Intelligent heat transfer prediction method and system for heat exchanger based on deep learning

Abstract

The application discloses a kind of heat exchanger intelligent heat transfer prediction method and system based on deep learning.The method first obtains input characteristic data containing fluid state and heat transfer state;Second, a cascade calibration architecture is used for prediction: an initial heat transfer performance prediction value is generated using a first machine learning model, and a second machine learning model is used to predict and correct the prediction residual of the former, and the final prediction value with high precision is obtained after superposition;Finally, based on the final prediction value and flow resistance data, an operation parameter adjustment scheme is generated by a multi-objective optimization algorithm, which takes into account heat transfer efficiency and energy consumption.The application combines physical simulation and AI, cascade residual calibration, dynamic trade-off optimization and self-adaptive updating mechanism, solves the problems of low prediction accuracy, inability to trade-off optimization and poor adaptability of the prior art, significantly improves the prediction accuracy, decision-making scientificity and reliability of long-term operation of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of heat transfer enhancement and energy saving, and in particular to a heat exchanger intelligent heat transfer prediction method and system based on deep learning. BACKGROUND

[0002] As a key device for energy conversion and utilization, heat exchangers are widely used in industrial production, energy development, environmental protection and other fields. Its heat transfer performance directly affects the energy efficiency and operating cost of the system. However, the existing heat transfer performance prediction and optimization methods of heat exchangers still have many limitations.

[0003] Firstly, for heat exchangers with built-in heat transfer enhancement elements such as spiral inserts, traditional prediction models rely on simplified theoretical assumptions or limited experimental data, lack of in-depth mechanism understanding and accurate quantitative means for complex three-dimensional flow patterns such as spiral flow, and are difficult to accurately characterize the internal flow mechanism, the formation and evolution of secondary flow, and the quantitative relationship between heat transfer enhancement and resistance increase.

[0004] Secondly, empirical formula models based on simplified theory have narrow applicable working condition range and low prediction accuracy. Traditional pure data-driven models, such as simple neural networks, lack physical law constraints, are prone to overfitting, have poor generalization ability for new working conditions outside the training set, and the prediction results are often unreliable.

[0005] Thirdly, how to effectively combine complex flow mechanism (physical world) with modern intelligent algorithms (data world) is a core challenge in this field. Lack of in-depth analysis of mechanism cannot provide reliable and physically meaningful data basis for intelligent prediction. On the contrary, lack of advanced intelligent prediction means also limits the in-depth exploration and utilization of the rules behind complex mechanisms.

[0006] In addition, heat transfer enhancement and reduction of flow resistance (i.e. reduction of pump power consumption) are a pair of eternal contradictions in heat exchanger design and operation. Existing methods often predict one of them in isolation, lack of intelligent decision-making tools that can consider both and dynamically optimize according to actual needs.

[0007] Finally, the actual running heat exchanger will change its heat transfer performance slowly over time (i.e. "concept drift") due to factors such as fouling, corrosion, and wear, and the operating conditions may also fluctuate at any time. The prediction accuracy of a one-time trained static model will decrease over time and cannot adapt to such dynamic changes.

[0008] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0009] Therefore, the present application provides a heat exchanger intelligent heat transfer prediction method and system based on deep learning to solve the problem of insufficient prediction accuracy of heat transfer performance in the prior art. By constructing a cascaded machine learning model, an initial heat transfer performance prediction value is generated using a first machine learning model, and the residual of the initial prediction value is predicted and calibrated using a second machine learning model, thereby improving the prediction accuracy of the heat transfer performance. On this basis, an operation parameter adjustment scheme of the heat exchanger is generated to realize the optimized control of the heat exchanger operation.

[0010] The present application embodiment provides a heat exchanger intelligent heat transfer prediction method based on deep learning, comprising the following steps:

[0011] Obtain the input feature data of the heat exchanger, and the input feature data includes parameters representing the fluid flow state and heat transfer state;

[0012] Using the first machine learning model, the input feature data is processed to generate an initial heat transfer performance prediction value;

[0013] Based on the input feature data and the initial heat transfer performance prediction value, the second machine learning model is trained to predict the residual between the initial heat transfer performance prediction value and the corresponding true heat transfer performance value;

[0014] When predicting new input feature data, a new initial heat transfer performance prediction value is obtained through the first machine learning model, and a corresponding prediction residual value is obtained through the second machine learning model;

[0015] The new initial heat transfer performance prediction value and the prediction residual value are superimposed to obtain the final heat transfer performance prediction value;

[0016] Based on the final heat transfer performance prediction value, the flow resistance data is combined to generate an operation parameter adjustment scheme of the heat exchanger.

[0017] In some optional embodiments, the input feature data is obtained by the following method:

[0018] Collect real-time operation data of the heat exchanger, and the real-time operation data includes fluid flow rate, temperature and pressure;

[0019] Based on the real-time operation data, the Reynolds number and the Nusselt number are calculated by a numerical simulation method, and the Reynolds number and the Nusselt number are used as part of the input feature data.

[0020] In some optional embodiments, the first machine learning model is a support vector machine model, and the second machine learning model is a random forest model.

[0021] In some optional embodiments, the step of generating the operation parameter adjustment scheme of the heat exchanger specifically comprises:

[0022] constructing a multi-objective optimization model aiming at maximizing the final heat transfer performance prediction value and minimizing the flow resistance;

[0023] solving the multi-objective optimization model by using a preset multi-objective optimization algorithm to obtain a set of Pareto optimal solution set representing different trade-off relationships;

[0024] selecting a target solution from the Pareto optimal solution set according to an externally input operating strategy weight, and determining an operating parameter adjustment scheme based on the target solution.

[0025] In some optional embodiments, the operating strategy weight is a value between 0 and 1, which is used to adjust the relative importance between the two objectives of maximizing the heat transfer performance and minimizing the flow resistance.

[0026] In some optional embodiments, after generating the final heat transfer performance prediction value, the following steps are further included:

[0027] monitoring the error flow between the final heat transfer performance prediction value and the actual measured value;

[0028] using a drift detection algorithm to analyze the statistical distribution of the error flow, and triggering a model update instruction when a preset degree of change in the statistical distribution is detected;

[0029] In response to the model update instruction, incrementally training the first machine learning model and the second machine learning model using newly collected data.

[0030] In some optional embodiments, the incremental training is performed by fine-tuning the original model parameters through transfer learning.

[0031] In some optional embodiments, before generating the initial prediction value using the first machine learning model, the following steps are further included:

[0032] using a cross-validation method to optimize the hyperparameters of the first machine learning model, including the penalty coefficient and the kernel function parameter.

[0033] In some optional embodiments, the second machine learning model is also used to calculate the contribution of each parameter in the input feature data to the prediction result, and generate a feature importance ranking.

[0034] The embodiment of the present application provides a heat exchanger intelligent heat transfer prediction system based on deep learning, which is used to implement the heat exchanger intelligent heat transfer prediction method based on deep learning as described above, and comprises:

[0035] a feature acquisition module for acquiring input feature data of the heat exchanger, the input feature data including parameters representing the fluid flow state and the heat transfer state;

[0036] a cascade prediction module, the cascade prediction module comprising: a first machine learning model configured to generate a new initial heat transfer performance prediction value based on new input feature data; and a second machine learning model configured to predict a residual between the new initial heat transfer performance prediction value and a corresponding actual heat transfer performance value based on the input feature data;

[0037] a result fusion module configured to superimpose the new initial heat transfer performance prediction value and the residual value predicted by the second machine learning model to output a final heat transfer performance prediction value;

[0038] a decision generation module configured to generate an operation parameter adjustment scheme of the heat exchanger based on the final heat transfer performance prediction value and a flow resistance value.

[0039] In some optional embodiments, the decision generation module is further configured to:

[0040] construct a multi-objective optimization model aiming to maximize the final heat transfer performance prediction value and minimize the flow resistance value;

[0041] obtain a set of Pareto optimal solutions by solving the multi-objective optimization model;

[0042] select a target solution from the Pareto optimal solutions according to an externally input operation strategy weight, and determine the operation parameter adjustment scheme based on the target solution.

[0043] In some optional embodiments, the system further comprises:

[0044] a model self-adaptive updating module configured to:

[0045] monitor an error stream between the final heat transfer performance prediction value and an actual measured value;

[0046] when a preset change in a statistical distribution of the error stream is detected, perform incremental training on the first machine learning model and the second machine learning model using newly collected data.

[0047] In some optional embodiments, the first machine learning model is a support vector machine model, and the second machine learning model is a random forest model; and the second machine learning model is further configured to calculate a contribution degree of each parameter in the input feature data to a prediction result.

[0048] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application.

[0049] The heat exchanger intelligent heat transfer prediction method and system based on deep learning has the following beneficial effects:

[0050] The application significantly improves the prediction accuracy of the heat transfer performance of the heat exchanger by constructing a cascade residual correction heat transfer performance prediction model, and overcomes the defects of low precision and poor generalization of traditional models. The method can dynamically balance the heat transfer benefit and flow resistance, select the best operation parameter adjustment scheme in the Pareto optimal solution set according to the user set operation strategy weight, realize the intelligent decision of the heat exchanger operation, improve the energy utilization efficiency, reduce the operation cost, and has remarkable economic benefits. BRIEF DESCRIPTION OF DRAWINGS

[0051] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments with reference to the drawings.

[0052] Figure 1 is a flow chart of a deep learning-based intelligent heat transfer prediction method of a heat exchanger according to an embodiment of the application;

[0053] Figure 2 is a structural schematic diagram of a deep learning-based intelligent heat transfer prediction system of a heat exchanger according to an embodiment of the application. DETAILED DESCRIPTION

[0054] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the implementations set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects to those skilled in the art. Features described in the description, examples, or claims can be combined in any suitable manner in one or more embodiments.

[0055] In addition, the accompanying drawings are only schematic and are non-limiting exact representations of embodiments of the application. Identical components have been given the same reference numerals in the various drawings and so will not be described again. Some of the blocks in the drawings are functional blocks that do not necessarily have to be implemented in the described way. The functional blocks can be implemented in software or hardware or a combination thereof.

[0056] The flowcharts shown in the drawings are only illustrative and do not necessarily include all the steps. For example, some steps can be further divided, and some steps can be combined or partially combined, so the actual execution order may be changed according to the actual situation.

[0057] The present application relates to a kind of based on deep learning's heat exchanger intelligent heat transfer prediction method, utilize machine learning model to accurately predict the heat transfer performance of heat exchanger, and realize the optimization adjustment of operating parameter based on prediction result.Heat transfer process is a complex nonlinear process, is influenced by multiple factors, and traditional prediction method based on physical model is difficult to accurately describe this complexity.On the other hand, machine learning method, especially deep learning method, has strong nonlinear fitting ability, can learn from a large amount of data and extract hidden rule, to realize the accurate prediction of heat transfer performance.The cascade prediction model used in the present application can further improve the prediction accuracy through the cooperative work of multiple machine learning models.Specifically, the first machine learning model is used to capture the global trend of heat transfer performance, and the second machine learning model is used to correct the prediction error of the first model, to realize the fine prediction of heat transfer performance.In addition, the present application also combines multi-objective optimization method, combines heat transfer performance prediction with flow resistance prediction, to realize the optimization adjustment of heat exchanger operating parameter, while ensuring heat transfer efficiency, as far as possible reduce energy consumption.Therefore, by applying machine learning and multi-objective optimization techniques, the present application can effectively solve the problems of low prediction accuracy, poor generalization ability and inability to consider heat transfer and resistance in the prior art, realize the accurate prediction of heat exchanger heat transfer performance and the optimization adjustment of operating parameter, to improve energy utilization efficiency and reduce operating cost.

[0058] As Figure 1 shown, the present application provides a kind of based on deep learning's heat exchanger intelligent heat transfer prediction method, including the following steps:

[0059] S100, the input characteristic data of heat exchanger is acquired, and input characteristic data includes the parameter of characterizing fluid flow state and heat transfer state.Input characteristic data is the key information of describing heat exchanger operating state, and its quality directly influences the accuracy of subsequent prediction model.In the field, fluid flow state is usually characterized by flow rate, Reynolds number and other parameters, and heat transfer state is characterized by temperature, pressure, Nusselt number and other parameters.In the present application, input characteristic data includes but is not limited to temperature, pressure, flow rate and other original data collected directly by sensor, and dimensionless parameters such as Reynolds number (Re), Prandtl number (Pr) and Nusselt number (Nu) calculated from these original data.It can also include the geometric parameters of heat exchanger, such as helical pipe diameter, pitch and the like.These parameters collectively reflect the flow and heat transfer characteristics inside heat exchanger.These characteristic data can be collected in real time by deploying high-precision sensor array at key positions of heat exchanger, and transmitted to cloud storage system through Internet of Things protocol.After the collected data is cleaned and pretreated, key physical characteristics are extracted in combination with numerical simulation results.

[0060] S200, using the first machine learning model, processing the input feature data to generate an initial heat transfer performance prediction value. The first machine learning model plays a role of preliminary prediction in the present application, which is used to extract the basic trend of heat transfer performance from the input feature data. The machine learning model can adopt support vector machine (SVM), neural network, decision tree and other algorithms. SVM model is good at processing small sample, high dimension and nonlinear problem, and can well capture the global basic trend in heat transfer process; the first machine learning model can adopt radial basis function (RBF kernel) to process the complex nonlinear relationship between input features and output, and through cross validation and grid search method, the hyperparameters of SVM are optimized.

[0061] S300, based on the input feature data and the initial heat transfer performance prediction value, training the second machine learning model to predict the residual error between the initial heat transfer performance prediction value and the corresponding real heat transfer performance value. By introducing the residual error calibration mechanism, the prediction accuracy can be effectively improved. The second machine learning model is specially used to learn and predict the residual error (i.e. error) of the first stage model, which can adopt random forest (RF), gradient boosting tree (GBDT) and other ensemble learning models. These models are not sensitive to high-dimensional data and complex interaction between features, and are very suitable for fitting the complex structure of residual error left by SVM model. Specifically, for each sample in the training set, the prediction residual error of SVM model is calculated, and a new training data set is constructed, whose input features are still original, but its training target becomes residual error. By training the RF model to learn the mapping relationship from input features to residual error, a residual prediction model is obtained.

[0062] S400, when predicting new input feature data, the new initial heat transfer performance prediction value is obtained by the first machine learning model, and the corresponding prediction residual value is obtained by the second machine learning model. For new input feature data, the preliminary prediction value is obtained by using the trained first machine learning model (such as SVM), and the possible error is predicted by using the trained second machine learning model (such as RF).

[0063] S500, superimposing the new initial heat transfer performance prediction value and the prediction residual value to obtain the final heat transfer performance prediction value. This step fuses the prediction results of the two models to realize accurate prediction of heat transfer performance. The preliminary prediction value and the prediction residual value are superimposed to obtain the final calibrated prediction value, which is the final heat transfer performance prediction result of the embodiment.

[0064] S600, based on the final heat transfer performance prediction value, combined with flow resistance data, generate the operation parameter adjustment scheme of the heat exchanger. This step applies the prediction result to the actual heat exchanger operation optimization. This step aims to solve the contradiction between heat transfer enhancement and flow resistance (energy consumption), and convert the prediction ability of the model into intelligent decision-making that can guide the actual operation. The flow resistance data can be obtained through CFD simulation results. Combined with the heat transfer performance prediction value and the flow resistance data, a multi-objective optimization model is established, for example, maximizing the heat transfer performance and minimizing the flow resistance. Multi-objective optimization algorithms such as non-dominated sorting genetic algorithm II (NSGA-II) can be used to solve the model and obtain a set of Pareto optimal solutions. Based on the Pareto optimal solution, combined with the user-set operation strategy weight, the operation parameter adjustment scheme of the heat exchanger is generated, such as the best flow rate, the best inlet temperature, etc.

[0065] The embodiment of the present application provides an intelligent heat transfer prediction method for overcoming the limitations of a single model through a two-stage cascade calibration architecture. The first stage model captures the global trend, and the second stage model fills in the details, achieving high-precision prediction. Through the solution of the multi-objective optimization model, decision support is provided for the intelligent operation of the heat exchanger, and the comprehensive performance of all targets is improved without sacrificing other targets.

[0066] In some embodiments, the input feature data is obtained by: collecting real-time operation data of the heat exchanger, the real-time operation data including fluid flow rate, temperature and pressure; based on the real-time operation data, the Reynolds number and the Nusselt number are calculated through a numerical simulation method, and the Reynolds number and the Nusselt number are taken as part of the input feature data. The acquisition of input feature data combines real-time operation data and numerical simulation, thereby obtaining more comprehensive and more physically meaningful feature parameters. The real-time operation data such as fluid flow rate, temperature and pressure directly reflect the instantaneous working condition of the heat exchanger, which is a necessary input of the data-driven model. The dimensionless parameters such as Reynolds number (Re) and Nusselt number (Nu) reflect the essential characteristics of fluid flow state and heat transfer efficiency, which can effectively improve the generalization ability of the model. Through CFD numerical simulation, the original operation data can be converted into these high-order physical characteristics, making up for the deficiency of the direct measurement ability of the sensor.

[0067] Specifically, real-time operational data can be acquired using various sensors. For example, turbine flow meters and electromagnetic flow meters can be used for fluid velocity; thermocouples and resistance temperature detectors (RTDs) can be used for temperature; and piezoresistive pressure sensors and capacitive pressure sensors can be used for pressure. Sensor signals are amplified, filtered, and converted from analog to digital before being transmitted to the data processing unit via the data acquisition system. For numerical simulation, commercial CFD software such as ANSYS Fluent and COMSOL Multiphysics, as well as open-source software such as OpenFOAM, can be used. When building the CFD model, the geometry of the heat exchanger must be accurately reproduced. For example, for a helical tube heat exchanger, the diameter and pitch of the helical tube must be accurately modeled, and boundary conditions such as inlet velocity and inlet temperature must be set according to the actual operating conditions. The choice of turbulence model has a significant impact on the accuracy of the simulation results; k-ε, k-ω, and SST k-ω models can be selected depending on the specific operating conditions. By solving the continuity equation, momentum equation, and energy equation, the flow field and temperature field distribution inside the heat exchanger can be obtained. Then, based on the simulation results, the Reynolds number and Nusselt number are calculated using the post-processing module. The Reynolds number is calculated using the formula Re = (ρ * v * D) / μ, and the Nusselt number is calculated using the formula Nu = (h * D) / k, where ρ is the fluid density, v is the fluid velocity, D is the characteristic dimension (e.g., pipe diameter), μ is the fluid dynamic viscosity, h is the convective heat transfer coefficient, and k is the fluid thermal conductivity.

[0068] By employing the above technical solution, the real-time operating status of the heat exchanger can be effectively combined with its internal flow mechanism, providing more representative and discriminative input features for machine learning models, thereby significantly improving the accuracy and reliability of heat transfer performance prediction. Compared to traditional methods that rely solely on sensor data, this solution can better capture the impact of complex flow phenomena, such as helical flow and secondary flow, on the heat transfer process, and is more robust to sensor noise and changes in operating conditions.

[0069] In some embodiments, the first machine learning model is a support vector machine (SVM) model, and the second machine learning model is a random forest (RF) model. These are the preferred types of machine learning models in this embodiment of the invention. Support vector machines (SVM) excel at handling small sample sizes, high dimensionality, and nonlinear problems, effectively capturing global trends in the heat transfer process and enabling preliminary predictions of heat transfer performance. Random forests (RF), as an ensemble learning model, effectively handle complex interactions between high-dimensional data and features, are insensitive to noise, and are better suited for fitting the structurally complex residuals present in the prediction results of SVM models, thereby achieving local calibration of the heat transfer performance prediction results.

[0070] As the kernel function of the SVM model, a radial basis function (RBF kernel), a polynomial kernel function, or a sigmoid kernel function can be selected, and the application does not limit this. The hyperparameters of the SVM, including the penalty coefficient C and the kernel function parameter gamma, are optimized by cross-validation and grid search methods.

[0071] Further, the random forest model is composed of multiple decision trees, and the generation process of each decision tree depends on the random selection of training samples and features, so as to ensure the generalization ability of the model. The number of decision trees in the random forest model can be adjusted according to actual conditions. Other hyperparameters of the random forest model, such as the maximum depth of the decision tree and the minimum number of samples required for splitting, can also be optimized by cross-validation and other methods.

[0072] The cascade machine learning prediction framework can effectively improve the prediction accuracy and robustness of the heat transfer performance by complementing the advantages of the SVM and RF models. Compared with using the SVM or RF model alone, the root mean square error (RMSE) can be reduced by 15%-30%, the adaptability to complex working conditions is stronger, and the prediction result is more accurate and reliable.

[0073] In some embodiments, the step of generating the operation parameter adjustment scheme of the heat exchanger specifically comprises:

[0074] A multi-objective optimization problem is constructed to maximize the final heat transfer performance prediction value and minimize the flow resistance;

[0075] A preset multi-objective optimization algorithm is used to solve the model to obtain a set of Pareto optimal solutions representing different trade-off relationships;

[0076] According to the externally input operation strategy weight, a target solution is selected from the set of Pareto optimal solutions, and the operation parameter adjustment scheme is determined based on the target solution.

[0077] The step of generating the operation parameter adjustment scheme of the heat exchanger in the embodiment of the application solves the inherent contradiction between the heat transfer performance and the flow resistance, i.e., improving the heat transfer efficiency usually increases energy consumption, and vice versa. The application constructs a multi-objective optimization model to integrate the two conflicting objectives into a unified mathematical framework. By using a preset multi-objective optimization algorithm, such as the non-dominated sorting genetic algorithm II (NSGA-II), the multi-objective particle swarm optimization algorithm (MOPSO), or the epsilon-constraint method, a set of Pareto optimal solutions can be searched in the decision variable space, and the solution set represents all possible schemes with the minimum flow resistance under different heat transfer performance levels. Each solution corresponds to a specific set of heat exchanger operation parameters, such as fluid flow rate, inlet temperature, or heat exchanger geometric structure parameters.

[0078] The Pareto optimal solution set constitutes the Pareto frontier, which intuitively demonstrates the trade-off relationship between heat transfer performance and flow resistance. The externally input operating strategy weight w (0 ≤ w ≤ 1) represents the degree of preference of the user or control system for the two objectives. When w is close to 1, it indicates that more emphasis is placed on maximizing heat transfer performance, allowing higher energy consumption; when w is close to 0, it indicates that more emphasis is placed on reducing energy consumption, which may sacrifice certain heat transfer performance. When w is equal to 0.5, it indicates equal emphasis on both. Based on this weight, a target solution is selected on the Pareto frontier, which represents the optimal balance point between heat transfer performance and flow resistance under the current operating strategy. Based on the target solution, the system can determine the corresponding operating parameter adjustment scheme, for example, adjusting the fluid flow rate to a specific value to achieve the optimal operating state under the strategy.

[0079] For example, if the goal is to maximize energy utilization efficiency, the weight can be set to a lower value, and the system will recommend a low flow rate operating scheme. If the goal is to maximize production efficiency, the weight can be set to a higher value, and a high flow rate operating scheme will be recommended. This dynamic adjustment allows the heat exchanger to flexibly adapt to different production demands and economic environments, achieving energy saving and cost reduction. By balancing heat transfer benefits and energy consumption costs and dynamically optimizing according to actual needs, the overall operating efficiency and economic benefits of the heat exchanger system can be improved, while meeting specific operating condition requirements and significantly reducing operating costs.

[0080] In some embodiments, the operating strategy weight w is a value between 0 and 1, which is used to adjust the relative importance between maximizing heat transfer performance and minimizing flow resistance. The technical meaning of the operating strategy weight w is that when w is close to 1, the system focuses on improving heat transfer performance, even if it means higher energy consumption; conversely, when w is close to 0, the system focuses more on reducing energy consumption, possibly sacrificing certain heat transfer performance. This trade-off mechanism allows operators or automatic control systems to flexibly adjust the operating state of the heat exchanger according to actual operating requirements, thereby maximizing economic benefits. For example, during periods of high electricity costs, the w value can be reduced to save energy, while during high-load operation, the w value can be increased to ensure heat transfer efficiency. The weight w can be manually set by the user or automatically optimized by the upper control system according to the economic model. By dynamically adjusting the operating strategy weight, the system can adapt to different operating conditions and economic environments, achieving intelligent energy management. This flexibility and adaptability significantly improves the economic efficiency and sustainability of the heat exchanger operation.

[0081] In the present application, after generating the final heat transfer performance prediction value, the following steps are included:

[0082] Monitoring the error flow between the final heat transfer performance prediction value and the actual measured value;

[0083] The statistical distribution of the error stream is analyzed by using a drift detection algorithm, and when a preset degree of change in the statistical distribution is detected, a model updating instruction is triggered;

[0084] In response to the model updating instruction, the first machine learning model and the second machine learning model are incrementally trained using newly collected data.

[0085] The embodiments of the present application aim to solve the performance degradation problem of heat exchangers caused by factors such as fouling and corrosion during long-term operation. By continuously monitoring the deviation between the predicted value and the actual measured value, the drift detection algorithm is used to determine whether the model prediction accuracy is reduced, and when a conceptual drift is determined, the model is updated, thereby ensuring that the model can adapt to changing environments and maintain long-term prediction accuracy.

[0086] In some embodiments, monitoring the error stream can be achieved in various ways. For example, a sliding window method can be used to periodically calculate the mean, variance, and other statistical quantities of the errors within the window. The actual measured value of the error can be measured by a high-precision sensor installed on the heat exchanger in real time. The drift detection algorithm can select DDM (Drift Detection Method), EDDM (Early Drift Detection Method), or ADWIN (Adaptive Windowing) algorithm, and select the appropriate drift detection algorithm according to the actual application scenario. The preset degree of change can be set according to actual needs, for example, the error mean can be set to trigger model updating when it exceeds three times the historical mean standard deviation. The model updating instruction can be sent to the model training module to trigger the incremental training process.

[0087] Incremental training, as an efficient implementation of model updating, avoids completely retraining the model and saves computing resources and time. Specifically, incremental training can use online learning, that is, whenever new data is generated, the model parameters are immediately updated with these data. Alternatively, batch updating can also be used, that is, a certain amount of new data is accumulated, and then the model is updated. The learning rate of incremental training can be set to be less than the initial training learning rate to avoid overfitting to new data.

[0088] In some embodiments, in addition to monitoring the statistical distribution of error streams, the drift detection algorithm can also make judgments based on expert knowledge. For example, when it is detected that the operating parameters of the heat exchanger are outside the normal range, expert experience can be used to determine whether model updating needs to be triggered. The drift detection algorithm can also use ensemble learning to combine multiple drift detectors to improve the accuracy and robustness of detection. The frequency of model updating can be adjusted according to actual conditions, for example, model updating can be performed during periods of low system load to reduce the impact on normal operation.

[0089] The technical effect of using error stream monitoring and adaptive model evolution mechanism is that it can effectively deal with the concept drift problem in the operation of the heat exchanger, ensure the long-term prediction accuracy of the model, reduce manual intervention, reduce maintenance costs, and improve the stability and reliability of the operation of the heat exchanger.

[0090] In some embodiments, incremental training is performed through transfer learning to fine-tune the original model parameters. This technical feature refers to using transfer learning strategy to update model parameters after concept drift occurs to quickly adapt to new data distribution. Specifically, load the current online serving SVM and RF model parameters, and use the incremental training set to retrain these two models with a small learning rate, thereby fine-tuning the original knowledge to adapt to the new data distribution. Transfer learning can specifically select fine-tuning strategy or feature extraction strategy.

[0091] This technical solution can avoid time-consuming and labor-intensive full retraining from scratch, and only requires a small amount of computing resources to quickly update the model, enabling the system to self-optimize more frequently and more lightweightly, and respond faster to changes. At the same time, the fine-tuning method ensures that the model adapts to the new environment while not "forgetting" the general physical laws learned from the vast amount of historical data, ensuring the long-term stability and high generalization ability of the model, and improving the accuracy and reliability of model prediction.

[0092] In some embodiments, to further improve the prediction accuracy of the global trend by the first machine learning model, such as a support vector machine model, before generating the initial prediction value by the first machine learning model, the method further comprises: optimizing the hyperparameters of the first machine learning model by a cross-validation method, the hyperparameters including a penalty coefficient and a kernel function parameter. The cross-validation is a model selection method that divides the data set into multiple subsets, uses one subset as the validation set and the rest as the training set, and evaluates the performance of the model multiple times to select the best hyperparameter combination. The penalty coefficient (C) is used to control the tolerance of the model to misclassified samples. The larger the C value, the more the model tends to accurately fit the training data, but it may lead to overfitting. The smaller the C value, the higher the tolerance of the model to misclassification, which helps to improve the generalization ability of the model. The kernel function parameter, such as γ in RBF kernel, determines the influence range of a single training sample on the model. The larger the γ value, the smaller the influence range, and the model may pay more attention to local details. The smaller the γ value, the larger the influence range, and the model pays more attention to global trends.

[0093] In specific implementations, the k-fold cross-validation method can be used, for example, the data set is divided into 5 or 10 parts, one part is selected as the validation set and the rest is used as the training set each time, and the training and validation process is repeated to finally select the hyperparameter combination with the smallest average validation error. The optimization algorithm can use grid search, random search or Bayesian optimization method to search for the optimal solution in the predefined hyperparameter space. For example, the value range of the penalty coefficient C can be [0.1, 1, 10, 100], and the value range of the kernel function parameter γ can be [0.01, 0.1, 1, 10]. The selection of hyperparameters is not limited to the penalty coefficient and the kernel function parameter, but can also include the selection of the loss function, the number of iterations, etc.

[0094] By optimizing the hyperparameters through cross-validation, the model overfitting or underfitting can be effectively avoided, the generalization ability and prediction accuracy of the model can be improved, and the first machine learning model can more accurately capture the global basic trend in the heat transfer process.

[0095] In some embodiments, the second machine learning model is also used to calculate the contribution of each parameter in the input feature data to the prediction result and generate a feature importance ranking. This technical feature aims to identify the key factors affecting the heat transfer performance of the heat exchanger and provide guidance for optimizing the operating parameters. By analyzing the information gain or Gini coefficient reduction of each feature in the decision tree splitting of the random forest model, the importance of each feature is quantified. For example, the Reynolds number Re can be identified as the most important feature, indicating that the flow rate has a significant impact on heat transfer. This feature importance ranking can be used to simplify the model, guide sensor deployment, and provide intuitive operation optimization suggestions for operators. In some embodiments, other feature selection algorithms such as Permutation Importance or SHAP values can also be used to calculate feature importance. By identifying key features, the operating parameters of the heat exchanger can be adjusted more effectively, thereby improving heat transfer efficiency while reducing energy consumption.

[0096] As shown in Figure 2 The embodiment of the present application also provides a deep learning-based intelligent heat transfer prediction system for heat exchangers, which is used to implement the deep learning-based intelligent heat transfer prediction method for heat exchangers described above. The system includes multiple modules that work together to achieve accurate heat transfer prediction and intelligent operation parameter optimization.

[0097] The feature acquisition module M100 collects key parameters during the operation of the heat exchanger as input for subsequent model prediction. These input feature data include parameters representing the flow state and heat transfer state of the fluid, such as flow rate, temperature, pressure, etc. These data are collected in real time by a high-precision sensor array and transmitted to a cloud storage system through an Internet of Things (IoT) protocol to form an original time series data set. This step is the basis for intelligent prediction, and the quality of the data directly affects the accuracy of the prediction model. More specifically, the flow state parameters can include the Reynolds number (Re), and the heat transfer state parameters can include the Nusselt number (Nu), the Prandtl number (Pr), etc. These parameters can more effectively describe the flow and heat transfer characteristics of the fluid.

[0098] A cascaded prediction module M200, which adopts an innovative two-stage cascaded architecture to significantly improve the prediction accuracy. The module includes a first machine learning model and a second machine learning model. The first machine learning model, such as a support vector machine (SVM), is configured to generate an initial heat transfer performance prediction value based on input feature data. SVM is good at handling small sample, high-dimensional nonlinear problems, and can well capture the global basic trend in the heat transfer process. The second machine learning model, such as a random forest (RF), is configured to predict the residual between the initial heat transfer performance prediction value output by the first machine learning model and the corresponding true heat transfer performance value based on input feature data. The RF model can learn and predict the residual (i.e. error) of the first stage model, and is not sensitive to the complex interaction between high-dimensional data and features through ensemble learning, and is suitable for fitting complex structure residuals.

[0099] A result fusion module M300, which functions to integrate the prediction results of the two stages of the cascaded prediction module M200. Specifically, the module superimposes the initial heat transfer performance prediction value and the residual value predicted by the second machine learning model to output the final heat transfer performance prediction value. This cascaded residual calibration architecture can fully exploit the complementary advantages of SVM and RF models, thereby improving the overall prediction accuracy and robustness.

[0100] A decision generation module M400, which receives the final heat transfer performance prediction value and a flow resistance value, and generates an operation parameter adjustment scheme for the heat exchanger based thereon. The flow resistance value can be obtained through CFD simulation or by establishing a corresponding prediction model. The core of the module is to solve the contradiction between heat transfer enhancement and flow resistance, and to convert the prediction ability of the model into intelligent decisions that can guide actual operation. The module can construct a multi-objective optimization model with the goal of maximizing the heat transfer performance prediction value and minimizing the flow resistance, and use efficient multi-objective optimization algorithms such as non-dominated sorting genetic algorithm II (NSGA-II) to solve the Pareto optimal front, and then select the best operation parameters according to the externally input operation strategy weight. The operation parameter adjustment scheme can be a strategy to adjust flow rate, inlet temperature, etc.

[0101] The heat exchanger intelligent heat transfer prediction system provided by the embodiments of the present application improves the prediction accuracy of heat transfer performance through the cascaded prediction module M200, and realizes intelligent optimization of the operation parameters of the heat exchanger through the decision generation module M400, thereby improving energy utilization efficiency and reducing industrial operation cost.

[0102] In some embodiments, the decision generation module M400 is further configured to: construct a multi-objective optimization model targeting at maximizing the final heat transfer performance prediction value and minimizing the flow resistance value; obtain a set of Pareto optimal solutions by solving the multi-objective optimization model; select a target solution from the Pareto optimal solutions according to an externally inputted operation strategy weight, and determine the operation parameter adjustment scheme based on the target solution.

[0103] Specifically, the decision generation module M400 aims to solve the inherent contradiction between heat transfer efficiency and energy consumption during the operation of the heat exchanger. The module first constructs a multi-objective optimization model, in which the final heat transfer performance prediction value, e.g., Nusselt number Nu, is maximized to represent the goal of improving heat transfer efficiency, and the flow resistance, e.g., friction factor f or pressure drop ΔP, is minimized to represent the goal of reducing energy consumption. The flow resistance can be obtained from CFD simulation calculation or predicted by establishing a separate machine learning model. The decision variables of the multi-objective optimization model include adjustable operation parameters such as fluid flow rate v. Subsequently, the module uses a multi-objective optimization algorithm, e.g., non-dominated sorting genetic algorithm NSGA-II, or other evolutionary algorithms, to solve the model and obtain a set of Pareto optimal solutions. Each solution in the Pareto optimal solution set represents a scheme that balances between heat transfer efficiency and energy consumption. Each solution corresponds to a specific setting value of fluid flow rate or other adjustable parameters. To facilitate the decision-making of the operator, the module introduces an externally inputted operation strategy weight w (0≤w≤1). The value of w can be dynamically adjusted by the user according to real-time electricity prices, production demands, etc. The closer w is to 1, the more emphasis on heat transfer performance, and the closer w is to 0, the more emphasis on energy saving. The module selects an optimal solution from the Pareto optimal solution set according to the value of w through weighted calculation or other decision rules, and sets the operation parameter setting value corresponding to the solution as the operation parameter adjustment scheme of the heat exchanger. For example, the solution that maximizes the value of the dynamic operation strategy indicator DOSI(x, w) = w * Nu_norm(x) - (1-w) * f_norm(x) can be selected.

[0104] In this way, the decision generation module M400 realizes the intelligent trade-off between heat transfer efficiency and energy consumption, so that the heat exchanger can operate in the most economical way under the premise of meeting production demands. This scheme simplifies the complex multi-objective optimization model into the adjustment of a single weight w, which is easy to operate and the result is intuitive and easy to understand.

[0105] In some embodiments, the model adaptive updating module is configured to monitor an error stream between the final heat transfer performance prediction value and the actual measured value, and when a preset change in the statistical distribution of the error stream is detected, incrementally train the first machine learning model and the second machine learning model using newly collected data. It aims to solve the performance degradation problem of heat exchangers caused by factors such as fouling and corrosion during long-term operation, and to cope with the problem of decreased model prediction accuracy caused by changes in working conditions. By monitoring the prediction error stream and performing drift detection, the system can timely detect the degradation of model performance and trigger model updating, thereby maintaining the accuracy and reliability of the prediction. Incremental training adopts a transfer learning strategy to fine-tune the original model, avoiding the computational resource consumption and time cost caused by training the model from scratch.

[0106] Specifically, the monitoring of the error stream can be achieved by calculating the mean square error (MSE), mean absolute error (MAE), or other suitable error indicators. The drift detection algorithm can choose Drift Detection Method (DDM), EarlyDrift Detection Method (EDDM), or Page-Hinkley Test, etc. The preset change can be that the error exceeds a preset threshold, for example, exceeds the historical error mean plus three times the standard deviation. Incremental training can use optimization algorithms such as gradient descent, and set a small learning rate to avoid destroying the knowledge of the original model. As a non-limiting example, the Adam optimizer can be selected, and the learning rate is set to 0.001. The newly collected data can be collected after detecting the drift, or the latest data collected in a sliding window manner.

[0107] Through the above-mentioned model adaptive updating module, the system can respond to changes in the operating state of the heat exchanger in real time, automatically adjust the prediction model, and ensure the long-term effectiveness of the model, thereby improving the reliability and efficiency of the heat exchanger operation and avoiding energy waste and safety hazards caused by model failure.

[0108] In some embodiments, the first machine learning model is a support vector machine model, and the second machine learning model is a random forest model; and the second machine learning model is further configured to calculate the contribution of each parameter in the input feature data to the prediction result.

[0109] In this embodiment, the support vector machine (SVM) model is good at processing nonlinear and high-dimensional data, and can capture the global trend of heat transfer performance prediction; the random forest (RF) model can effectively process local complex relationships and is used to calibrate the residual error of the SVM model to improve the prediction accuracy. The RF model can also perform feature importance analysis to quantify the influence of each input feature (such as Reynolds number, Prandtl number) on the heat transfer performance, providing a basis for optimizing the operating parameters. For example, the feature importance ranking such as [Re: 0.65, Pr: 0.25, spiral tube pitch: 0.10] can be output.

[0110] Alternatively, the first machine learning model can also use a gradient boosting decision tree (GBDT) model or the like. With this scheme, the key factors affecting the heat transfer performance can be effectively identified, and the heat exchanger design and operating parameters can be optimized accordingly, thereby improving the heat transfer efficiency and reducing the energy consumption.

[0111] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application should not be limited to these descriptions. For those of ordinary skill in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the scope of protection of the present application.

Claims

1. A deep learning-based heat exchanger intelligent heat transfer prediction method, characterized in that, The method comprises the following steps: obtaining input feature data of the heat exchanger, the input feature data comprising parameters representing fluid flow state and heat transfer state; processing the input feature data by using a first machine learning model to generate an initial heat transfer performance prediction value; training a second machine learning model based on the input feature data and the initial heat transfer performance prediction value to predict the residual between the initial heat transfer performance prediction value and the corresponding actual heat transfer performance value; when predicting new input feature data, obtaining a new initial heat transfer performance prediction value by using the first machine learning model and a corresponding predicted residual value by using the second machine learning model; superimposing the new initial heat transfer performance prediction value and the predicted residual value to obtain a final heat transfer performance prediction value; generating an operation parameter adjustment scheme of the heat exchanger based on the final heat transfer performance prediction value and flow resistance data; the step of generating the operation parameter adjustment scheme of the heat exchanger specifically comprises: constructing a multi-objective optimization model aiming to maximize the final heat transfer performance prediction value and minimize flow resistance; solving the multi-objective optimization model by using a preset multi-objective optimization algorithm to obtain a set of Pareto optimal solutions representing different trade-off relationships; selecting a target solution from the set of Pareto optimal solutions according to an externally input operation strategy weight, and determining the operation parameter adjustment scheme based on the target solution.

2. The deep learning-based heat exchanger intelligent heat transfer prediction method according to claim 1, characterized in that, The input feature data is obtained by: collecting real-time operation data of the heat exchanger, the real-time operation data including fluid flow rate, temperature and pressure; calculating Reynolds number and Nusselt number based on the real-time operation data by using a numerical simulation method, and taking the Reynolds number and Nusselt number as part of the input feature data.

3. The deep learning-based heat exchanger intelligent heat transfer prediction method according to claim 1 or 2, characterized in that, The first machine learning model is a support vector machine model, and the second machine learning model is a random forest model.

4. The deep learning-based heat exchanger intelligent heat transfer prediction method according to claim 1, characterized in that, After generating the final heat transfer performance prediction value, the following steps are further included: monitoring the error flow between the final heat transfer performance prediction value and the actual measured value; analyzing the statistical distribution of the error flow by using a drift detection algorithm, and triggering a model update instruction when a preset degree of change in the statistical distribution is detected; in response to the model update instruction, incrementally training the first machine learning model and the second machine learning model by using newly collected data.

5. The deep learning-based heat exchanger intelligent heat transfer prediction method according to claim 4, characterized in that, The incremental training is performed by fine-tuning the original model parameters in a manner of transfer learning.

6. The deep learning-based heat exchanger intelligent heat transfer prediction method according to claim 1, characterized in that, Before generating the initial prediction value by using the first machine learning model, the following steps are further included: optimizing hyperparameters of the first machine learning model by using a cross-validation method, the hyperparameters including penalty coefficient and kernel function parameters.

7. The deep learning-based heat exchanger intelligent heat transfer prediction method according to claim 1, characterized in that, The second machine learning model is also used to calculate the contribution of each parameter in the input feature data to the prediction result, and generate a feature importance ranking.

8. A deep learning-based heat exchanger intelligent heat transfer prediction system, characterized in that, A device for implementing the deep learning-based intelligent heat transfer prediction method of a heat exchanger according to any one of claims 1 to 7, comprising: a feature acquisition module configured to obtain input feature data of the heat exchanger, the input feature data comprising parameters representing fluid flow state and heat transfer state; a cascaded prediction module, which includes a first machine learning model configured to generate a new initial heat transfer performance prediction value based on new input feature data, and a second machine learning model configured to predict a residual between the new initial heat transfer performance prediction value and a corresponding real heat transfer performance value based on the input feature data; a result fusion module configured to superimpose the new initial heat transfer performance prediction value and the residual value predicted by the second machine learning model to output a final heat transfer performance prediction value; a decision generation module configured to generate an operation parameter adjustment scheme of the heat exchanger based on the final heat transfer performance prediction value and a flow resistance value, and further configured to: construct a multi-objective optimization model aiming to maximize the final heat transfer performance prediction value and minimize the flow resistance, solve the multi-objective optimization model by using a preset multi-objective optimization algorithm to obtain a set of Pareto optimal solutions representing different trade-off relationships, and select a target solution from the set of Pareto optimal solutions according to an externally input operation strategy weight, and determine the operation parameter adjustment scheme based on the target solution.

9. The system of claim 8, wherein, The system further includes: a model adaptive updating module configured to: monitor an error flow between the final heat transfer performance prediction value and an actual measured value; when detecting that a preset change occurs in a statistical distribution of the error flow, perform incremental training on the first machine learning model and the second machine learning model by using newly collected data.

Images