Circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and time convolutional neural network

Abstract

The invention discloses a circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and a time convolutional neural network. In the off-line preparation stage, firstly, original variable information in a multi-phase combustion field is obtained, the time average value of grid variables is calculated, and a result snapshot matrix is decentralized in time; and obtaining a POD mode and a corresponding mode coefficient thereof, and obtaining the optimal arrangement positions of the time convolutional neural network and the sensor. In the online prediction stage, the modal coefficient at the current moment is reconstructed based on the position of the sensor, the future modal coefficient is predicted based on the time convolutional neural network, and the flow field is reconstructed through linear combination with the POD modal. According to the method, on-line rapid prediction of full-field variables at the future moment is realized by utilizing sparse sensor data at the current moment, the defect that a traditional CFD method cannot be applied to industrial production rapid prediction is overcome, and the blank of multiphase combustion in the circulating fluidized bed in the aspect of on-line flow field rapid prediction is filled.

Description

Technical Field

[0001] The present invention belongs to the field of numerical simulation, and in particular relates to a circulating fluidized bed combustion reduction prediction method based on enhanced compressed sensing and temporal convolutional neural network. Background Art

[0002] In recent years, China has vigorously promoted the construction of an energy supply system based on clean, low-carbon energy, including promoting the clean and efficient use of coal, reducing environmental pollution, and promoting a diversified energy structure. Fluidized beds (FBDs) are one of the most important technologies for achieving these goals. However, multiphase combustion processes within fluidized beds involve interactions between particles and fluids, particles and particles, and particles and walls, as well as multi-physics couplings such as heat transfer, mass transfer, and chemical reactions, posing significant challenges to experimental measurement. With the advancement of computer technology, numerical simulation has become a powerful tool for studying multi-physics coupling in fluidization. Traditional computational fluid dynamics (CFD)-based methods, such as the Euler-Euler two-fluid model and the computational fluid dynamics-discrete element method (CFD-DEM), can accurately capture particle and grid-scale information, but they often require extremely small time steps to resolve transient particle displacements and collisions. Simulating industrial-scale fluidized bed equipment consumes significant computing resources. Industrial production processes place extremely high demands on simulation speed, often requiring full-scale real-time simulation. The slow computational speed of traditional CFD methods has limited their application and development in the fluidization industry.

[0003] With the rapid development of technologies such as artificial intelligence and big data, the concept of "digital interconnection," exemplified by the Internet of Things, has gradually emerged. Its goal is to achieve real-time simulation and prediction, bridging the physical world with the digital virtual realm. This concept offers a new approach for the rapid prediction of multiphase reaction flows in fluidized beds: by developing a rapid reduced-order prediction method for the reactor, real-time data from the physical world can be input into the virtual realm for efficient prediction, thereby providing real-time guidance for industrial production. The development of reduced-order prediction methods for circulating fluidized bed combustion faces two major challenges. First, the limited number of sensors in the physical world poses a major challenge in acquiring full-field data from the entire system within a limited sensor layout. Second, current traditional CFD (computational fluid dynamics) algorithms struggle to meet the requirements for rapid prediction. Developing efficient acceleration algorithms to ensure that reduced-order prediction methods can achieve real-time response is a key issue that needs to be addressed. The proper orthogonal decomposition method decomposes the spatiotemporal flow field into dominant spatial modes and corresponding temporal coefficients based on existing CFD snapshot data. The dominant spatial modes are used to locate sensor locations within a limited number of sensors, enabling compressed sensing of the flow field; the temporal coefficients are used for online, real-time prediction.

[0004] Based on the above background, it is necessary to carry out research on circulating fluidized bed combustion reduced-order prediction methods based on compressed sensing and time convolutional neural networks within the framework of the intrinsic orthogonal decomposition method, break through the limitations of traditional CFD algorithms in computing resources, and develop a rapid prediction platform based on real physical sensor data, which can be applied to industrial research on multiphase reaction flows. Summary of the Invention

[0005] In order to achieve online real-time prediction of circulating fluidized bed combustion, the present invention provides a circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and temporal convolutional neural network.

[0006] The object of the present invention is achieved through the following technical solution: a circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and temporal convolutional neural network, the method comprising the following steps:

[0007] (1) Obtain a snapshot of the flow field inside the multiphase combustion field, perform time averaging on the snapshot data, and obtain a decentralized flow field;

[0008] (2) Perform intrinsic orthogonal decomposition on the decentralized flow field to obtain the various modes of the velocity field and component field variables of the combustion process and their corresponding modal coefficients;

[0009] (3) According to the energy of each mode containing the flow field characteristics, the required number of intrinsically orthogonal POD modes are selected;

[0010] (4) Based on the selected POD mode, the optimal placement of sensors is obtained using the discrete empirical interpolation method and the column principal component-based QR decomposition method;

[0011] (5) Based on the selected POD mode and its coefficients, the temporal convolutional neural network is trained;

[0012] (6) In the online stage, sensor data is used as input to reconstruct the modal coefficients at the current moment and perform noise reduction optimization;

[0013] (7) The optimized modal coefficients at the current moment are used as the input of the temporal convolutional neural network to predict the modal time coefficients at the future moment;

[0014] (8) The predicted modal coefficients at future moments are combined with the POD modes to reconstruct the future flow field.

[0015] Furthermore, the flow snapshots inside the multiphase combustion field are decentralized as follows:

[0016] [x1,x2,…,x n ]=[w1-W0(x),w2-W0(x),…,w n -W0(x)] (1) where wn is the multiphase combustion field variable information at the nth moment, including velocity, pressure, temperature and component mass fraction; x n is the decentralized variable at the nth moment; W0(x) is the time average of the flow field variable information. The calculation process is as follows:

[0017]

[0018] Where x represents the grid label vector and N is the number of time snapshots.

[0019] Furthermore, the flow field snapshot is modally decomposed. In order to make the target mode contain as many flow field features as possible, it is necessary to minimize the projection distance from the decentralized original variable to the target mode, which is equivalent to maximizing the square of the inner product of the two:

[0020]

[0021] Among them, Φ k is the target mode, k is the order of the mode, and formula (3) is equivalent to:

[0022]

[0023] Where X=[x1,x2,…,x n ],XX T is a semi-positive definite matrix; at the same time, the flow field variable snapshot matrix is ​​projected onto Φ k When , follow the rule of formula (5):

[0024]

[0025] Among them, σ k (X T ) is X T Therefore, the target mode can be obtained by performing singular value decomposition (SVD) on the flow field snapshot, as follows: After SVD, the singular values ​​are arranged in descending order. Therefore, according to formula (5), when k = 1, Φ k The first-order mode contains the most features. As k increases, all POD modes are obtained, and the information of flow field variables gradually decreases. Therefore, the solution of σ based on SVD is k (X T ) and its eigenvectors, and then obtain the POD mode; SVD decomposes the matrix X of any shape into the product of multiple matrices: X=UΣV TThe orthogonal vectors in U are called left singular vectors. The elements on the diagonal in Σ are called singular values, which are arranged in descending order and are all 0 except for the diagonal elements. The orthogonal vectors in V are called right singular vectors. U can be regarded as the required modes. The time coefficient is obtained synchronously with the modes. Its amplitude is related to the eigenvalue size and is calculated by the following formula:

[0026] in Represents the coefficient of the kth flow field variable mode at time n during the transient process.

[0027] Furthermore, the required number of modes is selected according to the required accuracy, and the characteristic quantity ratio of the first K-order modes is given by the following formula:

[0028]

[0029] Where L is the total number of modes obtained; the number of modes is selected based on the required truncation error accuracy.

[0030] Furthermore, step (4) for determining the sensor position is as follows: determine the first sensor position p1, and use the discrete empirical interpolation method to determine the sensor position, requiring the first sensor position to be the position where the absolute value of the first-order mode is the maximum, that is, satisfying:

[0031]

[0032] Determine the second and subsequent sensor positions p j ; When you already have j-1 measurement points, define the projection matrix ρ j-1 =Ψ(B T Ψ) -1 B T , the jth sensor position p j Through the decision matrix χ j get:

[0033]

[0034] χ j =Φ j -ρ j-1 Φ j =Φ j -Ψ(B T Ψ) -1 B T Φ j (10)

[0035] Among them, B is the position information of the observation matrix composed of 0 and 1, and the modal matrix Ψ=[Φ1,Φ2,…,Φ j-1 ].

[0036] Furthermore, the temporal convolutional neural network is trained using the data at the existing time points to prepare for the online prediction process.

[0037] Furthermore, in the online prediction stage, for the two cases where the number of sensors is not less than or less than the required modal number, the data of the target sensor position is used as input, and the compressed sensing method is used to reconstruct the modal coefficients at the current moment:

[0038] α0=Θ -1 y (11.1)

[0039] α0=Θ T (ΘΘ T ) -1 y (11.2)

[0040] Among them, α0 is the modal coefficient without noise reduction processing, the permutation matrix Θ = BΦ, y is the sensor data, and B is the observation point position matrix.

[0041] Furthermore, the Savitzky-Golay denoising algorithm is used sg Smoothing the reconstructed modal coefficients can further improve the reconstruction accuracy:

[0042] α=f sg (α0) (12)

[0043] Where α is the modal coefficient after noise reduction.

[0044] Furthermore, based on the trained temporal convolutional neural network f TCN Convection field variable modal coefficients Perform time series forecasting as follows:

[0045] α output =f TCN (α input ) (13)

[0046] Among them, α input is the input coefficient of the prediction process, α output Output coefficients for the prediction process; for the selected POD mode, the existing POD mode coefficients are continuously used to predict future time steps.

[0047] Furthermore, according to the existing POD mode Φ k , combined with the predicted POD coefficient Reconstructing the flow field The calculation process is obtained by the following formula:

[0048]

[0049] Beneficial effects of the present invention: In response to the problem that traditional CFD algorithms are difficult to meet the requirements of fast real-time prediction, the present invention has developed a reduced-order prediction method for circulating fluidized bed combustion. The data at the sensor position is used as input to reconstruct the modal coefficients at the current moment. After the modal coefficients are denoised using a denoising algorithm, a time convolutional neural network is used to predict the modal coefficients at future time steps, and finally the flow field is reconstructed by linear combination with the existing POD mode. This method achieves online fast prediction of full-field variables at future moments using sparse sensor data at the current moment, overcomes the deficiency of traditional CFD methods in being unable to be applied to industrial production for fast prediction, and fills the gap in online fast full-field prediction of multiphase reaction flows. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The overall logic diagram of the circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and temporal convolutional neural network;

[0051] Figure 2 Schematic diagram of temporal convolutional network;

[0052] Figure 3 This is a schematic diagram comparing the prediction accuracy of the reduced-order prediction methods;

[0053] Figure 4 This is a schematic diagram comparing the acceleration effect of reduced-order prediction; DETAILED DESCRIPTION

[0054] The present invention will be further described below with reference to the accompanying drawings and examples.

[0055] The present invention provides a reduced-order prediction method for circulating fluidized bed combustion based on enhanced compressed sensing and temporal convolutional neural networks. Based on existing CFD snapshot data, the spatiotemporal flow field is decomposed into dominant spatial modes and corresponding time coefficients, and the modes are sorted from high to low according to the size of the eigenvalues ​​corresponding to the characteristic modes. The dominant spatial modes are used to find the sensor positions with a limited number of sensors to obtain compressed sensing of the flow field; the time coefficients are predicted online in real time by using a neural convolutional neural network. Ultimately, a rapid prediction platform based on real physical sensor data is developed, which can be applied to industrial research on multiphase reaction flows.

[0056] The following application is a circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and time convolutional neural network, such as Figure 1 The specific steps are as follows:

[0057] (1.1) Decentralize the flow snapshot inside the multiphase combustion field:

[0058] [x1,x2,…,x n ]=[w1-W0(x),w2-W0(x),…,wn -W0(x)] (1) where w n is the multiphase combustion field variable information at the nth moment, including velocity, pressure, temperature and component mass fraction; x n is the decentralized variable at the nth moment; W0(x) is the time average of the flow field variable information. The calculation process is as follows:

[0059]

[0060] Where x represents the grid label vector and N is the number of time snapshots.

[0061] (1.2) Clarify the POD decomposition principle. In order to make the target mode contain as many flow field features as possible, it is necessary to minimize the projection distance from the decentralized original variable to the target mode, which is equivalent to maximizing the square of the inner product of the two:

[0062]

[0063] Among them, Φ k is the target mode, k is the order of the mode, and formula (3) is equivalent to:

[0064]

[0065] Where X=[x1,x2,…,x n ],XX T is a positive semidefinite matrix.

[0066] Project the flow field variable snapshot matrix onto Φ k When , follow the rule of formula (5):

[0067]

[0068] Among them, σ k (X T ) is X T The singular values ​​of .

[0069] Furthermore, after the singular value decomposition SVD is performed on the flow field snapshot, the singular values ​​are arranged in descending order. Therefore, according to formula (5), when k = 1, Φ k The first-order mode contains the most features. As k increases, all POD modes are obtained, and the information of flow field variables gradually decreases. Therefore, σ can be solved based on SVD. k (X T ) and its eigenvectors, and then obtain the POD mode.

[0070] (1.3) Use SVD to perform POD decomposition and obtain the modes and their corresponding coefficients. SVD decomposes a matrix X of any shape into the product of multiple matrices: X = UΣVT The orthogonal vectors in U are called left singular vectors. The elements on the diagonal in Σ are called singular values, which are arranged in descending order, and all elements outside the diagonal are 0. The orthogonal vectors in V are called right singular vectors. Among them, U can be regarded as the required modes. The time coefficient is obtained synchronously with each mode. Its amplitude is related to the size of the eigenvalue and is calculated by the following formula:

[0071] in Represents the coefficient of the kth flow field variable mode at time n during the transient process.

[0072] (1.4) Extract characteristic modes. Select the required number of modes according to the required accuracy. The characteristic quantity ratio of the first K order modes is given by the following formula:

[0073]

[0074] Where L is the total number of modes obtained; the number of modes is selected based on the required truncation error accuracy.

[0075] (1.5) Determine the sensor positions. The number of sensors is determined by industrial production requirements. First, determine the first sensor position p1. Use the discrete empirical interpolation method to determine the sensor position. The first sensor position is required to be the position where the absolute value of the first-order mode is the largest, that is, it satisfies:

[0076]

[0077] Determine the second and subsequent sensor positions p j ; When you already have j-1 measurement points, define the projection matrix ρ j-1 =Ψ(B T Ψ) -1 B T , the jth sensor position p j Through the decision matrix χ j get:

[0078]

[0079] χ j =Φ j -ρ j-1 Φ j =Φ j -Ψ(B T Ψ) -1 B T Φ j (10) Where B is the position information of the observation matrix composed of 0 and 1, and the modal matrix Ψ = [Φ1, Φ2, …, Φ j-1 ].

[0080] (1.6) Use the existing time data to train the temporal convolutional neural network to prepare for the online prediction process of the reduced-order prediction method, such as Figure 2 shown.

[0081] like Figure 1 As shown, for the online prediction process, the specific steps are as follows:

[0082] (2.1) In the online prediction stage, for the two cases where the number of sensors is not less than or less than the required modal number, the data of the target sensor position is used as input, and the compressed sensing method is used to reconstruct the modal coefficients at the current moment:

[0083] α0=Θ -1 y (11.1)

[0084] α0=Θ T (ΘΘ T ) -1 y (11.2)

[0085] Among them, α0 is the modal coefficient without noise reduction processing, the permutation matrix Θ = BΦ, y is the sensor data, and B is the observation point position matrix.

[0086] (2.2) Using Savitzky-Golay denoising algorithm f sg Smoothing the reconstructed modal coefficients can further improve the reconstruction accuracy:

[0087] α=f sg (α0) (12)

[0088] Where α is the modal coefficient after noise reduction.

[0089] (2.3) Based on the trained temporal convolutional neural network f TCN Convection field variable modal time coefficient Make predictions as follows:

[0090] α output =f TCN (α input ) (13)

[0091] Among them, α input is the input coefficient of the prediction process, α output is the output coefficient of the prediction process.

[0092] (2.4) Repeat steps (2.1)(2.2)(2.3), and continuously use the existing sensor data to predict the modal coefficients of future time steps.

[0093] (2.5) According to the existing POD mode Φ k , combined with the predicted POD coefficient Reconstructing the flow field The calculation process is obtained by the following formula:

[0094]

[0095] The present invention provides a circulating fluidized bed combustion reduction prediction method based on enhanced compressed sensing and time convolutional neural network. The comparison of the reduction prediction method is shown in the figure below. Figure 3 As shown. Based on the existing CFD snapshot data, the space-time flow field is decomposed into dominant spatial modes and corresponding time coefficients, and the modes are sorted from high to low according to the size of the eigenvalues ​​corresponding to the characteristic modes. The dominant spatial modes are used to find the sensor positions under a limited number of sensors to obtain compressed perception of the flow field; the time coefficients rely on neural convolutional neural networks for online real-time prediction. Finally, a fast prediction platform based on real physical sensor data was developed, which can be applied to industrial research on multiphase reaction flows. By constructing a reduced-order prediction method for the combustion process data in a circulating fluidized bed, it was verified that the method of the present invention can accurately predict the data in the fluidized bed online and realize the online early warning function. The single-core operation acceleration effect is as follows. Figure 4 shown.

[0096] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and temporal convolutional neural network, characterized in that: The method comprises the following steps: (1) Obtain a snapshot of the flow field inside the multiphase combustion field, perform time averaging on the snapshot data, and obtain a decentralized flow field; (2) Perform intrinsic orthogonal decomposition on the decentralized flow field to obtain the various modes of the velocity field and component field variables of the combustion process and their corresponding modal coefficients; (3) According to the energy of each mode containing the flow field characteristics, the required number of intrinsically orthogonal POD modes are selected; (4) Based on the selected POD mode, the optimal placement of the sensors is obtained using the discrete empirical interpolation method or the column principal component-based QR decomposition method; (5) Based on the selected POD modal corresponding coefficients, the temporal convolutional neural network is trained; (6) In the online stage, sensor data is used as input to reconstruct the modal coefficients at the current moment and perform noise reduction optimization; (7) The optimized modal coefficients at the current moment are used as the input of the temporal convolutional neural network to predict the modal time coefficients at the future moment; (8) The predicted modal coefficients at future moments are combined with the POD modes to reconstruct the future flow field.

2. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1 is characterized in that: The snapshot of the multiphase reaction flow inside the multiphase combustion field is decentralized as follows: [x1,x2,…,x n ]=[w1-W0(x),w2-W0(x),…,w n -W0(x)] (1) where w n is the multiphase combustion field variable information at the nth moment, including velocity, pressure, temperature and component mass fraction; x n is the decentralized variable at the nth moment; W0(x) is the time average of the flow field variable information. The calculation process is as follows: Where x represents the grid label vector and N is the number of time snapshots.

3. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1 is characterized in that: Perform modal decomposition on the flow field snapshot. In order to make the target mode contain as many flow field features as possible, it is necessary to minimize the projection distance from the decentralized original variable to the target mode, which is equivalent to maximizing the square of the inner product of the two: Among them, Φ k is the target mode, k is the order of the mode, and formula (3) is equivalent to: Where X=[x1,x2,…,x n ],XX T is a semi-positive definite matrix; at the same time, the flow field variable snapshot matrix is ​​projected onto Φ k When , follow the rule of formula (5): Among them, σ k (X T ) is X T The singular value of ; Therefore, the target mode can be obtained by performing singular value decomposition SVD on the flow field snapshot, as follows: After SVD, the singular values ​​are arranged in descending order, so according to formula (5), when k = 1, Φ k The first-order mode contains the most features. As k increases, all POD modes are obtained, and the information of flow field variables gradually decreases. Therefore, the solution of σ based on SVD is k (X T ) and its eigenvectors, and then obtain the POD mode; SVD decomposes the matrix X of any shape into the product of multiple matrices: X=UΣV T The orthogonal vectors in U are called left singular vectors. The elements on the diagonal in Σ are called singular values, which are arranged in descending order and are all 0 except for the diagonal elements. The orthogonal vectors in V are called right singular vectors. U can be regarded as the required modes. The time coefficient is obtained synchronously with the modes. Its amplitude is related to the eigenvalue size and is calculated by the following formula: in Represents the coefficient of the kth flow field variable mode at time n during the transient process.

4. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1 is characterized in that: The required number of modes is selected according to the required accuracy, and the characteristic quantity ratio of the first K-order modes is given by the following formula: Where L is the total number of modes obtained; the number of modes is selected based on the required truncation error accuracy.

5. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1 is characterized in that: When the number of sensors is not less than the number of required modes, the step (4) for determining the sensor positions is as follows: (5.1) Determine the first sensor position p1 and use the discrete empirical interpolation method to determine the sensor position. The first sensor position is required to be the position where the absolute value of the first-order mode is the largest, that is, it satisfies: (5.2) Determine the second and subsequent sensor positions p j ; When you already have j-1 measurement points, define the projection matrix ρ j-1 =Ψ(B T Ψ) -1 B T , the jth sensor position p j Through the decision matrix χ j get: χ j =Φ j -ρ j-1 Φ j =Φ j -Ψ(B T Ψ) -1 B T Φ j (10) Where B is the position information of the observation matrix composed of 0 and 1, and the modal matrix Ψ = [Φ1, Φ2, …, Φ j-1 ].

6. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1, characterized in that: Use existing time-lapse data to train a temporal convolutional neural network in preparation for the online prediction process.

7. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1 or 5, characterized in that: In the online prediction stage, for the two cases where the number of sensors is not less than or less than the required modal number, the data of the target sensor position is used as input, and the compressed sensing method is used to reconstruct the modal coefficients at the current moment: α0=Θ -1 (11.1) α0=Θ T (TH) T ) -1 (11.2) Among them, α0 is the modal coefficient without noise reduction processing, the permutation matrix Θ = BΦ, y is the sensor data, and B is the observation point position matrix.

8. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to claim 1 or 7, characterized in that: Using Savitzky-Golay denoising algorithm f sg Smoothing the reconstructed modal coefficients can further improve the reconstruction accuracy: α=f sg (α0) (12)In the formula, α is the modal coefficient after noise reduction.

9. The method for reducing the order of combustion in a circulating fluidized bed based on enhanced compressed sensing and temporal convolutional neural network according to any one of claims 1, 6 or 8, characterized in that: Based on the trained temporal convolutional neural network f TCN Convection field variable modal time coefficient Make predictions as follows: α output ＝f TCN (α input ) (13) Among them, α input is the input coefficient of the prediction process, α output Output coefficients for the prediction process; for the selected POD mode, the existing POD mode coefficients are continuously used to predict future time steps.

10. A circulating fluidized bed combustion reduced-order prediction method based on enhanced compressed sensing and temporal convolutional neural network according to any one of claims 1, 3 or 9, characterized in that: According to the existing POD mode Φ k , combined with the predicted POD coefficient Reconstructing the flow field The calculation process is obtained by the following formula:

Images