unit of pressure drop is Pascal, the unit of flow rate is kilograms per second, and the unit of coefficient is Pascal-second per kilogram. The calculation tools can be Origin or Python's scipy library for curve fitting and differentiation.

[0077] The side admittance weight is the reciprocal of the linearized drag coefficient. It is calculated by dividing 1 by the linearized drag coefficient, and the unit is kilograms per second (Pascals per second). This parameter directly reflects the flow capacity of the pipe section. The larger the value, the smaller the obstruction to the water flow by the pipe section. After calculation, it needs to be stored in the parameter database of the control system and correspond one-to-one with the pipe section number.

[0078] The edge weight diagonal matrix is ​​a square matrix constructed from the edge admittance weights of all pipe segments. The dimension of the matrix is ​​equal to the number of edges in the edge set (i.e., the total number of pipe segments). The elements on the main diagonal of the matrix are the edge admittance weights of each pipe segment (arranged in order of pipe segment number). All off-diagonal elements are 0. The construction tool can be the numpy library in Python or the diag function in MATLAB. After construction, it is necessary to verify whether the dimension and element values ​​of the matrix are consistent with the number of pipe segments and the edge admittance weights.

[0079] The weighted Laplacian matrix is ​​a matrix that characterizes the global connectivity of a network. The construction steps are as follows: first, obtain the transpose of the incidence matrix (by swapping the rows and columns); then, perform matrix multiplication on the transpose matrix and the diagonal matrix of the edge weights; finally, perform matrix multiplication on the result and the original incidence matrix to obtain the weighted Laplacian matrix. The dimension of the matrix is ​​equal to the number of nodes in the node set. The calculation tools can be Python's NumPy library or MATLAB's matrix operation functions. After calculation, it is necessary to verify that the matrix is ​​a symmetric positive semi-definite matrix (with non-negative eigenvalues).

[0080] A single-point-failure topology is a graph model that simulates the network connection state after a pipe segment fails. It is constructed by deleting the edge corresponding to the pipe segment that needs to be simulated for failure from the original hydraulic resistance topology model (i.e., removing the pipe segment from the edge set and deleting the row corresponding to the pipe segment from the association matrix), while keeping the connection relationships of the remaining nodes and edges unchanged. The construction order is to simulate the failure of each pipe segment in sequence according to the pipe segment number, generating a single-point-failure topology equal to the number of pipe segments.

[0081] The single-point-fault topology weighted Laplace matrix is ​​the weighted Laplace matrix corresponding to the single-point-fault topology. The construction method is the same as that of the original pipeline network weighted Laplace matrix. The difference is that the correlation matrix (with the row corresponding to the failed pipe segment removed) and the edge weight diagonal matrix (with the edge admittance weight corresponding to the failed pipe segment removed) corresponding to the single-point-fault topology are used. After calculation, the symmetry and eigenvalue properties of the matrix also need to be verified.

[0082] The Moore-Penrose pseudo-inverse matrix is ​​a generalized inverse matrix calculated for the weighted Laplace matrix (usually a singular, non-invertible matrix) of a single-point fault topology. The calculation tool can be the pinv function of Python's numpy library or the pinv function of MATLAB. During the calculation, the tolerance is set to 1e-6 (used to determine the rank of the matrix). The calculation results need to be stored in a one-to-one correspondence with the fault topology number to facilitate the subsequent calculation of equivalent hydraulic impedance.

[0083] A source node is any node in the set of cold source-side nodes, that is, the node corresponding to a single chiller unit. It serves as the starting point for cold energy transfer. When calculating the equivalent hydraulic impedance, each node in the set of cold source-side nodes must be selected as a source node in sequence, and the impedance to the target critical load node must be calculated separately. The selection order of the source nodes is consistent with the numbering order of the set of cold source-side nodes.

[0084] A critical load node is any node in the set of critical load-side nodes, that is, the node corresponding to a single end critical load. As the endpoint of cold energy transfer, when calculating the equivalent hydraulic impedance, it is necessary to calculate the impedance from all source nodes to each critical load node. The selection order of critical load nodes is consistent with the numbering order of the set of critical load-side nodes.

[0085] The standard basis vector of the source node is a one-dimensional vector corresponding to the source node. The dimension of the vector is equal to the total number of nodes in the node set. The element values ​​of the vector are determined by the following rule: only the index position corresponding to the source node (in the order of node numbering, starting from 0 or 1, consistent with the indexing rule of matrix calculation tools) is 1, and all other positions are 0. For example, if the total number of nodes is 10 and the source node number is 3 (index 2), then the 3rd element of the vector (index 2) is 1, and the other 9 elements are 0. The construction tool can be the eye function of Python's numpy library.

[0086] The critical load node standard basis vector is a one-dimensional vector corresponding to the critical load node. The construction rules are the same as those of the source node standard basis vector, except that the index position corresponding to the critical load node is 1, and the rest are 0. For example, if the total number of nodes is 10 and the critical load node number is 5 (index 4), then the 5th element (index 4) of the vector is 1, and the other 9 elements are 0. After construction, it is necessary to verify whether the dimension of the vector and the position of the non-zero elements are correct.

[0087] Equivalent hydraulic impedance is the hydraulic transmission resistance from the source node to the critical load node under single-point failure conditions. The calculation steps are as follows: First, calculate the difference between the standard basis vector of the source node and the standard basis vector of the critical load node (subtract corresponding elements). Then, calculate the transpose of the difference vector. Perform matrix multiplication on the transpose vector and the Mohr-Penrose pseudo-inverse matrix. Finally, perform matrix multiplication on the result and the original difference vector. The scalar obtained is the equivalent hydraulic impedance, with the unit being Pascal-second per kilogram. The calculation tool is the same as the matrix operation tool. After calculation, the equivalent hydraulic impedance value from each source node to each critical load node for each fault topology must be recorded.

[0088] The effective supply impedance is only applicable to scenarios where the system has multiple cold source-side nodes. The calculation method involves first obtaining the equivalent hydraulic impedance from all cold source-side nodes to the target terminal critical load-side node under a single-point failure condition. Then, the smallest value among these impedance values ​​is selected as the effective supply impedance for that condition. During calculation, impedance values ​​must be extracted sequentially according to the cold source-side node numbers. The minimum value is determined using a comparison tool (such as Python's min function), and the result is stored in a data set linked to the failure condition number and the target load node number.

[0089] The most unfavorable fault condition impedance value is calculated for a single critical load-side node at the end. The calculation steps are as follows: First, iterate through all single-point fault conditions and collect the equivalent hydraulic impedance (single cooling source scenario) or effective supply impedance (multiple cooling source scenario) corresponding to the node under each condition. Second, select the largest value from all collected impedance values, which is the most unfavorable fault condition impedance value for that node. The calculation tool can be an Excel spreadsheet sorted or Python's max function. The results must correspond one-to-one with the node number and be stored in the system parameter library. The update cycle is after changes to the pipeline network structure (such as adding pipe sections or replacing equipment).

[0090] The fault-tolerant hydraulic bottleneck parameter set consists of the impedance values ​​of the most unfavorable fault conditions corresponding to all critical load-side nodes at the end, arranged in order of node number. The storage format is a one-dimensional array (e.g., [impedance value of node 1, impedance value of node 2, ...]). The construction tool can be a list in Python or an array in MATLAB. After the construction is completed, the impedance value of each node needs to be manually checked to ensure that there are no omissions or errors. It can be directly called when calculating the fault protection differential pressure benchmark index.

[0091] The current moment refers to the specific time point when the system collects operational data and calculates differential pressure requirements. It uses the Network Time Protocol (NTP) to synchronize with the cooling station monitoring system. The sampling period is preferably 1 to 10 seconds (set according to the system response speed requirements; 1 second is used for high-frequency scenarios such as data centers, and 10 seconds is used for ordinary buildings). Each current moment corresponds to a unique timestamp (format: year-month-day hour:minute:second.millisecond), which is used to associate all operational data at that moment.

[0092] Real-time flow data for nodes refers to the chilled water mass flow rate at the critical load-side node at the current moment. The measuring equipment is an electromagnetic flowmeter with an accuracy class of 0.5. The installation location is a straight pipe section where the node connects to the pipeline network (at least 5 times the pipe diameter away from elbows, valves, and other local resistance elements). During installation, ensure that the flowmeter is coaxial with the pipe section and that the pipe section is completely filled with fluid. The measured data is processed by the flowmeter's built-in temperature compensation module (the compensation is based on the correspondence between chilled water density and temperature within the range of 0-50℃), and the output unit is kilograms per second. It is transmitted to the control system in real time via the Modbus TCP protocol, with a data delay of no more than 100 milliseconds.

[0093] The node-level structural differential pressure requirement is calculated by multiplying the real-time flow data of a critical load-side node at the current moment by the impedance value of the node under the most unfavorable fault condition in the fault-tolerant hydraulic bottleneck parameter set. The product is the node-level structural differential pressure requirement, expressed in Pascals. During calculation, it is crucial to ensure the correct correspondence between the flow data and the impedance value, which can be verified by matching the node numbers. The calculation tool can be integrated into the PLC controller and automatically triggered upon completion of data acquisition at the current moment.

[0094] The calculation time of the system-level structural pressure difference envelope value is after the calculation of the node-level structural pressure difference requirements of all key load-side nodes at the current moment. Select the largest value from these requirement values, which is the system-level structural pressure difference envelope value, with the unit of Pascal. The calculation tool can use the maximum comparison instruction of the PLC or the max function of Python, and the result is bound and stored with the current moment timestamp as the safety bottom line for setting the supply and return water pressure difference in subsequent model prediction.

[0095] The fault protection pressure difference reference index consists of two parts. The first part is a one-dimensional vector formed by arranging the node-level structural pressure difference requirements of all key load-side nodes in the order of node numbers, and the second part is the system-level structural pressure difference envelope value at the current moment. The two are jointly stored in a dictionary format (such as {"node requirement vector": [value1, value2,...], "system envelope value": value}). This index is updated synchronously with the current moment and is transmitted to the time series attention neural network model in real time as the safety boundary constraint condition for model prediction.

[0096] The real-time flow data vector is a one-dimensional vector formed by arranging the real-time flow data of all key load-side nodes at the current moment in the order of node numbers. The vector dimension is equal to the number of key load-side nodes. For example, if the system has 5 key load nodes numbered from 1 to 5, the vector is [node 1 flow, node 2 flow,..., node 5 flow]. The construction tool can use the numpy array of Python or the array variable of the PLC. After construction, it is necessary to verify that the vector length is consistent with the number of nodes and the data order corresponds correctly to the node numbers.

[0097] The pump operating frequency data refers to the real-time operating frequency of the chilled water pump in the refrigeration station at the current moment. The measuring device is the frequency converter supporting the pump, and the measuring method is to read the output frequency parameter of the frequency converter, with an accuracy of ±0.1 Hz and the unit of Hz. The data is transmitted to the control system through the RS485 interface of the frequency converter, and the transmission protocol is ModbusRTU. The sampling period is the same as the current moment sampling period, and the value range is limited to 30 Hz to 50 Hz (to avoid insufficient lubrication caused by low-frequency operation of the pump or overload caused by high-frequency operation).

[0098] The external environment parameters include two core parameters, namely the outdoor dry-bulb temperature and relative humidity. The measuring device uses a temperature and humidity sensor with an accuracy of ±0.2 °C (temperature) and ±2%RH (humidity). The installation location is an area with good ventilation and no direct sunlight outside the refrigeration station (2 meters above the ground). The sampling period is the same as the current moment sampling period, with the temperature unit in degrees Celsius and the humidity unit in percentage. The data is transmitted to the control system in real time through the industrial bus to reflect the impact of the external environment on the terminal cooling load.

[0099] The structural enhancement feature vector is a comprehensive combination of features at the current moment. The construction order is: real-time flow data vector → pump operating frequency data → external environmental parameters → system-level structural differential pressure envelope value → node-level structural differential pressure demand vector. All elements are concatenated into a one-dimensional vector in this order. The vector dimension is "number of critical load nodes + 1 (pump frequency) + 2 (temperature and humidity) + 1 (system envelope value) + number of critical load nodes (node ​​demand)". The construction tool can use Python's `concatenate` function. After construction, the number of vector elements should be checked to ensure accuracy and avoid missing or redundant elements.

[0100] The preset time window length refers to the number of historical moments included when constructing the time-series input sequence. A preferred value is 30 to 60 sampling periods, determined based on the response lag time of the network's hydraulic balance (e.g., 60 periods for larger networks with slower responses, and 30 periods for smaller networks). For example, with a sampling period of 10 seconds and a time window length of 48, the input sequence includes historical operating conditions over the past 480 seconds (8 minutes). This parameter is set before model training and can be adjusted during operation based on system stability (the length can be increased appropriately if frequent fluctuations occur).

[0101] The time-series feature input sequence is a two-dimensional sequence composed of the structure-enhanced feature vectors of the current moment and all historical moments within a preset time window, arranged chronologically. The sequence shape is "time window length × structure-enhanced feature vector dimension". For example, if the time window length is 48 and the feature vector dimension is 20, the sequence will be a 48-row, 20-column matrix. During construction, the features are sorted from earliest to latest timestamp, with the most recent feature vector at the end of the sequence. The construction tool can be Python's pandas.DataFrame or MATLAB's matrix. The data is stored and named according to the current timestamp for easy model retrieval.

[0102] The learnable weight matrix is ​​used for feature extraction and dimension mapping of the fault-tolerant hydraulic bottleneck parameter set. The initial values ​​are set using the Xavier initialization method, and the matrix dimension is "the hidden layer dimension of the temporal attention neural network model × the dimension of the fault-tolerant hydraulic bottleneck parameter set". For example, if the hidden layer dimension is 64 and the parameter set dimension is 5, the matrix will be 64 rows and 5 columns. This matrix is ​​continuously updated during model training using the gradient descent algorithm, with the update frequency matching the model training batches. After training, it is permanently stored in the model file and directly loaded for use during inference.

[0103] The bias vector is a one-dimensional vector used in conjunction with the learnable weight matrix. It is initially set as an all-zero vector, with the vector dimension matching the hidden layer dimension of the temporal attention neural network model. For example, if the hidden layer dimension is 64, the bias vector is a vector consisting of 64 zeros. This vector is updated synchronously with the learnable weight matrix by multiplying the bias gradient calculated using the gradient descent algorithm by the learning rate and adding it to the original vector, ensuring the accuracy of feature mapping.

[0104] The ReLU function is preferred as the nonlinear activation function. The function rule is to output the input value when the input value is greater than 0, and to output 0 when the input value is less than or equal to 0. The calculation tool can be the torch.nn.ReLU module in Python or the tf.nn.relu function in TensorFlow. Its function is to perform nonlinear feature mapping on the fault-tolerant hydraulic bottleneck parameter set after linear transformation (learnable weight matrix weighting + bias vector bias), thereby enhancing the model's ability to fit complex relationships.

[0105] The network-wide topology constraint feature vector is generated from the fault-tolerant hydraulic bottleneck parameter set through linear transformation and nonlinear activation. The generation steps are as follows: First, perform matrix multiplication between the learnable weight matrix and the fault-tolerant hydraulic bottleneck parameter set; second, add the result of the multiplication to the bias vector; third, input the addition result into the nonlinear activation function, and the output is the network-wide topology constraint feature vector. The dimension of this vector is completely consistent with the dimension of the hidden layer of the temporal attention neural network model. After generation, it is synchronously input into the model with the temporal operating condition feature input sequence to constrain the safety boundary predicted by the model.

[0106] The hidden layer dimension of a temporal attention neural network model refers to the number of neurons in the hidden layer. The preferred values ​​are 64, 128, or 256, determined based on the dimension of the fault-tolerant hydraulic bottleneck parameter set and the length of the input sequence of temporal operating conditions (256 for larger parameter set dimensions and sequence lengths, and 64 for smaller ones). This parameter is set during model construction and needs to be verified through multiple experiments (e.g., testing the prediction accuracy and convergence speed of the model under different dimensions) to select the dimension with the best overall performance. Once set, this dimension remains unchanged during model training and inference.

[0107] The simulation dataset is a dedicated dataset for model training, containing various random load scenarios and environmental parameters. The construction steps are as follows: First, load values ​​are randomly generated within a range of 50% to 150% of the terminal load, and environmental parameters are randomly generated within a range of outdoor temperature -5℃ to 40℃ and relative humidity 30% to 90%. Second, for each combination of load and environmental parameters, the corresponding fault-tolerant hydraulic bottleneck parameter set is used to calculate the corresponding fault protection pressure differential benchmark index. Third, the load data, environmental parameters, and fault protection pressure differential benchmark index are arranged in time series (each scenario contains 1000 time steps) to form a single simulation sample. The final dataset contains 10,000 samples from different scenarios, stored in CSV format. Each sample includes fields for time step, load value, environmental parameters, and fault protection pressure differential benchmark index. The pandas library in Python can be used as the construction tool.

[0108] The numerical optimization solver is a computational tool used to solve for the theoretically optimal supply and return water pressure difference in a simulation scenario. Commercial software GAMS or open-source tool SCIP are preferred. Configuration parameters include selecting the interior-point method as the solution algorithm, setting the convergence tolerance to 1e-6, and setting the maximum number of iterations to 1000. When using it, the hydraulic constraints (N-1 constraints) and the pump power consumption objective function need to be converted into mathematical expressions recognizable by the solver. Inputs include the load, environmental parameters, and fault-tolerant pressure difference benchmark of the simulation scenario. After running, the output is the supply and return water pressure difference that satisfies the constraints and minimizes pump power consumption.

[0109] The pump power consumption objective function is a calculation expression that quantifies the energy consumption of the system's pumps. The calculation method involves dividing the shaft power of each pump by the efficiency of the corresponding motor, and then summing all the results. The pump shaft power is obtained from the performance curve (flow rate-shaft power relationship) on the pump's nameplate, and the motor efficiency is determined based on the motor's rated load efficiency curve (provided by the motor manufacturer), with a preferred value range of 0.85 to 0.92. The unit of calculation is kilowatts.

[0110] The theoretically optimal supply and return water pressure difference label represents the optimal supply and return water pressure difference value for each simulation scenario. It is obtained by the numerical optimization solver by minimizing the pump power consumption objective function while satisfying all single-point failure conditions and hydraulic constraints. The unit is Pascal. This label corresponds one-to-one with the load and environmental parameters of the simulation scenario, and is bound to the simulation samples during storage. It serves as a supervisory signal for model training, ensuring that the model learns the pressure difference setting logic that balances safety and energy saving.

[0111] Training labels are theoretically optimal supply and return water pressure difference labels, which are completely consistent with the former. They are the benchmark data used to fit the prediction results during model training. The storage format is a one-dimensional array, which corresponds one-to-one with the time step of the simulation dataset. When called, they are associated through the sample index.

[0112] The N-1 constraint is the core constraint condition for ensuring the hydraulic safety of the system. Specifically, it requires that for any pipe segment in the simulation scenario, after simulating its failure, the flow rate of all critical load-side nodes must still be maintained above 90% of the rated flow rate (the rated flow rate is determined by the design value of the end load). In the numerical optimization solver, this is achieved by setting a lower limit constraint on the flow rate, that is, the minimum flow rate variable of each critical load node is set to 90% of the rated flow rate.

[0113] The pre-training loss function is a loss calculation rule that guides the updating of model parameters. It consists of two parts: the main regression loss term and the structural safety margin loss term. The main regression loss term calculates the squared difference between the model's predicted value and the training label, while the structural safety margin loss term calculates the squared difference between the model's implicit safety margin and the preset target safety margin. The two are multiplied by weight coefficients and then summed to obtain the total loss, which is used to measure the accuracy and safety of the model's prediction results.

[0114] The model prediction is the supply and return water pressure difference prediction result of the temporal attention neural network during the training process, based on the input and output of the simulation dataset. The unit is Pascal. The calculation process is that after the model extracts features and performs temporal inference on the input simulation samples, it obtains the prediction value through the output layer. Each time step corresponds to a prediction value, which is stored synchronously with the training labels.

[0115] The weight coefficients are parameters that balance the two parts of the loss in the pre-training loss function. The preferred weight coefficient for the main regression loss term is 1.0, and the preferred weight coefficient for the structural safety margin loss term is 0.5. The values ​​are determined through multiple experiments (testing the prediction accuracy and safety margin achievement rate of the model under different weight combinations) to ensure both the model fitting effect and safety constraints. Once set, these values ​​remain unchanged during training.

[0116] The model implicit safety margin is the difference between the model prediction value and the system-level structural pressure differential envelope value in the fault tolerance benchmark index. It is calculated by subtracting the system-level structural pressure differential envelope value at the same time step from the model prediction value. The unit is Pascal, which directly reflects the degree to which the model prediction value exceeds the safety baseline. The larger the value, the stronger the fault tolerance capability.

[0117] The preset target safety margin is a pressure buffer value reserved to cope with pipeline pressure fluctuations and valve dynamic response lag. The preferred value range is 20,000 to 50,000 Pa. The specific value is determined according to the pipeline scale. When the pipeline diameter is less than DN200, 20,000 Pa is used. When the pipeline diameter is between DN200 and DN500, 35,000 Pa is used. When the pipeline diameter is greater than DN500, 50,000 Pa is used. It is manually configured by technicians before model training.

[0118] Gradient descent is an optimization algorithm used to minimize the pre-training loss function and update model parameters. The Adam optimizer is preferred, with the following configuration parameters: initial learning rate of 0.001, weight decay coefficient of 1e-5, momentum parameters β1 of 0.9, β2 of 0.999, and epsilon of 1e-8. During runtime, the loss gradient of each model parameter is calculated through backpropagation, and the parameter values ​​are adjusted according to the learning rate. The parameters are updated every 100 simulation samples until the loss function converges (the loss change is less than 1e-4 over 10 consecutive training batches).

[0119] Model parameters are the core learnable parameters in a temporal attention neural network, including learnable weight matrices, bias vectors, multi-head attention weights of the Transformer encoder, and hidden layer weights of the LSTM decoder. Initial values ​​are set using the Xavier initialization method and continuously updated during training using the gradient descent algorithm. After training, the parameters are stored as a .pth format file and loaded directly into the model during inference.

[0120] A time step is a single time unit contained in the input sequence. Each time step corresponds to a structure-enhanced feature vector. The interval between time steps is consistent with the system data acquisition cycle (preferably 1 to 10 seconds). For example, if the acquisition cycle is 5 seconds, then each time step represents 5 seconds of system operation status. The time steps in the input sequence are arranged in order from early to late.

[0121] The time-step structure enhancement feature vector is a comprehensive feature combination corresponding to each time step in the input sequence. It includes real-time flow data vector, pump operating frequency data, external environmental parameters, system-level structural pressure differential envelope value, and node-level structural pressure differential demand vector. The vector dimension is consistent with the structure enhancement feature vector, and the storage format is a one-dimensional array, with each element arranged in a fixed order.

[0122] The input weight matrix is ​​a learnable matrix that maps the time-step structure-enhanced feature vectors to high-dimensional embedding vectors. The initial values ​​are set using the Xavier initialization method, and the matrix dimension is "high-dimensional embedding vector dimension × time-step structure-enhanced feature vector dimension". The preferred value for the high-dimensional embedding vector dimension is 64 or 128 (chosen according to model complexity). For example, if the structure-enhanced feature vector dimension is 20 and the embedding dimension is 64, the matrix will be 64 rows and 20 columns. During training, the matrix element values ​​are updated using the gradient descent algorithm.

[0123] The time-step embedding vector is a high-dimensional feature vector obtained by mapping the time-step structure enhancement feature vector to the input weight matrix. It is generated by performing matrix multiplication between the time-step structure enhancement feature vector and the input weight matrix. The dimension of the output vector is the same as the number of rows of the input weight matrix (64 or 128). The unit has no actual physical meaning and is only used for feature processing within the model. The generation tool can be the torch.nn.Linear module of Python.

[0124] The positional encoding vector is used to identify the chronological order of time steps. It is generated by calculating the sine value for each time step at even-numbered indices (the input is the time step index divided by 10000, which is 2 × index / embedding dimension raised to the power of 10000), and the cosine value for each time step at odd-numbered indices (the input is the same as for even-numbered indices). The final result is a vector with the same dimensions as the time step embedding vector. For example, if the embedding dimension is 64 and the time step is 5, then for the 0th (even) position, sin(5 / 10000(0 / 64)) is calculated, and for the 1st (odd) position, cos(5 / 10000(2 / 64)) is calculated, and so on for all dimensions. The generation tool can be the numpy library in Python.

[0125] The hybrid feature sequence is a feature sequence that integrates dynamic temporal features and static topological constraints. It is generated by adding the time step embedding vector, the network topological constraint feature vector, and the location encoding vector at each time step element by element. Each time step generates a hybrid feature vector. The hybrid feature vectors of all time steps are arranged in chronological order to form a two-dimensional sequence (shaped as "time window length × embedding dimension"). The generation process is automatically completed by the model forward propagation process.

[0126] The Transformer encoder is a module for extracting long-range temporal dependencies in a mixed feature sequence. Its structure consists of three layers, each containing one multi-head self-attention sub-layer and one feedforward neural network sub-layer. The multi-head self-attention sub-layer has eight heads, with each head having a dimension equal to the embedding dimension / 8 (e.g., if the embedding dimension is 64, then each head has a dimension of 8). The hidden layer dimension of the feedforward neural network sub-layer is set to 256, and the activation function is ReLU. After inputting the mixed feature sequence, it undergoes layer normalization, multi-head self-attention calculation, and feedforward neural network operation, outputting a context state representation sequence.

[0127] The context state representation is the output of the Transformer encoder. It is a two-dimensional sequence (with the same shape as the mixed feature sequence). The context state representation vector of each time step integrates the feature association information of that time step with all other time steps. For example, the vector of the current time step contains the load change trend and environmental parameter fluctuation characteristics of all past time steps, providing input rich in association information for subsequent time series evolution calculations.

[0128] The LSTM decoder is a module that performs temporal evolution computation on the context state representation sequence. Its structure is configured with the number of hidden layer units set to match the embedding dimension (64 or 128), two layers, a dropout rate of 0.1, and the tanh activation function. After inputting the context state representation sequence, the hidden layer states are updated sequentially at time steps to smooth temporal fluctuations and output stable hidden states. The processing tool can be the `torch.nn.LSTM` module in Python.

[0129] The hidden state of the LSTM decoder is the output state of the LSTM decoder at the current moment. It is a one-dimensional vector (the dimension is the same as the number of hidden layer units) and contains all the temporal features and topological constraints of the context state representation sequence. For example, the vector element values ​​indirectly reflect historical load peaks and pressure difference requirements under extreme conditions. It is the core basis for generating the optimized setpoint for supply and return water pressure difference.

[0130] The output weight vector is a learnable vector that maps the hidden state of the LSTM decoder to the optimized set value of the supply and return water pressure difference. The initial value is set by the Xavier initialization method, and the vector dimension is consistent with the hidden state dimension of the LSTM decoder (64 or 128). During training, the element values ​​are updated by the gradient descent algorithm to ensure the accuracy of the mapping relationship.

[0131] The output bias scalar is a parameter used to fine-tune the hidden state mapping result of the LSTM decoder. It is initially set to 0 and is updated synchronously with the output weight vector during training. The update method is to accumulate the product of the bias gradient calculated by the gradient descent algorithm and the learning rate to the original value. The unit is Pascal, and it is used to compensate for the system error in the mapping process.

[0132] The optimized setpoint for the supply and return water pressure difference is the target value of the system supply and return water pressure difference output by the model. It is generated by performing a dot product operation between the hidden state of the LSTM decoder and the output weight vector, plus the output bias scalar, with units in Pascals. This value must satisfy the hydraulic safety boundary defined by the fault-tolerant hydraulic bottleneck parameter set (i.e., greater than the system-level structural pressure difference envelope value), while keeping pump power consumption at a low level. After generation, it is sent to the PLC controller via the Modbus TCP protocol, with an update cycle consistent with the current sampling cycle.

[0133] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An optimization control method for a fault-tolerant water-cooled chiller system based on a large model, characterized in that, include: Establish a topological model of hydraulic resistance of the pipeline network that maps the connection status of the water-cooled refrigeration station and the transmission and distribution pipeline network, and determine the linearized resistance coefficient of each pipeline segment under operating conditions. Based on the aforementioned hydraulic resistance topology model of the pipeline network, the equivalent hydraulic impedance from the cold source side to the terminal critical load side is calculated by traversing the single-point failure conditions of each pipe section in the transmission and distribution network, including: Using the correlation matrix and marginal admittance weights of the aforementioned pipeline hydraulic resistance topology model, a weighted Laplace matrix characterizing the global connectivity of the pipeline network is constructed. For each pipe segment in the transmission and distribution network, a corresponding single-point failure topology is constructed by removing the pipe segment from the pipeline hydraulic resistance topology model. For each single-point failure topology, the Mohr-Penrose pseudo-inverse matrix of its corresponding weighted Laplace matrix is ​​calculated. Using the Mohr-Penrose pseudo-inverse matrix, the quadratic form of the difference between the standard basis vector representing the cold source-side node and the standard basis vector representing the terminal critical load-side node is calculated, and the calculation result is used as the equivalent hydraulic impedance from the cold source side to the terminal critical load side under the single-point failure condition. Extract the maximum equivalent hydraulic impedance of each critical load side under all single-point failure conditions as the most unfavorable failure condition impedance value, and construct a fault-tolerant hydraulic bottleneck parameter set based on the most unfavorable failure condition impedance values ​​of all critical load sides. The system collects real-time flow data, calculates the fault protection differential pressure benchmark index using the fault-tolerant hydraulic bottleneck parameter set, generates an input sequence containing time-series operating condition characteristics, and maps the fault-tolerant hydraulic bottleneck parameter set into a network-wide topology constraint feature vector. The input sequence and the network topology constraint feature vector are input into a pre-trained temporal attention neural network model. The attention mechanism module in the temporal attention neural network model is used to fuse the network topology constraint feature vector into the time step of the input sequence. The optimized set value of the supply and return water pressure difference is output after forward calculation. The optimized setpoint for the supply and return water pressure difference is predicted by the temporal attention neural network model under the hydraulic safety boundary conditions defined by the fault-tolerant hydraulic bottleneck parameter set.

2. The optimization control method for a fault-tolerant water-cooled chiller system based on a large model as described in claim 1, characterized in that, Establish a topological model of the hydraulic resistance of the water-cooled refrigeration station and its distribution network, including: The water-cooled refrigeration station and its distribution network are abstracted as a directed or undirected graph model consisting of a set of nodes and a set of edges. The set of nodes includes cold source-side nodes representing the main unit of the refrigeration station, critical load-side nodes representing end users, and intermediate nodes representing branch points or diameter change points of the network. The set of edges includes fluid connection channels representing water distribution pipes, regulating valves, and heat exchange devices. An association matrix describing the connection relationship between the set of nodes and the set of edges is constructed based on the connection relationship.

3. The optimization control method for a fault-tolerant water-cooled chiller system based on a large model according to claim 2, characterized in that, Determine the linearized resistance coefficient of each pipe section under operating conditions, including: For each pipe segment in the edge set, the hydraulic characteristic relationship between pressure drop and flow rate under the current operating conditions is obtained; at the operating point of the current operating conditions, the nonlinear relationship between pressure drop and flow rate is linearized to obtain a proportionality coefficient characterizing the proportional relationship between the pressure difference at both ends of the pipe segment and the flow rate through the pipe segment, and the proportionality coefficient is determined as the linearized resistance coefficient; the reciprocal of the linearized resistance coefficient is calculated and used as the edge admittance weight of the pipe segment in the hydraulic resistance topology model of the pipe network.

4. The fault-tolerant water-cooled chiller system optimization control method based on a large model according to claim 3, characterized in that, The maximum equivalent hydraulic impedance of each critical load side under all single-point failure conditions is extracted as the most unfavorable failure condition impedance value. A fault-tolerant hydraulic bottleneck parameter set is constructed based on the most unfavorable failure condition impedance values ​​of all critical load sides, including: For each critical load-side node at the end, obtain the set of equivalent hydraulic impedances calculated under all single-point failure conditions. If there are multiple cold source-side nodes, under each single-point failure condition, select the minimum value among the equivalent hydraulic impedances from each cold source-side node to the critical load-side node at the end as the effective supply impedance under that condition. Select the largest value among the equivalent hydraulic impedances or effective supply impedances corresponding to all single-point failure conditions, and use it as the most unfavorable failure condition impedance value for that critical load-side node at the end. Combine the most unfavorable failure condition impedance values ​​corresponding to all critical load-side nodes at the end to form the fault-tolerant hydraulic bottleneck parameter set.

5. The optimization control method for a fault-tolerant water-cooled chiller system based on a large model according to claim 4, characterized in that, The system collects real-time flow data and uses the fault-tolerant hydraulic bottleneck parameter set to calculate the fault protection differential pressure benchmark index, including: For each critical load-side node, the real-time flow data of that node at the current moment is multiplied by the impedance value of the most unfavorable fault condition corresponding to the fault-tolerant hydraulic bottleneck parameter set to obtain the node-level structural differential pressure requirement at the current moment. At the current moment, the largest value among the node-level structural differential pressure requirements of all critical load-side nodes is selected as the system-level structural differential pressure envelope value. The node-level structural differential pressure requirements of all critical load-side nodes are combined with the system-level structural differential pressure envelope value to form the fault protection differential pressure benchmark index.

6. The optimization control method for a fault-tolerant water-cooled chiller system based on a large model according to claim 1, characterized in that, Generate an input sequence containing time-series operating condition characteristics, and map the fault-tolerant hydraulic bottleneck parameter set into a network-wide topology constraint feature vector, including: The real-time flow data, the calculated fault protection differential pressure benchmark index, the pump operating frequency data, and the external environmental parameters are combined according to the current time step to construct the structure-enhanced feature vector for the current moment. The structure-enhanced feature vectors containing the current moment and historical moments of a preset length are selected and arranged in chronological order to form the input sequence containing time-series operating condition features. Using a linear transformation layer with learnable weights and a nonlinear activation function, feature extraction and dimension mapping are performed on the fault-tolerant hydraulic bottleneck parameter set to generate the network-wide topology constraint feature vector that matches the hidden layer dimension of the time-series attention neural network model.

7. The optimization control method for a fault-tolerant water-cooled chiller system based on a large model according to claim 1, characterized in that, The pre-training process of a temporal attention neural network model includes: A simulation dataset containing various random load scenarios and environmental parameters is constructed, and the fault-tolerant hydraulic bottleneck parameter set is used to calculate the fault-protection pressure difference benchmark index under the simulation scenarios. For each simulation scenario, a numerical optimization solver is used to solve for the theoretical optimal supply and return water pressure difference that minimizes the system pump power consumption under hydraulic constraints that satisfy all single-point failure conditions, and this is used as the training label. A pre-training loss function is constructed, which consists of a main regression loss term for fitting the training label and a structural safety margin loss term for constraining the model output to be higher than the fault-protection pressure difference benchmark index. The temporal attention neural network model is trained using the simulation dataset and the training label. By minimizing the pre-training loss function, the model learns the mapping relationship between the fault-tolerant hydraulic bottleneck parameter set and the optimized setpoint of the supply and return water pressure difference.

8. The optimization control method for a fault-tolerant water-cooled chiller system based on a large model according to claim 1, characterized in that, The input sequence and the network-wide topology constraint feature vector are input into a pre-trained temporal attention neural network model. The attention mechanism module in the temporal attention neural network model fuses the network-wide topology constraint feature vector into the time step of the input sequence. After forward calculation, the optimized setpoint for the supply and return water pressure difference is output, including: The structure-enhanced feature vector of each time step in the input sequence is mapped to a high-dimensional time-step embedding vector using an input embedding layer. The network-wide topology constraint feature vector is used as global structural prior information and superimposed on the time-step embedding vector and position encoding vector of each time step to generate a hybrid feature sequence that incorporates physical topology constraints. The hybrid feature sequence is input into an encoder containing a multi-head self-attention mechanism for feature extraction to obtain a context state representation sequence. The context state representation sequence is input into a long short-term memory decoder for temporal evolution calculation, and the optimized setpoint of the supply and return water pressure difference at the current moment is generated through an output mapping layer.

Images