Method and system for predicting heat exchange coefficient of heat exchanger based on physical information neural network

Abstract

The invention belongs to the field of industrial thermal engineering and intelligent modeling, and discloses a heat exchanger heat exchange coefficient prediction method and system based on a physical information neural network. The method comprises the following steps: acquiring multi-dimensional operation data through a signal acquisition system, cleaning abnormal and blank values, standardizing, and segmenting into time sequence samples by adopting a sliding window method; a double-layer physical information long-short-term memory network is constructed, and a time sequence feature and a physical equation residual error are combined to generate a space-time fusion feature matrix. And a composite loss function including data loss, physical equation loss and physical consistency loss is designed, physical and data driving influences are balanced through hyper-parameter tuning, and accurate prediction of the heat exchange coefficient is achieved based on a gradient descent optimization model. The method combines field physical laws and data features, improves the reliability and physical interpretability of prediction, and is suitable for operation optimization of the heat exchanger of the desulfurization wastewater treatment system of the thermal power plant.

Description

Technical Field

[0001] The present invention belongs to the field of heat transfer coefficient prediction of heat exchangers in desulfurization wastewater treatment systems of thermal power plants, and particularly relates to a heat transfer coefficient prediction method of heat exchangers based on physical information neural networks. Background Art

[0002] As core equipment in the fields of energy, chemical engineering, refrigeration, etc., the heat transfer coefficient of heat exchangers directly affects the energy efficiency and operational safety of the system. Traditional heat transfer coefficient prediction methods rely on mechanism modeling, analyzing performance degradation by establishing heat transfer and fluid mechanics equations. However, parameter identification is difficult under complex operating conditions and the model generalization ability is insufficient. Data-driven methods such as long-short-term memory networks can mine data features, but they ignore the physical laws of heat exchanger operation, resulting in a disconnect between prediction results and actual physical constraints. When data noise is high or operating conditions change suddenly, the prediction accuracy drops significantly. Therefore, how to combine time series analysis with physical prior knowledge has become the key to improving the reliability of heat exchanger health prediction.

[0003] Physically-Informed Neural Networks (PINNs) fuse data-driven models with mechanistic models by introducing physical equation constraints into the loss function. In recent years, they have demonstrated their strengths in fluid dynamics and heat transfer process modeling. However, given the time series nature of heat exchanger operating data, traditional PINNs struggle to effectively capture long-term dependencies and fail to fully consider physical consistency across time and space (such as the coupled relationship between temperature and pressure changes at different locations). While long-short-term memory networks (LSTMs) excel at handling long-term dependencies in time series, their lack of physical constraints can lead to predictions that violate fundamental heat transfer laws. Therefore, there is an urgent need to propose a method that combines the time series modeling capabilities of LSTMs with the constraints of physical equations, namely, the Physically-Informed Long-Short-Term Memory Network (PILSTM), which balances the flexibility of data-driven models with the rigor of physical laws. Summary of the Invention

[0004] In response to the above content, the present invention provides a heat exchanger heat transfer coefficient prediction method based on physical information neural network, which screens variables of heat exchanger sensor data and extracts feature information, thereby improving the reliability and physical interpretability of heat transfer coefficient prediction and having practical engineering value.

[0005] To achieve the above objectives, the present invention provides the following technical solutions:

[0006] A method for predicting heat transfer coefficient of a heat exchanger based on a physical information neural network comprises the following steps:

[0007] Step 1: Use the signal acquisition system to obtain multi-dimensional operating information of the heat exchanger, including characteristic values ​​and target true values, and pre-process the raw data, including data cleaning and data standardization;

[0008] Step 2: Use the sliding window method to split the standardized feature data into time series samples, generate the time series feature matrix and target vector; construct a physical information long short-term memory network to capture the long-term dependencies in the time series and integrate physical laws into the model; convert the time series features into input tensors, calculate the derivatives of the physical equations, and perform feature splicing and physical model input;

[0009] Step 3: Introduce the physical information loss function, which includes the physical equation loss function and the physical consistency loss function. The weight factor is adjusted to balance the influence of the physical equation residual and the prediction trend consistency.

[0010] Step 4: Define the prediction of the heat exchanger heat transfer coefficient as a regression prediction problem; introduce a composite loss function that comprehensively considers data loss and physical information, allowing PILSTM to adapt to actual observation data while following physical laws.

[0011] A further improvement of the technical solution of the present invention is that step 1 includes the following steps:

[0012] Step 11: Use the signal acquisition system to obtain multi-dimensional operating information of the heat exchanger, including characteristic values ​​and target true values, and clean the data to remove outliers and blank values.

[0013] Step 12: Define the feature variable matrix X and use the normalizer to normalize the feature variables and target variables. The specific algorithm is as follows:

[0014]

[0015] Among them, X is the characteristic variable matrix, μ X and σ X are the mean and standard deviation of the characteristic variables, respectively.

[0016] A further improvement of the technical solution of the present invention is that step 2 includes the following steps:

[0017] Step 21: Use the sliding window method to divide the standardized feature data into time series matrix samples of length L;

[0018] Step 22: Construct a physical information long short-term memory network, consisting of a two-layer LSTM unit, to capture the long-term dependencies in the time series samples and output the time series feature matrix and target vector;

[0019] Step 23: Calculate the partial derivatives of the time series characteristic matrix with respect to time and the partial derivatives with respect to the spatial coordinates, and calculate the residuals of the physical equations;

[0020] Step 24: Concatenate the time series feature matrix and the physical equation residual along the channel dimension to form a spatiotemporal fusion feature matrix.

[0021] A further improvement of the technical solution of the present invention is that step 3 includes the following steps:

[0022] Step 31: Introduce the physical information loss function, which includes the physical equation loss function and the physical consistency loss function. The specific algorithm is as follows:

[0023] Loss consistency =αLoss PDE +βLoss physics

[0024] Among them, α and β are weight coefficients, which are used to adjust the proportion of the physical equation loss function and the physical consistency loss function in the physical information loss function.

[0025] Step 32: The physical equation loss function constrains the physical equation residual to approach zero. The specific algorithm is as follows:

[0026]

[0027] Where N is the number of samples, f i is the physical equation residual of the time series feature matrix, and the target value is the zero tensor.

[0028] Step 33: The physical consistency loss function mentioned above enhances the consistency of the change trend between the predicted value and the true value through the ReLU function. The specific algorithm is as follows:

[0029]

[0030] Among them, L is the length of the time series matrix sample, y i and The predicted value of the current time step and the next time step is ReLU. The function is to perform nonlinear transformation on the input data. The mathematical expression is ReLU(x)=max(0,x).

[0031] A further improvement of the technical solution of the present invention is that step 4 includes the following steps:

[0032] Step 41: Define the heat exchanger heat transfer coefficient prediction as a regression prediction problem;

[0033] Step 42: Introduce a composite loss function that comprehensively considers data loss and physical information. The specific algorithm is as follows:

[0034] Loss total =Loss data +Loss consistency

[0035] Step 43: The data loss function, i.e., the mean square error between the predicted value and the true value, has the following specific algorithm:

[0036]

[0037] in, The predicted value for the current time step.

[0038] Step 44: Input the spatiotemporal fusion feature space matrix obtained in step 24 into the designed heat exchanger heat transfer coefficient prediction module;

[0039] Step 45: Composite loss Loss based on each batch of samples total The back propagation algorithm is performed, and the model parameters are updated and adjusted according to the gradient descent algorithm. By repeating this process, the heat transfer coefficient of the heat exchanger is predicted.

[0040] Compared with the prior art, the heat exchanger heat transfer coefficient prediction method based on physical information neural network provided by the present invention has the following beneficial effects:

[0041] The present invention proposes a method for predicting the heat transfer coefficient of heat exchangers based on a physical information neural network. This method addresses the pain points of existing data-driven models, which lack physical constraints and rely on strong assumptions about mechanism models. By constructing time series samples through a sliding window, the method utilizes PILSTM to capture long-term dependencies and embed physical laws such as heat transfer and flow. Furthermore, a composite loss function is designed that incorporates prediction error, physical equation residuals, and consistency of changing trends. This method not only preserves the adaptability of data-driven models to complex operating conditions but also improves the reliability of prediction results through physical constraints. This method provides more precise technical support for real-time health monitoring and preventive maintenance of desulfurization wastewater heat exchangers, and has important engineering significance for improving the energy efficiency of heat exchangers in desulfurization wastewater treatment systems in thermal power plants and reducing operation and maintenance costs.

[0042] 1. Fusion of data-driven and physical constraints: The Physical Information Long Short-Term Memory (PILSTM) network integrates LSTM time series modeling capabilities with physical laws such as heat transfer and flow in heat exchangers. A sliding window is used to capture long-term dependencies. Combined with residual constraints of physical equations, this significantly improves the accuracy and robustness of health predictions under complex operating conditions.

[0043] 2. Composite loss function optimization: Design a composite loss function that includes data fitting error, physical equation residuals, and trend consistency. The mean square error ensures that the predicted value is close to the real data. The physical equation loss forces the basic physical constraints to be met. The ReLU function strengthens the consistency of the change trend and dynamically balances the weight factors, effectively improving the model's generalization ability and reliability.

[0044] 3. Deep mining of spatiotemporal features: This system segments time series samples based on a sliding window, combines a two-layer LSTM to calculate spatiotemporal derivatives, and constructs a spatiotemporal fusion matrix through feature concatenation. This system supports online model updates and real-time learning, accurately capturing heat exchanger performance degradation characteristics and providing an efficient, real-time prediction tool for preventive maintenance of equipment, thereby reducing operation and maintenance costs and downtime risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A flow chart of a method for predicting heat transfer coefficient of a heat exchanger based on a physical information neural network provided for the implementation of the present invention;

[0047] Figure 2 Provided for the implementation of the present invention Figure 1 Signal acquisition flow chart;

[0048] Figure 3 A diagram of a neural network structure based on physical information provided for the implementation of the present invention;

[0049] Figure 4 The overall architecture diagram of the physical information neural network provided for the implementation of the present invention;

[0050] Figure 5 A diagram showing the heat transfer coefficient prediction results provided for the implementation of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions of the present invention will be clearly and completely described below through specific implementation methods. It is obvious that the described embodiments are only some of the embodiments of the present invention, and not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0052] The present invention discloses a heat exchanger heat transfer coefficient prediction method based on physical information neural network. The multidimensional operating data of the heat exchanger is obtained through a signal acquisition system. After data cleaning to remove outliers and blank values, the characteristic variables and target variables are normalized by a standardization method to construct a data set. Secondly, the standardized characteristic data are divided into time series samples by a sliding window method to generate a feature matrix and a target vector. By constructing a two-layer physical information long-short-term memory network, the long-term dependency of the time series is captured, and the spatiotemporal partial derivatives of the physical equation are calculated to generate the physical equation residual. The time series features and the physical residuals are spliced ​​to form a spatiotemporal fusion feature matrix, thereby realizing the deep fusion of physical laws and data-driven models. A composite loss function including data loss, physical equation loss and physical consistency loss is designed. The weight factor is determined through hyperparameter tuning to balance the influence of data drive and physical constraints. The model parameters are optimized based on the gradient descent algorithm to realize the accurate prediction of the heat exchanger heat transfer coefficient. The present invention effectively combines domain knowledge and data characteristics, improves the reliability and physical interpretability of heat transfer coefficient prediction, and is suitable for the heat transfer coefficient prediction of heat exchangers in desulfurization wastewater treatment systems of thermal power plants.

[0053] The present invention first uses a distributed signal acquisition system to collect multidimensional key parameters of the heat exchanger during operation in real time, including characteristic information such as instantaneous flow rate, inlet and outlet temperatures, and pressure differential. The collected raw data then undergoes quality control. Robust anomaly detection methods are used to remove outliers, while linear interpolation and forward filling strategies are applied to repair missing measurements, ensuring the continuity and integrity of the input data along the timeline. This provides a highly reliable basic data source for subsequent modeling.

[0054] The cleaned dataset needs to be dimensionlessly normalized to unify the feature scale. Specifically, for each feature variable, its mean and standard deviation are calculated across the entire sample. The mean is then subtracted from the feature value and divided by the standard deviation to complete the normalization. The same normalization operation is performed on the target variable. This approach not only eliminates the imbalance caused by different dimensions but also effectively avoids the problems of vanishing or exploding gradients in deep learning training, improving the stability and convergence speed of neural network training.

[0055] The standardized data is segmented using a fixed-length sliding window to construct a sample set with time-dependent characteristics. Specifically, continuously sampled feature data is packaged into time series segments according to a set number of time steps, which serve as the model's input features. The target variable at the corresponding time is simultaneously extracted as the label output. By properly setting the window length and sliding step size, the temporal continuity of the sequence features is ensured while increasing the diversity and coverage of the samples, meeting the input requirements of long-short-term memory network modeling.

[0056] After obtaining the time series features, the first-order partial derivatives of the feature matrix are calculated along both the time and spatial axes to extract information about the temporal rate of change and spatial gradient. Based on the principle of energy conservation in heat exchanger thermodynamics, the residual value of the physical equation is defined and calculated as a key indicator of whether the model output conforms to physical laws. Subsequently, the time series features and the physical equation residual are concatenated along the channel direction to form a high-dimensional input matrix that combines data-driven features with physical consistency features for subsequent neural network learning.

[0057] To achieve the dual goals of modeling data fit and adhering to physical laws, this paper designs a composite loss function consisting of three parts. The first part is the data loss, which measures the mean squared error between the model's predictions and the true observations. The second part is the physical equation residual loss, which minimizes the physical equation residuals and improves physical consistency. The third part is the physical trend consistency loss, which uses a nonlinear transformation function to strengthen the consistency of the predicted results with the true trend change direction. These losses are fused through weighted hyperparameters to dynamically adjust the balance between data fitting and physical constraints.

[0058] The overall model uses mini-batch stochastic gradient descent for parameter optimization. During each training round, a compound loss is calculated based on batch samples, and the parameter weights of each layer of the neural network are updated via backpropagation. To accelerate convergence and improve generalization, the learning rate is dynamically adjusted during training, and a regularization strategy is introduced to prevent overfitting. After training, the model can accurately predict the heat transfer coefficient at a specific moment based on real-time data collected from the heat exchanger's operating characteristics, providing a scientific basis for heat exchanger status assessment and system optimization control.

[0059] like Figure 1 As shown, a method for predicting the heat transfer coefficient of a heat exchanger based on a physical information neural network is provided, comprising the following steps:

[0060] Step 1: If Figure 2 As shown in the signal acquisition system, the signal acquisition system is used to obtain multi-dimensional operating information of the heat exchanger, including characteristic values ​​and target true values, and pre-process the raw data, including data cleaning and data standardization. The specific operation steps are as follows:

[0061] Step 11: Use the signal acquisition system to obtain multi-dimensional operating information of the heat exchanger, including characteristic values ​​and target true values, and clean the data to remove outliers and blank values.

[0062] Step 12: Define the feature variable matrix X and use the normalizer to normalize the feature variables and target variables. The specific algorithm is as follows:

[0063]

[0064] Among them, X is the characteristic variable matrix, μ X and σ X are the mean and standard deviation of the characteristic variables, respectively.

[0065] Step 2: If Figure 3 The neural network structure based on physical information is shown in the figure, where p is the pressure measured by the pressure transmitter, q m is the mass flow rate measured by the flow transmitter, Δp is the pressure difference measured by the differential pressure transmitter, h is the material position measured by the level sensor, and σ is the activation function used to perform nonlinear transformation on the input. The standardized feature data is divided into time series samples using the sliding window method to generate the time series feature matrix and target vector. A physical information long short-term memory network is constructed to capture long-term dependencies in the time series and incorporate physical laws into the model. The time series features are converted into input tensors, the derivatives of the physical equations are calculated, and feature splicing and physical model input are performed. The specific steps are as follows:

[0066] Step 21: Use the sliding window method to divide the standardized feature data into time series matrix samples of length L;

[0067] Step 22: Construct a physical information long short-term memory network, consisting of a two-layer LSTM unit, to capture the long-term dependencies in the time series samples and output the time series feature matrix and target vector;

[0068] Step 23: Calculate the partial derivative of the time series feature matrix with respect to time and the partial derivatives with respect to the spatial coordinates And calculate the residual f of the physical equation i ;

[0069] Step 24: Concatenate the time series feature matrix and the physical equation residual along the channel dimension to form a spatiotemporal fusion feature matrix.

[0070] Step 3: If Figure 4 As shown in the overall architecture diagram of the physical information neural network, a physical information loss function is introduced, which includes the physical equation loss function and the physical consistency loss function. The influence of the physical equation residual and the prediction trend consistency is balanced by adjusting the weight factor.

[0071] Step 31: Introduce the physical information loss function, which includes the physical equation loss function and the physical consistency loss function. The specific algorithm is as follows:

[0072] Loss consistency =αLoss PDE +βLoss physics

[0073] Among them, α and β are weight coefficients, which are used to adjust the proportion of the physical equation loss function and the physical consistency loss function in the physical information loss function.

[0074] Step 32: The physical equation loss function constrains the physical equation residual to approach zero. The specific algorithm is as follows:

[0075]

[0076] Where N is the number of samples, f i is the physical equation residual of the time series feature matrix, and the target value is the zero tensor.

[0077] Step 33: The physical consistency loss function mentioned above enhances the consistency of the change trend between the predicted value and the true value through the ReLU function. The specific algorithm is as follows:

[0078]

[0079] Among them, L is the length of the time series matrix sample, y i and The predicted value of the current time step and the next time step is ReLU. The function is to perform nonlinear transformation on the input data. The mathematical expression is ReLU(x)=max(0,x).

[0080] Step 4: Define the prediction of the heat exchanger heat transfer coefficient as a regression prediction problem; introduce a composite loss function that comprehensively considers data loss and physical information, allowing PILSTM to adapt to actual observation data while following physical laws.

[0081] Step 41: Define the heat exchanger heat transfer coefficient prediction as a regression prediction problem;

[0082] Step 42: Introduce a composite loss function that comprehensively considers data loss and physical information. The specific algorithm is as follows:

[0083] Loss total =Loss data +Loss consistency

[0084] Step 43: The data loss function, i.e., the mean square error between the predicted value and the true value, has the following specific algorithm:

[0085]

[0086] in, The predicted value for the current time step.

[0087] Step 44: Input the spatiotemporal fusion feature space matrix obtained in step 24 into the designed heat exchanger heat transfer coefficient prediction module;

[0088] Step 45: Composite loss Loss based on each batch of samples total The back propagation algorithm is performed, and the model parameters are updated and adjusted according to the gradient descent algorithm. By repeating this process, the heat transfer coefficient of the heat exchanger is predicted.

[0089] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A heat exchanger heat transfer coefficient prediction method based on physical information neural network, characterized in that: The steps include: Collect multi-dimensional signal data during the operation of the heat exchanger, and perform outlier removal and missing value filling processing; Perform standardized preprocessing on the collected feature variables and target variables; Based on the sliding window method, the standardized data is divided into time series samples of fixed length; Construct a physical information neural network containing long short-term memory units and input the time series feature matrix; Calculate the partial derivatives of time series features with respect to time and space dimensions to generate the residuals of the physical equations; The time series feature matrix and the physical equation residual are concatenated along the channel dimension to form a fusion feature matrix; Define a composite loss function that includes data loss, physical equation residual loss, and physical consistency loss; Parameter optimization training is performed based on the composite loss function, and the heat transfer coefficient prediction results of the heat exchanger are output.

2. The method according to claim 1, characterized in that The data cleaning includes: Outlier data samples are eliminated through robust statistical methods, and missing data are filled using linear interpolation and forward filling.

3. The method according to claim 1, characterized in that The standardization process includes: Subtract the corresponding mean from each characteristic variable and divide it by the corresponding standard deviation; Subtract the target variable's mean and divide by its standard deviation.

4. The method according to claim 1, wherein The physical information neural network includes: A two-layer stacked long short-term memory unit, where each layer includes a gating mechanism to extract long-term dependency characteristics of time series.

5. The method according to claim 1, wherein The calculation of the physical equation residual includes: The first-order partial derivatives of the time series characteristic matrix in the time axis and space axis directions are calculated respectively, and the residual information is derived based on the thermodynamic conservation equation of the heat exchanger.

6. The method according to claim 1, characterized in that The composite loss function includes: The data loss term is defined based on the mean square error between the predicted value and the true value. The physical consistency loss term defined based on the residual sum of squares of the physical equations, And the physical consistency trend loss term defined based on the difference between the predicted value change trend and the true value change trend.

7. A heat exchanger heat transfer coefficient prediction system based on physical information neural network, characterized in that: include: Signal acquisition module, used to collect multi-dimensional operating data during the operation of the heat exchanger; A data preprocessing module, communicating with the signal acquisition module, performs data cleaning and standardization processing; A sample generation module is connected to the data preprocessing module and uses a sliding window method to generate a time series feature matrix and a target vector; A physical information neural network module, which is in communication with the sample generation module and includes a double-layer long short-term memory unit and a physical residual calculation unit; A feature fusion module is connected to the physical information neural network module to combine the time series features with the physical equation residuals along the channel dimension; A loss calculation module, connected to the feature fusion module, defines a composite loss function including data loss, physical equation residual loss and physical consistency loss; A parameter optimization module, connected to the loss calculation module, optimizes the neural network model weights based on a composite loss function; The prediction output module is connected to the parameter optimization module and outputs the prediction result of the heat transfer coefficient of the heat exchanger.

8. The system according to claim 7, characterized in that The signal acquisition module includes: Flow sensors, temperature sensors, and pressure sensors are installed at the inlet and outlet of the heat exchanger and the piping system, respectively, and are configured to synchronously collect time series data.

9. The system according to claim 7, wherein: The physical information neural network module includes: Double-layer stacked long short-term memory unit, each layer has a forget gate, input gate and output gate, The physical residual calculation unit is configured to calculate the first-order partial derivatives in the time axis and space axis directions according to the time series characteristics, and derive the physical equation residual based on the thermodynamic conservation relationship of the heat exchanger.

10. The system according to claim 7, wherein: The loss calculation module includes: The data error submodule defines the loss based on the mean square error between the predicted value and the true value; The physical residual submodule defines the loss based on the residual sum of squares of the physical equations; The physical consistency submodule defines the trend consistency loss by performing a nonlinear transformation on the difference between the predicted value change rate and the true value change rate; The loss values ​​output by each submodule are weighted and summed to construct a composite loss function.

Images