Intelligent modeling method for heavy-duty gas turbine based on xLSTM and Transformer network
Through intelligent modeling methods based on xLSTM and Transformer networks, combined with multimodal data and physical information constraints, the lack of dynamic performance prediction and complex operating conditions in heavy-duty gas turbine modeling is solved, and the modeling effect of high precision and high reliability is achieved.
Patent Information
- Application Number
- CN202510094086.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-08-22
- Estimated Expiration
- 2045-01-21
AI Technical Summary
The existing heavy-duty gas turbine modeling methods have shortcomings in dynamic performance prediction and complex operating conditions adaptability. Traditional data-driven models are difficult to capture the highly coupled nonlinear dynamic characteristics inside the gas turbine. The physical drive model has high computational complexity and is difficult to achieve both real-time and accuracy. Single-modal data modeling cannot effectively integrate multi-source data, resulting in limited model robustness and accuracy.
Using intelligent modeling methods based on xLSTM and Transformer networks, a coupled system model is constructed, combined with multimodal data, trained through xLSTM+Transformer data-driven model, and a loss function of physical information constraints is introduced to optimize the modeling process.
It improves the reliability and accuracy of the modeling method, enhances the generalization ability of the model, and is suitable for dynamic performance prediction and optimization requirements under complex operating conditions.
Smart Images
Figure CN120012580B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of the electric power industry, and in particular to an intelligent modeling method for a heavy-duty gas turbine based on an xLSTM and a Transformer network. Background Art
[0002] Heavy-duty gas turbines are currently the most efficient heat-to-power conversion power generation equipment. They are widely regarded as an important indicator of a country's heavy industry level and have a direct impact on national energy development. With the deepening of Industry 4.0 and digital transformation, the demand for intelligent and automated gas turbine operation has become increasingly urgent. However, due to the highly complex thermodynamic and mechanical characteristics of heavy-duty gas turbines, coupled with the complex coupling relationships between components and diverse operating conditions, existing heavy-duty gas turbine modeling methods have significant deficiencies in dynamic performance prediction and adaptability to complex operating conditions: traditional data-driven models are difficult to fully capture the highly coupled nonlinear dynamic characteristics within the gas turbine, and physical-driven models are difficult to achieve both real-time and accuracy due to their high parameterization and computational complexity. In addition, the modeling method of single modal data cannot effectively integrate multi-source data, resulting in limited robustness and accuracy of the model.
[0003] Based on this, the present invention proposes an intelligent modeling method for heavy-duty gas turbines based on xLSTM and Transformer networks on the basis of analyzing the dynamic characteristics of heavy-duty gas turbines, so as to achieve high-precision modeling of the dynamic performance of gas turbines under complex working conditions. Summary of the Invention
[0004] The purpose of the present invention is to provide an intelligent modeling method for heavy-duty gas turbines based on xLSTM and Transformer networks to solve the problems in the background technology.
[0005] To achieve the above objectives, the present invention provides a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network, comprising the following steps:
[0006] S1. Based on the dynamic characteristics of the steam-water process during the operation of the heavy-duty gas turbine, a coupled system model structure is constructed to determine the input and output variables;
[0007] S2: Combining multimodal data from heavy-duty gas turbines, and based on the relationship between rotor speed and turbine power, we construct an xLSTM+Transformer data-driven model with physical information constraints.
[0008] S3. Expand the input sequence dimension of the model constructed in S2 to 2, and the output sequence dimension to 4. Use the historical operating data of heavy-duty gas turbines to train the xLSTM+Transformer data-driven model with physical information constraints to complete the modeling.
[0009] Preferably, in said S1, the input variables are the intake air volume and intake guide vane opening during the power generation process of the heavy-duty gas turbine, and the output variables are the compressor outlet pressure, rotor speed, turbine power and exhaust temperature.
[0010] Preferably, the step S2 specifically includes the following steps:
[0011] S21. Obtain multimodal time series data from a heavy-duty gas turbine, perform data preprocessing, and generate an input sequence matrix and corresponding output labels;
[0012] S22. Input the result of S21 into the Transformer model containing the xLSTM module encoder for training. During the training process, introduce physical information constraints into the loss function to construct an xLSTM+Transformer data-driven model with physical information constraints.
[0013] Preferably, in said S21, the data preprocessing process is to obtain multimodal time series data in the heavy-duty gas turbine, use linear interpolation to fill in the missing time series data, and then perform standardization processing on the input variables and output variables;
[0014] The process of standardization is expressed as:
[0015] ;
[0016] in, is the mean, is the standard deviation;
[0017] Then, according to the dynamic characteristics of the gas turbine, the sliding window length and step size are determined, and the sliding window operation is performed on the original time series data to generate the input sequence matrix and the corresponding output labels.
[0018] Preferably, in S22, the core framework of the Transformer is built on the encoder-decoder architecture, and consists of an embedding layer, position encoding, and a multi-head attention mechanism. The processing process of the Transformer model is as follows:
[0019] 1) An embedding layer is used to convert the input sequence matrix obtained by S21 into a fixed-dimensional vector representation, representing the relationship between these features in a high-dimensional space; positional encoding is used to understand the sequential relationship in the input sequence;
[0020] ;
[0021] ;
[0022] in, represents the position in the sequence, is the dimension index;
[0023] 2) A multi-head attention mechanism is used to map the input vector into multiple different attention spaces. The results of different attention heads are calculated in parallel to capture different feature patterns in the input vector. The outputs of all attention heads are concatenated and linearly transformed to obtain the final output sequence of the multi-head attention.
[0024] Preferably, step 2) of the processing of the Transformer model is specifically as follows:
[0025] ① Map each input vector to query, key, and value spaces through linear transformation, and obtain :
[0026] , , ;
[0027] in, 、 、 is a trainable weight matrix;
[0028] Each attention head calculates the dot product of the query and key to measure the relevance, and obtains the normalized weight through scaling and Softmax function, which is expressed as:
[0029] ;
[0030] in, For query, is the key, For the value, is the dimension of the key vector; is a scaling factor used to prevent the gradient from disappearing due to excessive values;
[0031] ②Multi-head attention generates There are different attention heads, each of which captures different feature patterns in the input, represented as:
[0032] ;
[0033] in, For the The projection matrix of the head;
[0034] ③ After the outputs of all attention heads are concatenated and subjected to a linear transformation, the final output of the multi-head attention is obtained, which is expressed as:
[0035] ;
[0036] in, is the output weight matrix; the data is encoded through the embedding layer, normalized and then enters the decoder, and finally denormalized to generate the output sequence.
[0037] Preferably, the final output sequence of the multi-head attention in the Transformer model is input into the xLSTM module for processing, and the specific steps are as follows:
[0038] 1) Map the final output sequence of the multi-head attention to the hidden state space. Specifically, perform a linear transformation on the final output sequence matrix and map it to the query vector, key vector, and value vector through the embedding weight matrix and bias term. The calculation process is expressed as:
[0039] ;
[0040] ;
[0041] ;
[0042] in, is the query vector, is the key vector, is a value vector; are the embedding weight matrices, are the bias terms, is the scaling factor of the feature dimension, is the final output sequence;
[0043] 2) Dynamically update the memory state and control the flow of information through the gating mechanism (input gate, forget gate). Specifically, calculate the output of the forget gate and input gate, and combine the forget gate and input gate to update the memory state. The calculation process is expressed as:
[0044]
[0045] in, represents the memory state of the current time step, is the memory state of the previous time step; is the output of the forget gate, which controls the proportion of forgetting the memory of the previous moment. ; Is the output of the input gate, controlling the current input The ratio of information written, Represents matrix transpose; 、 Used to indicate the importance of the current input and how well it matches the memory state;
[0046] 3) Generate hidden state based on the current memory state and output gate. The calculation process is expressed as:
[0047] ;
[0048] in, is hidden state, is the element-wise product, is the normalized state.
[0049] Preferably, the input gate is expressed as:
[0050] ;
[0051] in, is the transpose of the weight matrix, is the bias term, both obtained through training, is an exponential function, ensuring that the output of the input gate is non-negative;
[0052] The forget gate is expressed as:
[0053] ;
[0054] in, is the weight matrix, is the bias term, both obtained through training, is the activation function; represents the transpose of the weight matrix;
[0055] Normalized state Expressed as:
[0056] ;
[0057] in, is the time step The standardized state at this time is expressed as:
[0058] ;
[0059] The output gate is represented as:
[0060] ;
[0061] in, is the weight vector, is the bias term;
[0062] The results processed by the xLSTM module are subsequently input into the "Add & Norm" layer. Combined with the context-dependent attention mechanism in the xLSTM and Transformer encoders, this enriches the feature information and enables the model to better focus on important input features.
[0063] Preferably, in S22, the specific steps of introducing physical information constraints into the loss function are:
[0064] 1) The physical relationship between rotor speed and gas turbine power is introduced into the loss function, and the total loss function is defined as:
[0065] ;
[0066] in, is the physical loss function, expressed as:
[0067] ;
[0068] in, represents the number of data samples, Indicates the samples, is the rotor speed, is the engine power, is the proportionality coefficient, is the power index;
[0069] is the data loss, expressed as:
[0070] ;
[0071] That is, the error between the predicted value and the actual value of the four output quantities;
[0072] in, P is the engine power, is the rotor speed, is the outlet pressure, T Exhaust temperature;
[0073] 2) Initialize the weight coefficient to 1. At the beginning of each training, calculate the current loss ratio and update the weight coefficient.
[0074] The loss ratio is expressed as:
[0075] ;
[0076] The updated weight coefficient is expressed as:
[0077] ;
[0078] in, 、 For the The loss value of the iteration, is the current weight coefficient, is the initial loss ratio, To adjust the speed coefficient, take 0.5;
[0079] 3) Determine whether the end condition is met. If so, output the result. If not, update the weight coefficient through return propagation.
[0080] In the weight coefficient update process, in order to prevent the weight coefficient from being too large or too small, the weight coefficient is regularized and expressed as:
[0081] ;
[0082] The updated weight coefficient is expressed as:
[0083] ;
[0084] in, 、 is the gradient magnitude of the loss term,
[0085] ;
[0086] ;
[0087] in, are model parameters.
[0088] Preferably, in S3, the air intake volume and the air intake guide vane opening during the power generation process of the heavy-duty gas turbine are used as model inputs, and the compressor outlet pressure, rotor speed, turbine power and exhaust temperature are used as model outputs. The input data is subjected to feature extraction and model calculation to output the final modeling result.
[0089] Therefore, the present invention provides an intelligent modeling method for heavy-duty gas turbines based on xLSTM and Transformer networks, which has the following beneficial effects:
[0090] (1) In the present invention, by introducing the physical relationship between rotor speed and gas turbine power into the loss function, a physical constraint is applied to the heavy-duty gas turbine model to eliminate unreasonable outputs and improve the reliability of the modeling method. In addition, the model is forced to follow this physical constraint during the training process. By minimizing the total loss function, the model will learn prediction results that conform to physical laws, thereby improving physical rationality and generalization ability.
[0091] (2) The present invention utilizes the high efficiency of the Transformer architecture in capturing global features and combines the advantages of xLSTM in processing local features of time series and computational efficiency. This modeling method improves the fitting accuracy of multivariable complex dynamic processes and enhances the generalization ability of the model, making it suitable for prediction and optimization needs in the actual operating environment of heavy-duty gas turbines.
[0092] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0093] Figure 1 A simplified structural diagram of a heavy-duty gas turbine operating process model according to an embodiment of the present invention;
[0094] Figure 2 This is a schematic diagram of the xLSTM+Transformer data-driven model structure in an embodiment of the present invention;
[0095] Figure 3 Schematic diagram of the model training process with physical information constraints in an embodiment of the present invention. DETAILED DESCRIPTION
[0096] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.
[0097] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.
[0098] Example
[0099] The present invention provides a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network, comprising the following steps:
[0100] S1, such as Figure 1 As shown in the figure, according to the dynamic characteristics of the steam-water process during the operation of the heavy-duty gas turbine, the coupling system model structure is constructed to determine the input variables and output variables; the input variables are the intake volume and intake guide vane opening during the power generation process of the heavy-duty gas turbine, and the output variables are the compressor outlet pressure, rotor speed, turbine power and exhaust temperature.
[0101] S2. Combining the multimodal data of heavy-duty gas turbines, and based on the relationship between rotor speed and turbine power, we construct an xLSTM+Transformer data-driven model with physical information constraints. Specifically:
[0102] S21, such as Figure 2 、3 As shown in the figure, multi-modal time series data of a heavy-duty gas turbine is obtained, the intake air volume and intake guide vane opening are used as the input sequence of the model, and the compressor outlet pressure, rotor speed, turbine power, and exhaust temperature are used as the output sequence of the model. The data is preprocessed to generate the input sequence matrix and the corresponding output labels.
[0103] The data preprocessing process involves obtaining multimodal time series data from heavy-duty gas turbines, including sensor data and operating conditions, using linear interpolation to complete missing time series data, and then standardizing the input and output variables.
[0104] The process of standardization is expressed as:
[0105] ;
[0106] in, is the mean, is the standard deviation;
[0107] Then, according to the dynamic characteristics of the gas turbine, the sliding window length and step size are determined, and the sliding window operation is performed on the original time series data to generate the input sequence matrix and the corresponding output labels.
[0108] S22. Input the results of S21 into the Transformer model containing the xLSTM module encoder for training. During the training process, introduce physical information constraints into the loss function to build an xLSTM+Transformer data-driven model with physical information constraints. The specific steps are as follows:
[0109] The core framework of the Transformer is built on an encoder-decoder architecture and consists of an embedding layer, positional encoding, and a multi-head attention mechanism. The processing of the Transformer model is as follows:
[0110] 1) An embedding layer is used to convert the input sequence matrix obtained by S21 into a fixed-dimensional vector representation, representing the relationship between these features in a high-dimensional space; positional encoding is used to understand the sequential relationship in the input sequence;
[0111] ;
[0112] ;
[0113] in, represents the position in the sequence, is the dimension index;
[0114] 2) A multi-head attention mechanism is used to map the input vector into multiple different attention spaces. The results of different attention heads are calculated in parallel to capture different feature patterns in the input vector. The outputs of all attention heads are concatenated and linearly transformed to obtain the final output sequence of the multi-head attention, which is:
[0115] ① Map each input vector to query, key, and value spaces through linear transformation, and obtain :
[0116] , , ;
[0117] in, 、 、 is a trainable weight matrix;
[0118] Each attention head calculates the dot product of the query and key to measure the relevance, and obtains the normalized weight through scaling and Softmax function, which is expressed as:
[0119] ;
[0120] in, For query, is the key, For the value, is the dimension of the key vector; is a scaling factor used to prevent the gradient from disappearing due to excessive values;
[0121] ②Multi-head attention generates There are different attention heads, each of which captures different feature patterns in the input, represented as:
[0122] ;
[0123] in, For the The projection matrix of the head;
[0124] ③ After the outputs of all attention heads are concatenated and subjected to a linear transformation, the final output sequence of the multi-head attention is obtained, which is expressed as:
[0125] ;
[0126] in, is the output weight matrix; the data is encoded through the embedding layer, normalized and then enters the decoder, and finally denormalized to generate the output sequence.
[0127] The final output sequence of the multi-head attention in the Transformer model is input into the xLSTM module for processing. xLSTM significantly improves the storage and computing capabilities of traditional LSTM through matrix storage and covariance optimization. The steps are as follows:
[0128] 1) Map the final output sequence of the multi-head attention to the hidden state space. Specifically, the input sequence matrix is linearly transformed and mapped to the query vector, key vector, and value vector through the embedding weight matrix and bias term. The calculation process is expressed as:
[0129] ;
[0130] ;
[0131] ;
[0132] in, is the query vector, is the key vector, is a value vector; are the embedding weight matrices, are the bias terms, is the scaling factor of the feature dimension, is the final output sequence;
[0133] 2) Dynamically update the memory state and control the flow of information through the gating mechanism (input gate, forget gate). Specifically, calculate the output of the forget gate and input gate, and combine the forget gate and input gate to update the memory state. The calculation process is expressed as:
[0134]
[0135] in, represents the memory state of the current time step, is the memory state of the previous time step; is the output of the forget gate, which controls the proportion of forgetting the memory of the previous moment. ; Is the output of the input gate, controlling the current input The ratio of information written, Represents matrix transpose, which is used to calculate the matching degree between key vector and value vector, so as to dynamically adjust the update of memory state; 、 Used to indicate the importance of the current input and how well it matches the memory state;
[0136] The input gate is represented as:
[0137] ;
[0138] in, is the weight matrix, is the bias term, both obtained through training, is an exponential function, ensuring that the output of the input gate is non-negative;
[0139] The forget gate is expressed as:
[0140] ;
[0141] in, is the weight matrix, is the bias term, both obtained through training, is the activation function;
[0142] 3) Generate hidden state based on the current memory state and output gate. The calculation process is expressed as:
[0143] ;
[0144] in, is hidden state, is the element-wise product, is the normalized state;
[0145] Normalized state Expressed as:
[0146] ;
[0147] in, is the time step The standardized state at this time is expressed as:
[0148] ;
[0149] The output gate is represented as:
[0150] ;
[0151] in, is the weight vector, is the bias term;
[0152] The results of the xLSTM module processing are then fed into the "Add & Norm" layer. After passing through the dense layer and the Dropout layer, they are fused with the output of the multi-head attention layer to produce the encoder output. Combining the context-dependent attention mechanism in the xLSTM and Transformer encoders enriches feature information, enabling the model to better focus on important input features.
[0153] like Figure 2As shown in the figure, after normalization, the input data flows simultaneously to the encoder stack and the decoder input module. In the encoder stack, the data passes through multiple encoder layers for feature extraction, while in the decoder input module, the data is vectorized through an embedding layer. The embedded data is then fused with the encoder output to form the final decoder input. The decoder output is then transformed and mapped through a transformation layer and a linear layer. Finally, a denormalization operation is performed to restore the data to its original scale, generating the final output sequence.
[0154] The specific steps of introducing physical information constraints into the loss function are:
[0155] 1) Change the rotor speed and gas turbine power The physical relationship between introduces the loss function, and the physical relationship is:
[0156] ;
[0157] in, is the proportionality coefficient, reflecting the mechanical characteristics, is the power index, which is fitted based on historical data, and the logarithm and linear relationship are obtained as follows:
[0158] ;
[0159] Fitting using linear regression and relationship, get and , the slope is , the intercept is ;
[0160] The above physical relationship is converted into a physical constraint function, which can be expressed as:
[0161] ;
[0162] In order to introduce physical constraints into the loss function, the physical constraint loss is defined as the square of the physical relationship residual, expressed as:
[0163] ;
[0164] Where N represents the number of data samples, i Indicates the i samples;
[0165] The total loss function consists of data loss and physical constraint loss, and the total loss function is defined as:
[0166] ;
[0167] in, is the physical loss function; is the weight coefficient used to balance the importance of data loss and physical loss, It is the data loss, which measures the difference between the model prediction value and the true value, expressed as:
[0168] ;
[0169] That is, the error between the predicted value and the actual value of the four output quantities;
[0170] in, P is the engine power, is the rotor speed, is the outlet pressure, T Exhaust temperature;
[0171] 2) Initialize the weight coefficient to 1. At the beginning of each training, calculate the current loss ratio and update the weight coefficient.
[0172] The loss ratio is expressed as:
[0173] ;
[0174] The updated weight coefficient is expressed as:
[0175] ;
[0176] in, 、 For the The loss value of the iteration, is the current weight coefficient, is the initial loss ratio, To adjust the speed coefficient, take 0.5;
[0177] 3) Determine whether the end condition is met. If so, output the result. If not, update the weight coefficient through return propagation.
[0178] In the process of updating the weight coefficient, in order to prevent the weight coefficient from being too large or too small, Apply a regularization term, expressed as:
[0179] ;
[0180] The updated weight coefficient is expressed as:
[0181] ;
[0182] in, 、 is the gradient amplitude of the loss term, balancing the training process, and λ is initialized to 1;
[0183] ;
[0184] ;
[0185] in, are model parameters;
[0186] In this step, by introducing the physical relationship between rotor speed and turbine power into the loss function, the model is forced to follow this physical constraint during training, allowing it to learn predictions that conform to physical laws. The collected data is then fed into the model for training. Through total loss minimization and a backpropagation algorithm, an adaptive weight coefficient λ is calculated based on a gradient balancing strategy, dynamically balancing the effects of the physical constraint loss and data-driven loss in model training. When the training ends, the iteration is terminated, denormalization is performed, and the model predictions are output.
[0187] S3. Expand the input sequence dimension of the model constructed in S2 to 2, and the output sequence dimension to 4. Use the intake volume and intake guide vane opening during the heavy-duty gas turbine power generation process as model input, and the compressor outlet pressure, rotor speed, turbine power, and exhaust temperature as model output. Use the historical operating data of the heavy-duty gas turbine to train the xLSTM+Transformer data-driven model with physical information constraints to complete the modeling.
[0188] In this example, 20,000 and 12,000 sets of historical operating data were collected from the actual peak-shaving process of a 300MW F-class heavy-duty gas turbine. The data included intake volume, intake guide vane opening, compressor outlet pressure, rotor speed, turbine power, and exhaust temperature. The above model was trained and verified, and the model prediction value was output and compared with the collected data.
[0189] Therefore, the present invention proposes an intelligent modeling method for heavy-duty gas turbines based on xLSTM and Transformer networks. This method utilizes the high efficiency of the Transformer architecture in capturing global features and combines the advantages of xLSTM in processing local features of time series and computational efficiency. This modeling method improves the fitting accuracy of multivariable complex dynamic processes and enhances the generalization ability of the model, making it suitable for the prediction and optimization needs in the actual operating environment of heavy-duty gas turbines.
[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network, characterized in that: The following steps are involved: S1. Based on the dynamic characteristics of the steam-water process during the operation of the heavy-duty gas turbine, a coupled system model structure is constructed to determine the input and output variables; S2. Combining the multimodal data of heavy-duty gas turbines, and based on the relationship between rotor speed and turbine power, an xLSTM+Transformer data-driven model with physical information constraints is constructed. This specifically includes the following steps: S21. Obtain multimodal time series data from a heavy-duty gas turbine, perform data preprocessing, and generate an input sequence matrix and corresponding output labels; S22. Input the result of S21 into the Transformer model containing the xLSTM module encoder for training. During the training process, physical information constraints are introduced into the loss function to build an xLSTM+Transformer data-driven model with physical information constraints. The processing of the Transformer model is as follows: 1) Use an embedding layer to convert the input sequence matrix obtained by S21 into a fixed-dimensional vector representation; use positional encoding to understand the order relationship in the input sequence; 2) A multi-head attention mechanism is used to map the input vector into multiple different attention spaces. The results of different attention heads are calculated in parallel to capture different feature patterns in the input vector. The outputs of all attention heads are concatenated and linearly transformed to obtain the final output sequence of the multi-head attention, which is: ① Map each input vector to the query, key, and value space through linear transformation. Each attention head calculates the dot product of the query and key to measure the relevance, and obtains the normalized weight through scaling and Softmax function. The formula is expressed as: ； in, For query, is the key, For the value, is the dimension of the key vector; ②Multi-head attention generates There are different attention heads, each of which captures different feature patterns in the input, represented as: ； in, For the The projection matrix of the head; ③ After the outputs of all attention heads are concatenated and subjected to a linear transformation, the final output sequence of the multi-head attention is obtained, which is expressed as: ； in, is the output weight matrix; The final output sequence of the multi-head attention in the Transformer model is input into the xLSTM module for processing. The specific steps are as follows: 1) Map the final output sequence of the multi-head attention to the hidden state space. Specifically, perform a linear transformation on the final output sequence matrix and map it to the query vector, key vector, and value vector through the embedding weight matrix and bias term. The calculation process is expressed as: ； ； ； in, is the query vector, is the key vector, is a value vector; are the embedding weight matrices, are the bias terms, is the scaling factor of the feature dimension, is the final output sequence; 2) Dynamically update the memory state and control the flow of information through the gating mechanism. Specifically, calculate the output of the forget gate and input gate, and combine the forget gate and input gate to update the memory state. The calculation process is expressed as: in, represents the memory state of the current time step, is the memory state of the previous time step; is the output of the forget gate, which controls the proportion of forgetting the memory of the previous moment. ; Is the output of the input gate, controlling the current input The ratio of information written; 3) Generate hidden state based on the current memory state and output gate. The calculation process is expressed as: ； in, is hidden state, is the element-wise product, is the normalized state; S3. Expand the input sequence dimension of the model constructed in S2 to 2, and the output sequence dimension to 4. Use the historical operating data of heavy-duty gas turbines to train the xLSTM+Transformer data-driven model with physical information constraints to complete the modeling.
2. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 1 is characterized in that: In S1, the input variables are the intake air volume and intake guide vane opening during the power generation process of the heavy-duty gas turbine, and the output variables are the compressor outlet pressure, rotor speed, turbine power and exhaust temperature.
3. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 2 is characterized in that: In S21, the data preprocessing process is to obtain multimodal time series data in the heavy-duty gas turbine, use linear interpolation to fill in the missing time series data, and then standardize the input variables and output variables; then, according to the dynamic characteristics of the gas turbine, determine the sliding window length and step size, perform a sliding window operation on the original time series data, and generate an input sequence matrix and corresponding output labels.
4. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 3 is characterized in that: The input gate is represented as: ； in, is the transpose of the weight matrix, is the bias term, is an exponential function, ensuring that the output of the input gate is non-negative; The forget gate is expressed as: ； in, is the transpose of the weight matrix, is the bias term, is the activation function; Normalized state Expressed as: ； in, is the time step The standardized state at this time is expressed as: ； The output gate is represented as: ； in, is the weight vector, is the bias term.
5. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 4 is characterized in that: In S22, the specific steps of introducing physical information constraints into the loss function are: 1) The physical relationship between rotor speed and gas turbine power is introduced into the loss function, and the total loss function is defined as: ； in, is the physical loss function, expressed as: ； in, represents the number of data samples, Indicates the samples, is the rotor speed, is the engine power, is the proportionality coefficient, is the power index; is the data loss, expressed as: ； in, is the outlet pressure, is the exhaust temperature; 2) Initialize the weight coefficient to 1. At the beginning of each training, calculate the current loss ratio and update the weight coefficient. The loss ratio is expressed as: ； The updated weight coefficient is expressed as: ； in, 、 For the The loss value of the iteration, is the current weight coefficient, is the loss ratio, is the initial loss ratio, To adjust the speed factor; 3) Determine whether the end condition is met. If so, output the result. If not, update the weight coefficient through return propagation. In the weight coefficient update process, in order to prevent the weight coefficient from being too large or too small, the weight coefficient is regularized and expressed as: ； The updated weight coefficient is expressed as: ； in, 、 is the gradient magnitude of the loss term, ； ； in, are model parameters.
6. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 1 is characterized in that: In S3, the air intake volume and the opening of the air intake guide vanes during the power generation process of the heavy-duty gas turbine are used as model inputs, and the compressor outlet pressure, rotor speed, turbine power and exhaust temperature are used as model outputs. The input data is subjected to feature extraction and model calculation, and the final modeling result is output.