Intelligent modeling method for heavy-duty gas turbine based on Mamba large model and Transformer network

Abstract

The present invention discloses an intelligent modeling method for heavy-duty gas turbines based on a Mamba large model and a Transformer network, belonging to the field of power engineering technology. The method comprises the following steps: S1. constructing a coupled system model structure and determining input and output variables based on the dynamic characteristics of the steam-water flow during the heavy-duty gas turbine peak-shaving process; S2. constructing a heavy-duty gas turbine peak-shaving process model based on the Mamba+Transformer multimodal large model; and S3. training and validating the heavy-duty gas turbine peak-shaving process model using historical operating data of the heavy-duty gas turbine to complete the modeling. The present invention fully considers the large span of variable data in the heavy-duty gas turbine peak-shaving process, and the constructed model has strong generalization capabilities. The model effectively integrates the characteristics of Mamba and Transformer, can efficiently process multivariate long time series, and can quickly identify and obtain an accurate heavy-duty gas turbine peak-shaving process model for industrial control. This is beneficial for improving the unit's deep peak-shaving capability and providing support for the high-proportion consumption of new energy power.

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 a Mamba large model and a Transformer network. Background Art

[0002] The high penetration of large-scale, random wind and photovoltaic power generation into China's power grid has led to an increasingly prominent demand for peak and frequency regulation in my country's power system. Against this backdrop, the need for flexible operation of thermal power generators under all operating conditions, with rapid load shedding and transition to islanded operation for power generation, has become an inevitable development trend. Heavy-duty gas turbines are highly favored in the peak-shaving market due to their high thermal efficiency, rapid start-up and shutdown, fast load response, and high cleanliness.

[0003] However, the rapid start-up and shutdown of heavy-duty gas turbines and the optimized control of variable loads require the accurate acquisition of a multivariable model of the power generation process. Therefore, developing intelligent modeling methods for heavy-duty gas turbines and promoting their participation in peak load regulation is a practical and effective way to improve the safe and stable operation of the power grid. To promote the participation of heavy-duty gas turbines in power grid peak load regulation, it is urgent to build a modeling method that can quickly obtain an accurate model of the heavy-duty gas turbine peak load regulation process, and this model should be able to serve the design of industrial process controllers. Traditional modeling methods cannot fully consider the large span of the variable data in the heavy-duty gas turbine peak load regulation process and cannot effectively integrate multi-source data.

[0004] Based on this, the present invention analyzes the dynamic characteristics of the heavy-duty gas turbine peak-shaving process, integrates the Mamba and Transformer large models, and designs a multivariable sequence intelligent modeling method to improve the deep peak-shaving capability of heavy-duty gas turbines from the perspective of modeling. Summary of the Invention

[0005] The purpose of the present invention is to provide a heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network to solve the problems in the background technology.

[0006] To achieve the above objectives, the present invention provides a heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network, comprising the following steps:

[0007] S1. Based on the dynamic characteristics of the steam-water process during the peak regulation of heavy-duty gas turbines, a coupled system model structure is constructed to determine the input and output variables.

[0008] S2. Obtain and process the multivariable mixed data sequence of the heavy-duty gas turbine peak-shaving process, and build a heavy-duty gas turbine peak-shaving process model based on the Mamba+Transformer multimodal large model;

[0009] S3. Use the historical operating data of heavy-duty gas turbines to train and verify the heavy-duty gas turbine peak-shaving process model to complete the modeling.

[0010] 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.

[0011] Preferably, the specific steps of S2 are as follows:

[0012] S21. Obtain a multivariate mixed data sequence for a heavy-duty gas turbine peak-shaving process, perform data construction, pooling operations, and multi-layer perception processing, and obtain multi-sequence input data;

[0013] The data construction process designs a unique pre-training mode based on multiple sequences of different dimensions and time scales, enabling the model to distinguish the temporal and spatial dependencies between multiple sequences;

[0014] The pooling operation maps gradually longer input sequences into fixed-length vectors. Adding the pooling operation to a large multimodal model can compress the gradually longer multivariate input sequence into a fixed-length vector that is easier to learn in subsequent modules.

[0015] The multilayer perceptron consists of two linear transformation layers and a nonlinear activation function. The first linear transformation layer increases the dimensionality of the input sequence so that the network can learn complex tasks. The second linear transformation layer restores the sequence dimension to the same size as the input sequence. Residual connections and layer normalization ensure stable model training. The nonlinear activation function helps large multimodal models capture complex patterns in long-term multi-sequence input data.

[0016] S22: Perform Mamba module processing, Transformer module processing, and normalization processing on the results of S21 to complete the construction of the Mamba+Transformer multimodal large model;

[0017] S23. Expand the input sequence dimension of the Mamba+Transformer multimodal large model to 2, and the output sequence dimension to 4, to complete the construction of the heavy-duty gas turbine peak-shaving process model.

[0018] Preferably, in S22, the processing process of the Mamba module is as follows:

[0019] 1) Use the state space model to process multiple sequence input data, expressed as:

[0020]

[0021] yk =Cx k ;

[0022] Among them, x k is the state variable at the kth moment, x k-1 is the state variable at the k-1th moment, u k The input vector for the model, y k Output vector for the model;

[0023] C is the parameter matrix of the state space model;

[0024]

[0025] Where e is the base of the exponential function, Δ is the increment, and I is the identity matrix;

[0026] 2) Perform convolution kernel calculation on the processed multi-sequence input data as follows:

[0027]

[0028] in, is the convolution kernel, N is the number of rows, is the transpose of the parameter matrix; for The Vandermonde matrix of

[0029] is the expansion matrix;

[0030] The purpose is to generate All power forms of , the state vector is mapped to a high-dimensional feature space through power expansion for convolution calculation;

[0031] 3) Output the predicted value, expressed as:

[0032]

[0033] Among them, u is the model input vector and y is the model output vector.

[0034] Preferably, in S22, the specific process of the Transformer model processing is: using the sinusoidal position encoding mode to learn the sequence, which is expressed as:

[0035]

[0036] Among them, pos is the position of the current semantics in the sequence, i is the dimension of semantic embedding, d model is the dimension of the input sequence.

[0037] Preferably, in S22, the sequence processed by the Transformer module enters the Mamba module again for reprocessing to accelerate the learning efficiency of complex sequences, and then the output value is normalized to output the model prediction value. The normalization operation of the prediction value before output is conducive to stabilizing and standardizing the data, ensuring stable transmission of data between network layers.

[0038] Preferably, in said S23, 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 outlet pressure of the compressor, the rotor speed, the turbine power and the exhaust temperature are used as model outputs.

[0039] Preferably, in said S3, step S2 is used for training and verification, the model prediction value is output and compared with the real data in the historical operation data, and the root mean square error, modeling error, and modeling accuracy are calculated to evaluate the model.

[0040] Preferably, the calculation formula of the root mean square error is:

[0041]

[0042] The calculation formula of modeling error is:

[0043] E(k)=|y(k)-r(k)|;

[0044] The calculation formula for modeling accuracy is:

[0045]

[0046] Among them, RMSE(k) is the root mean square error, E(k) is the modeling error, R 2 is the modeling accuracy, N represents the number of arrays, r(k) is the output variable of the heavy-duty gas turbine peak-shaving process model, and y(k) is the model prediction value.

[0047] Therefore, the heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network of the present invention has the following beneficial effects:

[0048] (1) The method protected by the present invention deeply explores the deep peak-shaving capability of heavy-duty gas turbines to promote the construction of a high-proportion new energy power system, fully considers the dynamic characteristics of the steam-water process in the peak-shaving process of heavy-duty gas turbines, and combines the Mamba+Transformer multimodal large model that can quickly and effectively capture the long-term and short-term dependencies of multivariable long-term series to design a multivariable power generation process intelligent modeling method. This method fully considers the characteristics of the large span of the variable data in the peak-shaving process of heavy-duty gas turbines, and the constructed model has the advantage of strong generalization ability.

[0049] (2) The Mamba+Transformer large model designed in this method has a stable selective state space model architecture and strong generalization ability, and can efficiently process multivariate long time series; in addition, the model can quickly identify and obtain accurate heavy-duty gas turbine peak-shaving processes for industrial control, while taking into account the coupling between the various variables in the heavy-duty gas turbine peak-shaving process and accurately capturing the mapping relationship, which helps to improve the peak-shaving capability of heavy-duty gas turbines, improve the deep peak-shaving capability of the unit, and provide support for the high proportion of new energy power consumption.

[0050] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a simplified structural diagram of a heavy-duty gas turbine peak-shaving process model according to an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of the Mamba+Transformer multimodal modeling in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0054] 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.

[0055] Example

[0056] This embodiment provides a heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network. It is based on 20,000 sets and 12,000 sets of historical operating data from the actual peak-shaving process of a 300MW F-class heavy-duty gas turbine. The data includes intake air flow, intake guide vane opening, compressor outlet pressure, rotor speed, turbine power, and exhaust temperature. The method includes the following steps:

[0057] S1. As the power grid's requirements for the depth and speed of heavy-duty gas turbine peak regulation become increasingly higher, the operating load range of heavy-duty gas turbines is becoming increasingly wider. As a result, the dynamic characteristics of the unit are becoming more and more obvious. The values ​​of various variables such as air intake, air intake guide vane opening, compressor outlet pressure, gas turbine power, and exhaust temperature vary greatly. The heat resistance of the blades inside the unit and the stable and safe operation of various components face great challenges. Therefore, based on the dynamic characteristics of the steam-water process during the peak regulation process of heavy-duty gas turbines, a simplified structure of the two-input and four-output coupled system model is obtained, as shown in the following figure: Figure 1As shown, the input variables are determined to be 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;

[0058] S2. Obtain and process the multivariable mixed data sequence of the heavy-duty gas turbine peak-shaving process, and build a heavy-duty gas turbine peak-shaving process model based on the Mamba+Transformer multimodal large model. The details are as follows:

[0059] S21. Obtain a multivariate mixed data sequence for a heavy-duty gas turbine peak-shaving process, perform data construction, pooling operations, and multi-layer perception processing, and obtain multi-sequence input data;

[0060] S22. Perform Mamba module processing, Transformer module processing, and normalization processing on the results of S21 to complete the construction of the Mamba+Transformer multimodal large model; integrating Mamba and Transformer to construct the Mamba+Transformer multimodal large model has practical significance for learning large-span and long-term multivariate sequences in the peak-shaving process of heavy gas turbines.

[0061] The Mamba+Transformer multimodal model includes the Mamba and Transformer models. The sequence from top to bottom first passes through the Mamba module and then enters the Transformer module to complete the model establishment of the heavy-duty gas turbine peak-shaving process. The ratio of the Mamba module to the Transformer module in the Mamba+Transformer multimodal model is 7:1, such as Figure 2 shown.

[0062] The underlying architecture of the Mamba large model is a selective state-space model, which excels at processing long time series quickly without requiring additional storage space. The processing process of the Mamba module is as follows:

[0063] 1) Use the state space model to process multiple sequence input data, expressed as:

[0064]

[0065] y k =Cx k ;

[0066] Among them, x k is the state variable at the kth moment, x k-1 is the state variable at the k-1th moment, u k The input vector for the model, y k Output vector for the model;

[0067] C is the parameter matrix of the state space model;

[0068]

[0069]

[0070] Where e is the base of the exponential function, Δ is the increment, and I is the identity matrix;

[0071] 2) Perform convolution kernel calculation on the processed multi-sequence input data as follows:

[0072]

[0073] in, is the convolution kernel, N is the number of rows, is the transpose of the parameter matrix; for The Vandermonde matrix of

[0074] is the expansion matrix;

[0075] The purpose is to generate All power forms of , the state vector is mapped to a high-dimensional feature space through power expansion for convolution calculation;

[0076] 3) Output the predicted value, expressed as:

[0077]

[0078] Among them, u is the model input vector and y is the model output vector.

[0079] The Transformer model consists of an attention mechanism and a feedforward neural network, with excellent multi-task parallel computing performance and strong generalization ability. The specific process of the Transformer model is to learn the sequence using a sinusoidal position encoding mode, which is expressed as:

[0080]

[0081] Among them, pos is the position of the current semantics in the sequence, i is the dimension of semantic embedding, d model is the dimension of the input sequence.

[0082] The sequence processed by the Transformer module enters the Mamba module for further processing to accelerate the learning efficiency of complex sequences. The output value is then normalized and the model prediction value is output. The normalization operation of the prediction value before output is conducive to stabilizing and standardizing the data and ensuring stable transmission of data between network layers.

[0083] S23. Expand the input sequence dimension of the Mamba+Transformer multimodal large model to 2, and use the intake volume and intake guide vane opening during the heavy-duty gas turbine power generation process as model input; expand the output sequence dimension to 4, and use the compressor outlet pressure, rotor speed, turbine power and exhaust temperature as model output; complete the construction of the heavy-duty gas turbine peak-shaving process model.

[0084] S3. Use 20,000 sets and 12,000 sets of historical operating data from the actual peak-shaving process of a 300MW F-class heavy-duty gas turbine to train and verify the heavy-duty gas turbine peak-shaving process model according to step S2. Output the model prediction value and compare it with the actual value of the unit to evaluate the performance of the modeling method.

[0085] Combine all loop model prediction values ​​and unit true values ​​to calculate the root mean square error RMSE, modeling error E, and modeling accuracy R 2 And modeling time T, to comprehensively evaluate the modeling performance. Modeling time T can be obtained by adding a timer to the model, RMSE, E and R 2 The calculation of is as follows:

[0086]

[0087] E(k)=|y(k)-r(k)|;

[0088]

[0089] Where N represents the number of arrays, r(k) is the output variable of the heavy-duty gas turbine peak-shaving process model, namely the actual values ​​of compressor outlet pressure, rotor speed, turbine power, and exhaust temperature, and y(k) is the model prediction value.

[0090] Therefore, the present invention proposes an intelligent modeling method for heavy-duty gas turbines based on the Mamba large model and the Transformer network, which deeply explores the deep peak-shaving capability of heavy-duty gas turbines to promote the construction of a high-proportion new energy power system, fully considers the dynamic characteristics of the steam-water process in the peak-shaving process of heavy-duty gas turbines, and combines the Mamba+Transformer multimodal large model that can quickly and effectively capture the long-term and short-term dependencies of multivariable long-term series to design an intelligent modeling method for multivariable power generation processes. This method fully considers the large span of the data of each variable in the peak-shaving process of heavy-duty gas turbines, and the constructed model has the advantage of strong generalization ability.

[0091] 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 the Mamba large model and the Transformer network, characterized in that: The following steps are involved: S1. Based on the dynamic characteristics of the steam-water process during the peak regulation of heavy-duty gas turbines, a coupled system model structure is constructed to determine the input and output variables. S2. Obtain and process the multivariable mixed data sequence of the heavy-duty gas turbine peak-shaving process, and build a heavy-duty gas turbine peak-shaving process model based on the Mamba+Transformer multimodal large model. The specific steps are as follows: S21. Obtain a multivariate mixed data sequence for a heavy-duty gas turbine peak-shaving process, perform data construction, pooling operations, and multi-layer perception processing, and obtain multi-sequence input data; S22: Perform Mamba module processing, Transformer module processing, and normalization processing on the results of S21 to complete the construction of the Mamba+Transformer multimodal large model; the processing process of the Mamba module is as follows: 1) Use the state space model to process multiple sequence input data, expressed as: ； ； in, For the The state variables at time , For the The state variables at time , The input vector for the model is, Output vector for the model; 、 、 are all parameter matrices of the state-space model; ； ； in, is the base of the exponential function, is the increment, is the identity matrix; 2) Perform convolution kernel calculation on the processed multi-sequence input data as follows: ； in, is the convolution kernel, is the number of rows, is the transpose of the parameter matrix, for The Vandermonde matrix of is the expansion matrix; 3) Output the predicted value, expressed as: ； in, The input vector for the model is, Output vector for the model; S23. Expand the input sequence dimension of the Mamba+Transformer multimodal large model to 2 and the output sequence dimension to 4 to obtain a heavy-duty gas turbine peak-shaving process model; S3. Use the historical operating data of heavy-duty gas turbines to train and verify the heavy-duty gas turbine peak-shaving process model to complete the modeling.

2. The heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network according to claim 1 is characterized by: 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 the Mamba large model and the Transformer network according to claim 1 is characterized by: In S22, the specific process of the Transformer model processing is: using the sinusoidal position encoding mode to learn the sequence, which is expressed as: ； in, is the position of the current semantics in the sequence, is the dimension of semantic embedding, is the dimension of the input sequence.

4. The heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network according to claim 1 is characterized by: In S22, the sequence processed by the Transformer module enters the Mamba module for further processing, and then the output value is normalized to output the model prediction value.

5. The heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network according to claim 1 is characterized by: In S23, the intake air volume and 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.

6. The heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network according to claim 1 is characterized by: In S3, step S2 is used for training and verification, the model prediction value is output and compared with the real data in the historical operation data, and the root mean square error, modeling error, and modeling accuracy are calculated to evaluate the model.

7. The heavy-duty gas turbine intelligent modeling method based on the Mamba large model and the Transformer network according to claim 6 is characterized in that: The calculation formula of the root mean square error is: ； The calculation formula of modeling error is: ； The calculation formula for modeling accuracy is: ； in, is the root mean square error, is the modeling error, For modeling accuracy, Represents the number of arrays, Output variables for the heavy-duty gas turbine peak-shaving process model, is the model's predicted value.

Images