Layered intelligent optimization decision-making method for circulating water system guided by large model

Abstract

The invention relates to the field of intelligent control, and aims to provide a large-model-guided hierarchical intelligent optimization decision-making method for a circulating water system. The method comprises the steps of calling a large model to generate a long-term operation strategy and a short-term operation strategy based on collected data, a set target and prediction information; after verification, establishing a multi-objective optimization model, solving in real time by adopting an optimization algorithm, and outputting an optimal control parameter combination; the optimal control parameter combination is converted into a control instruction, and the control instruction is issued to an execution mechanism; and analyzing and optimizing deviation between model output and actual response through a self-learning algorithm, and dynamically correcting control parameters output by the model on line. According to the method, the hierarchical reasoning capability and the deep learning advantage of a large model are utilized, intelligent collaborative decision making is performed on multiple time scales, and the defects that a traditional method is not intelligent in scheduling and difficult to meet multi-objective optimization at the same time are effectively overcome; the accuracy of the model and control is continuously improved through closed-loop feedback, and the energy-saving and operation efficiency of the circulating water system is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial circulating water system operation optimization and intelligent control, specifically to a method, system, and storage medium for hierarchical decision-making and optimization control using a large-scale artificial intelligence model. Background Technology

[0002] Circulating water systems (typically referring to industrial process cooling systems) are crucial infrastructure in many industrial plants and power plants, used for the recovery and recycling of cooling water. Traditional circulating water systems rely heavily on manual experience or simple rule-based algorithms for scheduling and control, resulting in low levels of intelligence and insufficient energy optimization.

[0003] In actual production, circulating water systems often operate under partial load. However, due to a lack of effective theoretical calculation support and refined management measures, enterprises frequently adopt full-load operation, leading to significant energy waste. Existing technologies, lacking theoretical calculation models and overall management systems, struggle to achieve refined operation management, further contributing to energy waste. Furthermore, while some optimization schemes can provide optimal water network scheduling, they fail to consider the comprehensive energy consumption of equipment such as cooling towers and chillers, lacking global optimization of key equipment energy consumption. Other schemes improve energy efficiency through modifications and upgrades, but fail to fully utilize historical operating data and real-time parameters, and lack mathematical model support, making them unable to cope with the impact of parameter uncertainties. These problems result in unintelligent scheduling strategies and insufficient energy consumption optimization in existing circulating water systems. Meanwhile, in fields such as the energy internet and microgrids, multi-timescale optimized scheduling has proven effective in improving system economy and renewable energy utilization. However, traditional circulating water systems generally lack the ability to coordinate and integrate multi-timescale and multi-level strategies.

[0004] In summary, existing circulating water system control technologies are insufficient for achieving intelligent scheduling, multi-objective coordination, and multi-timescale optimized operation, necessitating an innovative technical solution. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a hierarchical intelligent optimization decision-making method for circulating water systems guided by a large model.

[0006] To solve the technical problem, the solution of the present invention is:

[0007] A hierarchical intelligent optimization decision-making method for circulating water systems guided by a large model is provided, comprising the following steps:

[0008] S1. Collect historical and real-time operating status data of the circulating water system, and generate long-term operating strategies based on the set targets and prediction information by calling the large model.

[0009] S2. Based on the long-term operation strategy and current load information, call the large model to generate a short-term operation strategy; use the simulation model to evaluate whether the short-term operation strategy meets process constraints and safety requirements; if the verification fails, trigger the safety rollback mechanism.

[0010] S3. Based on the verified short-term operation strategy, establish a multi-objective optimization model, use an optimization algorithm to solve the optimization model in real time, and output the optimal control parameter combination; convert the optimal control parameter combination into control commands and send them to the actuators in the circulating water system for execution;

[0011] S4. Continuously collect system operation data after execution and compare it with the set optimization target; analyze the deviation between the output of the optimization model and the actual response through self-learning algorithm, dynamically correct the objective function weights, constraint parameters and structure in the optimization model, and correct the control parameters of the model output online.

[0012] Description of the invention principle:

[0013] The core innovation of this invention is the introduction of a large-scale artificial intelligence model (referred to as the "large model") into the operation and control of a circulating water system. Through hierarchical decision-making and closed-loop optimization, it achieves comprehensive optimization of system energy consumption, water consumption, and operational stability. The large model can be a large-scale language model or other deep learning model, used to guide the circulating water system in collaborative optimization scheduling and decision-making at the long-term, short-term, and control levels, thereby achieving efficient energy-saving operation and intelligent management.

[0014] Specifically, this invention establishes a hierarchical framework for long-term decision-making, short-term decision-making, and control decision-making: at the long-term decision-making level, based on predictive information and high-level objectives, a large model is used to conduct long-term operation planning and generate a global scheduling strategy for long-term operation; at the short-term decision-making level, based on the long-term operation strategy and real-time load information, the large model is invoked to infer short-term scheduling schemes; at the control level, a multi-objective optimization model is constructed based on the verified short-term operation strategy, and specific control parameters and instructions are obtained in real time.

[0015] Meanwhile, the invention performs simulation verification on the generated scheduling strategy before execution to ensure safety and feasibility. If the verification fails, a rollback mechanism is triggered to adopt a backup safety scheme to ensure stable system operation. During execution, a closed-loop feedback and self-learning mechanism are introduced into the system to continuously collect running data and compare it with the optimization target. When the execution effect is detected to deviate from the expectation, the parameters of the large model or the optimization model are automatically corrected through the self-learning algorithm, and the weights and constraints of the optimization objective function are dynamically updated to achieve online adaptive control.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] 1. This invention utilizes the hierarchical reasoning capabilities of large models and the advantages of deep learning to perform intelligent collaborative decision-making across multiple time scales, effectively overcoming the shortcomings of traditional methods such as unintelligent scheduling and difficulty in simultaneously satisfying multiple objectives.

[0018] 2. This invention ensures operational safety through execution verification and rollback mechanisms, and continuously improves the accuracy of the model and control through closed-loop feedback, significantly enhancing the energy saving and operational efficiency of the circulating water system.

[0019] 3. This invention utilizes a large model to learn the system's operating rules from historical and real-time data, providing strong cognitive support and cross-level planning capabilities for the decision-making process, making scheduling strategies more intelligent and efficient.

[0020] 4. This invention adopts a closed-loop feedback and self-learning mechanism to continuously optimize model parameters based on actual operating results, thereby improving the robustness and adaptability of the system. At the same time, the verification and rollback mechanisms ensure the safety of system operation.

[0021] 5. This invention can minimize the energy consumption and maximize the operating efficiency of the circulating water system, and achieve the interpretability and safety control of the system operation. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the system structure of the present invention.

[0023] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] Part 1 Overview of the Implementation Scheme of this Invention

[0026] 1. Layered Intelligent Optimization Decision-Making Method for Circulating Water Systems

[0027] The large-model-guided hierarchical intelligent optimization decision-making method for circulating water systems provided by this invention includes the following steps:

[0028] S1. Collect historical and real-time operating status data of the circulating water system, and generate long-term operating strategies by calling the large model based on the set goals and prediction information.

[0029] The large-scale model described in this invention is a deep learning-based language model or a hybrid intelligent model, which can be implemented using existing large-scale model solutions. For example, it can employ a large-scale language model (LLM) with a Transformer architecture, a GNN (Graph Neural Network) to process complex relational graph data, or a hybrid intelligent model combining multiple large-scale models, etc. Such large-scale models possess powerful representation and reasoning capabilities, enabling them to extract high-level correlation information from multi-source heterogeneous data; after training in a circulating water system operation scenario, they exhibit hierarchical decision-making capabilities.

[0030] S2. Based on the long-term operation strategy generated in step S1 and the current load information, call the large model to generate a short-term operation strategy.

[0031] When generating long-term and short-term operating strategies, the large model performs decision-making and reasoning directly within the abstract action space, without having to search for all lower-level control actions one by one.

[0032] S3. Verify the short-term operation strategy generated in step S2 before execution. Evaluate whether the short-term operation strategy meets the system's process constraints and safety requirements through a simulation model. If the verification fails, trigger the safety rollback mechanism and adopt the preset backup control mode.

[0033] The verification includes: simulating the execution of short-term operating strategies using a closed-loop simulation platform or digital twin device, evaluating the safety and feasibility of short-term operating strategies under various faults and extreme conditions, and confirming that the constraints regarding water supply temperature, safe pressure, and equipment operating limits are met.

[0034] S4. At the control level, a multi-objective optimization model is established based on the short-term operation strategy verified in step S3. The optimization objectives of this model include minimizing the weighted sum of the evaporation water consumption and total energy consumption of the circulating water system per unit time, with constraints such as water supply temperature as conditions. The multi-objective optimization model is solved in real time using an optimization algorithm, and the optimal combination of control parameters is output.

[0035] The optimization algorithm used in the multi-objective optimization model can be any one of genetic algorithm, particle swarm optimization or differential evolution; the optimization objective function is set as a weighted sum of evaporation water consumption and energy consumption, with the weights prioritizing minimizing evaporation water consumption; the constraints include the circulating water supply temperature range constraint, total energy consumption constraint and the parameter range constraint of each key equipment.

[0036] S5. The optimal control parameter combination obtained in step S4 is converted into control commands and sent to the actuators in the circulating water system for execution.

[0037] Based on long-term operation strategies, short-term operation strategies, and multi-objective optimization, multi-time-scale decision-making is formed, specifically including day-ahead scheduling, hourly rolling scheduling, and minute-level control. Among them, day-ahead or hourly trend planning is carried out at the long-term decision-making level, hourly or minute-level real-time correction is carried out at the short-term decision-making level, and minute-level precise adjustment is carried out at the control level, thereby achieving collaborative optimization across time scales.

[0038] S6. Continuously collect system operation data after execution and compare it with the set optimization target; analyze the deviation between the output of the optimization model and the actual response through self-learning algorithm, dynamically correct the objective function weights, constraint parameters and structure in the optimization model, and correct the control parameters of the model output online.

[0039] The self-learning algorithm is used to adjust the parameters of multi-objective optimization models and large models online. It is an incremental learning or reinforcement learning algorithm based on historical running data. By periodically evaluating the self-learning effect, it prevents model overfitting or correction errors.

[0040] S7. During execution, if the system operating data is detected to deviate significantly from the target range, a safety rollback mechanism will be triggered, and a pre-determined backup control mode will be adopted or a warning message to switch to manual control will be issued to ensure stable system operation.

[0041] As an example, when key indicators such as water supply temperature, cooling water flow rate, or energy consumption are found to continuously exceed predetermined thresholds, the system automatically switches to a standby control mode. This standby control mode calculates control parameters based on empirical rules or the latest reliable strategy and suspends large model inference and optimization until the system returns to normal.

[0042] Based on the understanding of those skilled in the art, the present invention also relates to a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, cause the processor to execute the aforementioned large-model-guided hierarchical intelligent optimization decision-making method for circulating water systems.

[0043] 2. Layered Intelligent Optimization Decision-Making System for Circulating Water System

[0044] The large-model-guided hierarchical intelligent optimization decision-making system for circulating water systems provided by this invention has the following structure: Figure 1 As shown. Specifically includes:

[0045] Large model server, used for deploying and running large-scale artificial intelligence models;

[0046] Sensor networks are used to collect operational data from the circulating water system and transmit the data to a large model server.

[0047] Edge computing devices are used to achieve the fusion and acquisition of multi-source heterogeneous data, and to execute local control strategies and implement safe rollback when large models are unavailable or delayed. When the edge computing device loses connection with the large model server, it can independently calculate control commands based on the most recently validated strategy or the local model to ensure continuous system operation.

[0048] The optimization decision-making module includes a long-term decision-making unit, a short-term decision-making unit, a strategy verification unit, a control decision-making unit, and a rollback control unit, each corresponding to different strategy scheduling functions. Specifically, the long-term decision-making unit obtains long-term operating strategies from the large model server based on historical and predictive data; the short-term decision-making unit obtains short-term operating strategies from the large model server based on the long-term operating strategies and real-time status; the strategy verification unit performs simulation verification of the scheduling strategies obtained from the large model server; the control decision-making unit establishes a multi-objective optimization model based on the verified short-term operating strategies and calculates control commands; and the rollback control unit invokes a pre-set safety strategy and outputs it to the execution mechanism when strategy verification fails or operational anomalies occur. The optimization decision-making module performs online analysis of the feedback execution data and updates the parameters of the large model and the optimization model in real time through a self-learning mechanism in the control decision-making unit to gradually optimize decision-making performance and system efficiency.

[0049] The actuators, including the circulating water pump controller, cooling tower fan controller, and valve actuators, are used to receive and execute instructions from the optimization decision module.

[0050] The communication module is used to connect the large model server, the optimization decision module, the sensor network, and the actuators to ensure real-time bidirectional transmission of data and instructions.

[0051] like Figure 1 As shown, the large model server is deployed in the cloud and is responsible for running large-scale artificial intelligence models; the units at each layer of the optimization decision module are located on edge or local servers, and perform policy reasoning and optimization solutions for different decision levels; the sensor network collects operating parameters of circulating water pumps, cooling towers, pipe networks and the environment; the actuators include circulating pumps, cooling tower fans, valves, etc., for executing control commands; the edge computing devices are used to execute local control and policy rollback when the large model is unavailable.

[0052] In the system architecture of this invention, the large model server and the optimization decision module are connected through an industrial control network to achieve collaborative operation. A sensor network is distributed across key points of the circulating water system, collaborating with edge computing devices via fieldbus or wireless sensor networks to achieve the fusion and acquisition of multi-source heterogeneous data. This data includes parameters such as circulating pump outlet flow and pressure, cooling tower inlet and outlet water temperatures, fan speed, motor power, pipe network temperature and flow, and ambient temperature and humidity. Sensor data is first transmitted to edge computing devices or local control nodes for data fusion and preprocessing, and then transmitted to the large model server via a data bus or gateway. Actuators are connected to the optimization decision module via programmable logic controllers (PLCs) or distributed control systems (DCS).

[0053] Part Two: Specific Implementation Cases

[0054] exist Figure 2 The flowchart of the method described in this invention is shown in the figure. The main implementation steps are as follows: data acquisition → large model inference → strategy verification → optimization solution → command issuance → execution feedback → online correction. Specifically, firstly, the sensor network collects the operating parameters of each node in the circulating water system; then, the data is input into a large model server, and inference is performed at both the long-term and short-term decision-making levels to generate multi-time-scale operating strategies; next, the generated strategies are simulated and verified to ensure that they meet safety and process constraints, and if the verification fails, a backoff mechanism is initiated; then, at the control level, a multi-objective optimization model is constructed based on the verified strategies, and the model is solved in real time to obtain accurate control parameters; the optimization results are converted into specific commands and issued to the actuators for implementation; finally, the system continuously collects feedback data after execution, compares and updates it through closed-loop feedback and self-learning algorithms, dynamically corrects the model parameters, and realizes online adaptive optimization control.

[0055] The following is a detailed explanation of the specific implementation process of the present invention through a specific implementation example.

[0056] 1. Data Acquisition and Status Awareness

[0057] 1.1 Establish an observation model to collect historical and real-time operating status data of the circulating water system.

[0058] Discrete-time index .

[0059] Define the system's true state vector as This may include: the outlet flow rate Qi of each pump, the pump head Hi, and the inlet and outlet water temperatures of the heat exchanger. Cooling tower inlet and outlet water temperatures Temperature and pressure difference at several nodes in the pipeline network, etc.

[0060] The control input vector is For example, pump speed, valve opening, fan frequency, etc.

[0061] The exogenous disturbance vector is For example, ambient dry-bulb / wet-bulb temperature, process load, inlet water temperature, electricity price, etc.

[0062] Sensor measurement vector is It satisfies the usual observation equation with respect to the actual state:

[0063] ,

[0064] Among them, C t For the observation matrix (which may vary over time to represent the visibility of different sensors); x t b is the system state vector, a minimal vectorized quantity used to describe the internal operating conditions of the system, such as pump speed and critical node temperature; t For known bias terms (such as zero-point deviation); v t To measure noise, it is usually assumed that ;

[0065] . represents a multivariate normal distribution (multidimensional Gaussian distribution), where It is the mean vector. It is the covariance matrix; The measurement noise vt is modeled as a multivariate Gaussian random vector with a mean of 0 and a covariance of Rt; R t It is the measurement noise covariance matrix, which describes the variance of sensor measurement error and the correlation between errors between sensors.

[0066] 1.2 Data Sampling and Synchronization

[0067] The sensor network can be either a fieldbus or a wireless sensor network, working in conjunction with edge computing devices to achieve the fusion and acquisition of multi-source heterogeneous data, used to collect real-time operational data, and then transmitted to a large model after preprocessing.

[0068] Field sensors in circulating water systems typically have different sampling periods (e.g., 1s, 5s, 60s). Timestamp alignment and resampling should be performed at the edge gateway to define a uniform control period. (For example =60s). All data is processed before uploading to the large model or performing optimization. Align as sequence .

[0069] 1.3 Denoising, Anomaly Detection, and State Estimation

[0070] The original measurements were subjected to combined filtering and statistical detection.

[0071] Assuming the system can be modeled using approximate dynamics. State estimation can be performed using either Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF).

[0072] For example, the EKF prediction and update steps are as follows:

[0073] predict:

[0074] in This is the Jacobian matrix of the dynamic model with respect to the state. Let be the process noise covariance.

[0075] This is the prediction of xt at time t-1; For predicting dynamic models (parameters) θ ), used for state progression or simulation; For the posterior estimate at time t-1; This is the control input vector at time t-1; To predict the covariance matrix; Let be the posterior covariance at time t-1; This is the transpose of the Jacobian matrix of the dynamic model with respect to the state.

[0076] renew:

[0077]

[0078] The Kalman gain matrix; for The transpose of the matrix; The Jacobian of the observation function with respect to the state (linearized observation matrix); To measure the noise covariance matrix; For measurement vectors; For the updated posterior state estimation; This is the updated posterior covariance matrix; It is an identity matrix.

[0079] When the data collected in the system is too complex or unidentifiable, the edges can be estimated using sliding window least squares estimation or a neural network-based observer (Observer-Net) to estimate unobservables (such as evaporation water consumption estimation).

[0080] Anomaly detection data can be tested using residual statistical tests:

[0081] ,

[0082] like If this occurs, it is considered an anomaly, triggering compensation or an alarm.

[0083] in, This is the residual vector (the difference between the measured and predicted measurements). For residual vectors The transpose of ; For residual covariance; The chi-square threshold; The chi-square distribution is at a significance level. Below, degrees of freedom The critical value is used to test whether the residual is significant.

[0084] 1.4 Feature Engineering and Multi-Scale Representation

[0085] To enable large models (such as Transformer / GNN) to effectively utilize information acquired from observation models, edge computing devices can be used to fuse and acquire heterogeneous data from multiple sources. For example, edge / feature engineering modules can be set up in edge computing devices to construct multi-scale feature vectors. ,Include:

[0086] posterior state estimate obtained by filtering ;

[0087] Statistical feature operators :recent Step mean, variance, maximum / minimum, slope (gradient), etc.;

[0088] Periodic coding Daily / Weekly Cycle Sine and Cosine Encoding ;

[0089] Topology coding : Sparse representation of the network connectivity matrix A or node embedding based on graph convolution precomputation;

[0090] Predicted perturbation sequence ;

[0091] Event markers: binary identifiers such as maintenance window and standby unit availability.

[0092] Vector concatenation operator .

[0093] Ultimately, the time-step embedding of the input large model is:

[0094]

[0095] in, This is a comprehensive feature vector; The eigenvector obtained after calculating the statistics; The timestamp of the current time t; Let A be the adjacency matrix of the system; For the predicted sequence of the perturbation variable.

[0096] 2. Large-scale model inference (long-term and short-term decision-making levels): Model structure, output, and training objectives

[0097] 2.1 Layered Responsibilities and Interfaces

[0098] At the long-term decision-making level, trend planning is carried out at the day-ahead or hour-ahead level. Based on long-term forecasts and business objectives, a set of long-term operating goals and alternative strategies are generated. The output is usually a high-level meta-command and a target weight vector.

[0099] At the short-term decision-making level, real-time adjustments are made at the hourly or minute level. Within a shorter time scale, an executable strategy draft is generated by combining real-time perturbations with long-term operational goals. The output is a set of candidate strategies along with their confidence and explanatory hints.

[0100] The two layers exchange information through a standardized interface: the long-term decision-making layer outputs goals. Key constraints With a priority list; short-term decision-making is based on these inputs and the current state to generate structured strategies.

[0101] 2.2 Architecture of the Large Model

[0102] This embodiment exemplifies the use of the following hybrid model architecture:

[0103] Temporal encoder: Based on Transformer multi-head self-attention, the input sequence is encoded to obtain the context representation H:

[0104]

[0105] Encoder output .

[0106] in, Output matrix for attention;

[0107] Q, K, V are the query / key / value matrices, respectively .

[0108] Let X be the linear projection parameter matrix, and let X be the input sequence matrix.

[0109] L×d means that H is a matrix with L rows and d columns; Let K be the transpose of the matrix. For the dimensions of Q and K;

[0110] Topology / Relationship Module: Considering the pipeline topology, a graph neural network (GNN) is used to perform message passing between the topology matrix A and node features to obtain the node embedding Z.

[0111]

[0112] in, , is an adjacency matrix with self-loops.

[0113] For degree matrix, Here, A is the layer parameter; A is the original adjacency matrix; I is the identity matrix; For the first The node feature matrix of the layer; This is the activation function.

[0114] Hybrid fusion layer: The encoder output H is fused with the node embedding Z (e.g., after stitching and projection) to obtain the global representation g.

[0115] Policy Decoder: The decoder can be an autoregressive Transformer (generating discrete policy tokens) or directly output a parameterized distribution (continuous actions):

[0116]

[0117] Confidence and Explanation Header: Calculate confidence scores and importance components based on g.

[0118]

[0119] in, For parameterized policy distribution, it represents the distribution of input features. Next action The conditional probability density; For action vectors; The strategy mean function; This is the fused global representation vector; To square the standard deviation vector and construct a diagonal covariance matrix; The standard deviation vector of the strategy; For confidence level scalar; This is the confidence head weight vector; for The transpose of ; For confidence head bias scalar; It is the logical sigmoid activation function.

[0120] 2.3 Model Types and Deployment

[0121] The large model in this example is a multimodal hybrid model, including a Transformer-based temporal encoder, a graph neural network-based topological encoder, and a Transformer-based text encoder. The outputs of each encoder are fused by cross-attention in the fusion layer, and the policy head generates a set of candidate policies, outputting confidence scores and explanations. The large model is deployed and runs on a large model server in the cloud. A lightweight, distilled copy is also provided and deployed on edge computing devices to support hierarchical decision-making and collaboration, as well as to execute local control policies and implement safe rollbacks when the large model is unavailable or delayed.

[0122] Based on the set goals and forecast information, the large model is invoked to generate a long-term operating strategy; further, based on the long-term operating strategy and current load information, the large model is invoked to generate a short-term operating strategy. In generating both long-term and short-term operating strategies, the large model performs decision-making reasoning directly within the abstract action space, without needing to search through all lower-level control actions one by one.

[0123] 2.4 Engineering and Safety Considerations

[0124] The large model generates long-term operating strategies based on historical and forecast data, and further generates short-term operating strategies based on the long-term operating strategies and real-time status. The strategies output by the large model must carry confidence and interpretable information to facilitate risk assessment by downstream strategy verification units.

[0125] The long-term decision-making output is a combination of "soft objectives + hard constraints" to ensure that short-term decision-making and control levels do not receive instructions that conflict with security rules during execution. For sensitive scenarios, a "human-in-the-loop" strategy can be further adopted: if the confidence level is low or the risk is high, an automatic alarm will be triggered and a prompt will be made to "submit the model-generated strategy to the operations and maintenance department for manual approval".

[0126] 3. Strategy Validation

[0127] The control layer establishes a multi-objective optimization model and calculates control commands based on the validated short-term operating strategy. Therefore, prior validation of the short-term operating strategy is essential.

[0128] This section describes the responsibilities, implementation, decision logic, and interfaces with upper and lower layers of the strategy validator.

[0129] 3.1 Verification Objectives and Principles

[0130] The goal of the strategy verification unit is to evaluate, through simulation models, whether the short-term operating strategies generated from the large model meet the system's process constraints and safety requirements before execution in the field. By filtering out potentially dangerous or unrealistic solutions through short-term simulation and constraint checks, it ensures that the strategies to be executed meet process safety constraints and equipment operating limits, and avoids inducing oscillations or frequent start-ups and shutdowns.

[0131] For example, a closed-loop simulation platform or digital twin device can be used to simulate the execution of short-term operating strategies, evaluate the safety and feasibility of short-term operating strategies under various faults and extreme conditions, and confirm that the constraints regarding water supply temperature, safe pressure, and equipment operating limits are met.

[0132] 3.2 Composition of the Policy Verification Unit (Verifier)

[0133] By using approximate models based on simplified physical laws or data-driven approaches instead of conventional simulators, several open-loop simulations can be completed in a short time (e.g., within 30 seconds to several minutes) to evaluate the short-term response of a strategy.

[0134] The constraint library includes hard constraints and soft constraints. Hard constraints are mandatory conditions that must be strictly adhered to and absolutely cannot be violated. Violation of these constraints will directly lead to serious consequences such as production safety accidents, process failures, permanent equipment damage, and substandard product quality. They are the core red line for strategy verification. If a short-term operating strategy generated by a large model touches a hard constraint, it will be directly judged as infeasible and trigger a safety rollback mechanism. Soft constraints are guiding conditions that are recommended to be followed, with minor short-term violations having no serious consequences. Violation of these constraints will not cause safety accidents or process failures, but will reduce system operating efficiency, increase energy and water consumption, accelerate equipment wear, and deviate from the optimal operating state. They serve as reference targets for strategy optimization.

[0135] Oscillation / Stability Detection: Check whether candidate strategies generate suspicious oscillations or trigger rapid switching in the simulation.

[0136] 3.3 Verification Process

[0137] Receive candidate strategies and current state; extract short-term disturbance predictions.

[0138] The τ-step response simulation of each candidate policy is performed in parallel in the simulator (τ can be the number of off-steps in 10–30 minutes, depending on the control cycle).

[0139] Constraint checks are performed on each simulation trajectory: if any hard constraint is triggered or exceeds the predefined margin, it is marked as "infeasible".

[0140] For strategies that pass hard constraints, perform secondary risk assessments, calculate simple risk scores, and sort them by score to return to the short-term decision-making level or directly to the control level.

[0141] If all candidate strategies are not feasible (verification fails), the fallback mechanism will be automatically triggered, and a preset security strategy will be adopted (such as executing a preset backup strategy or issuing an alarm message "handed over to manual judgment").

[0142] 4. Multi-objective optimization solution

[0143] This section describes how, at the control level, validated strategies are transformed into executable controls, and how to optimize problem structure, solution approaches, and practical constraints.

[0144] 4.1 The nature of the control problem

[0145] The task at the control level is to use optimization algorithms to solve the control sequence in the rolling time domain in real time, given the current state and the objectives of the short-term operating strategy (including objective weights, time windows and strong constraints), so that multiple objectives (such as minimizing energy consumption, minimizing evaporation water consumption, and maintaining water supply temperature) can be optimized under constraints to obtain the optimal combination of control parameters, and output the control commands to be executed immediately to the actuators in the circulating water system.

[0146] 4.2 Objectives and Constraints (Structured Description)

[0147] Objectives are typically given by the weights of higher-level objectives; typical objective terms include:

[0148] Energy consumption cost; water consumption / evaporation cost; water supply temperature deviation penalty (to ensure process quality); start-up, shutdown, and switching penalty (to reduce equipment wear). For example, the optimization objective function can be set as a weighted sum of evaporation water consumption and energy consumption, with the weights prioritizing minimizing evaporation water consumption.

[0149] The constraints cover: physical / process hard constraints (temperature, pressure difference, maximum / minimum flow rate), equipment capacity curves (pump characteristics, fan efficiency curves), and operational safety rules (minimum start-up and shutdown time, valve rate limits), etc. For example, constraints include circulating water supply temperature range constraints, total energy consumption constraints, and parameter range constraints for each key piece of equipment.

[0150] The main formulas are as follows:

[0151] I. Core Optimization Objective Function

[0152]

[0153] in, To comprehensively optimize the objectives; These are the weighting coefficients; This refers to the system's water consumption due to evaporation. This represents the total energy consumption of the system.

[0154] II. Main Constraints

[0155] a. Hard constraints of the process (water supply temperature)

[0156]

[0157] The actual water supply temperature, the lower limit of the water supply temperature. and upper limit It is set directly by the production process requirements.

[0158] b. Equipment capacity constraints (pump / fan power)

[0159]

[0160] , This refers to the actual power of the pump and fan; , This refers to the equipment's rated maximum power.

[0161] c. Operational safety constraints (total energy consumption)

[0162]

[0163] This represents the upper limit of the system's total energy consumption.

[0164] 4.3 Solution Methods and Engineering Trade-offs

[0165] The optimization algorithm used to solve the multi-objective optimization model can be a genetic algorithm, a particle swarm optimization algorithm, or a differential evolution algorithm.

[0166] For linear or linearizable cases, quadratic programming or sequential quadratic programming can provide an accurate solution in real time; this is optimal when the control cycle is short and the model can be approximated by linearization.

[0167] For problems involving discrete decisions (pump start / stop), integer variables need to be introduced and solved using mixed-integer linear programming or mixed-integer quadratic programming. If the solution time of mixed-integer linear programming exceeds the real-time budget, heuristic or two-stage methods can be used.

[0168] In high-dimensional or non-convex cases, gradient optimization and heuristics can be combined as backup solutions, but caution should be exercised in industrial deployments, and a strong constraint protection layer should be added.

[0169] 4.4 Engineering Implementation Details

[0170] Receding horizon: Only the first control action is issued for each control step, and then the solution is recalculated by time rolling.

[0171] Warm-start: Uses the solution from the previous step as an initial value to accelerate the solution in the next step.

[0172] Interpretable output: The optimizer exists as the algorithmic kernel of the optimization decision module and is its core component. The optimizer should return the Lagrange multipliers or sensitivity analysis results of key constraints so that upper management or operations personnel can understand why certain devices are constrained.

[0173] Real-time guarantee: When the solution times out, the system should use the most recent available solution or a simplification strategy to ensure continuous control.

[0174] 5. Command issuance and execution (interface, permissions, and local execution logic)

[0175] This section explains how to safely and reliably convert controller outputs into field-executable instructions, and how the local controller ensures safe execution.

[0176] 5.1 Instruction Encapsulation and Issuance Protocol

[0177] The control messages issued include: control actions, execution start and end times, priority, confidence level and risk label, message validity period and digital signature / authorization information.

[0178] It is recommended to use industry-standard protocols to send messages through secure channels (TLS or dedicated industrial VPN) to ensure message integrity and prevent tampering.

[0179] 5.2 Local Controller Reception Strategy

[0180] The actuators, including local circulating water pump controllers, cooling tower fan controllers, and valve actuators in the circulating water system, receive and execute instructions from the optimization decision module. At the control level, precise adjustments at the minute level are achieved based on local controllers, thereby enabling collaborative optimization across time scales in conjunction with long-term and short-term decision-making levels.

[0181] As an example, the actuator works in conjunction with an edge computing device deployed locally in a programmable logic controller (PLC) or distributed control system (DCS) to receive and execute instructions from the optimization decision module.

[0182] Upon receiving the policy, the local controller determines whether to execute it based on the following logic:

[0183] If the confidence level is high and no abnormalities are found in the local self-test, it will be executed automatically.

[0184] If the confidence level is moderate, then enter the "manual confirmation / semi-automatic" mode.

[0185] If the confidence level is low or there is a local hardware failure, execution will be refused and a local rollback will be triggered.

[0186] The local controller should also have rate limiting, minimum downtime limits, and execution confirmation reporting mechanisms.

[0187] 5.3 Execution Confirmation and Two-Way Feedback

[0188] Each issued instruction should have a traceable log on the local machine, including the execution confirmation time, the actual action value, and the status code indicating whether the execution was completed.

[0189] After receiving the execution report, the central system immediately incorporates the actual measurements into the next step of state estimation and modeling data pool for closed-loop updates.

[0190] 5.4 Redundancy and Disaster Recovery

[0191] The command delivery link should support dual redundancy (primary network / backup network) and allow edge nodes to continue executing the most recently verified policy or local rules when communication is interrupted.

[0192] To avoid equipment disruption caused by sudden commands, the instructions should include rate limiting information, requiring a smooth transition for the on-site frequency converter / PLC.

[0193] 6. Online feedback and self-learning (closed-loop iteration, model update strategy)

[0194] During the process of the actuator regulating the circulating water system based on the control commands of the optimal control parameters, it is necessary to continuously collect the system operation data after execution and compare it with the set optimization target; through the analysis of the deviation between the output of the optimization model and the actual response by the self-learning algorithm, the objective function weights, constraint parameters and structure in the optimization model are dynamically corrected to achieve online model correction.

[0195] This section explains how to use the execution results for continuous improvement, including when to update the model, how to ensure safe updates, and strategies for adjusting optimization weights online.

[0196] 6.1 Collection and Evaluation of Feedback Data

[0197] Real-time data collection: Real-time measurements (temperature, flow rate, energy consumption readings, etc.) after execution are transmitted back through the edge gateway; at the same time, the execution timestamp, command version number, confidence level and local anomaly flag are recorded.

[0198] Self-learning algorithms: These algorithms utilize incremental learning or reinforcement learning based on historical operating data to adjust the parameters of multi-objective optimization models and large models online; and prevent model overfitting or correction errors by periodically evaluating the self-learning effect.

[0199] Evaluation indicators: Calculate key performance indicators, such as actual energy consumption, evaporation water consumption, water supply temperature deviation, and number of equipment start-ups and shutdowns, and compare them with the optimization targets to determine whether the established improvement goals have been achieved.

[0200] 6.2 When is a model update triggered?

[0201] Periodic updates: For example, perform batch offline training once a week or month to enhance the model with the latest historical data.

[0202] Triggered update: Triggered when residuals or KPIs drop below a preset threshold. This update will first be verified in a sandbox.

[0203] Incremental updates: For minor drift, the model is fine-tuned online in small batches, and sandbox verification is required before and after the update.

[0204] 6.3 Update Strategies and Security Measures

[0205] Sandbox validation: All new models must first undergo before-and-after comparison experiments in a digital twin or simulation environment (covering common operating conditions and several extreme operating conditions).

[0206] Gray release: Through A / B testing or regional release, trial operation is first carried out on a small scale or low-impact units.

[0207] Version control and rollback: Each time a model is deployed, the version number and training data snapshot must be recorded so that a quick rollback can be performed if any problems occur.

[0208] Constraint protection: Regardless of how the model is updated, its output must be reviewed again by the policy verification unit, and hard constraints cannot be relaxed.

[0209] 6.4 Optimize online weight adjustment

[0210] At the long-term / short-term decision-making level, the weights of multiple objectives can be dynamically adjusted based on economic signals (such as real-time electricity prices) or operation and maintenance signals (such as spare parts availability). The adjustment strategy is usually given periodically by the long-term decision-making level, or issued temporarily in case of emergencies (severe fluctuations in electricity prices).

[0211] 7. Safety rollback mechanism

[0212] If the system operating data is detected to deviate significantly from the target range during the control execution of the circulating water system, a safety backoff mechanism is triggered, and a predetermined safety scheduling scheme is adopted to ensure stable system operation.

[0213] The safety rollback mechanism includes: when key indicators such as water supply temperature, cooling water flow rate, or energy consumption are found to continuously exceed the predetermined threshold, the system will automatically switch to the backup control mode. The backup control mode calculates control parameters based on empirical rules or the latest reliable strategy and suspends large model inference and optimization until the system returns to normal.

[0214] This section summarizes the triggering conditions of the rollback mechanism, the forms of rollback strategies, and the safeguards for its implementation, to ensure that the system can still operate safely under abnormal circumstances.

[0215] 7.1 Rollback Trigger Conditions

[0216] The rollback mechanism is triggered when any of the following conditions are met:

[0217] An irreparable hard constraint violation occurred during the strategy validation phase;

[0218] During the execution process, key indicators (such as water supply temperature and pump pressure difference) were detected to exceed the set threshold.

[0219] Communication interrupted or large model server unavailable;

[0220] The local device detected a fault or protection action (such as overcurrent / overtemperature);

[0221] Manual emergency stop / safety command.

[0222] 7.2 Rollback Strategy Types

[0223] Recent Reliable Strategy Rollback: Revert to the control strategy that was most recently validated and successfully implemented. This is usually the first choice, as it ensures continuity and minimizes disturbances.

[0224] Rule-based safety strategy: If no most recent reliable strategy is available, predefined empirical rules are used (e.g., "ensure the water supply temperature is within the range of X, and operate at a fixed pump frequency"). These rules are developed and validated in advance by process / operations experts.

[0225] Degradation mode (minimum risk mode): The system is placed in an operating state that minimizes risk, such as reducing switchovers, increasing safety margins, and pausing automatic optimization until manual intervention.

[0226] Manual takeover: In extreme or complex abnormal situations, the system switches to manual control, with on-duty engineers intervening to make decisions.

[0227] 7.3 Rollback Execution Process and Audit

[0228] The rollback command must include the triggering reason, execution time, execution subject (automatic / manual), and rollback strategy version number, and be fully recorded in the system log for post-event auditing.

[0229] After rollback, the system enters monitoring mode and continuously monitors for several periods of time to confirm whether it has returned to normal or continues to trigger higher-level responses (such as shutdown for maintenance).

[0230] 7.4 Verify the effectiveness of the rollback

[0231] The rollback action itself needs to be verified: after rollback, key indicators should be confirmed to return to the safe range through short-term simulation or field data; if rollback is ineffective, a higher priority safety process (such as manual shutdown or alarm escalation) should be initiated.

[0232] In summary, this invention achieves intelligent optimization scheduling of circulating water systems across multiple objectives and time scales through hierarchical optimization decision-making guided by a large model, combined with steps such as data acquisition, model inference, strategy verification, optimization solution, execution feedback, and online correction. Compared with traditional technologies, this invention has the following significant advantages: Firstly, the large model can learn system operation patterns from historical and real-time data, providing strong cognitive support and cross-level planning capabilities for the decision-making process, making the scheduling strategy more intelligent and efficient. Secondly, this invention adopts a closed-loop feedback and self-learning mechanism to continuously optimize model parameters based on actual operating results, improving the system's robustness and adaptability, while ensuring system operation safety through verification and rollback mechanisms. These innovations minimize energy consumption and maximize operating efficiency in circulating water systems, and achieve interpretability and controllability of system operation.

[0233] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A hierarchical intelligent optimization decision-making method for a circulating water system guided by a large model, characterized in that, Includes the following steps: S1. Collect historical and real-time operating status data of the circulating water system, and generate long-term operating strategies based on the set targets and prediction information by calling the large model. S2. Based on the long-term operation strategy and the current load information, call the large model to generate a short-term operation strategy; Use simulation models to evaluate whether short-term operating strategies meet process constraints and safety requirements; If verification fails, a security rollback mechanism will be triggered; S3. Based on the verified short-term operation strategy, establish a multi-objective optimization model, use an optimization algorithm to solve the optimization model in real time, and output the optimal control parameter combination; convert the optimal control parameter combination into control commands and send them to the actuators in the circulating water system for execution; S4. Continuously collect system operation data after execution and compare it with the set optimization target; analyze the deviation between the output of the optimization model and the actual response through self-learning algorithm, dynamically correct the objective function weights, constraint parameters and structure in the optimization model, and correct the control parameters of the model output online.

2. The method according to claim 1, characterized in that, The verification refers to: using a closed-loop simulation platform or digital twin device to simulate the execution of short-term operation strategies, evaluating the safety and feasibility of short-term operation strategies under various faults and extreme operating conditions, and confirming that the constraints regarding water supply temperature, safe pressure, and equipment operating limits are met.

3. The method according to claim 1, characterized in that, The large model is a language model or hybrid intelligent model based on deep learning, trained in the operation scenario of a circulating water system, and has hierarchical decision-making capabilities. When generating long-term and short-term operation strategies, the large model directly performs decision-making reasoning in the abstract action space, without having to search all lower-level control actions one by one.

4. The method according to claim 1, characterized in that, The optimization algorithm used in the multi-objective optimization model is any one of genetic algorithm, particle swarm optimization, or differential evolution; the optimization objective function is set as the weighted sum of the evaporation water consumption and total energy consumption of the circulating water system, with the weights prioritizing minimizing the evaporation water consumption; the constraints include the circulating water supply temperature range constraint, the total energy consumption constraint, and the parameter range constraints of each key equipment.

5. The method according to claim 1, characterized in that, The self-learning algorithm is an incremental learning or reinforcement learning algorithm based on historical running data, used to adjust the parameters of multi-objective optimization models and large models online; by periodically evaluating the self-learning effect, it prevents model overfitting or correction errors.

6. The method according to claim 1, characterized in that, During execution, if the short-term operational strategy fails to be verified or if system operational data deviates significantly from the target range, a safety rollback mechanism is triggered; specifically including: When key indicators such as water supply temperature, cooling water flow rate, or energy consumption are found to continuously exceed the predetermined threshold, the system automatically switches to standby control mode, calculates control parameters based on empirical rules or the latest reliable strategy, and suspends large model inference and optimization until the system returns to normal.

7. The method according to claim 1, characterized in that, Based on long-term operation strategies, short-term operation strategies, and multi-objective optimization, multi-time-scale decision-making is formed, specifically including day-ahead scheduling, hourly rolling scheduling, and minute-level control. Among them, day-ahead or hourly trend planning is carried out at the long-term decision-making level, hourly or minute-level real-time correction is carried out at the short-term decision-making level, and minute-level precise adjustment is carried out at the control level, thereby achieving collaborative optimization across time scales.

8. A hierarchical intelligent optimization decision-making system for a circulating water system guided by a large model, characterized in that, include: Large model server, used for deploying and running large-scale artificial intelligence models; Sensor networks are used to collect operational data from the circulating water system and transmit the data to a large model server. Edge computing devices are used to achieve the fusion and acquisition of multi-source heterogeneous data, and to execute local control strategies and safe rollback when large models are unavailable or delayed; The optimization decision-making module includes a long-term decision-making unit, a short-term decision-making unit, and a control decision-making unit, each corresponding to the scheduling function of different operating strategies. The long-term decision-making unit obtains long-term operating strategies from the large model server based on historical and predictive data. The short-term decision-making unit obtains short-term operating strategies from the large model server based on the long-term operating strategies and real-time status. The control decision-making unit establishes a multi-objective optimization model based on the verified short-term operating strategies and calculates control commands. The actuators, including the circulating water pump controller, cooling tower fan controller, and valve actuator, are used to receive and execute instructions from the control decision module. The communication module is used to connect the large model server, the optimization decision module, the sensor network, and the actuators to ensure real-time bidirectional transmission of data and instructions.

9. The system according to claim 8, characterized in that, The large model server and the optimization decision module are connected through an industrial control network; the sensor network is a fieldbus or wireless sensor network that works in conjunction with edge computing devices to achieve the fusion acquisition of multi-source heterogeneous data; the actuator is connected to the optimization decision module through a programmable logic controller (PLC) or a distributed control system (DCS).

10. The system according to claim 8, characterized in that, The optimization decision module also includes a strategy verification unit and a rollback control unit; the former is used to perform simulation verification of the scheduling strategy obtained from the large model server, and the latter is used to call the preset security strategy and output it to the execution mechanism when the strategy verification fails or the operation is abnormal; when the edge computing device is disconnected from the large model server, it independently calculates control instructions according to the most recently verified strategy or the local model to ensure the continuous operation of the system.

11. The system according to claim 8, characterized in that, The optimization decision module performs online analysis of the feedback execution data and updates the parameters of the large model and the optimization model in real time through the self-learning mechanism in the control decision unit, so as to gradually optimize the decision performance and system benefits.

12. A computer-readable storage medium, characterized in that, The storage medium stores computer-executable instructions, which, when executed by a processor, cause the processor to perform the large-model-guided hierarchical intelligent optimization decision-making method for circulating water systems as described in any one of claims 1 to 7.

Images