Calculation engine-based input-output method and system for coal chemical industry

Abstract

The invention discloses an input-output method and system realized based on a calculation engine in the coal chemical industry, and relates to the technical field of coal chemical industry, and the method comprises the steps: building a molecular model based on a reaction network of a coal molecular structure, and updating the parameters of the molecular model in real time through extended Kalman filtering; constructing a data fusion layer through a hybrid modeling technology, and constructing a coal chemical industry residual learning architecture by using a mechanism model and a neural network series structure; evaluating a data-driven model prediction confidence interval according to a Monte Carlo method; a digital twin verification platform is constructed based on hardware-in-the-loop test, and a data-driven adaptive iteration mechanism is constructed through an online and offline collaborative updating mechanism. The method has the beneficial effects that the medium component fluctuation is accurately captured, the output accuracy is improved, reliable evaluation dependence is provided, and the real-time response force and long-term stability of the system in a complex environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal chemical industry, in particular to an input-output method and system based on a computing engine for the coal chemical industry. BACKGROUND

[0002] Coal chemical industry involves gasification, liquefaction, coking and other processes, the composition of raw materials (coal) fluctuates greatly, and production parameters need to be dynamically adjusted to maximize energy efficiency; the input-output ratio of raw materials, energy and emissions needs to be accurately calculated to reduce costs and carbon emissions (such as the "double carbon" target requirement); traditional offline modeling cannot meet the rapid production scheduling needs under market fluctuations.

[0003] High-performance computing (HPC) and distributed architecture: using frameworks such as Spark and Flink to process massive process data and calculate material / energy balance in real time; numerical optimization algorithms: linear / nonlinear programming (such as CPLEX and Gurobi), genetic algorithms, etc., used to solve multi-objective optimization problems (such as minimum cost and maximum yield); industrial simulation engine: integration of process simulation software such as ASPEN Plus and gPROMS with custom algorithms to build a digital twin model.

[0004] The input-output software system of the coal chemical industry plays a key role in supporting production optimization, resource management and decision analysis, but currently has the following technical shortcomings: production data, laboratory information, equipment management, etc. systems run independently, data formats are not unified, manual cross-system integration is required, leading to multi-source data island problems; some software relies on periodic data import and cannot support real-time optimization, lacking real-time performance; there is a lack of effective time-series databases and machine learning models, making it difficult to mine the correlation between process parameters and energy efficiency, and the utilization rate of historical data is low; the kinetic model of core processes such as gasification and F-T synthesis often ignores the impact of coal quality fluctuations (such as ash melting point changes), and the simplified mechanism model leads to a simulation deviation of >10%; AI models trained based on single-plant data are difficult to migrate to other coal types or furnace types, and the small sample problem is prominent, and 5. The poor generalization ability of data-driven models leads to a high false alarm rate for abnormal detection. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] Therefore, the technical problem solved by the present application is that the existing input-output method implemented by the computing engine of the coal chemical industry has the problems of over-simplification of mechanism models, low real-time performance, and how to improve the generalization ability of data-driven models.

[0007] To solve the above technical problems, the present application provides the following technical solutions: an input-output method based on a computing engine for the coal chemical industry, comprising: constructing a molecular model based on a reaction network of a coal molecular structure, and updating molecular model parameters in real time using an extended Kalman filter; constructing a data fusion layer through a hybrid modeling technology, using a mechanism model and a neural network series structure to construct a coal chemical industry residual error learning architecture; evaluating a prediction confidence interval of a data-driven model according to a Monte Carlo method; constructing a digital twin verification platform based on a hardware-in-the-loop test, and constructing a data-driven adaptive iteration mechanism through an online and offline collaborative updating mechanism; the hybrid modeling technology combines high-fidelity mechanism modeling and data-driven compensation to construct a high-precision, interpretable process model of the input-output system of the coal chemical industry, and the mechanism and the data are deeply integrated through a hierarchical architecture.

[0008] As a preferred scheme of the input-output method based on a computing engine for the coal chemical industry, wherein: the construction of the molecular model comprises separating each group of the coal through hierarchical solvent extraction, analyzing and obtaining molecular structure parameters by combining spectral analysis technology, mapping process indicators of the coal into functional group types and parameters of aromaticity in the molecular model, and embedding medium characteristics in the molecular model.

[0009] As a preferred scheme of the input-output method based on a computing engine for the coal chemical industry, wherein: the updating of the molecular model parameters comprises dynamically augmenting to a state vector according to an extended Kalman filter algorithm, combining a discretized nonlinear model and a partial derivative output, online identifying coal molecular parameters and dynamically adjusting time-varying parameters in a coal molecular reaction network, and the time-varying parameters include functional group activity and reaction rate.

[0010] As a preferred scheme of the input-output method based on a computing engine for the coal chemical industry, wherein: the construction of the coal chemical industry residual error learning architecture comprises constructing a mechanism model output benchmark prediction value based on a physical equation, constructing a neural network to learn the residual error between the output of the real system and the prediction of the mechanism model, and optimizing the hybrid model through neural network training.

[0011] As a preferred scheme of the input-output method based on a computing engine for the coal chemical industry, wherein: the neural network training comprises an end-to-end training strategy, including a first stage and a second stage.

[0012] The first stage adopts a frozen mechanism model to train a neural network to compensate for residual errors, and the second stage adopts a joint fine-tuning to unfreeze the mechanism model parameters.

[0013] As a preferred scheme of the coal chemical industry input-output method based on the computing engine, wherein: the evaluation data-driven model prediction confidence interval comprises adding a Dropout layer in a neural network, keeping Dropout activation in the test stage, generating a prediction distribution through randomness, repeatedly executing forward propagation for the same input sample, generating different outputs each time Dropout, collecting continuous value prediction results for regression tasks, and quantifying the uncertainty of prediction through the statistical distribution characteristics of output prediction results.

[0014] As a preferred scheme of the coal chemical industry input-output method based on the computing engine, wherein: the construction of the digital twin verification platform comprises configuring a hardware system, developing a controlled object model based on a modeling tool and generating executable code for deployment to a real-time processor, configuring an I / O interface board card channel to connect a measured controller through a signal mapping table, integrating a fault injection module to simulate sensor short circuits, monitoring test data in real time, recording and analyzing results, and outputting an automated test report.

[0015] Another object of the present application is to provide a coal chemical industry input-output system based on a computing engine, which can construct a data fusion layer through a hybrid modeling technique, utilize a mechanism model and a neural network series structure, and construct a coal chemical industry residual error learning architecture, thereby solving the problem of excessive simplification of mechanism models in the current coal chemical industry computing engine implemented input-output method.

[0016] As a preferred scheme of the input-output system based on the computing engine in the coal chemical industry, wherein: including a model reconstruction layer module, a data fusion layer module, and an engineering implementation layer module; the model reconstruction layer module is used for separating various components of coal through hierarchical solvent extraction, obtaining structural parameters by combining spectral analysis, constructing a molecular model, mapping process indicators of coal into functional group characteristics in the molecular model, updating molecular model parameters in real time through extended Kalman filtering, augmenting time-varying parameters to a state vector, and performing dynamic updating of parameters through a discretized nonlinear model and a partial derivative output; the data fusion layer module is used for constructing a coal chemical industry residual error learning architecture through a mechanism model and a neural network series structure, outputting a benchmark prediction value based on a physical equation in the mechanism model, learning the residual error of a real system and the mechanism model in the neural network, adopting a two-stage training strategy, freezing the mechanism model in the first stage, training only the neural network to compensate for the residual error, and adopting joint fine-tuning in the second stage, unfreezing key parameters of the mechanism model, quantifying uncertainty by evaluating the prediction confidence interval of the data-driven model, adding a Dropout layer in the neural network as a regularization means, randomly discarding neurons according to a set probability, keeping the Dropout activation to generate a prediction distribution, performing repeated forward propagation on the same input sample, generating different outputs each time the Dropout is generated, and quantifying the prediction uncertainty according to the statistical distribution characteristics of the output prediction results; and the engineering implementation layer module is used for configuring a hardware system, constructing a digital twin verification platform, deploying a controlled object model developed based on a modeling tool to a real-time processor, configuring a signal mapping table to realize hardware connection with a measured controller, integrating a fault injection module to simulate abnormal working conditions, performing test execution and monitoring through configuration of test management software, monitoring test data records and analyzing results in real time, generating an automatic test report, establishing a trigger mechanism based on data distribution changes and cumulative data volume, and constructing a data-driven adaptive iteration mechanism through online and offline collaborative updates.

[0017] A computer device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the input-output method based on the computing engine in the coal chemical industry.

[0018] A computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the input-output method based on the computing engine in the coal chemical industry.

[0019] The coal chemical industry input-output method based on a computing engine provided by the application constructs a molecular model based on a reaction network of a coal molecular structure, updates molecular model parameters in real time by using an extended Kalman filter, accurately captures fluctuations in coal quality components, improves the accuracy of input and output, constructs a data fusion layer by using a hybrid modeling technology, constructs a coal chemical industry residual error learning architecture by using a mechanism model and a neural network series structure, improves the prediction accuracy and generalization ability of the mechanism model under complex working conditions, evaluates the prediction confidence interval of the data-driven model according to the Monte Carlo method, quantifies the inaccuracy of the prediction of key parameters of the coal chemical industry, provides reliable evaluation dependence, constructs a digital twin verification platform based on hardware-in-the-loop testing, constructs a data-driven adaptive iteration mechanism through an online and offline collaborative updating mechanism, and significantly improves the real-time response force and long-term stability of the system under complex environments. The application achieves better results in mechanism model construction, real-time performance, and improving the generalization ability of the data-driven model. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the 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 application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Fig. 1 The overall flowchart of the coal chemical industry input-output method based on a computing engine provided by the first embodiment of the application.

[0022] Fig. 2 The residual error learning architecture diagram of the coal chemical industry input-output method based on a computing engine provided by the second embodiment of the application.

[0023] Fig. 3 The adaptive iteration mechanism architecture diagram of the coal chemical industry input-output method based on a computing engine provided by the second embodiment of the application. DETAILED DESCRIPTION

[0024] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the application.

[0025] Embodiment 1, refer to Figs. 1-3For an embodiment of the present application, a coal chemical industry input-output method based on a computing engine is provided, comprising:

[0026] S1: Construct a molecular model based on a reaction network of coal molecular structure, and update the molecular model parameters in real time by using an extended Kalman filter.

[0027] Further, the construction of the molecular model comprises separating each family component of coal by hierarchical solvent extraction, obtaining molecular structure parameters by combining spectral analysis technology, mapping process indexes of coal into functional group types and aromaticity parameters in the molecular model, and embedding medium characteristics in the molecular model.

[0028] The coal macromolecule is connected by bridge chains such as benzene rings and heterocyclic rings, and the side chains comprise basic units such as alkyl groups, fluorine / sulfur functional groups, i.e., hydroxyl groups and carbonyl groups. 1 / 3 The process indexes of coal such as fixed carbon, volatile matter and sulfur content are mapped into functional group types and aromaticity parameters in the molecular model, so as to embed the quality characteristics of coal in the molecular model.

[0029] It should be noted that updating the molecular model parameters comprises dynamically augmenting to the state vector according to the extended Kalman filter algorithm, combining the discretized nonlinear model and the partial derivative output, online identifying the coal molecular parameters and dynamically adjusting the time-varying parameters in the coal molecular reaction network, and the time-varying parameters comprise functional group activity and reaction rate.

[0030] The time-varying parameters stator resistance (Rs) and inductance (Ld) are augmented to the state vector, the parameter dynamic model is assumed to be random walk, the parameter is allowed to slowly drift over time, the change rate of the parameter is controlled, the nonlinear model is discretized, then the Jacobian matrix is calculated, the state transition Jacobian needs to solve the partial derivative of the nonlinear function to each state component, i.e., the parameter, and finally the EKF iteration process is adopted.

[0031] It should also be noted that the molecular structure parameters are obtained by separating each family component of coal by hierarchical solvent extraction and combining spectral analysis to construct the molecular model, which solves the problems of static of traditional coal molecular model parameters and inability to dynamically respond, so that the coal molecular structure parameters can adaptively track and dynamically optimize the reaction network, and a more accurate real-time modeling tool at the molecular level is provided.

[0032] S2: Construct a data fusion layer by using a hybrid modeling technology, and construct a coal chemical industry residual error learning architecture by using a mechanism model and a neural network series structure.

[0033] Further, the coal chemical residual learning architecture is constructed by constructing a mechanism model to output a benchmark prediction value based on a physical equation, constructing a neural network to learn a residual between an output of an actual system and a prediction of the mechanism model, and optimizing the hybrid model through neural network training.

[0034] The coal chemical residual learning architecture is constructed as shown in Fig. 2 The current operating parameter oxygen-coal ratio and coal quality index input are first input into the trained mechanism model to generate a benchmark prediction value, the same parameters and coal quality characteristics, i.e., equipment state data, are input into the LSTM residual compensation network, the dynamic deviation not captured by the mechanism model is calculated, the output of the mechanism model is superimposed with the neural network compensation value, and the final corrected prediction result is output. In the training, the loss function is used for residual calculation, and the loss function is represented as:

[0035]

[0036] wherein, is the loss function, ɑ is the weight coefficient of the residual square term, Y is the actual monitoring value, Y m is the benchmark prediction value, Δ is the residual compensation, β is the weight coefficient of the regularization term, and ||Δ||1 is the norm of the residual compensation.

[0037] It should be noted that the neural network training includes an end-to-end training strategy, including a first stage and a second stage.

[0038] The first stage uses a frozen mechanism model to train the neural network compensation residual, and the second stage uses a joint fine-tuning to unfreeze the mechanism model parameters.

[0039] It should also be noted that the hybrid modeling technology combining the mechanism model and the neural network is used to construct the coal chemical residual learning architecture, the prediction correction is achieved by superimposing the compensation value, the two-stage training strategy is adopted, the problem of prediction value deviation caused by the simplification assumption of the traditional mechanism model is solved, the coal chemical process precision is higher, the dynamic adaptive prediction is achieved, and the operating condition optimization and control accuracy are significantly improved.

[0040] S3: Evaluate the prediction confidence interval of the data-driven model according to the Monte Carlo method.

[0041] Further, the evaluation of the prediction confidence interval of the data-driven model includes explicitly adding a Dropout layer in the neural network, keeping the Dropout activation in the test stage, generating a prediction distribution through randomness, repeatedly executing forward propagation for the same input sample, generating different outputs each time the Dropout is executed, collecting continuous value prediction results for the regression task, and quantifying the uncertainty of the prediction through the statistical distribution characteristics of the output prediction results.

[0042] It should be noted that the uncertainty quantification is achieved by evaluating the prediction confidence interval of the data-driven model, the Dropout layer is added in the neural network, the neurons are randomly discarded in the training stage according to the set probability of 0.2-0.5, the Dropout activation is maintained in the test stage, multiple (usually 50-100) forward propagations are performed on the same input sample, different outputs are generated each time due to Dropout, the output forms two categories, the first category is the softmax probability vector of the classification task, and the second category is the continuous value prediction of the regression task, the confidence interval calculation is performed, the prediction mean and the prediction variance are calculated respectively, the prediction mean is represented as:

[0043]

[0044] Wherein, μ is the prediction mean, T is the total number, y t is the prediction value of the tth, and t is the index value of the prediction.

[0045] The prediction variance is represented as:

[0046]

[0047] Wherein, σ 2 is the prediction variance, μ is the prediction mean, T is the total number, y t is the prediction value of the tth, and t is the index value of the prediction.

[0048] It should also be noted that the uncertainty quantification is achieved by evaluating the prediction confidence interval of the data-driven model through the Monte Carlo method, the mean and variance are calculated by using the distribution characteristics of the prediction results, the problem that the traditional prediction model cannot represent the uncertainty is solved, and the effect of enhancing the explainability and reliability of the model is achieved.

[0049] S4: Constructing a digital twin verification platform based on hardware-in-the-loop testing, constructing a data-driven adaptive iteration mechanism through online and offline collaborative updating mechanism.

[0050] Further, the construction of the digital twin verification platform includes configuring the hardware system, developing the controlled object model based on the modeling tool and generating executable code for deployment to the real-time processor, configuring the I / O interface board channel connection of the measured controller through the signal mapping table, integrating the fault injection module to simulate the short circuit of the sensor, monitoring the test data in real time, recording and analyzing the results, and outputting the automatic test report.

[0051] It should be noted that the hardware preparation includes the host computer, the real-time processor, the I / O interface board card, the real-time processor adopts the FPGA platform, and the simulation step length requirement is less than or equal to 1.25 microseconds for the power electronic system and less than or equal to 100 microseconds for the control system; the I / O resources are allocated, various types of board cards are configured according to the interface definition of the measured controller (ECU), and mechanical loads such as load motors and magnetic powder brakes are connected to support power level testing, a fault injection module is integrated, abnormal working conditions such as sensor short circuit and signal drift are simulated through a low-voltage fault injection box; on the software level, the controlled object model (battery thermal management model, motor dynamics model) is built based on MATLAB / Simulink or AMESim, code generation is supported and real-time constraints are met, test case editing, automatic execution, data monitoring and real-time analysis, and fault injection strategy configuration are realized through test management software, ECU pin definitions are parsed, signal mapping tables are generated and I / O board card channels are configured, adaptive harnesses are made, the compiled simulation model is deployed to the real-time processor, and key parameters such as voltage, current and temperature are monitored in real time, and the results are recorded and analyzed to generate an automatic test report.

[0052] Through online and offline collaborative updating, an adaptive iterative mechanism driven by data is constructed, as shown in Fig. 3 As shown, through the data pipeline, the online system and the offline platform are double-synchronized, when the data distribution deviation is greater than 5% or the cumulative data volume is greater than 100,000, incremental updating or full updating is triggered, and the updated model is uniformly managed through the version control warehouse.

[0053] It should also be noted that through the digital twin platform constructed based on hardware-in-the-loop testing, combined with the online and offline collaborative updating mechanism, the problem of insufficient dynamic adaptability of traditional test systems is solved, and combined with version control, iterative management is realized, achieving high-precision real-time simulation and full-coverage testing of abnormal working conditions, and improving the verification efficiency and reliability of complex systems.

[0054] Embodiment 2, which is an embodiment of the present application, provides an input-output system based on a computing engine in the coal chemical industry, comprising a model reconstruction layer module, a data fusion layer module, and an engineering implementation layer module.

[0055] The model reconstruction layer module is used to separate various components of coal by hierarchical solvent extraction, obtain structural parameters by combining spectral analysis, construct a molecular model, map process indicators of coal to functional group characteristics in the molecular model, update molecular model parameters in real time through extended Kalman filtering, and perform dynamic updating of parameters through discretization of nonlinear models and partial derivative outputs.

[0056] The data fusion layer module is used for constructing a coal chemical industry residual error learning architecture through a mechanism model and a neural network series structure, the mechanism model outputs a benchmark prediction value based on a physical equation, the neural network learns a residual error between a real system and the mechanism model, a two-stage training strategy is adopted, in a first stage, the mechanism model is frozen, and only the neural network is trained to compensate the residual error, in a second stage, joint fine-tuning is adopted, key parameters of the mechanism model are unfrozen, uncertainty is quantified by evaluating a prediction confidence interval of a data-driven model, a Dropout layer is added to the neural network as a regularization means, neurons are randomly discarded according to a set probability, a Dropout activation is kept to generate a prediction distribution, repeated forward propagation is performed on the same input sample, different outputs are generated by the Dropout each time, and prediction uncertainty is quantified according to statistical distribution characteristics of the output prediction results.

[0057] The engineering implementation layer module is used for configuring a hardware system, constructing a digital twin verification platform, deploying a controlled object model developed based on a modeling tool to a real-time processor, configuring a signal mapping table to realize hardware connection with a measured controller, integrating a fault injection module to simulate abnormal working conditions, performing test execution and monitoring through configuration of test management software, monitoring test data records and analyzing results in real time, generating an automatic test report, establishing a trigger mechanism based on data distribution changes and cumulative data volume, and constructing a data-driven adaptive iteration mechanism through online and offline collaborative updating.

Claims

1. A coal chemical industry input-output method based on a computing engine, characterized in that, The method comprises the following steps: a molecular model is constructed based on a reaction network of coal molecular structure, and a molecular model parameter is updated in real time by using an extended Kalman filter; a data fusion layer is constructed by using a hybrid modeling technology, a mechanism model and a neural network are connected in series, and a coal chemical industry residual error learning architecture is constructed; a prediction confidence interval of a data-driven model is evaluated according to a Monte Carlo method; a digital twin verification platform is constructed based on a hardware-in-the-loop test, an online-offline cooperative updating mechanism is constructed, and a data-driven adaptive iteration mechanism is constructed; the hybrid modeling technology combines high-fidelity mechanism modeling and data-driven compensation to construct a high-precision and interpretable process model of an input-output system of the coal chemical industry, and the mechanism and the data are deeply fused through a hierarchical architecture.

2. The input-output method based on a computing engine for coal chemical industry according to claim 1, characterized in that: The method comprises the following steps of constructing a molecular model, different groups of components of coal are separated by hierarchical solvent extraction, and molecular structure parameters are obtained by combining spectral analysis technology, process indexes of coal are mapped to functional group types and parameters of aromaticity in the molecular model, and medium characteristics are embedded in the molecular model.

3. The input-output method based on a computing engine for coal chemical industry according to claim 1, characterized in that: The method comprises the following steps of updating a molecular model parameter, a state vector is dynamically augmented according to an extended Kalman filter algorithm, a discrete nonlinear model and a partial derivative output are combined, coal molecular parameters are identified online, and time-varying parameters in a coal molecular reaction network are dynamically adjusted, the time-varying parameters include functional group activity and reaction rate.

4. The input-output method based on a computing engine for coal chemical industry according to claim 1, characterized in that: The method comprises the following steps of constructing a coal chemical industry residual error learning architecture, a mechanism model is constructed based on a physical equation to output a benchmark prediction value, a neural network is constructed to learn a residual error between an actual system output and a mechanism model prediction, and the hybrid model is optimized by neural network training.

5. The coal chemical industry based input-output method implemented by a computing engine of claim 4, wherein: The method comprises the following steps of neural network training, an end-to-end training strategy includes a first stage and a second stage; in the first stage, a neural network is trained to compensate for the residual error by freezing the mechanism model, and in the second stage, the mechanism model parameters are jointly fine-tuned by unfreezing the mechanism model.

6. The coal chemical industry based input-output method implemented by a computing engine of claim 1, wherein: The method comprises the following steps of evaluating a prediction confidence interval of a data-driven model, a Dropout layer is explicitly added to the neural network, the Dropout activation is kept in the test stage, a prediction distribution is generated by randomness, a forward propagation is repeatedly performed for the same input sample, different outputs are generated by Dropout each time, continuous value prediction results are collected for a regression task, and the uncertainty of prediction is quantified by the statistical distribution of output prediction results.

7. The coal chemical industry based input-output method implemented by a computing engine of claim 1, wherein: The method comprises the following steps of constructing a digital twin verification platform, a hardware system is configured, a controlled object model is developed based on a modeling tool and executable code is generated to be deployed to a real-time processor, an I / O interface board card channel is connected to a measured controller through a signal mapping table configuration, a fault injection module is integrated to simulate a sensor short circuit, test data are monitored in real time, results are recorded and analyzed, and an automatic test report is output.

8. A coal chemical industry input-output system based on a computing engine, which adopts the coal chemical industry input-output method based on a computing engine according to any one of claims 1 to 7. The method comprises the following steps of including a model reconstruction layer module, a data fusion layer module and an engineering implementation layer module; the model reconstruction layer module is used to separate different groups of components of coal by hierarchical solvent extraction, obtain structure parameters by combining spectral analysis, construct a molecular model, map process indexes of coal to functional group characteristics in the molecular model, update a molecular model parameter in real time by using an extended Kalman filter, augment time-varying parameters to a state vector, and dynamically update the parameters by using a discrete nonlinear model and a partial derivative output; The data fusion layer module is used to construct a coal chemical industry residual error learning architecture through a mechanism model and a neural network series structure, the mechanism model outputs a benchmark prediction value based on a physical equation, the neural network learns a residual error between a real system and the mechanism model, a two-stage training strategy is adopted, in the first stage, the mechanism model is frozen, only the neural network is trained to compensate the residual error, in the second stage, joint fine-tuning is adopted, key parameters of the mechanism model are unfrozen, uncertainty is quantified by evaluating a prediction confidence interval of a data-driven model, a Dropout layer is added to the neural network as a regularization means, neurons are randomly discarded according to a set probability, a Dropout activation is generated to maintain the prediction distribution, repeated forward propagation is performed on the same input sample, different outputs are generated by Dropout each time, and prediction uncertainty is quantified according to statistical distribution characteristics of the output prediction results; The engineering implementation layer module is used to configure a hardware system, construct a digital twin verification platform, deploy a controlled object model developed based on a modeling tool to a real-time processor, configure a signal mapping table to realize hardware connection with a measured controller, integrate a fault injection module to simulate abnormal working conditions, perform test execution and monitoring through configuration of test management software, monitor test data records in real time and analyze results, generate an automatic test report, establish a trigger mechanism based on data distribution changes and cumulative data volume, and construct a data-driven adaptive iteration mechanism through online and offline collaborative updates. 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 input-output method for the coal chemical industry based on the computing engine in 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 input-output method for the coal chemical industry based on the computing engine in any one of claims 1 to 7.

Images