Fine chemical product purity data prediction method based on deep learning

Abstract

The present application belongs to the technical field of chemical data prediction, and more specifically, the present application relates to a fine chemical product purity data prediction method based on deep learning, which comprises: calculating the life damage factor at each time based on historical performance data, the life damage factor being positively correlated with the current cumulative running time at the time and the total number of regenerations, and being negatively correlated with the rated initial maximum life value; calculating the physical correction index at a time, the physical correction index being positively correlated with the life damage factor, and being negatively correlated with the platinum load, the palladium load and the aluminum-titanium molar ratio; inputting real-time working condition parameters into a pre-trained deep learning model for feedforward calculation to obtain the target prediction purity value at the corresponding time; and based on the target prediction purity value and the physical correction index, the prediction purity value at the corresponding time is calculated, effectively improving the accuracy of the product purity data prediction result.

Description

Technical Field

[0001] This invention belongs to the field of chemical data prediction technology. More specifically, this invention relates to a method for predicting the purity data of fine chemical products based on deep learning. Background Technology

[0002] Deep learning is increasingly being applied to the field of fine chemical production for monitoring and predicting key data in the production process. Among these, product purity is a core indicator for measuring the efficiency of chemical process conversion. Therefore, building a high-precision purity prediction model is of great significance for achieving intelligent management and control in the fine chemical industry.

[0003] In specific continuous flow production scenarios in fine chemicals, such as the continuous gas-phase decarbonylation process of furfural to furan, the advancement of the process is highly dependent on the activity of supported noble metal catalysts. Chinese patent application CN115845837A discloses a catalyst for the continuous gas-phase decarbonylation of furfural to furan and its preparation method. Specifically, it discloses the use of impregnation to load platinum and palladium active components onto a support, followed by reduction activation and application in the furfural decarbonylation reaction. It also clearly describes the process characteristics of catalyst deactivation and the ability to regenerate the catalyst in situ more than five times after its activity decreases.

[0004] However, in actual long-term industrial operation, catalysts undergo carbon deposition and deactivation. While in-situ regeneration can restore some activity, irreversible sintering and loss of active sites occur with each regeneration cycle, leading to a gradual shortening of catalytic lifetime and affecting furan purity. Deep learning prediction methods can be used to predict furan purity, but directly using real-time operating conditions such as reaction temperature, pressure, and space velocity as input variables for deep learning models ignores the physical fatigue damage process of the catalyst. When predicting purity at the end of long-term operation, deep learning models lack evaluation of the deactivation and regeneration cycle logic, potentially resulting in significant lag and bias. Consequently, in the later stages of catalyst operation, the model's purity predictions are much higher than actual measurements, failing to provide accurate decision support for industrial production.

[0005] Therefore, the key to improving the accuracy of purity prediction in fine chemicals lies in how to construct a deep learning model that can integrate the physical fatigue state of catalysts to accurately obtain the purity prediction results. Summary of the Invention

[0006] To address the technical problem of how to construct a deep learning model that can integrate the physical fatigue state of catalysts to accurately predict furan purity, this invention proposes a deep learning-based method for predicting the purity of fine chemical products. This method includes the following steps: Step 1: Obtain the operating parameters of the catalyst under test. These parameters include real-time operating parameters and historical performance data for each moment. Historical performance data includes current cumulative operating time, total number of regenerations, rated initial maximum lifetime, platinum loading, palladium loading, and aluminum-titanium molar ratio. Real-time operating parameters include reaction temperature, reaction pressure, and mass hourly space velocity (MHV). Step 2: Calculate the lifetime damage factor for each moment based on the historical performance data. The lifetime damage factor is positively correlated with the current cumulative operating time and total number of regenerations at that moment, and negatively correlated with the rated initial maximum lifetime. Step 3: Calculate the physical correction index for a given moment. The physical correction index is positively correlated with the lifetime damage factor and negatively correlated with platinum loading, palladium loading, and aluminum-titanium molar ratio. Step 4: Input the real-time operating parameters into a pre-trained deep learning model for feedforward calculation to obtain the target predicted purity value for the corresponding moment. Based on the target predicted purity value and the physical correction index, calculate the predicted purity value for the corresponding moment.

[0007] This invention provides a deep learning-based method for predicting the purity of fine chemical products, which can effectively improve the accuracy of purity prediction results. In obtaining purity prediction results, this invention addresses the problem that existing technologies, relying solely on real-time operating conditions such as reaction temperature and pressure as input variables, are prone to lag and significant bias in end-stage predictions during long-term operations. Therefore, this invention incorporates historical catalyst performance data to calculate lifetime damage factors and physical correction indices, assesses the microscopic fatigue degradation process of the catalyst in physical space, and integrates this assessment into a deep learning model. This effectively solves the problem of overly high purity predictions in the later stages of continuous chemical production, providing high-precision purity predictions for long-term industrial operations, thereby supporting intelligent and precise control in actual production.

[0008] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the acquisition of the operating parameters of the catalyst to be tested includes: extracting the cumulative running time of the current reaction and the total number of regenerations corresponding to the current catalyst at each sampling time through a distributed control system; obtaining the rated initial maximum lifetime value, the platinum loading, palladium loading and aluminum-titanium molar ratio in the catalyst according to the catalyst loading material list; and collecting the reaction temperature, reaction pressure and mass hourly space velocity in real time through field instruments.

[0009] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the calculation of the lifetime damage factor at each moment based on historical performance data includes: obtaining the deviation of the current cumulative running time from the rated initial maximum lifetime value at a given moment; obtaining the regeneration penalty multiplication coefficient, which is used to characterize the irreversible grain agglomeration penalty caused by thermal shock; and using the product of the deviation and the regeneration penalty multiplication coefficient as the lifetime damage factor at the current moment. The regeneration penalty multiplication coefficient is obtained by using the natural constant as the base and the product of the initial activity decay constant and the current total number of regenerations as the exponent for power operation to obtain the regeneration penalty multiplication coefficient.

[0010] This invention constructs a physical-mathematical expression for a lifetime damage factor that includes the penalty effect of cumulative running time and regeneration cycles. This enables deep learning models to identify damage evolution such as irreversible sintering caused by multiple regeneration cycles, thereby effectively reducing the risk of purity predictions deviating significantly from actual performance.

[0011] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the method for obtaining the initial activity decay constant includes: conducting accelerated aging tests using a reference catalyst with the same noble metal loading as the catalyst under test, and determining the residual rate of active sites of the reference catalyst at different regeneration cycles; using the regeneration cycle as the independent variable and the logarithm of the residual rate of active sites as the dependent variable, performing linear regression using the least squares method, and extracting the absolute value of the slope of the regression line as the decay benchmark value; constructing a physical effective range of the activity decay constant based on multiple sets of decay benchmark values ​​of the reference catalyst at different reaction temperatures, and determining the upper and lower limits of its physical effective range; calculating the deviation between the initial aluminum-titanium molar ratio of the current catalyst under test and the standard molar ratio of the reference catalyst, and performing linear interpolation within the physical effective range based on the deviation to obtain the initial activity decay constant.

[0012] This invention obtains the residual rate of active sites through accelerated aging tests, performs linear regression using the least squares method, and interpolates within the physically effective range based on the actual deviation, accurately calibrating the initial activity decay constant of the catalyst under different operating conditions, thereby improving the robustness and universality of the prediction model in complex chemical application scenarios.

[0013] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the fine chemical product is furan prepared by continuous gas-phase decarbonylation of furfural; the catalyst is a supported platinum-palladium bimetallic catalyst.

[0014] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the calculation of the physical correction index at a given time step includes: ; in, For a moment The physical correction index, To preset the product degradation coefficient, For a moment Lifespan damage factors, For empirical coupling factors, , , These correspond to the platinum loading, palladium loading, and aluminum-titanium molar ratio in the catalyst, respectively.

[0015] This invention provides a precise physical correction index. It introduces a nonlinear damping denominator term composed of noble metal loading and aluminum-titanium ratio to construct the physical correction index. This not only evaluates the catalyst's resistance to deactivation but also transforms the degree of lifetime damage into a dynamic degradation penalty base that can be identified by a deep learning model, ensuring that the predicted data of the production process does not violate the fundamental laws of chemical kinetics.

[0016] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the step of calculating the predicted purity value at the corresponding time based on the target predicted purity value and the physical correction index includes: constructing a purity deviation term based on the product of the target predicted purity value and the physical correction index; and subtracting the purity deviation term from the target predicted purity value to obtain the predicted purity value at the current time.

[0017] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the method for obtaining the initial activity decay constant further includes: obtaining the measured purity value at the production site, using the difference between the predicted purity value and the measured purity value as a feedback error; and dynamically adjusting the initial activity decay constant based on the feedback error and the current total number of regenerations.

[0018] This invention addresses the problem that traditional fixed-parameter prediction methods are prone to large cumulative drift errors over long lifecycles. Based on this, the invention introduces an adaptive feedback error mechanism based on on-site chromatograph measurements. This mechanism can dynamically fine-tune the initial activity decay constant, effectively preventing and reducing the cumulative deviation caused by individual microscopic differences between different batches of catalysts and high-frequency noise fluctuations, thereby improving the steady-state high fidelity of purity prediction in the late stage of industrial operation.

[0019] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, the dynamic adjustment of the initial activity decay constant based on feedback error and the current total number of regenerations includes: ; in, For use in time updates The initial activity decay constant, For a moment The initial activity decay constant, The convergence gain factor, For a moment The predicted purity value, For a moment The measured purity value obtained by the chromatograph. For a moment The total number of regenerations.

[0020] According to the deep learning-based method for predicting the purity of fine chemical products provided by the present invention, after calculating the predicted purity value at the corresponding time, the method further includes: automatically switching to in-situ regeneration mode to burn off carbon deposits when the predicted purity value decreases to a preset activity threshold; after each in-situ regeneration program is triggered, the logic controller automatically sends a pulse counting signal to the distributed control system to update the total number of regenerations, and uses the updated total number of regenerations in the calculation of subsequent times.

[0021] The present invention has the following beneficial effects: Based on the above technical solution, the present invention provides a deep learning-based method for predicting the purity of fine chemical products. This method calculates a lifetime damage factor using historical catalyst performance data to assess irreversible grain sintering and active site loss caused by repeated in-situ regeneration. A physical correction index is constructed based on this lifetime damage factor and the catalyst's intrinsic deactivation resistance parameter. A nonlinear damping denominator term transforms the degree of microscopic fatigue degradation of the support framework into a dynamic penalty base for purity prediction. Real-time operating parameters are input into the deep learning model to obtain the target predicted purity value. A purity deviation term is constructed using the product of the target predicted purity value and the physical correction index. This deviation term is subtracted from the target predicted purity value to obtain the final predicted purity value, thus eliminating prediction lag and severe positive deviation caused by the lack of information on catalyst physical aging. This effectively solves the problem of predicted purity values ​​being much higher than actual measured values ​​due to neglecting catalyst physical fatigue damage, thereby significantly improving the accuracy of product purity prediction during long-term operation of fine chemicals. Attached Figure Description

[0022] Figure 1 A flowchart illustrating the steps of a deep learning-based method for predicting the purity of fine chemical products, as provided in an embodiment of the present invention. Figure 2 A schematic diagram of a process for the continuous gas-phase decarbonylation of furfural to prepare furan, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the trend of catalyst lifetime damage factors provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0024] Please see Figure 1 , Figure 1 A flowchart illustrating the steps of a deep learning-based method for predicting the purity of fine chemical products, as provided in this embodiment of the invention, is shown below. The method includes the following steps: S1: Obtain the operating parameters of the catalyst to be tested.

[0025] The operating parameters include real-time operating parameters and historical performance data for each moment. Historical performance data includes current cumulative running time, total number of regenerations, rated initial maximum lifetime, platinum loading, palladium loading, and aluminum-titanium molar ratio; real-time operating parameters include reaction temperature, reaction pressure, and mass hourly space velocity.

[0026] It is understandable that in the fine chemical production process of furfural gas-phase continuous decarbonylation to furan, the fine chemical product is furan prepared by furfural gas-phase continuous decarbonylation; the catalyst is a supported platinum-palladium bimetallic catalyst.

[0027] See also Figure 2 As shown, Figure 2 The schematic diagram of a continuous gas-phase decarbonylation process for preparing furan from furfural provided in this embodiment of the invention includes four modules: a gas mixing module, a liquid vaporization module, a catalytic reaction module, and a product collection module.

[0028] Specifically, during the production process, nitrogen gas, the purging gas, is supplied from an N2 cylinder. After being depressurized by a pressure reducing valve, it enters a mass flow meter for flow measurement. Hydrogen gas, the reaction gas, is supplied from an H2 cylinder. After being depressurized by a pressure reducing valve, it directly purges the downstream pipeline. One-way valves are installed at both ends to prevent gas backflow. Furfural vapor is generated by a precision vaporizer and precisely measured by a precision constant flow pump.

[0029] In the normal preparation process, hydrogen enters the vaporization mixer and is preheated and mixed with the steam generated from the vaporization of furfural. This mixture then flows downwards into a vertical reactor, which contains a catalyst bed containing a supported platinum-palladium bimetallic catalyst. During the reaction, the reactor employs a 30-segment programmed temperature control to ensure the reaction temperature remains within the range of 200 to 350°C, with a temperature control accuracy of ±1°C. A back pressure valve is installed at the reactor outlet to control the reaction pressure; a safety valve is installed upstream of the back pressure valve to prevent system malfunctions and excessive pressure. The reacted gas is condensed in a condenser and then enters a gas-liquid separator to remove waste gas. The product is collected and its purity is measured using an online chromatograph. The condenser uses 316L coils and is cooled by circulating water. The gas-liquid separator is placed in a low-temperature constant-temperature bath with a temperature ≤10°C.

[0030] When the predicted purity value is below 95%, nitrogen and a small amount of air are introduced for oxidation for 48 hours to burn off the carbon deposits.

[0031] For example, a distributed control system can extract the cumulative running time of the current reaction and the total number of regenerations corresponding to the current catalyst at each sampling moment. The rated initial maximum lifetime, the platinum loading, palladium loading, and aluminum-titanium molar ratio in the support are obtained based on the catalyst loading manifest. Reaction temperature, reaction pressure, and mass hourly space velocity are collected in real time using field instruments.

[0032] The data acquisition frequency of the field instruments can be set to once per second, and the high-frequency noise caused by valve action is filtered out by the moving average filtering algorithm to ensure that the operating parameters input to the neural network are smooth and reliable.

[0033] The initial maximum lifetime value can be set to 264 hours, the loading of platinum and palladium in the catalyst are dimensionless percentages, and the aluminum-titanium molar ratio is a dimensionless constant.

[0034] It is understandable that the raw sensor data may be out of sequence. If it is directly input into the neural network, it will cause logical conflicts in the feature space of different dimensions, resulting in serious lag and bias when predicting the purity at the end of a long-term operation.

[0035] Based on this, the embodiments of the present invention can also perform preprocessing such as time alignment and missing data interpolation on the various collected data. The specific settings can be made according to actual needs, and the embodiments of the present invention will not be elaborated here.

[0036] Thus, by collecting and filtering full-dimensional data on catalyst operating conditions and physical parameters, the embodiments of the present invention can effectively establish data features containing historical memories of the equipment, thereby providing an accurate data foundation for the subsequent construction of catalyst fatigue damage logic.

[0037] S2: Calculate the lifetime damage factor at each time point based on historical performance data.

[0038] Among them, the life damage factor is positively correlated with the current cumulative running time and the total number of regenerations at that moment, and negatively correlated with the rated initial maximum life value.

[0039] It should be noted that in the continuous flow production process of furfural gas-phase continuous decarbonylation to furan, the catalyst state includes a fresh state and a regenerated state. The fresh state is the initial high-activity state of the catalyst when it is first loaded into the reactor and has not undergone any regeneration. The regenerated state is the working state of the catalyst after multiple oxidations by air to remove carbon deposits.

[0040] In supported platinum-palladium bimetallic catalysts, platinum (Pt) and palladium (Pd), as active components, directly determine the conversion efficiency of furfural decarbonylation. The aluminum-titanium composite support provides skeletal support for the active metals, and its size determines the heat resistance and acid-base properties of the support. However, carbon deposition on the catalyst during production leads to a decrease in the purity of the furan product. In-situ regeneration technology can restore some of the activity lost due to carbon deposition and deactivation, but with increasing regeneration cycles, pore collapse and metal grain aggregation cause irreversible sintering and loss of active sites. If the original operating time is directly input into the prediction model, the model will be unable to identify the performance differences between catalysts in a fresh state and those in a repeatedly regenerated state under the same operating conditions.

[0041] Based on this, embodiments of the present invention can establish the evolution logic of catalyst active sites with operating time and regeneration frequency, calculate lifetime damage factors to assess the aging changes of such structures, thereby reducing the possibility that the purity prediction value given by the prediction model in actual industrial applications is much higher than the actual measured value.

[0042] Specifically, based on the theory of heterogeneous catalysis, it is known that the thermodynamic fatigue of the catalyst support skeleton continues throughout the entire life cycle. Therefore, after each in-situ regeneration, the cumulative operating time of the catalyst will continue to accumulate, so as to truly reflect the absolute aging degree of the support on a macroscopic time scale.

[0043] In supported bimetallic catalysts, in-situ regeneration through high-temperature carbon deposition can expose masked active sites, but it easily induces Ostwald ripening and agglomeration of metal grains, a typical thermal sintering deactivation mechanism. According to the thermodynamics of sintering deactivation, the permanent loss of metal active surface area increases non-linearly with the number of thermal shocks during regeneration. Therefore, a regeneration penalty multiplication factor can be constructed from the base characteristics of this sintering deactivation model to characterize the irreversible grain agglomeration penalty caused by thermal shock. Even without catalyst regeneration, the alumina support undergoes a slow crystal transformation, and the microporous structure gradually collapses due to thermodynamic fatigue; this change serves as the baseline for the catalyst's usable life. With each regeneration, the active metal components platinum and palladium undergo intense Ostwald ripening, and the metal particles continue to grow.

[0044] In this embodiment of the invention, the lifetime damage factor at each moment can be constructed by combining the absolute aging degree and the regeneration penalty multiplication factor on the above macroscopic time scale, wherein the lifetime damage factor is positively correlated with the loss of activity.

[0045] For example, when obtaining the lifetime damage factor at the current moment, the product of the deviation of the current cumulative running time from the initial maximum lifetime value of the catalyst and the regeneration penalty multiplication factor can be used as the lifetime damage factor at the current moment.

[0046] Understandably, although the removal of carbon deposits after regeneration restores the active components, the thermodynamic fatigue and metal sintering damage of the carrier skeleton accumulate monotonically over time. Therefore, cumulative time is used instead of single continuous time.

[0047] The construction logic of the regeneration penalty multiplication coefficient is based on the classical general deactivation mechanics model. First, according to the characteristics of thermal sintering deactivation, the change in activity with the number of thermal shocks is modeled as a first-order decay differential equation. Second, by performing an integral transformation on the first-order decay differential equation, an exponential decay mapping relationship between the catalyst residual activity and the number of regenerations is derived. Finally, the inverse change characteristic of the mapping relationship is extracted and defined as a penalty multiplication term to characterize the nonlinear lifetime loss aggravation effect caused by the increase of the number of regenerations. The algebraic form of the regeneration penalty multiplication coefficient after the final transformation is: using the natural constant as the base, and using the product of the initial activity decay constant and the current total number of regenerations as the exponent, the regeneration penalty multiplication coefficient is obtained by exponentiation.

[0048] The preset initial activity decay constant is used to characterize the thermal stability of the carrier skeleton.

[0049] Furthermore, when The current time indicates the initial moment of chemical production. To accurately obtain the initial activity decay constant in the above formula, this embodiment of the invention conducts an accelerated aging test on a reference catalyst with the same noble metal loading as the catalyst under test. The number of effective chemisorption sites on the catalyst surface is measured and the residual rate of active sites is recorded in real time under different regeneration cycles. With the number of regeneration cycles as the independent variable and the logarithm of the residual rate of active sites as the dependent variable, a linear regression is performed using the least squares method, and the absolute value of the slope of the regression line is extracted as the decay benchmark value. Based on multiple sets of decay benchmark values ​​of the reference catalyst at different reaction temperatures such as 250℃, 300℃, and 350℃, a physical effective range of the activity decay constant is constructed, and the upper and lower limits of the physical effective range are determined. The deviation between the initial aluminum-titanium molar ratio of the catalyst under test and the standard molar ratio of the reference catalyst is calculated, and linear interpolation is performed within the physical effective range based on the deviation to obtain the initial activity decay constant.

[0050] The maximum number of regenerations can be set to 5, and the specific number can be set according to actual needs.

[0051] To illustrate the calculation method of the above relationship: If the cumulative running time at a certain moment is 132 hours, the rated initial maximum lifetime is 264 hours, the initial activity decay constant obtained by interpolation is 0.2, and the current total number of regenerations is 3, substituting these values ​​into the calculation yields a lifetime damage factor of 0.911.

[0052] Please see Figure 3, Figure 3 This is a schematic diagram illustrating the trend of catalyst lifetime damage factors provided in an embodiment of the present invention.

[0053] As shown in the figure, the lifetime damage factor exhibits a nonlinear evolution trajectory consisting of multiple rising wave bands over time. The overall envelope shows a clear monotonically increasing trend, with each cycle showing a significant step increase compared to the previous cycle. As the number of regeneration cycles increases, the rising slope of the lifetime damage factor gradually becomes steeper, and the length of the cycle in which the peak reaches the limit value shortens successively, reflecting the nonlinear acceleration of the catalyst deactivation rate.

[0054] Thus, by extracting and calculating the features of the lifespan damage factor, the embodiments of the present invention can effectively transform the complex physical degradation process into a dynamic penalty base that can be read by the deep learning model, thereby providing an accurate data foundation for subsequent correction of the ideal purity prediction value.

[0055] S3: Calculate the physical correction exponent at each time step.

[0056] The physical correction index is positively correlated with the life damage factor, and negatively correlated with platinum loading, palladium loading and aluminum-titanium molar ratio.

[0057] It should be noted that, based on the steps described above, the degree of catalyst damage can be extracted. This step needs to convert this degree into a direct physical degradation penalty term for purity prediction. The total amount of noble metals in the bimetallic catalyst and their dispersion on a specific aluminum-titanium framework do not have an infinitely linear effect on the catalyst's anti-coking ability, but rather exhibit a surface site saturation effect at the microscopic level. If the degree of damage is not converted, the lifetime damage factor cannot be logically coupled with the product distribution predicted by subsequent deep learning. If the limited improvement of the sizing parameters against the coking limit is not evaluated, the correction results given by the model may violate the fundamental law of mass conservation and fail to reflect the true law that the large amount of furfural residue at the end of catalyst operation leads to a decrease in furan purity.

[0058] Based on this, embodiments of the present invention can introduce a nonlinear empirical saturation term with a lower bound convergence constraint to construct a physical correction exponent. In embodiments of the present invention, the relationship of the physical correction exponent is obtained as follows: ; in, For a moment The physical correction index is used to characterize the intensity of the degradation penalty applied to the final purity prediction, and its value ranges from 0 to 1; optionally, in the embodiments of the present invention, it can be set as a floating-point number that fluctuates dynamically over time, and can be set according to actual needs. This is the preset product degradation coefficient. For a moment Lifespan damage factors. This is the empirical coupling factor. For platinum loading, For palladium loading, The aluminum-titanium molar ratio is used to characterize the metallic activity of the catalyst and the acidity of the support.

[0059] The product degradation coefficient characterizes the natural degradation tendency of the catalyst and serves as a benchmark for purity degradation. Setting this value too small may overlook the actual drift at the end of the catalyst's lifespan; setting it too large may cause the predicted value to decay prematurely. This value can be obtained by continuously running the catalyst in its fresh state for 100 hours, collecting the measured product purity value every 10 hours. Using running time as the x-axis and purity as the y-axis, a linear fit is used to obtain the slope of purity decrease, and the absolute value of this slope is taken as the product degradation coefficient. Optionally, for the furfural-to-furan process, the value range in this embodiment can be set to 0.05 to 0.15.

[0060] The empirical coupling factor is used to calibrate the combined anti-coking strength gain of precious metal content and aluminum-titanium ratio. A larger empirical coupling factor results in a larger denominator value for the same metal loading, indicating a more robust physical structure and less purity penalty for the catalyst, thus significantly reducing the physical correction exponent. To obtain this value, a dataset can be constructed by collecting historical data from at least three complete catalytic cycles. Each cycle includes the complete process from fresh run to the regeneration threshold, in-situ regeneration, and a second run, specifically including temperature, pressure, space velocity, measured purity, number of regenerations, precious metal loading, and aluminum-titanium molar ratio. The candidate range for the empirical coupling factor is set to 0.1 to 0.5, with a step size of 0.05. For each candidate value, the root mean square error (RMSE) between the predicted purity and the measured purity of the entire dataset is calculated. The empirical coupling factor corresponding to the minimum RMSE is taken as the final value.

[0061] The denominator in the above equation provides a nonlinear damping term, which originates from the synergistic anti-coking limit effect between the bimetallic active sites and the acidic support. Since there is a physical upper limit to the dispersion of metal grains that the support surface can accommodate, increasing the noble metal loading and the aluminum-titanium ratio results in diminishing marginal returns to the gain in suppressing carbon deposition in the pores. This nonlinear damping structure effectively characterizes this physical limit, ensuring that even with extreme input parameters, the physical correction exponent remains controlled and convergent, without generating division-to-zero or divergence errors that violate the fundamental law of conservation of matter. Therefore, the larger the value of the denominator, the more the physical correction exponent, as a penalty term, will be suppressed, and the corresponding final data value will decrease, exhibiting a negative correlation.

[0062] Correspondingly, since the lifetime damage factor is the fundamental driving force of degradation, the larger the lifetime damage factor, the larger the numerator, and the larger the corresponding physical correction index data value, showing a positive correlation.

[0063] Based on the above steps, the physical correction index at each time step can be obtained. For example, to obtain the physical correction index at one time step: if the product degradation coefficient is 0.1, the empirical coupling factor is 0.2, the catalyst platinum loading is 0.5, the palladium loading is 0.2, and the aluminum-titanium molar ratio is 30, the calculated lifetime damage factor is 0.911. Then the denominator is... The final calculated physical correction index is approximately 0.0175.

[0064] Thus, by performing nonlinear mapping calculations on the physical correction index, this embodiment of the invention can effectively avoid division-to-zero errors and divergence risks caused by absolute single proportions, ensuring that the output does not deviate from the fundamental laws of chemical reaction engineering, thereby providing an accurate data foundation for subsequent feature-level fusion of neural networks.

[0065] S4: Obtain the predicted purity value at each time step.

[0066] Understandably, fine chemical production involves minute disturbances such as feed flow rates, necessitating the nonlinear fitting capabilities of deep learning to handle such dynamic, high-frequency information. Conventional deep learning models typically use a Softmax layer to output class probabilities, but furan purity is an independent normalized intensity value, not a probability distribution of mutually exclusive classes, making a Softmax layer unsuitable.

[0067] Based on this, embodiments of the present invention can employ a network structure with a Sigmoid activation layer to fuse real-time process parameters and physical correction exponents at the feature level, thereby effectively avoiding the possibility of semantic contradictions. The loss function can be mean squared error; the network structure can be a 3-5 layer fully connected network with 32-128 neurons per layer, and the activation function is the ReLU function; specific settings can be made according to actual needs.

[0068] For example, when pre-training a deep learning model, a historical operating dataset can be constructed first, which contains historical operating parameters for each historical moment. The historical operating parameters are used as input features, and the measured product purity value label is used as the target. The deep learning model is trained, and the model weight parameters are fixed to obtain a trained deep learning model that can be used to obtain the target predicted purity value. The target predicted purity value represents the theoretical product purity upper limit determined based on transient operating conditions, excluding catalyst aging damage.

[0069] To obtain the predicted purity value at the current moment, the real-time operating parameters, including reaction temperature, reaction pressure, and mass hourly space velocity, corresponding to that moment can be input into the deep learning model. The target predicted purity value is obtained by feedforward calculation using its weight matrix and bias term. The physical correction index at the current moment is obtained to characterize the degree of performance degradation that occurs during catalyst use. A purity deviation term is constructed based on the product of the target predicted purity value and the physical correction index. The predicted purity value at the current moment is calculated by subtracting the purity deviation term from the target predicted purity value.

[0070] Based on the above steps, the predicted purity value of the chemical product at each time point can be obtained.

[0071] Furthermore, due to microscopic differences in support porosity and metal dispersion, the decay rate of different batches of catalysts is not constant, leading to cumulative prediction drift over long-term operation. Therefore, to avoid cumulative drift caused by individual differences in different batches of catalysts, this invention introduces an adaptive feedback correction mechanism based on the prediction results. The residual term between the predicted purity value and the measured purity value is used as a feedback error signal, and a convergence gain factor is introduced for proportional adjustment, constructing a dynamic compensation closed loop. When the predicted value is higher than the measured value, the residual is positive, indicating that the model underestimates the degree of deactivation. Therefore, the decay constant is increased through positive compensation, and proportional adjustment is performed through the convergence gain factor. According to catalyst deactivation theory, as the number of regenerations increases, the catalyst's microstructure gradually transitions from the initial period of drastic change to the later steady-state degradation period. Therefore, the initial activity decay constant is reduced inversely based on the total number of regenerations. When the total number of regenerations is large in the later stages of operation, reducing the correction intensity can give higher confidence to historical cumulative characteristics, effectively avoiding random noise interference from chromatograph measurements.

[0072] Specifically, this includes: obtaining the measured purity value at the production site, using the difference between the predicted purity value and the measured purity value as feedback error; and dynamically adjusting the initial activity decay constant based on the feedback error and the current total number of regenerations.

[0073] In this embodiment of the invention, the relationship between the feedback error and the current total number of regenerations for dynamically adjusting the initial activity decay constant is as follows: ; in, For use in time updates The initial activity decay constant; For a moment The initial activity decay constant; For a moment The predicted purity value. For a moment The measured purity value obtained by the chromatograph is used to provide real-world benchmark data for on-site comparison. This is the convergence gain factor, used to balance sensitivity and stability. For a moment The total number of regenerations.

[0074] In obtaining the convergence gain factor, the prediction residuals of historical data can be fitted using the least squares method. This ensures that the obtained convergence gain factor can correct the initial activity decay constant at a reasonable rate, preventing system oscillation due to excessive correction or prediction drift due to excessively slow correction. In this embodiment of the invention, an empirical range of 0.01 to 0.05 is selected.

[0075] In the above relation, the formula uses fractional terms. Characterizes the decay of physical confidence. Due to prediction error. This represents the predicted spillover. A larger difference indicates an overly optimistic prediction. The formula adjusts the parameter through a positive increment to force a more severe catalyst aging process. Therefore, the larger the difference, the larger the compensation increment due to the positive mapping, resulting in the final updated data value. The value also increases accordingly, showing a positive correlation. At the same time, the more times the total number of regenerations increases, the more stable and inactive the microstructure becomes. Therefore, when the total number of regenerations increases, the correction step size should be reduced accordingly to resist random noise.

[0076] After obtaining the initial activity decay constant based on the above steps, the initial activity decay constant can be limited to the physically effective range based on the limiting function, that is, the value of the initial activity decay constant is kept within the physically effective range. The specific settings can be made according to actual needs, and this embodiment of the invention does not impose too many restrictions here.

[0077] Understandably, updating the total number of regenerations is an automatic counting process triggered by physical events, responding to asynchronous calculations triggered by in-situ regeneration pulse signals. If the system enters in-situ regeneration mode, the logic controller recognizes a pulse signal indicating the start of a regeneration cycle. Each time a complete coking and reduction activation process is completed, the total number of regenerations automatically increments by 1. The updated total number of regenerations is then used as the new total number of regenerations for the calculations described above. Therefore, from the time the catalyst is fresh until the first regeneration begins, a reference constant determined based on accelerated aging tests of a reference catalyst is used by default. The update formula is activated for closed-loop correction only when the system recognizes the end of the first regeneration process and the regeneration counter is updated to 1.

[0078] Furthermore, the updated initial activity decay constant is only used for calculating the lifetime damage factor and physical correction index at subsequent time steps, without changing the weight matrix and bias terms of the deep learning model, so as to reduce the waste of computing power caused by data training.

[0079] For example, in response to the predicted purity value decreasing to a preset activity threshold, the system automatically switches to in-situ regeneration mode to burn off carbon deposits. After each in-situ regeneration program is triggered, the logic controller automatically sends a pulse counting signal to the distributed control system to incrementally update the total number of regenerations, and uses the updated total number of regenerations in calculations at subsequent times.

[0080] The activity threshold can be set according to process requirements. If the downstream process has strict requirements on the purity of furan, the threshold can be set to the lowest value that can ensure the normal progress of subsequent reactions. Optionally, in the embodiments of the present invention, the activity threshold can be set to 95%.

[0081] When the predicted residual cannot be recovered even through correction, or when the number of regeneration cycles reaches its limit, the system will prompt for a complete catalyst replacement.

[0082] Thus, through the collaborative calculation of neural network feature fusion and adaptive feedback correction, the embodiments of the present invention can effectively eliminate batch differences and resist interference in the late stage of operation, improve the steady-state prediction accuracy of industrial control sites, and thus provide an accurate data foundation for realizing catalyst life early warning and precise catalyst replacement.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the purity of fine chemical products based on deep learning, characterized in that, include: Step 1: Obtain the operating parameters of the catalyst to be tested. The operating parameters include real-time operating parameters and historical performance data at each moment. The historical performance data includes the current cumulative running time, total number of regenerations, rated initial maximum lifetime value, platinum loading, palladium loading, and aluminum-titanium molar ratio. The real-time operating parameters include reaction temperature, reaction pressure, and mass hourly space velocity. Step 2: Calculate the lifetime damage factor at each time point based on historical performance data. The lifetime damage factor is positively correlated with the current cumulative running time and total number of regenerations at that time point, and negatively correlated with the rated initial maximum lifetime value. Step 3: Calculate the physical correction index at a given time point. The physical correction index is positively correlated with the lifetime damage factor and negatively correlated with platinum loading, palladium loading, and aluminum-titanium molar ratio. Step 4: Input the real-time operating condition parameters into the pre-trained deep learning model for feedforward calculation to obtain the target prediction purity value at the corresponding time. Based on the target predicted purity value and the physical correction index, the predicted purity value at the corresponding time is calculated.

2. The method for predicting the purity of fine chemical products based on deep learning according to claim 1, characterized in that, The acquisition of the operating parameters of the catalyst under test includes: The distributed control system extracts the cumulative running time of the current reaction and the total number of regenerations corresponding to the current catalyst at each sampling moment; it obtains the rated initial maximum lifetime value, the platinum loading and palladium loading in the catalyst, and the aluminum-titanium molar ratio in the support based on the catalyst loading material list; and it collects the reaction temperature, reaction pressure, and mass hourly space velocity in real time through field instruments.

3. The method for predicting the purity of fine chemical products based on deep learning according to claim 1, characterized in that, The calculation of the lifetime impairment factor at each time point based on historical performance data includes: Obtain the deviation of the current cumulative running time from the rated initial maximum lifetime value at a given moment; obtain the regeneration penalty multiplication factor, which is used to characterize the irreversible grain agglomeration penalty caused by thermal shock; multiply the deviation by the regeneration penalty multiplication factor as the lifetime damage factor at the current moment; wherein the regeneration penalty multiplication factor is obtained by using the natural constant as the base and the product of the initial activity decay constant and the current total number of regenerations as the exponent for power operation to obtain the regeneration penalty multiplication factor.

4. The method for predicting the purity of fine chemical products based on deep learning according to claim 3, characterized in that, The method for obtaining the initial activity decay constant includes: Accelerated aging tests were conducted using a reference catalyst with the same noble metal loading as the catalyst under test, and the residual active site rate of the reference catalyst was measured at different regeneration cycles. Using the regeneration cycle as the independent variable and the logarithm of the residual active site rate as the dependent variable, linear regression was performed using the least squares method, and the absolute value of the slope of the regression line was extracted as the attenuation baseline value. Based on multiple sets of attenuation baseline values ​​of the reference catalyst at different reaction temperatures, a physical effective range of the activity attenuation constant was constructed, and its upper and lower limits were determined. The deviation between the initial aluminum-titanium molar ratio of the current catalyst under test and the standard molar ratio of the reference catalyst was calculated, and linear interpolation was performed within the physical effective range based on the deviation to obtain the initial activity attenuation constant.

5. The method for predicting the purity of fine chemical products based on deep learning according to claim 1, characterized in that, The fine chemical product is furan prepared by continuous gas-phase decarbonylation of furfural; the catalyst is a supported platinum-palladium bimetallic catalyst.

6. The method for predicting the purity of fine chemical products based on deep learning according to claim 1, characterized in that, The calculation of the physical correction index at a given time includes: ； in, For a moment The physical correction index, To preset the product degradation coefficient, For a moment Lifespan damage factors, For empirical coupling factors, , , These correspond to the platinum loading, palladium loading, and aluminum-titanium molar ratio in the catalyst, respectively.

7. The method for predicting the purity of fine chemical products based on deep learning according to claim 1, characterized in that, The calculation of the predicted purity value at the corresponding time based on the target predicted purity value and the physical correction index includes: A purity deviation term is constructed based on the product of the target predicted purity value and the physical correction index; the predicted purity value at the current time is obtained by subtracting the purity deviation term from the target predicted purity value.

8. The method for predicting the purity of fine chemical products based on deep learning according to claim 3, characterized in that, The method for obtaining the initial activity decay constant also includes: The measured purity value at the production site is obtained, and the difference between the predicted purity value and the measured purity value is used as the feedback error. The initial activity decay constant is dynamically adjusted based on the feedback error and the current total number of regenerations.

9. The method for predicting the purity of fine chemical products based on deep learning according to claim 8, characterized in that, The dynamic adjustment of the initial activity decay constant based on feedback error and the current total number of regenerations includes: ； in, For use in time updates The initial activity decay constant, For a moment The initial activity decay constant, The convergence gain factor, For a moment The predicted purity value, For a moment The measured purity value obtained by the chromatograph. For a moment The total number of regenerations.

10. The method for predicting the purity of fine chemical products based on deep learning according to claim 1, characterized in that, The calculation yields the predicted purity value at the corresponding time point, and then includes: When the predicted purity value drops to the preset activity threshold, the system automatically switches to in-situ regeneration mode to burn off carbon deposits. After each in-situ regeneration program is triggered, the logic controller automatically sends a pulse counting signal to the distributed control system to update the total number of regenerations, and uses the updated total number of regenerations in subsequent calculations.

Images