Nonlinear random dynamic model parameter identification method for aromatic hydrocarbon combination unit

Abstract

The invention relates to a nonlinear stochastic dynamic model parameter identification method for an aromatic hydrocarbon combined unit, which adopts a stochastic differential equation to describe a continuous dynamic process of the aromatic hydrocarbon combined unit: dx (t) = f (x (t), u (t)) dt + G (x (t)) dW (t) + H (x (t)) d (t) dt, where: dx (t) = f (x (t), u (t)) dt + G (x (t)) dW (t) + H (x (t)) d (t) dt is a state vector; the vector is a control vector; f (.) is a nonlinear drift term and depicts deterministic nonlinear dynamic characteristics in the aromatic hydrocarbon device; w (t) is a d-dimensional Wiener process and is used for describing micro-disturbance continuous random noise; g (.) is a diffusion coefficient matrix and is used for measuring the strength and action mode of continuous random disturbance; h (.) is a disturbance correlation function, reflects the action relation of stochastic disturbance capable of modeling on the device state, and provides reliable model support for optimization operation and advanced control of the aromatic hydrocarbon combination device.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advanced control of petroleum chemical processes, and particularly relates to a nonlinear stochastic dynamic modeling and parameter identification method for an aromatic hydrocarbon combined device (covering core units such as catalytic reforming, aromatic extraction, disproportionation and isomerization), which is suitable for optimization of industrial processes containing strong nonlinearities and random disturbances. BACKGROUND

[0002] The aromatic hydrocarbon combined device is a key device in the petroleum chemical industry, and its running process has complex characteristics such as strong nonlinearity, multivariable coupling, large time delay and random disturbance. The device mainly includes multiple units such as catalytic reforming, aromatic extraction, xylene separation, etc., and the units are interrelated, making the dynamic behavior of the entire device extremely complex.

[0003] At present, the modeling and parameter identification methods for the aromatic hydrocarbon combined device are mostly based on linear assumptions or simplified mechanism models. However, in actual production, the reaction rate, material balance and energy balance of the aromatic hydrocarbon combined device all show significant nonlinear characteristics, and the model based on linear assumptions is difficult to accurately capture the nonlinear dynamic behavior of the device under different operating conditions. At the same time, there are many random factors in the production process, such as fluctuations in raw material composition, equipment operation noise, external environmental disturbance, etc., and the traditional method lacks an effective processing mechanism for these random factors, resulting in a large deviation between the established model and the actual production process, and the model cannot provide reliable theoretical basis for the optimization operation, fault diagnosis and advanced control of the device.

[0004] In addition, the existing parameter identification methods often ignore the influence of random disturbances or only use simple filtering methods for processing, resulting in poor accuracy and robustness of the parameter identification results. When the operating conditions of the device change or are disturbed by random factors, the model parameters are prone to drift, leading to a decrease in the prediction ability and adaptability of the model, and seriously affecting its application effect in actual engineering. Therefore, it is of great practical significance to develop a modeling and parameter identification method that can accurately describe the nonlinear stochastic dynamic characteristics of the aromatic hydrocarbon combined device. SUMMARY

[0005] The present application aims to overcome the deficiencies in the prior art and provide a nonlinear stochastic dynamic model parameter identification method for an aromatic hydrocarbon combined device, to solve the problems that traditional methods cannot accurately describe the nonlinear stochastic dynamic characteristics of the device, and the parameter identification accuracy and robustness are poor, thereby providing reliable model support for the optimization operation and advanced control of the aromatic hydrocarbon combined device.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a nonlinear stochastic dynamic model parameter identification method for an aromatic hydrocarbon combined device, which uses a stochastic differential equation (SDE) to describe the continuous dynamic process of the aromatic hydrocarbon device:

[0007] dx(t) = f(x(t),u(t))dt + G(x(t))dW(t) + H(x(t))d(t)dt

[0008] wherein:

[0009] x(t) is the state vector;

[0010] u(t) is the control vector;

[0011] f(·) is a nonlinear drift term, which characterizes the deterministic nonlinear dynamic characteristics in the aromatic hydrocarbon device;

[0012] W(t) is a d-dimensional Wiener process, which is used to describe the continuous random noise of small disturbances, and its random law is known;

[0013] G(·) is a diffusion coefficient matrix, which is used to measure the strength and mode of action of continuous random disturbances;

[0014] H(·) is a disturbance correlation function, which reflects the relationship between the discrete or modelable random disturbance and the state of the device.

[0015] Preferably, the specific function form of f(·) and the internal parameters are usually unknown, and the dual-scale LSTM algorithm with mechanism constraints of the aromatic hydrocarbon combined device is used to determine them by collecting the input and output data of the device.

[0016] Preferably, the function structure and parameters of G(·) are unknown, and the multi-scale mechanism-constrained Bayesian variational inference algorithm is used to identify them in combination with the statistical data of random disturbances of the device.

[0017] Preferably, the function form and parameters of H(·) are unknown, and the function identification method based on piecewise linearization and recursive least squares is used to determine them by using the device running data containing disturbance occurrence, so that the model can reasonably reflect the impact of the disturbance.

[0018] Preferably, the identification algorithm steps of f(·) follow: input layer→attention mechanism→LSTM dynamic layer→mechanism constraint layer→output layer mechanism, and the specific steps are as follows:

[0019] Input layer: input vector X(t) = [x(t) T , u(t) T , c(t) T ] T , extend the traditional state x(t), control u(t) input, add catalyst state c(t), and adapt the long-period dynamic of the aromatic hydrocarbon device catalyst;

[0020] Adaptive attention mechanism: time-varying attention weight α(t), assigns importance to input variables according to unit dynamic characteristics:

[0021] α(t) = Softmax(W a tanh(W h h(t-1) + W x X(t)))

[0022] where h(t-1) is the LSTM hidden state, W a ,W h ,W x are learnable weights;

[0023] Constraint rules

[0024] Reaction unit: when the catalyst activity c1(t) decreases, automatically increase the attention weight of temperature x1(t) and feed hydrocarbon composition u1(t), and strengthen the dynamic capture of "reaction rate decay";

[0025] Separation unit: when the solvent ratio u2(t) is adjusted, increase the weight of column plate temperature x2(t) and solvent water content x3(t) to adapt to the rapid change of phase equilibrium;

[0026] LSTM dynamic layer - multi-scale time memory

[0027] Dual-time scale LSTM (Dual-Time LSTM) is used to process fast time scale (second-minute level) h fast (t) and slow time scale (hour-day level) h slow (t);

[0028] Dual-scale hidden state fusion formula:

[0029] where is the fusion weight matrix, and Concat(·) is the vector concatenation operation, which concatenates the fast and slow hidden states into a new vector according to the dimension;

[0030] Mechanism constraint layer - chemical prior regularization

[0031] The core mechanism of the aromatic hydrocarbon device is converted into a constraint condition for network output, and the constraint equation of the "reaction kinetics + phase equilibrium" principle is constructed:

[0032] Reaction kinetics constraint (catalytic reforming dehydrogenation reaction)

[0033] Given the dehydrogenation reaction rate r deH (t) satisfies the Arrhenius equation:

[0034]

[0035] Where k0 is the pre-exponential factor, E a As the activation energy, x naphthene (t) represents the cycloalkane concentration, and x represents the reactor bed temperature. temp (t);

[0036] At the network output layer, the nonlinear drift term that forces the model's prediction is forced is... satisfy:

[0037]

[0038] Where λ∈[0,1] is the mechanism constraint coefficient. The LSTM output is driven purely by data.

[0039] Phase equilibrium constraint (NRTL model for aromatic extraction)

[0040] In the aromatics extraction tower, the phase equilibrium relationship between the solvent and aromatics satisfies the NRTL model:

[0041]

[0042] Where γ i (t) is the activity coefficient, τ ji G ji For model parameters, x k (t) represents the mole fraction of component k at time t;

[0043] Transform the phase balance constraint into a loss function regularization term:

[0044] in N(·) represents the drift term of the separated unit predicted by the model, and N(·) is the "theoretical state change-phase equilibrium mapping" derived based on the NRTL model, which forces the drift term prediction of the separated unit to satisfy the phase equilibrium law.

[0045] Error Function and Parameter Optimization – Multi-Task Constraint Optimization

[0046] Multi-task loss function

[0047] Define a three-objective loss, integrating "dynamic fitting", "process index prediction", and "mechanism constraints":

[0048] Dynamic fitting loss (Based data-driven loss):

[0049] Predicted losses based on process parameters

[0050] Select key indicators (aromatic purity y) purity(t), product yield y yield (t)) to build prediction branches:

[0051]

[0052] Loss is defined as:

[0053] Mechanism constraint loss (Embedding chemical prior):

[0054] Regular term integrating reaction kinetics, phase equilibrium constraints, etc.

[0055] Wherein is the reaction rate constraint loss, is the phase equilibrium constraint loss.

[0056] Preferably, the parameter identification implementation process:

[0057] Step one, data preparation, collecting process data, analyzing data, catalyst state of aromatic hydrocarbon device, constructing data set

[0058] Step two, pre-training (embedding mechanism), freeze LSTM weight, only train "mechanism constraint layer", make the model comply with chemical law;

[0059] Step three, fine-tuning (dynamic fitting), unfreeze the whole network, optimize the double loss function with Adam algorithm, learning rate is set to 1e-4, iterate 200 rounds, make the model fit dynamic and mechanism at the same time;

[0060] Step four, online update, when catalyst replacement or feed mutation occurs, use the newly collected 1 hour data incremental learning to adapt to the new working condition.

[0061] The technical effect of the present application is: the present application proposes a complete nonlinear random dynamic modeling and parameter identification method for the complex characteristics of aromatic hydrocarbon combination device, such as strong nonlinearity, multivariable coupling and significant random disturbance, which effectively overcomes the limitations of traditional linear model or simplified mechanism model in describing the dynamic behavior of the device, and at the same time solves the problems of low modeling accuracy, poor robustness and weak adaptability of traditional technology in complex industrial processes, and lays a foundation for efficient optimization and intelligent control of industrial processes. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0063] The present application is a nonlinear stochastic dynamic model parameter identification method for an aromatic hydrocarbon combination device, specifically as follows:

[0064] Nonlinear stochastic dynamic model construction

[0065] The aromatic hydrocarbon combination device has strong nonlinear dynamic characteristics (such as catalytic reaction kinetics, phase equilibrium relationship) and multiple source random disturbances (such as feed fluctuation, catalyst activity attenuation), and the traditional linear model cannot accurately describe its dynamic behavior. Therefore, the present application uses stochastic differential equation (SDE) to construct a continuous dynamic model, and is discretized by stochastic Runge-Kutta method to adapt to the real-time control requirements of industrial processes. The stochastic differential equation (SDE) is used to describe the continuous dynamic process of the aromatic hydrocarbon device:

[0066] dx(t)=f(x(t),u(t))dt+G(x(t))dW(t)+H(x(t))d(t)dt

[0067] Wherein:

[0068] is the state vector (including reaction temperature, tower liquid level, catalyst activity, etc.);

[0069] is the control vector (furnace power, reflux ratio, etc.);

[0070] f(·) is a nonlinear drift term, which describes the deterministic nonlinear dynamic characteristics in the aromatic hydrocarbon device, such as reaction kinetics (relationship between catalytic reaction rate and temperature, reactant concentration, etc.), material balance (relationship between material flow in the tower and liquid level, flow rate, etc.). The specific function form (such as polynomial form, exponential form, etc.) and internal parameters (such as reaction rate constant, mass transfer coefficient, etc.) are usually unknown, and the present application patent determines them by collecting input and output data of the device (such as changing the control variable u(t), measuring the corresponding state change x(t)), and designing a double-scale LSTM algorithm that adapts to the mechanism constraints of the aromatic hydrocarbon combination device;

[0071] W(t) is a d-dimensional Wiener process (Brownian motion) for describing continuous random noise such as fluid perturbation, which follows a random rule that is known knowledge (such as mean value of 0, variance related to time increment, etc.).

[0072] G(·) is the diffusion coefficient matrix, which measures the strength and mode of action of continuous random disturbances. Its functional structure (such as matrix function form related to state x(t)) and parameters are unknown. This patent combines statistical data of random disturbances during device operation (long-term acquisition of state fluctuation data, analysis of random parts after excluding deterministic disturbances), and identifies them through the design of multi-scale mechanism-constrained Bayesian variational inference algorithm;

[0073] H(·) is the disturbance correlation function, which represents the relationship between discrete or modelable random disturbances such as feed composition mutations, sudden external disturbances, and device states. Its functional form (such as matrix mapping relationship) and parameters (such as disturbance transmission coefficients) are unknown. This patent uses device operation data containing such disturbances (records disturbance events such as feed fluctuations and corresponding state responses), and determines them through a function identification method based on piecewise linearization and recursive least squares, so that the model can reasonably reflect the impact of such disturbances.

[0074] Model parameter identification

[0075] (1) f(·) is the nonlinear drift term, and its identification algorithm steps follow: input layer → attention mechanism → LSTM dynamic layer → mechanism constraint layer → output layer mechanism, and the specific steps are as follows:

[0076] ① Overall architecture

[0077] Input layer: input vector X(t) = [x(t) T ,u(t) T ,c(t) T ] T , extending the traditional state x(t), control u(t) input, adding catalyst state c(t) (such as activity decay coefficient, carbon deposition amount, obtained through online monitoring or soft measurement), adapting to the long-period dynamics of aromatic device catalysts.

[0078] Adaptive attention mechanism: design time-varying attention weight α(t), which assigns importance to input variables according to unit dynamic characteristics:

[0079] α(t) = Softmax(W a ·tanh(W h h(t-1) + W x X(t)))

[0080] Where h(t-1) is the LSTM hidden state, W a ,W h ,W x are learnable weights.

[0081] Constraint rules:

[0082] Reaction unit (e.g. reforming reactor): when catalyst activity c1(t) drops, automatically increase attention weights on temperature x1(t), feed hydrocarbon composition u1(t), and strengthen dynamic capture of "reaction rate decay";

[0083] Separation unit (e.g. aromatic extraction column): when solvent ratio u2(t) is adjusted, increase weights on tray temperature x2(t), solvent water content x3(t), and adapt to rapid changes in phase equilibrium.

[0084] ② LSTM dynamic layer - multi-scale time memory

[0085] Use dual-time scale LSTM (Dual-Time LSTM) to process:

[0086] Fast time scale (seconds to minutes) h fast (t): capture fast-changing dynamics such as heating furnace temperature fluctuations, pump flow shocks;

[0087] Slow time scale (hours to days) h slow (t): capture slow-changing characteristics such as catalyst activity decay, product quality cumulative deviation.

[0088] Dual-scale hidden state fusion formula:

[0089] Where is the fusion weight matrix (learnable parameters), and Concat(·) is the vector concatenation operation, which concatenates the fast and slow hidden states into a new vector.

[0090] Adapt to the characteristics of aromatic device:

[0091] In catalytic reforming reaction, reactor bed temperature x temp (t) changes rapidly with heating furnace power u fuel (t) (seconds), while catalyst activity c cat (t) decays in days, and dual-time scale LSTM can accurately fit both types of dynamics.

[0092] ③ Mechanism constraint layer - chemical prior regularization

[0093] Convert the core mechanism of the aromatic device into a constraint condition for network output, and construct a constraint equation based on the principles of "reaction kinetics + phase equilibrium":

[0094] Reaction kinetics constraints (catalytic reforming dehydrogenation reaction)

[0095] Given the dehydrogenation reaction rate r deH (t) satisfies the Arrhenius equation:

[0096]

[0097] where k0 is a pre-exponential factor, E a is the activation energy, x naphthene (t) is the concentration of cycloalkanes, reactor bed temperature x temp (t).

[0098] At the network output layer, the nonlinear drift term of the model prediction is forced to satisfy:

[0099]

[0100] where λ ∈ [0, 1] is the mechanism constraint coefficient (initially set to 0.3, self- adaptive adjustment during training), is the pure data-driven LSTM output.

[0101] Phase equilibrium constraints (aromatic extraction NRTL model);

[0102] In the aromatic extraction column, the phase equilibrium relationship between the solvent (such as sulfolane) and the aromatic hydrocarbon satisfies the NRTL model:

[0103]

[0104] where γ i (t) is the activity coefficient, τ ji , G ji are model parameters, x k (t) is the mole fraction of component k at time t.

[0105] Convert the phase equilibrium constraint into a loss function regular term:

[0106] where represents the model-predicted separation unit drift term (such as tray temperature, flow rate change), N(·) is the "theoretical state change-phase equilibrium mapping" derived based on the NRTL model, which forces the drift term prediction of the separation unit to satisfy the phase equilibrium law.

[0107] ④ Error function and parameter optimization - multi-task constraint optimization;

[0108] Multi-task loss function, define three target losses, fusion "dynamic fitting" "process index prediction" "mechanism constraint":

[0109] Dynamic fitting loss (basic data-driven loss):

[0110] Process index prediction loss (strengthen industrial relevance):

[0111] Select key indicators (such as aromatic purity y purity (t), product yield y yield (t)), build a prediction branch:

[0112]

[0113] Loss is defined as:

[0114] Mechanism constraint loss (chemical engineering priori embedding):

[0115] Integrate the regular terms of reaction kinetics, phase equilibrium constraints, etc. such as:

[0116] Where is the reaction rate constraint loss, is the phase equilibrium constraint loss.

[0117] ⑤Parameter identification implementation process

[0118] Step one, data preparation, collect process data (temperature, flow) of aromatic hydrocarbon device, analyze data (aromatic hydrocarbon composition), catalyst state (activity), and build data set

[0119] Step two, pre-training (embedding mechanism), freeze LSTM weight, only train "mechanism constraint layer", quickly make the model comply with chemical law (such as reaction rate formula);

[0120] Step three, fine-tuning (dynamic fitting). Thaw the whole network, optimize the double loss function with Adam algorithm, set the learning rate to 1e-4, iterate 200 rounds, and let the model fit dynamic and mechanism at the same time;

[0121] Step four, online update, when catalyst replacement or feed mutation occurs, use the newly collected 1 hour data incremental learning to quickly adapt to new working conditions.

[0122] G(·) is the diffusion coefficient matrix, and the specific implementation of the multi-scale mechanism constraint Bayesian variational inference algorithm is as follows:

[0123] Discrete mathematical model

[0124] The core equation of diffusion coefficient identification is discretized by random Runge-Kutta method:

[0125]

[0126] Where:

[0127] k1=f(x(k),u(k))Δt+G(x(k))ΔW k

[0128] k2 = f(x(k) + k1, u(k)) At + G(x(k) + k1) AW k

[0129] At is the sampling period (e.g., 1 minute),

[0130] After discretization of the core equation for diffusion coefficient identification, the covariance of the state increment satisfies:

[0131]

[0132] Left side: Cov(x(k+1) - x(k) | x(k), u(k)), represents the covariance of the state increment x(k+1) - x(k) under the condition of knowing the current state x(k) and input u(k).

[0133] Right side: G(x(k)) is the diffusion coefficient matrix corresponding to the state x(k), G(x(k)) T is its transpose; similarly, G(x(k) + k1) is the diffusion coefficient matrix corresponding to the state x(k) + k1 and its transpose. Here k1 is the relevant term in the random Runge-Kutta discretization, and At is the time step for discretization, used to relate the continuous-time model and the discretely sampled state changes.

[0134] Likelihood function construction with multi-scale mechanism constraints

[0135] Multi-scale data decomposition

[0136] Fast scale data (t ∈ [0, T fast ): short sampling frequency, e.g., 1 second, containing furnace temperature fluctuations, pump flow shocks, etc., corresponding to G fast (x);

[0137] Slow scale data (t ∈ [0, T slow ): long sampling frequency, e.g., 1 hour, containing catalyst activity decay noise, corresponding to G slow (x).

[0138] Decomposition operator: realized by wavelet transform, fast scale component takes high frequency coefficient, slow scale component takes low frequency coefficient.

[0139] Mechanism constraint term design

[0140] According to the physical laws of the core unit of the aromatic hydrocarbon device, the constraint conditions of G(x) are constructed:

[0141] ① Catalytic reforming reactor: diffusion coefficient is positively correlated with temperature (temperature rise intensifies molecular thermal motion):

[0142]

[0143] where x temp is the reactor temperature, G ii is the diagonal element (main noise channel);

[0144] ② Aromatics extraction column: diffusion coefficient is inversely proportional to solvent ratio u solv (high solvent ratio suppresses phase equilibrium fluctuations):

[0145] G(x)≤c solv ·(u solv ) -1 ·I(c solv is an empirical constant)

[0146] ③ Catalyst unit: slow-scale diffusion coefficient monotonically increases with activity decay (lower activity, stronger noise sensitivity):

[0147] G slow (x)∝exp(-λc cat )(λ>0,c cat is catalyst activity)

[0148] Likelihood function with constraints combines multiscale data and mechanism constraints to construct Bayesian posterior probability:

[0149]

[0150] where: is the multiscale data set; is the likelihood function, based on Gaussian assumption of equation (1):

[0151]

[0152] (μ k is deterministic increment,∑ k (G) is defined by equation (1));

[0153] p(G) is the prior distribution, taken as a Gaussian-Wishart distribution (adapted to matrix parameter characteristics);

[0154] is the indicator function of mechanism constraints, equal to 1 when constraints are satisfied, otherwise 0.

[0155] Variational inference and online optimization solution

[0156] Variational approximation

[0157] Since the posterior probability is difficult to solve analytically, variational inference is used to approximate it with an easily handled approximate distribution q(G);

[0158] q(G)=q fast (G fast )·q slow (G slow )

[0159] (Multiscale factorization) where q fast q slow All distributions are low-rank Gaussian distributions to reduce computational complexity.

[0160] Objective function: Lower bound of evidence (ELBO)

[0161] The optimization objective is to maximize ELBO.

[0162]

[0163] The first term: data fitting term, estimated through Monte Carlo sampling;

[0164] The second term: regularization term, which constrains the deviation of q(G) from the prior;

[0165] The third term: Mechanism constraint term, which is approximated by a penalty function (introducing a negative infinity penalty when the constraint is violated);

[0166] To enhance the real-time performance of the online incremental learning process in the industry, a sliding window incremental update mechanism is designed:

[0167] ① Window size N win =100 (minute level), an update is triggered every 10 new data sets collected;

[0168] ② Incremental corrections are made to the mean and covariance of q(G):

[0169]

[0170] Where α = 0.9 is the forgetting factor. This is an estimate of the new data.

[0171] Algorithm Flow Summary

[0172] 1. Offline initialization: Collect historical data (1 month) and decompose it into fast / slow scale datasets; initialize the mean μ0 (based on the process manual) and covariance ∑0 (empirical value) of q(G);

[0173] 2. Online identification:

[0174] Step 1: Real-time acquisition of state increment Δx = x(k+1) - x(k), and estimation of sample covariance;

[0175] Step 2: Substitute into equation (1), maximize ELBO using the Adam algorithm, and update q(G);

[0176] Step 3: Check mechanism constraints, if violated, revise q(G) by projection operator;

[0177] Step 4: Output the optimal estimation of G(x) every hour

[0178] 3. Feedback correction: dynamically adjust a by combining the state estimation residual of extended Kalman filter (EKF) (reduce a when residual is large, accelerate convergence).

[0179] H(·) is the disturbance correlation function, which reflects the effect of discrete or modelable random disturbances such as sudden changes in feed composition, external disturbances, etc. on the state of the device. Its function form (such as what kind of matrix mapping relationship) and parameters (such as disturbance transmission coefficient) are unknown, and need to be determined by designing a function identification method based on piecewise linearization and recursive least squares using device operation data containing such disturbance occurrence (record the disturbance events such as feed fluctuation and the corresponding state response), so that the model can reasonably reflect the impact of such disturbances.

[0180] ① Disturbance event detection and sample extraction

[0181] Disturbance detection: use a simple threshold method to calculate the change rate of disturbance variable d(t):

[0182]

[0183] Where T s is the sampling period, when (θ is an empirical threshold, such as 0.5 t / h / s for feed flow), it is determined as a disturbance event, and the disturbance start time t0 is recorded.

[0184] Sample extraction: for each disturbance event, extract the disturbance sequence and response sequence:

[0185] D = [d(t0), d(t0+1),..., d(t0+N)]

[0186] X = [x(t0+τ), x(t0+τ+1),..., x(t0+τ+N)]

[0187] Where N is the data quantity, and τ is the time delay parameter (set according to process experience, such as feed impact on temperature τ = 5).

[0188] ② Piecewise linear model structure

[0189] A three-section linear model is used to describe the disturbance correlation function:

[0190]

[0191] Where:

[0192] d1, d2 are segment thresholds (set according to operating range, e.g. d1 = 70 t / h, d2 = 110 t / h); k1, k2, k3 are the slopes of each segment (characterize disturbance sensitivity);

[0193] b1, b2, b3 are intercept terms, satisfying continuity condition: k1d1 + b1 = k2d1 + b2, k2d2 + b2 = k3d2 + b3.

[0194] ③FFRLS (forgetting factor recursive least square) based parameter identification

[0195] Objective function: for each linear model, define fitting error:

[0196]

[0197] Recursion formula:

[0198] Parameter update:

[0199]

[0200] Gain matrix:

[0201]

[0202] Covariance matrix:

[0203]

[0204] where θ = [k, b] T is the parameter to be estimated, φ(k) = [d(k), 1] T is the regression vector.

[0205] ④Segmented identification: classify samples according to d ≤ d1, d1 < d ≤ d2, d > d2, and call FFRLS algorithm to get parameters of each segment respectively.

[0206] ⑤Simple online update mechanism

[0207] Update trigger: when model prediction error e = |x meas (k·d+b)| > ε (ε is allowed error, e.g. 2℃ for temperature) and appears continuously for 3 times, trigger update.

[0208] Parameter correction: use sliding window (window size M = 50) to save latest samples, re-run FFRLS algorithm, and only update parameters of segment where current operating condition is located (e.g. current d = 80 t / h, only update k2, b2).

[0209] Stability check: make sure the updated slope meets the process mechanism (e.g. k>0 means that increasing feed leads to increasing temperature), otherwise discard this update.

[0210] In summary, the present application proposes a complete nonlinear stochastic dynamic modeling and parameter identification method for the complex characteristics of strong nonlinearity, multivariable coupling and significant random disturbance of the aromatic hydrocarbon combination device, effectively overcomes the limitations of traditional linear models or simplified mechanism models in describing the dynamic behavior of the device, and at the same time, solves the problems of low modeling accuracy, poor robustness and weak adaptability of traditional technologies in complex industrial processes, and lays a foundation for efficient optimization and intelligent control of industrial processes.

[0211] It should be noted that in this paper, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.

[0212] The principles and implementation modes of the present application are described by applying specific examples in this paper, and the above examples are only used to help understand the method of the present application and its core idea. The above description is only the preferred embodiment of the present application. It should be pointed out that due to the limitation of language expression, there are infinite specific structures, and for ordinary skilled persons in the technical field, some improvements, refinements or changes can be made without departing from the principles of the present application, or the above technical features can be combined in an appropriate way; these improvements, refinements, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, should be regarded as the protection scope of the present application.

Claims

1. A method for parameter identification of nonlinear stochastic dynamic models for aromatic hydrocarbon complexes, characterized in that, Step 1: Construct a dynamic model of the device using stochastic differential equations. The equation is dx(t)=f(x(t),u(t))dt+G(x(t))dW(t)+H(x(t))d(t)dt, where x(t)∈R n Let u(t) ∈ R be the state vector. m For control vectors; Step 2: Determine the functional form and parameters of the nonlinear drift term f(·) using a dual-scale LSTM algorithm adapted to the mechanism constraints of aromatics devices. Integrate the activation energy and reaction heat mechanism information in aromatics reaction kinetics into the gating mechanism of the LSTM network. By learning and memorizing data at short time scales (second-level) and long time scales (hour-day-level) respectively, the output results of the two time scale LSTM networks are weighted and summed according to the degree of influence to obtain the estimated value of the f(·) parameter. Step 3: The structure and parameters of the diffusion coefficient matrix G(·) are identified by the Bayesian variational inference algorithm with multi-scale mechanism constraints. Based on the process fluctuation range of the device and combined with historical operating data, the variational parameters are analyzed at the time scales of minutes, hours and days to determine G(·) by defining the variational distribution and constructing an objective function containing data likelihood terms and prior distribution terms. Step 4: Determine the functional form and parameters of the discrete disturbance correlation function H(·) by using a piecewise linearization and recursive least squares method. Divide the historical fault data of the device into multiple time periods according to the fault type and state change characteristics. In each time period, assume that the state of the device and the discrete disturbance are linearly related. Use the recursive least squares algorithm and introduce a forgetting factor to gradually update the parameter estimates to obtain H(·).

2. The method for parameter identification of nonlinear stochastic dynamic models for aromatic hydrocarbon complexes according to claim 1, characterized in that, The identification algorithm for f(·) follows the steps of: input layer → attention mechanism → LSTM dynamic layer → mechanism constraint layer → output layer mechanism, and the specific steps are as follows: Input layer: Input vector X(t) = [x(t)] T ,u(t) T ,c(t) T ] T Expand the traditional state x(t) and control u(t) inputs, add a catalyst state c(t), and adapt to the long-cycle dynamics of catalysts in aromatics units; Adaptive attention mechanism: Time-varying attention weight α(t) assigns importance to input variables based on unit dynamic characteristics. α(t)=Softmax(W a ·tanh(W h h(t-1)+W x X(t))) Where h(t-1) is the LSTM hidden state, W a W h W x These are learnable weights; Constraints Reaction unit: When the catalyst activity c1(t) decreases, the attention weight of temperature x1(t) and feed hydrocarbon composition u1(t) is automatically increased to enhance the dynamic capture of "reaction rate decay". Separation unit: When the solvent ratio u2(t) is adjusted, the weights of the tray temperature x2(t) and solvent water content x3(t) are increased to adapt to rapid changes in phase equilibrium; LSTM Dynamic Layers – Multi-Scale Temporal Memory A dual-timescale LSTM is used to process fast timescales (seconds to minutes) separately. fast (t) and; slow timescale (hours-days) h slow (t); Dual-scale hidden state fusion formula: in This refers to the fusion of the weight matrix. Concat(·) is a vector concatenation operation that concatenates the fast and slow hidden states into a new vector according to their dimensions. Mechanism constraint layer - chemical prior regularization The core mechanism of the aromatics unit is transformed into constraints for the network output, and constraint equations based on the principle of "reaction kinetics + phase equilibrium" are constructed: Reaction kinetic constraints (catalytic reforming dehydrogenation reaction) The dehydrogenation reaction rate r is known. deH (t) satisfies the Arrhenius equation: Where k0 is the pre-exponential factor, E a As the activation energy, x naphthene (t) represents the cycloalkane concentration, and x represents the reactor bed temperature. temp (t); At the network output layer, the nonlinear drift term that forces the model's prediction is forced is... satisfy: Where λ∈[0,1] is the mechanism constraint coefficient. The LSTM output is driven purely by data. Phase equilibrium constraint (NRTL model for aromatic extraction) In the aromatics extraction tower, the phase equilibrium relationship between the solvent and aromatics satisfies the NRTL model: Where γ i (t) is the activity coefficient, τ ji G ji For model parameters, x k (t) represents the mole fraction of component k at time t; Transform the phase balance constraint into a loss function regularization term: in N(·) represents the drift term of the separated unit predicted by the model, and N(·) is the "theoretical state change-phase equilibrium mapping" derived based on the NRTL model, which forces the drift term prediction of the separated unit to satisfy the phase equilibrium law. Error Function and Parameter Optimization – Multi-Task Constraint Optimization Multi-task loss function Define a three-objective loss, integrating "dynamic fitting", "process index prediction", and "mechanism constraint": Dynamic fitting loss (Based data-driven loss): Predicted losses based on process parameters Select key indicators (aromatic purity y) purity (t), Product yield y yield (t)), construct the prediction branch: Loss is defined as: Mechanism-constrained loss (Embedded chemical a priori): Regularization terms that integrate constraints such as reaction kinetics and phase equilibrium. in For reaction rate constraint loss, This is the loss due to phase balance constraint. According to claim 5, the method for parameter identification of nonlinear stochastic dynamic models for aromatic hydrocarbon complexes is characterized by using stochastic differential equations (SDEs) to describe the continuous dynamic process of the aromatic hydrocarbon complex: dx(t)=f(x(t),u(t))dt+G(x(t))dW(t)+H(x(t))d(t)dt in: It is a state vector; For control vectors; f(·) is a nonlinear drift term that characterizes the deterministic nonlinear dynamic characteristics in the aromatics unit; W(t) is a d-dimensional Wiener process used to describe continuous random noise with small perturbations, and the random laws it follows are given knowledge. G(·) is the diffusion coefficient matrix, used to measure the intensity and mode of action of continuous random disturbances; H(·) is the disturbance correlation function, which reflects the effect of discrete or modelable random disturbances on the state of the device.

3. The method for identifying nonlinear stochastic dynamic model parameters for aromatic hydrocarbon complexes according to claim 2, characterized in that, Parameter identification implementation process: Step 1: Data preparation. Collect process data, analytical data, and catalyst status from the aromatics unit to construct a dataset. Step 2: Pre-training (embedding mechanism), freeze the LSTM weights, and train only the "mechanism constraint layer" to make the model conform to chemical laws; Step 3: Fine-tuning (dynamic fitting), unfreeze the entire network, optimize the double loss function using the Adam algorithm, set the learning rate to 1e-4, iterate for 200 rounds, and allow the model to fit both dynamics and mechanisms simultaneously; Step 4: Online update. When the catalyst is replaced or the feed changes abruptly, the system uses newly collected 1-hour data increments to learn and adapt to the new operating conditions.

4. The method for identifying nonlinear stochastic dynamic model parameters for aromatic hydrocarbon complexes according to claim 3, characterized in that, Parameter identification implementation process: Step 1: Data Preparation. Collect process data, analytical data, and catalyst status from the aromatics unit to construct a dataset. Step 2: Pre-training (embedding mechanism), freeze the LSTM weights, and train only the "mechanism constraint layer" to make the model conform to chemical laws; Step 3: Fine-tuning (dynamic fitting), unfreeze the entire network, optimize the double loss function using the Adam algorithm, set the learning rate to 1e-4, iterate for 200 rounds, and allow the model to fit both dynamics and mechanisms simultaneously; Step 4: Online update. When the catalyst is replaced or the feed changes abruptly, the system uses newly collected 1-hour data increments to learn and adapt to the new operating conditions.

Images