A physical information neural network-based intelligent prediction method for refining devices

Abstract

The application discloses a kind of based on physical information neural network's intelligent prediction method of refining device, belong to intelligent prediction technical field, comprising: the real-time working condition data of refining device is obtained, space-time characteristic vector and device area mark are extracted from the real-time working condition data;The space-time characteristic vector is input into physical guide attention structure, and space-time attention weight is obtained;According to the combination of the space-time attention weight and the device area mark, determine the activated expert network set from expert network set;The space-time characteristic vector is input into the activated expert network set, and obtain local prediction value set and physical residual component set;The local prediction value set and the physical residual component set are input into bayesian output structure, and obtain prediction mean and calibrated prediction standard deviation;The prediction mean and the calibrated prediction standard deviation are output as target process parameter prediction result and uncertainty index.The application has high accuracy and traceability.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology, and specifically to an intelligent prediction method for refining and chemical equipment based on a physical information neural network. Background Technology

[0002] Physical Information Neural Networks (PINNs) have seen rapid development in industrial modeling in recent years. By embedding physical equations into neural network loss functions, they achieve a fusion of data-driven approaches and mechanistic constraints, demonstrating superior generalization and interpretability compared to pure data models in scenarios such as fluid mechanics, heat and mass transfer, and reaction engineering. Particularly in the energy and chemical engineering field, PINNs are being explored for tasks such as reactor temperature field reconstruction, pipeline pressure drop prediction, and catalyst activity inversion, gradually moving from academic simulation to engineering prototype verification, becoming a key technological path for "white-box AI" in smart factories.

[0003] However, PINN currently faces three major bottlenecks in prediction for refining and chemical plants: insufficient local accuracy: traditional PINN applies uniform physical constraints across the entire domain, making it difficult to focus on high-gradient regions such as reactor hotspots and phase change interfaces, leading to distortion in predictions of critical locations; poor mechanism adaptability: a single network structure cannot adapt to the coupling of multiple physics fields (such as reaction-heat transfer-flow) and regional heterogeneity within the plant, and forced uniform modeling causes mechanism conflicts; and unquantifiable risk: the output is only a point estimate, lacking uncertainty indicators linked to physical consistency, making it impossible for operators to judge the reliability of the prediction and hindering safety decision-making. These problems directly restrict the depth and reliability of PINN's application in complex industrial scenarios. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: how to achieve high-precision local focusing of multi-physics fields in refining and chemical equipment, adaptive mechanism modeling of heterogeneous regions, physical consistency constraints and uncertainty quantification output of prediction results through physical guidance attention mechanism, regional mechanism expert network and Bayesian residual calibration structure, while suppressing mechanism conflicts and prediction drift in high gradient regions caused by uniform modeling of the whole domain.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for intelligent prediction of refining and chemical equipment based on physical information neural network, comprising: acquiring real-time operating data of refining and chemical equipment, and extracting spatiotemporal feature vectors and equipment area identifiers from the real-time operating data; The spatiotemporal feature vector is input into the physical guided attention structure to obtain the spatiotemporal attention weights; Based on the combination of the spatiotemporal attention weights and the device region identifiers, an active expert network set is determined from the expert network set, with each expert network corresponding to a region and physical mechanism combination; The spatiotemporal feature vector is input into the set of activated expert networks to obtain a set of local predicted values ​​and a set of physical residual components. The set of local predicted values ​​and the set of physical residual components are input into the Bayesian output structure to obtain the predicted mean and the calibrated prediction standard deviation. The predicted mean and the calibrated predicted standard deviation are output as the prediction result and uncertainty index of the target process parameters.

[0007] As a preferred embodiment of the intelligent prediction method for refining and chemical processing equipment based on physical information neural networks described in this invention, the extraction of spatiotemporal feature vectors and equipment area identifiers includes: The raw data stream output by the sensor nodes is retrieved from the distributed control system and industrial IoT platform of the refining and chemical plant at 1-second intervals. Perform timestamp alignment and linear interpolation of missing values ​​on the original data stream to obtain a complete time-series dataset; Based on the device tag number and spatial coordinates, the time-series dataset is mapped to a preset region encoding table to generate a device region identifier; Sliding window slicing and Min-Max normalization are performed on the time series dataset to generate a spatiotemporal feature vector sequence of length 10; The spatiotemporal feature vector sequence is paired with the corresponding device region identifier to form a neural network input sample pair.

[0008] As a preferred embodiment of the intelligent prediction method for refining and chemical equipment based on a physical information neural network as described in this invention, the method for obtaining spatiotemporal attention weights includes: The spatiotemporal feature vector is input into the physical residual calculation layer, and the physical residual tensor is obtained according to the preset control equation. Perform gradient operator operations on the physical residual tensor to obtain the physical residual gradient tensor; The physical residual gradient tensor is input into the nonlinear mapping layer to obtain the initial attention weight tensor; Normalization is performed on the initial attention weight tensor to obtain the spatiotemporal attention weights.

[0009] As a preferred embodiment of the intelligent prediction method for refining and chemical equipment based on physical information neural networks described in this invention, the determination of the activated expert network set includes: The spatiotemporal attention weights are concatenated with the device region identifier to form a joint routing feature vector; The joint routing feature vector is input into the gating network to obtain the expert activation probability distribution; Based on the expert activation probability distribution, expert networks with probability values ​​greater than a preset threshold are selected from the expert network set to form the activated expert network set.

[0010] As a preferred embodiment of the intelligent prediction method for refining and chemical equipment based on a physical information neural network as described in this invention, the step of obtaining the set of local predicted values ​​and the set of physical residual components includes: The spatiotemporal feature vectors are input in parallel into each expert network in the set of activated expert networks to obtain the local prediction values ​​output by each expert network. The local predictions of each expert network are input into its embedded physical equation computation layer to obtain the corresponding physical residual components. The local predictions output by all expert networks are aggregated into the set of local predictions, and all physical residual components are aggregated into the set of physical residual components.

[0011] As a preferred embodiment of the intelligent prediction method for refining and chemical equipment based on physical information neural networks described in this invention, the step of obtaining the predicted mean and the calibrated predicted standard deviation includes: The set of local predicted values ​​is input into the uncertainty estimation layer to obtain the initial prediction standard deviation; The set of physical residual components is input into the 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 the weighted average of the local prediction set is used as the prediction mean.

[0012] As a preferred embodiment of the intelligent prediction method for refining and chemical processing equipment based on physical information neural networks described in this invention, wherein: the output of the predicted mean and the calibrated predicted standard deviation includes: The predicted mean and the calibrated predicted standard deviation are encapsulated into a structured prediction output object, which includes numerical values, timestamps, spatial locations, expert activation identifiers, and uncertainty level labels. The structured prediction output object is pushed to the refining unit operation decision interface as the rolling prediction result of the target process parameters and the corresponding uncertainty index.

[0013] This invention provides an intelligent prediction system for refining and chemical equipment based on a physical information neural network.

[0014] To solve the above-mentioned technical problems, the present invention further provides the following technical solution: an intelligent prediction system for refining and chemical equipment based on physical information neural network, comprising: a data sensing module, used to acquire real-time operating data of refining and chemical equipment, and extract spatiotemporal feature vectors and equipment area identifiers from the real-time operating data; The attention guidance module is used to input the spatiotemporal feature vector into the physical guidance attention structure to obtain the spatiotemporal attention weights; The expert routing module is used to determine the set of active expert networks from the set of expert networks based on the combination of the spatiotemporal attention weight and the device region identifier, wherein each expert network corresponds to a region and physical mechanism combination; The mechanism reasoning module is used to input the spatiotemporal feature vector into the set of activated expert networks to obtain a set of local predicted values ​​and a set of physical residual components; The reliable calibration module is used to input the set of local predicted values ​​and the set of physical residual components into the Bayesian output structure to obtain the predicted mean and the calibrated prediction standard deviation. The decision output module is used to output the predicted mean and the calibrated predicted standard deviation as the prediction result and uncertainty index of the target process parameters.

[0015] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the intelligent prediction method for refining and chemical equipment based on a physical information neural network.

[0016] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the intelligent prediction method for refining and chemical equipment based on a physical information neural network.

[0017] The beneficial effects of this invention are as follows: By constructing a three-layer coupled architecture of "physically guided attention—regional mechanism expert—residual calibration output," this invention systematically overcomes the bottlenecks of local distortion, mechanism conflict, and risk black box in the prediction of refining and chemical plants using traditional physical information neural networks. It significantly improves the modeling accuracy of key components such as reactors and heat exchangers by dynamically focusing high-gradient physical mismatch regions through spatiotemporal attention weights. Furthermore, it guides the expert network to adaptively activate according to "region-mechanism" combinations through device region identification, achieving divide-and-conquer modeling of heterogeneous physical fields such as reaction kinetics, heat transfer, and gas-liquid balance, thus avoiding parameter coupling errors. Finally, it synchronously outputs residual components through embedded physical equations and drives... The Bayesian layer dynamically scales the prediction standard deviation, directly linking uncertainty indicators to physical consistency, achieving a controllable risk output that is "more inaccurate, more cautious." Finally, through structured encapsulation and decision interface push, metadata such as prediction mean, calibration standard deviation, expert sources, and spatial location are delivered in an integrated manner. This supports operators and control systems in implementing hierarchical decision-making of "automatic optimization for low risks and manual intervention for high risks." Thus, while ensuring physical interpretability and industrial safety, it realizes a paradigm shift in the prediction of key parameters of refining and chemical plants from "black box point estimation" to "transparent risk quantification," providing smart factories with a prediction engine that combines high accuracy, strong generalization, traceability, and actionability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The overall flowchart of an intelligent prediction method for refining and chemical equipment based on a physical information neural network is provided in one embodiment of the present invention; Figure 2 A flowchart illustrating the risk prediction method for refining and chemical equipment based on a physical information neural network, as provided in one embodiment of the present invention; Figure 3 A safety alarm flowchart of a refining and chemical plant intelligent prediction method based on physical information neural network provided in one embodiment of the present invention; Figure 4 This is a computer device diagram illustrating an intelligent prediction method for refining and chemical equipment based on a physical information neural network, provided as an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides an intelligent prediction method for refining and chemical equipment based on a physical information neural network, comprising: S1: Obtain real-time operating data of the refining and chemical unit, and extract spatiotemporal feature vectors and unit area identifiers from the real-time operating data; S2: Input the spatiotemporal feature vector into the physical guided attention structure to obtain the spatiotemporal attention weights; S3: Based on the combination of the spatiotemporal attention weight and the device region identifier, determine the set of active expert networks from the expert network set, with each expert network corresponding to a region and physical mechanism combination; S4: Input the spatiotemporal feature vector into the activated expert network set to obtain the local predicted value set and the physical residual component set; S5: Input the set of local predicted values ​​and the set of physical residual components into the Bayesian output structure to obtain the predicted mean and the calibrated prediction standard deviation; S6: Output the predicted mean and the calibrated predicted standard deviation as the prediction result and uncertainty index of the target process parameters.

[0022] It should be noted that during the operation of refining and chemical plants, multiple physical fields are strongly coupled within equipment such as reactors, fractionation towers, and heat exchangers. Parameters such as temperature, pressure, composition, and flow rate undergo dramatic nonlinear evolution in both spatial and temporal dimensions. Traditional data-driven models are prone to distortion in local areas, while mechanistic models are difficult to update online due to time-varying parameters and ambiguous boundaries. At the same time, the long-term operation of the plant is affected by catalyst deactivation, coking, and corrosion, resulting in significant operating condition drift. A single model cannot balance accuracy and generalization. Furthermore, industrial decision-making requires the prediction results to be accompanied by credibility assessments. Existing AI models generally lack the ability to quantify uncertainty, leading to uncontrollable operational risks.

[0023] Therefore, to address the three core problems of low prediction accuracy, poor mechanism adaptability, and unknown risks, a neural network architecture integrating physical-guided attention, multi-mechanism expert networks, and Bayesian calibration is constructed through steps S1-S6. This architecture enables automatic focusing and modeling of high-gradient / strong nonlinear regions; achieves expert adaptive activation in regions with different physical mechanisms; and outputs confidence intervals with physical residual calibration for prediction results. This improves the prediction accuracy, operating condition adaptability, and operational decision safety of refining and chemical plants while ensuring physical consistency.

[0024] Example 2, refer to Figures 1-3 This is the second embodiment of the present invention, which provides an intelligent prediction method for refining and chemical equipment based on a physical information neural network.

[0025] S1: Obtain real-time operating data of the refining and chemical unit, and extract spatiotemporal feature vectors and unit area identifiers from the real-time operating data; It should be noted that the refining and chemical unit is a high-temperature, high-pressure, multiphase flow strongly coupled system. 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 feedstock. The unit has a large space, numerous measurement points, and significant regional differences in mechanisms. Data is easily interfered with or missing, and purely data-driven models are prone to distortion in key areas and lack reliable prediction output. Therefore, it is necessary to construct a prediction architecture that integrates physical attention, regional expert networks, and Bayesian calibration to achieve local focus, mechanism adaptation, and risk quantification, thereby improving prediction accuracy and decision reliability.

[0026] Specifically, step S1 extracts the spatiotemporal feature vector and device region identifier, including the following sub-steps S11 to S15: S11: Pull raw data streams from multiple sensor nodes at 1-second intervals from the distributed control system and industrial IoT platform of the refining and chemical plant; Specifically, the raw data stream is subscribed to from the DCS historical server via the OPC UA protocol. The subscription tag list is preloaded based on the device instrument index table. The acquisition period is fixed at 1000 milliseconds, and the acquisition range covers the sensor nodes of three types of process sections: reaction zone, separation zone, and heat exchange zone.

[0027] Specifically, each sensing node is configured with an independent data cache queue. If no new data is received within a certain period, the value of the previous period is retained and the timestamp is marked as not updated. When there are no updates for 5 consecutive periods, the "node offline" state is triggered, and the data of that node is skipped in subsequent steps and does not participate in feature stitching.

[0028] Furthermore, in the catalytic cracking riser area, a dual-channel acquisition mechanism is enabled for temperature measuring point TIC-101 and density measuring point DENS-102: the main channel acquires data at a 1-second cycle, and the backup channel acquires data at a 200-millisecond cycle and stores it independently; when the main channel data is marked as "interpolation" or "offline", it automatically switches to the most recent valid value of the backup channel to ensure the continuity of input in key areas.

[0029] Specifically, the data retrieval process incorporates a CRC check mechanism. Each data frame is accompanied by a 16-bit check code. When the receiving end fails to verify the data, it discards the frame and records the error count. If the accumulated errors exceed the threshold, the communication link self-diagnosis is triggered to prevent dirty data from entering the preprocessing process.

[0030] S12: Perform timestamp alignment and linear interpolation of missing values ​​on the original data stream to obtain a complete time-series dataset; Specifically, based on the DCS master clock, time alignment is performed on all node data within ±50 milliseconds. Data points outside this range are considered invalid and are filled by linear interpolation of adjacent valid points. The interpolated data is then marked with a "_I" suffix.

[0031] Furthermore, in the hydrogenation reactor area, if both the hydrogen flow rate FIC-301 and the feedstock oil flow rate FIC-302 are missing, the system calls the total flow rate at the reactor inlet and the preset hydrogen-oil ratio to calculate a substitute value. The calculation result is marked as "_C", and the ratio parameter used in the calculation is recorded for subsequent physical residual verification to exclude this point.

[0032] S13: Based on the device tag number and spatial coordinates, map the time series dataset to a preset region encoding table to generate a device region identifier; Specifically, the area code table is stored in a local configuration file. Each line contains three fields: device tag number prefix, spatial coordinate boundary, and area code. The system matches the smallest bounding box into which the sensor node falls based on the installation location coordinates bound to the sensor node and outputs the corresponding area code.

[0033] Furthermore, in the atmospheric and vacuum distillation tower area, a five-level area code is automatically generated based on the tower tray number TRAY-XX and the radial installation radius, such as "CDU-L12-R3", which represents the 12th tower tray and the 3rd radial zone. This code serves as the basis for subsequent expert network routing, ensuring that different tower layers activate different mass transfer or fluid dynamics experts.

[0034] S14: Perform sliding window slicing and Min-Max normalization on the time series dataset to generate a spatiotemporal feature vector sequence of length 10; Specifically, the sliding window length is fixed at 10 sampling points, with a step size of 1. Each sensing parameter within each window is independently normalized using Min-Max. The normalization baseline value is taken from the extreme value of the parameter over the past 8 hours, and the baseline is updated every 30 minutes.

[0035] Furthermore, independent normalization intervals are set for different physical quantities: temperature [50, 600], pressure [0.05, 5.0], flow rate [0, maximum design value × 1.1], and concentration [0, 100], to avoid imbalance of feature weights due to differences in dimensions. After normalization, the value range is unified as [0, 1].

[0036] Furthermore, in the top region of the fractionation column, for the three parameters "top temperature", "reflux ratio" and "light component concentration", the covariance matrix within the sliding window is first calculated. If the correlation coefficients between features are all greater than 0.8, principal component projection is performed to retain the first principal component, and then normalization is performed to compress redundant dimensions and strengthen the expression of coupled features.

[0037] S15: Pair the spatiotemporal 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 10×N dimensional feature vector and a region encoding string, where N is the number of valid sensing nodes in the window. If the number of nodes is less than the preset value, zeros are added to the end and the valid length is marked.

[0039] Furthermore, in the heat exchange network region, if the sample originates from "shell-side inlet temperature TE-401", "tube-side pressure drop PDIC-402", or "medium flow rate FIC-403", then the region identifier "HX-SHELL-IN" is bound, and the physical mechanism label "HEAT_XFER_FLUID" is attached, so that the subsequent gated network can identify that "heat transfer fluid dynamics expert" should be activated.

[0040] S2: Input the spatiotemporal feature vector into the physical guided attention structure to obtain the spatiotemporal attention weights; It should be noted that the evolution of the physical field within a refining and chemical plant exhibits strong nonlinearity and spatial heterogeneity. Traditional neural networks lack the ability to perceive local high gradients or regions of physical mismatch, easily leading to the accumulation of prediction biases. Simply relying on data fitting cannot guide the model to focus on key areas of physical inconsistency, while fixed-weight physical constraints are difficult to adapt to dynamic changes in operating conditions. Therefore, a physical-guided attention mechanism needs to be introduced, dynamically generating spatiotemporal attention weights based on the physical residual gradient. This allows the network to automatically focus on regions of conservation imbalance or drastic change during inference, thereby improving local modeling accuracy and physical consistency without increasing model complexity. This provides high-resolution regional guidance signals for subsequent expert routing and uncertainty calibration.

[0041] Specifically, step S2 obtains the spatiotemporal attention weights, including the following sub-steps S21 to S24: S21: Input the spatiotemporal feature vector into the physical residual calculation layer, and obtain the physical residual tensor according to the preset control equation; Specifically, the physical residual calculation layer calculates the imbalance between density, velocity, and rate of change of time in the input features based on the principle of mass conservation, and outputs residual values ​​aligned with the input spatial dimension. Each position represents the degree of deviation from the mass conservation equation at that point.

[0042] Furthermore, in the catalytic cracking reaction zone, the physical residual calculation layer simultaneously calculates the energy conservation residual and the component reaction kinetic residual, respectively, based on temperature changes, heat conduction, reaction exothermics, component concentration changes, convective transport, and chemical reaction rates, generating independent residual channels, which are then spliced ​​into a multi-channel tensor output.

[0043] S22: Perform gradient operator operations on the physical residual tensor to obtain the physical residual gradient tensor; Specifically, the gradient operator calculates the rate of change of the physical residual along the three spatial directions and the time direction, respectively. It adopts the central difference method, that is, it divides the difference between adjacent points by the sampling interval to obtain the partial derivatives in the four directions, forming the gradient tensor with the same dimension as the original residual tensor.

[0044] Furthermore, if the device has no structural changes in a certain spatial direction (e.g., an axial reactor has no radial gradient), the gradient in that direction is forcibly set to zero; before calculation, a three-dimensional median filter is performed on the residual tensor with a window size of 3×3×3 to suppress the interference of measurement noise on gradient calculation.

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

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

[0047] in, This represents the value of a single scalar element in the initial attention weight tensor. This is a trainable scaling factor, initially set to 1.0, used to adjust the amplification factor of the gradient on the attention. The magnitude of the physical residual gradient is obtained by taking the square root of the sum of the squares of the partial derivatives in the spatial and temporal directions. This is a trainable offset, initially set to 0.0, used to adjust the attention activation baseline. It is a nonlinear function whose output is always positive. It is defined as a composite of the natural logarithm and the exponential function to avoid negative weights.

[0048] Furthermore, to prevent excessive local gradients from causing abnormal weight increases, an upper limit truncation mechanism is set: if the mapped value exceeds the preset threshold of 5.0, it is forced to be 5.0; otherwise, the original value is retained to ensure numerical stability.

[0049] S24: Perform a normalization operation on the initial attention weight tensor to obtain the spatiotemporal attention weights; Specifically, the initial attention weights for each spatial location are normalized along the time dimension, so that the sum of the weights of that location at all time steps is 1, preserving the relative strength of spatial relationships, which makes it easier for the expert network to focus on key time periods according to the time distribution.

[0050] Furthermore, in the heat exchanger region, local maximum-minimum normalization is performed independently for each spatial location: the maximum value minus the minimum value in all time steps at that location is used as the denominator, the current value minus the minimum value is used as the numerator, and a very small value is added to prevent division by zero, so that the output value range is compressed to [0,1], highlighting the changes in local hot spots.

[0051] S3: Based on the combination of the spatiotemporal attention weight and the device region identifier, determine the set of active expert networks from the expert network set, with each expert network corresponding to a region and physical mechanism combination; It should be noted that the dominant physical mechanisms in different regions of a refining and chemical plant vary significantly, making it difficult for a single model to account for the constraints of multiple equations, including reaction kinetics, phase equilibrium, and heat transfer. Forcing global parameter sharing would lead to mechanism conflicts and prediction distortions. Therefore, it is necessary to construct a multi-expert network ensemble, where each expert is bound to a specific region and physical mechanism combination. Adaptive experts are dynamically activated based on spatiotemporal attention weights and region identifiers to achieve precise modeling "by region and by mechanism," avoiding mismatches between model structure and physical reality, and improving cross-condition generalization ability and local prediction reliability.

[0052] Specifically, step S3 involves determining the set of activated expert networks, including the following sub-steps S31 to S33: S31: Concatenate the spatiotemporal attention weights with the device region identifier to form a joint routing feature vector; Specifically, the device area identifier is a preset encoded string, which is first converted into a one-hot encoded vector, and then concatenated with the spatiotemporal attention weight in the feature dimension to form an enhanced routing feature, which is used to express "which areas are physically inconsistent at which time points".

[0053] Furthermore, in the atmospheric and vacuum distillation tower area, the device area identifier includes the tower layer number and radial partition. After being converted into a two-dimensional location embedding vector, it is weighted and added with spatiotemporal attention weights to generate spatially aware routing features.

[0054] S32: Input the joint routing feature vector into the gating network to obtain the expert activation probability distribution; Specifically, the gated network is a two-layer fully connected structure. The output dimension of the first layer is twice the total number of experts, and 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] Furthermore, the input layer of the gated network is supplemented with a learnable region bias vector, with different initial bias values ​​corresponding to different regions. For example, the bias of the reaction region is initialized to +0.5, and the bias of the separation region is initialized to -0.2, which guides the network to initially tend to activate the corresponding region experts.

[0056] Furthermore, if an expert's historical activation frequency in a specific region is less than 5%, a penalty term of -1.0 is applied to its output logarithmic probability before Softmax to reduce its selection probability, thereby achieving expert load balancing and resource optimization.

[0057] S33: Based on the expert activation probability distribution, select expert networks with probability values ​​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 an expert is greater than 0.1, it is added to the activation set to ensure that at least one expert is activated each time, and the maximum number of experts activated does not exceed 3, thus avoiding waste of computing resources.

[0058] Furthermore, during the operation of the catalytic cracking unit, if the spatiotemporal attention weight of the "regenerator temperature" region is detected to be continuously higher than 0.8, the "high-temperature oxidation reaction expert" will be forcibly activated, even if its probability is slightly lower than the threshold, to ensure that the critical region is not ignored.

[0059] S4: Input the spatiotemporal feature vector into the activated expert network set to obtain the local predicted value set and the physical residual component set; It should be noted that the activated expert network must independently output local predictions while maintaining physical constraints to avoid error propagation caused by mechanism coupling. If only the predicted values ​​are output while ignoring physical consistency verification, the core advantages of PINN will be lost. Therefore, each expert, while outputting local predicted values, simultaneously calculates the residual components of its embedded control equations, forming a "prediction-residual" paired output. This ensures the physical rationality of the local modeling and provides a quantitative basis for subsequent uncertainty calibration, achieving a closed-loop synergy between data-driven and mechanism-constrained approaches.

[0060] Specifically, step S4 obtains the set of local predicted values ​​and the set of physical residual components, including the following sub-steps S41 to S43: S41: Input the spatiotemporal feature vectors in parallel into each expert network in the set of activated expert networks to obtain the local prediction values ​​output by each expert network; Specifically, each expert network has a three-layer fully connected structure. The input layer receives the spatiotemporal feature vector, the hidden layer has 128 neurons with GELU activation function, and the output layer has the number of target process parameters, no activation function, and directly outputs continuous values ​​as local prediction values.

[0061] Furthermore, the "thermal reaction kinetics expert" activated in the hydrogenation reactor region has its reactor axial position code embedded in its network structure, enabling different conversion rate predictions to be output at different axial positions under the same inlet conditions, thus enhancing spatial resolution.

[0062] Furthermore, if an expert network has historically consistently predicted errors higher than a threshold in the current region, a decay coefficient of 0.9 is applied to its output in this inference process to reduce its weight in subsequent fusion, thereby achieving dynamic confidence adjustment.

[0063] S42: Input the local predictions of each expert network into its embedded physical equation computation layer to obtain the corresponding physical residual components; Specifically, the "thermal reaction kinetics expert" incorporates the Arrhenius reaction rate equation, calculates the theoretical reaction rate based on the output temperature and concentration predictions, and compares it with the predicted component change rate to obtain the reaction kinetic residual component.

[0064] Furthermore, the "steady-state heat transfer expert" incorporates Fourier's law of thermal conductivity, calculates the theoretical heat flux density based on the predicted temperature field and the thermal conductivity of the material, and compares it with the predicted rate of energy change to obtain the energy conservation residual component.

[0065] Furthermore, if the physical equations embedded in an expert network involve uninputted features (such as catalyst activity), then the historical moving average is used as a substitute input, and the residuals are calculated and marked as "containing estimated parameters" for weight reduction during subsequent uncertainty calibration.

[0066] S43: Aggregate the local predictions output by all expert networks into the set of local predictions, and aggregate all physical residual components into the set of physical residual components. Specifically, the local predicted value set is stored in list form, with each element containing an expert identifier, a predicted value vector, and a spatial location index; the physical residual component set stores the corresponding residual vectors in the same order to ensure that subsequent steps can be processed according to expert alignment.

[0067] Furthermore, outlier filtering is performed on local predictions before output: if an expert's prediction deviates from the mean of other experts by more than three standard deviations, it is marked as an "outlier prediction" and will not participate in the final fusion; only its residuals are retained for uncertainty analysis.

[0068] S5: Input the set of local predicted values ​​and the set of physical residual components into the Bayesian output structure to obtain the predicted mean and the calibrated prediction standard deviation; It should be noted that if the predictive model for refining and chemical plants only outputs point estimates, operational decisions will lack risk perception capabilities and are prone to misjudgment. While traditional Bayesian networks can quantify model uncertainty, they cannot perceive systematic biases caused by physical mismatches. Therefore, it is necessary to integrate physical residual information at the output layer and dynamically modulate the prediction standard deviation through a residual scaling factor, linking the uncertainty output to physical consistency—the larger the physical residual, the higher the standard deviation after calibration, and the less reliable the prediction. This provides intelligent outputs with "risk labels" for industrial operations, improving system safety margins and human-machine collaboration efficiency.

[0069] Specifically, step S5 obtains the predicted mean and the calibrated predicted standard deviation, including the following sub-steps S51 to S53: S51: Input the set of local predicted values ​​into the uncertainty estimation layer to obtain the initial prediction standard deviation; Specifically, the uncertainty estimation layer adopts a Monte Carlo Dropout structure, which performs five random dropout forward propagations on the local predicted value set during the inference phase, and calculates the standard deviation of the output value as the initial prediction standard deviation.

[0070] Furthermore, the uncertainty estimation layer is a variational Bayesian fully connected layer with weights following a Gaussian distribution. Multiple prediction branches are generated through reparameterized sampling, and the square root of the variance of the output prediction value is used as the initial prediction standard deviation.

[0071] Furthermore, if an expert network is labeled as "outlier prediction" or "containing estimated parameters", its predicted values ​​are weighted by 50% when calculating the standard deviation, thus avoiding outliers from interfering with 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 with the activation function being Softplus. The output value is used as the local scaling factor for the region corresponding to the expert, and then the global residual scaling factor is obtained by weighted averaging according to the expert activation probability.

[0073] Furthermore, in the reactor region, if the absolute value of the energy conservation residual component exceeds 0.5, the scaling factor is forcibly set to 1.5, indicating that the prediction confidence in this region should be reduced; if the residual is less than 0.1, the scaling factor is set to 0.8, indicating good physical consistency.

[0074] Furthermore, the residual calibration layer introduces the sliding historical residual mean as a benchmark. If the current residual deviates from the historical mean by more than twice the standard deviation, the scaling factor is linearly amplified to achieve dynamic anomaly detection.

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

[0076] Furthermore, if the predicted standard deviation after calibration exceeds the preset safety threshold of 1.2, a "high uncertainty" flag will be added to the output for the operator or upper control system to trigger manual review or a conservative strategy.

[0077] S6: Output the predicted mean and the calibrated predicted standard deviation as the prediction result and uncertainty index of the target process parameters.

[0078] It should be noted that if intelligent predictions for refining and chemical plants only output a single value, it is difficult to support operational decisions in safety-sensitive scenarios. Operators need to know the reliability of the predictions simultaneously to determine whether to adopt, intervene in, or review them. Therefore, the predicted mean and calibrated standard deviation need to be encapsulated into a structured output with risk labels, and associated with metadata such as spatial location, activated experts, and residual sources, and pushed to the control or human-machine interface to achieve a three-in-one delivery of "predicted value + uncertainty + interpretability." This upgrades intelligent prediction from "black box output" to "transparent decision support," improving system robustness and industrial applicability.

[0079] Specifically, step S6 outputs the predicted mean and the calibrated predicted standard deviation, including the following sub-steps S61 to S62: S61: Encapsulate the predicted mean and the calibrated predicted 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 the following fields: "predicted_value" stores the predicted mean, "uncertainty" stores the calibrated predicted standard deviation, "timestamp" stores the predicted time center point, "location" stores the spatial coordinates or device tag number, "experts_activated" stores the list of expert network identifiers activated in S3, and "uncertainty_level" is automatically labeled as "low / medium / high" according to the standard deviation.

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

[0081] Furthermore, the structured prediction output object is appended with metadata "input_quality_flag", the value of which is inherited from the data quality label in S15, and is used to trace whether the prediction uncertainty originates from defects in the original data.

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

[0083] Furthermore, if "uncertainty_level" is "high", a pop-up alarm will be pushed to the operator monitoring interface, containing the predicted value, standard deviation, activated expert, and the maximum source of physical residual, to assist manual judgment on whether to intervene.

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

[0085] Example 3 is the third embodiment of the present invention. This embodiment provides an intelligent prediction system for refining and chemical processing equipment based on a physical information neural network, comprising: The data sensing module is used to acquire real-time operating data of the refining and chemical unit, and extract spatiotemporal feature vectors and unit area identifiers from the real-time operating data. The attention guidance module is used to input the spatiotemporal feature vector into the physical guidance attention structure to obtain the spatiotemporal attention weights; The expert routing module is used to determine the set of active expert networks from the set of expert networks based on the combination of the spatiotemporal attention weight and the device region identifier, wherein each expert network corresponds to a region and physical mechanism combination; The mechanism reasoning module is used to input the spatiotemporal feature vector into the set of activated expert networks to obtain a set of local predicted values ​​and a set of physical residual components; The reliable calibration module is used to input the set of local predicted values ​​and the set of physical residual components into the Bayesian output structure to obtain the predicted mean and the calibrated prediction standard deviation. The decision output module is used to output the predicted mean and the calibrated predicted standard deviation as the prediction result and uncertainty index of the target process parameters.

[0086] Example 4, refer to Figure 4 This is the fourth embodiment of the present invention, which differs from the previous three embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

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

[0088] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0089] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0090] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent prediction of refining and chemical equipment based on physical information neural networks, characterized in that, include: Acquire real-time operating data of the refining and chemical unit, and extract spatiotemporal feature vectors and unit area identifiers from the real-time operating data; The spatiotemporal feature vector is input into the physical residual calculation layer, and the physical residual tensor is obtained according to the preset control equation. The preset control equation includes at least one of the mass conservation equation, energy conservation equation, or component reaction kinetic equation. The gradient operator operation is performed on the physical residual tensor to obtain the physical residual gradient tensor. The physical residual gradient tensor is input into the nonlinear mapping layer to obtain the initial attention weight tensor; the initial attention weight tensor is normalized to obtain the spatiotemporal attention weights. The spatiotemporal attention weights and the device region identifiers are concatenated to form a joint routing feature vector; the joint routing feature vectors are input into a gating network to obtain an expert activation probability distribution; based on the expert activation probability distribution, expert networks with probability values ​​greater than a preset threshold are selected from the expert network set to form an activated expert network set, with each expert network corresponding to a region and physical mechanism combination; The spatiotemporal feature vectors are input in parallel into each expert network in the set of activated expert networks to obtain the local predicted values ​​output by each expert network. The local predicted values ​​of each expert network are input into its embedded physical equation calculation layer. The theoretical reaction rate is calculated according to the Arrhenius reaction rate equation and compared with the predicted rate of change of components to obtain the reaction kinetic residual component, or the theoretical heat flux density is calculated according to Fourier's law of thermal conductivity and compared with the predicted rate of change of energy to obtain the energy conservation residual component. The local predicted values ​​output by all expert networks are aggregated into a set of local predicted values, and all physical residual components are aggregated into a set of physical residual components. The set of local predicted values ​​is input into an uncertainty estimation layer using a Monte Carlo Dropout structure or a variational Bayesian fully connected layer to obtain the initial prediction standard deviation; the set of physical residual components is input into a residual calibration layer, the absolute value of each physical residual component is taken and then input into a single-layer neural network and a local scaling factor is generated by the Softplus activation function, and then the global residual scaling factor is obtained by weighted averaging according to the expert activation probability distribution. The initial prediction standard deviation is multiplied by the global residual scaling factor to obtain the calibrated prediction standard deviation, and the local prediction set is weighted and averaged using the probability values ​​in the expert activation probability distribution as weights to obtain the prediction mean. The predicted mean and the calibrated predicted standard deviation are output as the prediction result and uncertainty index of the target process parameters.

2. The intelligent prediction method for refining and chemical equipment based on a physical information neural network as described in claim 1, characterized in that, The extracted spatiotemporal feature vector and device region identifier include: The raw data stream output by the sensor nodes is retrieved from the distributed control system and industrial IoT platform of the refining and chemical plant at 1-second intervals. Perform timestamp alignment and linear interpolation of missing values ​​on the original data stream to obtain a complete time-series dataset; Based on the device tag number and spatial coordinates, the time-series dataset is mapped to a preset region encoding table to generate a device region identifier; Sliding window slicing and Min-Max normalization are performed on the time series dataset to generate a spatiotemporal feature vector sequence of length 10; The spatiotemporal feature vector sequence is paired with the corresponding device region identifier to form a neural network input sample pair.

3. The intelligent prediction method for refining and chemical equipment based on a physical information neural network as described in claim 1, characterized in that, The output of the predicted mean and the calibrated predicted standard deviation includes: The predicted mean and the calibrated predicted standard deviation are encapsulated into a structured prediction output object, which includes numerical values, timestamps, spatial locations, expert activation identifiers, and uncertainty level labels. The structured prediction output object is pushed to the refining unit operation decision interface as the rolling prediction result of the target process parameters and the corresponding uncertainty index.

4. A refining and chemical plant intelligent prediction system based on a physical information neural network, employing the intelligent prediction method for refining and chemical plants based on a physical information neural network as described in any one of claims 1 to 3, characterized in that, include: The data sensing module is used to acquire real-time operating data of the refining and chemical unit, and extract spatiotemporal feature vectors and unit area identifiers from the real-time operating data. An attention-guided module is used to input the spatiotemporal feature vector into the physical residual calculation layer, obtain the physical residual tensor according to a preset control equation, wherein the preset control equation includes at least one of the mass conservation equation, energy conservation equation, or component reaction kinetic equation; and perform gradient operator operations on the physical residual tensor to obtain the physical residual gradient tensor. The physical residual gradient tensor is input into the nonlinear mapping layer to obtain the initial attention weight tensor; the initial attention weight tensor is normalized to obtain the spatiotemporal attention weights. The expert routing module is used to concatenate the spatiotemporal attention weights with the device region identifiers to form a joint routing feature vector; input the joint routing feature vectors into a gating network to obtain the expert activation probability distribution; and select expert networks with probability values ​​greater than a preset threshold from the expert network set according to the expert activation probability distribution to form an activated expert network set, with each expert network corresponding to a region and physical mechanism combination. The mechanism inference module is used to input the spatiotemporal feature vectors in parallel into each expert network in the set of activated expert networks to obtain the local predicted values ​​output by each expert network; input the local predicted values ​​of each expert network into its embedded physical equation calculation layer, calculate the theoretical reaction rate according to the Arrhenius reaction rate equation and compare it with the predicted rate of change of components to obtain the reaction kinetic residual component, or calculate the theoretical heat flux density according to Fourier's law of thermal conductivity and compare it with the predicted rate of change of energy to obtain the energy conservation residual component; and aggregate the local predicted values ​​output by all expert networks into a local predicted value set and aggregate all physical residual components into a physical residual component set. The reliable calibration module is used to input the set of local predicted values ​​into an uncertainty estimation layer that adopts a Monte Carlo Dropout structure or a variational Bayesian fully connected layer to obtain the initial prediction standard deviation; input the set of physical residual components into a residual calibration layer, take the absolute value of each physical residual component and input it into a single-layer neural network, generate a local scaling factor through the Softplus activation function, and then obtain the global residual scaling factor by weighted averaging according to the expert activation probability distribution. The initial prediction standard deviation is multiplied by the global residual scaling factor to obtain the calibrated prediction standard deviation, and the local prediction set is weighted and averaged using the probability values ​​in the expert activation probability distribution as weights to obtain the prediction mean. The decision output module is used to output the predicted mean and the calibrated predicted standard deviation as the prediction result and uncertainty index of the target process parameters.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent prediction method for refining and chemical equipment based on a physical information neural network as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent prediction method for refining and chemical equipment based on a physical information neural network as described in any one of claims 1 to 3.

Images