Real-time monitoring method for supercritical water hydrogen production reactor based on ROM and ECT

Abstract

The invention discloses a real-time monitoring method for a supercritical water hydrogen production reactor based on ROM and ECT, and belongs to the technical field of chemical reaction monitoring. The method comprises the following steps: determining key parameters, obtaining a multiphase flow field snapshot through electrical capacitance tomography, and obtaining a multiphase flow field sample matrix based on the multiphase flow field snapshot; obtaining a substrate and a current modal coefficient set based on intrinsic orthogonal decomposition; performing dynamic modal decomposition rolling prediction on the modal coefficient set to obtain a predicted modal coefficient set, and obtaining a global modal coefficient set; on the basis of the mapping relation between the key parameter variables and the modal coefficient set, a global modal coefficient set under unknown parameters is obtained through Gaussian process regression, and the high-order multiphase flow field snapshot is reconstructed on the basis of the substrate and the global modal coefficient set under the unknown parameters. The operation and maintenance state of energy equipment can be monitored, abnormal conditions can be detected, multiphase flow field data can be predicted in real time, and the operation efficiency of the reactor and the stability of the operation process can be improved through visualization of the whole process.

Description

Technical Field

[0001] The present invention relates to the intersection of chemical engineering and fluid dynamics, and in particular to a real-time monitoring method for a supercritical water hydrogen production reactor based on parameterized ROM (reduced order model) and ECT (electrical capacitance tomography). Background Art

[0002] Supercritical water gasification (SCWG), a clean and efficient coal conversion technology, has been widely studied and applied to coal conversion and clean utilization, leveraging the efficient coal gasification performance of supercritical water fluidized bed (SCWFB) reactors under high temperature and pressure. Traditional multiphase flow field testing methods for SCWFB reactors suffer from the following issues: full-order numerical models consume significant computational resources, making them difficult to meet real-time monitoring requirements; ECT experimental measurements result in high noise levels and fuzzy features in multiphase flow field data due to particle collisions and turbulent pulsations; and existing ROMs are primarily targeted at specific boundary conditions and lack parameter generalization capabilities. This results in inefficient and low-quality downstream tasks such as SCWFB reactor analysis, design, and optimization. Summary of the Invention

[0003] To address the aforementioned shortcomings of traditional multiphase flow field monitoring methods in terms of computational efficiency, noise suppression, and adaptability to operating conditions, the present invention provides a real-time monitoring method for a supercritical water hydrogen production reactor based on ROM and ECT. By combining advanced mathematical modeling with experimental testing techniques, the present invention enables real-time monitoring of the multiphase flow field within a SCWFB reactor.

[0004] The technical means adopted in the present invention are as follows:

[0005] A real-time monitoring method for a supercritical water hydrogen production reactor based on ROM and ECT, comprising:

[0006] S1. Determine key parameters during the operation of a supercritical water hydrogen production fluidized bed reactor, obtain a multiphase flow field snapshot by electrical capacitance tomography, and obtain a multiphase flow field sample matrix based on the multiphase flow field snapshot, wherein the key parameters include reactor outer wall temperature, supercritical water mass flow rate, and coal slurry concentration;

[0007] S2. performing proper orthogonal decomposition (POD) on the multiphase flow field sample matrix to obtain a basis and a current modal coefficient sequence;

[0008] S3. Perform dynamic mode decomposition (DMD) on the current modal coefficient set and perform rolling prediction to obtain the future modal coefficient sequence. The current modal coefficient sequence and the future modal coefficient sequence are combined to obtain the global modal coefficient set.

[0009] S4. Based on the mapping relationship between key parameters and modal coefficient sets, Gaussian process regression (GPR) is used to obtain the global modal coefficient set under unknown parameters;

[0010] S5. Reconstruct a snapshot of the high-order multiphase flow field based on the global modal coefficient set under the basis and unknown parameters.

[0011] Furthermore, the method further comprises:

[0012] S6. Compare the reconstructed high-order multiphase flow field snapshots with the actual multiphase flow field snapshots obtained by electrical capacitance tomography, and identify flow dead zones and local concentration anomaly areas based on the residual threshold method, where the residual is calculated according to the following formula:

[0013] e(x,t)=||Φa pred (x,t)-C meas (x,t)||,

[0014] Among them C meas (x, t) is the actual multiphase flow field snapshot obtained by electrical capacitance tomography, a pred (x, t) is the prediction coefficient, Φ is the spatial modal matrix obtained based on the existing multiphase flow field sample matrix POD;

[0015] When the local residual e(x, t) is greater than the preset threshold ε, it is considered that there may be a dead zone or concentration anomaly;

[0016] Calculate the mean μ and standard deviation σ of the current modal coefficient sequence. If the prediction coefficient a pred If any component of (x, t) exceeds μ±3σ, it indicates a local concentration anomaly. Based on the velocity field v(x, t) obtained by POD inversion, when v(x, t) at a certain location is continuously lower than 5% of the average flow velocity and the duration exceeds the preset time threshold τ, it indicates that this location is a flow dead zone.

[0017] Furthermore, obtaining a multiphase flow field sample matrix based on the multiphase flow field snapshot includes:

[0018] Based on ECT, the snapshot of the SCWFB reactor multiphase flow dynamic system at the jth discrete moment is: j∈[1,N t ], where Ns is the number of spatial nodes, N t is the number of sampling time steps;

[0019] Determination of N from the operation process of SCWFB reactor p design variables Get the working condition setting parameter α k , the spatial snapshot at the jth time step is u(x,t j ; α k );

[0020] Collect working condition setting parameter α k The corresponding spatial snapshots of the SCWFB reactor at all time steps are

[0021] Collect N p All snapshot matrices corresponding to group parameters As a multiphase flow field sample matrix.

[0022] Furthermore, POD is performed on the multiphase flow field sample matrix to obtain the basis and the current modal coefficient set, including:

[0023] Capturing parameters through learning The corresponding evolution information of the multiphase flow field of the SCWFB reactor is used to obtain the unknown parameter α j ∈P test The global multiphase flow field snapshot under k represents the kth training set, j represents the test set, and P train ∪P test =Ρ.

[0024] First, the multiphase flow field sample matrix is ​​subjected to POD. Based on the preset truncation order of L, the matrix Q = U T U's eigenvector Γ and eigenvalue matrix λ, where represents the multiphase flow sample matrix;

[0025] The snapshot set characteristics contained in the first L POD bases are:

[0026]

[0027] The snapshot set feature represents the ratio of the cumulative energy of the first L modes to the total energy. POD Falling below the preset threshold triggers new snapshot collection or base update;

[0028] Solve the current modal coefficient matrix according to the following formula:

[0029]

[0030] where a k , k∈[1,L] is a row vector with length N t N p .

[0031] Furthermore, a DMD rolling prediction is performed on the modal coefficient set, including:

[0032] Reorganize the column length of the current modal coefficient matrix A into the number of snapshots N of a single parameter t , the reorganized current modal coefficient matrix is ​​obtained as:

[0033]

[0034] in is a column vector of length LN p ;

[0035] The current modal coefficient matrix will be reorganized Assembled into two parts including N t -1 snapshot of the coefficient matrix but K is the recursive operator matrix, and the matrix Performing singular value decomposition (SVD) yields:

[0036]

[0037] Where M and N are unitary matrices obtained by SVD, and the matrix Σ is a diagonal matrix, then the similarity matrix of matrix K can be obtained

[0038]

[0039] use The eigenvector matrix W and the eigenvalue diagonal matrix Λ can define the modal matrix Ψ and modal amplitude Υ of the matrix K:

[0040]

[0041] Then the modal coefficient at any time can be written as:

[0042]

[0043] Where, ν i is a matrix The i-th eigenvalue of i is the amplitude of the ith mode.

[0044] Furthermore, based on the mapping relationship between key parameter variables and modal coefficient sets, Gaussian process regression (GPR) is used to obtain the global modal coefficient set under unknown parameters, including:

[0045] Will Reorganize by parameters The modal coefficients corresponding to each mode are recorded as For key parameter variables The corresponding modal coefficients The expression of the GPR model is:

[0046]

[0047] Where ε represents the noise variable, which obeys the Gaussian distribution N(0,σ 2 ), Obeying a Gaussian process:

[0048]

[0049] in is the mean function, is the covariance function,

[0050]

[0051] Among them, σ f ,l is the kernel function hyperparameter.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] (1) This invention enables non-contact multiphase flow field measurement using ECT technology, avoiding the limitations of traditional pressure sensors and optical probes in high-temperature and high-pressure environments. The non-interference characteristics of ECT technology enable it to operate stably in the high-temperature and high-pressure environment of a supercritical water reactor, thus adapting to complex gas-liquid-solid multiphase flows.

[0054] (2) The present invention shortens the prediction of full-order multiphase flow field to seconds through POD reduction, significantly improving the efficiency of real-time monitoring, effectively reducing noise interference in experimental data, and improving the identification accuracy of multi-flow field characteristics.

[0055] (3) The present invention combines the DMD and GPR methods. By predicting and optimizing the flow field modal coefficients, it can accurately predict the flow field state at future moments in the case of unknown parameters and automatically adjust the operation of the reactor to ensure that the reactor operates efficiently under optimal conditions.

[0056] (4) This invention designs an integrated data visualization platform based on a B / S architecture, enabling effective integration and visualization of multi-scale, multi-dimensional datasets through web applications. The system supports dynamic rendering and real-time interaction of flow field data. Operators can view information such as flow field evolution, optimization results, and hydrogen production analysis charts through a graphical interface, facilitating real-time monitoring and adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0058] Figure 1 Flowchart of a real-time monitoring method for a supercritical water hydrogen production reactor based on parameterized ROM and ECT in an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of the layout of the SCWFB reactor and the ECT sensor in an embodiment of the present invention;

[0060] Figure 3 2 is a diagram showing the mole fraction distribution of H2 phase in the SCWFB reactor at different times in an embodiment of the present invention;

[0061] Figure 4 The dynamic modal characteristics, frequency and growth rate diagram of the space-time field of the H2 phase mole fraction in the SCWFB reactor in the embodiment of the present invention;

[0062] Figure 5 Graph showing the prediction results of the mole fraction of H2 phase in the SCWFB reactor according to the embodiment of the present invention;

[0063] Figure 6 This is the operating interface of the SCWFB reactor reduction analysis platform based on the B / S architecture in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the solutions of the present invention, 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 embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0065] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0066] like Figure 1 As shown, the present invention provides a real-time monitoring method for a supercritical water hydrogen production reactor based on ROM and ECT, comprising the following steps:

[0067] S1. Key parameters are determined during the operation of a supercritical water fluidized bed (SCWFB) reactor, and a multiphase flow field sample matrix is ​​obtained using electrical capacitance tomography (ECT).

[0068] Specifically, this step is used for data acquisition and sample matrix construction. This application uses ECT technology to obtain multiphase flow field data from the operation process of the SCWFB reactor and construct a sample matrix. Key parameters Through experiments or numerical simulations, small disturbances are applied to each key parameter (such as outer wall temperature, supercritical water flow rate, coal slurry concentration, etc.), and the response amplitude of the multiphase flow field characteristics (pressure drop, local phase distribution, flow pattern transformation, etc.) is observed. The parameters that have the greatest impact on the flow field are regarded as "key parameters". The key parameters in this application include the outer wall temperature of the reactor, the supercritical water mass flow rate and the coal slurry concentration.

[0069] ECT technology uses non-contact measurement and is adaptable to high-temperature and high-pressure environments. It can capture the transient characteristics of gas-liquid-solid multiphase flows, such as flow pattern identification, particle motion trajectory, and three-dimensional phase distribution in SCWFB reactors. Based on ECT, the snapshot of the SCWFB reactor multiphase flow dynamic system at the jth discrete moment is: where N s is the number of spatial nodes, N t is the number of sampling time steps, which is determined from the operation of the SCWFB reactor. p design variables It can be assumed that the parameter α is set in the working condition k , the spatial snapshot at the jth time step is u(x,tj ; α k ). Collect working condition setting parameters α k The corresponding spatial snapshots of the SCWFB reactor at all time steps:

[0070]

[0071] Collect N p All snapshot matrices corresponding to group parameters

[0072] Usually, the key parameter dimensions of interest (such as reactor wall temperature T, supercritical water mass flow rate) are determined first according to the reactor process characteristics and control requirements. Coal slurry concentration C, etc.), thus The parameter domain Ω is defined in , and then random sampling methods such as full factorial design, Latin hypercube sampling, and Monte Carlo random sampling are used to obtain .

[0073] In any of the above sampling strategies, the working condition setting parameter α k , α k is the kth sampling point, that is, an N p dimensional vector, representing a set of specific working condition settings (e.g. For each α k By sampling the reactor under this working condition at ECT, the different time steps t j The spatial flow field snapshot u(x,t j ; α k ), forming a sub-matrix

[0074] Eventually all N p Submatrix U k Splice by column to get the overall sample matrix Used for subsequent POD decomposition.

[0075] S2. Perform proper orthogonal decomposition (POD) analysis based on the sample matrix to obtain the POD basis and the current modal coefficient set.

[0076] This step is based on POD, and the sample matrix is ​​reduced to extract the POD basis and modal coefficient set. The corresponding evolution information of the multiphase flow field of the SCWFB reactor is used to obtain the unknown parameter α j ∈P test The global multiphase flow field snapshot under k represents the kth training set, j represents the test set, and P train UP test=Ρ.

[0077] Evolution information refers to the multiphase flow field data obtained through ECT measurement, which includes time dimension, space dimension and parameter dimension. Time dimension: Evolution information is mainly obtained through the modal coefficient α of POD. i (t) is expressed as a function of time. This coefficient reflects the dynamic characteristics of the corresponding spatial mode over time. Spatial dimension: The spatial mode φ extracted by the POD method i (x) Describes the spatial distribution characteristics of the flow field. These modes themselves are static in space, while their coefficients evolve with time and parameters. Parameter dimensionality: As parameterized flow fields (such as reactor wall temperature, supercritical water mass flow rate, coal slurry concentration, and other design variables) change, the POD coefficients also generate distributions in parameter space. "Evolutionary information" does not directly refer to the key parameters themselves, but rather refers to the evolution of the POD coefficients in time and parameter space, as well as the spatial dimension of the spatial modes.

[0078] First, the matrix Perform POD reduction analysis and set the truncation order to L, and the matrix Q = U T The eigenvector Γ and eigenvalue matrix λ of U: QΓ = λΓ. The snapshot set characteristics contained in the first L POD bases can be characterized by the generalized "energy" they contain, defined as:

[0079]

[0080] Construct the autocorrelation matrix Calculate the autocorrelation matrix And perform eigendecomposition on Q to obtain the eigenvalue {λ j} and the eigenvector {Γ j Generate the spatial basis by Ω=UΓ to get the initial orthogonal basis matrix Then normalize each column vector to form the final POD spatial mode The first L POD basis of the snapshot set of the first L modes corresponds to λ (1) ,…,λ (L) The largest L eigenvalues ​​and their vectors represent the L directions with the highest energy in the flow field. These modes constitute the snapshot set features. Coefficient projection extraction projects the original snapshot U on the first L modes to calculate the modal coefficient matrix The snapshot set features (i.e., projection coefficients) of each mode at each time step can be obtained to find the initial orthogonal basis Ω = UΓ. The orthogonal basis is normalized by column to obtain the basis: Where L is the number of bases.

[0081] The calculation result of the snapshot set feature represents the proportion of the cumulative energy of the first L modes to the total energy. In practical applications, the energy threshold ∈ (for example, 0.9 or 0.95) can be set, and the minimum L can be selected so that E POD ≥∈, to ensure high-precision reconstruction in the lowest dimension. In real-time monitoring, when the system working condition changes and E POD When the value falls below the threshold, it can trigger new snapshot collection or base update to achieve adaptive maintenance of model dimensionality reduction. POD Changes can identify flow dead zones or local concentration anomalies: If the energy contribution distribution changes dramatically under a certain working condition, it means that the flow field structure has deviated significantly, which can be used as an abnormal warning indicator.

[0082] Then we can find the initial orthogonal basis Ω=UΓ and normalize the orthogonal basis by column The POD base is:

[0083]

[0084] Where L is the number of bases. Since Φ is an orthogonal matrix, the set A of POD reduction coefficients can be obtained by the following formula For multi-parameter transient flow problems, the approximate fitting form of high-fidelity data based on POD is U = ΦA. When the high-fidelity flow field snapshot under multi-parameters is decomposed by POD reduction, the POD basis can be obtained. and modal coefficients where a k (k∈[1,L]) is a row vector of length N t N p Through POD reduction, high-dimensional flow field data is mapped to low-dimensional space, significantly reducing the computational complexity while retaining the main characteristics of the flow field.

[0085] S3. Perform DMD on the current modal coefficient set and perform rolling prediction to obtain the future modal coefficient sequence, and merge the current modal coefficient sequence and the future modal coefficient sequence to obtain the global modal coefficient set.

[0086] Specifically, the modal coefficients of each mode at the current moment obtained by POD decomposition in step S2 are directly used to perform rolling prediction on the above current coefficient sequence using DMD or other time series prediction methods to obtain the modal coefficients for a period of time in the future. The global modal dataset is obtained according to the following method:

[0087] Global modal coefficient set = [sampled current coefficient sequence] ∪ [predicted future coefficient sequence]

[0088] This constructs a continuous coefficient sequence from t0 to t0+T, encompassing both actual observations and future estimates, providing complete dynamic information for subsequent reconstruction and monitoring. The "global modal coefficient set" consists of both the observed coefficients from the historical moment to the present and the predicted coefficients from the present moment to the predicted endpoint.

[0089] Specifically, this step uses DMD to perform DMD rolling prediction on the modal coefficient A. Since the row length of A has been reduced to L by the POD method, the order reduction function of the DMD method is no longer used here, and the focus is on the prediction ability of the future moment. The column length of the modal coefficient A is reorganized into the number of snapshots N of a single parameter. t :

[0090]

[0091] in is a column vector of length LN p . The input matrix Assembled into two parts including N t -1 snapshot of the coefficient matrix By the first N t -1 time multiphase flow field data composition, By the back N t -1 time multiphase flow field data, then That is the DMD recursive operator matrix, for the matrix Performing singular value decomposition, we can get:

[0092]

[0093] Where M and N are unitary matrices obtained by SVD, and matrix Σ is a diagonal matrix. Then we can get the similarity matrix of matrix K

[0094]

[0095] use The eigenvector matrix W and the eigenvalue diagonal matrix Λ can define the modal matrix Ψ and modal amplitude Υ of the matrix K:

[0096]

[0097] In summary, the modal coefficients at any time can be written as:

[0098]

[0099] Where, ν i is a matrix The i-th eigenvalue of iis the amplitude of the ith mode. DMD can accurately predict the evolution trend of the flow field by extracting the dynamic modes of the flow field and is suitable for real-time analysis of unsteady flows.

[0100] S4. Based on the mapping relationship between key parameter variables and modal coefficient sets, GPR is used to obtain the global modal coefficient set under unknown parameters.

[0101] This application uses GPR to learn key parameter variables (such as reactor wall temperature, supercritical water mass flow rate, coal slurry concentration) and the global flow field snapshot modal coefficient set The mapping relationship between them is used to obtain the unknown parameter α j The global modal coefficients under the above mapping relationship are constructed, including: extracting the modal basis matrix Φ and the corresponding measured modal coefficient set from the multiphase flow field sample matrix obtained in step S2 through POD decomposition. Taking the key parameter vector x = [wall temperature, supercritical water mass flow rate, slurry concentration, reactor pressure, etc.] as the independent variable and the POD modal coefficient as the dependent variable, the RBF (or Matern) kernel function is selected to construct the GPR model, and the kernel parameters and noise term hyperparameters are optimized by maximizing the marginal likelihood method. K-fold cross validation is used to evaluate the prediction accuracy of GPR under the reserved working conditions to ensure that the mean square error (MSE) is within an acceptable range. For the new unknown parameter x*, the trained GPR model is used to directly predict the corresponding global modal coefficient a*, completing the mapping from the key parameter to the coefficient set. In addition, the GPR mapping naturally has uncertainty assessment, which facilitates the use of confidence intervals to define thresholds in subsequent anomaly detection.

[0102] Since GPR establishes the mapping relationship between input variables and each modal coefficient, Reorganized into The modal coefficients corresponding to each mode are recorded as For key parameter variables The corresponding modal coefficients The expression of the Gaussian process regression model is:

[0103]

[0104] Where ε represents the noise variable, which obeys the Gaussian distribution N(0,σ 2 ). In GPR, it is assumed Obeying the Gaussian process, that is

[0105]

[0106] in is the mean function, is the covariance function, also known as the kernel function, Usually, the mean function is set to 0. The kernel function is the core of the Gaussian process. This paper chooses the radial basis function (RBF), whose basic form is as follows:

[0107]

[0108] Among them, σ f ,e is the kernel function hyperparameter. GPR is modeled through kernel functions (such as radial basis functions), which can effectively handle nonlinear relationships and output confidence intervals for prediction results.

[0109] Reorganize it "by parameters" into So that the kth column For all L modes at the kth time step, all N p Series projection coefficient under different parameter conditions. Will Rearrange again "by mode" and get The first line is all The prediction coefficient is denoted as Original matrix is the projection coefficient of all moments and parameters in the training set and Isomorphic construction process, the former is used for rolling forecast of historical data, and the latter is used for inference of new parameters or new time periods; and Similarly, the two differ only in the "row and column transposition idea", which is used to match the input and output dimensions of GPR.

[0110] In the embodiment of the present application, the offline preset training set parameters Pre-selected by experimental design (such as Latin hypercube, Monte Carlo or Greedy adaptive sampling) to cover the key parameter space (wall temperature, mass flow rate, slurry concentration, etc.). In the real-time monitoring stage of online measurement, the current operating condition α is obtained through the reactor's online sensors (thermocouples, flow meters, density meters, etc.) or process control system * , directly input the trained GPR model for coefficient prediction. If a new working condition that is not covered in the training set occurs in the non-preset situation, a small number of snapshots can be collected first for incremental training or triggering the base adaptive update (when E POD When the value falls below the threshold, GPR inference is performed to ensure prediction robustness.

[0111] S5. Reconstruct the modal coefficients corresponding to the unknown parameters into a high-order multiphase flow field snapshot.

[0112] This step is used for high-order flow field reconstruction and visualization. The predicted modal coefficient set is combined with the POD basis to reconstruct a snapshot of the high-order multiphase flow field. Based on the given trained GPR model, the unknown parameter α can be obtained. j The predicted mean and standard deviation of each order global modal coefficient are obtained by truncating the order L times according to the above GPR process and combining the predicted means to obtain the global modal coefficient set Thus, the global high-order flow field is reconstructed as

[0113] Furthermore, in this embodiment, this step also includes using data flow to convert data information into graphical elements. The visualization model pipeline includes three elements: data objects, operation objects, and the direction of data flow transmission. Front-end development and visualization rendering mainly implement web-based GUI development based on application frameworks such as Streamlit and Matplotlib, providing a friendly and easy-to-use human-computer interaction interface. The complete process from data generation to model training and post-processing visualization can be implemented from a single interface entry. In addition, the component-based development approach allows users to build and customize their own personal data analysis processes.

[0114] As a preferred embodiment of the present application, the above method may further include:

[0115] S6. Compare the reconstructed high-order multiphase flow field snapshots with the actual multiphase flow field snapshots obtained by electrical capacitance tomography to automatically identify flow dead zones or local concentration anomaly areas.

[0116] Specifically, based on the difference between the predicted flow field and the real-time measurement data, the residual threshold method is used to automatically identify flow dead zones and local concentration abnormal areas through statistical and time series analysis methods, and trigger early warning prompts. The residual is calculated according to the following formula:

[0117] e(x,t)=||Φa pred (x,t)-C meas (x,t)||,

[0118] Among them C meas (x,t) is the ECT measurement snapshot, a pred (x, t) is the prediction coefficient, and Φ is the spatial modal matrix obtained based on the intrinsic orthogonal decomposition of the existing multiphase flow field sample matrix.

[0119] When the local residual e>ε (which can be 2-3 times the overall mean square residual), it is considered that there is a significant difference between the measurement and prediction in this area, which may correspond to a dead zone or concentration anomaly.

[0120] For the historical modal coefficient set {a i}Calculate the mean μ and standard deviation σ. If the prediction coefficient a predIf any component in (x, t) exceeds μ±3σ, the corresponding spatial mode has abnormal energy at that moment, suggesting local concentration or structural abnormality. In practical applications, usually only the top k modes with the largest contribution (corresponding to the largest k eigenvalues) are retained to achieve dimensionality reduction, reconstruction and denoising. The prediction coefficient vector a pred Each component a of (x,t) k (x, t) is the corresponding k-th spatial mode Φ k The amplitude (or excitation coefficient) of (x, t) at time t reflects the strength of the mode at the current space-time position.

[0121] In addition, based on the velocity field v(x, t) obtained by POD inversion, when v(x, t) at a certain location is continuously lower than 5% of the average flow velocity and the duration exceeds τ (such as 0.5-1s), it is defined as a flow dead zone.

[0122] The scheme and effects of the present invention are further illustrated below through specific application examples.

[0123] This implementation includes multiple steps in the offline phase and the online phase, such as Figure 1 As shown, the offline phase is Figure 1 The solid line in the middle is the online stage. Figure 1 The dotted line shows the following process:

[0124] First, data acquisition and sample matrix construction are carried out, and multiphase flow field data in the SCWFB reactor are collected in real time through ECT technology. ECT does not need to invade the interior of the reactor, avoiding the risk of mechanical damage under high temperature and high pressure. Some commercial ECT systems (such as EVCT) can withstand high temperature (>500℃) and high pressure environments, and are suitable for supercritical water conditions. The capacitance resolution of the AC-ECT system can reach 0.0001pF, and the signal-to-noise ratio is 73dB, which meets the needs of fine measurement. A multi-section ECT sensor array is arranged on the outer wall of the SCWFB reactor to support three-dimensional phase distribution reconstruction (such as Figure 1 In a supercritical water environment (temperature>374℃, pressure>22.1MPa), the flow field snapshots are collected at high frame rates to capture the transient motion trajectory of particles (such as aggregation, dispersion and back-mixing phenomena), as shown in Figure 2. Figure 2 As shown. Sample matrix generation: The flow field snapshot at the jth discrete moment is recorded as j∈[1,N t ], where N s =512×32 is the number of spatial nodes, N t =1700 is the number of time steps. Collect N p = 17 sets of design variables (such as wall temperature T w , supercritical water mass flow rate T sw , coal slurry concentration C c ) corresponding snapshot matrix Secondly, based on the parameterized order reduction model, the POD order reduction analysis of the flow field is realized, the sample matrix U is POD-ed, and the first L order modes (energy proportion E POD ≥95%), and obtain an orthogonal basis and modal coefficient sets Recombined modal coefficients Use DMD to predict the modal coefficient set of the future moment Extract dynamic modal features (such as frequency, growth rate), such as Figure 4 As shown. Using GPR to establish key parameters (T w 、T sw 、C c ) and the nonlinear mapping relationship between the global modal coefficient set and the flow field evolution under unknown parameters. Finally, the predicted global modal coefficient set A * Combined with POD substrate, reconstructing high-order flow field snapshots

[0125] Comparing the predicted flow field with the real-time monitoring data, the dead zone or local concentration abnormality area is automatically identified and the early warning mechanism is triggered. The residual is calculated using the following formula:

[0126] e(x,t)=||Φa pred (x,t)-C meas (x,t)||,

[0127] Among them C meas is a snapshot of the ECT measurement. When the local residual e>ε (which can be 2-3 times the overall mean square residual), it is considered that there is a significant difference between the measurement and prediction in this area, which may correspond to a dead zone or concentration anomaly. i}Calculate the mean μ and standard deviation σ. If the prediction coefficient a pred If any component of exceeds μ ± 3σ, the corresponding spatial mode exhibits abnormal energy at that moment, suggesting a local concentration or structural anomaly. Furthermore, based on the velocity field v(x, t) obtained from POD inversion, a dead zone is defined when v(x, t) at a certain location is continuously below 5% of the average velocity for a duration exceeding τ (e.g., 0.5-1 s).

[0128] like Figure 5 As shown, at T W ∈[823K,923K], M SCW ∈[0.001kg / s,0.003kg / s], C C∈[23%,30%], the PSNR of the model-predicted flow field and the ECT result is >30dB, SSIM>0.9, the maximum error is ≤10%, and the response time is shortened to seconds. 3D unstructured grid data is rendered based on the Streamlit pipeline, supporting functions such as clipping plane operations and multi-window comparison. Analysis charts such as gas production bar charts and modal spectrum charts are generated through Matplotlib. Web applications are built based on B / S architecture (such as Figure 3 The following functions are realized: users upload parameter configuration, start model training, and export prediction results through the front-end interface (Matplotlib framework). The time slider controls the dynamic display of multiphase flow field cloud map and two-dimensional chart, and displays the change of H2 gas production in real time, such as Figure 6 shown.

[0129] The present invention adopts POD to reduce the order of the multi-parameter, multi-time-step snapshot matrix collected by ECT to extract the dominant spatial modes of the flow field and the corresponding modal coefficients.

[0130] POD calculates the eigenvalues ​​and eigenvectors of the autocorrelation matrix through SVD, projecting the high-dimensional flow field into a small number of modal subspaces with the highest energy, significantly reducing computational complexity. This technology has been widely used in the construction of fluid dynamics prediction models, reducing flow field prediction latency to seconds while maintaining an energy capture rate of over 90%.

[0131] DMD is a time series decomposition method based on Koopman operator approximation, which is used to extract the dynamic modes of the flow field and their frequencies and acceleration and deceleration rates.

[0132] Based on the DMD principle, the successive reconstructed modal coefficient matrices are decomposed using SVD to construct similarity matrices and recover the DMD operator, enabling rolling prediction of future coefficients. Using the DMD mode matrix and modal amplitudes, a global modal coefficient set is obtained for subsequent reconstruction.

[0133] The present invention uses GPR to establish a non-parametric mapping relationship between key parameters (such as wall temperature, mass flow rate, coal slurry concentration, etc.) and global modal coefficients, so as to predict the coefficient set under unknown working conditions.

[0134] The GPR model constructs the covariance matrix using a radial basis kernel function (RBF) or a Matern kernel. It automatically optimizes hyperparameters by maximizing marginal likelihood and provides prediction uncertainty assessment to enhance model robustness. In a leave-one-out cross-validation process, GPR has been shown to effectively capture the nonlinear relationship between coefficients and input variables in parameterized multiphase flow models, ensuring both prediction accuracy and generalization.

[0135] Based on the spatial modal matrix Φ and the global modal coefficient set A global , the reconstruction of high-order multiphase flow field snapshots is achieved through linear superposition. This reconstruction method uses POD-Galerkin or non-invasive projection technology to significantly reduce the amount of online calculation while ensuring an energy capture rate of more than 95%.

[0136] In the parameterized scenario, by comparing the reconstruction residual and the energy capture ratio E POD Dynamic monitoring can trigger base updates and anomaly detection in real time, enhancing system adaptability and stability.

[0137] The present invention designs an integrated data visualization platform based on the Browser / Server (B / S) architecture, which manages the database, application logic and browser interface in layers, reducing the burden on the client and facilitating maintenance and upgrades. The platform uses WebGL rendering frameworks such as Cesium to achieve real-time dynamic rendering and interactive viewing of multi-scale, multi-dimensional multiphase flow data, supporting functions such as flow field evolution, modal coefficient trends and hydrogen production analysis charts. At the same time, enterprise-level BI tools (such as Pentaho, etc.) are used to build draggable dashboards and report components, providing operators with a friendly interface for real-time monitoring of working conditions, alarms and optimization suggestions.

[0138] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time monitoring method for a supercritical water hydrogen production reactor based on ROM and ECT, characterized in that: include: S1. Determine key parameters during the operation of a supercritical water hydrogen production fluidized bed reactor, obtain a multiphase flow field snapshot by electrical capacitance tomography, and obtain a multiphase flow field sample matrix based on the multiphase flow field snapshot, wherein the key parameters include reactor outer wall temperature, supercritical water mass flow rate, and coal slurry concentration; S2. performing intrinsic orthogonal decomposition on the multiphase flow field sample matrix to obtain a basis and a current modal coefficient sequence; S3. Perform dynamic modal decomposition on the current modal coefficient set and perform rolling prediction to obtain a future modal coefficient sequence, and merge the current modal coefficient sequence and the future modal coefficient sequence to obtain a global modal coefficient set; S4. Based on the mapping relationship between key parameters and modal coefficient sets, the global modal coefficient set under unknown parameters is obtained using Gaussian process regression; S5. Reconstruct a snapshot of the high-order multiphase flow field based on the global modal coefficient set under the basis and unknown parameters.

2. The real-time monitoring method of a supercritical water hydrogen production reactor based on ROM and ECT according to claim 1, characterized in that: The method further comprises: S6. Compare the reconstructed high-order multiphase flow field snapshots with the actual multiphase flow field snapshots obtained by electrical capacitance tomography, and identify flow dead zones and local concentration anomaly areas based on the residual threshold method, where the residual is calculated according to the following formula: e(x,t)=||Φa pred (x,t)-C meas (x,t)||, Among them C meas (x, t) is the actual multiphase flow field snapshot obtained by electrical capacitance tomography, a pred (x, t) is the prediction coefficient, Φ is the spatial modal matrix obtained based on the intrinsic orthogonal decomposition of the existing multiphase flow field sample matrix; When the local residual e(x, t) is greater than the preset threshold ε, it is considered that there may be a dead zone or concentration anomaly; Calculate the mean μ and standard deviation σ of the current modal coefficient sequence. If the prediction coefficient a pred If any component of (x, t) exceeds μ±3σ, it indicates a local concentration anomaly. Based on the velocity field v(x, t) obtained by POD inversion, when v(x, t) at a certain location is continuously lower than 5% of the average flow velocity and the duration exceeds the preset time threshold τ, it indicates that this location is a flow dead zone.

3. The real-time monitoring method of a supercritical water hydrogen production reactor based on ROM and ECT according to claim 1, characterized in that: Acquiring a multiphase flow field sample matrix based on the multiphase flow field snapshot includes: The snapshot of the multiphase flow dynamic system of the SCWFB reactor at the jth discrete moment based on electrical capacitance tomography is: where N s is the number of spatial nodes, N t is the number of sampling time steps; Determination of N from the operation process of SCWFB reactor p design variables Get the working condition setting parameter α k , the spatial snapshot at the jth time step is u(x,t j ; α k ); Collect working condition setting parameters α k The corresponding spatial snapshots of the SCWFB reactor at all time steps are Collect N p All snapshot matrices corresponding to group parameters As a multiphase flow field sample matrix.

4. The real-time monitoring method of a supercritical water hydrogen production reactor based on ROM and ECT according to claim 3, characterized in that: Performing intrinsic orthogonal decomposition on the multiphase flow field sample matrix to obtain a basis and a current modal coefficient set, including: Capturing parameters through learning The corresponding evolution information of the multiphase flow field of the SCWFB reactor is used to obtain the unknown parameter α j ∈P test The global multiphase flow field snapshot under k represents the kth training set, j represents the test set, and P train ∪P test =P; Perform intrinsic orthogonal decomposition on the multiphase flow field sample matrix, and based on the preset truncation order L, the matrix can be obtained The eigenvector Γ and eigenvalue matrix λ of represents the multiphase flow sample matrix; The snapshot set characteristics contained in the first L POD bases are: The snapshot set feature represents the proportion of the cumulative energy of the first L modes to the total energy. POD Falling below the preset threshold triggers new snapshot collection or base update; Solve the current modal coefficient matrix according to the following formula: where a k is a length of N t N p The row vector of , k∈[1,L].

5. The real-time monitoring method of a supercritical water hydrogen production reactor based on ROM and ECT according to claim 4, characterized in that: Perform dynamic modal decomposition and rolling prediction on the modal coefficient set, including: Reorganize the column length of the current modal coefficient matrix A into the number of snapshots N of a single parameter t , the reorganized current modal coefficient matrix is ​​obtained as: in is a column vector of length LN p ; The current modal coefficient matrix will be reorganized Assembled into two parts including N t -1 snapshot of the coefficient matrix but K is the recursive operator matrix, and the matrix Performing singular value decomposition, we can get: Where M and N are unitary matrices obtained by singular value decomposition, and the matrix Σ is a diagonal matrix, then the similarity matrix of matrix K can be obtained use The eigenvector matrix W and the eigenvalue diagonal matrix Λ can define the modal matrix Ψ and modal amplitude Υ of the matrix K: Then the modal coefficient at any time can be written as: Where, ν i is a matrix The i-th eigenvalue of i is the amplitude of the ith mode.

6. The real-time monitoring method of a supercritical water hydrogen production reactor based on ROM and ECT according to claim 5, characterized in that: Based on the mapping relationship between key parameter variables and modal coefficient sets, Gaussian process regression is used to obtain the global modal coefficient set under unknown parameters, including: Will Reorganize by parameters The modal coefficients corresponding to each mode are recorded as For key parameter variables The corresponding modal coefficients The expression of the Gaussian process regression model is: Where ε represents the noise variable, which obeys the Gaussian distribution N(0,σ 2 ), Obeying a Gaussian process: in is the mean function, is the covariance function, Among them, σ f ,l is the kernel function hyperparameter.

Images