Heat transfer performance prediction and optimization method of multi-dimensional data weighted neural network

Abstract

The invention discloses a heat transfer performance prediction and optimization method of a multi-dimensional data weighted neural network, and belongs to the field of artificial intelligence and process equipment enhanced heat transfer. The prediction method comprises the following steps: acquiring multi-physical field heat transfer data, preprocessing the data, and dividing the data into a training set and a test set; performing multi-channel independent normalization processing on the preprocessed data; establishing a neural network prediction model; globally optimizing the weight and the bias parameter of the neural network by adopting a genetic algorithm; using the optimized parameters to initialize a neural network to carry out refined training; evaluating the model performance by using the test set and outputting a heat exchange performance prediction result; and carrying out heat exchange characteristic analysis and equipment optimization design based on a prediction result. According to the method, neural network hyper-parameters are optimized through the genetic algorithm, the problem that a traditional neural network is prone to falling into local optimum is solved, high-precision prediction of heat exchange performance parameters is achieved, and reliable technical support is provided for design and optimization of efficient heat exchange equipment.

Description

Technical Field

[0001] This invention relates to the field of machine learning and process equipment heat transfer enhancement technology, specifically to a method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network. Background Technology

[0002] Process equipment employing enhanced heat transfer technology serves as core equipment in energy, chemical, and power industries, and its performance directly impacts system energy efficiency, operating costs, and safety and reliability. In the petrochemical industry, a 5% increase in the heat transfer coefficient of shell-and-tube heat exchangers can reduce system energy consumption by 3-5%, resulting in significant economic benefits. In high-end equipment such as nuclear power plant steam generators and aerospace thermal management systems, accurate prediction of heat transfer performance is crucial for safe system operation and performance optimization.

[0003] Currently, the mainstream methods for predicting heat transfer performance include CFD simulation based on physical mechanisms and empirical formula methods based on experimental data. Although CFD methods can accurately simulate the flow and temperature field distributions, they consume a lot of computational resources, and their accuracy is greatly affected by factors such as the choice of turbulence model. While empirical formula methods are computationally simple, their applicability is limited, and they show significant prediction errors for novel heat transfer structures or complex flow regimes.

[0004] In recent years, neural network technology has shown potential in heat transfer performance prediction, but traditional methods have obvious limitations: First, traditional neural networks are extremely sensitive to the selection of hyperparameters, and different weight initialization strategies can significantly affect the model's convergence speed and prediction accuracy. Second, pure data-driven black-box models lack physical constraints, and the prediction results may deviate from engineering reality, making it difficult to reveal the inherent heat transfer laws. The model's generalization ability strongly depends on the quality and coverage of the training data, and the prediction performance will decrease when facing complex working conditions outside the training set. These models often ignore known physical conservation laws, which can lead to their prediction results being physically unrealistic or unreasonable, bringing additional uncertainties to engineering applications.

[0005] Therefore, developing a novel intelligent prediction method that can avoid local optima, coordinate multi-objective predictions, and incorporate prior physical knowledge has become an urgent need to improve the design level and operational efficiency of heat exchange equipment. Based on this, a multi-dimensional data weighted neural network-based method for heat transfer performance prediction and optimization is proposed. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network, so as to achieve high-precision and high-stability prediction of heat transfer parameters in complex multiphysics fields.

[0007] The technical solution of the present invention is as follows:

[0008] A method for predicting and optimizing the heat transfer performance of a multidimensional data weighted neural network, comprising the following steps:

[0009] S1. Acquire multiphysics heat transfer experimental data, perform data preprocessing, and divide the data into training and test sets;

[0010] S2. Perform multi-channel independent normalization on the pre-processed heat exchange parameter data;

[0011] S3. Establish a neural network prediction model that includes an input layer, a hidden layer, and an output layer;

[0012] S4. Use a genetic algorithm to globally optimize the weights and bias parameters of the neural network;

[0013] S5. Initialize the neural network using the optimized parameters and perform refined training;

[0014] S6. Use the test set to evaluate the model performance and output the heat transfer performance prediction results;

[0015] S7. Based on the prediction results, conduct heat exchange characteristic analysis and equipment optimization design.

[0016] Furthermore, the data preprocessing process in step S1 is as follows:

[0017] S11. Repair missing data through a hierarchical processing mechanism, fill isolated missing points with linear interpolation, and fill consecutive missing points with cubic spline interpolation.

[0018] S12. Set parameter thresholds based on the temperature characteristics of the heat exchange system to automatically filter abnormal data and instantaneous fault points that exceed the physical range.

[0019] S13. Use a band-stop filter to eliminate mechanical vibration noise and use the gradient threshold method to locate the effective physical signal.

[0020] S14. By calculating parameter gradients, velocity, and acceleration characteristics, construct a derived feature system with physical meaning;

[0021] S15. Use the sliding window technique to construct time series samples, convert the data into a supervised learning format, generate input vectors and output vectors, and construct a sample set; divide the sample set into training set and test set according to the proportion.

[0022] Furthermore, the multi-channel independent normalization process in step S2 is as follows:

[0023] To address the differences in the dimensions and numerical ranges of the heat transfer parameters in the multiphysics field, an independent normalization channel is established for each parameter. MinMaxScaler is used to linearly scale each parameter to [-1,1]. During the training phase, the extreme value parameters of each channel are fitted, and the transformation consistency is maintained during the testing phase. The normalized parameter sequence is saved to the Scaler object library to ensure that the data processing standards are consistent during the training and prediction phases.

[0024] Furthermore, the neural network prediction model in step S3 adopts a three-layer feedforward network structure, including: an input layer with 4 nodes to receive heat transfer feature parameters, a hidden layer with 5 neurons using the Sigmoid activation function to achieve nonlinear mapping, and an output layer with 3 nodes to output heat transfer performance parameters. Each layer is connected by a weight matrix and a bias vector to form a complete parameter optimization space.

[0025] Furthermore, the genetic algorithm optimization process in step S4 includes: parameter encoding, fitness evaluation, selection operation, crossover operation, mutation operation, and iterative optimization.

[0026] Furthermore, the refined training in step S5 employs a backpropagation algorithm that includes a learning rate, specifically including:

[0027] Forward propagation calculates the predicted value:

[0028] ;

[0029] ;

[0030] Where Hidden is the hidden layer output matrix, Output is the final output of the network, σ is the Sigmoid activation function, X is the input data matrix, W1 and W2 are weight matrices, and V1 and V2 are bias parameters;

[0031] Calculate the output error:

[0032] ;

[0033] in, For output error, For the actual data in the regression task, This represents the model's predicted value for the sample;

[0034] Backpropagate the error and update the parameters:

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] in, This is the error term for the hidden layer. Output layer weight matrix The transpose of , where η is the learning rate. For the input matrix The transpose of .

[0041] Furthermore, the model evaluation in step S6 adopts a multi-dimensional comprehensive evaluation system that includes regression accuracy indicators, classification performance indicators, and special engineering indicators. The numerical prediction accuracy is evaluated by mean absolute error and coefficient of determination, the working condition classification capability is verified by combining accuracy and F1 score, and the capture effect of key engineering phenomena is evaluated by special indicators such as peak capture accuracy and direction accuracy.

[0042] The formula for calculating the peak capture accuracy is as follows:

[0043] ;

[0044] In the formula, A spike For peak capture accuracy; This represents the total number of peaks in the actual set of peaks. This is the set of time points of actual temperature and pressure abrupt changes that occur in the heat exchange system; I is an indicator function, which takes the value of 1 when the condition is met, and 0 otherwise. For predicted peak features; t pred To predict the time step corresponding to the peak; t actual Δt is the time step corresponding to the actual peak; Δt is the allowable time deviation threshold, which is set according to the sampling frequency and is usually 3 to 10 time steps.

[0045] The formula for calculating the orientation accuracy is as follows:

[0046] ;

[0047] In the formula, A dir I represents the direction accuracy; I represents the indicator function; sign represents the sign function. The predicted value at time step t; t represents the actual value at time step t; T represents the total time step.

[0048] Furthermore, the heat transfer characteristic analysis and equipment optimization design in step S7 includes: calculating the heat transfer coefficient, Nusselt number and comprehensive performance evaluation factor based on the prediction results; identifying and locating the optimization area in the heat transfer process through performance bottlenecks; providing equipment optimization schemes for structural improvement and operation parameter adjustment accordingly; and verifying the optimization effect through comparison to achieve a complete closed loop from performance prediction to optimization design.

[0049] The present invention has the following beneficial effects:

[0050] (1) The prediction method of the present invention optimizes the initial parameters of the neural network globally through a genetic algorithm, which effectively avoids the problem of traditional gradient descent method being prone to getting trapped in local optima, and improves the accuracy and stability of heat transfer performance prediction;

[0051] (2) The prediction method of the present invention adopts multi-channel independent normalization processing, which eliminates the dimensional differences between multiple physical field parameters and enhances the model's ability to characterize complex heat transfer processes.

[0052] (3) The prediction method of the present invention constructs a multi-dimensional evaluation system that includes regression accuracy, classification performance and special engineering indicators, which comprehensively reflects the prediction ability of the model in actual engineering applications;

[0053] (4) The prediction method of the present invention enhances the model’s ability to capture sudden changes and trend changes in key operating conditions through special indicators such as peak capture accuracy and direction accuracy, providing a reliable guarantee for the safe operation of the heat exchange system.

[0054] (5) The prediction method of the present invention calculates the heat transfer coefficient, Nusselt number and comprehensive performance evaluation factor based on the prediction results, identifies and locates the optimization area in the heat exchange process through performance bottlenecks, and provides equipment optimization schemes for structural improvement and operation parameter adjustment accordingly, realizing a complete closed loop from performance prediction to optimization design, and providing scientific basis and technical support for the design and operation optimization of high-efficiency heat exchange equipment. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method for parallel prediction of efficient heat transfer performance based on data weight hyperparameter optimization of neural networks according to an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the genetic algorithm optimization process described in an embodiment of the present invention;

[0057] Figure 3 This is a comparison chart of the test set prediction results and actual values ​​described in the embodiments of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0059] Reference Figure 1 and Figure 2 This embodiment proposes a method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network, specifically including the following steps S1 to S7. This method enables high-precision prediction and optimization of the heat exchange system performance.

[0060] S1. Obtain multiphysics heat transfer experimental data, perform data preprocessing to construct a sample set, and divide the sample set into a training set and a test set.

[0061] The data preprocessing process in step S1 is as follows:

[0062] S11. Missing data is repaired through a hierarchical processing mechanism. Isolated missing points are filled with linear interpolation, and consecutive missing points are filled with cubic spline interpolation. More than 3 consecutive missing points are marked as abnormal segments and not used for training, so as to ensure the continuity and integrity of the data.

[0063] S12. Based on the temperature characteristics of the heat exchange system, set parameter thresholds to automatically filter abnormal data and instantaneous fault points that exceed the physical range.

[0064] S13. Use a band-stop filter to eliminate mechanical vibration noise and use the gradient threshold method to locate the effective physical signal.

[0065] S14. By calculating the parameter gradient, velocity, and acceleration characteristics, construct a physically meaningful derived feature system and perform feature derivation processing on the data.

[0066] S15. Use the sliding window technique to construct time-series samples, convert the data into a supervised learning format, generate input and output vectors, and construct a sample set. Divide the sample set into training and test sets proportionally.

[0067] The input vector is:

[0068] ;

[0069] The output vector is:

[0070] ;

[0071] Among them, X t Let t be the input vector, l be the time step, and r be the prediction step size.

[0072] After data preprocessing, 70% of the data in the sample set is used as the training set, and the remaining 30% is used as the test set, ensuring that the training set covers all flow regimes of laminar, turbulent, and transitional flow.

[0073] S2. Perform multi-channel independent normalization on the pre-processed heat exchange parameter data.

[0074] To address the differences in the dimensions and numerical ranges of the multiphysics heat transfer parameters, an independent normalization channel is established for each parameter. MinMaxScaler is used to linearly scale each parameter to [-1,1], and the extreme value parameters of each channel are fitted during the training phase, while maintaining the consistency of transformation during the testing phase. The normalized parameter sequence is saved to the Scaler object library to ensure that the data processing standards are consistent during the training and prediction phases. Specifically, for temperature, pressure, and flow rate, an independent normalization channel is established for each parameter.

[0075] The normalization formula is:

[0076] ;

[0077] Where x norm x is the normalized variable; x is a sample value in the original data; x min x is the minimum value in the original dataset; max The maximum value in the original dataset;

[0078] During the training phase, the extreme parameter x of each channel is fitted. min and x max During the testing phase, the same extreme value parameters are used for transformation to ensure the uniformity of data processing; the normalized Scaler object is serialized and saved for easy use in subsequent predictions.

[0079] S3. The neural network prediction model adopts a three-layer feedforward network structure, including: an input layer with 4 nodes to receive heat transfer feature parameters, a hidden layer with 5 neurons using the sigmoid activation function to achieve nonlinear mapping, and an output layer with 3 nodes to output heat transfer performance parameters. Each layer is connected by a weight matrix and a bias vector to form a complete parameter optimization space, simplifying the network structure while ensuring the model's expressive power. A neural network prediction model containing an input layer, hidden layers, and an output layer is established.

[0080] The neural network model uses a three-layer feedforward structure:

[0081] Specifically, the input layer has four nodes, each receiving one of the four feature parameters: temperature gradient, flow rate, pressure difference, and heat flux density. The hidden layer has five neurons, using the sigmoid activation function.

[0082] ;

[0083] σ(x) is the output of the Sigmoid function, with a range of (0,1), which can map the input real number x to this interval;

[0084] Output layer: Configured with 3 nodes, which output three performance parameters respectively: heat transfer coefficient, heat exchange efficiency, and pressure drop coefficient;

[0085] Network parameters include: W1∈R 4×5 V1∈R is the weight matrix from the input layer to the hidden layer. 1×5 W2 represents the hidden layer bias vector; W2∈R 5×3 V1 is the weight matrix from the hidden layer to the output layer, and V2 is the bias vector of the output layer.

[0086] S4. Use a genetic algorithm to globally optimize the weights and bias parameters of the neural network.

[0087] The optimization process of a genetic algorithm includes: parameter encoding, fitness evaluation, selection operation, crossover operation, mutation operation, and iterative optimization.

[0088] S5. Initialize the neural network using the optimized parameters and perform refined training.

[0089] The optimal chromosome obtained by the genetic algorithm is decoded into neural network parameters (W1, V1, W2, V2), which are used as the initial parameters of the neural network. A backpropagation algorithm incorporating a learning rate is used for refined training, with a dynamic learning rate, until convergence.

[0090] The training process specifically includes:

[0091] Forward propagation: Calculate the hidden layer and output layer outputs for each training sample.

[0092] ;

[0093] ;

[0094] Where Hidden is the hidden layer output matrix, Output is the final output of the network, σ is the Sigmoid activation function, X is the input data matrix, W1 and W2 are weight matrices, and V1 and V2 are bias parameters;

[0095] Error Calculation: Calculate the output layer error:

[0096] ;

[0097] in, For output error, For the actual data in the regression task, This represents the model's predicted value for the sample;

[0098] Backpropagation: Calculate hidden layer errors and update network parameters.

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] in, This is the error term for the hidden layer. Output layer weight matrix The transpose of , where η is the learning rate. For the input matrix The transpose of the loss function; the performance of the test set is verified every 50 rounds during training, and the iteration ends when the loss function converges and reaches the preset value.

[0105] S6. A multi-dimensional comprehensive evaluation system is adopted, including regression accuracy indicators, classification performance indicators, and special engineering indicators. The accuracy of numerical prediction is evaluated by mean absolute error and coefficient of determination. The classification capability of working conditions is verified by combining accuracy and F1 score. Special indicators such as peak capture accuracy and direction accuracy are used to evaluate the capture effect of key engineering phenomena, thereby providing a comprehensive and reliable evaluation basis for the performance prediction of heat exchange system. The test set is used to evaluate the model performance and output the heat exchange performance prediction results.

[0106] Regression accuracy metrics:

[0107] ;

[0108] ;

[0109] ;

[0110] Where MAE is the mean absolute error; n is the sample size; y i The true value of the i-th sample; The predicted value of the i-th sample;

[0111] Classification performance metrics: These metrics convert continuous outputs into categories by setting a threshold. The calculation is as follows:

[0112] Accuracy: The proportion of samples that are correctly classified;

[0113] Precision: The proportion of actual positive classes predicted in positive class predictions;

[0114] Recall: The proportion of positive classes that are correctly predicted.

[0115] F1 score: Harmonic mean of precision and recall;

[0116] Special engineering specifications:

[0117] Peak capture accuracy A spike :

[0118] ;

[0119] In the formula, A spike For peak capture accuracy; This represents the total number of peaks in the actual set of peaks. This is the set of time points of actual temperature and pressure abrupt changes that occur in the heat exchange system; I is an indicator function, which takes the value of 1 when the condition is met, and 0 otherwise. For predicted peak features; t pred To predict the time step corresponding to the peak; t actual Δt is the time step corresponding to the actual peak; Δt is the allowable time deviation threshold, which is set according to the sampling frequency and is usually 3 to 10 time steps.

[0120] Directional accuracy A dir :

[0121] ;

[0122] In the formula, A dir I represents the direction accuracy; I represents the indicator function; sign represents the sign function. The predicted value at time step t; t represents the actual value at time step t; T represents the total time step.

[0123] S7. Heat transfer characteristic analysis includes: calculating the heat transfer coefficient, Nusselt number, and comprehensive performance evaluation factor based on the prediction results; identifying and locating optimization areas in the heat transfer process through performance bottlenecks; and providing equipment optimization solutions, including structural improvements and adjustments to operating parameters. Finally, the optimization effect is verified through comparison, achieving a complete closed loop from performance prediction to optimization design, thereby improving the design efficiency and performance of the heat exchange system. Heat transfer characteristic analysis and equipment optimization design are performed based on the prediction results.

[0124] The formulas for calculating the heat transfer coefficient, Nusselt number, and comprehensive performance evaluation factor are as follows:

[0125] ;

[0126] ;

[0127] ;

[0128] Performance bottleneck identification: By analyzing the heat transfer coefficient distribution, flow velocity field, and temperature field, the optimization areas in the heat exchange process are located, including high resistance areas, low heat transfer coefficient areas, and flow dead zones.

[0129] Optimization plan formulation: Based on the bottleneck analysis results, a comprehensive plan is formulated, covering fin shape and flow channel structure optimization, selection of high thermal conductivity and corrosion resistant materials, and adjustment of flow rate and temperature setpoints.

[0130] Effect verification: By comparing the performance prediction results before and after optimization, the effectiveness of the optimization scheme is verified, and a complete closed loop from performance prediction to optimization design is achieved.

[0131] To verify the effectiveness of the method of this invention, experimental verification was conducted on a typical shell-and-tube heat exchanger dataset. The dataset contains 1000 sets of multiphysics heat transfer experimental data. The input features include inlet temperature, flow rate, pressure difference, and heat flux density, and the output targets include heat transfer coefficient, heat transfer efficiency, and pressure drop coefficient.

[0132] Reference Figure 3 Experimental results show that the prediction performance of the method of the present invention on the test set is as follows:

[0133] Experimental results show that the method of the present invention achieves a mean absolute error of 0.023, a mean squared error of 0.0012, and a coefficient of determination of 0.945 in regression metrics on the test set; for classification metrics, the accuracy is 0.932, the precision is 0.918, the recall is 0.926, and the F1 score is 0.922; and for special engineering metrics, the peak capture accuracy reaches 0.947, and the direction accuracy reaches 0.961.

[0134] Compared with traditional methods such as BP neural networks, support vector machines, and random forests, the method of this invention has improvements in prediction accuracy, stability, and engineering applicability.

[0135] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network, characterized in that, Including the following steps: S1. Acquire multiphysics heat transfer experimental data, perform data preprocessing, and divide the data into training and test sets; S2. Perform multi-channel independent normalization on the pre-processed heat exchange parameter data; S3. Establish a neural network prediction model that includes an input layer, a hidden layer, and an output layer; S4. Use a genetic algorithm to globally optimize the weights and bias parameters of the neural network; S5. Initialize the neural network using the optimized parameters and perform refined training; S6. Use the test set to evaluate the model performance and output the heat transfer performance prediction results; S7. Based on the prediction results, conduct heat exchange characteristic analysis and equipment optimization design.

2. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The data preprocessing process in step S1 is as follows: S11. Repair missing data through a hierarchical processing mechanism, fill isolated missing points with linear interpolation, and fill consecutive missing points with cubic spline interpolation. S12. Set parameter thresholds based on the temperature characteristics of the heat exchange system to automatically filter abnormal data and instantaneous fault points that exceed the physical range. S13. Use a band-stop filter to eliminate mechanical vibration noise and use the gradient threshold method to locate the effective physical signal. S14. By calculating parameter gradients, velocity, and acceleration characteristics, construct a derived feature system with physical meaning; S15. Use the sliding window technique to construct time series samples, convert the data into a supervised learning format, generate input vectors and output vectors, and construct a sample set. The sample set is divided into a training set and a test set according to the proportion.

3. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The multi-channel independent normalization process in step S2 is as follows: To address the differences in the dimensions and numerical ranges of the heat transfer parameters in the multiphysics field, an independent normalization channel was established for each parameter. MinMaxScaler was used to linearly scale each parameter to [-1,1]. During the training phase, the extreme value parameters of each channel were fitted, and the transformation consistency was maintained during the testing phase. The normalized parameter sequence is saved to the Scaler object library to ensure that the data processing standards are consistent during the training and prediction phases.

4. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The neural network prediction model in step S3 adopts a three-layer feedforward network structure, including: an input layer with 4 nodes to receive heat transfer feature parameters, a hidden layer with 5 neurons using the Sigmoid activation function to achieve nonlinear mapping, and an output layer with 3 nodes to output heat transfer performance parameters. Each layer is connected by a weight matrix and a bias vector to form a complete parameter optimization space.

5. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The genetic algorithm optimization process in step S4 includes: parameter encoding, fitness evaluation, selection operation, crossover operation, mutation operation, and iterative optimization.

6. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The refined training in step S5 employs a backpropagation algorithm that includes a learning rate, specifically including: Forward propagation calculates the predicted value: ； ； Where Hidden is the hidden layer output matrix, Output is the final output of the network, σ is the Sigmoid activation function, X is the input data matrix, W1 and W2 are weight matrices, and V1 and V2 are bias parameters; Calculate the output error: ； in, For output error, For the actual data in the regression task, This represents the model's predicted value for the sample; Backpropagate the error and update the parameters: ； ； ； ； ； in, This is the error term for the hidden layer. Output layer weight matrix The transpose of , where η is the learning rate. For the input matrix The transpose of .

7. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The model evaluation in step S6 adopts a multi-dimensional comprehensive evaluation system that includes regression accuracy indicators, classification performance indicators and special engineering indicators. The numerical prediction accuracy is evaluated by mean absolute error and coefficient of determination, the working condition classification capability is verified by combining accuracy and F1 score, and the peak capture accuracy and direction accuracy special indicators are used to evaluate the capture effect of key engineering phenomena. The formula for calculating the peak capture accuracy is as follows: ； In the formula, A spike For peak capture accuracy; This represents the total number of peaks in the actual set of peaks. This is the set of time points of actual temperature and pressure abrupt changes that occur in the heat exchange system; I is an indicator function, which takes the value of 1 when the condition is met, and 0 otherwise. For the predicted peak features; t pred To predict the time step corresponding to the peak; t actual Δt is the time step corresponding to the actual peak; Δt is the allowable time deviation threshold, which is set according to the sampling frequency and is usually 3 to 10 time steps. The formula for calculating the orientation accuracy is as follows: ； In the formula, A dir I represents the direction accuracy; I represents the indicator function; sign represents the sign function. The predicted value at time step t; t represents the actual value at time step t; T represents the total time step.

8. The method for predicting and optimizing heat transfer performance using a multidimensional data weighted neural network according to claim 1, characterized in that, The heat transfer characteristic analysis and equipment optimization design in step S7 include: calculating the heat transfer coefficient, Nusselt number and comprehensive performance evaluation factor based on the prediction results; identifying and locating the optimization area in the heat transfer process through performance bottlenecks; providing equipment optimization schemes for structural improvement and operation parameter adjustment accordingly; and verifying the optimization effect through comparison to achieve a complete closed loop from performance prediction to optimization design.

Images