Refining device intelligent prediction method based on physical information neural network

Abstract

The invention discloses a refining device intelligent prediction method based on a physical information neural network, and belongs to the technical field of intelligent prediction, and the method comprises the steps: obtaining the real-time working condition data of a refining device, and extracting a spatial-temporal feature vector and a device region identifier from the real-time working condition data; inputting the spatio-temporal feature vector into a physical guidance attention structure to obtain a spatio-temporal attention weight; determining an activated expert network set from expert network sets according to the combination of the space-time attention weight and the device area identifier; inputting the spatial-temporal feature vector into the activated expert network set to obtain a local predicted value set and a physical residual component set; inputting the local predicted value set and the physical residual component set into a Bayesian output structure to obtain a predicted mean value and a predicted standard deviation after calibration; and outputting the prediction mean value and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty index. The method has high precision and traceability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent prediction, in particular to a refining device intelligent prediction method based on a physical information neural network. BACKGROUND

[0002] Physical information neural network (PINN) has developed rapidly in the field of industrial modeling in recent years. By embedding physical equations into the loss function of neural network, the fusion of data-driven and mechanism constraint is realized, and the generalization and interpretability better than pure data model are shown in the scenes of fluid mechanics, heat and mass transfer, reaction engineering, etc. Especially in the field of energy and chemical industry, PINN is tried to be used for tasks such as reactor temperature field reconstruction, pipeline pressure drop prediction, catalyst activity inversion, and gradually moves from academic simulation to engineering prototype verification, becoming a key technology path of "white box AI" in intelligent factory.

[0003] However, PINN still faces three major bottlenecks in refining device prediction: local precision is insufficient; traditional PINN applies uniform physical constraints to the whole domain, which is difficult to focus on high gradient areas such as reactor hot spots and phase change interfaces, resulting in distorted prediction at key positions; mechanism adaptability is poor; a single network structure cannot adapt to the coupling of multiple physical fields (such as reaction-heat-flow) and regional heterogeneity in the device, and forced unified modeling causes mechanism conflicts; risk is not quantifiable; the output is only point estimation, lacking uncertainty indicators consistent with physics, operators cannot judge the credibility of the prediction, and it is difficult to support safety decisions. The above problems directly restrict the depth and reliability of PINN in complex industrial scenarios. SUMMARY

[0004] In view of the above problems, the present application is proposed.

[0005] Therefore, the technical problem solved by the present application is: how to realize local high-precision focusing of multiple physical fields in refining devices, adaptive mechanism modeling of heterogeneous regions, physical consistency constraint and uncertainty quantization output of prediction results by physical guided attention mechanism, regional mechanism expert network and Bayesian residual calibration structure, while suppressing mechanism conflicts and high gradient area prediction drift caused by global uniform modeling.

[0006] To solve the above technical problems, the present application provides the following technical scheme: a refining device intelligent prediction method based on a physical information neural network, comprising: acquiring real-time working condition data of a refining device, extracting a space-time feature vector and a device region identifier from the real-time working condition data; inputting the space-time feature vector into a physical guided attention structure to obtain a space-time attention weight; determining an activated expert network set from an expert network set according to the combination of the space-time attention weight and the device region identifier, each expert network corresponding to a region and a physical mechanism combination; inputting the spatio-temporal feature vector into the activated expert network set to obtain a local prediction value set and a physical residual component set; inputting the local prediction value set and the physical residual component set into a Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; outputting the prediction mean and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty index.

[0007] As a preferred scheme of the intelligent prediction method for a refining device based on a physical information neural network, the extraction of the spatio-temporal feature vector and the device region identifier includes: pulling raw data streams output by a sensor node at a 1-second cycle from a refining device distributed control system and an industrial Internet of Things platform; performing timestamp alignment and linear interpolation of missing values on the raw data streams to obtain a complete time series data set; mapping the time series data set to a preset region coding table according to equipment position number and spatial coordinates to generate a device region identifier; performing sliding window slicing and Min-Max normalization on the time series data set to generate a spatio-temporal feature vector sequence with a length of 10; pairing the spatio-temporal feature vector sequence with the corresponding device region identifier to form a neural network input sample pair.

[0008] As a preferred scheme of the intelligent prediction method for a refining device based on a physical information neural network, the obtaining of the spatio-temporal attention weight includes: inputting the spatio-temporal feature vector into a physical residual calculation layer to obtain a physical residual tensor according to a preset control equation; performing gradient operator operation on the physical residual tensor to obtain a physical residual gradient tensor; inputting the physical residual gradient tensor into a nonlinear mapping layer to obtain an initial attention weight tensor; performing normalization operation on the initial attention weight tensor to obtain the spatio-temporal attention weight.

[0009] As a preferred scheme of the intelligent prediction method for a refining device based on a physical information neural network, the determination of the activated expert network set includes: splicing the spatio-temporal attention weight and the device region identifier into a joint routing feature vector; inputting the joint routing feature vector into a gating network to obtain an expert activation probability distribution; According to the expert activation probability distribution, an expert network with a probability value greater than a preset threshold is selected from the expert network set to form the activated expert network set.

[0010] As a preferred scheme of the intelligent prediction method for a refining device based on a physical information neural network, the obtaining of the local prediction value set and the physical residual component set comprises: The spatio-temporal feature vector is input into each expert network in the activated expert network set in parallel to obtain a local prediction value output by each expert network; The local prediction value of each expert network is input into the embedded physical equation calculation layer thereof to obtain a corresponding physical residual component; The local prediction values output by all the expert networks are collected into the local prediction value set, and all the physical residual components are collected into the physical residual component set.

[0011] As a preferred scheme of the intelligent prediction method for a refining device based on a physical information neural network, the obtaining of the prediction mean value and the calibrated prediction standard deviation comprises: The local prediction value set is input into an uncertainty estimation layer to obtain an initial prediction standard deviation; The physical residual component set is input into a residual calibration layer to generate a residual scaling factor; The initial prediction standard deviation is multiplied by the residual scaling factor to obtain the calibrated prediction standard deviation, and a weighted average value of the local prediction value set is taken as the prediction mean value.

[0012] As a preferred scheme of the intelligent prediction method for a refining device based on a physical information neural network, the outputting of the prediction mean value and the calibrated prediction standard deviation comprises: The prediction mean value and the calibrated prediction standard deviation are packaged into a structured prediction output object, the structured prediction output object comprising a numerical value, a time stamp, a spatial position, an expert activation identifier and an uncertainty level label; The structured prediction output object is pushed to a refining device operation decision interface as a rolling prediction result of a target process parameter and a corresponding uncertainty index.

[0013] The present application provides an intelligent prediction system for a refining device based on a physical information neural network.

[0014] To solve the above technical problems, the present application further provides the following technical scheme: an intelligent prediction system for a refining device based on a physical information neural network, comprising: a data perception module configured to acquire real-time working condition data of the refining device and extract a spatio-temporal feature vector and a device region identifier from the real-time working condition data. An attention guiding module is configured to input the spatio-temporal feature vector into a physical guiding attention structure to obtain a spatio-temporal attention weight; An expert routing module is configured to determine an activated expert network set from an expert network set according to a combination of the spatio-temporal attention weight and the device region identifier, each expert network corresponding to a region and a physical mechanism combination; A mechanism reasoning module is configured to input the spatio-temporal feature vector into the activated expert network set to obtain a local prediction value set and a physical residual component set; A credible calibration module is configured to input the local prediction value set and the physical residual component set into a Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; A decision output module is configured to output the prediction mean and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty indicator.

[0015] The application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the intelligent prediction method for refining device based on physical information neural network when executing the computer program.

[0016] The application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the intelligent prediction method for refining device based on physical information neural network.

[0017] The beneficial effects of the present application: the present application constructs a three-layer coupled architecture of "physical attention guidance - regional mechanism expert - residual calibration output", which systematically breaks through the local distortion, mechanism conflict and risk black box bottleneck of traditional physical information neural network in refining device prediction: through the dynamic focusing of high gradient physical mismatch area by space-time attention weight, the modeling accuracy of key parts such as reactor and heat exchanger is significantly improved; through the device region identification to guide the expert network to activate adaptively according to "region-mechanism" combination, the heterogeneous physical field modeling of reaction dynamics, heat transfer, gas-liquid equilibrium and other heterogeneous physical fields is realized, and the parameter coupling error is avoided; through the embedded physical equation synchronous output residual component, and driving the Bayesian layer to dynamically scale the prediction standard deviation, the uncertainty index is directly linked with the physical consistency, and the risk controllable output of "the more inaccurate, the more cautious" is realized; finally, through the structured packaging and decision interface pushing, the prediction mean, calibration standard deviation, expert source, spatial position and other metadata are integrated to support the operators and control system to implement the hierarchical decision of "low-risk automatic optimization and high-risk manual intervention", so as to realize the paradigm transition of refining device key parameter prediction from "black box point estimation" to "transparent risk quantification" under the premise of guaranteeing physical interpretability and industrial safety, and provide a prediction engine with high precision, strong generalization, traceability and actionability for intelligent factory. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 The overall flowchart of a refining device intelligent prediction method based on a physical information neural network provided by an embodiment of the present application is shown in the figure. Figure 2 The risk pushing flowchart of a refining device intelligent prediction method based on a physical information neural network provided by an embodiment of the present application is shown in the figure. Figure 3 The safety alarm flowchart of a refining device intelligent prediction method based on a physical information neural network provided by an embodiment of the present application is shown in the figure. Figure 4 The computer device diagram of a refining device intelligent prediction method based on a physical information neural network provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0020] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0021] Embodiment 1, reference Figure 1 For the first embodiment of the present application, the embodiment provides a physical information neural network-based intelligent prediction method for a refining device, comprising: S1: acquiring real-time working condition data of the refining device, extracting a space-time feature vector and a device region identifier from the real-time working condition data; S2: inputting the space-time feature vector into a physical guided attention structure to obtain a space-time attention weight; S3: determining an activated expert network set from an expert network set according to a combination of the space-time attention weight and the device region identifier, each expert network corresponding to a region and a physical mechanism combination; S4: inputting the space-time feature vector into the activated expert network set to obtain a local prediction value set and a physical residual component set; S5: inputting the local prediction value set and the physical residual component set into a Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; S6: outputting the prediction mean and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty index.

[0022] It should be noted that during the operation of the refining device, the reactor, the fractionating column, the heat exchanger and other equipment are strongly coupled in multiple physical fields, and parameters such as temperature, pressure, components and flow rate evolve dramatically and nonlinearly in the space and time dimensions. Traditional data-driven models are prone to distortion in local regions, and mechanism models are difficult to update online due to parameter time variation and boundary ambiguity. At the same time, the long-term operation of the device is significantly affected by catalyst deactivation, coking and corrosion, and the working condition drift is significant. A single model is difficult to balance accuracy and generalization. In addition, industrial decision-making requires prediction results with reliability evaluation. Existing AI models generally lack uncertainty quantification capability, resulting in uncontrollable operation risks.

[0023] Therefore, for the above three core problems of low prediction accuracy, poor mechanism adaptation, and unknown risk, through steps S1-S6, a neural network architecture is constructed that fuses physical guiding attention, multi-mechanism expert network, and Bayesian calibration: automatic focusing modeling of high gradient / strong nonlinear regions is achieved; expert adaptive activation of different physical mechanism regions is achieved; and confidence interval output of the prediction result with physical residual correction is achieved, thereby improving the prediction accuracy, operating condition adaptability, and operation decision safety of the key parameters of the refining device under the premise of ensuring physical consistency.

[0024] Embodiment 2, refer to Figures 1-3 The second embodiment of the present application provides a neural network-based intelligent prediction method for a refining device.

[0025] S1: Obtain real-time operating condition data of the refining device, and extract a space-time feature vector and a device region identifier from the real-time operating condition data; It should be noted that the refining device is a high-temperature and high-pressure, multi-phase flow strongly coupled system, and its operating state is affected by multiple physical fields such as reaction, heat transfer, and flow, and continuously drifts with catalyst deactivation, coking, and changes in raw materials. The device has a large space, many measuring points, and significant differences in regional mechanisms, and data is easily disturbed or missing. Pure data-driven models are prone to distortion in key regions and lack of prediction confidence output. Therefore, a prediction architecture that fuses physical attention, regional expert network, and Bayesian calibration needs to be constructed to achieve local focusing, mechanism adaptation, and risk quantification, and to improve prediction accuracy and decision reliability.

[0026] Specifically, the extraction of the space-time feature vector and the device region identifier in step S1 includes the following sub-steps S11-S15: S11: From the distributed control system and the industrial Internet of Things platform of the refining device, pull the raw data stream output by multiple sensor nodes at a 1-second cycle; Specifically, the raw data stream is subscribed from the DCS history server through the OPC UA protocol, the subscription tag list is preloaded according to the device instrument index table, the collection period is fixed at 1000 milliseconds, and the collection range covers sensor nodes in reaction, separation, and heat exchange process sections.

[0027] Specifically, each sensor node is configured with an independent data buffer queue. If no new data is received within a certain period, the last period value is retained and the timestamp is marked as not updated. When there is no update for 5 consecutive periods, the "node offline" state is triggered, and the node data is skipped in the subsequent steps and does not participate in feature splicing.

[0028] Further, in the catalytic cracking riser region, the double-channel acquisition mechanism is enabled for the temperature measuring point TIC-101 and the density measuring point DENS-102: the main channel acquires data at a period of 1 second, the standby channel acquires data at a period of 200 milliseconds and stores independently; when the main channel data is marked as "interpolation" or "offline", the standby channel automatically switches to the latest valid value to ensure the continuity of the key region input.

[0029] Specifically, the data pulling process embeds a CRC check mechanism, each frame of data is attached with a 16-bit check code, and the receiving end discards the frame and records the error count when the verification fails; when the cumulative error exceeds the threshold, the communication link self-diagnosis is triggered to avoid dirty data entering the preprocessing process.

[0030] S12: performing timestamp alignment and missing value linear interpolation on the original data stream to obtain a complete time series dataset; Specifically, taking the DCS master clock as the reference, time alignment within ±50 milliseconds is performed on all node data, and data points exceeding the range are considered invalid, and linear interpolation is used to fill in the adjacent valid points before and after, and the interpolated data is marked with a "_I" suffix.

[0031] Further, in the hydrogenation reactor region, if the hydrogen flow FIC-301 and the raw oil flow FIC-302 are missing at the same time, the system calls the total flow at the reactor inlet and the preset hydrogen oil ratio to calculate the replacement value, and the calculation result is marked as "_C", and the ratio parameter used for calculation is recorded for subsequent physical residual check to exclude this point.

[0032] S13: mapping the time series dataset to a preset region coding table according to the equipment position number and the spatial coordinates to generate a device region identifier; Specifically, the region coding table is stored in the local configuration file, and each row contains three fields of equipment position number prefix, spatial coordinate boundary, and region code; the system matches the smallest spatial bounding box according to the installation position coordinates bound by the sensing node, and outputs the corresponding region code.

[0033] Further, in the atmospheric and vacuum tower region, five-level region codes are automatically generated according to the tray number TRAY-XX and the radial installation radius, such as "CDU-L12-R3", indicating the 12th tray and the 3rd radial partition; the code serves as the basis for subsequent expert network routing to ensure that different tower layers activate different mass transfer or fluid dynamics experts.

[0034] S14: performing sliding window slicing and Min-Max normalization on the time series dataset to generate a time-space feature vector sequence with a length of 10; Specifically, the sliding window length is fixed at 10 sampling points, the step is 1, each window independently executes Min-Max normalization, and the normalization reference value is taken from the past 8-hour running extreme value of the parameter. The reference is updated every 30 minutes.

[0035] Further, independent normalization intervals are set for different physical quantities: temperature [50, 600], pressure [0.05, 5.0], flow [0, maximum design value x 1.1], and concentration [0, 100], to avoid feature weight imbalance caused by dimensional differences, and the normalized value range is unified to [0, 1].

[0036] Further, in the overhead region of the fractionating column, the "overhead temperature", "reflux ratio" and "light component concentration" are calculated first, and if the correlation coefficient between the features is greater than 0.8, the first principal component is retained by principal component projection, then normalized, compressed and coupled to enhance the expression of the feature.

[0037] S15: Pair the spatio-temporal feature vector sequence with the corresponding device region identifier to form a neural network input sample pair.

[0038] Specifically, each sliding window generates an input sample pair, which contains a 10xN-dimensional feature vector and a region encoding string. N is the number of valid sensor nodes in the window. If the number of nodes is less than the preset value, zero is appended to the end and the effective length is marked.

[0039] Further, in the heat exchange network region, if the sample is derived from "shell inlet temperature TE-401", "tube pressure drop PDIC-402" and "medium flow rate FIC-403", the region identifier "HX-SHELL-IN" is bound, and the physical mechanism label "HEAT_XFER_FLUID" is attached, which is used for subsequent gate network to identify and activate "heat transfer fluid dynamics expert".

[0040] S2: Input the spatio-temporal feature vector into the physical guided attention structure to obtain the spatio-temporal attention weight; It should be noted that the physical field evolution in the refining and chemical plant has strong nonlinearity and spatial heterogeneity. Traditional neural networks lack the ability to perceive local high gradients or physical mismatch regions, which can easily lead to prediction bias accumulation. Simply relying on data fitting cannot guide the model to focus on key regions with inconsistent physics, and fixed weight physical constraints are difficult to adapt to dynamic changes in operating conditions. Therefore, a physical guided attention mechanism is introduced to dynamically generate spatio-temporal attention weights based on physical residual gradients, so that the network can automatically focus on regions with conservation imbalance or rapid changes during inference, thereby improving local modeling accuracy and physical consistency without increasing model complexity, and providing high-resolution regional guidance signals for subsequent expert routing and uncertainty calibration.

[0041] Specifically, the spatio-temporal attention weight obtained in step S2 includes the following sub-steps S21 to S24: S21: inputting the spatio-temporal feature vector into a physical residual calculation layer to obtain a physical residual tensor according to a preset control equation; Specifically, the physical residual calculation layer calculates the imbalance between the density, velocity and time change rate in the input feature according to the principle of mass conservation, and outputs residual values aligned with the input spatial dimension, each position representing the deviation degree of the mass conservation equation at the point.

[0042] Further, in the catalytic cracking reaction zone, the physical residual calculation layer synchronously calculates the energy conservation residual and the component reaction kinetics residual, respectively generates independent residual channels according to the temperature change, heat conduction, reaction heat release, and component concentration change, convective transport and chemical reaction rate, and splices them into a multi-channel tensor output.

[0043] S22: performing gradient operator operation on the physical residual tensor to obtain a physical residual gradient tensor; Specifically, the gradient operator calculates the change rate of the physical residual along the three spatial directions and the time direction respectively, adopts a central difference method, that is, uses the difference value of adjacent points before and after to divide the sampling interval to obtain the partial derivative of four directions, and composes a gradient tensor with the same dimension as the original residual tensor.

[0044] Further, if the device has no structural change in a certain spatial direction (such as an axial reactor without radial gradient), the gradient in this direction is forced to be zero; before calculation, the residual tensor is subjected to three-dimensional median filtering with a window size of 3x3x3 to suppress the interference of measurement noise on gradient calculation.

[0045] S23: inputting the physical residual gradient tensor into a nonlinear mapping layer to obtain an initial attention weight tensor; Specifically, the nonlinear mapping layer inputs the physical residual gradient module length into a nonlinear function with trainable parameters, and outputs a positive value as the initial attention intensity, and the function form ensures that the output is always greater than zero, and the greater the gradient, the higher the output value.

[0046] Further, the nonlinear mapping layer performs the following transformation:

[0047] wherein, is a single scalar element value in the initial attention weight tensor, is a trainable scaling factor, the initial value is set to 1.0, and is used to adjust the amplification multiple of the gradient on the attention, is the physical residual gradient module length, which is obtained by square root of the sum of squares of the partial derivatives in the spatial and time directions, The trainable bias is initially set to 0.0, which is used to adjust the baseline of attention activation. The nonlinear function is defined as the composition of natural logarithm and exponential function to avoid negative weights.

[0048] Furthermore, to prevent the weight from abnormally rising due to the excessively large local gradient, an upper limit truncation mechanism is set: if the mapped value exceeds the preset threshold 5.0, 5.0 is forced to be taken, and the original value is retained in other cases, ensuring the numerical stability.

[0049] S24: performing a normalization operation on the initial attention weight tensor to obtain the spatio-temporal attention weight; Specifically, the normalization operation is performed on the initial attention weight of each spatial position along the time dimension, so that the weight sum of the position at all time steps is 1, the relative strength relationship between spaces is preserved, and the subsequent expert network can focus on the key time period according to the time distribution.

[0050] Further, in the heat exchanger region, the local maximum and minimum normalization is independently performed for each spatial position: the maximum value minus the minimum value in all time steps of the position is used as the denominator, the current value minus the minimum value is used as the numerator, and a minimum value is added to prevent division by zero, so that the output value range is compressed to [0, 1], and the local hotspot change is highlighted.

[0051] S3: determining an activated expert network set from the expert network set according to the combination of the spatio-temporal attention weight and the device region identifier, each expert network corresponding to a region and a physical mechanism combination; It should be noted that the dominant physical mechanisms in different regions of the refining and chemical device are different, and a single model cannot consider multiple types of equation constraints such as reaction kinetics, phase equilibrium, heat transfer, and conduction. If the parameters are globally shared, mechanism conflicts and prediction distortion will occur. Therefore, a multiple expert network set needs to be constructed, each expert network is bound to a specific region and physical mechanism combination, and the adaptive expert is activated according to the spatio-temporal attention weight and the region identifier, so as to realize "regional and mechanism" precise modeling, avoid mismatch between model structure and physical reality, and improve the cross-working-condition generalization ability and local prediction reliability.

[0052] Specifically, the determination of the activated expert network set in step S3 includes the following sub-steps S31 to S33: S31: concatenating the spatio-temporal attention weight and the device region identifier into a joint routing feature vector; Specifically, the device region identifier is a preset encoding string, which is first converted into a one-hot encoding vector, and then concatenated with the spatio-temporal attention weight in the feature dimension to form an enhanced routing feature, which is used to express "which regions are physically inconsistent at which time points".

[0053] Further, in the atmospheric-vacuum tower region, the device region identifier contains the tower layer number and the radial partition, and after being converted into a two-dimensional position embedding vector, the spatial attention weight is weighted and added to generate a space-aware routing feature.

[0054] S32: input the joint routing feature vector into the gating network to obtain an expert activation probability distribution; Specifically, the gating network is a two-layer fully connected structure, the output dimension of the first layer is twice the total number of experts, the activation function is ReLU, the output dimension of the second layer is equal to the total number of experts, and the activation function is Softmax. The output value represents the probability of each expert being activated.

[0055] Further, the input layer of the gating network is additionally provided with a learnable regional bias vector, different regions correspond to different bias initial values, for example, the bias of the reaction zone is initialized to +0.5, and the bias of the separation zone is initialized to -0.2. The guide network tends to activate the experts in the corresponding region at the initial stage.

[0056] Further, if the activation frequency of a certain expert in a specific region is lower than 5% historically, a penalty term -1.0 is applied to the output log probability of the expert before Softmax, reducing the probability of being selected, and achieving expert load balancing and resource optimization.

[0057] S33: According to the expert activation probability distribution, select the expert network with a probability value greater than a preset threshold from the expert network set to form the activated expert network set; Specifically, the preset threshold is 0.1, if the activation probability of a certain expert is greater than 0.1, it will be added to the activated set, ensuring that at least one expert is activated each time, and the maximum number of activated experts does not exceed 3, avoiding waste of computing resources.

[0058] Further, during the operation of the catalytic cracking device, if it is detected that the spatial-temporal attention weight in the "regenerator temperature" region is continuously higher than 0.8, the "high-temperature oxidation reaction expert" is forced to be activated, even if its probability is slightly lower than the threshold, to ensure that the key region is not ignored.

[0059] S4: input the spatial-temporal feature vector into the activated expert network set to obtain a local prediction value set and a physical residual component set; It should be noted that the activated expert network needs to independently output the local prediction under the premise of maintaining the physical constraint, to avoid error propagation caused by mechanism coupling. If only the prediction value is output and the physical consistency check is ignored, the core advantage of PINN will be lost. Therefore, each expert synchronously calculates the residual component of the embedded control equation while outputting the local prediction value, forming a "prediction-residual" pair output, which not only guarantees the physical reasonableness of local modeling, but also provides a quantitative basis for subsequent uncertainty calibration, realizing the closed-loop cooperation of data-driven and mechanism constraint.

[0060] Specifically, obtaining the set of local prediction values and the set of physical residual components in step S4 includes the following sub-steps S41-S43: S41: input the spatiotemporal feature vector into each expert network in the set of activation expert networks in parallel to obtain the local prediction value output by each expert network; Specifically, each expert network is a three-layer fully connected structure, the input layer receives the spatiotemporal feature vector, the number of hidden layer neurons is 128, the activation function is GELU, the number of output layer neurons is the number of target process parameters, there is no activation function, and continuous values are directly output as local prediction values.

[0061] Further, the "thermal reaction kinetics expert" activated in the hydrogenation reactor area has a network structure embedded with reactor axial position coding, so that different axial positions output different conversion rate predictions under the same inlet conditions, enhancing the spatial resolution capability.

[0062] Still further, if the prediction error of a certain expert network in the current area is continuously higher than the threshold value in history, a decay coefficient of 0.9 is applied to its output in this inference to reduce its weight in subsequent fusion, realizing dynamic confidence adjustment.

[0063] S42: input the local prediction value of each expert network into its embedded physical equation calculation layer to obtain the corresponding physical residual component; Specifically, the "thermal reaction kinetics expert" has an Arrhenius reaction rate equation embedded therein, which calculates the theoretical reaction rate according to the predicted temperature and concentration values, and compares it with the predicted component change rate to obtain the reaction kinetics residual component.

[0064] Further, the "steady-state heat transfer expert" has a Fourier heat conduction law embedded therein, which calculates the theoretical heat flux density according to the predicted temperature field and material thermal conductivity, and compares it with the predicted energy change rate to obtain the energy conservation residual component.

[0065] Still further, if the physical equation embedded in a certain expert network involves an input feature (such as catalyst activity) that is not input, the historical moving average value is called as a substitute input, and after the residual is calculated, it is marked as "containing estimated parameters" for subsequent weight reduction when uncertainty is calibrated.

[0066] S43: collect the local prediction values output by all expert networks into the set of local prediction values, and collect all physical residual components into the set of physical residual components; Specifically, the set of local prediction values is stored in the form of a list, each element containing an expert identifier, a prediction value vector, and a spatial position index; the set of physical residual components is stored in the same order as the corresponding residual vector, ensuring that the subsequent steps can be processed in alignment with the experts.

[0067] Further, outlier filtering is performed on the local prediction values before output: if a certain expert prediction value deviates from the mean of other experts by more than three standard deviations, it is marked as an "outlier prediction" and does not participate in the final fusion, only its residual is retained for uncertainty analysis.

[0068] S5: input the set of local prediction values and the set of physical residual components into the Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; It should be noted that if the refining device prediction model only outputs point estimates, the operation decision will lack risk perception ability and is easy to cause misjudgment. Although the traditional Bayesian network can quantify the uncertainty of the model, it cannot perceive the systematic bias caused by physical mismatch. Therefore, the physical residual information needs to be fused in the output layer, and the prediction standard deviation is dynamically modulated by the residual scaling factor, so that the uncertainty output is linked to the physical consistency - the larger the physical residual, the higher the calibrated standard deviation, and the less reliable the prediction, thereby providing an "risk-labeled" intelligent output for industrial operation and improving the system safety margin and human-machine cooperation efficiency.

[0069] Specifically, in step S5, the prediction mean and the calibrated prediction standard deviation are obtained, including the following sub-steps S51 to S53: S51: input the set of local prediction values into the uncertainty estimation layer to obtain an initial prediction standard deviation; Specifically, the uncertainty estimation layer adopts a Monte Carlo Dropout structure, and in the inference stage, the set of local prediction values is subjected to 5 times of random dropout forward propagation, and the standard deviation of the output value is calculated as the initial prediction standard deviation.

[0070] Further, the uncertainty estimation layer is a variational Bayesian fully connected layer, the weights of which are subject to Gaussian distribution, a plurality of prediction branches are generated by reparameterization sampling, and the variance of the output prediction value is taken as the initial prediction standard deviation.

[0071] Further, if a certain expert network is marked as an "outlier prediction" or "contains estimated parameters", the weight of its prediction value is reduced by 50% when calculating the standard deviation, so as to avoid the interference of abnormal values on the overall uncertainty estimation.

[0072] S52: input the set of physical residual components into the residual calibration layer to generate a residual scaling factor; Specifically, the residual calibration layer takes the absolute value of each physical residual component and inputs it into a single-layer neural network, the activation function is Softplus, and the output value is taken as the local scaling factor of the corresponding area of the expert, and then the global residual scaling factor is obtained by weighted average according to the expert activation probability.

[0073] Further, in the reactor area, if the absolute value of the energy conservation residual error component exceeds 0.5, the scaling factor is forced to be 1.5, indicating that the prediction reliability of the area should be adjusted downward; if the residual error is less than 0.1, the scaling factor is set to 0.8, indicating that the physical consistency is good.

[0074] Further, the residual calibration layer introduces a sliding historical residual mean as a reference. If the current residual deviates from the historical mean by more than twice the standard deviation, the scaling factor is linearly enlarged to realize dynamic anomaly perception.

[0075] S53: multiply the initial prediction standard deviation by the residual scaling factor to obtain the calibrated prediction standard deviation, and take the weighted average of the local prediction value set as the prediction mean; Specifically, the weights of the weighted average come from the expert activation probability output by the gating network, ensuring that high-confidence experts contribute more to the prediction mean.

[0076] Further, if the calibrated prediction standard deviation exceeds the preset safety threshold 1.2, a "high uncertainty" label is appended to the output for the operator or upper control system to trigger manual review or conservative strategy.

[0077] S6: output the prediction mean and the calibrated prediction standard deviation as the target process parameter prediction result and uncertainty indicator.

[0078] It should be noted that intelligent prediction of refining and chemical plants is difficult to support operation decisions in safety-sensitive scenarios if it only outputs a single value. The operator needs to know the prediction confidence simultaneously to determine whether to adopt, intervene or review. Therefore, the prediction mean and the calibrated standard deviation need to be packaged as a structured output with a risk label, and associated with metadata such as spatial location, activated expert, and residual source, and pushed to the control or human-machine interface to realize the "prediction value + uncertainty + explainability" trinity delivery, so that intelligent prediction is upgraded from "black box output" to "transparent decision support", improving system robustness and industrial applicability.

[0079] Specifically, the output of the prediction mean and the calibrated prediction standard deviation in step S6 includes the following sub-steps S61 to S62: S61: package the prediction mean and the calibrated prediction standard deviation into a structured prediction output object, including numerical value, timestamp, spatial location, expert activation identifier and uncertainty level label; Specifically, the structured prediction output object is a JSON format data body, containing fields: "predicted_value" stores the predicted mean value, "uncertainty" stores the calibrated prediction standard deviation, "timestamp" stores the time center point corresponding to the prediction, "location" stores the spatial coordinates or device position number, "experts_activated" stores the list of activated expert network identifiers in S3, and "uncertainty_level" is automatically labeled as "low / medium / high" according to the standard deviation size.

[0080] Further, if the calibrated prediction standard deviation is greater than 1.0, "uncertainty_level" is forcibly marked as "high", and a "physics_residual_alert: true" field is appended to the data body, prompting that there is a significant physical inconsistency in the predicted area.

[0081] Further, the structured prediction output object is appended with metadata "input_quality_flag", whose value is inherited from the data quality label in S15, used to trace whether the prediction uncertainty is caused by original data defects.

[0082] S62: Push the structured prediction output object to the refining device operation decision interface as the rolling prediction result of the target process parameter and the corresponding uncertainty index; Specifically, the operation decision interface is an advanced process control (APC) system data bus, which pushes the latest prediction object every 5 seconds through OPC UA protocol for the controller to reference and adjust the set value.

[0083] Further, if "uncertainty_level" is "high", a pop-up window alarm is pushed to the operator monitoring interface at the same time, containing the prediction value, standard deviation, activated experts, and the largest source of physical residual, to assist manual judgment of whether to intervene.

[0084] Further, the structured prediction output object is synchronously written into the historical database for subsequent online model updating and performance evaluation, forming a "prediction-feedback-optimization" closed loop.

[0085] Embodiment 3 is a third embodiment of the present application, which provides a refining device intelligent prediction system based on a physical information neural network, comprising: A data perception module is configured to acquire real-time working condition data of a refining device, extract a space-time feature vector and a device region identifier from the real-time working condition data; An attention guiding module is configured to input the space-time feature vector into a physical guiding attention structure to obtain a space-time attention weight; an expert routing module configured to determine an activated expert network set from a set of expert networks according to a combination of the spatio-temporal attention weight and the device region identifier, each expert network corresponding to a region and a physical mechanism combination; a mechanism reasoning module configured to input the spatio-temporal feature vector into the activated expert network set to obtain a set of local prediction values and a set of physical residual components; a credible calibration module configured to input the set of local prediction values and the set of physical residual components into a Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; a decision output module configured to output the prediction mean and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty indicator.

[0086] Embodiment 4, with reference to Figure 4 As a fourth embodiment of the present application, different from the first three embodiments, the function, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application or the part of the technical solution that essentially contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0087] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processor, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions, or in conjunction with these instructions execution systems, apparatuses, or devices. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport programs for use by an instruction execution system, apparatus, or device, or in conjunction with these instruction execution systems, apparatuses, or devices.

[0088] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer readable medium can be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory in a suitable manner.

[0089] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following techniques, which are well known in the art of hardware implementation, can be used: a combination of logic gates in a logic circuit, a combination of processor(s) and memory that forms a special purpose machine and has appropriate software with instructions for performing the desired functions, a combination of application specific integrated circuits (ASICs) and / or other hardware components, a combination of one or more Central Processing Units (CPUs) and / or other suitable processors and storage.

[0090] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, but not limit the present application, although the present application is described in detail with reference to the preferred embodiments, it should be apparent for those skilled in the art that the technical solutions of the present application can be modified or equivalent replaced, without departing from the spirit and scope of the present application, and all of them should be included in the scope of the claims of the present application.

Claims

1. A method for intelligent prediction of refining and chemical equipment based on physical information neural networks, characterized in that: The method comprises: acquiring real-time working condition data of a refining device; extracting a space-time feature vector and a device region identifier from the real-time working condition data; inputting the space-time feature vector into a physical guided attention structure to obtain a space-time attention weight; determining an activated expert network set from an expert network set according to a combination of the space-time attention weight and the device region identifier, each expert network corresponding to a region and a physical mechanism combination; inputting the space-time feature vector into the activated expert network set to obtain a local prediction value set and a physical residual component set; inputting the local prediction value set and the physical residual component set into a Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; outputting the prediction mean and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty index.

2. The intelligent prediction method for refining devices based on physical information neural network according to claim 1, characterized in that: The method comprises: pulling raw data streams output by a sensor node from a refining device distributed control system and an industrial Internet of Things platform at a 1-second cycle; performing timestamp alignment and linear interpolation of missing values on the raw data streams to obtain a complete time series data set; mapping the time series data set to a preset region coding table according to equipment position numbers and spatial coordinates to generate a device region identifier; performing sliding window slicing and Min-Max normalization on the time series data set to generate a space-time feature vector sequence with a length of 10; pairing the space-time feature vector sequence with a corresponding device region identifier to form a neural network input sample pair.

3. The intelligent prediction method for refining and chemical plants based on physical information neural networks according to claim 2, characterized in that: The method comprises: inputting the space-time feature vector into a physical residual calculation layer to obtain a physical residual tensor according to a preset control equation; performing gradient operator operation on the physical residual tensor to obtain a physical residual gradient tensor; inputting the physical residual gradient tensor into a nonlinear mapping layer to obtain an initial attention weight tensor; performing normalization operation on the initial attention weight tensor to obtain the space-time attention weight.

4. The intelligent prediction method for refining and chemical plants based on physical information neural networks according to claim 3, characterized in that: The method comprises: splicing the space-time attention weight and the device region identifier into a joint routing feature vector; inputting the joint routing feature vector into a gating network to obtain an expert activation probability distribution; selecting expert networks with a probability value greater than a preset threshold value from the expert network set according to the expert activation probability distribution to form the activated expert network set.

5. The intelligent prediction method for refining and chemical plants based on physical information neural networks according to claim 4, characterized in that: The method comprises: inputting the space-time feature vector into each expert network in the activated expert network set in parallel to obtain a local prediction value output by each expert network; inputting the local prediction value of each expert network into its embedded physical equation calculation layer to obtain a corresponding physical residual component; pooled into the local prediction value set, and all physical residual components are pooled into the physical residual component set.

6. The intelligent prediction method for refining plants based on physical information neural networks according to claim 5, characterized in that: The method comprises: inputting the local prediction value set into an uncertainty estimation layer to obtain an initial prediction standard deviation; inputting the physical residual component set into a residual calibration layer to generate a residual scaling factor; multiplying the initial prediction standard deviation by the residual scaling factor to obtain the calibrated prediction standard deviation, and taking a weighted average of the set of local prediction values as the prediction mean.

7. The intelligent prediction method for refining and chemical plants based on physical information neural networks according to claim 6, characterized in that: The outputting the prediction mean and the calibrated prediction standard deviation comprises: encapsulating the prediction mean and the calibrated prediction standard deviation into a structured prediction output object, the structured prediction output object comprising a numerical value, a time stamp, a spatial location, an expert activation identification, and an uncertainty level label; pushing the structured prediction output object to a refinery operation decision interface as a rolling prediction result of a target process parameter and a corresponding uncertainty indicator.

8. A physical information neural network-based intelligent prediction system for a refining plant, applying the physical information neural network-based intelligent prediction method according to any one of claims 1 to 7. comprises: a data perception module configured to obtain real-time operating data of a refinery, extract a spatio-temporal feature vector and a device region identification from the real-time operating data; an attention guiding module configured to input the spatio-temporal feature vector into a physical guiding attention structure to obtain a spatio-temporal attention weight; an expert routing module configured to determine an activated expert network set from a set of expert networks according to a combination of the spatio-temporal attention weight and the device region identification, each expert network corresponding to a region and a physical mechanism combination; a mechanism reasoning module configured to input the spatio-temporal feature vector into the activated expert network set to obtain a set of local prediction values and a set of physical residual components; a reliable calibration module configured to input the set of local prediction values and the set of physical residual components into a Bayesian output structure to obtain a prediction mean and a calibrated prediction standard deviation; a decision output module configured to output the prediction mean and the calibrated prediction standard deviation as a target process parameter prediction result and an uncertainty indicator. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the intelligent prediction method for a refinery based on a physical information neural network according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the intelligent prediction method for a refinery based on a physical information neural network according to any one of claims 1 to 7.

Images