An asynchronous multi-frequency hot hydrogen temperature soft measurement method for pure hydrogen metallurgy

Abstract

The present application belongs to the technical field of industrial process modeling and optimization, and relates to an asynchronous multi-frequency hot hydrogen temperature soft measurement method for pure hydrogen metallurgy, comprising: constructing a time-labeled multi-source operation data set; respectively constructing asynchronous multi-frequency sampling structure features, process variable statistical state features, front and rear stage heat state transmission features and variable stage time lag features, and combining to obtain an asynchronous multi-frequency time lag coupling enhanced input vector; based on a reliability constraint adaptive composite kernel function, a training sample set and a sample reliability weight, establishing a temperature soft measurement model; based on a real-time collected multi-source operation data set, constructing an asynchronous multi-frequency time lag coupling enhanced real-time input vector, and outputting a hydrogen gas temperature prediction value at the outlet of an electric heater through the temperature soft measurement model. The present application avoids dynamic distortion caused by forced resampling of multi-source variables, and improves the adaptability of the model to asynchronous sampling, time lag response, heat state transmission and working condition drift.

Description

Technical Field

[0001] This invention relates to the field of industrial process modeling and optimization technology, and in particular to an asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy. Background Technology

[0002] Direct reduction of iron in a pure hydrogen shaft furnace is an important technological direction for achieving deep decarburization in the steel industry. Unlike the traditional blast furnace ironmaking process, the direct reduction process in pure hydrogen uses hydrogen as the main reducing medium. The temperature, pressure, and flow rate of hot hydrogen directly affect the reduction reaction rate, hydrogen utilization efficiency, metallization rate, and system energy consumption. To meet the requirements of the pure hydrogen shaft furnace for the inlet hot hydrogen temperature, a hot hydrogen stepped heating process is usually set up before the furnace. The hydrogen is preheated by a heat exchanger, and then further heated by an electric heater to ensure that the hydrogen temperature at the outlet of the electric heater reaches the process setting requirements.

[0003] However, while existing temperature prediction methods, including mechanistic modeling, possess physical significance, they heavily rely on parameters such as heat transfer coefficients and medium properties that are difficult to obtain online. Furthermore, the frequent changes in the boundary conditions of pure hydrogen metallurgy lead to high modeling costs and insufficient adaptability. Data-driven methods, while eliminating the need for explicit mechanistic parameters, typically neglect the dynamic time delays and thermal inertia of hot hydrogen transfer along heat exchangers, connecting pipelines, and electric heaters, making it difficult to characterize the continuous transfer mechanism between preheating, pipeline transmission, and subsequent heat replenishment. Simultaneously, multi-source operating data in the pure hydrogen-hot hydrogen cascade heating process generally suffers from inconsistent sampling frequencies, asynchronous acquisition times, and data refresh delays. Simply aligning to a unified timescale results in the loss of dynamic information for high-frequency variables, false synchronization of low-frequency variables, and distortion of thermal state transfer relationships. Moreover, existing support vector machine and least squares support vector machine methods struggle to adapt to the complex dynamic characteristics and long-term operating condition drift of the pure hydrogen-hot hydrogen cascade heating process, leading to decreased prediction accuracy during operating condition switching, power adjustment, or fluctuations in inlet conditions. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide an asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy, in order to solve the problems of existing technologies such as lack of dynamic transmission mechanism, distortion of asynchronous sampling information, difficulty in characterizing the dynamic relationship of cascade thermal processes, and low prediction accuracy.

[0005] On one hand, embodiments of the present invention provide an asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy, comprising: Based on pre-collected heat exchanger operation data, electric heater operation data, and connecting pipeline operation data, a time-stamped multi-source operation dataset is constructed. Based on the multi-source running dataset, asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are constructed respectively, and combined to obtain the asynchronous multi-frequency time lag coupling enhanced input vector; A training sample set and a reliability-constrained adaptive composite kernel function are constructed based on the asynchronous multi-frequency time-delay coupling enhanced input vector; a temperature soft measurement model is established based on the reliability-constrained adaptive composite kernel function, the training sample set, and sample reliability weights. Based on the real-time operating data of the heat exchanger, the electric heater, and the connecting pipeline, an asynchronous multi-frequency time-delay coupled enhanced real-time input vector is constructed, and the predicted value of the hydrogen temperature at the outlet of the electric heater at the next moment is obtained through the temperature soft measurement model.

[0006] Furthermore, the heat exchanger operating data includes at least the following operating variables: heat exchanger inlet hydrogen temperature, heat exchanger inlet hydrogen pressure, heat exchanger inlet hydrogen flow rate, heat exchanger outlet hydrogen temperature, heat exchanger outlet hydrogen pressure, and heat exchanger valve opening. The operating data of the electric heater includes at least the following operating variables: electric heating power, electric heater inlet hydrogen temperature, electric heater inlet hydrogen pressure, electric heater outlet hydrogen temperature, and electric heater outlet hydrogen pressure. The operating data of the connecting pipeline includes at least the following operating variables: connecting pipeline temperature and connecting pipeline pressure.

[0007] Furthermore, the construction of asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features based on the multi-source operational dataset includes: Based on the target prediction time of the hydrogen temperature at the outlet of the electric heater, the original sampling sequence of each operating variable in its respective historical period is extracted from the multi-source operating dataset to form the historical dynamic sample window corresponding to each operating variable; based on each of the historical dynamic sample windows, a multi-source asynchronous sampling sample is constructed. Based on the multi-source asynchronous sampling samples, asynchronous multi-frequency sampling structural features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are constructed respectively.

[0008] Furthermore, the features for constructing the asynchronous multi-frequency sampling structure include: Based on the multi-source asynchronous sampling samples, obtain the valid sampling timestamps of each running variable within the corresponding historical dynamic sample window, and calculate the sampling time interval of each running variable based on adjacent valid sampling timestamps; Based on the sampling time intervals, the sampling frequency characteristics of each running variable are constructed, and the sampling frequency characteristics are combined to obtain the asynchronous multi-frequency sampling structure characteristics; The sampling frequency feature is calculated according to the following formula: ; in, Indicates the first The running variable in the first... Sampling frequency characteristics within a sample window; Indicates the first The running variable in the first... The average sampling interval within each sample window; Indicates the first The running variable in the first... The maximum sampling interval within a sample window; Indicates the first The running variable in the first... Standard deviation of the sampling interval within each sample window; Indicates the first The running variable in the first... The number of valid sampling points within a sample window; Indicates the first Each running variable relative to the target prediction time Data freshness.

[0009] Further, constructing the statistical state characteristics of the process variables includes: Based on the multi-source asynchronous sampling samples, at least one of the following is statistically analyzed for each operating variable within the corresponding historical dynamic sample window: window mean, window maximum, window minimum, window standard deviation, window first-to-last difference, window change slope, time-weighted mean, and time-weighted rate of change; the statistical results of each operating variable are combined to obtain the statistical state characteristics of the process variable. The time-weighted average is calculated according to the following formula: ; in, Indicates the first The running variable in the first... Time-weighted mean within each sample window; Indicates the first Within the sample window, the first The running variable in the first... The values ​​at each valid sampling time. This represents the time interval corresponding to the r-th sampling point; The time-weighted rate of change is calculated according to the following formula: ; in, Indicates the first The running variable in the first... Time-weighted rate of change within each sample window; Indicates the first Within the sample window, the first The first valid sampled value of each running variable; Indicates the first Within the sample window, the first The last valid sampled value of each running variable; Indicates the first Within the sample window, the first The first valid sampling time of each running variable, Indicates the first Within the sample window, the first The last valid sampling time of each running variable.

[0010] Furthermore, constructing the thermal state transfer characteristics between the preceding and following stages includes: Based on the multi-source asynchronous sampling samples, at least one of the following is calculated: heat exchanger state change characteristics, heat exchanger to electric heater transfer characteristics, electric heater historical temperature rise characteristics, valve regulation characteristics, and connecting pipeline transfer characteristics. The calculation results are then combined to obtain the front and rear stage heat state transfer characteristics. The heat exchanger state change characteristics are calculated based on the temperature rise amplitude and pressure loss on the hydrogen side of the heat exchanger; the heat exchanger-to-electric heater transfer characteristics are calculated based on the temperature and pressure transfer differences with hysteresis compensation between the heat exchanger outlet and the electric heater inlet; the electric heater historical temperature rise characteristics are calculated based on the historical temperature rise amplitude, pressure change, and historical unit electric heating power temperature rise efficiency of the electric heater; the valve regulation characteristics are calculated based on the change in the valve opening of the heat exchanger; and the connecting pipeline transfer characteristics are calculated based on the transfer difference between the connecting pipeline temperature and the heat exchanger outlet temperature, and the transfer difference between the electric heater inlet temperature and the connecting pipeline temperature.

[0011] Further, constructing the variable-level time lag feature includes: Based on the multi-source asynchronous sampling samples, the lag time of each operating variable relative to the hydrogen temperature at the outlet of the electric heater is determined; Based on the lag time, construct variable-level lag characteristics for each operating variable at the target lag time, and construct multi-order lag characteristics based on the dynamic change characteristics of each operating variable. Valid sampling points of each operating variable are selected in the neighborhood of the target lag time, and the estimated value of the target lag time is calculated based on each valid sampling point using a time kernel weighted aggregation method, so as to obtain the time kernel weighted estimation features of each operating variable. The variable-level time lag features are obtained by combining the variable-level lag features, multi-order lag features, and time kernel weighted estimation features of each operating variable.

[0012] Furthermore, the construction of the reliability-constrained adaptive composite kernel function includes: A training sample set is constructed based on the asynchronous multi-frequency time-delay coupling enhanced input vector and its corresponding measured value of hydrogen temperature at the outlet of the electric heater. Based on the training sample set, the predefined operating condition variable kernel, time lag kernel, thermal state transfer kernel, sampling frequency structure kernel, and interaction kernel are weighted and combined, and after being corrected by the sampling reliability coefficient, the reliability constraint adaptive composite kernel function is obtained.

[0013] Furthermore, the establishment of the temperature soft measurement model includes: weighting the kernel similarity of each training sample in the training sample set through the reliability constraint adaptive composite kernel function, and weighting the prediction error penalty intensity of each training sample through the sample reliability weight; The sample reliability weight is composed of time recentity weight, operating condition similarity weight, sampling frequency reliability weight, and thermal state transmission consistency weight.

[0014] Furthermore, the temperature soft measurement model is implemented based on a double-weighted least squares support vector machine, and the optimization objective of the temperature soft measurement model is expressed in the following form: ; The constraints of the temperature soft sensing model are expressed in the following form: ; in, Represents the weight vector; Indicates the bias term; Represents the modeling error variable In the The modeling error value for each training sample; Indicates the first Reliability weights for each training sample; Represents the regularization parameter; Indicates the first Measured hydrogen temperature at the outlet of the electric heater for each training sample; Represents a nonlinear mapping function. Indicates the number of training samples; Indicates the first Asynchronous multi-frequency time-delay coupling enhanced input vector for training samples.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: First, unlike related technologies which lack dynamic transmission mechanisms and suffer from asynchronous sampling information distortion, this invention retains the original sampling timestamps of each operating variable and constructs an asynchronous multi-frequency dynamic sample window based on the target prediction time of the hydrogen temperature at the electric heater outlet. Compared to traditional uniform interpolation or equal-interval resampling methods, this invention avoids the loss of dynamic information of high-frequency variables, false synchronization of low-frequency variables, and distortion of thermal state transmission relationships, enabling the sample input to more realistically reflect the actual operating state of the heat exchanger, connecting pipelines, and electric heater series system.

[0016] Secondly, unlike related technologies which struggle to characterize the dynamic relationships of cascade thermal processes, this invention unifies the transfer relationships of thermal states between stages, variable-level time lag relationships, and changes in operating states during heat exchanger preheating, connecting pipeline transmission, and electric heater reheating into explicit, modelable features. These features include asynchronous multi-frequency sampling structure characteristics, process variable statistical state characteristics, transfer characteristics of thermal states between stages, and variable-level time lag characteristics. Through this feature system, the model can simultaneously characterize differences in sampling structures, changes in operating conditions, the heat-hydrogen transfer process, and the lag in regulatory effects, avoiding the difficulty of traditional static black-box models in characterizing the dynamic relationships of cascade thermal processes.

[0017] Third, unlike related technologies which suffer from low prediction accuracy, this invention introduces a reliability-constrained adaptive composite kernel function and thermal state transfer consistency sample weights into the least squares support vector machine modeling framework, forming a dual-weighted temperature soft measurement model. The composite kernel function measures sample similarity from multiple perspectives, including operating condition variables, time lag, thermal state transfer, and sampling frequency structure, while simultaneously using sampling reliability coefficients to correct the kernel matrix. The sample reliability weights comprehensively consider time recentity, operating condition similarity, sampling frequency reliability, and thermal state transfer consistency to adjust the sample error penalty intensity.

[0018] Therefore, this invention can reduce the interference of long-term samples, low-reliability sampling samples and abnormal heat transfer samples on model training, and improve the accuracy, stability and adaptability of hydrogen temperature prediction at the outlet of the electric heater.

[0019] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of an asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the technical route of the asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy according to an embodiment of the present invention. Figure 3 This is a schematic diagram of information transmission in the pure hydrogen metallurgical hot hydrogen stepped heating process according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the features and integration of the technical route in an embodiment of the present invention; Figure 5 This is a schematic diagram of the temperature soft measurement model of the double-weighted least squares support vector machine according to an embodiment of the present invention. Detailed Implementation

[0021] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0022] The technical concept of this invention is as follows: First, for a pure hydrogen metallurgical thermal hydrogen cascade heating process consisting of a heat exchanger, connecting pipelines, and an electric heater, the hydrogen temperature at the outlet of the electric heater is used as the prediction target. Multi-source operating data from the heat exchanger, electric heater, and connecting pipelines are collected, and the original sampling timestamps of each operating variable are retained. Second, a differentiated historical dynamic sample window is constructed based on the target prediction time. Asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are extracted to form an asynchronous multi-frequency time-delay coupling enhanced input vector. Third, a reliability-constrained adaptive composite kernel function is constructed, introducing operating condition variable kernels, time lag kernels, thermal state transfer kernels, sampling frequency structure kernels, and interaction kernels. The kernel matrix is ​​corrected based on the sampling reliability coefficient. Finally, a consistent sample weight for thermal state transfer is constructed, and a double-weighted least squares support vector machine temperature soft measurement model is established to achieve online prediction of the hydrogen temperature at the outlet of the electric heater. Therefore, this invention can improve the accuracy, stability, and engineering applicability of soft measurement of hydrogen temperature at the outlet of the electric heater under conditions such as asynchronous sampling of multiple source variables, thermal state transfer between stages, variable lag, and changes in operating conditions.

[0023] A specific embodiment of the present invention discloses an asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy, such as... Figure 1 and Figure 2 As shown, the steps S1 to S4 are as follows: Step S1: Based on the pre-collected heat exchanger operation data, electric heater operation data, and connecting pipeline operation data, construct a time-stamped multi-source operation dataset.

[0024] In practice, this embodiment of the invention takes a heat exchanger, connecting pipeline and electric heater arranged in series as the object, and focuses on the online prediction of hydrogen temperature at the outlet of the electric heater. It collects multi-source operating data including heat exchanger operating data, electric heater operating data and connecting pipeline operating data, and retains the original sampling timestamp of each variable.

[0025] refer to Figure 3 As shown, from a process structure perspective, the pure hydrogen thermal hydrogen cascade heating process typically consists of a heat exchanger, connecting pipelines, and an electric heater in sequence. The hydrogen inlet temperature, pressure, and flow rate determine the initial thermal and flow states of the hydrogen before it enters the heat exchanger; the hydrogen outlet temperature and pressure, transmitted through the connecting pipelines, form the inlet boundary of the electric heater; the valve opening of the heat exchanger affects the flow rate and heat exchange regulation process on the heat exchanger side; the electric heating power determines the intensity of the electric heater's supplemental heating of the hydrogen; and the temperature and pressure of the connecting pipelines reflect the transmission state between the upstream and downstream equipment. Therefore, the hydrogen outlet temperature of the electric heater is not determined by a single local variable of a single equipment, but rather by the combined effects of the heat exchanger preheating state, the transmission state of the connecting pipelines, the electric heater inlet boundary, the electric heating power input, and the thermal inertia of the equipment, forming the output of a cascaded system.

[0026] Specifically, the heat exchanger operating data includes at least the following operating variables: heat exchanger inlet hydrogen temperature, heat exchanger inlet hydrogen pressure, heat exchanger inlet hydrogen flow rate, heat exchanger outlet hydrogen temperature, heat exchanger outlet hydrogen pressure, and heat exchanger valve opening. The operating data of the electric heater includes at least the following operating variables: electric heating power, electric heater inlet hydrogen temperature, electric heater inlet hydrogen pressure, electric heater outlet hydrogen temperature, and electric heater outlet hydrogen pressure. The operating data of the connecting pipeline includes at least the following operating variables: connecting pipeline temperature and connecting pipeline pressure.

[0027] Specifically, the inlet hydrogen temperature, inlet hydrogen pressure, and inlet hydrogen flow rate of the heat exchanger are used to characterize the initial thermal and flow states of the hydrogen before it enters the heat exchanger; the outlet hydrogen temperature and outlet hydrogen pressure of the heat exchanger are used to characterize the thermal and pressure states of the hydrogen after it has been preheated by the heat exchanger; the valve opening of the heat exchanger is used to characterize the flow rate or heat exchange regulation state on the heat exchanger side; the electric heating power is used to characterize the heat input intensity of the electric heater to the hydrogen; the inlet hydrogen temperature and inlet hydrogen pressure of the electric heater are used to characterize the state of the hot hydrogen before it enters the electric heater; the outlet hydrogen temperature and outlet hydrogen pressure of the electric heater are used to characterize the output state of the hydrogen after it has been heated by the electric heater; and the connecting pipeline temperature and connecting pipeline pressure are used to characterize the pipeline transmission state between the heat exchanger and the electric heater, and between the electric heater and the pure hydrogen vertical furnace.

[0028] Furthermore, for any running variable The original sampling sequence is preserved according to its actual collection time, and is represented in the following form: (1) in, Indicates the first The first running variable One original sampling timestamp, Indicates the first The number of sampling points for each running variable within the corresponding sampling period.

[0029] In this embodiment, there is no need to pre-force interpolation, resample at equal intervals, or align the time stamps of each running variable. Instead, the original sampling frequency, the actual acquisition time, and the asynchronous sampling relationship of each running variable are preserved so that asynchronous multi-frequency sampling features, time lag features, and thermal state transfer features between stages can be constructed based on the original timestamps.

[0030] Step S2: Based on the multi-source running dataset, construct asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features respectively, and combine them to obtain the asynchronous multi-frequency time lag coupling enhanced input vector.

[0031] Specifically, refer to Figure 4 As shown, unlike the method of forcibly interpolating or uniformly resampling different variables, this embodiment uses the target prediction time of the hydrogen temperature at the outlet of the electric heater as the benchmark, and constructs historical dynamic sample windows for different operating variables to retain the true sampling frequency, acquisition time and asynchronous relationship of multi-source variables.

[0032] In the sample characterization process, asynchronous multi-frequency sampling structure features are used to characterize the sampling interval, the number of effective sampling points, and data freshness; process variable statistical state features are used to describe the operating status of the heat exchanger, electric heater, and connecting pipeline within the historical window; front-to-back stage thermal state transfer features are used to characterize the temperature changes, pressure changes, valve regulation effects, and electric heating supplementary effects during the transfer of hot hydrogen along the "heat exchanger-connecting pipeline-electric heater" process; and variable-level time lag features are used to describe the delayed impact of each operating variable on the hydrogen temperature at the outlet of the electric heater. Furthermore, when the input features involve the hydrogen temperature at the outlet of the electric heater, only historical values, lag values, window statistical values, or the most recent effective sampling value available before the target prediction time are used; the measured value of the hydrogen temperature at the outlet of the electric heater at the target prediction time is not used as the model input. Based on this, the above features are combined to form an asynchronous multi-frequency time-lag coupled enhanced input vector, and the hydrogen temperature at the outlet of the electric heater at the target prediction time is used as the model output to construct a temperature soft measurement sample set.

[0033] In practice, the specific steps include: using the target prediction time of the hydrogen temperature at the outlet of the electric heater as a benchmark, extracting the original sampling sequence of each operating variable in its respective historical period from the multi-source operating dataset to form a historical dynamic sample window corresponding to each operating variable; and constructing multi-source asynchronous sampling samples based on each of the historical dynamic sample windows. It is understood that in this embodiment, each running variable corresponds to a historical dynamic sample window, which consists of the original sampling sequence of the variable in the historical period before the target prediction time; when all running variables are at the same target prediction time, their respective historical dynamic sample windows together constitute a multi-source asynchronous sampling sample.

[0034] Furthermore, based on the aforementioned multi-source asynchronous sampling samples, asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are constructed respectively. It should be noted that each of these four features is a feature set containing multiple statistical or physical quantities. For example, each element in each feature set is calculated for the same target prediction time (i.e., the k-th sampling sample), and is used to describe the operational characteristics of that sample from different aspects such as sampling structure, statistical state, thermal transfer relationship, and time lag correlation.

[0035] In some implementations, reference continues. Figure 4 As shown, constructing an asynchronous multi-frequency time-delay coupling enhanced input vector specifically includes the following stages: First, target time alignment and dynamic window construction.

[0036] The target prediction time is the hydrogen temperature at the outlet of the electric heater (the temperature variable to be predicted). Based on the original sampling sequences of each operating variable collected above, a historical dynamic sample window corresponding to each operating variable is constructed to form a multi-source asynchronous sampling sample for temperature soft measurement modeling.

[0037] For the One running variable The corresponding historical dynamic sample window is constructed as follows: (2) in, Indicates the first The historical window length corresponds to each operating variable. The historical window length can be the same for different operating variables, or it can be differentiated based on the variable's sampling frequency, dynamic response speed, transmission lag time, thermal inertia characteristics, and its influence on the temperature variable to be predicted. For example, for operating variables with fast dynamic response, direct adjustment effects, or high sampling frequencies, a relatively short historical window can be set to highlight their recent impact on the temperature state at the current target prediction time; for operating variables strongly affected by thermal inertia, transmission lag, or equipment heat storage effects, a relatively long historical window can be set to retain their delayed and cumulative effects on the temperature variable to be predicted; for operating variables with long sampling periods or data refresh delays, the corresponding window length can be determined based on the number of effective sampling points and data freshness.

[0038] Using the above method, without forcing equal-interval resampling or uniform time-scale alignment of each operating variable, a method is constructed based on the target prediction time. Based on the principle of asynchronous multi-frequency dynamic samples, which can preserve the original sampling frequency and real time relationship of multi-source variables.

[0039] For example, the temperature variable to be predicted is defined as , indicating the target prediction time The outlet hydrogen temperature of the electric heater is used as the output label during the model training phase when the temperature variable to be predicted has measured data. When the temperature variable to be predicted cannot be directly measured or there is a measurement delay in the subsequent online operation phase, the real-time sample is input into the temperature soft measurement model to obtain the temperature prediction value at the corresponding target prediction time.

[0040] Second, the construction of asynchronous multi-frequency sampling structure features.

[0041] In practice, this includes: obtaining the valid sampling timestamps of each running variable within the corresponding historical dynamic sample window based on the multi-source asynchronous sampling samples, and calculating the sampling time interval of each running variable based on adjacent valid sampling timestamps; Based on the sampling time intervals, the sampling frequency characteristics of each running variable are constructed, and the sampling frequency characteristics are combined to obtain the asynchronous multi-frequency sampling structure characteristics.

[0042] For example, in cases where there are inconsistent sampling frequencies, asynchronous acquisition times, and local data refresh delays among various operating variables, an asynchronous multi-frequency sampling feature is constructed based on the aforementioned historical dynamic sample window to characterize the sampling time structure of each operating variable.

[0043] For the One running variable In the Within the historical dynamic sample window corresponding to each sample, obtain the valid sampling timestamp sequence of the runtime variable, represented in the following form: (3) in, Indicates the first Within the sample window, the first The first running variable One valid sampling timestamp Indicates the first Within the sample window, the first The number of valid sampling points for each running variable.

[0044] Furthermore, the number of sampling timestamps is calculated based on adjacent sampling timestamps. The running variable in the first... The sampling time interval within a sample window is expressed in the following form: (4) Furthermore, based on the aforementioned sampling time interval, the first... The running variable in the first... Sampling frequency characteristics in a sample It is represented in the following form: (5) in, Indicates the first The running variable in the first... The average sampling interval within each sample window; Indicates the first The running variable in the first... The maximum sampling interval within a sample window; Indicates the first The running variable in the first... Standard deviation of the sampling interval within each sample window; Indicates the first The running variable in the first... The number of valid sampling points within a sample window; Indicates the first Each running variable relative to the target prediction time Data freshness.

[0045] Furthermore, the data freshness Indicates the target prediction time With the The time difference between the most recent valid sampling time of each running variable before the target time is calculated using the following formula: (6) in, Indicates the first Within the sample window, the first The distance of each running variable from the target prediction time Most recent valid sampling timestamp.

[0046] By combining the sampling frequency features corresponding to each running variable, we obtain the first... The asynchronous multi-frequency sampling characteristics of each sample are represented in the following form: (7) in, Indicates the number of runtime variables involved in the modeling; , , They represent the 1st, 2nd, and 3rd respectively. The sampling frequency characteristic value of each running variable within the k-th sample window.

[0047] Therefore, by using the above method, the differences in sampling frequency, sampling interval fluctuation, number of effective sampling points, and data freshness among different operating variables can be transformed into explicit features that can be input into the temperature soft measurement model. This enables the model to identify the asynchronous sampling structure of multi-source operating data while retaining the original timestamp information, and reduces the distortion caused by forced resampling or unified timescale alignment to the dynamic characteristics of thermal state.

[0048] Third, constructing statistical state characteristics of process variables.

[0049] In practice, this includes: based on the multi-source asynchronous sampling samples, statistically analyzing at least one of the following for each operating variable within the corresponding historical dynamic sample window: window mean, window maximum, window minimum, window standard deviation, window first-to-last difference, window change slope, time-weighted mean, and time-weighted rate of change; combining the statistical results of each operating variable to obtain the statistical state characteristics of the process variables. That is, based on the aforementioned constructed historical dynamic sample window, statistical features are extracted from the multi-source operating variables to form a set of statistical features of process variables used to characterize the operating state of the thermal hydrogen cascade heating process.

[0050] It is worth noting that when the hydrogen temperature at the outlet of the electric heater... When the temperature variable is to be predicted, the characteristics of the process variable statistical state features involving the hydrogen temperature at the electric heater outlet are all based on the target prediction time. The target prediction time is constructed using previously available historical sampled values, lagged sampled values, window statistics, or the most recent valid sampled value. The measured value of the hydrogen temperature at the outlet of the electric heater was used as the model input.

[0051] For example, for the first One running variable In the Within the historical dynamic sample window corresponding to each sampled sample, at least one of the following is extracted as the statistical state feature of the process variable corresponding to that operating variable: window mean, window maximum, window minimum, window standard deviation, window first-to-last difference, window change slope, time-weighted mean, and time-weighted change rate.

[0052] The time-weighted mean is used to characterize the window average state of the running variable under non-equal interval sampling conditions. For the first... One running variable In its first The time-weighted mean within each sample window is calculated using the following formula: (8) in, Indicates the first The running variable in the first... Time-weighted mean within each sample window; Indicates the first Within the sample window, the first The running variable in the first... The values ​​at each valid sampling time. This represents the time interval corresponding to the r-th sampling point; The time-weighted rate of change is used to characterize the trend of the operating variable over time within the historical dynamic sample window, and is calculated according to the following formula: (9) in, Indicates the first The running variable in the first... Time-weighted rate of change within each sample window; Indicates the first Within the sample window, the first The first valid sampled value of each running variable; Indicates the first Within the sample window, the first The last valid sampled value of each running variable; Indicates the first Within the sample window, the first The first valid sampling time of each running variable, Indicates the first Within the sample window, the first The last valid sampling time of each running variable.

[0053] By combining the statistical state characteristics of the process variables corresponding to each operating variable, we obtain the first... Statistical state characteristics of process variables in a sample It is represented in the following form: (10) in, They represent the 1st, 2nd, and 3rd respectively. Statistical state characteristics of process variables within the kth sample window; This indicates the number of runtime variables involved in the modeling process.

[0054] By using the above method, while preserving the original sampling time structure of each operating variable, the operating states of the heat exchanger, electric heater, and connecting pipeline, such as temperature, pressure, flow rate, valve opening, and electric heating power, are transformed into statistical features that can participate in temperature soft measurement modeling. At the same time, the bias caused by non-equal interval sampling to the window state characterization results is reduced.

[0055] Third, the characteristics of thermal state transfer between the preceding and following stages are constructed.

[0056] In practice, the specific steps include: based on the multi-source asynchronous sampling samples, calculating at least one of the following: heat exchanger state change characteristics, heat exchanger to electric heater transfer characteristics, electric heater historical temperature rise characteristics, valve regulation characteristics, and connecting pipeline transfer characteristics, and combining the calculation results to obtain the front-to-back stage heat state transfer characteristics.

[0057] Specifically, based on a process structure consisting of a heat exchanger, connecting pipelines, and an electric heater connected in series, a front-to-back stage thermal state transfer characteristic is constructed to characterize the transfer process of hot hydrogen between the front and back stages. This front-to-back stage thermal state transfer characteristic reflects the combined influence of the heat exchanger preheating state, heat exchanger valve regulation state, connecting pipeline transmission state, electric heating power input state, and the historical temperature rise state of the electric heater on the hydrogen outlet temperature of the electric heater.

[0058] Furthermore, the heat exchanger state change characteristics are calculated based on the temperature rise amplitude and pressure loss on the hydrogen side of the heat exchanger; the heat exchanger-to-electric heater transfer characteristics are calculated based on the temperature and pressure transfer differences with hysteresis compensation between the heat exchanger outlet and the electric heater inlet; the electric heater historical temperature rise characteristics are calculated based on the historical temperature rise amplitude, pressure change, and historical unit electric heating power temperature rise efficiency of the electric heater; the valve regulation action characteristics are calculated based on the change in the heat exchanger valve opening; and the connecting pipeline transfer characteristics are calculated based on the transfer difference between the connecting pipeline temperature and the heat exchanger outlet temperature, and the transfer difference between the electric heater inlet temperature and the connecting pipeline temperature. Wherein: (1) The characteristics of the heat exchanger state change include the temperature rise of the hydrogen side of the heat exchanger and the pressure loss of the hydrogen side of the heat exchanger.

[0059] The temperature rise on the hydrogen side of the heat exchanger This characterizes the preheating effect of the heat exchanger on hydrogen, and is expressed in the following form: (11) in, This indicates the hydrogen temperature at the heat exchanger inlet. This indicates the hydrogen temperature at the heat exchanger outlet. This indicates the hysteresis time of the heat exchanger's thermal response.

[0060] Furthermore, the hydrogen side pressure loss of the heat exchanger This is used to characterize the pressure changes and flow state of hydrogen gas as it passes through a heat exchanger, and is expressed in the following form: (12) in, This indicates the hydrogen pressure at the heat exchanger inlet. This indicates the hydrogen pressure at the heat exchanger outlet.

[0061] (2) The heat exchanger to electric heater transmission characteristics include the temperature transmission difference and pressure transmission difference from the heat exchanger outlet to the electric heater inlet.

[0062] The temperature transfer difference is: (13) in, This indicates the inlet hydrogen temperature of the electric heater. This indicates the thermal state transfer lag time between the heat exchanger outlet and the electric heater inlet. The temperature transfer difference characterizes the temperature change of hydrogen gas at the heat exchanger outlet before it enters the electric heater after being transported through the connecting pipeline.

[0063] The pressure transmission difference is: (14) in, This indicates the hydrogen pressure at the inlet of the electric heater. The pressure transmission difference is used to characterize the pressure change and pipeline flow conditions between the heat exchanger outlet and the electric heater inlet.

[0064] (3) The historical temperature rise characteristics of the electric heater include the historical temperature rise range of the electric heater, the pressure change of the electric heater, and the historical unit electric heating power temperature rise efficiency.

[0065] The historical temperature rise of the electric heater This characterizes the historical temperature rise of the electric heater that can be observed before the target time, and is expressed in the following form: (15) in, This indicates the hydrogen temperature at the outlet of the electric heater relative to the target prediction time. The available lag time, and satisfying ; This indicates the hysteresis time of the internal thermal response of the electric heater.

[0066] The pressure change of the electric heater This is used to characterize the pressure changes and operating status of hydrogen flow through the electric heater, and is expressed in the following form: (16) in, This indicates the hydrogen pressure at the outlet of the electric heater.

[0067] The historical unit electric heating power temperature rise efficiency This is used to characterize the effective contribution of the electric heating power to the temperature rise of hydrogen before the target time, and is expressed in the following form: (17) in, Indicates the electric heating power. This indicates the lag time in which the electric heating power affects the temperature of the hydrogen gas. To prevent extremely small positive numbers with a denominator of zero.

[0068] (4) The valve regulation characteristics include the lag value of the heat exchanger valve opening and the change in the heat exchanger valve opening; these are used to characterize the influence of the heat exchanger-side regulation action on the subsequent heat state transfer process.

[0069] The hysteresis value of the heat exchanger valve opening is expressed as follows: (18) The change in the opening degree of the heat exchanger valve is expressed as follows: (19) in, Indicates the valve opening degree of the heat exchanger. This indicates the lag time at which changes in the opening degree of a heat exchanger valve affect the temperature, pressure, or flow rate of hydrogen at the heat exchanger outlet.

[0070] (5) The transmission characteristics of the connecting pipeline include the transmission difference between the temperature of the connecting pipeline and the outlet temperature of the heat exchanger, and the transmission difference between the inlet temperature of the electric heater and the temperature of the connecting pipeline; these are used to characterize the influence of the connecting pipeline on the temperature transmission process from the outlet of the heat exchanger to the inlet temperature of the electric heater.

[0071] The temperature difference between the connecting pipeline and the heat exchanger outlet temperature is: (20) The temperature difference between the inlet temperature of the electric heater and the temperature of the connecting pipe is: (twenty one) in, Indicates the temperature of the connecting pipes. This indicates the lag time of heat transfer in the connecting pipeline.

[0072] As can be seen from the above, combining one or more of the above features will yield the first... Thermal state transfer characteristics of each sample before and after stage It is represented in the following form: (twenty two) in, They represent the first The 1st, 2nd, and 3rd samples in the sampled data Characteristics of thermal state transfer between preceding and following stages.

[0073] As can be seen from the above, by using the above method, the heat exchanger preheating state, heat exchanger valve adjustment state, connecting pipeline transmission state, electric heater historical temperature rise state, and electric heating power input state are transformed into explicit coupled features that can participate in temperature soft measurement modeling. This allows the model to characterize the transmission process of thermal state along the series system of heat exchanger, connecting pipeline, and electric heater, while avoiding the misrepresentation of the target prediction time. The measured value of the hydrogen temperature at the outlet of the electric heater was directly input into the model.

[0074] Fourth, the construction of variable-level time lag features.

[0075] In practice, this specifically includes: determining the lag time of each operating variable relative to the hydrogen temperature at the outlet of the electric heater based on the multi-source asynchronous sampling samples; Based on the lag time, construct variable-level lag characteristics for each operating variable at the target lag time, and construct multi-order lag characteristics based on the dynamic change characteristics of each operating variable. Valid sampling points of each operating variable are selected in the neighborhood of the target lag time, and the estimated value of the target lag time is calculated based on each valid sampling point using a time kernel weighted aggregation method, so as to obtain the time kernel weighted estimation features of each operating variable. The variable-level time lag features are obtained by combining the variable-level lag features, multi-order lag features, and time kernel weighted estimation features of each operating variable.

[0076] Specifically, considering the flow transport lag, thermal response lag, and regulation lag that exist in the process of hot hydrogen transport in heat exchangers, connecting pipelines, and electric heaters, the lag time of each operating variable relative to the temperature variable to be predicted is determined, and variable-level time lag characteristics are constructed.

[0077] For the One running variable Its relative to the temperature variable to be predicted The lag time is denoted as The lag time It can be determined by at least one of the following methods: (1) Mechanism estimation based on hydrogen flow rate, pipeline length, effective equipment volume, gas transmission path and equipment structural parameters; (2) Calculate the operating variables based on historical operating data. With the temperature variable to be predicted The correlation under different lag times was determined, and the lag time that satisfies the preset conditions was selected. (3) Determined based on the operating experience, control cycle, sensor placement and equipment thermal inertia characteristics of the pure hydrogen thermal hydrogen stepped heating process; (4) The goal is to minimize the prediction error of the temperature soft measurement model on the validation sample, which is determined by grid search, particle swarm optimization, genetic algorithm or Bayesian optimization.

[0078] For example, when using cross-correlation analysis to determine the lag time, the first... The lag time corresponding to each running variable can be expressed as: (twenty three) in, Represents the correlation measurement function. Indicates the first The running variables in the lag time The value to be taken below, This indicates the temperature variable to be predicted at time [time]. The value of .

[0079] Determine the lag time Then, for the target prediction time Construct the first The variable-level time lag characteristics of each running variable are expressed in the following form: (twenty four) Furthermore, based on the dynamic characteristics of the operating variables and the model input requirements, a multi-order lag feature is constructed, expressed in the following form: (25) in, This indicates the lag order.

[0080] Preferably, when there is asynchronous sampling and inconsistent sampling frequencies among the multi-source operating variables, if the target lag time... If no corresponding precise sample value exists, then the first sample value is selected from the neighborhood of the target's lag time. The effective sampling points of each running variable are used, and the estimated value of the target lag time is calculated using a time kernel weighted aggregation method. The formula includes: (26) in, Indicates the first The running variables at the target lag time Time-weighted estimate; Indicates the first Within the sample window, the first The first running variable One valid sampling timestamp; Indicates the corresponding sampled value; Indicates the first Within the sample window, the first The number of valid sampling points for each running variable; Indicates the first The time kernel decay parameters corresponding to each running variable.

[0081] As shown above, the time kernel weighted aggregation method assigns higher weights to sampling points closer to the target lag time and lower weights to sampling points farther from the target lag time. Therefore, this method can obtain effective estimates of the operating variables at the target lag time without forcing uniform time-scale alignment or equal-interval interpolation.

[0082] Furthermore, the variable-level lag characteristics, multi-order lag characteristics, and / or time kernel-weighted estimation characteristics corresponding to each operating variable are combined to form the first... Variable-level time lag characteristics of each sample It is represented in the following form: (27) in, They represent the 1st, 2nd, and 3rd samples in the kth sample, respectively. Individual variable-level time lag characteristics.

[0083] Therefore, by using the above method, the delayed effects of the heat exchanger inlet state, heat exchanger outlet state, heat exchanger valve opening, electric heating power, electric heater inlet state, electric heater outlet state, and connecting pipeline state on the temperature variable to be predicted are transformed into explicit features that can participate in temperature soft measurement modeling, thereby improving the model's ability to characterize the lag in thermal hydrogen transport, equipment thermal response lag, and regulation lag.

[0084] Fifth, constructing input vectors to enhance coupling.

[0085] Specifically, the asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features obtained from the aforementioned construction are combined to form the first... Asynchronous multi-frequency time-delay coupling enhanced input vector of each sampled sample : (28) in, Indicates the first The statistical state characteristics of process variables of each sample are used to characterize the operating status of heat exchangers, electric heaters and connecting pipelines within the historical dynamic sample window. Indicates the first The asynchronous multi-frequency sampling structure features of each sample are used to characterize the sampling frequency, sampling interval, number of effective sampling points, and data freshness of each running variable; Indicates the first The thermal state transfer characteristics of each sample before and after the stage are used to characterize the state transfer relationship of hot hydrogen between the heat exchanger, connecting pipeline and electric heater. Indicates the first The variable-level time lag characteristics of each sample are used to characterize the delayed effect of each operating variable on the hydrogen temperature at the outlet of the electric heater.

[0086] Step S3: Construct a training sample set and a reliability-constrained adaptive composite kernel function based on the asynchronous multi-frequency time-delay coupling enhanced input vector; establish a temperature soft measurement model based on the reliability-constrained adaptive composite kernel function, the training sample set, and sample reliability weights.

[0087] When implementing, refer to Figure 5 As shown, the construction of the reliability-constrained adaptive composite kernel function includes: A training sample set is constructed based on the asynchronous multi-frequency time-delay coupling enhanced input vector and its corresponding measured value of hydrogen temperature at the outlet of the electric heater. Specifically, the above asynchronous multi-frequency time-delay coupling enhanced input vector As input to the temperature soft sensing model, the target prediction time is... The corresponding hydrogen temperature at the outlet of the electric heater is used as the model's output label to form training samples, represented in the following form: (29) Among them, the model output variables Indicates the first The expression for the hydrogen temperature at the outlet of the electric heater corresponding to each training sample is: (30) in, Indicates the target prediction time The outlet hydrogen temperature of the electric heater.

[0088] As shown above, during the model training phase, the measured value of the hydrogen temperature at the electric heater outlet is used as the supervised learning label. In the subsequent online soft measurement phase, the real-time constructed asynchronous multi-frequency time-delay coupled enhanced real-time input vector can be input into the trained temperature soft measurement model to obtain the predicted value of the hydrogen temperature at the electric heater outlet corresponding to the target prediction time. Through this method, multi-source asynchronous sampling data, thermal state transfer relationships, and the effects of variable time delays are uniformly expressed as model inputs, enabling the temperature soft measurement model to perform online prediction of the hydrogen temperature at the electric heater outlet.

[0089] In some implementations, reference continues. Figure 5 As shown, a reliability-constrained adaptive composite kernel function is constructed, specifically including: based on the training sample set, a weighted combination of predefined operating condition variable kernel, time lag kernel, thermal state transfer kernel, sampling frequency structure kernel, and interaction kernel is performed, and after correction by sampling reliability coefficient, the reliability-constrained adaptive composite kernel function is obtained, so as to weaken the influence of low-reliability sampling samples on similarity measurement; at the same time, it can effectively address the problem that ordinary least squares support vector machines are difficult to simultaneously represent asynchronous sampling structure, thermal state transfer relationship, and time lag coupling characteristics.

[0090] For example, based on the first Asynchronous multi-frequency time-delay coupling enhanced input vector of each sample We construct an adaptive composite kernel function suitable for weighted least squares support vector machines.

[0091] Regarding the first The first sample and the first Each sample is used to construct a kernel of operating condition variables. Time lag kernel Thermal state transfer nuclei and sampling frequency structure kernel : (31) in, Let represent the statistical state characteristics of the process variables for the i-th and j-th samples, respectively; Let represent the variable-level time lag features of the i-th and j-th samples, respectively; These represent the thermal state transfer characteristics between the i-th and j-th samples, respectively. Let i and j represent the asynchronous multi-frequency sampling structure features of the i-th and j-th samples, respectively. , , and These represent the kernel width parameters for the operating condition variable kernel, time lag kernel, thermal state transfer kernel, and sampling frequency structure kernel, respectively.

[0092] Preferably, to enhance the model's ability to represent asynchronous multi-frequency sampling structures and thermal state transfer processes, an interactive kernel function is further constructed. The interactive kernel function includes a time lag-thermal state transfer interactive kernel. and sampling frequency-thermal state transfer interaction kernel : (32) in, represents the asynchronous multi-frequency time-delay coupling enhancement input vectors of the i-th and j-th samples, respectively; the time lag-thermal state transfer interaction kernel is used to characterize the joint similarity of the two samples in the thermal state transfer process and its lag structure; the sampling frequency-thermal state transfer interaction kernel is used to characterize the joint similarity of the two samples in the sampling time structure and thermal state transfer relationship. Based on this, a reliability-constrained adaptive composite kernel function is constructed. : (33) in, , , , , , These represent the weights of the operating condition variable kernel, time lag kernel, thermal state transfer kernel, sampling frequency structure kernel, time lag-thermal state transfer interaction kernel, and sampling frequency-thermal state transfer interaction kernel, respectively. The combined weights of the aforementioned kernel functions can adaptively adjust the contribution of various kernel functions to temperature prediction. They are adaptively determined based on the validation set prediction error, the alignment between the kernel matrix and the temperature output variable, or the comprehensive evaluation index of the temperature soft sensing model, satisfying the following relationship: (34) In some preferred embodiments, the reliability constraint adaptive composite kernel function is modified based on the sample sampling reliability to obtain the composite kernel function of the sampling reliability constraint. : (35) in, and They represent the first The first sample and the first The sampling reliability coefficient of each sample.

[0093] For the Each sample, its sampling reliability coefficient It can be represented as: (36) in, Indicates the number of runtime variables involved in the modeling; Indicates the first In the nth sample The data freshness of each running variable relative to the target time; Indicates the first In the nth sample The maximum sampling interval of each running variable within the historical dynamic sample window; Indicates the first In the nth sample Standard deviation of the sampling interval for each running variable; , and They represent the first The data freshness weight coefficient, the maximum sampling interval weight coefficient, and the sampling interval fluctuation weight coefficient are corresponding to the running variables; This represents the sampling reliability attenuation parameter. It corresponds to the sampling reliability coefficient when the data freshness of the running variables in the sample is low, the maximum sampling interval is large, or the sampling interval fluctuates significantly. Decrease; when the sample of running variables is sufficiently sampled, the data is updated in a timely manner, and the sampling interval is relatively stable, the corresponding sampling reliability coefficient decreases. Increase.

[0094] In this way, the reliability-constrained adaptive composite kernel function can not only measure the similarity of operating conditions, time lag, thermal state transfer, and sampling frequency structure among samples, but also further characterize the coupling relationship between time lag and thermal state transfer, and between sampling frequency structure and thermal state transfer, through the interactive kernel function. Simultaneously, by modifying the kernel function with the sampling reliability coefficient, the similarity contribution of samples with insufficient sampling, insufficient data freshness, or large fluctuations in sampling intervals in the kernel matrix is ​​weakened, thereby improving the robustness of the subsequent least squares support vector machine model to asynchronous multi-frequency sampled data.

[0095] In some implementations, considering the drift in operating conditions, inconsistent sampling frequencies, differences in data freshness, and fluctuations in the thermal state transfer process of the pure hydrogen-thermal hydrogen stepped heating process, it is necessary to construct a sample reliability weight for heat transfer consistency to adjust the error penalty intensity of different training samples in the weighted least squares support vector machine model.

[0096] Preferably, the sample reliability weight is composed of time recentity weight, operating condition similarity weight, sampling frequency reliability weight, and thermal state transmission consistency weight.

[0097] For example, the Sample reliability weights for each training sample Represented as: (37) in, Indicates the weight of time recentity. Indicates the weight of similarity in working conditions. Represents the reliability weight of the sampling frequency. This represents the consistency weight for thermal state propagation.

[0098] Preferably, the time recency weight is: (38) in, Indicates the current prediction time; Indicates the first The target time corresponding to each sample; This represents the time decay parameter.

[0099] The similarity weight of the working conditions is: (39) in, This represents the statistical state characteristics of process variables under the current operating conditions. Indicates the first Statistical state characteristics of the working process variables of each training sample; This represents the parameter for attenuation of operating condition similarity.

[0100] The reliability weight of the sampling frequency is: (40) in, Indicates the first In the training samples, the th The time interval between the most recent sampling of each running variable and the target time. Indicates the first In the training samples, the th The maximum sampling interval for each running variable and They represent the first The data freshness weight coefficient and sampling interval weight coefficient corresponding to each running variable This represents the reliability attenuation parameter of the sampling frequency. Indicates the number of variables used in the operation.

[0101] The thermal state transfer consistency weight is used to evaluate whether the transfer relationship between the heat exchanger outlet state, the connecting pipeline state, and the electric heater inlet state in the sample conforms to the continuous transfer law of the thermal hydrogen stepped heating process; its expression is: (41) in, Indicates the first Thermal state transfer bias of each training sample This represents the thermal state transfer consistency attenuation parameter.

[0102] Preferably, the thermal state transmission deviation can be composed of both temperature transmission deviation and pressure transmission deviation, and can be expressed in the following form: (42) in, and These represent the weighting coefficients for temperature transmission deviation and pressure transmission deviation, respectively.

[0103] When considering the influence of the temperature of the connecting pipeline, the thermal state transfer deviation can also be expressed as: (43) in, Indicates the temperature of the connecting pipes. This indicates the lag time of heat transfer in the connecting pipeline.

[0104] As can be seen from the above, by using the sample reliability weights, the model pays more attention to recent samples, samples with similar working conditions, samples with reliable sampling structures, and samples with consistent thermal state transmission relationships during the training process, thereby reducing the impact of long-term samples, insufficiently sampled samples, abnormal transmission samples, and low-reliability samples on the solution of model parameters.

[0105] In some implementations, establishing the temperature soft measurement model includes: weighting the kernel similarity of each training sample in the training sample set using the reliability-constrained adaptive composite kernel function, and weighting the prediction error penalty intensity of each training sample using the sample reliability weights; the two correction methods are applied to the kernel matrix layer and the objective function layer respectively, thereby forming a dual-weighted modeling mechanism for asynchronous multi-frequency time-delay coupled data; specifically including: Based on the asynchronous multi-frequency time-delay coupling enhanced input vector constituting the training samples Measured hydrogen temperature at the outlet of the electric heater Composite kernel function with sampling reliability constraints and sample reliability weights A double-weighted least squares support vector machine (LS-SVM) temperature soft measurement model was established. Wherein: (1) Composite kernel function constrained by sampling reliability The kernel similarity between training samples is weighted and corrected. (2) Through sample reliability weights , for the The prediction error penalty intensity of each training sample is weighted and corrected.

[0106] Furthermore, the optimization objective of the temperature soft-sensing model implemented based on a double-weighted least squares support vector machine is expressed in the following form: (44) The constraints of the temperature soft sensing model are expressed in the following form: (45) in, Represents the weight vector; Indicates the bias term; Represents the modeling error variable In the The modeling error value for each training sample; Indicates the first Reliability weights for each training sample; Represents the regularization parameter; Indicates the first Measured hydrogen temperature at the outlet of the electric heater for each training sample; Represents a nonlinear mapping function. Indicates the number of training samples; Indicates the first Asynchronous multi-frequency time-delay coupling enhanced input vector for training samples.

[0107] Preferably, by introducing Lagrange multipliers The above model optimization problem is transformed into a system of linear equations for solution: (46) in, Represents the kernel matrix, whose elements , This represents the sample weight regularization diagonal matrix, where the first... The diagonal elements are represented as ; Represents the total number of training samples corresponding to all training samples. The column vector formed; Therefore, the bias term is obtained by solving. and Lagrange multipliers Subsequently, a soft temperature measurement model was developed to predict the hydrogen temperature at the outlet of the electric heater: (47) in, Indicates based on input samples The predicted value of hydrogen temperature at the outlet of the electric heater was obtained.

[0108] In some preferred embodiments, the joint optimization process of model parameters and kernel weights for the temperature soft sensing model includes: Cross-validation, grid search, particle swarm optimization, genetic algorithms, or Bayesian optimization methods are used to jointly optimize model parameters, kernel width parameters, sample weight parameters, and kernel function combination weights. Parameters to be optimized include: regularization parameters. Operating condition variable kernel width Time lag kernel width thermal state transfer kernel width Sampling frequency, structure, kernel width Recentness decay parameter Operating condition similarity attenuation parameters Sampling frequency reliability attenuation parameter Thermal state transfer consistency attenuation parameter Historical window length Lag order and kernel function combined weights , , , , , .

[0109] Preferably, the parameter optimization objective is to minimize the prediction error of the hydrogen temperature at the electric heater outlet on the validation set; the joint optimization objective function for model parameters and kernel weights is: (48) in, This represents the prediction error on the validation set. Indicates the first The weights are combined using kernel functions. The variance representing the sample reliability weights and These represent the regularization coefficients.

[0110] Therefore, the above joint optimization method enables the model to adaptively adjust the contribution of the operating condition variable kernel, time lag kernel, thermal state transfer kernel, sampling frequency structure kernel and their interaction kernel to temperature prediction according to the data characteristics of different operating stages.

[0111] Step S4: Based on the real-time operating data of the heat exchanger, the electric heater, and the connecting pipeline, an asynchronous multi-frequency time-delay coupling enhanced real-time input vector is constructed, and the predicted value of the hydrogen temperature at the outlet of the electric heater at the next moment is output through the temperature soft measurement model.

[0112] Specifically, during the online operation phase, multi-source operational data and their original timestamps for the pure hydrogen thermal hydrogen cascade heating process are collected in real time, including real-time operational data of the heat exchanger with timestamps, real-time operational data of the electric heater, and real-time operational data of the connecting pipelines. Based on the multi-source real-time operational dataset, an asynchronous multi-frequency time-delay coupling enhanced real-time input vector for the current moment is constructed in the same manner as the aforementioned steps. After that, the aforementioned Input the trained temperature soft sensing model to obtain the predicted hydrogen temperature at the electric heater outlet at the next moment. Meanwhile, once the measured value of the hydrogen temperature at the outlet of the new electric heater is entered into the database, the measured value and its corresponding multi-source operating data, original timestamp, asynchronous multi-frequency sampling characteristics, process variable statistical characteristics, front-to-back stage thermal state transfer characteristics, and time lag characteristics are combined to form a new sample and added to the sample library.

[0113] Furthermore, model update triggering conditions are set based on sampling structure drift, thermal state transfer deviation, and online prediction error. A model update is triggered if at least one of the following conditions is met: (49) in, This represents the sampling structure drift index of the current sample. This indicates the thermal state transfer deviation of the current sample. This indicates the predicted hydrogen temperature at the outlet of the electric heater at the next moment. This indicates the measured value of the hydrogen temperature at the outlet of the electric heater at the next moment. , and These represent the sampling structure drift threshold, the thermal state propagation deviation threshold, and the prediction error threshold, respectively.

[0114] Furthermore, when a model update is triggered, the sample library, kernel function combination weights, sample reliability weights, and model parameters are updated using sliding window updates, incremental updates, or periodic retraining methods, so that the model can adapt to changes in the sampling structure, thermal state transfer relationships, operating condition switching, and long-term operation drift of the pure hydrogen-hot hydrogen cascade heating process.

[0115] Therefore, this embodiment can construct a feature system that reflects the sampling structure, operating state, thermal state transfer between upstream and downstream stages, and variable-level hysteresis effect while retaining the original timestamps and asynchronous sampling relationships of multi-source operating variables. Furthermore, it introduces reliability constraints and adaptive kernel function mechanisms into a modeling framework suitable for small and medium-sized sample industrial data, thereby achieving accurate and stable online prediction of the hydrogen temperature at the outlet of the electric heater. This provides technical support for the operation monitoring, power regulation, and optimized control of the pure hydrogen heating process.

[0116] It is understood that the above embodiments are only for ease of understanding and simplification of description, and should not be construed as limiting the present invention. The present invention does not specifically limit the type of multi-source running dataset, the construction method of the temperature soft measurement model, the calculation method, etc.

[0117] Therefore, the embodiments of the present invention, on the one hand, can achieve stable prediction of the hydrogen temperature at the outlet of the electric heater under complex conditions such as inconsistent sampling frequencies of multiple source variables, cascade coupling between the heat exchanger and the electric heater, changes in the transmission status of connecting pipelines, and lag in the dynamic response of hot hydrogen. On the other hand, compared with methods relying on manual experience adjustment, static mechanism models, or ordinary data-driven models, the present invention can make fuller use of multi-source asynchronous data in the field, and more accurately characterize the dynamic relationship between preheating, pipeline transmission, and post-heating, providing reliable model support for electric heating power adjustment, online monitoring of outlet temperature, early warning of temperature anomalies, and optimized control of the hot hydrogen heating process. Furthermore, the present invention is more suitable for engineering scenarios with limited operating samples, frequent changes in operating conditions, and high requirements for real-time online prediction in pure hydrogen metallurgical plants, providing engineering-feasible technical means for the stable operation, energy efficiency improvement, and intelligent control of hot hydrogen heating systems in pure hydrogen metallurgy.

[0118] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An asynchronous multi-frequency thermal hydrogen temperature soft measurement method for pure hydrogen metallurgy, characterized in that, include: Based on pre-collected heat exchanger operation data, electric heater operation data, and connecting pipeline operation data, a time-stamped multi-source operation dataset is constructed. Based on the multi-source running dataset, asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are constructed respectively, and combined to obtain the asynchronous multi-frequency time lag coupling enhanced input vector; Based on the aforementioned asynchronous multi-frequency time-delay coupling enhanced input vector, a training sample set and a reliability-constrained adaptive composite kernel function are constructed. A soft temperature measurement model is established based on the aforementioned reliability-constrained adaptive composite kernel function, the aforementioned training sample set, and the sample reliability weights. Based on the real-time operating data of the heat exchanger, the electric heater, and the connecting pipeline, an asynchronous multi-frequency time-delay coupled enhanced real-time input vector is constructed, and the predicted value of the hydrogen temperature at the outlet of the electric heater at the next moment is obtained through the temperature soft measurement model.

2. The method according to claim 1, characterized in that, The heat exchanger operating data includes at least the following operating variables: heat exchanger inlet hydrogen temperature, heat exchanger inlet hydrogen pressure, heat exchanger inlet hydrogen flow rate, heat exchanger outlet hydrogen temperature, heat exchanger outlet hydrogen pressure, and heat exchanger valve opening. The operating data of the electric heater includes at least the following operating variables: electric heating power, electric heater inlet hydrogen temperature, electric heater inlet hydrogen pressure, electric heater outlet hydrogen temperature, and electric heater outlet hydrogen pressure. The operating data of the connecting pipeline includes at least the following operating variables: connecting pipeline temperature and connecting pipeline pressure.

3. The method according to claim 2, characterized in that, Based on the multi-source operational dataset, asynchronous multi-frequency sampling structure features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are constructed, including: Based on the target prediction time of the hydrogen temperature at the outlet of the electric heater, the original sampling sequence of each operating variable in its respective historical period is extracted from the multi-source operating dataset to form the historical dynamic sample window corresponding to each operating variable; based on each of the historical dynamic sample windows, a multi-source asynchronous sampling sample is constructed. Based on the multi-source asynchronous sampling samples, asynchronous multi-frequency sampling structural features, process variable statistical state features, front-to-back stage thermal state transfer features, and variable-level time lag features are constructed respectively.

4. The method according to claim 3, characterized in that, The features for constructing the asynchronous multi-frequency sampling structure include: Based on the multi-source asynchronous sampling samples, obtain the valid sampling timestamps of each running variable within the corresponding historical dynamic sample window, and calculate the sampling time interval of each running variable based on adjacent valid sampling timestamps; Based on the sampling time intervals, the sampling frequency characteristics of each running variable are constructed, and the sampling frequency characteristics are combined to obtain the asynchronous multi-frequency sampling structure characteristics; The sampling frequency feature is calculated according to the following formula: ； in, Indicates the first The running variable in the first... Sampling frequency characteristics within a sample window; Indicates the first The running variable in the first... The average sampling interval within each sample window; Indicates the first The running variable in the first... The maximum sampling interval within a sample window; Indicates the first The running variable in the first... Standard deviation of the sampling interval within each sample window; Indicates the first The running variable in the first... The number of valid sampling points within a sample window; Indicates the first Each running variable relative to the target prediction time Data freshness.

5. The method according to claim 3, characterized in that, Constructing the statistical state characteristics of the process variables includes: Based on the multi-source asynchronous sampling samples, at least one of the following is statistically analyzed for each operating variable within the corresponding historical dynamic sample window: window mean, window maximum, window minimum, window standard deviation, window first-to-last difference, window change slope, time-weighted mean, and time-weighted rate of change; the statistical results of each operating variable are combined to obtain the statistical state characteristics of the process variable. The time-weighted average is calculated according to the following formula: ； in, Indicates the first The running variable in the first... Time-weighted mean within each sample window; Indicates the first Within the sample window, the first The running variable in the first... The value at each valid sampling time. This represents the time interval corresponding to the r-th sampling point; The time-weighted rate of change is calculated according to the following formula: ； in, Indicates the first The running variable in the first... Time-weighted rate of change within each sample window; Indicates the first Within the sample window, the first The first valid sampled value of each running variable; Indicates the first Within the sample window, the first The last valid sampled value of each running variable; Indicates the first Within the sample window, the first The first valid sampling time of each running variable, Indicates the first Within the sample window, the first The last valid sampling time of each running variable.

6. The method according to claim 3, characterized in that, Constructing the preceding and following stage thermal state transfer characteristics includes: Based on the multi-source asynchronous sampling samples, at least one of the following is calculated: heat exchanger state change characteristics, heat exchanger to electric heater transfer characteristics, electric heater historical temperature rise characteristics, valve regulation characteristics, and connecting pipeline transfer characteristics. The calculation results are then combined to obtain the front and rear stage heat state transfer characteristics. The heat exchanger state change characteristics are calculated based on the temperature rise amplitude and pressure loss on the hydrogen side of the heat exchanger; the heat exchanger-to-electric heater transfer characteristics are calculated based on the temperature and pressure transfer differences with hysteresis compensation between the heat exchanger outlet and the electric heater inlet; the electric heater historical temperature rise characteristics are calculated based on the historical temperature rise amplitude, pressure change, and historical unit electric heating power temperature rise efficiency of the electric heater; the valve regulation characteristics are calculated based on the change in the valve opening of the heat exchanger; and the connecting pipeline transfer characteristics are calculated based on the transfer difference between the connecting pipeline temperature and the heat exchanger outlet temperature, and the transfer difference between the electric heater inlet temperature and the connecting pipeline temperature.

7. The method according to claim 3, characterized in that, Constructing the variable-level time lag feature includes: Based on the multi-source asynchronous sampling samples, the lag time of each operating variable relative to the hydrogen temperature at the outlet of the electric heater is determined; Based on the lag time, construct variable-level lag characteristics for each operating variable at the target lag time, and construct multi-order lag characteristics based on the dynamic change characteristics of each operating variable. Valid sampling points of each operating variable are selected in the neighborhood of the target lag time, and the estimated value of the target lag time is calculated based on each valid sampling point using a time kernel weighted aggregation method, so as to obtain the time kernel weighted estimation features of each operating variable. The variable-level time lag features are obtained by combining the variable-level lag features, multi-order lag features, and time kernel weighted estimation features of each operating variable.

8. The method according to any one of claims 1-7, characterized in that, The construction of the reliability-constrained adaptive composite kernel function includes: A training sample set is constructed based on the asynchronous multi-frequency time-delay coupling enhanced input vector and its corresponding measured value of hydrogen temperature at the outlet of the electric heater. Based on the training sample set, the predefined operating condition variable kernel, time lag kernel, thermal state transfer kernel, sampling frequency structure kernel, and interaction kernel are weighted and combined, and after being corrected by the sampling reliability coefficient, the reliability constraint adaptive composite kernel function is obtained.

9. The method according to claim 8, characterized in that, The establishment of the temperature soft measurement model includes: weighting the kernel similarity of each training sample in the training sample set through the reliability constraint adaptive composite kernel function, and weighting the prediction error penalty intensity of each training sample through the sample reliability weight. The sample reliability weight is composed of time recentity weight, operating condition similarity weight, sampling frequency reliability weight, and thermal state transmission consistency weight.

10. The method according to claim 9, characterized in that, The temperature soft measurement model is implemented based on a double-weighted least squares support vector machine, and the optimization objective of the temperature soft measurement model is expressed in the following form: ； The constraints of the temperature soft sensing model are expressed in the following form: ； in, Represents the weight vector; Indicates the bias term; Represents the modeling error variable In the The modeling error value for each training sample; Indicates the first Reliability weights for each training sample; Represents the regularization parameter; Indicates the first Measured hydrogen temperature at the outlet of the electric heater for each training sample; Represents a nonlinear mapping function. Indicates the number of training samples; Indicates the first Asynchronous multi-frequency time-delay coupling enhanced input vector for training samples.

Images