Gas turbine dynamic parameter regression prediction method based on fusion network

Abstract

The invention relates to the field of running state monitoring of gas turbines for power generation, in particular to a gas turbine dynamic parameter regression prediction method based on a fusion network, which comprises the following steps: S1, data acquisition: acquiring historical data of a gas turbine running atlas; s2, data preprocessing: performing steady-state data screening and abnormal value elimination on the collected historical data, determining input parameters and output parameters of a prediction model, setting a step length for reading the historical data each time, segmenting the data through a sliding window method, dividing data of a training set and a test set, and the like; s3, the prediction model extracts a spatial feature vector through a convolutional layer in a convolutional neural network CNN, then fully extracts time information through a mode of combining Transform and an LSTM network, and obtains potential spatial-temporal correlation; and S4, training test and storage of the prediction model: training the prediction model through the training data, judging whether the trained prediction model meets the requirement or not through the test result of the test set, and storing the prediction model meeting the requirement.

Description

Technical Field

[0001] The invention relates to the technical field of operating status monitoring of gas turbines for power generation, and more specifically to a gas turbine dynamic parameter regression prediction method based on a fusion network. Background Art

[0002] Gas turbines are important equipment for achieving efficient energy conversion and clean utilization. They are widely used in many fields and are an important manifestation of the level of industrialization. However, due to the long-term operation of gas turbines under high speed, high temperature, high pressure and high load conditions, the failure rate is high, which affects the stability, economy and safety of their operation. With the advancement of sensor technology and the massive increase in gas turbine operation monitoring data, establishing an online dynamic identification model through gas turbine operating parameters has become one of the important research directions for gas turbine health management. Deep learning algorithms perform well in dealing with nonlinear complex system problems, making it possible to establish a dynamic identification model for gas turbine operating parameters. The model can predict the baseline values ​​of the parameters at each measuring point during the operation of the gas turbine in real time, providing an important reference for the operation status monitoring of the gas turbine. This not only improves the operating efficiency of the gas turbine, but also effectively prevents potential failures and ensures the long-term stable operation of the equipment. Summary of the invention

[0003] The purpose of the present invention is to provide a gas turbine dynamic parameter regression prediction method based on a fusion network, which can perform dynamic parameter regression prediction on different gas turbine models based on historical operating data, and can realize online real-time identification of the dynamic change process of monitoring parameters such as speed, pressure, and temperature, so as to provide important decision support for online anomaly detection of gas turbine monitoring parameters.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A gas turbine dynamic parameter regression prediction method based on a fusion network, the method comprising the following steps:

[0006] Step S1: data collection, collecting historical data of gas turbine operation spectrum;

[0007] Step S2: Data preprocessing, screening the collected historical data for steady-state data, eliminating outliers, determining the input parameters and output parameters of the prediction model, setting the step size for reading historical data each time, splitting the data by sliding window method, dividing the training set and test set data, etc.;

[0008] Steady-state data screening For gas turbine generator sets, steady-state data screening is performed based on the range of unit power and speed and the rate of change of speed;

[0009] In data preprocessing, the input and output parameters are normalized to the maximum and minimum;

[0010] The split of the training set and the test set cannot be completely random, and the temporal consistency of the time series data should be maintained;

[0011] Step S3: The prediction model extracts spatial feature vectors through the convolutional layer in the convolutional neural network CNN, and then fully extracts temporal information by combining the Transformer and LSTM networks to obtain potential spatiotemporal correlations;

[0012] Step S4: training, testing and saving of the prediction model. The prediction model is trained with the training data, and the test results of the test set are used to determine whether the trained prediction model meets the requirements, and the prediction model that meets the requirements is saved;

[0013] In the step S1, the operation history data of the gas turbine includes casing pressure, casing temperature, thrust bearing temperature, fuel control instruction, main gas valve outlet temperature, speed, generator power, fuel main oil line pressure, lubricating oil tank pressure, lubricating oil tank temperature, inlet air temperature, turbine exhaust temperature average, turbine exhaust temperature 1, turbine exhaust temperature 2 and turbine exhaust temperature 3; the monitoring data of the gas turbine under different operating conditions is collected, sensors can be arranged for data collection, and monitoring data collected by the gas turbine control system can also be read, and the sampling time interval is 1 second;

[0014] In step S2, the steady-state data screening specifically includes:

[0015] Utilized unit power x i,kW , speed signal x i,Tgp The change range is used to select typical working conditions and the rate of change of speed. The calculation formula is as follows;

[0016]

[0017] The steady-state data is filtered by the rate of change of the speed, and the steady-state threshold is set to filter the steady-state data of the gas turbine. Let N represent the length of the data data, the window size is w, and the steady-state threshold is θ; define the number of windows M as: M = N-w+1; define the result matrix I as a vector of length M, where each element I i Indicates whether the i-th window is steady-state data; for each window i (from 1 to M), the calculation process is as follows:

[0018] d i ={data j |i≤j≤i+w-1}

[0019]

[0020] Among them, d iis the extracted window data, μ i , σ i is the mean and standard deviation of the window data; if Then I i =1, which means it is determined to be steady-state data;

[0021] In step S2, the outlier elimination method is a mean interpolation algorithm, a regression interpolation algorithm, a multiple interpolation algorithm or an isolation forest algorithm;

[0022] In step S2, the input parameters and output parameters of the prediction model specifically include: the input parameters of the prediction model need to be able to reflect the changes in the outside air temperature and atmospheric pressure, and include at least one parameter related to the operating conditions; the output parameters of the prediction model are the gas turbine measurement point parameters that need to be monitored for operating conditions;

[0023] In step S2, the sliding window method is:

[0024] The monitoring parameters that can characterize the changes in the external environment and the operating conditions of the gas turbine after data processing are used as input parameters of the prediction model and input into the prediction model; the expression form of the input parameters is shown in the formula;

[0025]

[0026] In the formula, x i t It represents the value of the i-th thermal parameter collected at the current time t, k represents the window size, and through the sliding window, the data of i*k dimensions are input each time;

[0027] In step S2, the training set and the data set are divided into 90% of the data as the training set and 10% of the data as the test set according to different working conditions;

[0028] Sliding window method, for time series data with multiple steps and multiple parameters as input and single step and multiple parameters as output, each time series of window size is regarded as a feature. Since the model is applied to single-step prediction, the sliding window size is 1. If the number of sample points of each monitoring parameter in the training set is N, the window size is k, and the sliding step size is 1, then each monitoring parameter can generate N-k+1 training samples. Assuming that the number of input monitoring points is i, the dimension of each input is i*(k+1);

[0029] In step S3, the specific calculation process of the convolutional neural network CNN is as follows:

[0030] Assuming that the model input sequence is X, T represents the length of the input sequence, the calculation of local features through the convolution layer is as follows:

[0031]

[0032] Among them, X is the input feature map after data tiling, the size is H*W*C, H and W are the height and width of the input feature map, C is the number of channels of the input feature map; W is the convolution kernel, M and N are the height and width of the convolution kernel, K is the number of channels of the output feature map; Y is the output feature map, the size is H'*W'*K'; B k is the bias term;

[0033] The purpose of the pooling layer is to downsample the feature map, reduce the amount of calculation and suppress overfitting; the selected maximum pooling calculation is as follows:

[0034]

[0035] Where: X: input feature map, Y: output feature map, M, N: height and width of the pooling window. The pooling layer reduces the size of the feature map by selecting the maximum value or average value in the pooling window.

[0036] The fully connected layer flattens the input feature map into a vector, multiplies it with the weight matrix, adds the bias term, and obtains the output:

[0037] Y=W·X+b

[0038] Where: X: input vector, W: weight matrix, b: bias term, Y: output vector, the fully connected layer is used to map the extracted features to the output space;

[0039] In step S3, the specific calculation processes of the self-attention mechanism and the multi-head attention mechanism of the Transformer are as follows: after the convolutional neural network extracts the features, the self-attention mechanism captures the global dependency in the sequence by calculating the attention weights between the elements in the features;

[0040]

[0041] Where: Q: Query Matrix K: Key Matrix V: Value Matrix k : Dimension of the key

[0042] The multi-head self-attention mechanism concatenates the outputs of multiple self-attention layers:

[0043] MultiHead(Q,K,V)=Concat(head1,...,head n )W O

[0044] in:

[0045] head i =Attention(QW i Q ,KW i K ,VW i v )

[0046] Q=XW Q +b Q

[0047] K=XW K +b K

[0048] V=XW V +b V

[0049] W i Q ,W i K ,W i V : weight matrix of different heads;

[0050] b Q ,b K ,b V : bias term;

[0051] W O : Linear transformation weight matrix after multi-head attention;

[0052] In step S3, the gating mechanism of LSTM, i.e., the forget gate, input gate and output gate are calculated as follows: t =σ(W f ·[h t-1 ,x t ]+b f )

[0053] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0054] C t =tanh(W C ·[h t-1 ,x t ]+b C )

[0055] C t =f t ·C t-1 +i t ·C t

[0056] h t =o t tanh(C t )

[0057] Where: x t : input vector; h t-1 : The hidden state of the previous moment; C t : cell state; f t ,i t ,o t : The activation values ​​of the forget gate, input gate, and output gate; W f ,W i ,W C ,W o : weight matrix; b f ,b i ,b C ,b o : bias term;

[0058] The number of convolutional layers of the convolutional neural network CNN is 2, and the sizes of the convolution kernels are 3*1 and 2*1 respectively. The extracted local features are input into the Transformer through the regression output layer. The Transformer contains a position encoding layer, an addition layer, and two self-attention layers. Each self-attention layer contains 128 channels, and each channel has 4 heads. The features extracted by the Tansformer are input into the LSTM layer and predicted and output through the regression layer.

[0059] In step S3, the model parameters of transformer and LSTM are optimized by Bayesian optimization algorithm, and the model parameters are optimized and solved by calculating the root mean square error between the predicted value and the true value to obtain the best hyperparameter combination with the minimum error. The objective function is calculated as follows:

[0060]

[0061] Where N is the number of samples in the training set, yi is the true value of the i-th sample, and yi-hat is the predicted value of the i-th sample; the Bayesian optimization algorithm optimizes the number of hidden layer nodes, the initialization learning rate, and the L2 regularization coefficient in the model to obtain the best hyperparameter combination with the minimum error;

[0062] In step S4, the predicted results of the test set are compared with the actual values ​​and the error is calculated and analyzed, and the average relative error is used to measure the identification results:

[0063]

[0064] Among them, Act represents the actual value, Pred represents the predicted value, and n is the number of samples; when the test prediction results meet the requirements, the model parameters are saved; when the identification results are measured and the average relative errors of each prediction parameter meet the requirements, the model parameters are saved.

[0065] The beneficial effects of the present invention are:

[0066] By combining the convolutional neural network CNN, Transformer and LSTM models, the spatiotemporal dependencies can be better captured; the convolution layer extracts local features, the Transformer layer captures global dependencies through the self-attention mechanism, the LSTM layer captures the dependencies of long time series, and the fully connected layer performs the final output mapping; this multi-level feature extraction and dependency capture makes the prediction model more expressive when processing complex spatiotemporal data, and can better realize the regression prediction of gas turbine dynamic parameters, providing a basis for parameter anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0068] Figure 1 It is a schematic diagram of the steps of a gas turbine dynamic parameter regression prediction method based on a fusion network of the present invention;

[0069] Figure 2 It is a schematic diagram of the overall model architecture of a gas turbine dynamic parameter regression prediction method based on a fusion network of the present invention;

[0070] Figure 3 It is a schematic diagram of the optimization process of the Yes optimization algorithm of the present invention;

[0071] Figure 4 It is a schematic diagram of the regression prediction result of the exhaust gas temperature 01 after the turbine of a certain type of power generation gas turbine of the present invention;

[0072] Figure 5 It is a schematic diagram of the prediction residual of the exhaust temperature 01 after the turbine of a certain type of gas turbine for power generation according to the present invention. DETAILED DESCRIPTION

[0073] The present invention is further described in detail below with reference to the accompanying drawings.

[0074] like Figures 1 to 5As shown, in order to achieve the technical effect of "being able to perform dynamic parameter regression prediction on different gas turbine models based on historical operation data, being able to realize online real-time identification of the dynamic change process of monitoring parameters such as speed, pressure, and temperature, and providing important decision support for online abnormal detection of gas turbine monitoring parameters", the steps and functions of a gas turbine dynamic parameter regression prediction method based on a fusion network are described in detail below;

[0075] like Figure 1 As shown, a gas turbine dynamic parameter regression prediction method based on a fusion network comprises the following steps:

[0076] A gas turbine dynamic parameter regression prediction method based on a fusion network, the method comprising the following steps:

[0077] Step S1: data collection, collecting the historical data of the operation spectrum of the SGT400 dual-shaft gas turbine; the historical operation data of the gas turbine includes casing pressure, casing temperature, thrust bearing temperature, fuel control command, main gas valve outlet temperature, speed, generator power, fuel main oil line pressure, lubricating oil tank pressure, lubricating oil tank temperature, inlet air temperature, turbine exhaust temperature average value, turbine exhaust temperature 1, turbine exhaust temperature 2 and turbine exhaust temperature 3; collecting the monitoring data of the gas turbine under different operating conditions, sensors can be arranged for data collection, and the monitoring data collected by the gas turbine control system can also be read, and the sampling time interval is 1 second;

[0078] Step S2: Data preprocessing, screening the collected historical data for steady-state data, eliminating outliers, determining the input parameters and output parameters of the prediction model, setting the step size for reading historical data each time, splitting the data by sliding window method, dividing the training set and test set data, etc.;

[0079] Steady-state data screening For gas turbine generator sets, steady-state data screening is performed based on the range of unit power and speed and the rate of change of speed;

[0080] In data preprocessing, the input and output parameters are normalized to the maximum and minimum;

[0081] The split of the training set and the test set cannot be completely random, and the temporal consistency of the time series data should be maintained;

[0082] In step S2, the steady-state data screening specifically includes:

[0083] Utilized unit power x i,kW , speed signal x i,Tgp The change range is used to select typical working conditions and the rate of change of speed. The calculation formula is as follows;

[0084]

[0085] The steady-state data is filtered by the rate of change of the speed, and the steady-state threshold is set to filter the steady-state data of the gas turbine. Let N represent the length of the data data, the window size is w, and the steady-state threshold is θ; define the number of windows M as: M = N-w+1; define the result matrix I as a vector of length M, where each element I i Indicates whether the i-th window is steady-state data; for each window i (from 1 to M), the calculation process is as follows:

[0086]

[0087] Among them, d i is the extracted window data, μ i , σ i is the mean and standard deviation of the window data; if Then I i =1, which means it is determined to be steady-state data;

[0088] In step S2, the outlier elimination method is a mean interpolation algorithm, a regression interpolation algorithm, a multiple interpolation algorithm or an isolation forest algorithm;

[0089] In step S2, the input parameters and output parameters of the prediction model specifically include: the input parameters of the prediction model need to be able to reflect the changes in the outside air temperature and atmospheric pressure, and include at least one parameter related to the operating conditions; the output parameters of the prediction model are the gas turbine measurement point parameters that need to be monitored for operating conditions;

[0090] In step S2, the sliding window method is:

[0091] The monitoring parameters that can characterize the changes in the external environment and the operating conditions of the gas turbine after data processing are used as input parameters of the prediction model and input into the prediction model; the expression form of the input parameters is shown in the formula;

[0092]

[0093] In the formula, x i t It represents the value of the i-th thermal parameter collected at the current time t, k represents the window size, and through the sliding window, the data of i*k dimensions are input each time;

[0094] In step S2, the training set and the data set are divided into 90% of the data as the training set and 10% of the data as the test set according to different working conditions;

[0095] Sliding window method, for time series data with multiple steps and multiple parameters as input and single step and multiple parameters as output, each time series of window size is regarded as a feature. Since the model is applied to single-step prediction, the sliding window size is 1. If the number of sample points of each monitoring parameter in the training set is N, the window size is k, and the sliding step size is 1, then each monitoring parameter can generate N-k+1 training samples. Assuming that the number of input monitoring points is i, the dimension of each input is i*(k+1);

[0096] Utilized unit power x i,kW , speed signal x i,Tgp The change range is used to screen typical operating conditions and the rate of change of the speed, where Pmin=2000, Tmin=99, Tmax=102, the window size w=200, and the rate of change θ=0.01.

[0097] In a specific embodiment, 90% of the data are selected as a training set and 10% of the data are selected as a test set according to different working conditions;

[0098] Step S3: The prediction model extracts spatial feature vectors through the convolutional layer in the convolutional neural network CNN, and then fully extracts temporal information by combining the Transformer and LSTM networks to obtain potential spatiotemporal correlations;

[0099] In step S3, the specific calculation process of the convolutional neural network CNN is as follows:

[0100] Assuming that the model input sequence is X, T represents the length of the input sequence, the calculation of local features through the convolution layer is as follows:

[0101]

[0102] Among them, X is the input feature map after data tiling, the size is H*W*C, H and W are the height and width of the input feature map, C is the number of channels of the input feature map; W is the convolution kernel, M and N are the height and width of the convolution kernel, K is the number of channels of the output feature map; Y is the output feature map, the size is H'*W'*K'; B k is the bias term;

[0103] The purpose of the pooling layer is to downsample the feature map, reduce the amount of calculation and suppress overfitting; the selected maximum pooling calculation is as follows:

[0104]

[0105] Where: X: input feature map, Y: output feature map, M, N: height and width of the pooling window. The pooling layer reduces the size of the feature map by selecting the maximum value or average value in the pooling window.

[0106] The fully connected layer flattens the input feature map into a vector, multiplies it with the weight matrix, adds the bias term, and obtains the output:

[0107] Y=W·X+b

[0108] Where: X: input vector, W: weight matrix, b: bias term, Y: output vector, the fully connected layer is used to map the extracted features to the output space, usually for classification or regression tasks;

[0109] In step S3, the specific calculation processes of the self-attention mechanism and the multi-head attention mechanism of Transformer are as follows: the self-attention mechanism captures the global dependency in the sequence by calculating the attention weights between the elements in the input sequence;

[0110]

[0111] Where: Q: Query Matrix K: Key Matrix V: Value Matrix k : Dimension of the key

[0112] The multi-head self-attention mechanism concatenates the outputs of multiple self-attention layers:

[0113] MultiHead(Q,K,V)=Concat(head1,...,head n )W O

[0114] in:

[0115] head i =Attention(QW i Q ,KW i K ,VW i v )

[0116] Q=XW Q +b Q

[0117] K=XW K +b K

[0118] V=XW V +b V

[0119] W i Q ,W i K,W i V : weight matrix of different heads;

[0120] b Q ,b K ,b V : bias term;

[0121] W O : Linear transformation weight matrix after multi-head attention;

[0122] In step S3, the gating mechanism of LSTM, i.e., the forget gate, input gate and output gate are calculated as follows: t =σ(W f ·[h t-1 ,x t ]+b f )

[0123] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0124] C t =tanh(W C ·[h t-1 ,x t ]+b C )

[0125] C t =f t ·C t-1 +i t ·C t

[0126] h t =o t tanh(C t )

[0127] Where: x t : input vector; h t-1 : The hidden state of the previous moment; C t : cell state; f t ,i t ,o t : The activation values ​​of the forget gate, input gate, and output gate; W f ,W i ,W C ,W o : weight matrix; b f ,b i ,b C ,b o : bias term;

[0128] The number of convolutional layers of the convolutional neural network CNN is 2, and the sizes of the convolution kernels are 3*1 and 2*1 respectively. The extracted local features are input into the Transformer through the regression output layer. The Transformer contains a position encoding layer, an addition layer, and two self-attention layers. Each self-attention layer contains 128 channels, and each channel has 4 heads. The features extracted by the Tansformer are input into the LSTM layer and predicted and output through the regression layer.

[0129] In step S3, the model parameters of transformer and LSTM are optimized by Bayesian optimization algorithm, and the model parameters are optimized and solved by calculating the root mean square error between the predicted value and the true value to obtain the best hyperparameter combination with the minimum error. The objective function is calculated as follows:

[0130]

[0131] Where N is the number of samples in the training set, yi is the true value of the i-th sample, and yi-hat is the predicted value of the i-th sample; the Bayesian optimization algorithm optimizes the number of hidden layer nodes, the initialization learning rate, and the L2 regularization coefficient in the model to obtain the best hyperparameter combination with the minimum error;

[0132] Step S4: training, testing and saving of the prediction model. The prediction model is trained with the training data, and the test results of the test set are used to determine whether the trained prediction model meets the requirements, and the prediction model that meets the requirements is saved;

[0133] In step S4, the predicted results of the test set are compared with the actual values ​​and the error is calculated and analyzed, and the average relative error is used to measure the identification results:

[0134]

[0135] Among them, Act represents the actual value, Pred represents the predicted value, and n represents the number of samples. When the test prediction results meet the requirements, the model parameters are saved. When the identification results are measured and the average relative errors of each prediction parameter meet the requirements, the model parameters are saved.

[0136] In a specific embodiment, there are 4 input parameters of the prediction model, namely, casing pressure, fuel control command, generator power, and inlet air temperature. The casing pressure and inlet air temperature reflect changes in the external environment, and the fuel control command and generator power reflect changes in load.

[0137] In a specific embodiment, the input of the prediction model is 5 historical moments of 4 parameters and the current moment, with a total of 6 steps, and the input layer size of the convolutional neural network is 24*1*1.

[0138] In a specific embodiment, the hyperparameters of the prediction model after optimization by the Bayesian optimization algorithm are as follows: the number of hidden layer nodes is 96, the initialization learning rate is 0.00157, and the L2 regularization coefficient is 0.001.

[0139] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A gas turbine dynamic parameter regression prediction method based on a fusion network, characterized in that: The method comprises the following steps: Step S1: data collection, collecting historical data of gas turbine operation spectrum; Step S2: Data preprocessing, screening the collected historical data for steady-state data, eliminating outliers, determining the input parameters and output parameters of the prediction model, setting the step size for reading historical data each time, splitting the data by sliding window method, dividing the training set and test set data, etc.; Step S3: The prediction model extracts spatial feature vectors through the convolutional layer in the convolutional neural network CNN, and then fully extracts temporal information by combining the Transformer and LSTM networks to obtain potential spatiotemporal correlations; Step S4: training, testing and saving of the prediction model. The prediction model is trained with the training data, and the test results of the test set are used to determine whether the trained prediction model meets the requirements, and the prediction model that meets the requirements is saved.

2. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1 is characterized in that: In step S1, the operating history data of the gas turbine includes casing pressure, casing temperature, thrust bearing temperature, fuel control instructions, main gas valve outlet temperature, speed, generator power, fuel main oil line pressure, lubricating oil tank pressure, lubricating oil tank temperature, inlet air temperature, turbine exhaust temperature average, turbine exhaust temperature 1, turbine exhaust temperature 2 and turbine exhaust temperature 3.

3. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1 is characterized in that: In step S2, the steady-state data screening specifically includes: Utilized unit power x i,kW , speed signal x i,Tgp The change range is used to select typical working conditions and the rate of change of speed. The calculation formula is as follows; The steady-state data is filtered by the rate of change of the speed, and the steady-state threshold is set to filter the steady-state data of the gas turbine. Let N represent the length of the data data, the window size is w, and the steady-state threshold is θ; define the number of windows M as: M = N-w+1; define the result matrix I as a vector of length M, where each element I i Indicates whether the i-th window is steady-state data; for each window i (from 1 to M), the calculation process is as follows: d i ={data j |i≤j≤i+w-1} Among them, d i is the extracted window data, μ i , σ i is the mean and standard deviation of the window data; if Then I i =1, which means it is judged as steady-state data.

4. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1 is characterized in that: In step S2, the outlier elimination method is a mean interpolation algorithm, a regression interpolation algorithm, a multiple interpolation algorithm or an isolation forest algorithm.

5. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1, characterized in that: In step S2, the input parameters and output parameters of the prediction model specifically include: the input parameters of the prediction model must be able to reflect the changes in the outside air temperature and atmospheric pressure, and include at least one parameter related to the operating conditions; the output parameters of the prediction model are the gas turbine measurement point parameters that need to be monitored for operating status.

6. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1, characterized in that: In step S2, the sliding window method is: The monitoring parameters that can characterize the changes in the external environment and the operating conditions of the gas turbine after data processing are used as input parameters of the prediction model and input into the prediction model; the expression form of the input parameters is shown in the formula; In the formula, x i t It represents the value of the i-th thermal parameter collected at the current time t, k represents the window size, and through the sliding window, data of i*k dimensions are input each time.

7. The method for regression prediction of gas turbine dynamic parameters based on fusion network according to claim 1, characterized in that: In step S2, the training set and the data set are divided into 90% of the data as the training set and 10% of the data as the test set according to different working conditions.

8. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1, characterized in that: In step S3, the specific calculation process of the convolutional neural network CNN is as follows: Assuming that the model input sequence is X, T represents the length of the input sequence, the calculation of local features through the convolution layer is as follows: Among them, X is the input feature map after data tiling, the size is H*W*C, H and W are the height and width of the input feature map, C is the number of channels of the input feature map; W is the convolution kernel, M and N are the height and width of the convolution kernel, K is the number of channels of the output feature map; Y is the output feature map, the size is H'*W'*K'; B k is the bias term; The purpose of the pooling layer is to downsample the feature map, reduce the amount of calculation and suppress overfitting; the selected maximum pooling calculation is as follows: Where: X: input feature map, Y: output feature map, M, N: height and width of the pooling window. The pooling layer reduces the size of the feature map by selecting the maximum value or average value in the pooling window. The fully connected layer flattens the input feature map into a vector, multiplies it with the weight matrix, adds the bias term, and obtains the output: Y=W·X+b Where: X: input vector, W: weight matrix, b: bias term, Y: output vector, the fully connected layer is used to map the extracted features to the output space; In step S3, the specific calculation processes of the self-attention mechanism and the multi-head attention mechanism of the Transformer are as follows: after the convolutional neural network extracts the features, the self-attention mechanism captures the global dependency in the sequence by calculating the attention weights between the elements in the features; Where: Q: Query Matrix K: Key Matrix V: Value Matrix k : Dimension of the key The multi-head self-attention mechanism concatenates the outputs of multiple self-attention layers: MultiHead(Q,K,V)=Concat(head1,...,head n )W O in: head i =Attention(QW i Q ,KW i K ,VW i v ) Q=XW Q +b Q K=XW K +b K V=XW V +b V Weight matrices for different heads; b Q ,b K ,b V : bias term; W O : Linear transformation weight matrix after multi-head attention; In step S3, the gating mechanism of LSTM, i.e., the forget gate, input gate and output gate are calculated as follows: t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) C t =tanh(W C ·[h t-1 ,x t ]+b C ) C t =f t ·C t-1 +i t ·C t h t =o t ·tanh(C t ) Where: x t : input vector; h t-1 : The hidden state of the previous moment; C t : cell state; f t ,i t ,o t : The activation values ​​of the forget gate, input gate, and output gate; W f ,W i ,W C ,W o : weight matrix; b f ,b i ,b C ,b o : Bias term.

9. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1, characterized in that: In step S3, the model parameters of transformer and LSTM are optimized by Bayesian optimization algorithm, and the model parameters are optimized and solved by calculating the root mean square error between the predicted value and the true value to obtain the best hyperparameter combination with the minimum error. The objective function is calculated as follows: Where N is the number of samples in the training set, yi is the true value of the i-th sample, and yi-hat is the predicted value of the i-th sample.

10. The method for gas turbine dynamic parameter regression prediction based on fusion network according to claim 1, characterized in that: In step S3, the predicted results of the test set are compared with the actual values ​​and the error is calculated and analyzed, and the average relative error is used to measure the identification results: Among them, Act represents the actual value, Pred represents the predicted value, and n is the number of samples; when the test prediction results meet the requirements, the parameters of the model are saved.

Images