Heavy-duty gas turbine intelligent modeling method based on xLSTM and Transform network

Abstract

The invention discloses a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transform network, and belongs to the technical field of electric power industry, and the method comprises the following steps: S1, determining an input variable and an output variable according to the dynamic characteristics of a steam-water flow in the operation process of a heavy-duty gas turbine; s2, combining the multi-modal data of the heavy duty gas turbine, and constructing an xLSTM + Transform data driving model with physical information constraint according to the relationship between the rotor speed and the gas turbine power; and S3, training the model by adopting historical operation data of the heavy duty gas turbine to finish modeling. According to the method, the efficient capability of a Transform architecture in global feature capture is utilized, the advantages of xLSTM in the aspects of time sequence local feature processing and calculation efficiency are combined, and physical information constraints are introduced into a loss function of model training, so that the modeling method improves the fitting precision of a multivariable complex dynamic process, enhances the generalization capability of the model, and improves the modeling efficiency. The method is suitable for prediction and optimization requirements in the actual operation environment of the heavy-duty gas turbine.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power industry, and in particular to an intelligent modeling method for a heavy-duty gas turbine based on xLSTM and Transformer networks. 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 in-depth advancement 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, as well as the complex coupling relationships between components and the diverse operating conditions, the 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 of gas turbines, and physical drive 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 a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network to solve the problems in the background technology.

[0005] To achieve the above object, the present invention provides a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network, comprising the following steps:

[0006] S1. 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;

[0007] S2. Combining the multimodal data of heavy-duty gas turbines, an xLSTM+Transformer data-driven model with physical information constraints is constructed based on the relationship between rotor speed and gas turbine power.

[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 S1, the input variables are the air intake volume and the air intake guide vane opening during the power generation process of the heavy-duty gas turbine, and the output variables are the outlet pressure of the compressor, the rotor speed, the gas turbine power and the exhaust temperature.

[0010] Preferably, the S2 specifically comprises the following steps:

[0011] S21, obtaining multimodal time series data in a heavy-duty gas turbine, performing data preprocessing, and generating 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 the physical information constraints into the loss function to construct an xLSTM+Transformer data-driven model with physical information constraints.

[0013] Preferably, in S21, the data preprocessing process is to obtain multi-modal time series data in the heavy-duty gas turbine, use linear interpolation to complete the missing time series data, and then standardize the input variables and output variables;

[0014] The process of standardization is expressed as:

[0015]

[0016] Among them, μ is the mean and σ 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 Transformer is built on the codec architecture, consisting 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 vector representation of fixed dimension, representing the relationship between these features in a high-dimensional space; position encoding is used to understand the order relationship in the input sequence;

[0020]

[0021] Among them, pos represents the position in the sequence, and i is the dimension index;

[0022] 2) A multi-head attention mechanism is used to map the input vector to multiple different attention spaces, and 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.

[0023] Preferably, step 2) of the processing process of the Transformer model is specifically as follows:

[0024] ① Map each input vector to the query, key, and value spaces through linear transformations to obtain Q, K, and V respectively:

[0025] Q=XW Q , K = XW K , V = XW V ;

[0026] Among them, W Q , W K , W V is a trainable weight matrix;

[0027] Each attention head calculates the dot product of the query and the key to measure the relevance, and obtains the normalized weight through scaling and Softmax function, which is expressed as:

[0028]

[0029] Where Q is the query, K is the key, V is the value, and d k is the dimension of the key vector; is a scaling factor used to prevent the gradient from disappearing due to excessive values;

[0030] ②Multi-head attention generates h different attention heads through multiple different linear transformations, each of which captures different feature patterns in the input, expressed as:

[0031]

[0032] in, is the projection matrix of the i-th head;

[0033] ③ After the outputs of all attention heads are concatenated, a linear transformation is performed to obtain the final output of the multi-head attention, which is expressed as:

[0034] MultHead(Q,K,V)=Concat(head1,…,head h )W o ;

[0035] Among them, W o 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.

[0036] 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:

[0037] 1) Map the final output sequence of 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:

[0038] q t =W q x t +b q ;

[0039]

[0040] v t =Wvx t +b v ;

[0041] Among them, q t is the query vector, k t is the key vector, v t is a value vector; W q , Wk, Wv are embedding weight matrices respectively, b q , bk, b v are the bias terms, d is the scaling factor of the feature dimension, x t is the final output sequence;

[0042] 2) Dynamically update the memory state and control the information flow through the gating mechanism (input gate, forget gate). Specifically, calculate the output of the forget gate and input gate, and update the memory state in combination with the forget gate and input gate. The calculation process is expressed as:

[0043] C t =f t C t-1 +i t v t k t T

[0044] Among them, C t represents the memory state of the current time step, C t-1 is the memory state of the previous time step; f t is the output of the forget gate, which controls the proportion of the memory at the previous moment to be forgotten, f t ∈[0,1]; it Is the output of the input gate, controlling the current input x t The ratio of information written, T represents the matrix transpose; k t 、v t Used to indicate the importance of the current input and how well it matches the memory state;

[0045] 3) Generate hidden state based on the current memory state and output gate. The calculation process is expressed as:

[0046]

[0047] Among them, ht is the hidden state, ⊙ is the element-wise product, is the normalized state.

[0048] Preferably, the input gate is expressed as:

[0049] i t =exp(wi T x t +bi);

[0050] Among them, wi T is the transpose of the weight matrix, bi is the bias term, both of which are obtained through training, and exp is the exponential function to ensure that the output of the input gate is non-negative;

[0051] The forget gate is expressed as:

[0052] f t =σ(wf T x t +bf);

[0053] Among them, wf is the weight matrix, bf is the bias term, both of which are obtained through training, and σ is the activation function; wf T represents the transpose of the weight matrix;

[0054] Normalized state It is expressed as:

[0055]

[0056] Among them, n t is the normalized state at time step t, expressed as:

[0057] n t =f t n t-1 +i t k t ;

[0058] The output gate is represented as:

[0059] o t =σ(wo T x t +b o );

[0060] Among them, w o is the weight vector, b o is the bias term;

[0061] The results processed by the xLSTM module will be subsequently input into the "Add&Norm" layer, combined with the context-dependent attention mechanism in the xLSTM and Transformer encoders, enriching the feature information and enabling the model to better focus on important input features.

[0062] Preferably, in S22, the specific steps of introducing physical information constraints into the loss function are:

[0063] 1) The physical relationship between rotor speed and engine power is introduced into the loss function, and the total loss function is defined as:

[0064] L total =L data +λ·L physics ;

[0065] Among them, L physics is the physical loss function, expressed as:

[0066]

[0067] Wherein, N represents the number of data samples, i represents the i-th sample, n represents the rotor speed, P represents the engine power, H represents the proportional coefficient, and α represents the power index;

[0068] L data is the data loss, expressed as:

[0069]

[0070] That is, the error between the predicted value and the actual value of the four output quantities;

[0071] Among them, P is the engine power, n is the rotor speed, Pe is the outlet pressure, and T is the exhaust temperature;

[0072] 2) Initialize the weight coefficient to 1. At the beginning of each training, calculate the current loss ratio and update the weight coefficient;

[0073] The loss ratio is expressed as:

[0074]

[0075] The updated weight coefficient is expressed as:

[0076]

[0077] Among them, L data，t , L physics，t is the loss value of the tth iteration, λt is the current weight coefficient, r0 is the initial loss ratio, β is the adjustment speed coefficient, which is 0.5;

[0078] 3) Determine whether the end condition is met, if so, output the result, if not, update the weight coefficient by return propagation;

[0079] In the process of updating the weight coefficient, in order to prevent the weight coefficient from being too large or too small, the weight coefficient is regularized, expressed as:

[0080] L total =L data +e-λ·L physics +λ;

[0081] The updated weight coefficient is expressed as:

[0082]

[0083] Among them, Gdate, G physics is the gradient magnitude of the loss term,

[0084]

[0085] Among them, θ is the model parameter.

[0086] Preferably, in S3, the air intake volume and the opening of the air intake guide vane during the power generation process of the heavy-duty gas turbine are used as model inputs, and the outlet pressure of the compressor, the rotor speed, the engine power and the exhaust temperature are used as model outputs. The input data is subjected to feature extraction and model calculation to output the final modeling result.

[0087] Therefore, the heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network of the present invention has the following beneficial effects:

[0088] (1) In the present invention, by introducing the physical relationship between the rotor speed and the gas turbine power into the loss function, a physical constraint is applied to the heavy-duty gas turbine model to eliminate unreasonable outputs, thereby improving 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 the prediction results that conform to the physical laws, thereby improving the physical rationality and generalization ability.

[0089] (2) The present invention utilizes the high efficiency of the Transformer architecture in capturing global features and combines the advantages of xLSTM in local feature processing and computational efficiency of time series. This modeling method improves the fitting accuracy of multivariable complex dynamic processes and enhances the generalization ability of the model, which is suitable for the prediction and optimization needs in the actual operating environment of heavy-duty gas turbines.

[0090] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 A simplified structural diagram of a heavy-duty gas turbine operation process model according to an embodiment of the present invention;

[0092] Figure 2 This is a schematic diagram of the structure of the xLSTM+Transformer data-driven model according to an embodiment of the present invention;

[0093] Figure 3 Schematic diagram of a model training process with physical information constraints according to an embodiment of the present invention. DETAILED DESCRIPTION

[0094] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0095] 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.

[0096] Example

[0097] The present invention provides a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network, comprising the following steps:

[0098] S1, such as Figure 1 As shown, 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.

[0099] S2. Combining the multimodal data of heavy-duty gas turbines, an xLSTM+Transformer data-driven model with physical information constraints is constructed based on the relationship between rotor speed and gas turbine power. Specifically:

[0100] S21, such as Figure 2 ,3 As shown in the figure, the multi-modal time series data of the heavy-duty gas turbine is obtained, the intake volume and intake guide vane opening are used as the input sequence of the model, the compressor outlet pressure, rotor speed, gas turbine power, and exhaust temperature are used as the output sequence of the model, and the data is preprocessed to generate the input sequence matrix and the corresponding output label;

[0101] The data preprocessing process is to obtain multi-modal time series data in heavy-duty gas turbines, including sensor data and operating conditions, use linear interpolation to complete the missing time series data, and then standardize the input variables and output variables;

[0102] The process of standardization is expressed as:

[0103]

[0104] Among them, μ is the mean and σ is the standard deviation;

[0105] 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.

[0106] S22. Input the result of S21 into the Transformer model containing the xLSTM module encoder for training. During the training process, introduce the physical information constraint into the loss function to build an xLSTM+Transformer data-driven model with physical information constraints. The specific steps are as follows:

[0107] The core framework of Transformer is built on the encoder-decoder architecture, which consists of an embedding layer, position encoding, and a multi-head attention mechanism. The processing process of the Transformer model is as follows:

[0108] 1) An embedding layer is used to convert the input sequence matrix obtained by S21 into a vector representation of fixed dimension, representing the relationship between these features in a high-dimensional space; position encoding is used to understand the order relationship in the input sequence;

[0109]

[0110] Among them, pos represents the position in the sequence, and i is the dimension index;

[0111] 2) The multi-head attention mechanism is used to map the input vector to multiple different attention spaces, and 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:

[0112] ① Map each input vector to the query, key, and value spaces through linear transformations to obtain Q, K, and V respectively:

[0113] Q=XWQ, K=XWK, V=XWV;

[0114] Among them, WQ, WK, and WV are trainable weight matrices;

[0115] Each attention head calculates the dot product of the query and the key to measure the relevance, and obtains the normalized weight through scaling and Softmax function, which is expressed as:

[0116]

[0117] Where Q is the query, K is the key, V is the value, and dk is the dimension of the key vector; is a scaling factor used to prevent the gradient from disappearing due to excessive values;

[0118] ②Multi-head attention generates h different attention heads through multiple different linear transformations, each of which captures different feature patterns in the input, expressed as:

[0119]

[0120] in, is the projection matrix of the i-th head;

[0121] ③ After the outputs of all attention heads are concatenated, a linear transformation is performed to obtain the final output sequence of the multi-head attention, which is expressed as:

[0122] MultiHead(Q,K,V)=Concat(head1,…,head h )W o ;

[0123] Among them, W o 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.

[0124] 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 the traditional LSTM through matrix storage and covariance optimization. The steps are as follows:

[0125] 1) Map the final output sequence of multi-head attention to the hidden state space. Specifically, perform a linear transformation on the input 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:

[0126] q t =Wq x t +b q ;

[0127]

[0128] v t =Wvx t +b v ;

[0129] Among them, q t is the query vector, k t is the key vector, v t is a value vector; W q , Wk, Wv are embedding weight matrices respectively, b q , bk, b v are the bias terms, d is the scaling factor of the feature dimension, x t is the final output sequence;

[0130] 2) Dynamically update the memory state and control the information flow through the gating mechanism (input gate, forget gate). Specifically, calculate the output of the forget gate and input gate, and update the memory state in combination with the forget gate and input gate. The calculation process is expressed as:

[0131] C t =f t C t-1 +i t v t k t T

[0132] Among them, C t represents the memory state of the current time step, C t-1 is the memory state of the previous time step; f t is the output of the forget gate, which controls the proportion of the memory at the previous moment to be forgotten, f t ∈[0,1]; i t Is the output of the input gate, controlling the current input x t The writing ratio of information, T represents the matrix transpose, which is used to calculate the matching degree between the key vector and the value vector, so as to dynamically adjust the update of the memory state; k t 、v t Used to indicate the importance of the current input and how well it matches the memory state;

[0133] The input gate is represented as:

[0134] i t =exp(w i T x t +b i );

[0135] Among them, w i is the weight matrix, b i is the bias term, all obtained through training, exp is the exponential function, which ensures that the output of the input gate is non-negative;

[0136] The forget gate is expressed as:

[0137] f t =σ(w f T x t +b f );

[0138] Among them, w f is the weight matrix, b f is the bias term, all obtained through training, and σ is the activation function;

[0139] 3) Generate hidden state based on the current memory state and output gate. The calculation process is expressed as:

[0140]

[0141] Among them, h t is the hidden state, ⊙ is the element-wise product, is the normalized state;

[0142] Normalized state It is expressed as:

[0143]

[0144] Among them, n t is the normalized state at time step t, expressed as:

[0145] n t =f t n t-1 +i t k t ;

[0146] The output gate is represented as:

[0147] o t =σ(w o T x t +b o );

[0148] Among them, w o is the weight vector, b o is the bias term;

[0149] The result of the xLSTM module processing will be subsequently input into the "Add&Norm" layer, and after passing through the dense layer and the Dropout layer, it is fused with the output of the multi-head attention layer to obtain the encoder output. Combining the context-dependent attention mechanism in the xLSTM and Transformer encoders enriches the feature information, enabling the model to better focus on important input features.

[0150] like Figure 2 As shown in the figure, after the input data is normalized, it flows to the encoder stack and the decoder input module at the same time. In the encoder stack, the data is extracted through multiple layers of encoders, while in the decoder input module, the data is vectorized through the embedding layer. Subsequently, the embedded data is fused with the encoder output to form the final decoder input. The decoder output is transformed and mapped through the conversion layer and the linear layer, and finally the original scale of the data is restored through the denormalization operation to generate the final output sequence.

[0151] The specific steps of introducing physical information constraints into the loss function are:

[0152] 1) The physical relationship between the rotor speed n and the engine power P is introduced into the loss function. The physical relationship is:

[0153] P=K·n α ;

[0154] Among them, K is the proportional coefficient, which reflects the mechanical characteristics, and α is the power index. Based on the historical data fitting, the logarithm is taken and the linear relationship is obtained:

[0155] lnP=lnK+αln n;

[0156] Use linear regression method to fit the relationship between lnP and ln n, and get lnK and α, the slope is α, and the intercept is lnK;

[0157] The above physical relationship is converted into a physical constraint function, which is expressed as:

[0158] f physics (n, P) = PK·n α =0;

[0159] 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:

[0160]

[0161] Where N represents the number of data samples, and i represents the i-th sample;

[0162] The total loss function consists of data loss and physical constraint loss, and the total loss function is defined as:

[0163] L total =L data +λ·L physics ;

[0164] Among them, L physics is the physical loss function; λ is the weight coefficient, which is used to balance the importance of data loss and physical loss. data It is the data loss, which measures the difference between the model prediction value and the true value, expressed as:

[0165]

[0166] That is, the error between the predicted value and the actual value of the four output quantities;

[0167] Among them, P is the engine power, n is the rotor speed, P e is the outlet pressure, T is the exhaust temperature;

[0168] 2) Initialize the weight coefficient to 1. At the beginning of each training, calculate the current loss ratio and update the weight coefficient;

[0169] The loss ratio is expressed as:

[0170]

[0171] The updated weight coefficient is expressed as:

[0172]

[0173] Among them, L data，t , L physics，t is the loss value of the tth iteration, λ t is the current weight coefficient, r0 is the initial loss ratio, β is the adjustment speed coefficient, which is 0.5;

[0174] 3) Determine whether the end condition is met, if so, output the result, if not, update the weight coefficient by return propagation;

[0175] In the process of updating the weight coefficient, in order to prevent the weight coefficient from being too large or too small, a regularization term can be applied to λ, which is expressed as:

[0176] L total =L data +e -λ ·L physics +λ;

[0177] The updated weight coefficient is expressed as:

[0178]

[0179] Among them, Gdate , G physics is the gradient amplitude of the loss term, balancing the training process, and λ is initialized to 1;

[0180]

[0181] Among them, θ is the model parameter;

[0182] In this step, by introducing the physical relationship between rotor speed and engine power into the loss function, the model is forced to follow this physical constraint during training, and the model will learn prediction results that conform to physical laws; the collected data is input into the model training, and the adaptive weight coefficient λ is calculated according to the gradient balance strategy through the minimization of the total loss and the back propagation algorithm, and the role of the physical constraint loss and data-driven loss in model training is dynamically balanced. When the training meets the end condition, the iteration is terminated, the denormalization operation is performed, and the model prediction value is output.

[0183] S3. Expand the input sequence dimension of the model constructed in S2 to 2, and the output sequence dimension to 4. Use the air intake volume and air intake guide vane opening during the power generation process of the heavy-duty gas turbine as the model input, and the compressor outlet pressure, rotor speed, gas turbine power and exhaust temperature as the 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.

[0184] In this embodiment, 20,000 and 12,000 groups of historical operation data of a 300MW F-class heavy-duty gas turbine in the actual peak-shaving process are collected. The data include intake volume, intake guide vane opening, compressor outlet pressure, rotor speed, gas turbine power, and exhaust temperature. The above model is trained and verified, and the model prediction value is output and compared with the collected data.

[0185] Therefore, the present invention proposes an intelligent modeling method for heavy-duty gas turbines based on xLSTM and Transformer networks, which utilizes the efficient ability of the Transformer architecture in capturing global features and combines the advantages of xLSTM in local feature processing and computational efficiency of time series. The modeling method improves the fitting accuracy of multivariable complex dynamic processes and enhances the generalization ability of the model, and is suitable for the prediction and optimization needs in the actual operating environment of heavy-duty gas turbines.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. 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 solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution 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. 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; S2. Combining the multimodal data of heavy-duty gas turbines, an xLSTM+Transformer data-driven model with physical information constraints is constructed based on the relationship between rotor speed and gas turbine power. 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. According to claim 1, a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network 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 outlet pressure of the compressor, the rotor speed, the engine power and the exhaust temperature.

3. According to claim 1, a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network is characterized in that: The S2 specifically includes the following steps: S21, obtaining multimodal time series data in a heavy-duty gas turbine, performing data preprocessing, and generating 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, introduce the physical information constraints into the loss function to construct an xLSTM+Transformer data-driven model with physical information constraints.

4. According to claim 3, a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network 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 complete 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.

5. According to claim 3, a heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network is characterized in that: In S22, the processing process of the Transformer model is as follows: 1) Use an embedding layer to convert the input sequence matrix obtained by S21 into a vector representation of fixed dimension; use position encoding to understand the order relationship in the input sequence; 2) A multi-head attention mechanism is used to map the input vector to multiple different attention spaces, and 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.

6. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 5 is characterized in that: Step 2) of the processing process of the Transformer model is specifically as follows: 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: Where Q is the query, K is the key, V is the value, and d k is the dimension of the key vector; ②Multi-head attention generates h different attention heads through multiple different linear transformations, each of which captures different feature patterns in the input, expressed as: in, is the projection matrix of the i-th head; ③ After the outputs of all attention heads are concatenated, a linear transformation is performed to obtain the final output sequence of the multi-head attention, which is expressed as: MultiHead(Q,K,V)=Concat(head1,...,head h )W o ; Among them, W o is the output weight matrix.

7. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 6 is characterized in that: 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 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: q t =W q x t +b q ; v t =W v x t +b v ; Among them, q t is the query vector, k t is the key vector, v t is a value vector; W q , W k , W v are the embedding weight matrices, b q , b k , b v are the bias terms, d is the scaling factor of the feature dimension, x t is the final output sequence; 2) Dynamically update the memory state and control the information flow through the gating mechanism. Specifically, calculate the output of the forget gate and the input gate, and combine the forget gate and the input gate to update the memory state. The calculation process is expressed as: C t =f t C t-1 +i t v t k t T Among them, C t represents the memory state of the current time step, C t-1 is the memory state of the previous time step; f t is the output of the forget gate, which controls the proportion of the memory at the previous moment to be forgotten, f t ∈[0,1];i t Is the output of the input gate, controlling the current input x t The proportion of information written; 3) Generate hidden state based on the current memory state and output gate. The calculation process is expressed as: Among them, h t is the hidden state, ⊙ is the element-wise product, is the normalized state.

8. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 7 is characterized in that: The input gate is represented as: i t =exp(w i T x t +b i ); Among them, w i T is the transpose of the weight matrix, b i is the bias term, exp is the exponential function, which ensures that the output of the input gate is non-negative; The forget gate is expressed as: f t =σ(w f T x t +b f ); Among them, w f T is the transpose of the weight matrix, b f is the bias term, σ is the activation function; Normalized state It is expressed as: Among them, n t is the normalized state at time step t, expressed as: n t =f t n t-1 +i t k t ; The output gate is represented as: the t =σ(w o T x t +b o ); Among them, w o is the weight vector, b o is the bias term.

9. The heavy-duty gas turbine intelligent modeling method based on xLSTM and Transformer network according to claim 8 is characterized in that: In S22, the specific steps of introducing the physical information constraint into the loss function are: 1) The physical relationship between rotor speed and engine power is introduced into the loss function, and the total loss function is defined as: L total =L data +λ·L physics ; Among them, L physics is the physical loss function, expressed as: Wherein, N represents the number of data samples, i represents the i-th sample, n represents the rotor speed, P represents the engine power, H represents the proportional coefficient, and α represents the power index; L data is the data loss, expressed as: Among them, P e is the outlet pressure, T 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: Among them, L data，t , L physics，t is the loss value of the tth iteration, λ t is the current weight coefficient, r t is the loss ratio, r0 is the initial loss ratio, and β is the adjustment speed coefficient; 3) Determine whether the end condition is met, if so, output the result, if not, update the weight coefficient by return propagation; In the process of updating the weight coefficient, in order to prevent the weight coefficient from being too large or too small, the weight coefficient is regularized, expressed as: L total =L data +e -λ ·L physics +λ; The updated weight coefficient is expressed as: Among them, G date , G physics is the gradient magnitude of the loss term, Among them, θ is the model parameter.

10. 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 vane during the power generation process of the heavy-duty gas turbine are used as model inputs, and the outlet pressure of the compressor, the rotor speed, the engine power and the exhaust temperature are used as model outputs. The input data is subjected to feature extraction and model calculation, and then the final modeling result is output.

Images