Hydrogen purification breakthrough trend intelligent prediction and early warning method based on time series model

Abstract

The application discloses a hydrogen purification breakthrough trend intelligent prediction and early warning method based on a time sequence model and specifically relates to the technical field of prediction and alarm, which comprises the following steps: collecting impurity concentration at the outlet of an adsorption tower, operating condition variables and cumulative running time to form multi-source time sequence data; obtaining an input sequence containing target variable historical values, operating condition historical values and life parameters through preprocessing; generating a baseline prediction sequence through continuous time state evolution and generating a residual prediction sequence through time sequence feature learning; obtaining two hidden states, dynamically determining contribution degree configuration in combination with process context information and synthesizing a breakthrough index prediction result; determining a prediction inflection point through trend analysis and generating a graded early warning signal in combination with a safety judgment benchmark. Through double-channel decoupling and dynamic contribution degree configuration, the application realizes advanced prediction and graded early warning of the breakthrough inflection point and improves accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of predictive alarm technology, and more specifically, to a method for intelligent prediction and early warning of hydrogen purification penetration trend based on a time-series model. Background Technology

[0002] In the hydrogen pressure swing adsorption (PSA) purification process, the impurity concentration at the adsorption tower outlet changes over time, forming a breakthrough curve. When the impurity concentration approaches the safety upper limit, the adsorption tower must be switched in a timely manner to prevent the product hydrogen purity from exceeding the standard. Predicting the inflection point of the breakthrough curve in advance and issuing an early warning is crucial to ensuring the safe operation of the hydrogen purification process. Existing predictive alarm methods mainly employ single data-driven models, such as time-series predictive networks based on LSTM or Transformer. These methods take historical impurity concentration and operating condition data as input, directly outputting predicted values ​​of future breakthrough indicators and generating alarm signals accordingly.

[0003] However, in practical use, it still has some shortcomings, such as encoding the physical trend of the penetration process mixed with the fluctuations caused by operating conditions, the model is highly dependent on data quality, and the prediction accuracy drops significantly when the adsorbent performance deteriorates or the operating conditions drift; it fails to dynamically adjust the prediction strategy according to the current penetration stage, and the sparse training samples near the inflection point can easily lead to prediction lag, resulting in untimely warnings; it lacks adaptability to the changes in the state of the adsorbent throughout its entire life cycle, and it is difficult to maintain stable warning performance at different aging stages of the adsorbent. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a method for intelligent prediction and early warning of hydrogen purification penetration trend based on a time-series model, which addresses the problems mentioned in the background art through the following scheme.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent prediction and early warning of hydrogen purification penetration trend based on a time-series model, comprising: During the hydrogen purification process, the concentration of impurities at the outlet of the adsorption tower, operating condition variables, and the cumulative running time of the adsorbent are collected to form multi-source time-series monitoring data. The multi-source time-series monitoring data is preprocessed to form an input sequence for prediction. The input sequence includes historical values ​​of the target variable, historical values ​​of the operating condition, and adsorbent lifetime parameters. Based on the historical values ​​of the target variable and the adsorbent lifetime parameters, a baseline prediction sequence characterizing the penetration trend is generated through continuous time state evolution processing; based on the historical values ​​of the operating conditions, a residual prediction sequence characterizing the data residuals is generated through time-series feature learning processing. The current first hidden state corresponding to the continuous time state evolution processing and the current second hidden state corresponding to the temporal feature learning processing are obtained. Combined with the process context information representing the current penetration stage, the contribution configuration of the baseline prediction sequence and the residual prediction sequence is dynamically determined. According to the contribution configuration, the baseline prediction sequence and the residual prediction sequence are synthesized to obtain the penetration index prediction result for the future preset step size; Trend analysis is performed on the penetration index prediction results to extract the evolution trend characteristics of the penetration index over time. The prediction inflection point time is determined based on the evolution trend characteristics, and a graded early warning signal is generated in combination with the safety judgment benchmark.

[0006] The technical effects and advantages of this invention are as follows: 1. This invention achieves a decoupled representation of physical trends and data residuals by assigning the historical values ​​of target variables and adsorbent lifetime parameters in the input sequence to continuous-time state evolution processing and assigning the historical values ​​of operating conditions to time-series feature learning processing. This avoids the problem of physical laws being masked by noise in a single black-box model and enables the baseline prediction sequence to stably reflect the monotonic evolution trend of the penetration process.

[0007] 2. This invention introduces process context information representing the current penetration stage and combines it with the dynamic determination of contribution configuration of the dual-channel hidden state, enabling sequence synthesis to adaptively adjust with the penetration stage. When far from the inflection point, it makes full use of the data-driven short-term fluctuation characterization capability, and automatically increases the contribution of the physical trend channel when the inflection point is approaching and at the end of the adsorbent's lifespan. This effectively avoids the prediction lag and prediction jump problems caused by sample sparsity near the inflection point, and improves the timeliness and accuracy of the early warning.

[0008] 3. This invention embeds a parameterized function based on adsorption-through kinetics into the continuous-time state evolution processing and uses the adsorbent lifetime parameter as a modulation condition, so that the prediction results can adaptively track the changes in the breakthrough curve morphology throughout the entire life cycle of the adsorbent. It can maintain stable early warning performance even after the adsorbent performance deteriorates and is replaced, without the need for manual model retraining. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0010] Figure 2 This is a schematic diagram of the data acquisition and preprocessing structure of the present invention.

[0011] Figure 3 This is a schematic diagram of the generation structure of the baseline prediction sequence and the residual prediction sequence of the present invention.

[0012] Figure 4This is a schematic diagram of the contribution configuration and sequence synthesis structure of the present invention.

[0013] Figure 5 This is a schematic diagram of the trend analysis and hierarchical early warning structure of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] refer to Figures 1-5 The intelligent prediction and early warning method for hydrogen purification penetration trend based on time-series models shown includes: The concentration of impurities at the outlet of the adsorption tower, operating condition variables, and cumulative operating time of the adsorbent are collected during the hydrogen purification process to form multi-source time-series monitoring data. The multi-source time-series monitoring data is preprocessed to form an input sequence for prediction. The input sequence includes historical values ​​of the target variable, historical values ​​of the operating condition, and adsorbent lifetime parameters.

[0016] The concentration of impurities at the adsorption tower outlet, used as the target variable for predictive alarms, is continuously measured by online analytical instruments installed on the adsorption tower outlet pipeline and recorded in a time series format, forming a target variable time series. Operating condition variables refer to external process parameters affecting the breakthrough rate and morphology, including at least inlet flow rate, adsorption pressure, and bed temperature. These variables are synchronously collected by corresponding flow meters, pressure transmitters, and temperature sensors, with each variable aligned to the same timestamp, forming a synchronously recorded multidimensional time series. The cumulative adsorbent operating time refers to the total working time experienced by the adsorbent since it was put into operation, obtained from the unit's operating log or the cumulative timing module in the distributed control system, with the current cumulative value read synchronously at each data acquisition moment. These three types of data together constitute multi-source time-series monitoring data.

[0017] The multi-source time-series monitoring data is preprocessed to form an input sequence for prediction. Preprocessing includes: outlier detection and removal for target and operating condition variables; data points exceeding a preset physical range are marked and removed; missing values ​​due to removal or instrument malfunction are filled using interpolation; and standardization is performed on all variables, converting each variable value into a standard score on a uniform scale. After standardization, the input sequence is constructed using a sliding window approach. A fixed-length historical window is set. For the current moment, the standardized data from each time step within the window constitute the input sequence, where the data vector for each time step includes the standardized target variable value, operating condition variable values, and adsorbent lifetime parameter values ​​for that moment. The length of the historical window is typically set to 120 time steps, which can be adjusted according to the switching cycle and sampling frequency of the adsorption tower.

[0018] The input sequence includes historical values ​​of the target variable, historical values ​​of operating conditions, and adsorbent lifetime parameters. The historical values ​​of the target variable are a standardized sequence of impurity concentrations at the adsorption tower outlet within a preset time step, reflecting the historical evolution of the breakthrough process. The historical values ​​of the operating conditions are standardized sequences of operating condition variables such as inlet flow rate, adsorption pressure, and bed temperature within a preset time step, reflecting the historical variation patterns of operating conditions. The adsorbent lifetime parameters are the current adsorbent lifetime state values ​​after standardization, obtained by converting and standardizing the cumulative operating time of the adsorbent, reflecting the aging stage of the adsorbent.

[0019] Furthermore, the preprocessing of multi-source time-series monitoring data includes outlier detection and removal, missing value imputation, and standardization, among which: Outlier detection is based on the preset physical range of each variable. The impurity concentration should not be negative and should not exceed the feed concentration. The adsorption pressure should be within the design operating pressure range of the device. Values ​​that exceed the range are marked as outliers and removed. The removed vacancies are processed together in the missing value filling step.

[0020] Missing value imputation addresses data gaps caused by instrument malfunctions, communication interruptions, etc. For short-term missing values, linear interpolation or the mean of the effective values ​​before and after the missing values ​​is used for imputation. For missing values ​​over a longer period, regression is used to imput the missing values ​​based on the changing trend of related variables during the same period.

[0021] The standardization process employs the Z-score standardization method, which calculates the mean and standard deviation of the historical data for each variable, converting each variable value into a standard score with a mean of zero and a standard deviation of one, in order to eliminate differences in units and numerical ranges between different variables.

[0022] Furthermore, the adsorbent lifetime parameter in the input sequence is generated as follows: The cumulative operating time of the adsorbent is converted into the proportion of adsorbent used, i.e., the ratio of the current cumulative operating time to the designed total lifetime. This ratio is then standardized to obtain a dimensionless adsorbent lifetime parameter. This parameter serves as a quantitative characterization of the adsorbent's aging state and is used to modulate kinetic parameters in subsequent continuous-time state evolution processing, enabling baseline predictions to reflect the impact of adsorbent performance degradation on the breakthrough trend.

[0023] Based on the historical values ​​of the target variable and the adsorbent lifetime parameters, a baseline prediction sequence characterizing the penetration trend is generated through continuous-time state evolution processing; based on the historical values ​​of the operating conditions, a residual prediction sequence characterizing the data residuals is generated through time-series feature learning processing.

[0024] The two processing steps described above are executed in parallel, processing data components with different properties in the input sequence respectively. The historical values ​​of the target variable and the adsorbent lifetime parameter together reflect the intrinsic evolutionary law of the penetration process and the long-term impact of adsorbent aging on the penetration trend. Therefore, they are input into continuous-time state evolution processing to generate a baseline prediction sequence that conforms to the physical monotonic trend. The historical values ​​of operating conditions reflect the short-term perturbations of the penetration process caused by changes in external operating conditions. Therefore, they are input into time-series feature learning processing to generate a residual prediction sequence that characterizes the data residuals. The baseline prediction sequence and the residual prediction sequence are generated independently, representing the penetration process from the two dimensions of physical trend and data residuals, respectively.

[0025] Furthermore, the continuous-time state evolution process includes hidden state initialization, continuous-time evolution solving, and baseline decoding projection, wherein: Hidden state initialization involves encoding the final time-state of the target variable's historical values ​​in the input sequence into an initial hidden state vector. Specifically, the sequence of historical target variable values ​​is mapped to an initial hidden state through an encoding network. This initial hidden state contains an abstract representation of the current penetration stage and the degree of adsorbent aging.

[0026] The continuous-time evolution solution involves inputting the initial hidden state and adsorbent lifetime parameters into a parameterized continuous-time evolution function, and then using a differential equation solver to calculate the hidden state trajectory at each future time step.

[0027] Specifically, set Let be the hidden state vector at time t. Let the adsorbent lifetime parameter be used, then the parameterized continuous-time evolution function is defined as:

[0028] in It is a trainable neural network.

[0029] Given an initial hidden state The future hidden state is solved by: The continuous-time evolution function models the right-hand side of the adsorption-throughput kinetics equation as a trainable parameterized function. The adsorbent lifetime parameter serves as a conditional input to modulate the evolution function, enabling the evolution trajectory to adaptively reflect the differences in penetration rates at different aging stages. The differential equation solver can employ an adaptive step-size solver, generating a sequence of hidden states corresponding to each time point based on a preset future time step.

[0030] Baseline decoding projection maps the hidden state trajectory obtained from the evolution solution to a baseline prediction sequence through a decoding network. This baseline prediction sequence characterizes the monotonic evolution trend of the penetration index over time, considering only the adsorption penetration physical process and adsorbent aging factors. Its shape is a smooth S-shaped growth curve, and it does not include short-term fluctuations caused by operating condition variations.

[0031] Furthermore, the parameterized continuous-time evolution function in the continuous-time state evolution processing is constructed based on the linear driving force model of fixed-bed adsorption, where: A linear driving force model describes the relationship between the rate of change of impurity concentration at the adsorbent bed outlet and the current concentration and saturated adsorption capacity; that is, the rate of change is proportional to the difference between the current adsorption capacity and the equilibrium adsorption capacity. Replacing the right-hand side of this kinetic equation with a small, fully connected network, which accepts the current hidden state and adsorbent lifetime parameters as input and outputs the rate of change of the hidden state, specifically, this small, fully connected network... The expression is:

[0032] in This represents vector concatenation. and For trainable parameters, The activation function is used, and the network output is the rate of change of the hidden states. The adsorbent lifetime parameter is input into the network as a conditional variable, enabling the network to learn the slow modulation effect of adsorbent aging on kinetic parameters—as the cumulative operating time of the adsorbent increases, the kinetic parameters are adjusted accordingly, so that the inflection point position and steepness of the generated baseline prediction sequence can adaptively reflect the changes in the breakthrough curve shape caused by the degradation of adsorbent performance.

[0033] Through the above continuous-time state evolution processing, a baseline prediction sequence is obtained that is constrained by the physical laws of adsorption penetration and depends only on time evolution and aging effects. This sequence has a monotonically increasing trend with a gradually increasing slope, providing a stable trend reference for subsequent contribution configuration and inflection point detection.

[0034] Furthermore, temporal feature learning processing includes temporal coding and residual decoding, wherein: Temporal coding extracts features from historical operating conditions in the input sequence using a temporal coding network. This network employs a temporal attention mechanism to model the long-range dependencies between time steps in the historical operating condition value sequence, capturing the lag effect patterns of operating condition changes on the penetration process. After encoding, a second hidden state is generated for the current moment, which contains an abstract representation of the operating condition fluctuation patterns and their impact on the penetration process.

[0035] Residual decoding maps the second hidden state to a residual prediction sequence through a decoding network. This residual prediction sequence characterizes the amount by which operating condition fluctuations correct for the penetration index—that is, the direction and magnitude of the penetration index's deviation from the baseline trend under actual operating conditions. The values ​​of the residual prediction sequence can be positive or negative; positive values ​​indicate that operating condition factors accelerate penetration, while negative values ​​indicate that operating condition factors delay penetration. This sequence is then synthesized with the baseline prediction sequence in subsequent steps through contribution configuration to reflect the complete evolution of the actual penetration process.

[0036] The current first hidden state corresponding to the continuous time state evolution processing and the current second hidden state corresponding to the temporal feature learning processing are obtained. Combined with the process context information representing the current penetration stage, the contribution configuration of the baseline prediction sequence and the residual prediction sequence is dynamically determined.

[0037] The continuous-time state evolution processing generates the current first hidden state during execution. This hidden state is an abstract encoding of the physical trend of the penetration process, and its numerical characteristics implicitly contain information about the current evolutionary stage of the penetration. The temporal feature learning processing generates the current second hidden state during execution. This hidden state is an abstract encoding of the operating condition fluctuation pattern, and its numerical characteristics implicitly contain information about the degree of influence of the current operating condition disturbance on the penetration process. These two hidden states provide the basis for judging the reliability of their respective predictions from the two dimensions of physical trend and data residuals.

[0038] The first hidden state, the second hidden state, and process context information representing the current penetration stage are input into a gating network. The gating network calculates and outputs the contribution coefficients of the baseline prediction sequence and the residual prediction sequence. The determined result of the contribution configuration is a set of normalized contribution coefficients, where the sum of the contribution coefficients of the baseline prediction sequence and the residual prediction sequence is one. This configuration result is not fixed, but dynamically adjusted with each sliding window update.

[0039] Furthermore, the process context information characterizing the current penetration stage includes the relationship between the current penetration index value and the safety judgment benchmark, as well as the adsorbent lifetime parameter. The relationship between the current penetration index value and the safety judgment benchmark reflects the urgency of the current penetration process, i.e., how close the current concentration value is to the safety judgment benchmark. The closer the distance, the higher the penetration risk, and the more necessary it is to rely on stable extrapolation of physical trends for early judgment. The adsorbent lifetime parameter reflects the aging stage of the adsorbent. Towards the end of the adsorbent's lifetime, the penetration curve shape undergoes a systematic change, and the baseline trend's ability to predict inflection points is relatively more reliable. The gating network comprehensively utilizes the above process context information, as well as the first and second hidden states, to evaluate the credibility of the two prediction channels at the current moment and dynamically allocate contribution coefficients accordingly.

[0040] Furthermore, the dynamic determination of contribution configuration includes channel confidence assessment and contribution coefficient generation, wherein: Channel confidence assessment is a quantitative evaluation of the reliability of the baseline prediction channel and the residual prediction channel at the current moment. For the baseline prediction channel, the local curvature information of the first hidden state is used to assess trend stability. When the curvature change of the hidden state trajectory is stable and conforms to the monotonically accelerating characteristics of the S-shaped curve of the penetration process, the confidence of the baseline prediction channel is high; conversely, when the hidden state trajectory exhibits abnormal fluctuations inconsistent with physical laws, its confidence decreases. The assessment of trend stability using the local curvature information of the first hidden state can be specifically achieved by calculating the second-order difference of the hidden state trajectory over continuous time steps. The smaller the second-order difference, the more stable the curvature change and the higher the trend stability. For the residual prediction channel, the prediction variance corresponding to the second hidden state is used to assess uncertainty. The smaller the prediction variance, the higher the reliability of the data-driven residual prediction under the current operating condition; conversely, when encountering operating conditions not fully covered in the training data, the prediction variance increases, and the confidence of the residual prediction channel decreases accordingly.

[0041] The contribution coefficient generation involves processing the first hidden state, the second hidden state, the channel confidence assessment results, and the process context information through a lightweight gating network. The gating network outputs two values, which are then normalized to obtain the contribution coefficients for the baseline prediction sequence and the residual prediction sequence, respectively. The normalization process ensures that the sum of the two contribution coefficients is one, keeping the synthesized prediction result within a reasonable numerical range.

[0042] Furthermore, the dynamic adjustment of contribution allocation follows these patterns: In the early stages of the breakthrough process, the current breakthrough index value in the process context information is far from the safety judgment benchmark, and the impurity concentration at the adsorption tower outlet is in a slow rising period. At this time, the historical data of the operating conditions is relatively abundant, the residual prediction channel has a strong ability to characterize short-term fluctuations, and the gating network tends to assign a higher contribution coefficient to the residual prediction sequence, so that the synthetic prediction results can more sensitively reflect changes in the operating conditions.

[0043] As the penetration process approaches the inflection point, the distance between the current penetration index value in the process context information and the safety judgment benchmark decreases, and the penetration curve is about to enter a rapid upward phase. At this time, data-driven residual prediction faces increased uncertainty due to the sparse sample size near the inflection point. Meanwhile, the baseline prediction channel, with its monotonic trend extrapolation capability constrained by physical laws, provides a more stable judgment on the inflection point location and subsequent evolution direction. After sensing these changes, the gating network automatically increases the contribution coefficient of the baseline prediction sequence, making the synthesized prediction result more reliant on the stable extrapolation of physical trends, thereby avoiding prediction lag or jumps near the inflection point.

[0044] When the adsorbent reaches the end of its lifespan, the adsorbent lifetime parameter in the process context information indicates a deep aging process, and the inflection point and steepness of the breakthrough curve undergo a systematic shift. Under these conditions, the gating network further enhances the contribution of the baseline prediction channel by modulating the adsorbent lifetime parameter, enabling the prediction results to adaptively track changes in the breakthrough trend caused by adsorbent performance degradation. When the adsorbent is replaced with a brand new batch, the adsorbent lifetime parameter is reset, and the continuous-time state evolution function, modulated by the lifetime parameter, will automatically adjust its kinetic parameters to adapt the generated baseline prediction sequence to the breakthrough curve morphology of the new adsorbent in the early stages of activation.

[0045] According to the contribution configuration, the baseline prediction sequence and the residual prediction sequence are synthesized to obtain the penetration index prediction result for the future preset step size.

[0046] The preset future step size can typically be set to 60 time steps to provide sufficient early warning lead time, which can be determined according to the early warning response time required by the process. Both the baseline prediction sequence and the residual prediction sequence are prediction data for a predetermined future time step, and they correspond one-to-one in the time dimension—that is, at the same future time step, the baseline prediction sequence provides a baseline prediction value, and the residual prediction sequence provides a residual correction value. Sequence synthesis integrates the two prediction sequences according to the contribution coefficients determined by the contribution configuration to obtain the penetration index prediction values ​​for each future time step.

[0047] Specifically, for each future time step, the predicted value of that time step in the baseline prediction sequence and the predicted value of that time step in the residual prediction sequence are multiplied by their respective contribution coefficients and then added together to obtain the composite predicted value for that time step. Arranging the composite predicted values ​​of each time step in chronological order forms the penetration index prediction result. This penetration index prediction result is a prediction sequence covering a preset future step length, integrating stable extrapolation of physical trends and short-term correction of data residuals, and can relatively completely reflect the expected evolution of the penetration process over a future period.

[0048] The penetration index prediction results after sequence synthesis differ from both the baseline prediction sequence containing only smoothed trends and the residual prediction sequence containing only fluctuation corrections. Instead, they represent an organic integration of both under a dynamic contribution configuration. Because the contribution configuration is dynamically adjusted with each sliding window, the prediction results after sequence synthesis can adaptively balance the long-term stability of physical trends and the short-term sensitivity of data-driven approaches, depending on the current penetration stage.

[0049] Furthermore, the contribution configuration during sequence synthesis functions as follows: During the stable operation phase of the penetration process, which is far from the inflection point, the contribution coefficient of the residual prediction sequence in the contribution configuration is relatively high. The prediction results after sequence synthesis can more sensitively reflect the short-term impact of operating condition fluctuations on the penetration index, so that the prediction sequence closely follows the actual operating condition changes.

[0050] As the penetration process approaches its inflection point or nears the end of the adsorbent's lifespan, the contribution coefficient of the baseline prediction sequence in the contribution configuration is automatically increased. The prediction results after sequence synthesis rely more on the extrapolation of monotonic trends constrained by physical laws. The synthesized sequence exhibits smoother trend characteristics, providing a stable prediction basis for early prediction of the inflection point.

[0051] Through the sequence synthesis processing under the above dynamic contribution configuration, the penetration index prediction results achieve optimal integration of physical trends and data fluctuations at different stages of the penetration process, providing a high-quality predictive data foundation for subsequent trend analysis and early warning generation.

[0052] Trend analysis is performed on the penetration index prediction results to extract the evolution trend characteristics of the penetration index over time. The prediction inflection point time is determined based on the evolution trend characteristics, and a graded early warning signal is generated in combination with the safety judgment benchmark.

[0053] The penetration index prediction result is a time series containing the penetration index values ​​at each time point within a predetermined future step. Trend analysis of this time series involves extracting quantitative features that characterize the evolution of the penetration process, rather than simply relying on single numerical predictions. These evolutionary trend features include at least the direction, rate, and acceleration of the penetration index value's change. The direction of change indicates whether the penetration index is rising or falling; the rate of change reflects the speed of the penetration process; and the acceleration reveals whether the penetration rate itself is accelerating or slowing down.

[0054] The determination of the predicted inflection point is based on the changing acceleration in the evolutionary trend characteristics. As the penetration curve approaches the inflection point, its changing acceleration exhibits a transition from positive acceleration to decreasing acceleration. At the zero-crossing position where the acceleration sign changes from positive to negative, the slope of the penetration curve reaches its maximum value; this position is the inflection point of the penetration curve. By calculating and determining the sign of the changing acceleration at each time step in the penetration index prediction results, the time position where the acceleration sign changes is determined as the predicted inflection point.

[0055] The generation of tiered early warning signals is based on the predicted inflection point time and the predicted penetration index results, combined with a comprehensive judgment based on safety criteria. The safety criteria are the upper limit of the allowable impurity concentration set according to process requirements. When the predicted penetration index result exceeds this criterion value at a future time step, it indicates that the purity of the product hydrogen will face the risk of exceeding the standard. The tiered early warning signals generate different levels of warning signals based on the time distance between the current time and the predicted inflection point time, the closeness of the predicted penetration index result to the safety criteria, and the continuous trend of the rate of change.

[0056] Furthermore, the extraction of the evolutionary trend characteristics of the penetration indicator over time includes: The direction of change is determined by judging the sign of the difference between adjacent time steps of the penetration index prediction sequence. If the difference is consistently positive, it indicates that the penetration index is in an upward trend. The rate of change is obtained by calculating the first difference of the penetration index prediction sequence. The magnitude of the first difference reflects the magnitude of the change of the penetration index per unit time. The acceleration of change is obtained by calculating the second difference of the penetration index prediction sequence. The magnitude of the second difference reflects the increase or decrease of the rate of change itself.

[0057] The extraction of the aforementioned evolutionary trend features can be performed on the baseline prediction sequence. Since the baseline prediction sequence is constrained by physical laws, its shape is smooth and it has good monotonicity. The trend features calculated on it are more stable and can avoid interference from residual fluctuations in the judgment of inflection points.

[0058] Furthermore, the method for determining the inflection point is as follows: During the rising phase of the penetration curve, the acceleration gradually decreases from a positive value. When the acceleration crosses from a positive value to zero or a negative value, the corresponding time position is the inflection point of the penetration curve. To eliminate the interference of high-frequency noise introduced by numerical calculation on the inflection point judgment, the calculated second-order difference sequence can first be processed by moving average filtering. Then, by traversing the acceleration values ​​at each time step within the future preset step size, the position of the change in the sign of acceleration can be detected, and the time step in which the sign first changes can be determined as the predicted inflection point.

[0059] If no change in the sign of acceleration is detected within the preset future step size, it indicates that the penetration curve has not yet reached the inflection point within the current prediction range. At this time, based on the decreasing trend of the changing acceleration and the level of the current rate of change, the approximate range of the inflection point can be estimated by linear extrapolation or template matching based on the shape of the historical penetration curve, and marked as a state to be confirmed.

[0060] Furthermore, the generation rules for tiered early warning signals are as follows: The first-level warning signal corresponds to the attention level. It is triggered when the penetration indicator prediction result shows a continuously increasing rate of change, and the time distance between the predicted inflection point and the current time is greater than a first preset duration. This indicates to the operator that the penetration process has entered an accelerated phase and that attention should be paid to the development of the penetration trend. The first preset duration is typically set to 30 minutes, but can be specifically set according to the safety margin requirements in the process operation procedures.

[0061] The second-level warning signal corresponds to the preparation level. It is triggered when the time distance between the predicted inflection point and the current time is reduced to within the first preset time period, and the predicted result of the penetration index will reach the preset proportion of the safety judgment benchmark within the future preset time step. This signal reminds the operator that the penetration inflection point is about to arrive and that preparations should be made for switching the adsorption tower or related operations.

[0062] The third-level warning signal corresponds to the switching level. It is triggered when the time distance between the predicted inflection point and the current time narrows to within the second preset time period, or when the predicted result of the penetration index will exceed the safety judgment benchmark within a future preset time step. The second preset time period can typically be set to 10 minutes, and can be set according to the safety margin requirements in the process operation procedure. If the second preset time period is less than the first preset time period, it prompts the operator to take immediate intervention measures to prevent the hydrogen purity of the product from exceeding the standard.

[0063] After the warning signal is generated, the warning level, the predicted inflection point time, the prediction results of the penetration indicators, and the evolution trend characteristics will be output together for operators or related systems to make decisions and respond.

[0064] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, 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 spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent prediction and early warning of hydrogen purification penetration trend based on a time-series model, characterized in that, include: The concentration of impurities at the outlet of the adsorption tower, operating condition variables, and cumulative operating time of the adsorbent during the hydrogen purification process are collected to form multi-source time-series monitoring data. The multi-source time-series monitoring data is preprocessed to form an input sequence for prediction; wherein the input sequence includes historical values ​​of the target variable, historical values ​​of operating conditions, and adsorbent lifetime parameters. Based on the historical values ​​of the target variable and the adsorbent lifetime parameters, a baseline prediction sequence characterizing the penetration trend is generated through continuous-time state evolution processing. Based on the historical values ​​of the operating conditions, a residual prediction sequence representing the data residuals is generated through time-series feature learning processing. The current first hidden state corresponding to the continuous time state evolution processing and the current second hidden state corresponding to the temporal feature learning processing are obtained. Combined with the process context information representing the current penetration stage, the contribution configuration of the baseline prediction sequence and the residual prediction sequence is dynamically determined. According to the contribution configuration, the baseline prediction sequence and the residual prediction sequence are synthesized to obtain the penetration index prediction result for the future preset step size; Trend analysis is performed on the penetration index prediction results to extract the evolution trend characteristics of the penetration index over time. The prediction inflection point time is determined based on the evolution trend characteristics, and a graded early warning signal is generated in combination with the safety judgment benchmark.

2. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The multi-source time-series monitoring data undergoes preprocessing, including: outlier detection and removal, missing value imputation, and standardization, wherein: Outlier detection is based on the preset physical range of each variable. The impurity concentration should not be negative and should not exceed the feed concentration. The adsorption pressure should be within the design operating pressure range of the device. Values ​​that exceed the range are marked as outliers and removed. The removed vacancies are processed together in the missing value filling step. Missing value imputation addresses data gaps caused by instrument malfunctions or communication interruptions. For short-term missing values, linear interpolation or the mean of previous and subsequent valid values ​​is used for imputation. For missing values ​​over a longer period, regression is used to imputation based on the changing trend of related variables during the same period. The standardization process employs the Z-score standardization method, which calculates the mean and standard deviation of the historical data for each variable, converting each variable value into a standard score with a mean of zero and a standard deviation of one, in order to eliminate differences in units and numerical ranges between different variables.

3. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The generation of the adsorbent lifetime parameters includes: The cumulative operating time of the adsorbent is converted into the proportion of adsorbent already used, that is, the ratio of the current cumulative operating time to the designed total lifetime. Then, the ratio is standardized to obtain a dimensionless adsorbent lifetime parameter. This parameter serves as a quantitative characterization of the aging state of the adsorbent and is used to modulate the kinetic parameters in the subsequent continuous-time state evolution processing so that the baseline prediction can reflect the impact of adsorbent performance degradation on the breakthrough trend.

4. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The continuous-time state evolution process includes: hidden state initialization, continuous-time evolution solving, and baseline decoding projection, wherein: Hidden state initialization involves encoding the final state of the target variable's historical values ​​in the input sequence into an initial hidden state vector. Specifically, the target variable's historical value sequence is mapped to an initial hidden state through an encoding network. This initial hidden state contains an abstract representation of the current penetration stage and the degree of adsorbent aging. The continuous-time evolution solution involves inputting the initial hidden state and adsorbent lifetime parameters into a parameterized continuous-time evolution function, and then using a differential equation solver to calculate the hidden state trajectory for each future time step. This continuous-time evolution function models the right-hand side of the adsorption-breakthrough kinetics equation as a trainable parameterized function, with the adsorbent lifetime parameters serving as conditional inputs to modulate the evolution function, enabling the evolution trajectory to adaptively reflect the differences in breakthrough rates at different aging stages. The differential equation solver employs an adaptive step-size solver, generating a sequence of hidden states corresponding to each time point based on a preset future time step. Baseline decoding projection maps the hidden state trajectories obtained from evolutionary solutions to baseline prediction sequences through a decoding network.

5. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The temporal feature learning process includes: temporal encoding and residual decoding, wherein: Temporal coding extracts features from historical operating conditions in the input sequence through a temporal coding network. This temporal coding network can employ a temporal attention mechanism to model the long-range dependencies between time steps in the historical operating condition value sequence, capturing the lag effect pattern of operating condition changes on the penetration process. After encoding, a second hidden state is generated at the current moment, which contains an abstract representation of the operating condition fluctuation pattern and its impact on the penetration process. Residual decoding maps the second hidden state to a residual prediction sequence through a decoding network. This residual prediction sequence represents the amount of correction that operating condition fluctuations make to the penetration index. Positive values ​​indicate that operating condition factors accelerate penetration, while negative values ​​indicate that operating condition factors delay penetration.

6. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The process context information includes: the relationship between the current penetration index value and the safety judgment benchmark, and the adsorbent lifetime parameters; Among them, the relationship between the current penetration index value and the safety judgment benchmark is used to reflect the urgency of the current penetration process, that is, how close the current concentration value is to the safety judgment benchmark. The closer the distance is, the higher the penetration risk. The adsorbent lifetime parameter is used to reflect the aging stage of the adsorbent.

7. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The dynamic determination of the contribution configuration includes: Obtain the current first hidden state generated by the continuous-time state evolution process, evaluate the trend stability using the local curvature information of the first hidden state, and generate baseline channel confidence. The current second hidden state generated by the time-series feature learning process is obtained, and the uncertainty is evaluated using the prediction variance corresponding to the second hidden state to generate the residual channel confidence. The first hidden state, the second hidden state, the baseline channel confidence, the residual channel confidence, and the process context information representing the current penetration stage are input into the gating network. After normalization, the contribution coefficients of the baseline prediction sequence and the residual prediction sequence are output.

8. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The generation of the penetration index prediction results includes: For each future time step, the predicted value of that time step in the baseline prediction sequence and the predicted value of that time step in the residual prediction sequence are multiplied by their respective contribution coefficients and then added together to obtain the composite predicted value of that time step. The composite predicted values ​​of each time step are arranged in chronological order to form the penetration index prediction result.

9. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The determination of the predicted inflection point time includes: During the rising phase of the penetration curve, the acceleration decreases from a positive value. When the acceleration crosses from a positive value to zero or a negative value, the corresponding time position is the inflection point of the penetration curve. To eliminate the interference of high-frequency noise introduced by numerical calculation on the inflection point judgment, the calculated second-order difference sequence is first processed by moving average filtering. Then, by traversing the acceleration values ​​of each time step within the future preset step size, the position of the change in the sign of acceleration is detected, and the time step in which the sign first changes is determined as the predicted inflection point. If no change in the sign of acceleration is detected within the preset future step size, it indicates that the penetration curve has not yet reached the inflection point within the current prediction range. At this time, the range of the inflection point moment can be estimated by linear extrapolation or template matching based on the historical penetration curve shape, based on the decreasing trend of the changing acceleration and the level of the current rate of change, and marked as pending confirmation.

10. The intelligent prediction and early warning method for hydrogen purification penetration trend based on a time-series model according to claim 1, characterized in that, The generation rules for the graded early warning signals include: The first-level warning signal corresponds to the attention level. It is triggered when the penetration indicator prediction results show that the rate of change continues to increase and the time distance between the predicted inflection point and the current time is greater than the first preset duration. This indicates to the operator that the penetration process has entered the acceleration phase and that attention should be paid to the development of the penetration trend. The first preset duration is set to 30 minutes. The second-level warning signal corresponds to the preparation level. It is triggered when the time distance between the predicted inflection point and the current time is reduced to within the first preset time period, and the predicted result of the penetration index reaches the preset proportion of the safety judgment benchmark within the future preset time step. It reminds the operators that the penetration inflection point is about to arrive and that they need to prepare for the switching of the adsorption tower or related operations. The third-level warning signal corresponds to the switching level. It is triggered when the time distance between the predicted inflection point and the current time is reduced to within the second preset time period, or when the prediction result of the penetration indicator will exceed the safety judgment benchmark within a future preset time step. The second preset time period is set to 10 minutes, which is less than the first preset time period. This prompts the operator to take immediate intervention measures.

Images