Search

CN-121763783-B - Fault-tolerant water-cooling refrigeration station system optimization control method based on large model

CN121763783BCN 121763783 BCN121763783 BCN 121763783BCN-121763783-B

Abstract

The invention relates to the technical field of industrial control and discloses a fault tolerant water cooling station system optimization control method based on a large model, which comprises the steps of firstly establishing a pipe network hydraulic resistance topology model, traversing all single pipe failure working conditions to calculate equivalent hydraulic impedance from a cold source to the tail end, and extracting impedance values under the most unfavorable working conditions to construct a fault tolerant hydraulic bottleneck parameter set; and the model utilizes an attention mechanism to fuse the topological constraint into time sequence deduction, and directly predicts a water supply and return pressure difference set value meeting a potential fault safety boundary. The method effectively solves the problem that the traditional control strategy causes system paralysis due to insufficient safety margin when the pipe network breaks down, and ensures the continuity of key load cooling and the economical efficiency of system operation.

Inventors

  • LI HUI

Assignees

  • 南京深度智控科技有限公司

Dates

Publication Date
20260508
Application Date
20260304

Claims (8)

  1. 1. The fault tolerant water-cooling refrigeration station system optimization control method based on the large model is characterized by comprising the following steps: Establishing a pipe network hydraulic resistance topology model for mapping the connection states of the water-cooling refrigeration station and the transmission pipe network, and determining linearization resistance coefficients of each pipe section under the operation working condition; based on the pipe network hydraulic resistance topology model, traversing the single-point fault working condition of each pipe section failure in the transmission and distribution pipe network, and calculating the equivalent hydraulic resistance from the cold source side to the tail end key load side, wherein the method comprises the following steps: Constructing a corresponding single-point fault topology by removing each pipe section in the pipe network hydraulic resistance topology model according to each pipe section in the transmission pipe network, calculating a molar-Peng Resi pseudo-inverse matrix of the corresponding weighted Laplace matrix according to each single-point fault topology, calculating a quadratic form of the difference between a standard base vector representing a cold source side node and a standard base vector representing a terminal key load side node according to the molar-Peng Resi pseudo-inverse matrix, and taking the calculation result as the equivalent hydraulic impedance from the cold source side to the terminal key load side under the single-point fault working condition; Extracting the maximum equivalent hydraulic impedance of each key load side under all single-point fault working conditions as the most unfavorable fault working condition impedance value, and constructing a fault tolerant hydraulic bottleneck parameter set based on the most unfavorable fault working condition impedance value of all key load sides; Collecting real-time flow data of a system, calculating a fault guarantee pressure difference reference index by using the fault-tolerant hydraulic bottleneck parameter set, generating an input sequence containing time sequence working condition characteristics, and mapping the fault-tolerant hydraulic bottleneck parameter set into a full-network topology constraint characteristic vector; Inputting the input sequence and the full-network topological constraint feature vector into a pre-trained time sequence attention neural network model, fusing the full-network topological constraint feature vector into a time step of the input sequence by using an attention mechanism module in the time sequence attention neural network model, and outputting a supply and return water pressure difference optimization set value through forward calculation; The supply and return water pressure difference optimization set value is obtained by predicting the time sequence attention neural network model under the condition of meeting the hydraulic safety boundary condition defined by the fault tolerance hydraulic bottleneck parameter set.
  2. 2. The method for optimizing control of a large model-based fault tolerant water-cooled refrigeration station system according to claim 1, wherein establishing a network hydraulic resistance topology model mapping the connection states of the water-cooled refrigeration station and the transmission network comprises: The method comprises the steps of abstracting a water-cooling refrigeration station and a transmission and distribution pipe network into a directed graph or undirected graph model formed by a node set and an edge set, wherein the node set comprises a cold source side node representing a host machine of the refrigeration station, a key load side node representing an end user and an intermediate node representing a pipe network branch point or a reducing point, the edge set comprises a fluid connection channel representing a transmission and distribution pipe, an adjusting valve and a heat exchange device, and an association matrix describing the connection relation between the node set and the edge set is constructed based on the connection relation.
  3. 3. The method for optimizing control of a large model-based fault tolerant water-cooled refrigeration station system of claim 2, wherein determining the linearization resistance coefficient of each pipe section under the operating condition comprises: The method comprises the steps of obtaining the hydraulic characteristic relation between pressure drop and flow of each pipe section in the edge set under the current operation working condition, carrying out linear approximation processing on the nonlinear relation between the pressure drop and the flow at the operation point of the current operation working condition to obtain a proportionality coefficient representing the proportional relation between the pressure difference at two ends of the pipe section and the flow passing through the pipe section, determining the proportionality coefficient as the linear resistance coefficient, calculating the reciprocal of the linear resistance coefficient, and taking the reciprocal as the edge admittance weight of the pipe section in the pipe network hydraulic resistance topological model.
  4. 4. The method for optimizing control of a large model-based fault tolerant water-cooled chiller plant system of claim 3, wherein extracting the maximum equivalent hydraulic impedance of each critical load side under all the single point fault conditions as the most unfavorable fault condition impedance value, and constructing a fault tolerant hydraulic bottleneck parameter set based on the most unfavorable fault condition impedance values of all the critical load sides comprises: The method comprises the steps of obtaining a set of equivalent hydraulic impedance calculated under all single-point fault working conditions of each end key load side node, selecting the minimum value of the equivalent hydraulic impedance from each cold source side node to the end key load side node as effective supply impedance under the working conditions under each single-point fault working condition if a plurality of cold source side nodes exist, selecting the maximum value of the equivalent hydraulic impedance or the effective supply impedance corresponding to all the single-point fault working conditions as the most unfavorable fault working condition impedance value of the end key load side node, and combining the most unfavorable fault working condition impedance values corresponding to all the end key load side nodes to form the fault-tolerant hydraulic bottleneck parameter set.
  5. 5. The method for optimizing control of a large model-based fault tolerant water-cooled refrigeration station system of claim 4, wherein collecting real-time flow data of the system, calculating a fault assurance pressure difference reference index using the fault tolerant hydraulic bottleneck parameter set, comprises: The method comprises the steps of multiplying real-time flow data of each key load side node at the current moment by the least adverse fault working condition impedance value corresponding to a fault tolerance hydraulic bottleneck parameter set to obtain node level structure differential pressure requirements of the node at the current moment, selecting the node level structure differential pressure requirement of all key load side nodes with the largest value as a system level structure differential pressure envelope value at the current moment, and combining the node level structure differential pressure requirements of all key load side nodes with the system level structure differential pressure envelope value to form the fault guarantee differential pressure reference index.
  6. 6. The method for optimizing control of a large model-based fault tolerant water-cooled refrigeration station system of claim 1, wherein generating an input sequence comprising time sequence operating condition characteristics and mapping the fault tolerant hydraulic bottleneck parameter set to a full network topology constraint feature vector comprises: The real-time flow data, the calculated fault guarantee differential pressure reference index, the water pump operating frequency data and the external environment parameters are combined according to the current time step to construct a structural enhancement feature vector at the current moment, the structural enhancement feature vectors at the historical moment comprising the current moment and the past preset length are selected and arranged according to the time sequence to form the input sequence comprising the time sequence working condition characteristics, and the linear transformation layer and the nonlinear activation function with the learning weight are utilized to conduct feature extraction and dimension mapping on the fault tolerance hydraulic bottleneck parameter set to generate the full-network topology constraint feature vector matched with the hidden layer dimension of the time sequence attention neural network model.
  7. 7. The method for optimizing control of a large model-based fault tolerant water-cooled refrigeration station system of claim 1, wherein the pre-training process of the time series attention neural network model comprises: The method comprises the steps of constructing a simulation data set containing various random load scenes and environment parameters, calculating a fault guarantee pressure difference reference index under the simulation scenes by utilizing the fault tolerance hydraulic bottleneck parameter set, solving a theoretical optimal water supply and return pressure difference which enables the power consumption of a system pump to be minimum under the hydraulic constraint condition of meeting all single-point fault working conditions by utilizing a numerical optimization solver for each simulation scene, taking the theoretical optimal water supply and return pressure difference as a training label, constructing a pre-training loss function, wherein the pre-training loss function consists of a main regression loss item for fitting the training label and a structural safety margin loss item for constraining model output to be higher than the fault guarantee pressure difference reference index, training the time sequence attention neural network model by utilizing the simulation data set and the training label, and enabling the model to learn the mapping relation between the fault tolerance hydraulic bottleneck parameter set and a water supply and return pressure difference optimization set value by minimizing the pre-training loss function.
  8. 8. The optimization control method of a large model-based fault tolerant water-cooled refrigeration station system according to claim 1, wherein the input sequence and the full-network topology constraint feature vector are input into a pre-trained time sequence attention neural network model, the full-network topology constraint feature vector is fused into a time step of the input sequence by an attention mechanism module in the time sequence attention neural network model, and a supply and return water pressure difference optimization set value is output through forward calculation, and the method comprises the steps of: The method comprises the steps of mapping a structural enhancement feature vector of each time step in an input sequence into a high-dimensional time step embedded vector by using an input embedded layer, overlapping the full-network topology constraint feature vector serving as global structure priori information on the time step embedded vector and a position coding vector of each time step to generate a mixed feature sequence fused with physical topology constraint, inputting the mixed feature sequence into an encoder comprising a multi-head self-attention mechanism to perform feature extraction to obtain a context state representation sequence, inputting the context state representation sequence into a long-period memory decoder to perform time sequence evolution calculation, and generating the water supply return differential pressure optimization set value at the current moment through an output mapping layer.

Description

Fault-tolerant water-cooling refrigeration station system optimization control method based on large model Technical Field The invention relates to the technical field of industrial control, in particular to a fault-tolerant water-cooling refrigeration station system optimization control method based on a large model. Background In modern large public buildings and industrial parks, regional cooling systems (DistrictCoolingSystem) typically employ complex hydraulic configurations with multi-source ring networks, multi-stage nips, and pressure differential control coordination. The system adjusts the flow of the chilled water by utilizing a variable pressure difference control strategy through the cooperative work of a plurality of refrigeration hosts and an annular transmission and distribution pipe network so as to adapt to the dynamic change of the end load. Conventional control logic generally relies on real-time feedback regulation, i.e., by monitoring the real-time state of the most unfavorable thermodynamic circuit (the area where the end differential pressure is most difficult to meet) in a pipe network, to dynamically adjust the supply and return water differential pressure set value of a refrigeration station water collector. The mode can reduce the energy consumption of the water pump as much as possible while meeting the terminal refrigeration requirement under the ideal stable working condition, and is a main stream means in the field of automatic control of the current heating ventilation air conditioner. However, with the increase of the requirements on the reliability of the system, especially in the scenes such as data centers, which have extremely high requirements on the continuity of cooling, the conventional control method gradually exposes serious potential safety hazards. In the prior art, when calculating an optimal differential pressure set value, the hydraulic characteristic of the pipe network is often optimized only based on the current moment and the current normal connection state. Most rule-based or common data-driven (e.g., conventional neural networks, reinforcement learning) optimization algorithms, limited in their field of view to historical data and current topology that have already occurred, tend to compress differential pressure very much in the normal operating mode in pursuit of energy savings. However, a large number of bypass or standby pipe sections with extremely small or even zero flow at ordinary times exist in a complex annular pipe network, and the bypass or standby pipe sections are not obviously important in normal operation, but are the only life lines for maintaining the connectivity of the system in case of sudden pipe network faults (such as breakage of key pipe sections and locking of valves). Because the existing control model cannot sense the minimum cut set or potential bottleneck in the graph theory sense, the calculated differential pressure set value is often clung to the lower limit of the current working condition. Once the system fails at a single point (N-1 condition), the hydraulic flow direction is forced to reconstruct, and the resistance to flow through the backup path increases instantaneously. At this time, if the system still maintains the original low pressure difference setting, the sudden increase of the pipe network impedance cannot be overcome, so that the critical tail end is instantaneously cooled, and serious safety accidents are caused. Disclosure of Invention The invention provides a fault-tolerant water-cooling refrigeration station system optimization control method based on a large model, which solves the technical problems in the background technology. The invention provides a fault tolerant water-cooling refrigeration station system optimization control method based on a large model, which comprises the following steps: Establishing a pipe network hydraulic resistance topology model for mapping the connection states of the water-cooling refrigeration station and the transmission pipe network, and determining linearization resistance coefficients of each pipe section under the operation working condition; Based on the network hydraulic resistance topology model, traversing single-point fault conditions of each pipe section failure in a transmission and distribution network, calculating equivalent hydraulic resistance from a cold source side to a tail end key load side, extracting the maximum equivalent hydraulic resistance of each key load side under all the single-point fault conditions as a most unfavorable fault condition resistance value, and constructing a fault tolerance hydraulic bottleneck parameter set based on the most unfavorable fault condition resistance values of all the key load sides; Collecting real-time flow data of a system, calculating a fault guarantee pressure difference reference index by using the fault-tolerant hydraulic bottleneck parameter set, generating an input sequence containing time sequen