NOx emission prediction method of gas-steam combined cycle generator set

Abstract

The invention discloses a NOx emission prediction method for a gas-steam combined cycle generator set, and belongs to the technical field of gas-steam combined cycle power generation, and the method comprises the steps: data collection and data preprocessing, and construction of a source-target domain input pair; the method comprises the following steps: establishing a CATAs-DSN model, and training the CATAs-DSN model; nOx emission prediction and model evaluation are carried out. Learning domain common and private time features of load data of a source domain and a target domain by using an attention mechanism through a feature extractor based on a CATans layer, and carrying out knowledge migration by using the common features; the feature reconstructor is used as an auxiliary module to reconstruct data of a source domain and a target domain; and converting the learned features into predicted values by a regression predictor. Finally, the NOx emission prediction model obtained through training on the source domain and the target domain is directly used for NOx emission prediction, NOx emission prediction is achieved, and low-carbon and environment-friendly operation of the unit is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas-steam combined cycle power generation, and in particular, relates to a method for predicting NOx emissions of a gas-steam combined cycle power generation unit. Background Art

[0002] With the transformation of energy structure, higher requirements are placed on the environmental performance of power generation equipment. Gas-steam combined cycle power generation units fueled by clean energy natural gas have become the primary technology for the development of clean energy power generation. However, such units will still emit a certain amount of NOx during operation, posing a potential threat to the environment and human health. Therefore, accurate prediction and optimization of NOx emissions are crucial to the formulation of effective environmental policies and clean production strategies.

[0003] A prediction model based on sparrow search algorithm (SSA) to optimize temporal pattern attention mechanism long short-term memory neural network (TPA-LSTM) is proposed in the existing literature ("Jin Xiuzhang, Zhang Jin, Chen Jiazheng. Prediction of NOx emissions from thermal power plants based on sparrow search algorithm to optimize TPA-LSTM [J / OL]. Control Engineering, 1-9 [2024-08-12]. https: / / doi.org / 10.14107 / j.cnki.kzgc.20230640"). The variables related to SCR inlet NOx are selected through mechanism analysis, and the auxiliary variables are screened by the least absolute shrinkage and selection operator (LASSO) algorithm. The maximum information coefficient (MIC) is used to calculate the delay time, and the hyperparameters of the TPA-LSTM model are optimized by SSA. Finally, a comparative experiment is conducted.

[0004] The existing literature first screens out the features related to NOx emission indicators based on the LASSO algorithm and mechanism analysis, calculates the delay time by the maximum information coefficient, and finally uses SSA to optimize the hyperparameters of the TPA-LSTM model to obtain the NOx emission prediction model. This method cannot solve the problem of limited training data sets, and has weak generalization ability, and the prediction effect of NOx emissions under different working conditions is poor. Traditional transfer learning uses the data and models of existing related tasks to assist in the prediction of new tasks and improve model performance. It is especially suitable for new tasks with limited data and across working conditions. In the study of deep transfer learning, traditional algorithms mainly focus on how to eliminate the differences in feature distribution between different domains to achieve effective knowledge transfer. However, each domain contains common features and private features. If private features are also included in the migration process, negative migration may occur, thereby affecting the generalization ability of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a NOx emission prediction method for a gas-steam combined cycle power generation unit based on CATrans-DSN, which aims to solve the measurement problem of NOx emissions, an environmental indicator, during the operation of the gas-steam combined cycle power generation unit under different operating conditions, realize NOx emission prediction, and ensure low-carbon and environmentally friendly operation of the unit.

[0006] To achieve the above object, the present invention adopts the following technical solution: A method for predicting NOx emissions of a gas-steam combined cycle generator set comprises the following steps:

[0007] Step 1: Data collection and data preprocessing, constructing source-target input pairs formed by source domain data and target domain data;

[0008] Step 2: Establish the CATrans-DSN model and train it;

[0009] The CATrans-DSN model includes a feature extractor, a regression predictor and a feature reconstructor, wherein the feature extractor is composed of at least one CATrans layer, and the regression predictor and the feature reconstructor are both composed of at least one fully connected layer;

[0010] The feature extractor is a feature extractor based on the CATrans layer, which is used to extract the private features of the input source domain and target domain data through the self-attention mechanism, and to identify and separate the common features in the source domain and target domain data through the cross-attention mechanism;

[0011] The regression predictor is used to find the mapping relationship between the extracted temporal features and NOx emissions based on the source domain and target domain data, and convert the extracted features into accurate NOx emission prediction values ​​through regression loss training;

[0012] The feature reconstructor is used to reconstruct the samples composed of private features and common features to obtain the reconstructed source domain and target domain;

[0013] Step 3: NOx emission prediction and model evaluation

[0014] The target operating condition test data is input into the trained CATrans-DSN model to predict NOx emissions, and the final predicted value is output. The prediction performance of the model is evaluated by comparing it with the actual NOx emission data.

[0015] Furthermore, in step 1, the missing data of the source and target conditions are first supplemented by linear interpolation and random forest; then, the data is normalized and the sliding window method is used to process the original time series data.

[0016] Furthermore, the feature extractor includes three weight-sharing branches, namely a source branch, a source-target branch, and a target branch. The source branch includes a self-attention mechanism and a feedforward network, the source-target branch includes a cross-attention mechanism and a feedforward network, and the target branch includes a self-attention mechanism and a feedforward network. The source domain data and the target domain data are input into the feature extractor. The self-attention mechanism is responsible for learning the representation of a specific domain and outputting the private features of the source domain and the target domain respectively; the cross-attention mechanism is responsible for aligning the features of the two domains and outputting the common features of the source and target domains.

[0017] Furthermore, the feature extractor extracts temporal features from the training data of the source and target conditions through a weight-sharing three-branch structure, including private features and common features of the two domains. During the training process, the feature reconstructor is used to ensure that the original information is not lost after feature extraction; the regression predictor is responsible for converting the extracted features into accurate NOx emission prediction values.

[0018] Furthermore, the process of finding the mapping relationship between the extracted temporal features and NOx emissions is as follows: the mapping relationship between the source domain screening features and the NOx emission indicators is learned through full connection layer training.

[0019] The NOx emission prediction method of the gas-steam combined cycle generator set, the training optimization loss L includes the difference loss L difference , regression loss L regre and the reconstruction loss L recon , the training goal is to minimize the loss L;

[0020] L=L difference +k1L regre +k2L recon

[0021] Among them, k1 and k2 are the weight parameters of regression loss and reconstruction loss, respectively, which are used to balance the contribution of different losses;

[0022] Difference loss L difference It is used to calculate the difference between the extracted common features and private features. The stronger the orthogonality between the two, the greater the difference and the smaller the loss. It is defined as:

[0023]

[0024] Among them, H S+T represents the common feature representation, H S and H T Represent the private feature representations of source domain data and target domain data samples respectively; is the square of the Frobenius norm, which is 0 when the two matrices are orthogonal;

[0025] Regression loss L regre Used to train the model to predict the final performance results, expressed as mean square error:

[0026]

[0027] Where N is the batch size of the training data, y i and They represent the actual and predicted values ​​of the target domain performance respectively;

[0028] Reconstruction loss L recon It is used to ensure the integrity of the reconstructed source domain and target domain data, and is represented by cross entropy H:

[0029]

[0030] in, and are the original source domain data and the reconstructed source domain data, respectively. Reconstruct loss for the source domain; and are the original target domain data and the reconstructed target domain data, respectively. Reconstruction loss for the target domain, N S and N T is the number of samples in the source domain and the target domain.

[0031] Through the above design scheme, the present invention can bring the following beneficial effects: the NOx emission prediction method of the gas-steam combined cycle power generation unit proposed in the present invention solves the problem of limited accuracy of the existing deep learning prediction model when the data is insufficient based on the CATrans-DSN model. The feature extractor based on CATrans uses the attention mechanism to learn the domain common and private time features of the source domain and target domain load data, and uses the common features for knowledge transfer; the feature reconstructor acts as an auxiliary module to reconstruct the source domain and target domain data; the regression predictor converts the learned features into predicted values. Finally, the NOx emission prediction model trained on the source domain and target domain is directly used for NOx emission prediction to achieve NOx emission prediction and ensure the low-carbon and environmentally friendly operation of the unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the NOx simulation result diagram under different fuel flow conditions;

[0033] Figure 2 It is the NOx simulation result diagram under different IGV working conditions;

[0034] Figure 3 This is a comparison chart between transfer learning and traditional machine learning;

[0035] Figure 4 This is the network structure diagram of the feature extractor based on the CATrans layer;

[0036] Figure 5 It is a structural diagram based on the CATrans-DSN model;

[0037] Figure 6 This is an analysis chart of factors affecting NOx. DETAILED DESCRIPTION

[0038] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the following is a detailed description of the present invention in combination with the embodiments of the present invention. Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 , the technical solution in the present invention is clearly and completely described. Obviously, the present invention is not limited by the following embodiments, and the specific implementation method can be determined according to the technical solution of the present invention and the actual situation. In order to avoid confusing the essence of the present invention, the known methods, processes and procedures are not described in detail.

[0039] During the operation of gas-steam combined cycle generator sets, the actuator mainly adjusts the fuel flow to meet the performance requirements under different working conditions. The inlet guide vanes (IGV) in the gas turbine are not only responsible for controlling the airflow entering the machine, but also affect the efficiency of the turbine by changing the turbine inlet temperature. Figure 1 and Figure 2 The NOx simulation results under different fuel flow rates and IGV conditions are shown respectively.

[0040] Depend on Figure 1 and Figure 2 It can be seen that under different working conditions, NOx emissions have extremely strong nonlinearity. Considering that transfer learning is suitable for data feature learning under different working conditions, the present invention aims to solve the measurement problem of NOx emissions, an environmental indicator, during the operation of gas-steam combined cycle generator sets under different working conditions, realize NOx emission prediction, and ensure low-carbon and environmentally friendly operation of the units.

[0041] Principle of the method:

[0042] Gas turbine NOx emissions are affected by many factors, including air-fuel ratio, fuel composition, premixing degree of fuel and air, combustion temperature, residence time, and ambient temperature. When calculating NOx emissions, the empirical formula method is used.

[0043] Adiabatic flame temperature calculation formula:

[0044]

[0045] Where A, α, β and λ are constants, given by table lookup, π = p / p ref is dimensionless pressure, p is pressure, unit is MPa, p ref is the reference pressure, the value is 0.1013MPa, θ=T / T ref is the dimensionless temperature, T is the temperature, unit K, T ref is the reference temperature, which is 300K, ψ=H / C is the atomic ratio of hydrogen and carbon, and φ is the equivalent ratio. When φ≤1, σ=φ, and when φ≥1, σ=φ-0.7. * ,y * , z* is a quadratic function of σ and is calculated according to the following formula:

[0046] x * =a1+b1σ+c1σ 2

[0047] y * =a2+b2σ+c2σ 2

[0048] z * =a3+b3σ+c3σ 2 (2)

[0049] Wherein a1, a2, a2, b1, b2, b3, c1, c2, c3, σ and the constants in formula (1) are calculated by looking up Table 1.

[0050] Table 1 Summary of constants in equations

[0051]

[0052]

[0053] Using the calculated adiabatic flame temperature, the NOx emission level is estimated according to the following formula:

[0054]

[0055] Where, T f is the flame temperature, τ=0.022s is the flame residence time, p2 is the combustion chamber inlet pressure, Δp2 / p2 is the dimensionless pressure drop of the combustion chamber, and EI stands for the emission index (Emission Index), which is used to evaluate the emission level of nitrogen oxides NOx.

[0056] Based on the above formula, this paper proposes a method for predicting NOx emissions from a gas-steam combined cycle generator set based on CATrans-DSN. In order to solve the problem that the accuracy of deep learning prediction models is limited when data is insufficient, a deep transfer learning method combining the cross attention mechanism of Transformer (CATrans) and the domain separation network (DSN) is proposed, namely CATrans-DSN. Figure 3 This is a comparison chart between transfer learning and traditional machine learning; it is used for NOx emission prediction of gas-steam combined cycle generator sets. Feature extractor G based on CATrans layer f The attention mechanism is used to learn the domain-shared and private temporal features of the source domain and target domain load data, and the shared features are used for knowledge transfer; the feature reconstructor G c As an auxiliary module, the source domain and target domain data are reconstructed; the regression predictor G r The learned features are converted into prediction values. Finally, the NOx emission prediction model trained on the source domain and the target domain is directly used for NOx emission prediction.

[0057] In order to avoid negative transfer and improve the generalization ability of the model, the present invention adopts Domain Separation Networks (DSN). The core work of the domain separation network is divided into two parts: first, extracting common features between different domains; second, using these common features for knowledge transfer. The domain separation network splits the samples of the source domain and the target domain into two parts, one of which is to encode the specific information in each domain through the encoder private to the two domains, and the other is to extract the common features through the encoder shared by the two domains. In the classification task, it is attempted to make the private features and the common features orthogonal to reflect their irrelevance, and this goal is achieved through the difference loss of the two domains. At the same time, the similarity loss is used to ensure that the common features generated by the two domains are as similar as possible. The outputs of the three encoders are decoded by a common decoder to ensure that the private features and the common features can restore the previous samples as much as possible through the decoder, and are optimized by the reconstruction loss of the two domains. After the model training is completed, a classifier is trained using the generated common features, which can be effectively applied to the target domain and show good performance.

[0058] In the study of deep transfer learning, previous methods mainly focus on finding the mapping between the source domain representation and the target representation, or exploring the shared representation between the two domains. Although these methods have been successful in practice, the shared representation may be disturbed by the underlying shared distribution noise, affecting the performance of the model. Therefore, the present invention improves on the basis of the DSN feature extraction module and designs a feature extractor G based on the CATrans layer. f , to extract the temporal characteristics of NOx emission data in the source and target domains, and combine them with the regression predictor G r Carry out prediction tasks.

[0059] like Figure 4 As shown, the feature extractor G based on the CATrans layer f It includes three weight-sharing branches, namely the source branch, the source-target branch, and the target branch. S+T represents the hidden shared representation in the source domain data and the target domain data samples, H S and H T Represent the private feature representation of source domain data and target domain data samples respectively. The source domain data and target domain data are input to the feature extractor G based on the CATrans layer. f In the above, the self-attention mechanism is responsible for learning the representation of a specific domain and outputting the private features of the source domain and the target domain respectively; the cross-attention mechanism is responsible for aligning the features of the two domains and outputting the common features of the source and target domains. fIt can not only extract private features and common features of the source domain and target domain with strong orthogonality, but also ensure high accuracy and robustness of the target NOx emission prediction.

[0060] The structural diagram based on the CATrans-DSN model is as follows Figure 5 As shown, it includes a feature extractor G f , regression predictor G r and feature reconstructor G c . Feature Extractor G f It consists of at least one CATrans layer, the regression predictor G r and feature reconstructor G c Each of them consists of at least one fully connected layer.

[0061] (1) Feature extractor G based on CATrans layer f , extracting private and common features of the input source and target domain data. By utilizing the cross-attention mechanism, the feature extractor G based on the CATrans layer f It can effectively identify and separate the key features in the source domain and target domain data, laying a solid foundation for subsequent prediction tasks.

[0062] (2) Regression predictor G r The purpose is to find the mapping relationship between the extracted temporal features and NOx emissions based on the source domain and target domain data. The mapping relationship between the source domain screening features and the NOx emission index is learned through the fully connected layer training. After the regression loss training reaches the prediction accuracy, the target domain data is tested. Therefore, the regression predictor G r Convert the extracted features into accurate NOx emission predictions.

[0063] (3) Feature Reconstructor G c It is used to reconstruct samples composed of private features and shared features to obtain the reconstructed source domain and target domain. The reconstruction loss ensures that the feature representation learned by the model can retain the essential information of the source domain and target domain data as much as possible, while reducing unnecessary noise and domain specificity.

[0064] Indicates the sample size N S The source domain dataset, and Represent the input and output of source domain data respectively, Indicates the sample size N T The target domain dataset, and Represent the input and output of the target domain data respectively. The training optimization loss L of the CATrans-DSN model includes the difference loss L difference , regression loss L regre and the reconstruction loss Lrecon , the training goal is to minimize the loss L.

[0065] L=L difference +k1L regre +k2L recon (4)

[0066] Among them, k1 and k2 are the weight parameters of regression loss and reconstruction loss, respectively, which are used to balance the contribution of different losses.

[0067] Difference loss L difference It is used to calculate the difference between the extracted common features and private features. The stronger the orthogonality between the two, the greater the difference and the smaller the loss. It is defined as:

[0068]

[0069] Among them, H S+T represents the common feature representation, H S and H T Represent the private feature representations of source domain data and target domain data samples respectively. is the square of the Frobenius norm, which is 0 when the two matrices are orthogonal.

[0070] Regression loss L regre Used to train the model to predict the final performance results, expressed as mean square error:

[0071]

[0072] Where N is the batch size of the training data, y i and They represent the actual and predicted values ​​of the target domain performance respectively.

[0073] Reconstruction loss L recon It is used to ensure the integrity of the reconstructed source domain and target domain data, and is represented by cross entropy H:

[0074]

[0075] in, are the original source domain data and the reconstructed source domain data, respectively. Reconstruct loss for the source domain; are the original target domain data and the reconstructed target domain data, respectively. Reconstruction loss for the target domain, N S and N T is the number of samples in the source domain and the target domain.

[0076] Through this comprehensive loss function design, the DSN model can effectively perform domain adaptation and extract key features that are helpful for NOx emission prediction while ensuring data integrity and prediction accuracy.

[0077] The implementation of this method is:

[0078] The specific process of the NOx emission prediction method for gas-steam combined cycle power generation unit based on CATrans-DSN proposed in the present invention is as follows:

[0079] Step 1: Data collection and data preprocessing. First, the missing data of the source and target conditions are supplemented by linear interpolation and random forest. Then, the data is normalized to ensure that the data is on the same scale for easy model processing. The sliding window method is used to process the original time series data and construct the source-target input pair, which helps the model learn the long-term dependencies in the time series data.

[0080] Step 2: CATrans-DSN model training. Using the CATrans-based feature extractor G f , extracts temporal features from the training data of the source and target conditions through a weight-sharing three-branch structure, including private and shared features of the two domains, providing rich information for subsequent prediction and reconstruction tasks. During the training process, the feature reconstructor G c To ensure that the original information is not lost after feature extraction, the regression predictor G r Responsible for converting the extracted features into accurate NOx emission prediction values.

[0081] Step 3: NOx emission prediction and model evaluation. Input the target operating condition test data into the trained CATrans-DSN model to predict NOx emissions and output the final predicted value. Evaluate the prediction performance of the model by comparing it with the actual NOx emission data.

[0082] Beneficial effects brought by the technical solution of the present invention:

[0083] In order to clearly reveal the factors affecting NOx emissions, the atmospheric temperature (302.15K), pressure (101325Pa) and other environmental factors are kept constant, and the fuel flow rate is changed for simulation analysis. Figure 6 shown.

[0084] Depend on Figure 6 It can be seen that as the fuel flow rate decreases, the compressor pressure ratio and equivalence ratio decrease. When the equivalence ratio is greater than 1, the adiabatic flame temperature increases and NOx emissions increase; when the equivalence ratio is less than 1, the adiabatic flame temperature decreases and NOx emissions decrease.

[0085] According to the above mechanism analysis, the input quantities of the gas-steam combined cycle generator set: fuel flow, IGV, ambient temperature and state quantities: unit power, turbine front temperature, gas turbine output power, turbine input flow are selected. After feature screening, the NOx emission prediction is completed based on CATrans-DSN. The specific process is as follows. The operating data of a typical operating condition C of the gas-steam combined cycle generator set is selected as the target domain data, which is dataset C. The source domain data selects the operating data of the same unit near the typical operating condition, which are dataset A and dataset B respectively. In order to maintain the continuity of the time series data, the missing data is first filled, and then the z-score normalization method is used to convert the data into a distribution with a mean of 0 and a standard deviation of 1, in which extreme values ​​and noise are reduced, and the distribution characteristics of the data are retained. Since the model cannot automatically identify the input and output variables, the original source domain and target domain data cannot be directly input into the model, so the sliding window is used to process the original input time series.

[0086] After data preprocessing, the model parameters are determined, including the number of fully connected layers, the number of units per layer, the learning rate, the activation function, and the weight optimization during the model training process. These parameters need to be adjusted according to the specific task and data set to achieve the best training effect. The specific parameters are shown in Table 1. Among them, in the CATrans-DSN model, it is necessary to construct a source-target domain data input pair, so the batch size of the source domain and target domain data is set to 512 and 40 respectively. After 6 iterations, all source domain and target training data can be traversed to ensure that the model can fully learn the data features during the training process.

[0087] Table 2 Model network parameter configuration

[0088]

[0089]

[0090] In order to verify the effectiveness of the proposed NOx emission prediction method based on the CATrans-DSN model, two transfer learning NOx emission prediction experiments were set up: Experiment 1 (dataset A→dataset C); Experiment 2 (dataset B→dataset C).

[0091] The mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of variation-root mean square error (CV-RMSE) were used to evaluate the model performance.

[0092]

[0093] The performance evaluation indicators of the four different models in Experiment 1 and Experiment 2 are shown in Table 2. As can be seen from the table, compared with other methods, the model D based on the CATrans-DSN prediction method proposed in the present invention can greatly improve the NOx emission prediction effect.

[0094] Table 3 Prediction performance indicators of four different models

[0095]

Claims

1. A method for predicting NOx emissions from a gas-steam combined cycle power generation unit, characterized in that: The steps include: Step 1: Data collection and data preprocessing, constructing source-target input pairs formed by source domain data and target domain data; Step 2: Establish the CATrans-DSN model and train it; The CATrans-DSN model includes a feature extractor, a regression predictor and a feature reconstructor, wherein the feature extractor is composed of at least one CATrans layer, and the regression predictor and the feature reconstructor are both composed of at least one fully connected layer; The feature extractor is a feature extractor based on the CATrans layer, which is used to extract the private features of the input source domain and target domain data through the self-attention mechanism, and to identify and separate the common features in the source domain and target domain data through the cross-attention mechanism; The regression predictor is used to find the mapping relationship between the extracted temporal features and NOx emissions based on the source domain and target domain data, and convert the extracted features into accurate NOx emission prediction values ​​through regression loss training; The feature reconstructor is used to reconstruct the samples composed of private features and common features to obtain the reconstructed source domain and target domain; Step 3: NOx emission prediction and model evaluation The target operating condition test data is input into the trained CATrans-DSN model to predict NOx emissions, and the final predicted value is output. The prediction performance of the model is evaluated by comparing it with the actual NOx emission data.

2. The NOx emission prediction method for a gas-steam combined cycle power generation unit according to claim 1, characterized in that: In step 1, the missing data of the source and target conditions are first supplemented by linear interpolation and random forest; then, the data is normalized and the sliding window method is used to process the original time series data.

3. The NOx emission prediction method for a gas-steam combined cycle power generation unit according to claim 1, characterized in that: The feature extractor includes three weight-sharing branches, namely the source branch, the source-target branch, and the target branch. The source branch includes a self-attention mechanism and a feedforward network, the source-target branch includes a cross-attention mechanism and a feedforward network, and the target branch includes a self-attention mechanism and a feedforward network. The source domain data and the target domain data are input into the feature extractor. The self-attention mechanism is responsible for learning the representation of a specific domain and outputting the private features of the source domain and the target domain respectively. The cross-attention mechanism is used to align the features of the two domains and output the common features of the source and target domains.

4. The NOx emission prediction method for a gas-steam combined cycle power generation unit according to claim 3, characterized in that: The feature extractor extracts temporal features from the training data of the source and target conditions through a weight-sharing three-branch structure, including private features and common features of the two domains. During the training process, the feature reconstructor is used to ensure that the original information is not lost after feature extraction; the regression predictor is responsible for converting the extracted features into accurate NOx emission prediction values.

5. The NOx emission prediction method for a gas-steam combined cycle power generation unit according to claim 4, characterized in that: The process of finding the mapping relationship between the extracted temporal features and NOx emissions is as follows: the mapping relationship between the source domain screening features and the NOx emission indicators is learned through full connection layer training.

6. The NOx emission prediction method for a gas-steam combined cycle power generation unit according to claim 5, characterized in that: The training optimization loss L of the CATrans-DSN model includes the difference loss L difference , regression loss L regre and the reconstruction loss L recon , the training goal is to minimize the loss L; L=L difference +k1L regre +k2L recon Among them, k1 and k2 are the weight parameters of regression loss and reconstruction loss, respectively, which are used to balance the contribution of different losses; Difference loss L difference It is used to calculate the difference between the extracted common features and private features. The stronger the orthogonality between the two, the greater the difference and the smaller the loss. It is defined as: Among them, H S+T represents the common feature representation, H S and H T Represent the private feature representations of source domain data and target domain data samples respectively; is the square of the Frobenius norm, which is 0 when the two matrices are orthogonal; Regression loss L regre Used to train the model to predict the final performance results, expressed as mean square error: Where N is the batch size of the training data, y i and They represent the actual and predicted values ​​of the target domain performance respectively; Reconstruction loss L recon It is used to ensure the integrity of the reconstructed source domain and target domain data, and is represented by cross entropy H: in, and are the original source domain data and the reconstructed source domain data, respectively. Reconstruct loss for the source domain; and are the original target domain data and the reconstructed target domain data, respectively. Reconstruction loss for the target domain, N S and N T is the number of samples in the source domain and the target domain.

Images