A method and system for predicting main steam temperature of generator sets based on deep learning models

Abstract

This invention discloses a method and system for predicting main steam temperature of generator sets based on a deep learning model. First, time-series data of the generator set's main steam temperature and related operating parameters are acquired. Principal component analysis is used to reduce the dimensionality of the input variables, identifying feature variables highly correlated with the main steam temperature. A convolutional neural network is then used to extract short-term local features, which are flattened and input into a bidirectional long short-term memory network to capture long-term dependencies in the time series from both positive and negative directions. Subsequently, an attention mechanism is used to calculate the importance weights of features at different time steps, generating the main steam temperature prediction result. This invention employs a snow ablation optimization algorithm to dynamically optimize the hyperparameters of the BiLSTM model, effectively balancing the exploration and development capabilities of the search, accelerating model convergence, and improving prediction accuracy. This achieves high-precision, low-latency, and deployable dynamic prediction of main steam temperature, and also features visualization and alarm linkage functions.

Description

Technical Field

[0001] This invention relates to the field of power industry measurement and control and intelligent prediction technology, specifically to a method and system for predicting the main steam temperature of generator sets based on a deep learning model. Background Technology

[0002] Main steam temperature is a critical operating parameter affecting the safety and economy of generator units. Its variation is influenced by multiple factors, including superheater / reheater desuperheating water regulation, boiler airflow, gas valve opening, load power, and ambient temperature, exhibiting strong nonlinearity, strong time-varying characteristics, and multi-scale temporal series features. Existing methods, including ARIMA, grey models, and single LSTM / GRU, generally suffer from insufficient ability to extract local short-term patterns, inadequate long-term dependency capture, reliance on human experience for hyperparameters, slow convergence, and susceptibility to local optima. Therefore, there is an urgent need for a main steam temperature prediction technology that balances local and long-term dependencies, adaptively optimizes, and is suitable for industrial online applications. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for predicting the main steam temperature of generator sets based on a deep learning model, which can achieve high-precision, low-latency, and deployable dynamic prediction of the main steam temperature, and provide visualization and alarm linkage capabilities to overcome the shortcomings of the existing technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for predicting the main steam temperature of a generator set based on a deep learning model includes the following steps:

[0006] S1: The on-site distributed control system (DCS) collects multi-dimensional time series data in real time during the operation of the generator set;

[0007] S2: Perform outlier detection and removal on the data collected in S1, and use the sliding window method for time alignment and interpolation completion;

[0008] S3: Normalize the preprocessed data in S2 to unify different physical quantities into the same dimension range;

[0009] S4: Principal Component Analysis (PCA) algorithm is used to reduce the dimensionality of the normalized multidimensional input variables to obtain a set of principal component eigenvectors;

[0010] S5: By calculating the correlation coefficient between each principal component and the main steam temperature change, the principal feature components that contribute highly to the prediction of the main steam temperature are selected to form the final input feature matrix.

[0011] S6: Input the final input feature matrix into the convolutional neural network (CNN), and extract local short-term dependent features in the time series by setting multiple one-dimensional convolution kernels to perform convolution operations on data at adjacent time steps.

[0012] S7: The extracted local short-term dependency features are compressed and noise suppressed by pooling layers to form a high-dimensional feature mapping matrix containing local dynamic patterns;

[0013] S8: Input the high-dimensional feature mapping matrix into the bidirectional long short-term memory network BiLSTM. BiLSTM consists of a forward LSTM subnetwork and a backward LSTM subnetwork, which are used to capture the deep time feature sequences of the time series in the forward and backward directions, respectively.

[0014] S9: Input the deep temporal feature sequence output by BiLSTM into the Attention layer of the attention mechanism. By calculating the attention weight coefficients of the hidden state vector at each time step, the importance of different time steps is weighted and fused to highlight the dominant role of key moments and key feature parameters in the change of main steam temperature.

[0015] S10: The attention-weighted feature representation is used as the final temporal feature output and input into the fully connected layer for regression calculation to obtain the main steam temperature prediction output value. Historical sample data is divided into training and testing sets, and the parameters of the CNN-BiLSTM-Attention hybrid model are trained and optimized using a mini-batch iterative approach.

[0016] S11: The optimized CNN-BiLSTM-Attention hybrid prediction model is applied to the new sample set, and real-time generator set operation data is input. The main steam temperature is predicted in real time through the hybrid prediction model to obtain the predicted value. ;

[0017] S12: Using the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination Three prediction accuracy evaluation indicators are used to comprehensively evaluate the prediction results, and the training strategy of the model is adjusted or further optimized based on the evaluation results to achieve high-precision dynamic prediction of the main steam temperature change trend.

[0018] Furthermore, the multidimensional time series data in S1 includes at least the following operational characteristic variables: inlet steam temperature of the superheater secondary desuperheater, main steam pressure, outlet steam temperature of the superheater secondary desuperheater, inlet flue gas temperature of the selective catalytic reduction (SCR) unit, outlet steam temperature of the superheater primary desuperheater, inlet steam temperature of the reheater desuperheater, flow rate of the superheater primary desuperheating water, flow rate of the superheater secondary desuperheating water, feedwater flow rate, total boiler air volume, gas valve opening, gas flow rate, unit load, electrical power output, and ambient temperature.

[0019] Furthermore, the convolution operation formula in S6 is as follows:

[0020]

[0021] in, For the first Each convolutional kernel in Output eigenvalues Input features within the time window, For convolution kernel weights, It is a non-linear activation function. This is a bias term.

[0022] Furthermore, the formulas for calculating the forward and backward hidden states of the BiLSTM bidirectional long short-term memory network in S8 are as follows:

[0023]

[0024]

[0025] in, For the positive LSTM at time step The hidden state, For the first Input features at each time step For the positive LSTM at the previous time step The hidden state For inverse LSTM at time step The hidden state, For the inverse LSTM at the previous time step The hidden state;

[0026] By splicing together the forward and reverse hidden states, a deep time feature sequence reflecting the operating state of the generator set can be obtained:

[0027]

[0028] in, For BiLSTM at time steps The final output feature vector.

[0029] Furthermore, the calculation and weighted output of the attention weights in S9 are as follows:

[0030]

[0031]

[0032] in, For learnable weight vectors, For time step Raw attention score, Used to score attention Convert to non-negative values. To sum the exponential scores over all time steps, a normalization constant is obtained. For time step Attention weights To perform a weighted summation of the hidden states over all time steps, This is a weighted feature representation.

[0033] Furthermore, in S10, during the process of dividing historical sample data into training and test sets, the mean squared error (MSE) is used as the loss function, which is defined as follows:

[0034]

[0035] in, For the first The true main steam temperature value of each sample. For the corresponding predicted value, This represents the number of samples.

[0036] Furthermore, in S10, during model training, the Snow Ablation Optimization Algorithm (SAO) is used to adaptively search and update the model's hyperparameters globally. The optimization variables include the kernel size in the convolutional neural network. Regularization coefficients in Bidirectional Long Short-Term Memory (BiLSTM) networks The SAO algorithm performs a global search by simulating the gradual melting mechanism of ice and snow during energy transfer. Its energy balance and perturbation update formula is expressed as:

[0037]

[0038] in, For the first System energy in the next iteration As a learning factor, This is the maximum temperature control parameter. The current search temperature, For parameter perturbation terms;

[0039] The SAO algorithm continuously updates the energy state. With parameter vector In the search space, a balance is struck between local perturbation and global convergence to minimize the loss function. For the goal:

[0040]

[0041] The optimal parameter combination obtained by optimizing the SAO algorithm Substituting these values ​​into the CNN-BiLSTM-Attention model results in an optimized hybrid deep learning prediction model, thereby improving the model's convergence speed and achieving optimal prediction accuracy.

[0042] Furthermore, in S12, the Mean Absolute Error (MAE) is used to measure the average absolute difference between the predicted and actual values, and the calculation formula is as follows:

[0043]

[0044] The root mean square error (RMSE) measures the square root of the prediction error and reflects its magnitude. The calculation formula is as follows:

[0045]

[0046] Coefficient of determination The formula used to measure a model's ability to explain data variation is as follows:

[0047] .

[0048] This invention provides another technical solution: a generator set main steam temperature prediction system based on a deep learning model, comprising:

[0049] The data acquisition module is used to acquire real-time operating parameters and main steam temperature time series data of the gas generator set based on the field distributed control system (DCS); the data acquisition module also includes a data caching unit for time synchronization and formatting of the acquired high-frequency data stream.

[0050] The data preprocessing module is used to perform outlier detection and removal, time interpolation completion and normalization on the collected data, and to use principal component analysis (PCA) to reduce the dimensionality of the input variables and extract principal component features that are highly correlated with the main steam temperature change in order to construct the model input matrix.

[0051] The deep feature extraction and temporal modeling module is used to input the feature variables after PCA dimensionality reduction into a convolutional neural network (CNN), extract local short-term dependency features of the time series through multi-layer convolution and pooling operations, and form a high-dimensional feature map; and input the feature map output by the CNN into a bidirectional long short-term memory network (BiLSTM) to capture long-term dependencies from both forward and backward time dimensions, and calculate the time step weights through the attention mechanism, thereby obtaining a weighted and fused deep temporal feature representation;

[0052] The model training and optimization module is used to adaptively optimize the hyperparameters of the CNN-BiLSTM-Attention model using the Snow Ablation Optimization Algorithm (SAO). The mean squared error (MSE) is used as the loss function to automatically optimize the kernel size of the CNN and the regularization coefficient of the BiLSTM, thereby obtaining the globally optimal combination of hyperparameters and improving the model's convergence speed and prediction accuracy.

[0053] The prediction and performance evaluation module applies the optimized model to real-time data input to generate predicted main steam temperature values, and uses mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. Three evaluation metrics are used to verify the prediction performance, thereby comprehensively evaluating the prediction accuracy and model stability.

[0054] The system deployment and interaction module deploys the system on industrial edge computing nodes or cloud server platforms and connects to the unit control system through the human-machine interface module to realize the visualization of prediction results, historical trend playback and abnormal alarm functions; at the same time, it supports remote model updates and automated parameter tuning to achieve closed-loop control of intelligent prediction of main steam temperature and optimization of operating status.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] The present invention relates to a generator set main steam temperature prediction method and system based on a deep learning model. It employs the Snow Ablation Optimization (SAO) algorithm to dynamically optimize the hyperparameters of the BiLSTM model, effectively balancing the exploration and development capabilities of the search, accelerating model convergence and improving prediction accuracy. It outperforms traditional models in both prediction accuracy and convergence speed. When applied to the generator set main steam temperature prediction process, it can achieve high-precision, low-latency, and deployable dynamic prediction of main steam temperature. Furthermore, in conjunction with the interactive module deployed in the system, it has visualization and alarm linkage functions. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the prediction method of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 This invention provides a method for predicting the main steam temperature of generator sets based on a deep learning model. This method achieves high-precision, low-latency, and deployable dynamic prediction of the main steam temperature of generator sets, and provides visualization and alarm linkage capabilities. The method includes the following steps:

[0060] S1: The distributed control system (DCS) based on the field collects multi-dimensional time series data in real time during the operation of the generator set. The multi-dimensional time series data includes at least fifteen operating characteristic variables: inlet steam temperature of the superheater secondary desuperheater, main steam pressure, outlet steam temperature of the superheater secondary desuperheater, inlet flue gas temperature of the selective catalytic reduction (SCR) unit, outlet steam temperature of the superheater primary desuperheater, inlet steam temperature of the reheater desuperheater, flow rate of superheater primary desuperheating water, flow rate of superheater secondary desuperheating water, feedwater flow rate, total boiler air volume, gas valve opening, gas flow rate, unit load, electrical power output, and ambient temperature.

[0061] S2: Perform outlier detection and removal on the data collected in S1, and use the sliding window method for time alignment and interpolation completion;

[0062] S3: Normalize the preprocessed data in S2 to unify different physical quantities to the same dimension range, so as to reduce the scale difference between features.

[0063] S4: Principal component analysis (PCA) algorithm is used to reduce the dimensionality of the normalized multidimensional input variables to obtain a set of principal component eigenvectors;

[0064] S5: By calculating the correlation coefficients between each principal component and the main steam temperature change, the principal feature components that contribute highly to the prediction of the main steam temperature are selected to form the final input feature matrix; thereby improving the feature extraction efficiency and prediction accuracy of the subsequent model while ensuring the integrity and representativeness of the data.

[0065] S6: The final input feature matrix is ​​fed into a convolutional neural network (CNN). Multiple one-dimensional convolutional kernels are used to perform convolution operations on data from adjacent time steps to extract local short-term dependent features from the time series. The convolution operation can be represented as...

[0066]

[0067] in, For the first Each convolutional kernel in Output eigenvalues Input features within the time window, For convolution kernel weights, It is a non-linear activation function. For bias terms;

[0068] S7: The extracted local short-term dependency features are compressed and noise suppressed by pooling layers to form a high-dimensional feature mapping matrix containing local dynamic patterns;

[0069] S8: Input the high-dimensional feature mapping matrix into a bidirectional long short-term memory (BiLSTM) network. The BiLSTM consists of a forward LSTM subnetwork and a backward LSTM subnetwork, used to capture long-term dependencies of the time series in the forward and backward directions, respectively. The formulas for calculating the forward and backward hidden states of the BiLSTM network are as follows:

[0070]

[0071]

[0072] in, For the positive LSTM at time step The hidden state, For the first Input features at each time step For the positive LSTM at the previous time step The hidden state For inverse LSTM at time step The hidden state, For the inverse LSTM at the previous time step The hidden state.

[0073] By splicing together the forward and reverse hidden states, a deep time feature sequence reflecting the operating state of the generator set can be obtained:

[0074] ;

[0075] in, For BiLSTM at time steps The final output feature vector.

[0076] S9: The deep temporal feature sequence output by BiLSTM is input into the attention mechanism layer. By calculating the attention weight coefficients of the hidden state vector at each time step, the importance of different time steps is weighted and fused to highlight the dominant role of key moments and key feature parameters in the main steam temperature change. The calculation and weighted output of the attention weights are as follows:

[0077]

[0078]

[0079] in, For learnable weight vectors, For time step Raw attention score, Used to score attention Convert to non-negative values. To sum the exponential scores over all time steps, a normalization constant is obtained. For time step Attention weights To perform a weighted summation of the hidden states over all time steps, For weighted feature representation,

[0080] S10: The feature representation weighted by Attention is used as the final temporal feature output and input into the subsequent prediction layer for fitting and predicting the main steam temperature. This achieves multi-scale feature extraction and long- and short-term dynamic modeling of the gas generator set's operating parameters, improving the temporal awareness and feature representation accuracy of the main steam temperature prediction model. The feature representation weighted by the Attention layer is then input into the fully connected layer for regression calculation to obtain the predicted main steam temperature output value. Historical sample data is divided into training and testing sets. The CNN-BiLSTM-Attention hybrid model is trained and optimized using a mini-batch iterative approach. The mean squared error (MSE) is used as the loss function, defined as follows:

[0081]

[0082] in, For the first The true main steam temperature value of each sample. For the corresponding predicted value, This represents the number of samples.

[0083] During the training of the aforementioned model, the Snow Ablation Optimization (SAO) algorithm is used to adaptively search and update the model's hyperparameters. The optimization variables include the kernel size in the convolutional neural network. Regularization coefficients in Bidirectional Long Short-Term Memory (BiLSTM) networks The SAO algorithm performs a global search by simulating the gradual melting mechanism of ice and snow during energy transfer. Its energy balance and perturbation update can be represented as:

[0084]

[0085] in, For the first System energy in the next iteration As a learning factor, This is the maximum temperature control parameter. The current search temperature, This is the parameter perturbation term.

[0086] The SAO algorithm continuously updates the energy state. With parameter vector In the search space, a balance is struck between local perturbation and global convergence to minimize the loss function. For the goal:

[0087]

[0088] The optimal parameter combination obtained through SAO optimization Substituting these values ​​into the CNN-BiLSTM-Attention model results in an optimized hybrid deep learning prediction model, thereby improving the model's convergence speed and achieving optimal prediction accuracy.

[0089] S11: The SAO-optimized CNN-BiLSTM-Attention hybrid prediction model is applied to the new sample set, and real-time collected generator set operating data is input. The hybrid prediction model is used to predict the main steam temperature in real time to obtain the predicted value. ;

[0090] S12: Using the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination ( Three prediction accuracy evaluation indicators are used to comprehensively verify the prediction results. MAE is used to measure the average absolute difference between the predicted value and the actual value; RMSE measures the square root of the prediction error, which can reflect the magnitude of the prediction error. Used to measure the model's ability to explain data variation; denoted as:

[0091]

[0092]

[0093]

[0094] The prediction accuracy and model stability are comprehensively evaluated based on the above three indicators. The training strategy of the model is adjusted or further optimized based on the evaluation results to achieve high-precision dynamic prediction of the main steam temperature change trend.

[0095] To further explain the implementation process of the above method, this embodiment of the invention also provides a prediction system for implementing the above method, which includes a data acquisition module, a data preprocessing module, a deep feature extraction and temporal modeling module, a model training and optimization module, a prediction and performance evaluation module, and a system deployment and interaction module; wherein:

[0096] The data acquisition module is used to collect real-time operating parameters and main steam temperature time series data of the gas generator set based on the field distributed control system (DCS). The operating parameters include multi-dimensional characteristic variables such as the inlet steam temperature of the superheater secondary desuperheater, the main steam pressure, the outlet steam temperature of the superheater secondary desuperheater, the inlet flue gas temperature of the selective catalytic reduction (SCR) unit, the inlet steam temperature of the reheater desuperheater, the flow rate of the superheater primary and secondary desuperheating water, the feedwater flow rate, the total boiler air volume, and the opening degree of the gas valves. The module also includes a data caching unit, which is used to synchronize and format the acquired high-frequency data stream to ensure data continuity and stability.

[0097] The data preprocessing module is used to detect and remove outliers, perform time interpolation and normalization on the collected data, and uses principal component analysis (PCA) to reduce the dimensionality of the input variables and extract principal component features that are highly correlated with the main steam temperature change in order to construct the model input matrix.

[0098] The deep feature extraction and temporal modeling module is used to input the feature variables after PCA dimensionality reduction into a convolutional neural network (CNN), extract local short-term dependency features of the time series through multi-layer convolution and pooling operations, and form a high-dimensional feature map. The feature map output by the CNN is input into a bidirectional long short-term memory network (BiLSTM) to capture long-term dependencies from both forward and backward time dimensions, and calculate the time step weights through an attention mechanism, thereby obtaining a weighted and fused deep temporal feature representation.

[0099] The model training and optimization module uses the Snow Ablation Optimization (SAO) algorithm to adaptively optimize the hyperparameters of the CNN-BiLSTM-Attention model, with mean squared error (MSE) as the loss function. This module automatically optimizes the kernel size of the CNN and the regularization coefficients of the BiLSTM to obtain the globally optimal hyperparameter combination, improving the model's convergence speed and prediction accuracy.

[0100] The prediction and performance evaluation module applies the optimized model to real-time data input, generates predicted main steam temperature values, and verifies the prediction performance using the following three evaluation indicators. This allows for a comprehensive assessment of prediction accuracy and model stability, achieving high-precision dynamic prediction of the main steam temperature variation trend of the gas turbine generator set.

[0101] The system deployment and interaction module is used for data interaction and system deployment. Its system can be deployed on industrial edge computing nodes or cloud server platforms, supports high-frequency real-time data stream access and online model inference, and connects to the unit control system through the human-machine interaction interface module to realize the visualization of prediction results, historical trend playback and abnormal alarm functions. At the same time, it supports remote model updates and automated parameter tuning to achieve closed-loop control of intelligent prediction of main steam temperature and optimization of operating status.

[0102] In summary, this invention provides a method and system for predicting main steam temperature of generator sets based on a deep learning model. It employs the Snow Ablation Optimization (SAO) algorithm to dynamically optimize the hyperparameters of the BiLSTM model, effectively balancing the exploration and development capabilities of the search, accelerating model convergence, and improving prediction accuracy. It outperforms traditional models in both prediction accuracy and convergence speed. Applying this method to the prediction process of main steam temperature of generator sets enables high-precision, low-latency, and deployable dynamic prediction of main steam temperature. Furthermore, the system's interactive module provides visualization and alarm linkage functions.

[0103] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the main steam temperature of a generator set based on a deep learning model, characterized in that, Includes the following steps: S1: The on-site distributed control system (DCS) collects multi-dimensional time series data in real time during the operation of the generator set; S2: Perform outlier detection and removal on the data collected in S1, and use the sliding window method for time alignment and interpolation completion; S3: Normalize the preprocessed data in S2 to unify different physical quantities into the same dimension range; S4: Principal Component Analysis (PCA) algorithm is used to reduce the dimensionality of the normalized multidimensional input variables to obtain a set of principal component eigenvectors; S5: By calculating the correlation coefficient between each principal component and the main steam temperature change, the principal feature components that contribute highly to the prediction of the main steam temperature are selected to form the final input feature matrix. S6: Input the final input feature matrix into the convolutional neural network (CNN), and extract local short-term dependent features in the time series by setting multiple one-dimensional convolution kernels to perform convolution operations on data at adjacent time steps. S7: The extracted local short-term dependency features are compressed and noise suppressed by pooling layers to form a high-dimensional feature mapping matrix containing local dynamic patterns; S8: Input the high-dimensional feature mapping matrix into the bidirectional long short-term memory network BiLSTM. BiLSTM consists of a forward LSTM subnetwork and a backward LSTM subnetwork, which are used to capture the deep time feature sequences of the time series in the forward and backward directions, respectively. S9: Input the deep temporal feature sequence output by BiLSTM into the Attention layer of the attention mechanism. By calculating the attention weight coefficients of the hidden state vector at each time step, the importance of different time steps is weighted and fused to highlight the dominant role of key moments and key feature parameters in the change of main steam temperature. S10: The attention-weighted feature representation is used as the final temporal feature output and input into the fully connected layer for regression calculation to obtain the main steam temperature prediction output value. Historical sample data is divided into training and testing sets, and the parameters of the CNN-BiLSTM-Attention hybrid model are trained and optimized using a mini-batch iterative approach. S11: The optimized CNN-BiLSTM-Attention hybrid prediction model is applied to the new sample set, and real-time generator set operation data is input. The main steam temperature is predicted in real time through the hybrid prediction model to obtain the predicted value. ; S12: Using the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination Three prediction accuracy evaluation indicators are used to comprehensively evaluate the prediction results, and the training strategy of the model is adjusted or further optimized based on the evaluation results to achieve high-precision dynamic prediction of the main steam temperature change trend.

2. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 1, characterized in that: The multidimensional time series data in S1 includes at least the following operational characteristic variables: inlet steam temperature of the superheater secondary desuperheater, main steam pressure, outlet steam temperature of the superheater secondary desuperheater, inlet flue gas temperature of the selective catalytic reduction (SCR) unit, outlet steam temperature of the superheater primary desuperheater, inlet steam temperature of the reheater desuperheater, flow rate of superheater primary desuperheating water, flow rate of superheater secondary desuperheating water, feedwater flow rate, total boiler air volume, gas valve opening, gas flow rate, unit load, electrical power output, and ambient temperature.

3. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 1, characterized in that: The convolution operation formula in S6 is as follows: in, For the first Each convolutional kernel in Output eigenvalues Input features within the time window, For convolution kernel weights, It is a non-linear activation function. This is a bias term.

4. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 3, characterized in that: The formulas for calculating the forward and backward hidden states of the BiLSTM bidirectional long short-term memory network in S8 are as follows: in, For the positive LSTM at time step The hidden state, For the first Input features at each time step For the positive LSTM at the previous time step The hidden state For inverse LSTM at time step The hidden state, For the inverse LSTM at the previous time step The hidden state; By splicing together the forward and reverse hidden states, a deep time feature sequence reflecting the operating state of the generator set can be obtained: ； in, For BiLSTM at time steps The final output feature vector.

5. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 4, characterized in that: The calculation and weighted output of attention weights in S9 are as follows: in, For learnable weight vectors, For time steps Raw attention score, Used to score attention Convert to non-negative values. To sum the exponential scores over all time steps, a normalization constant is obtained. For time steps Attention weights To perform a weighted summation of the hidden states over all time steps, This is a weighted feature representation.

6. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 5, characterized in that: In S10, during the process of dividing historical sample data into training and test sets, the mean squared error (MSE) is used as the loss function, which is defined as follows: in, For the first The true main steam temperature value of each sample. For the corresponding predicted value, This represents the number of samples.

7. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 6, characterized in that: In S10, during model training, the Snow Ablation Optimization (SAO) algorithm is used to adaptively search and update the model's hyperparameters globally. The optimization variables include the kernel size in the convolutional neural network. Regularization coefficients in Bidirectional Long Short-Term Memory (BiLSTM) networks The SAO algorithm performs a global search by simulating the gradual melting mechanism of ice and snow during energy transfer. Its energy balance and perturbation update formula is expressed as: in, For the first System energy in the next iteration As a learning factor, This is the maximum temperature control parameter. The current search temperature, For parameter perturbation terms; The SAO algorithm continuously updates the energy state. With parameter vector In the search space, a balance is struck between local perturbation and global convergence to minimize the loss function. For the goal: The optimal parameter combination obtained by optimizing the SAO algorithm Substituting these values ​​into the CNN-BiLSTM-Attention model results in an optimized hybrid deep learning prediction model, thereby improving the model's convergence speed and achieving optimal prediction accuracy.

8. The method for predicting the main steam temperature of a generator set based on a deep learning model as described in claim 7, characterized in that: In S12, the Mean Absolute Error (MAE) measures the average absolute difference between the predicted and the actual values. The calculation formula is as follows: The root mean square error (RMSE) measures the square root of the prediction error and reflects its magnitude. The calculation formula is as follows: Coefficient of determination The formula used to measure a model's ability to explain data variation is as follows: 。 9. A generator set main steam temperature prediction system based on a deep learning model, used to implement the generator set main steam temperature prediction method based on a deep learning model as described in claim 1, characterized in that, include: The data acquisition module is used to acquire real-time operating parameters and main steam temperature time series data of the gas generator set based on the field distributed control system (DCS). The data acquisition module also includes a data caching unit for time synchronization and formatting of the acquired high-frequency data stream; The data preprocessing module is used to perform outlier detection and removal, time interpolation completion and normalization on the collected data, and to use principal component analysis (PCA) to reduce the dimensionality of the input variables and extract principal component features that are highly correlated with the main steam temperature change in order to construct the model input matrix. The deep feature extraction and temporal modeling module is used to input the feature variables after PCA dimensionality reduction into a convolutional neural network (CNN), extract local short-term dependency features of the time series through multi-layer convolution and pooling operations, and form a high-dimensional feature map; and input the feature map output by the CNN into a bidirectional long short-term memory network (BiLSTM) to capture long-term dependencies from both forward and backward time dimensions, and calculate the time step weights through the attention mechanism, thereby obtaining a weighted and fused deep temporal feature representation; The model training and optimization module is used to adaptively optimize the hyperparameters of the CNN-BiLSTM-Attention model using the Snow Ablation Optimization Algorithm (SAO). The mean squared error (MSE) is used as the loss function to automatically optimize the kernel size of the CNN and the regularization coefficient of the BiLSTM, thereby obtaining the globally optimal combination of hyperparameters and improving the model's convergence speed and prediction accuracy. The prediction and performance evaluation module applies the optimized model to real-time data input to generate predicted main steam temperature values, and uses mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. Three evaluation metrics are used to verify the prediction performance, thereby comprehensively evaluating the prediction accuracy and model stability. The system deployment and interaction module deploys the system on industrial edge computing nodes or cloud server platforms and connects to the unit control system through the human-machine interface module to realize the visualization of prediction results, historical trend playback and abnormal alarm functions; at the same time, it supports remote model updates and automated parameter tuning to achieve closed-loop control of intelligent prediction of main steam temperature and optimization of operating status.

Images