Heat distribution control method for heat exchanger based on multi-objective optimization

Abstract

The application relates to the field of intelligent control of heat exchangers, and discloses a heat exchanger heat distribution control method based on multi-target optimization, which collects operation data under multiple working conditions and carries out labeling; nonlinear decoupling mapping of heat exchanger multi-working condition thermodynamic parameters and low-dimensional working condition feature extraction; flow distribution multi-target optimization based on double pheromone cooperation and thermodynamic constraint projection; online decision and valve opening degree instruction dynamic smoothing mapping based on variable weight target heart closeness; heat distribution closed-loop control based on variable weight decision and valve dynamic smoothing mapping. The application has the beneficial effect that the sensitivity of entropy production rate to each operation parameter is used as physical guiding information to reweight the original operation state vector in the sense of thermodynamics, so that the subsequent low-dimensional feature extraction and optimization process preferentially pays attention to key variables that truly affect irreversible loss, instead of simply relying on numerical compression.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of heat exchangers, specifically a heat exchanger heat distribution control method based on multi-objective optimization. Background Technology

[0002] As core equipment in energy systems and chemical processes, heat exchangers directly impact the energy efficiency and operational economy of the entire system through heat distribution control. In typical applications such as centralized heating, air conditioning, and petrochemical production, heat exchangers typically employ a multi-branch parallel structure. Each branch distributes the flow on the primary or secondary side via regulating valves to meet the heat load demands of different areas or process stages. However, heat exchangers face a severe multi-objective conflict during variable load operation: on the one hand, it is necessary to increase the total heat exchange to ensure heating quality or process requirements; on the other hand, it is desirable to reduce the power consumption of the circulating pump to save energy. These two objectives are physically mutually restrictive. Increasing the flow in a certain branch can usually enhance the convective heat transfer in that branch, but it will also increase flow resistance and may crowd out the flow in other branches, resulting in overall performance where more uniform flow is not necessarily better.

[0003] 1. Existing methods typically input sensor quantities directly into optimization algorithms when performing data-driven modeling. This lacks differentiation of the thermodynamic importance of different physical quantities. Variables such as flow rate, temperature, and pressure drop interfere with each other due to differences in dimensions and amplitudes. Optimization algorithms are easily misled by high-amplitude but low-value data and find it difficult to consistently grasp the key directions affecting heat distribution efficiency.

[0004] 2. Traditional linear dimensionality reduction techniques use a uniform projection matrix for all operating conditions, which cannot adapt to the changes in the dominant contradictions of heat exchangers under different modes such as high load and large temperature difference and low load and small temperature difference. This results in the extracted low-dimensional features either biased towards heat exchange improvement while ignoring resistance costs, or vice versa, failing to provide adaptive state input for multi-objective optimization.

[0005] 3. Conventional continuous ant colony algorithm or particle swarm optimization algorithm usually only uses a single pheromone or fitness function in multi-objective optimization, which easily causes the population to gather in a single-objective optimal region. The generated solution set is unevenly distributed or missing ends on the Pareto front. Moreover, the handling of constraints is limited to simple boundary clipping, and it cannot identify and avoid those thermally unreasonable regions that meet numerical constraints but have deteriorated heat transfer and sudden increase in entropy production.

[0006] 4. When selecting points online from the Pareto candidate set, existing control systems often use fixed weights or decision rules based solely on the current instantaneous performance, lacking smooth constraints on the execution results of the previous cycle. This makes them prone to frequent jumps between different candidate solutions when operating conditions fluctuate, leading to repeated large valve movements. At the same time, valve opening commands are usually generated using uniform or fixed rate limits, which cannot adaptively adjust the amplitude of the action according to the speed of changes in operating conditions. This results in either slow response or over-adjustment, affecting control quality and the lifespan of the actuator. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a heat exchanger heat distribution control method based on multi-objective optimization, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention employs the following technical solution: A heat exchanger heat distribution control method based on multi-objective optimization includes the following steps: S1. Collect operating data of the heat exchanger under multiple operating conditions and label the collected data; S2. Construct a unified operating state vector from the historical and current operating data of the heat exchanger, calculate the influence of each operating parameter on the entropy production rate, and use this influence as physical guidance information to weight the operating state vector. Then, extract a low-dimensional operating condition feature vector through a modal mapping method. The low-dimensional operating condition feature vector is used to characterize the dominant change direction of heat exchange potential and resistance cost under the current operating condition. S3. A dual pheromone field is used to track the high heat transfer solution and the low pump work solution respectively, and thermodynamic constraint projection and entropy production correction are applied to the candidate flow distribution coefficient vector after each generation. S4. Automatically adjust the target weights based on the current low-dimensional operating condition feature vector, and add the difference between the candidate solution and the execution result of the previous cycle as a smoothing constraint into the decision score to obtain the optimal flow distribution coefficient, and convert it into a valve opening command, and output a smooth control command through dynamic rate limiting. S5. Based on the final valve opening command vector obtained from S4, a closed-loop regulation process is completed in each control cycle.

[0009] Furthermore, the collected operational data includes the total primary flow rate, the total secondary flow rate, the primary inlet temperature, the primary outlet temperature, the secondary inlet temperature, the secondary outlet temperature, the inlet and outlet temperatures of each branch, the pressure drop of each branch, the current opening degree of each branch valve, and the flow distribution coefficient of each branch calculated from the flow rate of each branch.

[0010] Furthermore, S2 specifically refers to: The operating data of the heat exchanger in the current control cycle is collected to form the original operating state vector. ; For the original running state vector Each component in the vector is normalized to obtain a normalized running state vector. ; Based on the normalized running state vector Calculate the global entropy yield using thermal measurement data ; Evaluate the normalized running state vectors one by one The yield of each component to global entropy The degree of influence is used to obtain the entropy production sensitivity vector. ; Based on the entropy production sensitivity vector For the normalized running state vector Perform weighted operations to obtain the entropy-generated weighted running state vector. ; A multi-condition mode center is established using historical operating data, and the entropy-weighted operating state vector is calculated. Perform a submodal low-dimensional mapping to obtain a low-dimensional operating condition feature vector. .

[0011] Furthermore, S3 specifically refers to: S31. In addition to the mass conservation constraint, an entropy yield threshold judgment and gradient correction are introduced to make the randomly generated candidate flow allocation coefficient vector simultaneously satisfy numerical feasibility and thermal rationality. S32. Based on the low-dimensional working condition feature vector, adaptively determine the search focus of the two targets, and add repulsion drift to make the ants actively move away from the already over-dense area and expand the coverage of the Pareto front. S33, External Archive Maintenance and Pareto Candidate Set Output: Through non-dominated sorting and hypervolume contribution truncation, more representative non-dominated candidate solutions are retained.

[0012] Furthermore, S4 specifically refers to: S41. Candidate scheme selection based on variable weighted target proximity; S42. Convert the flow distribution ratio into the branch target flow, obtain the valve opening target based on the valve characteristic curve, and finally perform rate limiting and smoothing output.

[0013] Furthermore, S31 specifically refers to: S311, Define the flow allocation coefficient vector As an optimization variable; S312. Initialize the ant colony within the feasible region and establish a heat exchange pheromone field and a pump power pheromone field respectively; wherein, the heat exchange pheromone field is used to record the vector distribution of the flow allocation coefficients that rank first in total heat exchange in history; the pump power pheromone field is used to record the vector distribution of the flow allocation coefficients that rank last in total pump power consumption in history. S313. Generate the original candidate flow allocation coefficient vector based on dual pheromone field sampling. ; S314. Divide the original candidate flow allocation coefficient vector Project the result onto the feasible region that satisfies the mass conservation and minimum branch flow constraints, and perform the feasible region projection operation to obtain a first correction result; S315. Determine whether the global entropy productivity corresponding to the first correction result exceeds the threshold; if it exceeds the threshold, perform a second correction in the direction of reducing entropy productivity to obtain the final candidate flow allocation coefficient vector. ; S316. Calculate the final candidate flow allocation coefficient vector. The corresponding total heat exchange and total pump power consumption will be used as the basis for subsequent ranking.

[0014] Furthermore, S32 specifically refers to: S321. Determine the fusion weights of the heat exchange pheromone field and the pump power pheromone field based on the current low-dimensional operating condition feature vector. The fusion weights are used to characterize whether the current operating condition is more inclined to increase heat exchange or more inclined to reduce pump power consumption. S322. Select elite centers from the heat exchange pheromone field and the pump power pheromone field respectively, and generate random perturbations near each elite center. S323. Merge the two types of sampling results according to the fusion weight to obtain a new original candidate flow allocation coefficient vector. The generated new solutions will automatically shift towards the more desired objectives depending on the operating conditions; S324. Assign coefficient vectors to new original candidate flows based on the density of neighboring solutions in the external archives. Increase repulsive drift; S325. Perform feasible region projection and entropy production correction in step S31 on the original candidate flow allocation coefficient vector after adding the repulsion drift to obtain the final candidate flow allocation coefficient vector. And calculate the target value, that is, calculate the total heat exchange and total pump power consumption;

[0015] S326. Update the heat exchange pheromone field and pump work pheromone field based on the target performance of the candidate solutions in this round.

[0016] Furthermore, S33 specifically refers to: S331. Convert the vector of all final candidate flow allocation coefficients obtained in the current iteration. Merge with existing external file A; S332. Perform non-dominated sorting on the merged solution set, and prioritize retaining candidate solutions with higher non-dominated levels. S333, When external files When the number of non-dominated candidate solutions exceeds the capacity limit, candidate solutions with smaller contributions are deleted in ascending order of hypervolume contribution. The hypervolume contribution is used to measure the contribution of a candidate solution to the current Pareto front coverage area.

[0017] Furthermore, S41 specifically refers to: S411. Read the Pareto candidate set of the current control cycle and extract the target value and flow allocation coefficient vector for each candidate solution; S442. Generate dynamic target weights based on the current low-dimensional working condition feature vector; S443. Determine the positive and negative ideal target points based on the current candidate set, and calculate the bullseye proximity for each candidate solution. ; S444. Based on the actual execution results of the previous control cycle, a smoothing penalty is added to each candidate solution to obtain a corrected score. ; S445. Select the corrected score from all candidate solutions. The largest candidate solution is used as the target flow allocation coefficient vector for the current control cycle. .

[0018] Furthermore, S42 specifically refers to: S421. Based on the current total flow demand and target flow allocation coefficient vector Calculate the target flow rate for each branch; S422. Calculate the original valve opening command vector based on the valve flow characteristics of each branch. ; S423. Determine the maximum allowable opening change in this control cycle based on the current rate of change of operating conditions and the actuator's operating capability; S424, Original valve opening command vector Perform branch-by-branch amplitude limiting to obtain the final valve opening command vector. ; S425, Valve opening command vector based on actual branch flow feedback. Make minor closed-loop corrections; S426, Output the final valve opening command vector This enables online decision-making and execution for the current control cycle.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using the sensitivity of entropy production rate to various operating parameters as physical guidance information, the original operating state vector is reweighted in a thermodynamic sense, so that the subsequent low-dimensional feature extraction and optimization process will prioritize the key variables that truly affect irreversible loss, rather than simply relying on numerical compression.

[0020] 2. A modal local projection strategy is adopted, and dimensionality reduction mappings are established for different operating conditions. This avoids the uniform linear projection from averaging the dominant contradictions between high-load and low-load operating conditions, and enables the low-dimensional feature vector to adaptively reflect the true trade-off between "heat exchange potential" and "resistance cost" under the current operating conditions.

[0021] 3. In the multi-objective optimization, two independent pheromone fields of heat exchange and pump work were constructed. When generating candidate flow distribution coefficients, entropy yield threshold judgment and gradient correction projection were introduced. The second law of thermodynamics was embedded as a hard constraint into the search process to actively filter out mathematically feasible but thermally poor invalid solutions.

[0022] 4. In the online decision-making process, a point selection mechanism combining variable weight target proximity and historical execution smoothing penalty is proposed. The working condition adaptive dynamic rate limit is used to convert the flow distribution coefficient into valve command, realizing an online trade-off between performance pursuit and action smoothness. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention; Figure 2 It is a probability density map of the heat exchange pheromone field sampling; Figure 3 It is a sampling probability density map of the pump power pheromone field; Figure 4 This is a schematic diagram of the feasible region projection operation; Figure 5 This is a scatter plot of the target spatial distribution between total heat exchange and total pump power consumption generated during the optimization process; Figure 6 It is a graph showing the changes between the originally calculated valve opening command and the actual command issued after dynamic rate limiting within a continuous control cycle. Detailed Implementation

[0024] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0025] like Figure 1As shown, this invention proposes a heat exchanger heat distribution control method based on multi-objective optimization, the main contents of which are as follows: S1. Data Acquisition and Labeling Preprocessing for Multi-condition Operation of Heat Exchangers The heat exchanger heat distribution control method described in this invention relies on a comprehensive understanding of the heat exchanger's multi-condition thermal state and the accumulation of historical operating data. Before implementing optimized control, systematic data acquisition is required. The data acquisition targets include temperature sensors, flow sensors, pressure sensors, and valve position sensors installed on the primary side pipe network, secondary side pipe network, and each parallel branch.

[0026] A total flow meter and inlet / outlet temperature sensors are installed on the primary side main pipe, and similarly on the secondary side main pipe. Each branch is equipped with a branch flow meter, inlet / outlet temperature sensor, branch differential pressure sensor, and valve opening feedback transmitter. The sensor signals are synchronously acquired through a distributed data acquisition module or programmable logic controller at a uniform sampling period. The sampling period is generally set between a few seconds and tens of seconds based on the thermal inertia of the heat exchanger, ensuring that the dynamic process of load changes is captured without introducing excessive high-frequency noise.

[0027] The collected raw data includes the total primary flow rate, the total secondary flow rate, the primary inlet temperature, the primary outlet temperature, the secondary inlet temperature, the secondary outlet temperature, the inlet and outlet temperatures of each branch, the pressure drop of each branch, the current valve opening of each branch, and the flow distribution coefficient of each branch calculated from the flow rate of each branch.

[0028] To ensure the reliability and consistency of training data and online computation data in subsequent steps, preprocessing of the raw signals is necessary during the data acquisition phase. First, outlier removal is performed. For abrupt changes significantly exceeding the physically possible range, such as a sudden drop in flow rate to zero or a short-term temperature jump exceeding the medium's operating range, median filtering based on a sliding window or a three-standard-deviation discrimination method is used to remove them, and the gaps are filled with the previous valid value or local interpolation results. Second, time alignment processing is performed on data from different sensors. Since there may be slight differences in the response time and communication delay of each sensor, the physical quantities acquired within the same control cycle are aligned to a unified time using nearest neighbor matching or linear interpolation, using the control system's reference clock as a reference. This eliminates the problem of mismatched thermodynamic parameters caused by time misalignment.

[0029] The data collection scope should cover typical operating conditions of the heat exchanger, including high-load, large-temperature-difference operation periods, low-load, small-temperature-difference operation periods, transition periods during partial branch shutdown or regulation, and typical daily operating curves for different seasons. The collected historical operating data will be stored in a structured format, with each data record containing a timestamp, all sensor measurements, and performance evaluation metrics calculated from the measurements, such as total heat exchange, total pump power consumption, and global entropy productivity.

[0030] To support the modal clustering for low-dimensional operating condition feature extraction, the training of the lightweight operating condition identification network, and the historical reference for dual pheromone field search in subsequent steps, the collected historical operating data needs to be labeled accordingly. The labeling categories mainly include two types: one is unsupervised labels for operating condition modal clustering, which are operating condition modal numbers automatically assigned by clustering algorithms. These numbers are not given manually in advance, but are obtained by performing clustering analysis using entropy-weighted operating state vectors after historical data has accumulated to a certain scale. The clustering results are stored in the database as modal labels for historical operating states, and are used to quickly determine which typical mode the current operating condition belongs to during online operation. The other type is supervised labels for training lightweight operating condition identification networks. The input of this network is a low-dimensional operating condition feature vector, and the output is the heat exchange pheromone field weights. The construction method of its training labels is as follows: In the offline stage, for the low-dimensional operating condition feature vector at each historical operating moment, combined with the load change trend before and after that moment, whether the current state is a heat exchange bottleneck, and the expert's judgment on "whether to focus more on increasing heat exchange or saving pump power at this time", the reasonable weight value of heat exchange to be assigned under this operating condition is determined by using posterior optimality analysis or manual rule assignment. This weight value is the label for supervised learning.

[0031] In addition, to support the screening and historical reference of non-dominated solutions in external archives, it is also necessary to mark the dominance relationships of candidate allocation schemes in historical operating data. That is, based on the two objectives of total heat exchange and total pump power consumption, mark which allocation schemes do not dominate each other and which schemes are dominated, forming a sample library of non-dominated solutions, which serves as a reference benchmark for optimizing algorithm parameter tuning and historical comparison.

[0032] S2. Nonlinear decoupling mapping and low-dimensional feature extraction of thermodynamic parameters under multiple operating conditions of heat exchangers During the operation of a heat exchanger under varying loads, there is a clear coupling relationship between the total flow rate on the primary side, the total flow rate on the secondary side, the inlet and outlet temperatures, the pressure drop in the branches, and the valve opening. The redistribution of flow not only changes the intensity of convective heat transfer in each branch, but also changes the pressure drop distribution and the local temperature difference distribution, thereby causing the heat transfer performance and flow resistance to change synchronously.

[0033] Conventional linear dimensionality reduction methods typically only compress the original sample data numerically, and cannot distinguish "which variables change mainly affect heat exchange and which variables change mainly increase irreversible losses". Therefore, subsequent optimization is prone to searching in high-dimensional mixed data, and it is difficult to find an effective adjustment direction that balances heat exchange and energy consumption.

[0034] This invention first constructs a unified operating state vector from historical and current operating data of the heat exchanger. Then, it calculates the influence of each operating parameter on the entropy production rate, using this influence as physical guidance information to weight the operating state vector. Next, it extracts a low-dimensional operating condition feature vector through a modal mapping method. This low-dimensional feature vector characterizes the dominant change direction of "heat exchange potential" and "resistance cost" under the current operating condition. Subsequent multi-objective optimization is directly performed based on this low-dimensional feature vector, thereby reducing invalid searches. The specific steps are as follows: Traditionally, data from sensors such as flow rate, temperature, and differential pressure are directly spliced ​​together and input into the optimization algorithm. While this method is simple, different physical quantities have different dimensions, different rates of change, and different impacts on heat exchanger performance. This can easily lead to the optimization algorithm being interfered with by high-amplitude but low-value data.

[0035] This invention first calculates the irreversible loss, then reweights the input using the sensitivity of each variable to the irreversible loss, and finally extracts a low-dimensional feature vector of the operating condition. This allows the subsequent optimization process to prioritize the key directions that truly affect heat distribution. The specific steps are as follows: 1> Collect the operating data of the heat exchanger in the current control cycle to form the original operating state vector. , This represents the original running state vector at the current moment, with dimension . This is used to fully characterize the current thermal state of the heat exchanger; in, This includes the total primary flow rate, total secondary flow rate, primary inlet temperature, primary outlet temperature, secondary inlet temperature, secondary outlet temperature, pressure drop of each branch, current valve opening of each branch, and flow distribution coefficient of each branch. All attributes must be in a unified unit (e.g., converted to SI).

[0036] Furthermore, to ensure the stability of subsequent calculations, the sampled data first undergoes outlier removal and time alignment processing. Outlier removal can be performed using the 3σ discriminant method, median filtering, or upper and lower limit judgment based on a sliding window. Time alignment can be performed by matching the most recent time or linear interpolation of data from different sensors according to a unified sampling period, so that temperature, flow rate, and pressure drop data at the same time correspond in time.

[0037] 2> For the original running state vector Each component in the vector is normalized to obtain a normalized running state vector. , This represents a normalized running state vector with dimension . It is used to eliminate the differences in dimensions of different physical quantities, so that flow rate, temperature and pressure drop can participate in subsequent calculations on the same numerical scale.

[0038] In practical implementation, minimum-maximum normalization or mean-standard deviation normalization can be used for each dimension of data. For parameters with relatively stable operating ranges, such as valve opening degree and flow distribution coefficient, minimum-maximum normalization can be used. For parameters that are greatly affected by operating condition fluctuations, such as temperature difference and pressure drop, mean-standard deviation normalization can be used. In order to avoid normalization distortion caused by new operating conditions exceeding the historical range, the upper and lower limits of normalization can be determined by the upper and lower limits of design operating conditions and historical statistical intervals.

[0039] 3> Based on the normalized running state vector Calculate the global entropy yield using thermal measurement data , This represents the global entropy production rate of the heat exchanger under its current operating conditions, used to characterize the intensity of irreversible losses caused by both flow resistance and finite temperature difference heat transfer. The larger the value, the greater the thermodynamic loss under the current distribution state.

[0040] In practical implementation, the global entropy yield consists of two parts: one part is the flow irreversible loss and the other part is the heat transfer irreversible loss. The flow irreversible loss can be obtained by multiplying and accumulating the pressure drop of each branch with the volumetric flow rate; the heat transfer irreversible loss can be estimated by the heat exchange of each branch and the temperature difference between the wall temperature and the average fluid temperature.

[0041] In one embodiment, for example, assuming the heat exchanger has two branches, the following measurements are taken under the current operating conditions: Branch 1: Volumetric Flow Rate Pressure drop heat exchange Fluid average temperature Average wall temperature ; Branch 2: Volumetric flow rate Pressure drop heat exchange Fluid average temperature Average wall temperature ; Then the flow is irreversible. The calculation method is expressed as follows: ; And the heat transfer loss is irreversible. The calculation method is expressed as follows: ; Furthermore, global entropy productivity The calculation method is expressed as follows: This value represents the intensity of irreversible loss under the current allocation state.

[0042] In practical implementation, if the wall temperature cannot be directly measured, the wall temperature can be estimated based on the branch inlet temperature, branch outlet temperature, heat exchange capacity, and heat exchange area, combined with the heat balance relationship of a conventional heat exchanger. In a method that is easy to implement in engineering, the wall temperature can be approximated by using the logarithmic mean temperature difference of the branch and the heat exchange capacity of the branch to construct the irreversible heat transfer loss, thereby reducing the amount of online calculation.

[0043] 4> Evaluate the normalized running state vectors one by one The yield of each component to global entropy The degree of influence is used to obtain the entropy production sensitivity vector. , dimension Entropy production sensitivity vector The The component characterizes the first The strength of the change in entropy production rate caused by a small change in an operating parameter; the greater the entropy production sensitivity, the more worthy the corresponding operating parameter is to focus on in feature extraction and subsequent optimization. For example, if a slight increase in the total primary flow rate results in a significant decrease in the global entropy yield, it indicates that this parameter is more sensitive to improving irreversible losses under the current operating conditions and will receive higher weight in subsequent feature construction.

[0044] In practical implementation, a finite difference method can be used for calculation, that is, fixing other components and only calculating the difference between the components. Apply a small perturbation to one of the components and recalculate the global entropy yield. Then, the sensitivity of the component is obtained by dividing the difference in entropy yield before and after the perturbation by the amount of perturbation.

[0045] In practical implementation, to prevent noise amplification, the disturbance amount can be taken as 0.5% to 2% of the normalized range of the corresponding variable. For parameters with relatively obvious noise, such as flow rate and pressure drop, the sampled values ​​can be averaged for 3 to 5 periods before calculating the sensitivity.

[0046] In one embodiment, as an example, assume a normalized running state vector In the figure, the component representing "total primary flow" is... The current global entropy production rate is Then, only for Apply a small positive perturbation, such as the perturbation value (Taking 1% of the normalized range of 1.0), the perturbed components Recalculate along with other unchanged components to obtain a new entropy production rate. Then the component Corresponding entropy production sensitivity for The result is negative, indicating that under the current operating conditions, increasing the total primary flow rate can effectively reduce the global entropy productivity. This parameter deserves special attention in feature extraction. Its absolute value... This will serve as the basis for subsequent weighting.

[0047] 5> Based on the entropy production sensitivity vector For the normalized running state vector Perform weighted operations to obtain the entropy-generated weighted running state vector. , dimension This is used to highlight variables that have a significant impact on irreversible losses and weaken variables that are less related to current heat allocation decisions.

[0048] In one implementation, the entropy production sensitivity vector is first... Normalize to weight coefficients, and then multiply these weight coefficients dimension by dimension into the normalized running state vector. The entropy-weighted operating state vector is obtained from the above. Based on this, although variables such as temperature, flow rate, and pressure drop are all retained, their contributions in subsequent mappings have been reordered according to thermodynamic significance.

[0049] In one embodiment, as an example, the sensitivity vector Take the absolute value of each component, and then use the minimum-maximum normalization method to map it to... The interval is used as a weighting coefficient. For example, if the minimum absolute value of all components is 0.1 and the maximum is 7.0, then the absolute value of the k-th component... Then its weight Finally, the weight coefficient vector composed of all weights and the normalized running state vector are combined. Element-wise multiplication yields the entropy-weighted operating state vector. .

[0050] It should be noted that if a component has very low long-term sensitivity but is a variable that must be monitored for operational safety, such as the anti-freezing constraint variable corresponding to the lowest branch flow, then that component will still be retained, but its weight will not be too high to avoid omitting safety boundary information. For example, even if some variables (such as the flow of a certain branch) are not sensitive to changes in entropy production rate under the current operating conditions (i.e., ... Although it is very small, it still needs to be retained in the eigenvector because it involves system safety (e.g., preventing heat exchanger tube bundles from freezing or scaling), and its weight cannot be set to zero and discarded.

[0051] 6> Establish a multi-condition mode center using historical operating data, and weight the entropy-generated operating state vector. Perform a submodal low-dimensional mapping to obtain a low-dimensional operating condition feature vector. , dimension ,and This is used as the state input for subsequent multi-objective optimization, preferably. 2 to 5 are acceptable.

[0052] Specifically, clustering is first performed on historical operating data to obtain multiple operating condition mode centers. Here, "operating condition mode centers" represent the typical operating states of the heat exchanger under different loads, temperature differences, and flow distribution levels. Clustering can be performed using... To improve consistency with thermal performance, mean clustering, Gaussian mixture clustering, or hierarchical clustering are preferred methods. Entropy-weighted operating state vectors are preferred as the clustering input.

[0053] In practice, the steps for extracting modal centers from historical operational data are as follows: a> Collect historical operating data of the heat exchanger under different seasons and loads (processed into an entropy-weighted operating state vector). (in the form of) constitute a sample set.

[0054] b) The sample set is divided into groups using the k-means clustering algorithm. There are several clusters, and the clustering objective is to minimize the sum of squared distances from each sample to the center of its cluster (i.e., the modality center).

[0055] After clustering is completed, the geometric center of each cluster (the mean of all samples in each dimension) is a "condition mode center". For example, if we set... The algorithm will automatically find 8 of the most representative ones. The vectors represent typical states such as "high load and large temperature difference" and "low load and small temperature difference".

[0056] After clustering is completed, a local dimensionality reduction mapping is established for each working condition mode. The local dimensionality reduction mapping can be achieved by principal component analysis, small-scale autoencoder or local linear embedding. In an easy-to-implement approach, principal component analysis is performed on the samples in each mode, and the top few principal components with a cumulative contribution rate of 85% to 95% are retained to form the local projection matrix of that mode.

[0057] Furthermore, the current entropy-weighted operating state vector is... The similarity is calculated by comparing the modal centers of each working condition with the model centers, and multiple local projection results are weighted and fused according to the similarity to obtain a low-dimensional working condition feature vector. The calculation method is expressed as ; in, Indicates the number of operating modes; Indicates the current operating condition and the first The similarity weights for each working condition mode range from 0 to 1, and the sum of all similarity weights is 1. Indicates the first The local projection matrix corresponding to each operating mode.

[0058] In practical implementation, similarity weight Based on the current entropy-weighted operating state vector Modal centers for various operating conditions Euclidean distance The calculations show that the smaller the distance, the higher the similarity and the greater the weight.

[0059] In one implementation, the number of operating modes The value can be 6 to 12; when the current operating condition is close to the high-load, large-temperature-difference mode, the similarity weight corresponding to this mode is larger, and the output low-dimensional operating condition feature vector is obtained. It will further highlight the "heat exchange improvement" related directions; when the current operating condition is close to the low load and low resistance mode, the output low-dimensional operating condition feature vector It will place greater emphasis on the "energy-saving operation" direction.

[0060] It should be noted that the dominant contradictions of a heat exchanger differ under different operating conditions. High-load conditions prioritize heat exchange potential, while low-load conditions prioritize pump power cost. Using a uniform linear projection would easily average out these differences. Therefore, this invention does not simply perform dimensionality reduction, but first rearranges the importance of variables using entropy production sensitivity, and then performs local projection according to the operating mode. Based on this, a low-dimensional operating condition feature vector is obtained. It has clear thermodynamic significance and can directly serve subsequent multi-objective optimization.

[0061] S3. Multi-objective optimization of flow allocation based on dual pheromone synergy and thermodynamic constraint projection Obtain low-dimensional working condition feature vectors Then, it is necessary to find candidate solutions that simultaneously take into account heat exchange and pump power consumption in the continuous decision space formed by the flow distribution coefficients of each branch. The multi-objective optimization of heat exchanger flow distribution includes at least two objectives: one is to increase the total heat exchange and the other is to reduce the total pump power consumption.

[0062] Conventional continuous ant colony algorithms typically use only a single pheromone model to drive the search, which can easily lead to excessive clustering on a certain target. Furthermore, after generating candidate assignments, they often only perform simple boundary pruning and cannot actively avoid high-entropy production regions. As a result, although a large number of candidate solutions are numerically feasible, their thermal performance is poor.

[0063] This invention employs a dual pheromone field to track the "high heat transfer solution" and the "low pump work solution" separately, and adds thermodynamic constraint projection and entropy production correction after each generation of candidate flow distribution coefficient vectors. This ensures that the optimization process is both multi-objective-oriented and does not deviate from the actual thermal constraints of the heat exchanger. The specific steps are as follows: S31, Construction of Dual Pheromone Fields and Projection of Thermodynamically Feasible Domain Traditional constraint processing typically only guarantees that the sum of the flow distribution coefficients equals 1, but cannot guarantee that candidate solutions avoid the heat transfer degradation zone. This invention, in addition to the mass conservation constraint, introduces entropy yield threshold judgment and gradient correction, ensuring that the randomly generated candidate flow distribution coefficient vector simultaneously satisfies both "numerical feasibility" and "thermally reasonable" requirements. The specific steps are as follows: 1> Define the flow allocation coefficient vector As an optimization variable, among which, Represents the vector of flow allocation coefficients. Indicates the number of branches. Indicates the first The proportion of traffic allocated to each branch is between 0 and 1, and the sum of the traffic proportions of all branches is 1.

[0064] Furthermore, to ensure minimum antifreeze flow or minimum flushing flow, a minimum allocation lower limit is also set for each branch. Based on this, the target search space for optimization is defined as: the flow allocation coefficient vector. In satisfying At the same time, it also needs to meet The constraints are then defined; subsequent projection operations are used to ensure that the generated candidate solutions always lie within the feasible region.

[0065] 2. Initialize the ant colony within the feasible region and establish a heat exchange pheromone field and a pump power pheromone field, respectively. The heat exchange pheromone field is used to record the vector distribution of flow allocation coefficients that rank high in total heat exchange in history; the pump power pheromone field is used to record the vector distribution of flow allocation coefficients that rank low in total pump power consumption in history. "Ranking high in heat exchange" means that in the current iteration population or historical records, the target value of heat exchange of the candidate solution is in the high-ranking interval. For example, sort all solutions in the population from large to small in total heat exchange, and select the top 10%-20% of solutions as the elite set of the heat exchange pheromone field. These solutions are considered "high-ranking". Similarly, "ranking low in pump power consumption" means that the target value of total pump power consumption is in the low-ranking interval (the smaller the better). Select the top 10%-20% of solutions with the smallest pump power consumption as the elite set of pump power consumption. For example, if the current population contains 50 solutions and the total heat exchange is between 80kW and 120kW, then solutions with a heat exchange greater than 115kW (corresponding to the top 20%) are considered "ranked high"; if the total pump power consumption is between 30W and 70W, then solutions with a pump power less than 38W are considered "ranked low".

[0066] In practical implementation, based on the ant colony algorithm, initial ant positions can be generated uniformly within the feasible region, and each initial ant position is an initial flow distribution coefficient vector. Then, several solutions with the highest ranking for heat exchange target are selected from the current population as the heat exchange elite set, and several solutions with the highest ranking for pump power target are selected as the pump power elite set.

[0067] Furthermore, by establishing Gaussian sampling models centered on the two types of elite solutions respectively, the heat transfer pheromone field tends to sample near solutions with high heat transfer, while the pump work pheromone field tends to sample near solutions with low pump work, providing a foundation for subsequent dual-objective collaborative search.

[0068] In one embodiment, for example, the heat exchange target (i.e., total heat exchange) ) and pump power target (i.e., total pump power consumption) () is a vector of candidate flow allocation coefficients. The model obtained by substituting the input into the performance calculation model of the heat exchanger can be a simplified thermodynamic model, such as a lumped parameter model based on the ε-NTU method (number of heat transfer units method). Given boundary conditions such as the current total flow, the model can calculate the corresponding... and ; The lumped parameter model based on the ε-NTU method (heat transfer unit number method) takes into account parameters such as the current total flow rate, the flow distribution coefficient of each branch, and the inlet temperature. The output is the heat transfer and pressure drop of each branch. For example, for a single branch, given the heat capacity flow rate... Heat transfer coefficient Heat exchange area Then the number of heat transfer units Heat exchange efficiency heat exchange The pressure drop is calculated using an empirical formula. , The drag coefficient, The volumetric flow rate is given; when multiple branches are connected in parallel, the total heat exchange is the sum of the heat exchange of each branch, and the total pump work is the sum of the heat exchange of each branch. The sum is divided by the pump efficiency.

[0069] In one embodiment, for example, regarding the Gaussian sampling model, suppose that the three solutions with the highest heat exchange rates are selected from the current population as an elite set, and their flow distribution coefficient vectors are respectively... , , Calculate the mean vector of the elite set. And calculate the variance of the elite set in each dimension to form a variance vector. Then, the heat exchange pheromone field changes from a mean of variance is Gaussian distribution definition ( Represented by vector (The elements in the matrix are used as the diagonal elements of the diagonal matrix). During subsequent ant sampling, new vectors can be directly extracted from this distribution, meaning that the new solution will likely appear in the distribution. The distribution range is consistent with the discreteness of the elite solution itself.

[0070] 3> Generate the original candidate flow allocation coefficient vector based on dual pheromone field sampling. , This represents the original candidate flow allocation coefficient vector randomly generated by the ant in the current iteration, which has not yet been constrained and corrected.

[0071] In practice, each ant first samples two candidate centers from the heat exchange pheromone field and the pump work pheromone field, respectively. Then, random perturbations are superimposed on the vicinity of the two candidate centers. Finally, the samples are fused according to the weights of the current operating conditions to form the original candidate flow distribution coefficient vector. .

[0072] In one embodiment, as an example, suppose the Gaussian sampling mean of the heat exchange pheromone field is... variance vector The mean value of the pump power pheromone field is variance vector The current operating condition weight is then determined. , .

[0073] Based on this, each new solution generated retains both the "search towards higher heat exchange" and "search towards lower pump power" trends simultaneously, rather than favoring only one of them.

[0074] 4> Assign the original candidate flow coefficient vector Projecting onto the feasible region that satisfies the constraints of mass conservation and minimum branch flow, the feasible region projection operation is performed to obtain a correction result.

[0075] In the specific implementation, first put Components smaller than the minimum allocation limit of each branch are directly raised to the corresponding limit. Then, the remaining allocable proportion is calculated and redistributed according to the relative size of other components, so that the sum of all components is equal to 1 again. If some components still exceed the limit after redistribution, the above process is repeated until all components are within the allowable range.

[0076] It should be noted that the above "truncation-redistribution" iterative process is essentially equivalent to transforming the original candidate flow allocation coefficient vector... It pulls the object back into the constrained simplex, and its implementation is simple and easy to execute online.

[0077] 5> Determine whether the global entropy productivity corresponding to the first correction result exceeds the threshold; if it exceeds the threshold, perform a second correction in the direction of reducing entropy productivity to obtain the final candidate flow allocation coefficient vector. , This represents the candidate flow allocation coefficient vector ultimately used for target evaluation; entropy production rate threshold. This represents the risk boundary of high entropy production.

[0078] In practical implementation, the result of the first correction is first substituted into a pre-trained shallow neural network (such as a single hidden layer neural network, with inputs including the flow rate of each branch, inlet temperature, and valve opening, and outputs including the global entropy productivity, total heat exchange, and total pump work; for online use, the forward calculation is performed directly). The global entropy productivity under this allocation state is calculated. If the global entropy productivity is not higher than the threshold... Then, the result of the first correction is directly used as the final candidate flow allocation coefficient vector. If the global entropy productivity is higher than the threshold Then, a small perturbation is applied to the flow distribution coefficient of each branch, and the changing trend of entropy production rate with respect to the flow distribution coefficient of each branch is calculated to obtain the entropy production gradient direction. Then, a small-step correction is made to the first correction result along the direction of decreasing entropy production. This correction method is expressed as follows: ; in, This represents the feasible region projection operation; This indicates the correction step size, used to control the magnitude of the secondary correction (preferably 0.05). The entropy production gradient vector (calculated based on the finite difference method) is used to characterize the direction of the influence of the increase or decrease of the flow distribution coefficient of each branch on the entropy production rate.

[0079] It should be noted that after the second correction, a feasible region projection needs to be performed again to ensure the final candidate flow allocation coefficient vector. It still meets the total amount constraint and the lower limit constraint of a single branch.

[0080] 6> Calculate the final candidate flow allocation coefficient vector The corresponding total heat exchange and total pump power consumption will be used as the basis for subsequent ranking, where: Total heat exchange is used to measure the heat transfer effect of the candidate solution, and total pump power consumption is used to measure the flow cost of the candidate solution.

[0081] In practice, the total heat exchange can be obtained by multiplying the mass flow rate, specific heat capacity and temperature difference between the inlet and outlet of each branch and then summing them; the total pump power consumption can be obtained by multiplying the volume flow rate and pressure drop of each branch and then summing them, and then converting them in combination with the pump efficiency.

[0082] In one embodiment, assuming a two-branch system, the working fluid is water (specific heat capacity...). ): Branch 1: Mass Flow Inlet and outlet temperature difference Volumetric flow rate Pressure drop ; Branch 2: Mass Flow Inlet and outlet temperature difference Volumetric flow rate Pressure drop ; Pump efficiency ; Total heat exchange ; And the total pump power consumption .

[0083] It should be noted that if a candidate solution causes the flow rate of the edge branch to be too low, although the local pressure drop may decrease, it will cause the local temperature difference to deteriorate and the entropy production rate to increase rapidly. This invention will actively pull the flow rate of the branch back to a more reasonable range through secondary correction, thereby reducing invalid solutions that are obviously infeasible but are easily generated by random search.

[0084] It should also be noted that the heat distribution problem of heat exchangers is not a purely mathematical constraint problem. Although many distribution schemes meet the flow rate and ratio constraints, they may cause local heat exchange deterioration or a sudden increase in irreversible losses. This invention extends the "feasible region processing" from the traditional boundary trimming to "constraint projection + entropy production correction". By introducing the entropy production rate threshold and entropy production gradient correction, regions with poor thermal performance can be filtered out in advance during the search phase.

[0085] In one embodiment, such as Figure 2-3 As shown, this figure uses two-dimensional contour lines to display the sampling probability density of two pheromone fields. The first subplot represents the heat exchange pheromone field (red, favoring high flow rate in branch 1 and low flow rate in branch 2), while the right subplot represents the pump power pheromone field (blue, favoring low flow rate in branch 1 and high flow rate in branch 2). These dual pheromone fields guide the search towards "high heat exchange" and "low pump power," respectively, avoiding bias caused by a single pheromone. The horizontal and vertical axes represent the flow rate distribution coefficients of branch 1, respectively. Flow distribution coefficient of branch 2 (Dimensionless, 0-1), the contour line color represents the sampling probability density.

[0086] In one embodiment, a feasible region projection operation is performed, such as... Figure 4 As shown, 20 original candidate points (red dots) for three-branch flow distribution coefficients are transformed into feasible points (blue dots) after "truncation-redistribution" projection, with the following constraints: The gray plane represents the constrained simplex. Simple boundary clipping cannot satisfy the simultaneous constraint and lower bound, while the projection operation can pull any original point back into the feasible region. The coordinate axes λ1, λ2, and λ3 (dimensionless, 0-1) represent the flow distribution coefficients of branch 1, branch 2, and branch 3, respectively.

[0087] S32. Ant movement based on dual pheromone weight fusion and repulsion drift A single pheromone field can easily lead to ant colonies concentrating near a certain type of target solution, resulting in incomplete Pareto front coverage. This invention addresses this issue by using low-dimensional eigenvectors... The search focus of the two targets is adaptively determined, and a repulsion drift is added to encourage ants to actively move away from overly dense areas, thereby expanding the coverage of the Pareto front. The specific steps are as follows: 1> Based on the current low-dimensional operating condition feature vector Determine the fusion weights of the heat exchange pheromone field and the pump power pheromone field. The fusion weights are used to characterize whether the current operating condition is more inclined to "increase heat exchange" or "reduce pump power consumption".

[0088] In one implementation, a lightweight operating condition identification network is used to weigh the output heat transfer weights. And set the pump power weight to ,in, Indicates the weight of the heat exchange pheromone field. This represents the weight of the pump power pheromone field.

[0089] To facilitate engineering implementation, a small neural network needs to be trained offline beforehand. The training data comes from historical operating data, and the input is a low-dimensional feature vector of operating conditions at various historical moments. The output labels are the weights that should be assigned to the heat exchange under this operating condition, determined based on expert experience or posterior optimality analysis. The network structure adopts a conventional multilayer perceptron structure, with the number of nodes in the input layer being equal to... The dimensions are the same, there are 2 hidden layers, each with 16 nodes, and the activation function is ReLU. The output layer has 1 node and uses the Sigmoid activation function, outputting a scalar between 0 and 1 (i.e., ...). During training, mean squared error is used as the loss function for supervised learning.

[0090] 2> Elite centers are selected from the heat exchange pheromone field and the pump work pheromone field respectively, and random perturbations are generated near each elite center. This allows the dual pheromone sampling to both "develop around high-quality solutions" and retain a certain degree of random exploration capability.

[0091] In practice, each ant first selects an elite center in each of the two types of pheromone fields (heat exchange pheromone field and pump work pheromone field) (randomly selecting a solution from the set as the center), and then determines the sampling variance based on the dispersion of the solutions near the corresponding elite center. The more concentrated the solutions are near the elite center, the more stable the local area is, and the sampling variance can be appropriately reduced. The more dispersed the solutions are near the elite center, the more the area needs to be explored, and the sampling variance can be appropriately increased.

[0092] 3> Fuse the two types of sampling results according to the fusion weights to obtain a new original candidate flow allocation coefficient vector. The generated new solutions will automatically lean towards the more needed target depending on the operating conditions. For example, the new solutions generated under high load conditions are closer to the high heat exchange solution, while the new solutions generated under low load stable conditions are closer to the low pump power solution.

[0093] In one implementation, candidate results from the heat exchange pheromone field and candidate results from the pump work pheromone field can be linearly combined according to weights, and then a small random perturbation can be superimposed to obtain a new original candidate flow distribution coefficient vector. .

[0094] In one embodiment, for example, suppose , The elite centers selected from the heat exchange field are The elite center selected from the pump power field is Add perturbations near the two centers (e.g., add Gaussian noise with a mean of 0 and a standard deviation of 0.05). ), obtained the first disturbance elite center Second Disruption Elite Center Then, the weighted fusion is obtained. .

[0095] 4> Assign coefficient vectors to new original candidate flows based on the density of neighboring solutions in the external archive. Increase repulsive drift.

[0096] In practice, we first find several non-dominated solutions that are closest to the current ant's position in the external archive, and then calculate the relative directions of these neighboring solutions to the current ant's position. If the neighboring solutions are too concentrated, it means that the region has been fully explored. In this case, when generating a new solution, we add a small displacement in the direction away from the center of the neighboring solutions.

[0097] It should be noted that rejection drift is a mechanism to prevent premature convergence of the algorithm and increase the diversity of the solution set. When the algorithm finds that a large number of similar non-dominated solutions have gathered in a certain area, it will actively "push" the newly generated candidate solutions away from the area. The role of rejection drift is to reduce the probability that a large number of ants will repeatedly land in the same local area, so that the front end and the middle area can be searched.

[0098] It should also be noted that the external archive is a memory space independent of the ant colony, used to store all excellent non-dominated solutions found in the algorithm iterations so far. In one embodiment, as an example, the external archive... One of the elements is a structure containing: Flow allocation coefficient vector: ; Total heat exchange: ; Total pump power consumption: ; This can be represented as a tuple: ( , , External archives It is a list of multiple such tuples, where any two solutions are mutually exclusive.

[0099] In one embodiment, for example, suppose the current ant position is Find the three closest non-dominated solutions in external file A, for example... , , Then calculate the centers of these neighboring solutions. And calculate the repulsion direction, that is, from point to Direction vector Then, in generating new Then, an additional displacement along the repulsive direction is added to achieve the repulsive drift, denoted as... ,in, Describing the L2 norm, It is a small step size (such as 0.01), based on which new solutions tend to move to regions where the solution set is sparser.

[0100] 5> Perform feasible region projection and entropy production correction in step S301 on the original candidate flow allocation coefficient vector after adding the repulsion drift to obtain the final candidate flow allocation coefficient vector. And calculate the target value (i.e., calculate the total heat exchange and total pump power consumption).

[0101] Based on this, the movement of each ant is not only guided by the dual-target pheromone field, but also by thermal feasibility constraints and frontier uniformity constraints, which can significantly reduce the problem of "concentrated search but single result".

[0102] 6> Update the heat exchange pheromone field and pump work pheromone field based on the target performance of the candidate solutions in this round.

[0103] In practical implementation, the original pheromones are first evaporated to simulate the evaporation of pheromones in nature over time. Specifically, at the beginning of each iteration, the pheromone field (i.e., the mean vector and variance matrix of Gaussian sampling) is attenuated by a certain proportion to reduce the influence of the old pheromones. Then, the candidate solutions that perform well in this round are added to the pheromone field of the corresponding target. If a candidate solution has a better ranking in total heat exchange, it will contribute more to the heat exchange pheromone field. If a candidate solution has a lower total pump power, it will contribute more to the pump power pheromone field. For non-dominated solutions that simultaneously take into account two targets, pheromones can be added to both pheromone fields at the same time.

[0104] Based on this, by alternating between pheromone evaporation and replenishment, the search direction can be continuously updated according to changes in operating conditions, instead of remaining near outdated solutions for a long time.

[0105] S33, External File Maintenance and Pareto Candidate Set Output The result of multi-objective optimization is not a single solution, but a set of non-dominated candidate solutions. If these candidate solutions are not organized, the database will expand rapidly, increasing computational load and making subsequent decisions unstable. This invention retains more representative non-dominated candidate solutions through non-dominated sorting and hypervolume contribution truncation. The specific steps are as follows: 1> Convert the vector of all final candidate flow allocation coefficients obtained in the current iteration. Merge with existing external files, where external files This represents the set used for long-term storage of non-dominated candidate solutions. Each element in the set includes a flow distribution coefficient vector, total heat exchange, and total pump power consumption.

[0106] In one embodiment, as an example, suppose the current iteration produces a non-dominated solution. Its total heat exchange Main pump power Existing solutions in external archives The range is , The range is The solution is in The solution is superior to most solutions. If the above is also better, then the update method is: Add the solution to the elite set of the heat exchange pheromone field (according to...) (Sort the top few), then recalculate the mean. and variance Simultaneously, this solution is added to the elite set of the pump power pheromone field (according to...). (Sort the top few), then recalculate. and If the solution is significantly better than the worst solution in each of the two elite sets for both objectives, then it contributes to both pheromone fields.

[0107] 2> Perform non-dominated sorting on the merged solution set, and prioritize retaining candidate solutions with higher non-dominated levels.

[0108] In practice, if a candidate solution is not inferior to another candidate solution in terms of both total heat exchange and total pump power consumption, and is better in at least one of the objectives, then the former is considered to dominate the latter. After sorting according to this rule, the solutions in the first non-dominated layer are retained first.

[0109] 3> When external files When the number of non-dominated candidate solutions exceeds the capacity limit, candidate solutions with smaller contributions are deleted according to their hypervolume contribution, from smallest to largest. Hypervolume contribution measures the magnitude of a candidate solution's contribution to the current Pareto front coverage area. In the hypervolume contribution ranking, "smaller" means that when the external archive capacity exceeds the limit, all non-dominated solutions are sorted by their hypervolume contribution value from smallest to largest, and the solutions with the smallest contribution value at the top of the ranking are deleted first, until the number of archives drops to the capacity limit. For example, in a bi-objective optimization scenario, after sorting the non-dominated solutions by a certain objective, the area increment of the rectangle enclosed by adjacent solutions and the reference point is calculated. If the proportion of the adjacent area increment of a solution to the total front area is lower than a preset lower limit (e.g., lower than 1%), or its contribution value is in the bottom 20th percentile of the archives, then the solution's independent contribution to the front coverage is considered "small," and it should be prioritized for deletion.

[0110] In one implementation, under a dual-objective scenario, the hypervolume contribution can be calculated in a way that is easy to implement: first, sort the non-dominated candidate solutions according to one of the objectives, and then calculate the area increment of the rectangle enclosed by the adjacent candidate solutions and the reference point. The smaller the area increment, the smaller the independent contribution of the candidate solution to the frontier coverage area, and it can be deleted first. The reference point can be determined by expanding outward by a certain margin based on the worst objective value in the current file. For example, the reference point can be "the coordinates corresponding to the current maximum pump power consumption and the current minimum heat exchange rate, expanded outward by 5%".

[0111] 4> Truncate the external file As the Pareto candidate set output for the current control cycle, this Pareto candidate set covers both "high heat exchange scheme", "low pump power scheme" and the compromise scheme between the two, providing an optional basis for subsequent online decision-making.

[0112] It should be noted that the present invention ensures the representativeness and uniformity of the solution set by maintaining external archives, so that subsequent online decision-making will not only select from a few densely similar solutions, but will make a robust selection on a more complete candidate set.

[0113] S4. Online decision-making based on variable-weight target proximity and dynamic smooth mapping of valve opening commands. In each control cycle, the actual control system can only issue a unique set of valve opening commands to the actuator. If the solution closest to a fixed target is directly selected from the Pareto candidate set, it is easy to frequently jump between different candidate solutions when the load changes, causing the valve to repeatedly and significantly move. Conventional fixed-weight decision-making methods are also difficult to adapt to the changes in emphasis between high-load and low-load conditions.

[0114] This invention automatically adjusts the target weights based on the current low-dimensional operating condition feature vector, then incorporates the difference between the candidate solution and the execution result of the previous cycle as a smoothing constraint into the decision scoring, and finally converts the optimal flow distribution coefficient vector into a valve opening command, and outputs a smooth control command through dynamic rate limiting. The specific steps are as follows: S41. Candidate scheme selection based on variable weighted target proximity

[0115] 1> Read the Pareto candidate set for the current control cycle and extract the target value and flow allocation coefficient vector for each candidate solution. Specifically: Suppose the current Pareto candidate set contains The candidate solution, the th The vector of flow allocation coefficients corresponding to the candidate solutions is denoted as . , Indicates the first The flow distribution ratio of each candidate solution on each branch; The target performance for each candidate solution includes total heat exchange and total pump power consumption.

[0116] To facilitate the joint evaluation of different objectives, the two objectives are first normalized and aligned. Specifically, the total heat exchange is treated as an evaluation quantity that is "the larger the better" and the total pump power consumption is treated as an evaluation quantity that is "the smaller the better". Then, they are mapped to the range of 0 to 1 respectively.

[0117] 2> Based on the current low-dimensional operating condition feature vector Generate dynamic target weights; specifically, this can be achieved by transforming low-dimensional feature vectors of operating conditions. The first two principal features are used as decision inputs. The first principal feature usually reflects the overall load level, and the second principal feature usually reflects the temperature difference utilization status or resistance change trend.

[0118] Furthermore, heat exchange weight and pump power weight are generated based on the positions of these two main features in the preset rule table. When the first main feature is high and the second main feature shows a large temperature difference utilization space, the heat exchange weight is increased. When the first main feature is low and the system is already in a stable operating range, the pump power weight is increased.

[0119] 3> Determine the positive and negative ideal target points based on the current candidate set, and calculate the bullseye proximity for each candidate solution. , Indicates the first The bullseye proximity of a candidate solution is used to measure how close the candidate solution is to the current ideal target. The larger the value, the more suitable the candidate solution is as the execution target of the current control cycle.

[0120] In practical implementation, firstly, the optimal values ​​of two objectives are selected from the current candidate set to form a positive ideal objective point, and then the worst values ​​of two objectives are selected to form a negative ideal objective point; then, the i-th... The differences between each candidate solution and the positive and negative ideal target points are then converted into weighted grey relational degrees by combining dynamic target weights, ultimately yielding the bullseye proximity. , represented as ,in, Indicates the first The weighted absolute difference between each candidate solution and the positive ideal objective point is used to measure the degree to which the solution deviates from the positive ideal point. The dynamic weights representing the total heat exchange are derived from the low-dimensional eigenvectors of the operating conditions. The decision reflects the degree of importance attached to the heat exchange target under the current operating conditions, and the range of values ​​is [not specified]. ; Indicates the first Total heat exchange of each candidate solution Total heat exchange at the ideal point The absolute value of the difference; Indicates the first The total heat exchange corresponding to each candidate solution; This represents the total heat transfer at the ideal target point, taking the maximum heat transfer value among all solutions in the current candidate set. The dynamic weight representing the total pump power consumption is preferably defined as follows: This reflects the current level of emphasis on energy conservation (reducing pump power) in the operating conditions; Indicates the first Total pump power consumption for each candidate solution Total pump power consumption at the ideal point The absolute value of the difference; Indicates the first The total pump power consumption corresponding to each candidate solution; This represents the total pump power consumption at the ideal target point, taking the minimum pump power value among all solutions in the current candidate set; Indicates the first The weighted absolute difference between each candidate solution and the negative ideal objective point is used to measure the degree to which the solution deviates from the negative ideal point. This represents the total heat transfer at the negative ideal target point, taking the minimum heat transfer value among all solutions in the current candidate set; This represents the total pump power consumption at the negative ideal target point, taking the maximum pump power value among all solutions in the current candidate set.

[0121] In practical implementation, the positive ideal point takes the maximum total heat exchange and the minimum total pump power consumption, while the negative ideal point takes the minimum total heat exchange and the maximum total pump power consumption.

[0122] 4> Based on the actual execution result of the previous control cycle (i.e., the flow distribution coefficient vector corresponding to the command calculated by step S402 and finally issued to the valve in the previous control cycle), a smoothing penalty is added to each candidate solution to obtain a corrected score. , Indicates the first The corrected score of each candidate solution is used to control the trade-off between "performance priority" and "smooth operation". In specific implementation, define Indicates the proximity to the bullseye. Indicates the first The cost of the difference between each candidate solution and the result of the previous control cycle. The positive score represents the smoothing penalty coefficient. pass The calculation is performed in the following manner, where, It can be calculated by the weighted sum of squares of the changes in the flow distribution coefficients of each branch.

[0123] In practical implementation, if the flow change of a certain branch has a greater impact on the total heat exchange, then the branch has a higher weight in the differential cost; if the flow change of a certain branch has a smaller impact on the total heat exchange, then the branch has a lower weight in the differential cost. Based on this, the system can avoid frequently adjusting key branches in pursuit of minor performance improvements.

[0124] In practical implementation, the smoothing penalty coefficient The system can be set based on the thermal inertia of the heat exchanger. A higher thermal inertia indicates that the system is less sensitive to short-term actions. A larger value can be selected to further suppress unnecessary back-and-forth adjustments.

[0125] 5> Select the corrected score from all candidate solutions The largest candidate solution is used as the target flow allocation coefficient vector for the current control cycle. , This represents the target flow distribution coefficient vector selected in the current control cycle, which is used for subsequent valve opening command calculations.

[0126] For example, if two candidate solutions are close to each other, but one of the candidate solutions is significantly less different from the result of the previous cycle, the system will prioritize the candidate solution with smaller action changes, thereby reducing valve impact and improving control continuity.

[0127] It should be noted that online control of heat exchangers not only needs to consider the current instantaneous target, but also the actuator operation cost and system thermal inertia. This invention does not simply select the "best performing" solution from the Pareto candidate set, but incorporates the current operating condition preference and historical execution continuity into the decision-making process. By introducing variable weights and historical penalties, decision jumps can be significantly reduced.

[0128] In one embodiment, such as Figure 5 The graph, presented as a scatter plot, illustrates the spatial distribution of the total heat exchange (horizontal axis) and total pump power consumption (vertical axis) generated during the optimization process. Green dots represent non-dominated solutions (stored in an external archive), gray dots represent dominated solutions, and the green dashed line outlines the Pareto front. The graph also labels positive ideal points (maximum heat exchange, minimum pump power) and negative ideal points (minimum heat exchange, maximum pump power). The dual-pheromone collaborative search can obtain a broad set of non-dominated solutions, providing a complete selection space from "high heat exchange schemes" to "low pump power schemes" for subsequent online decision-making. The horizontal axis represents total heat exchange (kW); the vertical axis represents total pump power consumption (W).

[0129] S42. Dynamic smooth mapping from the target flow distribution coefficient vector to the valve opening command. Field actuators typically receive valve opening commands, not flow distribution ratios. Therefore, it is necessary to first convert the flow distribution ratio into the branch target flow, then determine the valve opening target based on the valve characteristic curve, and finally apply rate limiting and output smoothing. The specific steps are as follows: 1> Based on the current total traffic demand and target traffic allocation coefficient vector Calculate the target flow rate for each branch, where: The current total traffic demand is denoted as This indicates the setpoint for the total distributed flow rate of the heat exchanger or the measured value of the total circulating flow rate in the current control cycle; the first... The target flow rate of each branch road is denoted as This indicates the number of times the current control cycle is completed. The target flow rate that each branch road should achieve.

[0130] In specific implementation, the first Target flow of branch road Target flow allocation coefficient vector Compared with the current total traffic demand Multiplying these results in the proportional information in the Pareto candidate solutions being transformed into on-site executable flow targets.

[0131] In one embodiment, it is assumed that the target flow allocation coefficient vector selected in the current control cycle is: And the current total flow demand (e.g., the total flow measurement determined by the main circulation pump) is The target flow rate for each branch is calculated as follows: Target flow of branch 1: ; Target flow of branch 2: ; Target flow of branch 3: .

[0132] 2> Calculate the original valve opening command vector based on the valve flow characteristics of each branch. Original valve opening command vector The The component represents the first... The ideal opening value that a branch valve should achieve.

[0133] In practical implementation, a calibration table or fitting curve of "valve opening degree - branch flow rate" can be established in advance for each branch; during calibration, the stable branch flow rate under different valve opening degrees is recorded to obtain a one-to-one correspondence; during online operation, according to the first... Target flow of branch road By consulting a table or performing piecewise linear interpolation, the corresponding original valve opening can be obtained.

[0134] 3> Based on the current rate of change in operating conditions and the actuator's operational capability, determine the maximum permissible change in opening degree for this control cycle, where the first... The maximum permissible change in opening degree of the branch in the current control cycle is denoted as: , used to limit the The variation range of valve opening in a branch circuit within a single control cycle.

[0135] In practical implementation, It is determined by two parts: one part comes from the maximum safe operating rate of the valve actuator itself, and the other part comes from the rate of change of the current operating conditions.

[0136] In one embodiment, for example, assuming a valve actuator has a full stroke time of 30 seconds and a control cycle of 1 second, the physically maximum permissible change in opening is... (i.e., 3.3%), at which point, if the operating conditions are stable (characteristic rate of change is small), the system may... Set to smaller (i.e., 1%), to prevent overshoot and oscillation; if the load changes drastically (large characteristic rate of change), the system may... Relaxed to near the physical limit To quickly keep up with the set value.

[0137] Furthermore, low-dimensional operating condition feature vectors can be used. The first principal characteristic change rate characterizes how fast the current load changes. When the first principal characteristic change rate is large, it indicates that the system is undergoing significant changes in operating conditions. At this time, the maximum allowable opening change should be appropriately increased so that the valve can track the target faster. When the first principal characteristic change rate is small, it indicates that the system is close to steady state. At this time, the maximum allowable opening change should be reduced to suppress frequent actions caused by small fluctuations.

[0138] In one implementation, dynamic rate limiting is expressed as follows: ,in, This represents the basic change (the maximum allowable change in the basic opening can be taken as 0.01 to 0.03). This represents the correction factor (which can range from 0.3 to 0.8). This indicates the rate of change of operating conditions.

[0139] 4> Original valve opening command vector Perform branch-by-branch amplitude limiting to obtain the final valve opening command vector. , Indicates the first The valve opening command vector issued at the end of each control cycle.

[0140] In specific implementation, the first The smooth output of this branch can be expressed as: ; in, Indicates the first The branch road in the The final valve opening command for each control cycle. Indicates the first The branch road in the The final valve opening command for each control cycle; Indicates the first Original valve opening command for each branch; This represents the limiting function, used to restrict the change in opening degree within the allowable range; Indicates the first The branch road in the The maximum allowable change in opening degree per control cycle.

[0141] It should be noted that, This indicates that the desired valve actuation range will be strictly limited to... Within this safe rate range.

[0142] It should be noted that branch-by-branch limiting is more suitable for multi-branch control of heat exchangers than uniform limiting, because different branches may have different valve diameters, flow sensitivities, and response speeds. Individual limiting makes it easier to balance control accuracy and equipment lifespan.

[0143] 5> Adjust valve opening command vector based on actual branch flow feedback Make minor closed-loop corrections.

[0144] In practical implementation, in the valve opening command vector After the data is sent, the actual branch flow for the next sampling period is read and compared with the data from the first sampling period. Target flow of branch road For comparison, if a stable deviation persists in an individual branch, a small compensation amount is added to the valve opening lookup result for that branch. The compensation amount can be obtained by proportional correction, integral correction, or segmented empirical correction. Based on this, long-term errors caused by valve wear, changes in medium properties, or sensor deviations can be compensated, so that the target flow distribution ratio can be implemented more accurately.

[0145] 6> Output the final valve opening command vector This completes the online decision-making and execution for the current control cycle. In one embodiment, as an example, assuming the system has 3 branches, the output at the end of the control cycle might be a data packet in the following format: {"timestamp":1712563200,"valve_cmds":[45.2,78.5,60.0]}; Its physical meaning is: send a 45.2% opening command to the valve in branch 1, a 78.5% opening command to the valve in branch 2, and a 60.0% opening command to the valve in branch 3; where "timestamp" is a Unix timestamp (in seconds) indicating the time when the command was issued, used for recording and synchronization; "valve_cmds" is a floating-point array that lists the opening commands for each valve in branch order.

[0146] In one embodiment, such as Figure 6 As shown, the curves comparing the original calculated valve opening command (red dashed line) with the actual issued command after dynamic rate limiting (blue solid line) within a continuous control cycle reveal that the original command exhibits short-term high-frequency fluctuations or abrupt changes, while the smoothed and limited command effectively suppresses excessively rapid jumps while preserving the main adjustment trend. This invention achieves a balance between rapidly tracking changes in operating conditions and protecting the actuator, preventing overshoot oscillations. In the figure, the horizontal axis represents the control cycle number (unitless); the vertical axis represents the valve opening (unit: %).

[0147] S5. Closed-loop control of heat distribution based on variable weight decision-making and dynamic smoothing mapping of valves.

[0148] Based on the final valve opening command vector obtained from the aforementioned steps, the heat exchanger heat distribution control system completes a full closed-loop regulation process in each control cycle.

[0149] In practical engineering implementation, the controller first acquires the total primary flow rate, total secondary flow rate, inlet and outlet temperatures of each branch, branch pressure drop, and current valve opening feedback values ​​in real time through the data acquisition system, forming the current operating state vector. Then, it extracts a low-dimensional operating condition feature vector according to the method described in step S2, and calculates the global entropy yield under the current operating condition as a thermodynamic state monitoring index. Next, the controller inputs the low-dimensional operating condition feature vector into a pre-trained lightweight operating condition identification network to obtain the heat exchange pheromone field weights and pump work pheromone field weights to be used in the current control cycle. Based on this, it calls the dual-pheromone collaborative optimization algorithm established in step S3 to generate a set of Pareto candidate solutions covering different preferences within the feasible region of the flow distribution coefficient. The optimization process utilizes historical operating data and current operating condition information. It guides the search direction towards high heat exchange regions through a heat exchange pheromone field and towards low pump power consumption regions through a pump power pheromone field. Furthermore, it actively applies entropy-yield constraint projections when generating candidate flow allocation coefficient vectors, eliminating inferior solutions that, while satisfying flow and ratio constraints, cause localized heat exchange deterioration or excessive irreversible losses. After the optimization iteration, the non-dominated solutions stored in the external archive constitute the candidate allocation scheme set for the current cycle.

[0150] After obtaining the Pareto candidate set, the controller executes the variable-weight target proximity online decision-making process described in step S4. The decision module reads the flow allocation coefficient vector of each scheme in the current candidate set, along with its corresponding total heat exchange and total pump power consumption. Based on the load level and temperature difference utilization status reflected by the current low-dimensional operating condition feature vector, it dynamically generates heat exchange weights and pump power weights. Simultaneously, the decision module obtains the flow allocation coefficient vector actually executed in the previous control cycle, calculates the difference cost between the current candidate scheme and the execution result of the previous cycle, and adds it as a smoothing penalty term to the target proximity score. Based on this, the controller strikes a balance between "pursuing the current optimal thermal performance" and "maintaining smooth and continuous valve operation," selecting the target flow allocation coefficient vector with the highest correction score as the optimal allocation scheme for this cycle. Then, combined with the current total flow demand value, the target flow allocation coefficient vector is converted into the target flow setpoint for each branch. According to the pre-calibrated valve flow characteristic curves or interpolation tables for each branch, the controller converts the target flow of each branch into the corresponding original valve opening command.

[0151] To ensure the safe operation of the actuators and the stability of the system, the controller implements a branch-by-branch dynamic rate limit on the original valve opening command. The rate limit is adaptively adjusted based on the physical actuation capability of the valve actuator and the rate of change of the current low-dimensional operating condition feature vector: when the operating conditions change drastically and the load rises or falls rapidly, the upper limit of the valve actuation rate is appropriately relaxed, enabling the system to quickly track the target distribution ratio; when the operating conditions tend to be stable, the rate limit is tightened to avoid frequent valve actuation caused by small fluctuations. The valve opening command vector after the limit processing serves as the final control command, which is sent to the positioner or electric actuator of each branch control valve through the fieldbus or analog output module. After receiving the opening command, the valve actuator drives the valve core to move to the designated position, thereby changing the actual flow area of ​​each branch and realizing the redistribution of flow. In the next control cycle after the command is issued, the data acquisition system reads the actual flow feedback value of each branch again and compares it with the target flow. If a branch experiences a persistent flow deviation due to valve characteristic nonlinearity, changes in medium parameters, or sensor deviation, the controller will introduce a small compensation correction during the valve opening lookup process in subsequent cycles. This compensation can be achieved using integral-separated proportional-integral control or a static correction table based on historical deviation statistics to gradually eliminate steady-state errors.

[0152] Through the cyclical execution of the above steps, the heat exchanger heat distribution control system can achieve a complete closed loop from operating condition perception, multi-objective optimization, online decision-making to valve command generation and feedback correction. During variable load operation, the system can automatically adjust the flow distribution ratio of each branch according to the real-time thermal status, effectively reducing the total pump power consumption and irreversible losses while ensuring that the total heat exchange meets process requirements, and avoiding frequent and large valve movements, thus extending the service life of the actuators.

Claims

1. A heat exchanger heat distribution control method based on multi-objective optimization, characterized in that, Includes the following steps: S1. Collect operating data of the heat exchanger under multiple operating conditions and label the collected data; S2. Construct a unified operating state vector from the historical and current operating data of the heat exchanger, calculate the influence of each operating parameter on the entropy production rate, and use this influence as physical guidance information to weight the operating state vector. Then, extract a low-dimensional operating condition feature vector through a modal mapping method. The low-dimensional operating condition feature vector is used to characterize the dominant change direction of heat exchange potential and resistance cost under the current operating condition. S3. A dual pheromone field is used to track the high heat transfer solution and the low pump work solution respectively, and thermodynamic constraint projection and entropy production correction are applied to the candidate flow distribution coefficient vector after each generation. S4. Automatically adjust the target weights based on the current low-dimensional operating condition feature vector, and add the difference between the candidate solution and the execution result of the previous cycle as a smoothing constraint into the decision score to obtain the optimal flow distribution coefficient, and convert it into a valve opening command, and output a smooth control command through dynamic rate limiting. S5. Based on the final valve opening command vector obtained from S4, a closed-loop regulation process is completed in each control cycle.

2. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 1, characterized in that, The collected operational data includes the total primary flow rate, the total secondary flow rate, the primary inlet temperature, the primary outlet temperature, the secondary inlet temperature, the secondary outlet temperature, the inlet and outlet temperatures of each branch, the pressure drop of each branch, the current valve opening of each branch, and the flow distribution coefficient of each branch calculated from the flow rate of each branch.

3. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 1, characterized in that, S2 specifically refers to: The operating data of the heat exchanger in the current control cycle is collected to form the original operating state vector. ; For the original running state vector Each component in the vector is normalized to obtain a normalized running state vector. ; Based on the normalized running state vector Calculate the global entropy yield using thermal measurement data ; Evaluate the normalized running state vectors one by one The yield of each component to global entropy The degree of influence is used to obtain the entropy production sensitivity vector. ; Based on the entropy production sensitivity vector For the normalized running state vector Perform weighted operations to obtain the entropy-generated weighted running state vector. ; A multi-condition mode center is established using historical operating data, and the entropy-weighted operating state vector is calculated. Perform a submodal low-dimensional mapping to obtain a low-dimensional operating condition feature vector. .

4. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 1, characterized in that, S3 specifically refers to: S31. In addition to the mass conservation constraint, an entropy yield threshold judgment and gradient correction are introduced to make the randomly generated candidate flow allocation coefficient vector simultaneously satisfy numerical feasibility and thermal rationality. S32. Based on the low-dimensional working condition feature vector, adaptively determine the search focus of the two targets, and add repulsion drift to make the ants actively move away from the already over-dense area and expand the coverage of the Pareto front. S33, External Archive Maintenance and Pareto Candidate Set Output: Through non-dominated sorting and hypervolume contribution truncation, more representative non-dominated candidate solutions are retained.

5. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 1, characterized in that, S4 specifically refers to: S41. Candidate scheme selection based on variable weighted target proximity; S42. Convert the flow distribution ratio into the branch target flow, obtain the valve opening target based on the valve characteristic curve, and finally perform rate limiting and smoothing output.

6. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 4, characterized in that, S31 specifically refers to: S311, Define the flow allocation coefficient vector As an optimization variable; S312. Initialize the ant colony within the feasible region and establish a heat exchange pheromone field and a pump power pheromone field respectively; wherein, the heat exchange pheromone field is used to record the vector distribution of the flow allocation coefficients that rank first in total heat exchange in history; the pump power pheromone field is used to record the vector distribution of the flow allocation coefficients that rank last in total pump power consumption in history. S313. Generate the original candidate flow allocation coefficient vector based on dual pheromone field sampling. ; S314. Divide the original candidate flow allocation coefficient vector Project the result onto the feasible region that satisfies the mass conservation and minimum branch flow constraints, and perform the feasible region projection operation to obtain a first correction result; S315. Determine whether the global entropy productivity corresponding to the first correction result exceeds the threshold; if it exceeds the threshold, perform a second correction in the direction of reducing entropy productivity to obtain the final candidate flow allocation coefficient vector. ; S316. Calculate the final candidate flow allocation coefficient vector. The corresponding total heat exchange and total pump power consumption will be used as the basis for subsequent ranking.

7. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 4, characterized in that, S32 specifically refers to: S321. Determine the fusion weights of the heat exchange pheromone field and the pump power pheromone field based on the current low-dimensional operating condition feature vector. The fusion weights are used to characterize whether the current operating condition is more inclined to increase heat exchange or more inclined to reduce pump power consumption. S322. Select elite centers from the heat exchange pheromone field and the pump power pheromone field respectively, and generate random perturbations near each elite center. S323. Merge the two types of sampling results according to the fusion weight to obtain a new original candidate flow allocation coefficient vector. The generated new solutions will automatically shift towards the more desired objectives depending on the operating conditions; S324. Assign coefficient vectors to new original candidate flows based on the density of neighboring solutions in the external archives. Increase repulsive drift; S325. Perform feasible region projection and entropy production correction in step S31 on the original candidate flow allocation coefficient vector after adding the repulsion drift to obtain the final candidate flow allocation coefficient vector. And calculate the target value, that is, calculate the total heat exchange and total pump power consumption; S326. Update the heat exchange pheromone field and pump work pheromone field based on the target performance of the candidate solutions in this round.

8. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 4, characterized in that, S33 specifically refers to: S331. Convert the vector of all final candidate flow allocation coefficients obtained in the current iteration. Merge with existing external file A; S332. Perform non-dominated sorting on the merged solution set, and prioritize retaining candidate solutions with higher non-dominated levels. S333, When external files When the number of non-dominated candidate solutions exceeds the capacity limit, candidate solutions with smaller contributions are deleted in ascending order of hypervolume contribution. The hypervolume contribution is used to measure the contribution of a candidate solution to the current Pareto front coverage area.

9. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 5, characterized in that, S41 specifically refers to: S411. Read the Pareto candidate set of the current control cycle and extract the target value and flow allocation coefficient vector for each candidate solution; S442. Generate dynamic target weights based on the current low-dimensional working condition feature vector; S443. Determine the positive and negative ideal target points based on the current candidate set, and calculate the bullseye proximity for each candidate solution. ; S444. Based on the actual execution results of the previous control cycle, a smoothing penalty is added to each candidate solution to obtain a corrected score. ; S445. Select the corrected score from all candidate solutions. The largest candidate solution is used as the target flow allocation coefficient vector for the current control cycle. .

10. The heat exchanger heat distribution control method based on multi-objective optimization according to claim 5, characterized in that, S42 specifically refers to: S421. Based on the current total flow demand and target flow allocation coefficient vector Calculate the target flow rate for each branch; S422. Calculate the original valve opening command vector based on the valve flow characteristics of each branch. ; S423. Determine the maximum allowable opening change in this control cycle based on the current rate of change of operating conditions and the actuator's operating capability; S424, Original valve opening command vector Perform branch-by-branch amplitude limiting to obtain the final valve opening command vector. ; S425, Valve opening command vector based on actual branch flow feedback. Make minor closed-loop corrections; S426, Output the final valve opening command vector This enables online decision-making and execution for the current control cycle.

Images