Bi-lstm-pinn-admm-based multi-objective dynamic collaborative optimization method for industrial circulating water system

Abstract

This invention discloses a multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM. Addressing the challenges of high energy consumption, strong coupling among multiple devices, and stringent safety constraints in industrial circulating water systems, this method constructs an optimization architecture of "global multi-objective programming + distributed collaborative iteration." The method utilizes Bi-LSTM combined with an attention mechanism to predict heat load and wet-bulb temperature; constructs a physical information neural network (PINNs) model based on mechanism and data fusion; employs the NSGA-III algorithm to solve for the Pareto optimal solution set for energy efficiency and safety at the global level; utilizes the alternating direction multiplier method (ADMM) to decouple the global objectives and distribute them to distributed device agents; and finally, achieves millisecond-level closed-loop command execution through nonlinear model predictive control (NMPC). This invention solves the problems of model inaccuracy, poor global coordination, and difficulty in balancing multi-objective conflicts in traditional control methods under non-stationary operating conditions, significantly improving the system's energy efficiency and operational stability.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control and energy management technology, and relates to a multi-objective dynamic collaborative optimization decision-making method for industrial circulating water systems that utilizes the integration of artificial intelligence and control theory. Specifically, it relates to a multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM. Background Technology

[0002] Industrial circulating water systems, acting as the "blood circulation system" in heavy industrial production processes such as power generation, chemical engineering, and metallurgy, encompass the entire heat transfer process, from heat absorption by industrial equipment, pipeline transportation, cooling tower heat dissipation, to water treatment. Their operational stability and heat exchange efficiency directly determine the safe and continuous production of core production equipment (such as reactors, turbine condensers, and blast furnaces) and the overall energy consumption level of the plant. However, in actual operation, industrial circulating water systems are prone to problems such as hydraulic imbalance, condensation pressure fluctuations, localized decreases in heat exchange efficiency, or persistently high energy consumption. Therefore, achieving global dynamic optimization and precise coordinated control of the circulating water system is of great significance for ensuring process safety and reducing operating costs.

[0003] Currently, control and optimization methods for circulating water systems mainly rely on traditional PID feedback control, manual empirical adjustment, or single-objective mathematical programming methods. However, these methods have the following significant shortcomings when facing the complex requirements of modern large-scale factories:

[0004] 1. The disconnect between physical mechanisms and data-driven approaches: Existing pure mechanism models often suffer from parameter calibration difficulties and model inaccuracies due to factors such as pipe scaling and equipment aging; while purely data-driven models (such as traditional neural networks) are prone to outputting control commands that violate fluid dynamics or the law of conservation of energy when faced with extreme working conditions with sparse data, lacking safety and interpretability for engineering applications.

[0005] 2. Disconnect between global coordination and local response: The components (pumps, valves, towers) within the system are usually controlled independently, lacking system-level overall planning and coordination. For example, in order to meet the pressure demand at the local end, the pumping station often adopts an oversaturated operation mode of "high flow rate and high head", resulting in huge throttling losses in the pipeline network and failing to resolve the contradiction between "oversized pumps for small loads" and local hydraulic imbalance.

[0006] 3. Lack of multi-objective dynamic trade-off capability: The system's operational objectives are inherently conflicting. For example, "reducing transmission energy consumption" often sacrifices "heat exchange safety margin." Existing methods struggle to calculate the Pareto optimal solution in real time under dynamically changing production loads and weather conditions, which balances the minimum total system power consumption with the safe operation of critical equipment, causing the system to deviate from its optimal operating range for extended periods.

[0007] Therefore, there is an urgent need for an optimization decision-making method that can deeply integrate physical mechanism constraints, possess global multi-objective optimization capabilities, and enable distributed device collaborative execution, in order to break through the bottlenecks of existing technologies in terms of security, collaboration, and energy efficiency. Summary of the Invention

[0008] This invention provides a multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM, which solves the problems of physical mechanism model parameter drift, lack of safety constraints in pure data-driven models, and system-level hydraulic imbalance and low energy efficiency caused by independent control of single equipment in existing circulating water system control methods.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] In a first aspect, this invention provides a multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM. This method constructs an optimization architecture of "global multi-objective programming + distributed collaborative iteration," and includes the following steps:

[0011] S1. A time series prediction model based on a bidirectional long short-term memory network and an attention-weighted mechanism is used to mine historical features of the collected operating parameters and environmental meteorological data of the circulating water system, thereby obtaining the system heat load demand sequence and the environmental wet-bulb temperature sequence within a preset future time period. The construction method of the time series prediction model based on the bidirectional long short-term memory network and the attention-weighted mechanism is as follows: the operating parameters and environmental meteorological data of the circulating water system, as well as the system heat load demand sequence and the environmental wet-bulb temperature sequence within a preset future time period (relative to the collection time of the operating parameters and environmental meteorological data) are collected; the operating parameters and environmental meteorological data of the circulating water system are used as input, and the system heat load demand sequence and the environmental wet-bulb temperature sequence within a preset future time period (relative to the collection time of the operating parameters and environmental meteorological data) are used as output to train the time series prediction model based on the bidirectional long short-term memory network and the attention-weighted mechanism (Bi-LSTM-Attention model).

[0012] S2. Establish a physical information neural network model (PINNs+GNN model) based on the fusion of mechanism and data. The main control equations of fluid mechanics and heat transfer are added to the loss function of the model as physical constraints. The model is trained by minimizing the physical residual and the data fitting error, so that the model still conforms to physical laws in the data sparse region.

[0013] S3. Input the system heat load demand sequence and the environmental wet-bulb temperature sequence obtained in step S1 into the physical information neural network model based on mechanism and data fusion established in step S2, perform system-level multi-objective global planning, and obtain the global operating setpoint.

[0014] S4. The global operating settings determined in step S3 are distributed and iteratively processed using the Alternating Direction Multiplier Method (ADMM). Specifically, the local load allocation command that satisfies both coupling constraints and local constraints and is optimal for each subsystem is calculated by iteratively exchanging boundary coupling variables among the device nodes. All subsystems together constitute the industrial circulating water system.

[0015] S5. Each subsystem's equipment nodes receive local load allocation commands and implement closed-loop rolling optimization using a nonlinear model predictive control strategy. Specifically, at each sampling time, a dynamic response state-space model of the pump-pipeline is constructed based on the pipeline pressure and instantaneous flow rate. Under the conditions of control magnitude constraints, control increment rate constraints, and state safety constraints, the optimal control increment sequence that minimizes the tracking error is solved. Only the first element of the optimal control increment sequence is issued as a control command. At the next sampling time, the optimal control increment sequence is refreshed using the latest pipeline pressure and instantaneous flow rate. The prediction time domain is shifted backward, and the above optimization process is repeated. This rolling feedback mechanism effectively compensates for model prediction errors and environmental disturbances.

[0016] In the above technical solution, further, in step S1, the real-time operating parameters of the circulating water system include the inlet and outlet water temperature difference and flow rate; the working method of the time series prediction model based on bidirectional long short-term memory network and attention weighting mechanism is as follows: First, the bidirectional long short-term memory network is used to extract bidirectional time series features from the real-time operating parameters of the circulating water system and environmental meteorological data to obtain the hidden state of the historical time step; then, the attention weighting mechanism is used to calculate the correlation weight between the hidden state of the historical time step and the current state, and key historical features are adaptively extracted to obtain the heat load sequence and wet-bulb temperature sequence for the next T time steps.

[0017] Further, in step S2, the training loss function of the physical information neural network model based on mechanism and data fusion is defined as:

[0018]

[0019] in, The mean square error between the output of the physical information neural network model based on mechanism and data fusion and the sensor measured data is given. The physical residuals that violate Bernoulli's equation or the law of conservation of energy are calculated based on the governing equations of fluid mechanics and heat transfer. The weighting coefficients are used to balance the weighting of data items and physical constraint items.

[0020] Further, in step S3, the specific operation of the system-level multi-objective global planning is as follows: taking the minimum total energy consumption of the system and the maximum heat exchange safety margin of key equipment as joint optimization objectives, performing multi-objective evolutionary search calculation to obtain the Pareto optimal solution set, thereby obtaining the global operating set value;

[0021] The joint optimization objective specifically includes:

[0022] Objective function 1: Minimize the total energy consumption of the system. Specifically, it is expressed as:

[0023]

[0024] in, This refers to the total shaft power of the pump set. The total shaft power of the cooling tower fan;

[0025] Objective function 2: Maximize the heat transfer safety margin of critical equipment. Specifically, it is expressed as:

[0026]

[0027] in, This is a collection of key heat exchangers for the entire plant; This is the upper limit of the process alarm temperature for the kth heat exchanger. The outlet temperature of the kth heat exchanger is calculated based on the physical model in step 2 and is related to the wet-bulb temperature.

[0028] Constraints: Pressure at the end of the pipeline network And total water supply flow ;

[0029] in, Minimum pressure limit at the end of the pipeline. This represents the minimum total traffic requirement.

[0030] The multi-objective evolutionary search computation is implemented based on the NSGA-III algorithm, which selects a unique global operating setting from the Pareto optimal solution set according to the current energy price weight.

[0031] The specific operation of step S4 is as follows:

[0032] Step 4.1: Establish a mathematical model for distributed load allocation, specifically expressed as follows:

[0033] (1) The objective function is to minimize the total energy consumption of the subsystems:

[0034]

[0035] in, The local flow load allocated to the i-th device. This represents the private characteristic parameters of the i-th device. Let N be the energy consumption cost function of the i-th water pump, and N represent the total number of devices. The global load allocation vector is composed of the local flow loads of all devices.

[0036] (2) Taking the global flow supply and demand balance as the global coupling constraint:

[0037]

[0038] in, Set a value for the total flow rate;

[0039] (3) Using the safe operating range of the equipment as a local constraint:

[0040] ;

[0041] in, Let be the lower limit of the allowable flow rate for the i-th device. Let i be the maximum allowed traffic volume for the i-th device;

[0042] Step 4.2: Construct the augmented Lagrangian function

[0043]

[0044] in, The local flow load allocated to the i-th device. This is the dual vector of the global coupling constraint. The parameter is the quadratic penalty term for the coupling constraint;

[0045] Step 4.3: Implement distributed cooperative iteration based on the alternating direction multiplier method

[0046] Treat each device controller as a computing node and perform the following iterative process:

[0047] (1) Local load autonomous optimization

[0048] ;

[0049] in, This represents an estimate of the traffic at other nodes; Indicates the optimal flow for node i;

[0050] (2) Broadcasting and aggregation of boundary coupled variables

[0051] Each node will broadcast the calculated optimal flow (i.e., boundary coupling variables) to its neighboring nodes or the virtual coordinator; and calculate the current global flow total deviation.

[0052] ;

[0053] in, This represents the global flow deviation at the (k+1)th iteration;

[0054] (3) Synchronous update of dual variables

[0055] Update the locally maintained dual variable based on the global traffic sum deviation:

[0056]

[0057] in, The dual variable used in the k-th iteration, The dual variable used in the (k+1)th iteration;

[0058] (4) Convergence determination

[0059] Examine the two residual indices: the original residual and the dual residual, where the original residual... Dual residuals are used to measure the degree to which physical constraints are satisfied. Used to measure the stability of an allocation scheme. and These are the global load allocation vectors for the (k+1)th and kth iterations, respectively; when and hour, The iteration ends when the preset tolerance is reached.

[0060] Step 4.4: Output and Execution

[0061] After the iteration converges, the optimal flow of all nodes calculated in the last iteration is summarized to obtain a set of optimal local load distribution instructions. ,in, This represents the optimal local flow load allocation value finally determined for the i-th device.

[0062] Further, in step S5, the closed-loop rolling optimization specifically includes:

[0063] (1) State sampling and feedback correction: The latest state variables of the system are obtained through sensors. The measured values ​​are used to correct the initial state (i.e., the predicted values) of the prediction model;

[0064] (2) Online rolling optimization: Based on the constructed nonlinear dynamic prediction model (i.e., the dynamic response state space model of the pump-pipeline network), under the premise of satisfying hard constraints such as the amplitude of the control variable and the rate of control increment, the future is solved. The optimal control increment sequence for each control step size ;

[0065] (3) First element execution mechanism: Only the first element in the optimal control increment sequence is extracted. As the actual control command, calculate the absolute control quantity. And it is sent to the underlying actuator (such as frequency converter) via PLC;

[0066] (4) Horizon shift and feedback correction: Enter the next control time k+1, shift the prediction time domain (prediction window) one step backward, and repeat the state sampling and optimization execution process of the above steps (1)-(3) to compensate for model prediction error and environmental disturbance.

[0067] Secondly, the present invention also provides a multi-objective dynamic collaborative optimization decision system for industrial circulating water systems. This system is used to execute the aforementioned multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM, and specifically includes the following modules:

[0068] The data acquisition and load forecasting module is used to execute step S1. It is responsible for collecting real-time operating parameters of the circulating water system and environmental meteorological data, and using a time series forecasting model based on bidirectional long short-term memory network and attention weighting mechanism to calculate the system heat load demand sequence and environmental wet-bulb temperature sequence for a future preset period.

[0069] The physical fusion state assessment module is used to execute step S2 and is responsible for running the physical information neural network model based on mechanism and data fusion.

[0070] The global optimization scheduling module, deployed in the global multi-objective planning layer, is used to execute the above step S3. It is responsible for inputting the system heat load demand sequence and the environmental wet-bulb temperature sequence within the future preset time period into the physical information neural network model based on mechanism and data fusion, and executing system-level multi-objective global planning to obtain the global operation setting value.

[0071] The distributed computing and allocation module, deployed in the distributed collaborative execution layer, is used to execute step S4. It is responsible for using the alternating direction multiplier method to perform distributed collaborative iterative optimization of the global running settings. Specifically, it calculates the local load allocation instruction that satisfies the global coupling constraints and local constraints and is optimal for each subsystem objective by iteratively exchanging boundary coupling variables between each device node.

[0072] A rolling optimization control module is implemented to execute step S5. It is responsible for solving the dynamic response state-space model of the pump-pipeline network at each sampling time, under the conditions of satisfying the control magnitude constraint, control increment rate constraint, and state safety constraint, to obtain the optimal control increment sequence that minimizes the tracking error. Only the first element in the optimal control increment sequence is issued as a control command, and the optimal control increment sequence is refreshed at the next sampling time using the latest pipeline pressure and instantaneous flow rate. The above optimization process is repeated after shifting the prediction time domain backward.

[0073] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect above.

[0074] Fourthly, the present invention also provides an electronic device comprising: one or more processors; and a memory for storing one or more programs;

[0075] When the one or more programs are executed by the one or more processors, the one or more processors perform the method described in the first aspect above.

[0076] The beneficial effects of this invention are as follows:

[0077] 1. This invention introduces a time series prediction model based on a bidirectional long short-term memory network and an attention weighting mechanism (Bi-LSTM-Attention), which enables a leap from "post-event passive adjustment" to "pre-event proactive perception", significantly improving the control accuracy under non-stationary operating conditions.

[0078] 2. This invention constructs a Physical Information Neural Network (PINNs) model, which incorporates the master equations of fluid mechanics and heat transfer as physical constraints into the model loss function, greatly improving the engineering safety of the system and the interpretability of the results.

[0079] 3. This invention employs the NSGA-III algorithm at the global multi-objective programming layer, using the minimum total system energy consumption and the maximum heat exchange safety margin of key equipment as joint optimization objectives for evolutionary search. This breaks the deadlock of the difficulty in balancing multi-objective conflicts and achieves Pareto optimality for "ultimate energy saving" and "process safety".

[0080] 4. This invention introduces the Alternating Directional Multiplier Method (ADMM), which decouples the global optimization problem and distributes it to each pump control node. Each node can achieve collaborative autonomous optimization by iteratively exchanging boundary coupling variables, thus eliminating the computational bottleneck and single-point failure risk of large-scale systems. Attached Figure Description

[0081] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0082] Figure 1 This is a schematic diagram of the structure of a dynamic collaborative optimization method for multi-objective industrial circulating water systems according to the present invention;

[0083] Figure 2 This is a schematic diagram of the process of a dynamic collaborative optimization method for multi-objective industrial circulating water systems according to the present invention. Detailed Implementation

[0084] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0085] Example 1

[0086] A multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM is proposed, and the implementation process is as follows: Figure 2 As shown.

[0087] This embodiment describes the application scenario of a public utility circulating water system in a large chemical industrial park. This system is responsible for cooling core production units such as the ethylene cracking unit and the aromatics complex. The system has a complex physical architecture, including four large variable frequency circulating water pumps operating in parallel (rated flow rate 12000 m³ / h). 3 / h, power 2500kW), 3 mechanical ventilation counterflow cooling towers (equipped with variable frequency fans), and 12 key heat exchangers (HE-01 to HE-12) distributed at the end of a long-distance pipeline network.

[0088] To address the problems of severe load fluctuations, hydraulic imbalance in the pipeline network, high energy consumption, and inaccuracies of traditional control models under extreme conditions in this system, this embodiment adopts the following... Figure 1 The two-layer optimization architecture shown, consisting of a "global multi-objective planning layer + distributed collaborative execution layer", achieves closed-loop optimization control of the system through the following five specific implementation steps.

[0089] Step 1: Multidimensional State Sensing and Dynamic Load Trend Prediction

[0090] The core of this step lies in overcoming the large hysteresis characteristic of the circulating water system. Traditional feedback control adjusts only based on the current deviation, often leading to system overshoot or oscillation. This embodiment introduces deep learning algorithms to mine the temporal characteristics of historical data, achieving "advanced perception" of future operating conditions.

[0091] Based on the real-time operating parameters of the circulating water system and environmental meteorological data, a time series prediction model based on a bidirectional long short-term memory network and an attention weighting mechanism is used to predict the system heat load demand sequence and the environmental wet-bulb temperature sequence for a future preset period.

[0092] The method for constructing a time series prediction model based on a bidirectional long short-term memory network and an attention-weighted mechanism is as follows:

[0093] 1) Data Acquisition and Cleaning

[0094] First, the distributed control system (DCS) collects three types of operational data at a frequency of once per minute: system operating parameters (including total water supply flow). Total return water pressure Water supply temperature Return water temperature Speed ​​of each water pump With power Meteorological data (dry bulb temperature) With relative humidity ) and heat load data (temperature difference between inlet and outlet of each heat exchanger) With the critical process medium outlet temperature Then, the raw data is preprocessed. For missing values ​​caused by sensor failure, interpolation is used to fill in the missing values. Outliers that exceed physical limits (such as temperature readings exceeding 100℃) are removed according to statistical rules. Finally, the data is standardized to improve quality.

[0095] 2) Calculation of wet-bulb temperature

[0096] wet-bulb temperature This represents the theoretical limit of the cooling tower's heat dissipation capacity and a key input for optimization decisions. This embodiment does not rely on expensive wet-bulb temperature sensors; instead, it utilizes collected dry-bulb temperature and relative humidity to perform high-precision soft-measurement using Stull's empirical formula:

[0097]

[0098] The wet-bulb temperature and the meteorological data collected in step 1.1 together constitute the environmental meteorological data.

[0099] 3) Construct a time series prediction model based on bidirectional long short-term memory network and attention weighting mechanism (Bi-LSTM-Attention prediction model)

[0100] The input to the bidirectional long short-term memory network is a multidimensional feature vector matrix of the past 24 hours (1440 time steps). D is the feature dimension, and its value is related to the types of real-time operating parameters and environmental meteorological data. A bidirectional LSTM is used to extract bidirectional time-series features from the real-time operating parameters and environmental meteorological data of the circulating water system, obtaining the hidden state at each historical time step. A forward LSTM captures the historical evolution trend, and a backward LSTM captures the inverse dependencies. For time step t, the hidden state h... tThe calculation is as follows:

[0101]

[0102] in This represents the input feature vector at time t; This represents the forward hidden state at time t. This represents the forward hidden state at time t-1. Indicates a backward hidden state. This represents the backward hidden state at time t-1. This represents vector concatenation. The LSTM unit internally uses a forget gate. Input gate and output gate This effectively solves the gradient vanishing problem in long sequence training. To highlight the impact of key historical moments on the future (e.g., load mutation patterns at the same time yesterday), an attention mechanism is introduced to calculate weights. :

[0103]

[0104] Where v T W represents the transpose of the weight vector. h Let b represent the weight matrix. h Denotes the bias vector, e t The original attention score at time step t, and the weighted context vector. Finally, the preset time period is output through a fully connected layer. Minutes of total system heat load sequence and wet-bulb temperature sequence .

[0105] 4) Use the Adam optimizer for offline training and periodically fine-tune it online to ensure prediction accuracy. This will give you a well-trained Bi-LSTM-Attention prediction model.

[0106] The mean squared error (MSE) is used as the loss function during training.

[0107]

[0108] in This represents the true value of the i-th sample (specifically, heat load or wet-bulb temperature data). This represents the model's prediction for the i-th sample.

[0109] Step 2: Establish a physical information neural network model based on the fusion of mechanism and data.

[0110] To address the challenges of complex pipeline networks in chemical industrial parks, the lack of sensor monitoring at some nodes, and the potential for pure data models to violate physical principles, this step involves constructing a Physical Information Neural Network (PINNs) model based on the fusion of mechanisms and data.

[0111] Step 2.1: Define physical constraint equations

[0112] The system operation must strictly follow the laws of fluid mechanics and thermodynamics, which constitute the regularization terms of the PINNs model.

[0113] (1) Nodal mass conservation (continuity equation): For any node i, the algebraic sum of the inflow and outflow mass flow rates is zero.

[0114]

[0115] in, This represents the mass conservation residual of the i-th node; Let q represent the set of all upstream neighbor nodes that flow to node i. ji This represents the water flow rate from upstream node j to the current node i. Let q represent the set of all downstream neighbor nodes flowing out of node i. ik This represents the water flow rate from the current node i to the downstream node k.

[0116] (2) Conservation of pipeline momentum (pressure drop equation): Pressure drop of fluid in pipe ij With traffic Satisfying the nonlinear drag relationship (using the Hazen-Williams formula):

[0117]

[0118] in, This represents the momentum conservation residual of pipe ij. This indicates the boost pressure provided by the water pump. Indicates the length of the pipe. Indicates the water flow rate in the pipe. Represents the roughness coefficient. Indicates the pipe diameter.

[0119] (3) Energy conservation equation: Heat exchange satisfies thermal equilibrium:

[0120]

[0121] in, This represents the energy conservation residual of the system. This indicates the specific heat capacity of water. This indicates the density of water. This indicates the total circulating water flow rate of the system. Indicates the total return water temperature of the system. Indicates the total water supply temperature of the system. The sum of the total heat loads of all production equipment (heat exchangers).

[0122] Step 2.2: Training of the Physical Information Neural Network Model (PINNs model) based on mechanism and data fusion

[0123] Construct a fully connected deep neural network, with boundary conditions (pump speed, valve opening) as input and the network state variables (pressure at each node) as output. Flow rate of each pipe section Define the loss function. :

[0124]

[0125] in, This indicates the number of observation points, that is, the total number of nodes and pipe sections where sensors are installed. This indicates that only the set of sensors is traversed. , This is the physical weighting coefficient (value 2.0), used to strengthen physical constraints. This represents the mass conservation residual of the j-th node. This represents the momentum conservation residual for the j-th pipe segment. Through training, the model not only fits the measured values ​​at sensor locations but also derives a reasonable pressure and flow distribution based on physical equations at sensorless locations, thus resolving the issues of uninterpretable and unsafe models.

[0126] Step 3: System-level multi-objective global planning

[0127] This step aims to predict the heat load demand sequence and the ambient wet-bulb temperature sequence based on the time series prediction model based on the bidirectional long short-term memory network and attention weighting mechanism in step 1, and the physical state mapping relationship constructed in step 2 (based on the physical information neural network model of mechanism and data fusion). By solving a multi-objective mathematical programming problem with strong nonlinearity and multivariate coupling characteristics, the optimal operating condition point of the system in the future control cycle is determined.

[0128] Step 3.1: Construct a multi-objective mathematical programming model for system operation

[0129] To address the inherent contradiction between "energy efficiency" and "process safety" in the operation of circulating water systems, the following multi-objective nonlinear programming model is established:

[0130] (1) Definition of decision variables

[0131] In the optimization problem, the decision variables include the setpoint of the total water supply flow of the circulating water system ( ), water supply main pipe pressure setting value ( ), and the operating frequency settings for each of the M cooling tower fans ( ).

[0132] (2) Establish the first objective function: total energy consumption of the system minimize

[0133] The goal is to reduce the overall system's power consumption cost. Based on the similarity laws of fluid machinery and the characteristics of motors, a total power calculation model is constructed:

[0134]

[0135] in, This refers to the total shaft power of the pump set. N represents the total shaft power of the cooling tower fan. P Let ρ be the total number of water pumps, ρ be the density of the circulating water, and g be the acceleration due to gravity. Let be the head of the i-th pump (i.e., the pressure head provided by the pump). Let be the hydraulic efficiency of the i-th pump. Let be the motor efficiency of the i-th water pump. Let M be the inverter efficiency of the i-th water pump, and M be the total number of cooling tower fans. The rated power of the j-th wind turbine, The actual operating frequency of the j-th wind turbine. This is the rated frequency. The specific calculation method is as follows:

[0136] Based on the fluid machinery similarity law and experimental data, the flow-head (QH) characteristic curve and flow-efficiency (Q-η) characteristic curve of each pump are polynomial-fitted using the least squares method. Specifically, for the i-th pump, its head function is fitted as a quadratic polynomial of the flow rate:

[0137]

[0138] Its hydraulic efficiency function is fitted to a cubic polynomial of the flow rate:

[0139]

[0140] in, and These are the fitting coefficients obtained through calibration using experimental data.

[0141] Meanwhile, considering inverter efficiency With operating frequency To address the nonlinear changes, a frequency conversion correction coefficient function is introduced:

[0142]

[0143] in The rated frequency (usually 50Hz). The polynomial fitting coefficients represent the efficiency curve of the frequency converter.

[0144] Finally, the formula for calculating the shaft power of a single water pump is:

[0145]

[0146] in, Let be the shaft power of the i-th water pump.

[0147] The total electrical power of all operating cooling tower fans follows the similarity laws of fluid machinery, where fan power is proportional to the cube of the rotational speed (or frequency). For the j-th cooling tower fan, its power calculation model is as follows:

[0148]

[0149] in, Let j be the rated shaft power of the j-th fan at the rated frequency. To optimize the given operating frequency command for the j-th wind turbine, Let be the motor efficiency of the j-th fan.

[0150] (3) Establish the second objective function: maximize the heat exchange safety margin J2 of the key equipment.

[0151] The goal is to improve the system's robustness in the face of operating condition fluctuations, and to define the "heat exchange safety margin" as the difference between the process alarm temperature of the heat exchanger and the outlet temperature of the heat exchanger.

[0152]

[0153] in This is a collection of key heat exchangers for the entire plant; This is the upper limit of the process alarm temperature for the kth heat exchanger. The outlet temperature of the kth heat exchanger, calculated using the physical information neural network model based on mechanism and data fusion in step 2, is related to the wet-bulb temperature.

[0154] (4) Constraints

[0155] Pipeline terminal pressure And total water supply flow ;

[0156] in, Minimum pressure limit at the end of the pipeline. This represents the minimum total traffic requirement.

[0157] Step 3.2: Optimization solution based on NSGA-III

[0158] Since the above model is a typical high-dimensional, non-convex, multimodal optimization problem, and there is a complex implicit coupling between the objective function J2 and the decision variables, the traditional gradient descent method is prone to getting trapped in local optima and cannot handle discrete variables. Therefore, this embodiment uses the reference point-guided non-dominated ranking evolution strategy NSGA-III to solve the problem.

[0159] (1) Population initialization and encoding

[0160] Using real number encoding, generate a number containing N pop The initial population P0 consists of x individuals (combinations of decision variables). Each individual x i This represents a potential system operating condition.

[0161] (2) Quick Non-Dominated Sort

[0162] Calculate the two objective function values ​​J1 and J2 for each individual in the population. Based on Pareto dominance, stratify the population into different frontiers F1, F2, ... . If individual p is not inferior to individual q on all objectives, and is superior to q on at least one objective, then p is said to dominate q. The F1 layer contains all undominated solutions, representing the current set of optimal trade-off solutions.

[0163] (3) Microhabitat preservation based on reference points

[0164] To ensure that the optimization results provide diverse decision options, the system pre-defines a set of uniformly distributed reference points. After generating these reference points in the standardized target space, each candidate solution is associated with the nearest reference point, and the number of solutions associated with each reference point is counted. Candidate solutions in regions with fewer associated solutions are retained first to prevent the algorithm from converging to an undesirable solution too early.

[0165] (4) Evolution operator operations

[0166] In the optimization algorithm, new solutions are generated by mimicking the crossover process of biological inheritance to maintain the diversity of the solution set. At the same time, parameters are randomly fine-tuned with low probability to enhance the fine search capability in local regions and help the algorithm escape the trap of suboptimal solutions.

[0167] (5) Decision-making and output

[0168] After multiple iterations, the algorithm outputs a set of non-dominated solutions (Pareto optimal solutions). The system calculates a weighted score based on the current "time-of-use electricity price" and "production priority".

[0169]

[0170] Where w cost As a cost weight, w safe For safety weights, The normalized total system energy consumption The normalized heat transfer safety margin for critical equipment is determined; after obtaining the optimal solution based on the score, it is used as the global setpoint at the current moment. and .

[0171] Step 4: Distributed Coordinated Load Allocation

[0172] To address the risks of single-point failures in traditional centralized distribution methods and the efficiency mismatch caused by differences in characteristics among multiple parallel devices, this embodiment adopts a distributed collaborative computing architecture based on the Alternating Direction Multiplier Method (ADMM).

[0173] Step 4.1: Establish a mathematical model for distributed load distribution

[0174] The N pumps operating in parallel are considered as N distributed intelligent nodes with independent computing capabilities. The global load allocation problem is modeled as the following separable convex optimization problem:

[0175] (1) Subsystem objective function: Minimize the total energy consumption of the subsystem

[0176]

[0177] in, The local flow load allocated to the i-th water pump. This represents the private characteristic parameters of the i-th water pump. Let N be the energy consumption cost function of the i-th water pump, and let N represent the total number of water pumps. The global load allocation vector is composed of the local flow loads of all devices.

[0178] (2) Global coupling constraint: global flow supply and demand balance

[0179] The sum of the local traffic of all distributed nodes must be exactly equal to the total traffic setting value issued by the global planning layer. :

[0180]

[0181] (3) Local constraints: safe operating range of equipment

[0182] Each piece of equipment must operate within its respective flow range to avoid surge or cavitation:

[0183]

[0184] in, Let be the lower limit of the allowable flow rate for the i-th device. Let i be the maximum allowed traffic volume for the i-th device;

[0185] Step 4.2: Construct the augmented Lagrangian function

[0186] For step 4.1 which contains global coupling constraints ( The optimization problem of [the problem] cannot be solved directly in a distributed manner due to the coupling of variables among the devices. To decouple the problem, dual variables (Lagrange multipliers) are introduced. and secondary penalty term parameters If the value is greater than 0, the global coupling constraint is moved into the objective function as a penalty term to construct the augmented Lagrangian function:

[0187]

[0188] By constructing this function, the original global optimization problem with strong coupling constraints is transformed into an unconstrained subproblem that can be solved independently by each device node. This function serves as the basic objective function for subsequent iterative calculations using the Alternating Direction Multiplier Method (ADMM), enabling the complex global optimization process to be decomposed into multiple parallel local optimization calculations and dual variable update processes.

[0189] Step 4.3: Distributed Cooperative Iteration

[0190] Using the ADMM algorithm framework, the complex global optimization problem is decomposed into N parallel local subproblems and a global aggregation problem, and collaboration is achieved through iterative communication between nodes.

[0191] Treating each pump controller as a computing node, the following iterative process is executed based on the ADMM algorithm:

[0192] (1) Local load autonomous optimization

[0193] Each node i receives the flow load calculated in the previous step. Independently solve for its own optimal flow rate :

[0194]

[0195] in This is an estimate of the traffic at other nodes. This step involves very little computation and can be completed in milliseconds within the edge controller.

[0196] (2) Broadcasting and aggregation of boundary coupled variables

[0197] Each node will calculate the optimal flow (i.e., the boundary coupling variables). Broadcast to neighboring nodes (or the virtual coordinator) via Industrial Ethernet. Calculate the current global traffic summation deviation:

[0198]

[0199] in, This represents the global flow deviation at the (k+1)th iteration;

[0200] (3) Synchronous update of dual variables

[0201] Each node updates its locally maintained dual variable based on the global traffic sum deviation:

[0202]

[0203] in, The dual variable used in the k-th iteration, The dual variable used in the (k+1)th iteration;

[0204] (4) Convergence determination

[0205] Examine the two residual indices: the original residual and the dual residual, where the original residual... Dual residuals are used to measure the degree to which physical constraints are satisfied. Used to measure the stability of an allocation scheme, where and These are the global load allocation vectors for the (k+1)th and kth iterations, respectively. Let represent the Euclidean norm. When and When the iteration ends, This is the preset tolerance.

[0206] Step 4.4: Output and Execution

[0207] After the iteration converges, the optimal flow of each node calculated in the last iteration is summarized to obtain a set of optimal local load distribution instructions. ,in, This represents the optimal local flow load allocation value finally determined for the i-th device. This scheme has the following engineering characteristics: the marginal energy cost of all pumps is equal, and the total system energy consumption is minimized; the total flow demand is strictly met and no pump exceeds its limit; if a pump suddenly fails and goes offline, the remaining nodes will automatically sense the flow gap and automatically share the load of the failed pump through ADMM iteration, achieving self-healing.

[0208] Through this algorithm, the system achieves globally optimal collaboration among multiple pumps without uploading all device curves to the cloud (protecting privacy / reducing bandwidth), automatically balancing the efficiency differences between new and old pumps.

[0209] Step 5: Implement closed-loop control with rolling time-domain optimization

[0210] This step aims to allocate the traffic load. Converted into frequency command for inverter Unlike traditional PID control, which only adjusts the current error after the fact, this embodiment uses a nonlinear model predictive control (NMPC) strategy, which can predict the future dynamics of the system in a forward-looking manner. Under the premise of explicitly handling hard constraints (such as water hammer protection and motor safety range), it calculates the optimal control trajectory and realizes a smooth and accurate conversion from flow target to frequency command.

[0211] Step 5.1: Construct the objective function of the nonlinear dynamic prediction model

[0212] The nonlinear dynamic prediction model is specifically a dynamic response state-space model of the pump-pipeline network, which is expressed as follows:

[0213]

[0214] in and These represent the state variables at times k and k+1, respectively. The state variables include pipeline pressure and instantaneous flow rate. Let k be the control variable at time k, and the control variable is specifically the inverter frequency.

[0215] The solution method for the dynamic response state-space model of the pump-pipeline network is as follows:

[0216] At each control time k, the local flow setpoint issued in step 4 is tracked. The primary objective is to suppress drastic fluctuations in the control quantity, with the secondary objective being to achieve this within a finite prediction time domain. Internally solve the following optimization problem:

[0217]

[0218] The first item is the tracking error item, which is used to ensure the actual traffic. Approaching the set value quickly and without steady-state error The second item is a penalty for maintaining stability. This represents the change in frequency, used to prevent excessively rapid adjustments from causing network oscillations; Q and R are the corresponding weight matrices, respectively. To predict the time domain, To control the time domain.

[0219] Step 5.2: Mathematical Modeling and Processing of Hard Constraints in Complex Engineering Projects

[0220] In industrial settings, to ensure the lifespan of actuators and the physical safety of pipeline networks, the controller output must be strictly limited within the permissible feasible range. Unlike traditional methods that rely on external limiting circuits, this method directly embeds the following three types of constraints into the optimization solution process of the nonlinear dynamic prediction model:

[0221] (1) Control variable amplitude constraint

[0222] Limiting the operating frequency of the variable frequency drive Always keep the equipment within its allowed high-efficiency and safe range to prevent motor overload or pump surge:

[0223]

[0224] in, and These are the upper and lower limits of the equipment's permissible high-efficiency and safe range, respectively.

[0225] (2) Controlling incremental rate constraints

[0226] Limit the frequency variation within adjacent control cycles To avoid water hammer or motor current surges caused by excessively rapid adjustments:

[0227]

[0228] in, Specifically, it indicates the maximum allowable frequency variation of the system within a single control cycle;

[0229] (3) State safety constraints

[0230] The system's critical state variables (such as pump outlet pressure) must be limited to ensure they do not exceed the pipeline's pressure limit. This constraint is typically an implicit function of the control quantity. :

[0231]

[0232] Step 5.3: Execution logic of closed-loop rolling optimization

[0233] A rolling time-domain strategy of "prediction-optimization-execution-correction" is adopted to transform open-loop optimization into closed-loop feedback control, with the specific loop as follows:

[0234] (1) State sampling and feedback correction: At the current time k, the latest state variables of the system are obtained through the sensor. The measured values ​​are used to correct the initial state (predicted value) predicted by the model, thereby eliminating the cumulative error caused by model mismatch or environmental disturbances (such as the drift of pipeline resistance coefficient over time).

[0235] (2) Online rolling optimization: Based on the objective function of the nonlinear dynamic prediction model constructed in 5.1, under the premise of satisfying all the hard constraints of complex engineering described in 5.2, the future is solved. Optimal control increment sequence for each control step:

[0236]

[0237] This sequence represents the optimal operation path within a future window, from the current perspective. This means calculating the optimal frequency change amplitude at the current time k and the future time k+j.

[0238] (3) First element execution mechanism: Based on the rolling optimization principle, only the first element in the optimal control increment sequence is extracted. As the actual control command, calculate the absolute control quantity. The command is then sent from the PLC to the frequency converter for execution.

[0239] (4) Horizon shift and feedback correction: Enter the next time step k+1, shift the prediction window one step backward, and repeat the above "sampling-prediction-optimization-execution" steps. This rolling mechanism of "taking one step, looking one step, and correcting one step" enables the system to dynamically adjust the control strategy in real time when facing nonlinear operating condition fluctuations, and has extremely strong robustness and anti-interference ability compared with traditional fixed parameter control.

[0240] Example 2

[0241] A multi-objective dynamic collaborative optimization decision system for an industrial circulating water system is provided, used to execute the multi-objective dynamic collaborative optimization method for an industrial circulating water system based on Bi-LSTM-PINNs-ADMM described in Example 1. The system specifically includes:

[0242] The data acquisition and load forecasting module is used to perform step 1. It is responsible for collecting real-time operating parameters of the circulating water system and environmental meteorological data, and using a time series forecasting model based on bidirectional long short-term memory network and attention weighting mechanism to calculate the system heat load demand sequence and environmental wet-bulb temperature sequence for a future preset period.

[0243] The physical fusion state assessment module is used to perform step 2 and is responsible for running the physical information neural network model based on mechanism and data fusion.

[0244] The global optimization scheduling module, deployed in the global multi-objective planning layer, is used to execute step 3. It is responsible for inputting the system heat load demand sequence and the environmental wet-bulb temperature sequence within the future preset time period into the physical information neural network model based on mechanism and data fusion, and executing system-level multi-objective global planning to obtain the global operating setpoint.

[0245] The distributed computing and allocation module, deployed in the distributed collaborative execution layer, is used to execute step 4. It is responsible for using the alternating direction multiplier method to perform distributed collaborative iterative optimization of the global running settings. Specifically, it calculates the local load allocation instruction that satisfies the global coupling constraints and local constraints and is optimal for each subsystem objective by iteratively exchanging boundary coupling variables between each device node.

[0246] A rolling optimization control module is implemented to execute step 5. It is responsible for solving the dynamic response state-space model of the pump-pipeline network at each sampling time, under the conditions of satisfying the control magnitude constraint, control increment rate constraint, and state safety constraint, to obtain the optimal control increment sequence that minimizes the tracking error. Only the first element in the optimal control increment sequence is issued as a control command, and the optimal control increment sequence is refreshed at the next sampling time using the latest pipeline pressure and instantaneous flow rate. The above optimization process is repeated after shifting the prediction time domain backward.

[0247] Example 3

[0248] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM described in Example 1.

[0249] Example 4

[0250] An electronic device comprising:

[0251] One or more processors;

[0252] Memory, used to store one or more programs;

[0253] When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in Embodiment 1.

[0254] experiment:

[0255] 1. To verify the effectiveness and advancement of the method proposed in this invention, a hardware-in-the-loop simulation experiment was conducted on the circulating water system of a large chemical industrial park described in Example 1. The experimental data came from the historical operation records of the system's DCS, covering two typical operating conditions: high temperature and high load in summer and medium load in spring and autumn, with a time span of 14 days and a data sampling interval of 1 minute. The system configuration parameters are as follows: 4 x 2500kW variable frequency circulating water pumps (rated flow rate 12000 m³ / h). 3The system consists of three mechanically ventilated cooling towers ( / h), with 12 key heat exchangers connected at the end. The simulation platform is built using Python, employing PyTorch to implement the Bi-LSTM-Attention prediction model and the PINNs state evaluation model, and utilizing MATLAB / Simulink to construct the dynamic environment of the fluid network. The experiment aims to evaluate the system's performance in terms of energy efficiency, temperature control stability, and computational real-time performance.

[0256] 2. Ablation test

[0257] To verify the contribution of key modules (bidirectional temporal attention prediction, physical information neural networks (PINNs), and distributed collaborative mechanism) to the overall performance of this invention, the following ablation experiment was designed. A high-fluctuation operating condition segment (24 hours) was selected from the test set, and the following variant models were run. The average operating power (kW), heat exchanger temperature violation rate (%, indicating the percentage of time exceeding the process alarm temperature), and average computation time (ms) of the system were statistically analyzed. The experimental results are shown in Table 1.

[0258] Table 1 Results of the ablation experiment

[0259]

[0260] The Base-Model serves as a baseline for subsequent complex models. It consists of a basic Long Short-Term Memory (LSTM) network and a centralized single-objective optimization. It utilizes only the LSTM to mine the temporal dependencies of historical data for simple load forecasting and schedules the system with the single objective of minimizing energy consumption.

[0261] Model-A: This model introduces a "Bidirectional Long Short-Term Memory Network (Bi-LSTM) and an attention mechanism" on top of the base model. By capturing the bidirectional dependencies of time-series data through the bidirectional structure, the attention mechanism adaptively weights key historical moments (such as sudden load fluctuations), thus enhancing the model's ability to predict heat load and wet-bulb temperature under non-stationary operating conditions compared to the Base-Model.

[0262] Model-B: Building upon Model-A, it introduces "Physical Information Neural Networks (PINNs)". It constructs a state assessment model based on the fusion of mechanism and data, incorporating fluid mechanics (such as Bernoulli's equation) and the laws of heat transfer as physical constraints into the training loss function. By minimizing the physical residuals, it ensures that the model can still output state estimates that conform to physical conservation laws in sparse data regions, thus improving safety compared to a pure data model.

[0263] Model-C: Based on Model-B, it introduces the "Non-Dominated Sorting Genetic Algorithm (NSGA-III)". It constructs a global multi-objective optimization architecture, aiming to find the Pareto optimal solution set with the lowest energy consumption and the largest equipment safety margin at the same time. It solves the problem of single-objective optimization losing sight of one aspect, but still retains the centralized computing architecture and does not perform distributed decoupling.

[0264] This invention presents a complete model that integrates "Bi-LSTM-Attention, PINNs, NSGA-III" and "Distributed Cooperative Execution Layer (ADMM+NMPC)". Based on global multi-objective programming, it utilizes the Alternating Direction Multiplier (ADMM) method to decouple and distribute computational tasks to each distributed agent, and achieves rolling temporal optimization through Nonlinear Model Predictive Control (NMPC).

[0265] The results show that the introduction of Bi-LSTM and attention mechanisms significantly improves prediction accuracy (reduces MAPE), while the addition of PINNs (Physical Information Neural Networks) drastically reduces the heat exchanger temperature violation rate, demonstrating the importance of mechanistic constraints in ensuring system safety. Furthermore, compared to the centralized computation of Model-C, the Alternating Directional Multiplier Method (ADMM) introduced in this invention compresses the computation time from 1200ms to 250ms, effectively solving the computational bottleneck of real-time optimization of large-scale complex systems and achieving a perfect balance between energy efficiency, safety, and response speed.

[0266] 3. Comparative Experiment

[0267] To further verify the comprehensive advantages of this invention in practical industrial applications, a comparative experiment was conducted using mainstream control strategies currently employed in industry. The experiment simulated continuous operation under varying conditions for 7 days (168 hours), comparing the performance of each method in terms of total system energy saving (relative to power frequency operation), temperature control variance of key equipment (reflecting stability), and hydraulic misalignment coefficient. The results are shown in Table 2.

[0268] Table 2 Results of the comparative experiment

[0269]

[0270] Among them, manual experience control is the traditional operating mode. It relies on "PID feedback control" and the operator's experience for adjustment. The set values ​​of pump frequency and fan frequency are fixed or rarely change, making it difficult to adapt to real-time fluctuations in operating conditions.

[0271] Traditional RTO: Real-time optimization method based on steady-state models. It usually optimizes the setpoint on an hourly basis based on steady-state mechanism models, which does not adequately consider the lag of dynamic processes and has a low update frequency.

[0272] Centralized MPC: This includes a centralized solver and model predictive control (MPC). It solves for the control variables of all devices in a unified manner. Although it can handle multivariable constraints, the computational load is too high in large-scale systems, resulting in high response delays.

[0273] In summary, the experimental results fully demonstrate the effectiveness of the global multi-objective programming and distributed cooperative iterative optimization of this invention, outperforming existing mainstream methods on all key performance indicators. This verifies the innovation and practical value of this invention, which can be used to solve:

[0274] 1. The problem of dynamic trade-offs among multiple objectives

[0275] Traditional RTOs or manual control often suffer from trade-offs, making it difficult to ensure the temperature control safety margin of critical equipment while reducing energy consumption. This invention achieves Pareto optimal balance.

[0276] 2. Real-time response issues in large-scale systems

[0277] Centralized MPC takes too long to compute when dealing with complex pipeline networks, and cannot meet the millisecond-level control requirements. The distributed collaborative mechanism (ADMM) of this invention significantly reduces response latency.

[0278] 3. Model mismatch and robustness issues

[0279] Human experience and traditional models are insufficient to cope with nonlinear changes under extreme operating conditions. This invention integrates physical mechanisms and data-driven approaches (PINNs) to ensure control accuracy and stability under all operating conditions.

[0280] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

[0281] The embodiments described above are merely some preferred embodiments of the present invention, and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM, characterized in that, Includes the following steps: S1. Based on the real-time operating parameters of the circulating water system and environmental meteorological data, a time series prediction model based on a bidirectional long short-term memory network and an attention weighting mechanism is used to predict the system heat load demand sequence and the environmental wet-bulb temperature sequence within a preset future period. S2. The master equations of fluid mechanics and heat transfer are added as physical constraints to the loss function of the physical information neural network model based on mechanism and data fusion. The model is trained by minimizing the physical residual and the data fitting error. S3. Input the system heat load demand sequence and the ambient wet-bulb temperature sequence obtained in step S1 into the physical information neural network model based on mechanism and data fusion, perform system-level multi-objective global planning, and obtain the global operating setpoint. S4. The global operating settings determined in step S3 are optimized by distributed collaborative iteration using the alternating direction multiplier method to calculate the local load allocation command that satisfies both global coupling constraints and local constraints and is optimal for the objectives of each subsystem. S5. Each subsystem's equipment node receives a local load allocation command and implements closed-loop rolling optimization using a nonlinear model predictive control strategy. During the closed-loop rolling optimization process, only the first element in the optimal control increment sequence is used as the control command for issuance.

2. The multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in claim 1, characterized in that, In step S1, the real-time operating parameters of the circulating water system include the inlet and outlet water temperature difference and flow rate. The working method of the time series prediction model based on bidirectional long short-term memory network and attention weighting mechanism is as follows: First, the bidirectional long short-term memory network is used to extract bidirectional time series features from the real-time operating parameters of the circulating water system and environmental meteorological data to obtain the hidden state of the historical time step. Then, the attention weighting mechanism is used to calculate the correlation weight between the hidden state of the historical time step and the current state, and key historical features are adaptively extracted to obtain the heat load sequence and wet-bulb temperature sequence for the next T time steps.

3. The multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in claim 1, characterized in that, In step S2, the training loss function of the physical information neural network model based on mechanism and data fusion is defined as: in, The mean square error between the output of the physical information neural network model based on mechanism and data fusion and the sensor measured data is given. The physical residuals that violate Bernoulli's equation or the law of conservation of energy are calculated based on the governing equations of fluid mechanics and heat transfer. The weighting coefficients are used to balance the weighting of data items and physical constraint items.

4. The multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in claim 1, characterized in that, In step S3, the specific operation of the system-level multi-objective global planning is as follows: taking the minimum total energy consumption of the system and the maximum heat exchange safety margin of key equipment as joint optimization objectives, performing multi-objective evolutionary search calculation, obtaining the Pareto optimal solution set, and thus obtaining the global operating set value; The joint optimization objective specifically includes: Objective function 1: Minimize the total energy consumption of the system. Specifically, it is expressed as: in, This refers to the total shaft power of the pump set. The total shaft power of the cooling tower fan; Objective function 2: Maximize the heat transfer safety margin of critical equipment. Specifically, it is expressed as: in, This is a collection of key heat exchangers for the entire plant; This is the upper limit of the process alarm temperature for the kth heat exchanger. The outlet temperature of the kth heat exchanger is calculated based on the physical model in step 2 and is related to the wet-bulb temperature. Constraints: Pressure at the end of the pipeline network And total water supply flow ; in, Minimum pressure limit at the end of the pipeline. This represents the minimum total traffic requirement. The multi-objective evolutionary search computation is implemented based on the NSGA-III algorithm, which selects a unique global operating setting from the Pareto optimal solution set according to the current energy price weight.

5. The multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in claim 1, characterized in that, The specific operation of step S4 is as follows: Step 4.1: Establish a mathematical model for distributed load allocation, specifically expressed as follows: (1) The objective function is to minimize the total energy consumption of the subsystems: in, The local flow load allocated to the i-th device. This represents the private characteristic parameters of the i-th device. Let N be the energy consumption cost function of the i-th water pump, and N represent the total number of devices. The global load allocation vector is composed of the local flow loads of all devices. (2) Taking the global flow supply and demand balance as the global coupling constraint: in, Set a value for the total flow rate; (3) Using the safe operating range of the equipment as a local constraint: ； in, Let be the lower limit of the allowable flow rate for the i-th device. Let i be the maximum allowed traffic volume for the i-th device; Step 4.2: Construct the augmented Lagrangian function in, The local flow load allocated to the i-th device. This is the dual vector of the global coupling constraint. The parameter is the quadratic penalty term for the coupling constraint; Step 4.3: Implement distributed cooperative iteration based on the alternating direction multiplier method Treat each device controller as a computing node and perform the following iterative process: (1) Local load autonomous optimization ； in, This represents an estimate of the traffic at other nodes; Indicates the optimal flow for node i; (2) Broadcasting and aggregation of boundary coupled variables Each node broadcasts its calculated optimal flow, i.e., the boundary coupling variables, to its neighboring nodes or the virtual coordinator; it also calculates the current global flow total deviation. ； in, This represents the global flow deviation at the (k+1)th iteration; (3) Synchronous update of dual variables Update the locally maintained dual variable based on the global traffic sum deviation: in, The dual variable used in the k-th iteration, The dual variable used in the (k+1)th iteration; (4) Convergence determination Examine the two residual indices: the original residual and the dual residual, where the original residual... Dual residuals are used to measure the degree to which physical constraints are satisfied. Used to measure the stability of an allocation scheme. and These are the global load allocation vectors for the (k+1)th and kth iterations, respectively; when and hour, The iteration ends when the preset tolerance is reached. Step 4.4: Output and Execution After the iteration converges, the optimal flow of all nodes calculated in the last iteration is summarized to obtain a set of optimal local load distribution instructions. ,in, This represents the optimal local flow load allocation value finally determined for the i-th device.

6. The multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in claim 1, characterized in that, In step S5, the nonlinear model predictive control strategy is used to implement closed-loop rolling optimization. Specifically, the following method is employed: a dynamic response state-space model of the pump-pipeline is constructed based on the pipeline pressure and instantaneous flow rate. At each sampling time, under the conditions of satisfying the control magnitude constraint, control increment rate constraint, and state safety constraint, the dynamic response state-space model of the pump-pipeline is solved to obtain the optimal control increment sequence that minimizes the tracking error. Only the first element in the optimal control increment sequence is used as the control command and issued. At the next sampling time, the optimal control increment sequence is refreshed using the latest pipeline pressure and instantaneous flow rate. The prediction time domain is then shifted backward, and the above optimization process is repeated.

7. The multi-objective dynamic collaborative optimization method for industrial circulating water systems based on Bi-LSTM-PINNs-ADMM as described in claim 6, wherein the closed-loop rolling optimization specifically comprises: (1) At the current time k, obtain the latest state variables of the system through the sensor. The predicted values ​​are corrected using the measured values; (2) Solving the future based on a nonlinear dynamic prediction model Optimal control increment sequence for each control step: in, This means calculating the optimal frequency change amplitude at the current time k+j in the future; (3) Extract only the first element in the optimal control increment sequence. As the actual control command, calculate the absolute control quantity. And the data is sent to the frequency converter via the PLC; (4) Enter the next time k+1, shift the prediction window one step backward, and repeat the above steps (1)-(3).

8. A multi-objective dynamic collaborative optimization decision-making system for an industrial circulating water system, characterized in that, For performing the method as described in any one of claims 1-6, the system comprises: The data acquisition and load forecasting module is used to execute step S1. It is responsible for collecting real-time operating parameters of the circulating water system and environmental meteorological data, and using a time series forecasting model based on bidirectional long short-term memory network and attention weighting mechanism to calculate the system heat load demand sequence and environmental wet-bulb temperature sequence for a future preset period. The physical fusion state assessment module is used to execute step S2 and is responsible for running the physical information neural network model based on mechanism and data fusion. The global optimization scheduling module, deployed in the global multi-objective planning layer, is used to execute the above step S3. It is responsible for inputting the system heat load demand sequence and the environmental wet-bulb temperature sequence within the future preset time period into the physical information neural network model based on mechanism and data fusion, and executing system-level multi-objective global planning to obtain the global operation setting value. The distributed computing and allocation module, deployed in the distributed collaborative execution layer, is used to execute step S4. It is responsible for using the alternating direction multiplier method to perform distributed collaborative iterative optimization of the global running settings. Specifically, it calculates the local load allocation instruction that satisfies the global coupling constraints and local constraints and is optimal for each subsystem objective by iteratively exchanging boundary coupling variables between each device node. A rolling optimization control module is implemented to execute step S5. It is responsible for solving the dynamic response state-space model of the pump-pipeline network at each sampling time, under the conditions of satisfying the control magnitude constraint, control increment rate constraint, and state safety constraint, to obtain the optimal control increment sequence that minimizes the tracking error. Only the first element in the optimal control increment sequence is issued as a control command, and the optimal control increment sequence is refreshed at the next sampling time using the latest pipeline pressure and instantaneous flow rate. The above optimization process is repeated after shifting the prediction time domain backward.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method steps as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

Images