Dynamic quantitative feeding control system and method based on real-time working condition feedback of reaction kettle

Abstract

The invention discloses a dynamic quantitative feeding control system and method based on real-time working condition feedback of a reaction kettle, and the method comprises the steps: collecting the temperature, pressure, flow and other data of the reaction kettle, carrying out the denoising and standardization, and generating a working condition input data set; calculating heat release intensity, an energy residual error and a material residual error based on the working condition input data set, and constructing a physical feature vector; inputting the physical characteristics into an improved echo state network, and outputting temperature and pressure prediction; according to a prediction result and equipment limitation, temperature, pressure and rate boundary conditions are constructed to generate a security constraint set; self-evolution waveform optimization is adopted based on safety constraints, a candidate waveform set is generated, and an optimal feeding curve is selected; and inputting the optimal waveform to control feeding, establishing a trust propagation graph, calculating score gating and triggering safety interlocking. According to the invention, accurate prediction and safety control of reaction kettle feeding are realized by improving the echo state network and optimizing the self-evolution waveform.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control and process control technology, and in particular to a dynamic quantitative feeding control system and method based on real-time operating condition feedback of a reactor. Background Technology

[0002] With the rapid development of the hydrogen energy industry, hydrogen production equipment, hydrogen fuel cell systems, and hydrogen refueling stations are increasingly widely used in industry, transportation, and energy storage. Hydrogen production and reaction processes generally involve high pressure, high temperature, and complex catalytic reaction systems. One of the core pieces of equipment—the reactor—plays a crucial role in the synthesis of hydrogen production materials, the preparation of hydrogen storage media, and the production of electrode active materials. However, in the hydrogen production reaction, the reaction exothermics are intense, and the heat conduction coupling is complex; the feed rate directly affects the reaction stability and hydrogen production efficiency. Traditional feed control methods mostly use fixed-rate or proportional-integral-derivative (PID) control, which cannot dynamically adjust according to the real-time operating conditions of the reactor. This leads to overshoot, lag, or fluctuations in temperature and pressure, which in turn affect product purity and equipment safety.

[0003] Currently, while some companies have introduced intelligent sensing and data monitoring methods in the integration of hydrogen energy systems and the manufacturing of hydrogen production equipment, their operational condition modeling remains at the static control level, making it difficult to accurately capture the nonlinear characteristics of the reaction system under different loads and thermodynamic states. Especially in catalytic hydrogenation and high-temperature decomposition hydrogen production processes, the feedstock conversion rate and exothermic characteristics exhibit significant time variations, and traditional models cannot provide forward-looking predictions and adaptive adjustments for such dynamic processes. Existing safety interlocking mechanisms are typically based on threshold judgments, resulting in delayed responses and a lack of intelligent discrimination capabilities. This can easily lead to an imbalance between false alarms and missed alarms, affecting the continuity and reliability of hydrogen production equipment.

[0004] Therefore, how to provide a dynamic quantitative feeding control system and method based on real-time operating condition feedback of the reactor is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a dynamic quantitative feeding control system and method based on real-time operating condition feedback of a reactor. This invention comprehensively utilizes intelligent control technologies such as improved echo state network modeling, dynamic fusion of physical features, multi-objective self-evolutionary waveform optimization, and trust propagation closed-loop control. Addressing the problems of difficulty in ensuring feeding accuracy, parameter lag, and slow response of safety interlocks under complex operating conditions, it constructs a self-learning, adaptive, and predictable dynamic control system. This invention generates the optimal feeding waveform online by modeling the temperature, pressure, and energy characteristics of the reactor in real time, and achieves safety gating and model self-calibration through a trust propagation mechanism during execution, thereby realizing high-precision adjustment and active safety control of the feeding process. This invention possesses advantages such as high predictive accuracy, fast control response, strong safety, and superior system self-learning capabilities, improving the stability of reactor operation and the automation level of the production process.

[0006] The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to embodiments of the present invention includes: Collect real-time operating data of the reactor, preprocess the real-time operating data, and form an operating condition input dataset; Based on the operating condition input dataset, according to the relationship between energy conservation and mass conservation, the reaction exothermic intensity, energy balance residual and material balance residual are calculated and added to the operating condition input dataset to construct a physical feature vector containing the exothermic intensity index, energy residual and material residual. Using physical feature vectors as input, an improved echo state network model is established. Multi-timescale operating condition modeling is achieved through the state interaction between the fast and slow reservoirs. The memory weights are dynamically adjusted using the cross-layer migration matrix driven by energy balance error, and the temperature and pressure prediction results are output. Based on the temperature and pressure prediction results, and combined with the upper and lower limits of the allowable temperature, pressure and rate of change of the equipment, safety constraints are constructed and a set of safety constraints is generated. Based on the safety constraint set and temperature and pressure prediction results, a multi-objective self-evolution waveform generation mechanism is adopted to parameterize the feeding waveform, generate a candidate waveform set, and optimize the candidate waveforms through morphological variation operator, constraint correction operator and smoothness preservation operator, calculate the comprehensive cost function, and select the optimal feeding waveform. The optimal feeding waveform is input into the control execution unit to control the variable frequency pump or electric valve to perform dynamic quantitative feeding operation. A trust propagation diagram of the cognitive decision layer, sensor layer and actuator layer is constructed. The comprehensive trust score is calculated and gating control is performed. When the comprehensive trust score is lower than the threshold, the safety interlock is triggered and the echo state network model parameters are updated based on the feedback data.

[0007] Optionally, the real-time operating data includes the temperature, pressure, liquid level, stirring speed, feed flow rate, jacket temperature, and feed temperature inside the reactor.

[0008] Optionally, the preprocessing of real-time operating data includes synchronous sampling, noise reduction filtering and standardization of real-time operating data, removal of abnormal data and interpolation compensation for missing data.

[0009] Optionally, the construction of a physical feature vector comprising an exothermic intensity index, energy residual, and material residual includes: Read temperature, pressure, liquid level, feed flow rate, jacket temperature and feed temperature from the working condition input dataset by timestamp; The temperature change rate and liquid level change rate are determined based on the difference between samples at adjacent time points and the sampling time interval. The total amount of material in the vessel is determined using the liquid level, the equivalent cross-sectional area of ​​the vessel body, and the material density. The feed mass flow rate is determined using the feed flow rate and the feed density. The material accumulation rate is determined by the change in the total amount of material at adjacent time points and the sampling time interval. Based on the principle of heat conservation, the reaction exothermic intensity is obtained by subtracting the enthalpy flow carried by the feed from the sum of the heat change of the contents of the vessel and the heat exchange between the vessel and the jacket. The heat change is determined by the total amount of material in the vessel, the specific heat at constant pressure, and the rate of temperature change. The heat exchange is determined by the heat transfer coefficient, the heat transfer area, and the difference between the temperature inside the vessel and the temperature in the jacket. The enthalpy flow is determined by the feed mass flow rate, the specific heat at constant pressure of the feed, and the difference between the feed temperature and the temperature inside the vessel. Based on the principle of mass conservation, the material balance residual is obtained by the difference between the material accumulation rate and the feed mass flow rate. Based on the principle of heat conservation, the energy balance residual is determined by the difference between the sum of the heat change of the contents of the vessel and the heat transfer, minus the heat carried by the feed, and the intensity of the exothermic reaction. Align the reaction exothermic intensity, energy balance residual, and material balance residual with their corresponding timestamps to form a physical feature vector that includes the exothermic intensity index, energy residual, and material residual.

[0010] Optionally, the output temperature and pressure prediction results include: Read the working condition input dataset and physical feature vector according to the timestamp, and concatenate the two to form a comprehensive feature vector; An improved echo state network model is established, which consists of a fast dynamic expression layer, a slow time-domain memory layer, and a cross-layer memory transfer-readout fusion layer. Initial values ​​are set for the state dimension, leakage rate, spectral radius, and inter-layer coupling weight of each layer. The comprehensive feature vector is input into the fast dynamic expression layer. An impact suppression buffer is set to limit the amplitude and slope of high-frequency mutations. A residual sensitive gate is set to trigger or suppress the state update of the fast dynamic expression layer according to the energy balance residual, and the fast dynamic state is output. The comprehensive feature vector and the fast dynamic state are input into the slow time domain memory layer, multi-resolution sliding aggregation is performed to form trend features, and the memory retention time is controlled based on the material balance residual scheduling adaptive forgetting factor. Change point detection is set to refresh part of the memory window when a sudden change in the working condition is detected, and the slow time domain memory state is output. Within the cross-layer memory migration-readout fusion layer, cross-layer migration weights are calculated based on energy balance residuals and material balance residuals. Memory migration is performed between the fast dynamic state and the slow time-domain memory state to form a fusion state. Cross-layer consistency verification and rollback mechanisms are executed, and the temperature and pressure prediction values ​​for the next moment are output. Starting with the predicted temperature and pressure values, multi-step rolling predictions are performed, and the predicted temperature and pressure results for multiple future time steps are generated step by step according to the prediction step size.

[0011] Optionally, generating the set of security constraints includes: Obtain the temperature and pressure prediction results and uncertainty range, and read the sampling period and prediction step size; Based on the safe operating range of the equipment, set the lower temperature limit, upper temperature limit, lower pressure limit, and upper pressure limit. Check the temperature and pressure values ​​at each predicted time point by point. If the predicted value exceeds the set upper and lower limits, record the time as the boundary risk point and generate the corresponding temperature and pressure boundary constraints. Based on the upper limits of the allowable temperature change rate and pressure change rate, calculate the temperature change rate and pressure change rate between adjacent prediction times, determine whether they are within the allowable range, and if they exceed the allowable rate, record the time as the rate risk point and generate the corresponding rate constraint conditions. Based on the uncertainty information of the prediction results, set the temperature uncertainty amplification factor and the pressure uncertainty amplification factor, add the corresponding safety margin to the predicted temperature and pressure values ​​and then perform boundary checks again. Set thresholds and check the energy balance residuals and material balance residuals at each time step. If the two residuals exceed the thresholds, they are marked as physical anomalies, and robustness constraints and physical constraints are generated. The temperature and pressure boundary constraints, rate constraints, robustness constraints, and physical constraints are integrated to form a set of safety constraints.

[0012] Optionally, the calculation of the comprehensive cost function and selection of the optimal feeding waveform includes: The temperature and pressure prediction results, uncertainty information, and safety constraint set are obtained as input conditions for waveform generation and verification. A waveform representation method for a multi-objective self-evolutionary waveform generation mechanism is established. A morphological syntax library consisting of segmented rising segments, plateau segments, falling segments, S-shaped segments, and micro-pulse segments is used to parameterize the feeding waveform. A feeding quota scheduler is set up to allocate and recover the feeding quota of each segment so that the total feeding amount meets the allowable deviation of the target feeding amount. Perform a two-channel evolution process of self-evolution operators on the candidate waveform set: The shape channel is used to add, delete, and rearrange shape segments, and switch segment types; the parameter channel is used to adjust the amplitude, duration, and slope of each segment. The morphological mutation operator, constraint correction operator, and smoothness preservation operator are applied in sequence, and a safety margin embedding operator is introduced to automatically reduce the local slope or insert a buffer segment when the prediction is close to the boundary. A falsification backoff operator is introduced to eliminate candidate waveforms that do not meet the set of safety constraints in advance, and an intra-segment micro-pulse shaping operator is introduced to refine the micro-pulse duty cycle and interval; For each candidate waveform after evolution, the improved echo state network model is called to perform look-ahead evaluation to obtain the temperature and pressure response trajectory within the prediction step. The model is checked time by time according to the set of safety constraints. Candidate waveforms that violate the constraints are directly eliminated, and the comprehensive cost is calculated for candidate waveforms that meet the constraints. Among the candidate waveforms that meet the safety constraints, the one with the lowest overall cost is selected as the optimal feeding waveform. When there are ties for the best, the one with the lower peak temperature and the smoother flow rate change is selected first, and the optimal feeding waveform is output.

[0013] Optionally, the construction of the trust propagation graph of the cognitive decision layer, sensor layer, and actuator layer, the calculation of the comprehensive trust score, and the gating control include: The optimal feeding waveform is received and sent to the control execution unit to establish a trust propagation graph. The trust propagation graph consists of nodes of the cognitive decision layer, each measuring point node of the sensor layer, and each execution mechanism node of the actuator layer. Directed relation edges are set between the nodes to transmit and attenuate the trust degree. During the execution of the optimal feeding waveform, real-time feedback data is collected, and the predicted temperature and pressure values, energy balance residuals and material balance residuals at the corresponding times are read. Based on the consistency between the actual values ​​and the predicted values, and the closeness between the residuals and the thresholds, the observation consistency index of each node is generated and written into the trust propagation graph. Update the relation edge weights and node trust levels in the trust propagation graph based on the observation consistency index, calculate the current comprehensive trust score, and set a first gate threshold and a second interlocking threshold, with the second interlocking threshold being less than the first gate threshold; Gating control is executed based on the comprehensive trust score, the first gating threshold, and the second interlocking threshold. When the overall trust score is lower than the first gate threshold and not lower than the second interlocking threshold, it switches to conservative feeding mode to limit the upper limit of feeding flow and the rate of flow change. When the overall trust score is lower than the second interlock threshold, the safety interlock is triggered, the feeding is immediately stopped, and the cooling device and the pressure relief device are activated in a preset sequence. The control commands after gate control or interlocking are sent to the variable frequency pump and electric valve, and the corresponding node and constraint trigger information are recorded. Based on real-time feedback data and gating execution results, the improved echo state network model is updated online, including recursive correction of readout parameters and adjustment of cross-layer memory migration configuration.

[0014] The dynamic quantitative feeding control system based on real-time operating condition feedback of the reactor according to an embodiment of the present invention includes the following modules: The data acquisition module is used to collect real-time operating data of the reactor and preprocess it to generate an operating condition input dataset. The physical feature construction module is used to calculate the reaction exothermic intensity, energy residual, and material residual based on the operating condition input dataset, and form a physical feature vector. The modeling and prediction module is used to build an improved echo state network model, realize multi-timescale modeling through state interaction, and output temperature and pressure prediction results. The safety constraint construction module is used to set upper and lower limits for temperature, pressure and rate of change based on temperature and pressure prediction results and the equipment's allowable range, and generate a set of safety constraints. The waveform optimization module is used to generate candidate feeding waveforms and select the optimal feeding waveform through morphological variation, constraint correction and smoothness preservation. The trust control module is used to send out the optimal feeding waveform, construct a trust propagation graph, calculate the comprehensive trust score and execute gating control. When the score is lower than the threshold, a safety interlock is triggered.

[0015] The beneficial effects of this invention are: This invention introduces an improved echo state network model to achieve multi-timescale dynamic modeling of complex reactor operating conditions. Utilizing the synergistic effect of fast and slow reservoirs and the cross-layer migration matrix, it not only captures the rapid fluctuations in parameters such as temperature and pressure during the reaction process but also effectively characterizes the long-term thermodynamic trends of the system, thereby achieving high-precision prediction of reaction dynamics. Compared to traditional static models or single-layer neural network structures, the model of this invention has stronger temporal memory and generalization performance, improving the predictability and stability of the reaction process.

[0016] This invention employs a multi-objective self-evolving waveform generation mechanism to enable adaptive optimization of the feeding strategy under safety constraints. This mechanism comprehensively considers multiple factors such as temperature and pressure change rates, and actuator constraints. Through self-evolving operators including morphological variation, constraint correction, and smoothness preservation, it dynamically optimizes the feeding waveform, thereby achieving a balance between stable feeding and energy release in the reaction process while meeting the total quantitative requirements. Compared to traditional fixed waveforms or manually set strategies, this invention can automatically generate the optimal feeding curve under complex reaction conditions, significantly reducing the risk of temperature and pressure overshoot and improving the controllability and consistency of the reaction.

[0017] This invention introduces a trust-propagation closed-loop control mechanism in the control execution phase. By constructing a trust propagation graph across the cognitive decision-making layer, sensor layer, and actuator layer, it calculates a comprehensive trust score in real time and executes gating and interlocking controls. This mechanism can automatically switch to a conservative feeding mode or trigger safety interlocks when it detects a decrease in system trust or an abnormal trend, achieving proactive safety protection during the feeding process. By dynamically updating the echo state network model parameters based on real-time feedback data, the system possesses self-learning and self-correction capabilities, thereby achieving a comprehensive effect of high precision, strong robustness, and high safety in reactor feeding control. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows the dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor proposed in this invention. Figure 2 This is a schematic diagram of the dynamic quantitative feeding control system based on real-time operating condition feedback of the reactor proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 A dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor includes: Collect real-time operating data of the reactor, preprocess the real-time operating data, and form an operating condition input dataset; Based on the operating condition input dataset, according to the relationship between energy conservation and mass conservation, the reaction exothermic intensity, energy balance residual and material balance residual are calculated and added to the operating condition input dataset to construct a physical feature vector containing the exothermic intensity index, energy residual and material residual. Using physical feature vectors as input, an improved echo state network model is established. Multi-timescale operating condition modeling is achieved through the state interaction between the fast and slow reservoirs. The memory weights are dynamically adjusted using the cross-layer migration matrix driven by energy balance error, and the temperature and pressure prediction results are output. Based on the temperature and pressure prediction results, and combined with the upper and lower limits of the allowable temperature, pressure and rate of change of the equipment, safety constraints are constructed and a set of safety constraints is generated. Based on the safety constraint set and temperature and pressure prediction results, a multi-objective self-evolution waveform generation mechanism is adopted to parameterize the feeding waveform, generate a candidate waveform set, and optimize the candidate waveforms through morphological variation operator, constraint correction operator and smoothness preservation operator, calculate the comprehensive cost function, and select the optimal feeding waveform. The optimal feeding waveform is input into the control execution unit to control the variable frequency pump or electric valve to perform dynamic quantitative feeding operation. A trust propagation diagram of the cognitive decision layer, sensor layer and actuator layer is constructed. The comprehensive trust score is calculated and gating control is performed. When the comprehensive trust score is lower than the threshold, the safety interlock is triggered and the echo state network model parameters are updated based on the feedback data.

[0021] In this embodiment, the real-time operating data includes the temperature, pressure, liquid level, stirring speed, feed flow rate, jacket temperature, and feed temperature inside the reactor.

[0022] In this embodiment, the preprocessing of real-time operating data includes synchronous sampling, noise reduction filtering and standardization of real-time operating data, removal of abnormal data and interpolation compensation for missing data.

[0023] In this embodiment, the construction of a physical feature vector comprising an exothermic intensity index, energy residual, and material residual includes: Read temperature, pressure, liquid level, feed flow rate, jacket temperature and feed temperature from the working condition input dataset by timestamp; The temperature change rate and liquid level change rate are determined based on the difference between samples at adjacent time points and the sampling time interval. The total amount of material in the vessel is determined using the liquid level, the equivalent cross-sectional area of ​​the vessel body, and the material density. The feed mass flow rate is determined using the feed flow rate and the feed density. The material accumulation rate is determined by the change in the total amount of material at adjacent time points and the sampling time interval. Based on the principle of heat conservation, the reaction exothermic intensity is obtained by subtracting the enthalpy flow carried by the feed from the sum of the heat change of the contents of the vessel and the heat exchange between the vessel and the jacket. The heat change is determined by the total amount of material in the vessel, the specific heat at constant pressure, and the rate of temperature change. The heat exchange is determined by the heat transfer coefficient, the heat transfer area, and the difference between the temperature inside the vessel and the temperature in the jacket. The enthalpy flow is determined by the feed mass flow rate, the specific heat at constant pressure of the feed, and the difference between the feed temperature and the temperature inside the vessel. Based on the principle of mass conservation, the material balance residual is obtained by the difference between the material accumulation rate and the feed mass flow rate. Based on the principle of heat conservation, the energy balance residual is determined by the difference between the sum of the heat change of the contents of the vessel and the heat transfer, minus the heat carried by the feed, and the intensity of the exothermic reaction. Align the reaction exothermic intensity, energy balance residual, and material balance residual with their corresponding timestamps to form a physical feature vector that includes the exothermic intensity index, energy residual, and material residual.

[0024] In this embodiment, the output temperature and pressure prediction results include: Read the working condition input dataset and physical feature vector according to the timestamp, and concatenate the two to form a comprehensive feature vector; An improved echo state network model is established, which consists of a fast dynamic expression layer, a slow time-domain memory layer, and a cross-layer memory transfer-readout fusion layer. Initial values ​​are set for the state dimension, leakage rate, spectral radius, and inter-layer coupling weight of each layer. The comprehensive feature vector is input into the fast dynamic expression layer. An impact suppression buffer is set to limit the amplitude and slope of high-frequency mutations. A residual sensitive gate is set to trigger or suppress the state update of the fast dynamic expression layer according to the energy balance residual, and the fast dynamic state is output. The comprehensive feature vector and the fast dynamic state are input into the slow time-domain memory layer, and multi-resolution sliding aggregation is performed to form trend features. An adaptive forgetting factor based on material balance residual scheduling controls the memory retention time. A change point detection is set to refresh part of the memory window when a sudden change in operating conditions is detected, and the slow time-domain memory state is output. Specifically, the multi-resolution sliding aggregation is performed as follows: The input fast dynamic state sequence is segmented and slid aggregated at different time scales to calculate the statistical features within the short-term, medium-term and long-term windows respectively, and to capture the changing trends at different time levels. Within a sliding window at each time scale, local fluctuation features and global change features are fused by weighted averaging and differential accumulation to generate a multi-resolution trend description vector. The aggregation results at different time scales are dynamically weighted and combined according to time weights to form a multi-resolution aggregation output with time-level correlation, which serves as the input feature of the slow temporal memory layer. Within the cross-layer memory migration-readout fusion layer, cross-layer migration weights are calculated based on energy balance residuals and material balance residuals. Memory migration is performed between the fast dynamic state and the slow time-domain memory state to form a fusion state. Cross-layer consistency verification and rollback mechanisms are executed, and the temperature and pressure prediction values ​​for the next moment are output. Starting with the predicted temperature and pressure values, multi-step rolling predictions are performed, and the predicted temperature and pressure results for multiple future time steps are generated step by step according to the prediction step size.

[0025] In this embodiment, generating the set of security constraints includes: Obtain the temperature and pressure prediction results and uncertainty range, and read the sampling period and prediction step size; Based on the safe operating range of the equipment, set the lower temperature limit, upper temperature limit, lower pressure limit, and upper pressure limit. Check the temperature and pressure values ​​at each predicted time point by point. If the predicted value exceeds the set upper and lower limits, record the time as the boundary risk point and generate the corresponding temperature and pressure boundary constraints. Based on the upper limits of the allowable temperature change rate and pressure change rate, calculate the temperature change rate and pressure change rate between adjacent prediction times, determine whether they are within the allowable range, and if they exceed the allowable rate, record the time as the rate risk point and generate the corresponding rate constraint conditions. Based on the uncertainty information of the prediction results, set the temperature uncertainty amplification factor and the pressure uncertainty amplification factor, add the corresponding safety margin to the predicted temperature and pressure values ​​and then perform boundary checks again. Set thresholds and check the energy balance residuals and material balance residuals at each time step. If the two residuals exceed the thresholds, they are marked as physical anomalies, and robustness constraints and physical constraints are generated. The temperature and pressure boundary constraints, rate constraints, robustness constraints, and physical constraints are integrated to form a set of safety constraints.

[0026] In this embodiment, the calculation of the comprehensive cost function and the selection of the optimal feeding waveform include: The temperature and pressure prediction results, uncertainty information, and safety constraint set are obtained as input conditions for waveform generation and verification. A waveform representation method for a multi-objective self-evolutionary waveform generation mechanism is established. A morphological syntax library consisting of segmented rising segments, plateau segments, falling segments, S-shaped segments, and micro-pulse segments is used to parameterize the feeding waveform. A feeding quota scheduler is set up to allocate and recover the feeding quota of each segment so that the total feeding amount meets the allowable deviation of the target feeding amount. Perform a two-channel evolution process of self-evolution operators on the candidate waveform set: The shape channel is used to add, delete, and rearrange shape segments, and switch segment types; the parameter channel is used to adjust the amplitude, duration, and slope of each segment. The morphological mutation operator, constraint correction operator, and smoothness preservation operator are applied in sequence, and a safety margin embedding operator is introduced to automatically reduce the local slope or insert a buffer segment when the prediction is close to the boundary. A falsification backoff operator is introduced to eliminate candidate waveforms that do not meet the set of safety constraints in advance, and an intra-segment micro-pulse shaping operator is introduced to refine the micro-pulse duty cycle and interval; For each candidate waveform after evolution, an improved echo state network model is invoked for look-ahead evaluation to obtain the temperature and pressure response trajectories within the prediction step size. The trajectories are then checked time-by-time according to the set of safety constraints. Candidate waveforms that violate the constraints are directly eliminated, while the comprehensive cost is calculated for candidate waveforms that satisfy the constraints. Specifically, the calculation of the comprehensive cost for candidate waveforms that satisfy the constraints involves: Based on the sum of squares of the deviations of the predicted trajectories for temperature and pressure, the stability cost under operating conditions is calculated to measure the steady-state control performance of the waveform during the dynamic response process. Based on the variation in feeding rate and waveform smoothness, the smoothing cost is calculated to evaluate the controllability of the waveform and the risk of equipment wear during the execution phase. By combining reaction completion time, energy utilization rate and safety margin factor, the comprehensive economic cost is calculated. The three cost items are normalized and weighted and summed according to their weights to obtain the comprehensive cost of the candidate waveform, which is used to guide the selection of the optimal waveform. Among the candidate waveforms that meet the safety constraints, the one with the lowest overall cost is selected as the optimal feeding waveform. When there are ties for the best, the one with the lower peak temperature and the smoother flow rate change is selected first, and the optimal feeding waveform is output.

[0027] In this embodiment, the construction of the trust propagation graph of the cognitive decision-making layer, sensor layer, and actuator layer, the calculation of the comprehensive trust score, and the gating control include: The optimal feeding waveform is received and sent to the control execution unit to establish a trust propagation graph. The trust propagation graph consists of nodes of the cognitive decision layer, each measuring point node of the sensor layer, and each execution mechanism node of the actuator layer. Directed relation edges are set between the nodes to transmit and attenuate the trust degree. During the execution of the optimal feeding waveform, real-time feedback data is collected, and the predicted temperature and pressure values, energy balance residuals and material balance residuals at the corresponding time are read. Based on the consistency between the actual value and the predicted value, and the closeness between the residual and the threshold, the observation consistency index of each node is generated and written into the trust propagation diagram. The real-time feedback data includes temperature, pressure, feeding flow rate and valve opening. The relationship edge weights and node trust levels in the trust propagation graph are updated based on the observation consistency index. The current comprehensive trust score is calculated, and a first gating threshold and a second interlocking threshold are set, wherein the second interlocking threshold is less than the first gating threshold. The calculation of the current comprehensive trust score is specifically as follows: The real-time state confidence of nodes in the integrated sensor layer, actuator layer, and cognitive decision layer is used to calculate local trust scores based on signal consistency and delay deviation between nodes. In the trust propagation graph, the trust scores of each node are weighted, accumulated and normalized along the information transmission path to obtain the global trust vector after inter-layer propagation. The weighted average of each node in the global trust vector is used as the comprehensive trust score output to reflect the overall consistency and operational reliability of the system. Gating control is executed based on the comprehensive trust score, the first gating threshold, and the second interlocking threshold. When the overall trust score is lower than the first gate threshold and not lower than the second interlocking threshold, it switches to conservative feeding mode to limit the upper limit of feeding flow and the rate of flow change. When the overall trust score is lower than the second interlock threshold, the safety interlock is triggered, the feeding is immediately stopped, and the cooling device and the pressure relief device are activated in a preset sequence. The control commands after gate control or interlocking are sent to the variable frequency pump and electric valve, and the corresponding node and constraint trigger information are recorded. Based on real-time feedback data and gating execution results, the improved echo state network model is updated online, including recursive correction of readout parameters and adjustment of cross-layer memory migration configuration.

[0028] refer to Figure 2 The dynamic quantitative feeding control system based on real-time operating condition feedback of the reactor includes the following modules: The data acquisition module is used to collect real-time operating data of the reactor and preprocess it to generate an operating condition input dataset. The physical feature construction module is used to calculate the reaction exothermic intensity, energy residual, and material residual based on the operating condition input dataset, and form a physical feature vector. The modeling and prediction module is used to build an improved echo state network model, realize multi-timescale modeling through state interaction, and output temperature and pressure prediction results. The safety constraint construction module is used to set upper and lower limits for temperature, pressure and rate of change based on temperature and pressure prediction results and the equipment's allowable range, and generate a set of safety constraints. The waveform optimization module is used to generate candidate feeding waveforms and select the optimal feeding waveform through morphological variation, constraint correction and smoothness preservation. The trust control module is used to send out the optimal feeding waveform, construct a trust propagation graph, calculate the comprehensive trust score and execute gating control. When the score is lower than the threshold, a safety interlock is triggered.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a liquid organic hydrogen storage material preparation system of a hydrogen energy equipment manufacturing company. The effective volume of the reactor in this system is 2.5 m³, used for catalytic hydrogenation reaction, mainly producing hydrogen-containing energy carrier material (LOHC). This process has typical strong exothermic characteristics, with extremely high requirements for feed rate and thermal management. Traditional PID control systems often fail to respond promptly to changes in reaction rate, leading to temperature overshoot, pressure lag, and uneven heat release. This can reduce hydrogen production efficiency or even trigger safety interlock shutdowns, affecting the continuous operation of the production line.

[0030] In this hydrogen production system, the dynamic quantitative feeding control method proposed in this invention is adopted. The system configuration includes a data acquisition module, a physical feature construction module, a modeling and prediction module, a safety constraint construction module, a waveform optimization module, and a trust control module. The reactor collects data in real time using six types of sensors: temperature, pressure, liquid level, flow rate, and stirring speed, with a sampling frequency of 1Hz. The acquired multidimensional signals undergo data denoising and time synchronization by a preprocessing module to form a standardized operating condition input dataset. Subsequently, the physical feature construction module calculates the reaction exothermic intensity, energy balance residual, and material balance residual based on the energy and mass conservation relationships, adding these to the input dataset to construct the reactor's physical feature vector.

[0031] In the modeling phase, the system employs an improved echo state network model, achieving multi-timescale prediction through collaborative modeling of fast and slow memory pools and cross-layer dynamic memory transfer units. This model can dynamically adjust memory weights based on energy balance errors, thereby enhancing prediction accuracy under nonlinear time-varying conditions. Experimental results show that the model's prediction errors for temperature and pressure over the next 30 seconds are controlled within ±1.2℃ and ±0.013MPa, respectively, representing an accuracy improvement of approximately 65% ​​compared to traditional neural networks.

[0032] Subsequently, the safety constraint construction module automatically sets a set of safety constraints based on the prediction results and equipment design limits, including a temperature upper limit of 320℃, a pressure upper limit of 0.65MPa, and a rate of change threshold. The waveform optimization module employs a multi-objective self-evolutionary waveform generation mechanism to generate multiple candidate feeding waveforms and performs evolutionary optimization through morphological variation, constraint correction, and smoothness preservation operators. The system performs look-ahead evaluation and cost calculation for each candidate waveform, ultimately selecting the feeding waveform with the lowest overall cost that satisfies the safety constraints as the control command.

[0033] During the execution phase, the control system transmits the optimal waveform signal to the variable frequency pump, achieving continuous and gradual dynamic quantitative feeding. Simultaneously, the trust control module establishes a trust propagation diagram across the cognitive decision-making layer, sensor layer, and actuator layer, monitoring the consistency and delay of node signals and calculating a comprehensive trust score. When signal drift or sudden changes in operating conditions cause the score to drop below a threshold, the system immediately executes gating control, reducing the feeding rate or temporarily halting operation, and initiating a safety interlock mechanism. Through feedback data, the model automatically updates its parameters, achieving self-learning and closed-loop optimization.

[0034] During a pilot-scale test running continuously for 72 hours, the system completed six full reaction batches. Compared with the results of traditional PID control, the method of this invention significantly improved the stability of temperature and pressure control.

[0035] Table 1 Comparison of Dynamic Quantitative Feeding Control Performance of Reactors

[0036] As shown in Table 1, the dynamic quantitative feeding control performance of the method of this invention in the hydrogen energy production reaction system is significantly better than that of the traditional PID control method. Regarding temperature and pressure control, the method of this invention achieves higher process stability. The average reaction temperature is 316.5℃, which is basically consistent with the 316.2℃ under traditional PID control, but the temperature fluctuation range is significantly reduced from ±6.1℃ to ±2.5℃, a decrease of 59%. Simultaneously, the pressure fluctuation range decreases from ±0.10MPa to ±0.05MPa, a decrease of 50%. This indicates that the improved echo state network model of this invention can effectively predict changes in reaction thermodynamics and adjust the feeding strategy in advance, thereby avoiding overshoot in temperature and pressure and improving the thermal stability and safety of the reaction system.

[0037] In terms of production efficiency and energy efficiency, the method of this invention achieves higher reaction rates and energy utilization without increasing energy consumption. The average feeding rate increased from 4.2 kg / min to 4.3 kg / min, and the total feeding time was shortened from 51 minutes to 46 minutes, resulting in an overall process efficiency improvement of approximately 10%. Energy utilization increased from 83.4% to 87.8%, indicating that the system can maintain high heat recovery and transfer efficiency during dynamic feeding adjustments. Regarding model prediction accuracy, the temperature prediction error decreased from ±3.4℃ to ±1.2℃, and the pressure prediction error decreased from ±0.038 MPa to ±0.013 MPa, representing improvements of 65% and 66% respectively, verifying the high-precision modeling capability of the improved echo state network under dynamic conditions across multiple time scales.

[0038] Finally, in terms of safety and intelligent control, the method of this invention exhibits significant robustness and adaptability. Traditional PID control systems trigger safety interlocks an average of 2.1 times during batch operation, while the method of this invention did not exhibit any false triggers in any batch, fully demonstrating that the trust propagation-based safety control mechanism can dynamically adjust the system response based on node signal consistency and trust scores, fundamentally reducing misjudgments and delays. The system's average trust score reaches 93.1%, demonstrating a high level of signal consistency and decision reliability. In summary, this invention not only improves the feeding accuracy and thermodynamic control stability of the reactor in the hydrogen energy production process but also takes into account energy efficiency, safety, and intelligence levels, possessing broad industrial promotion and application value.

[0039] 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 equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic quantitative feeding control method based on real-time operating condition feedback of a reactor, characterized in that, include: Collect real-time operating data of the reactor, preprocess the real-time operating data, and form an operating condition input dataset; Based on the operating condition input dataset, according to the relationship between energy conservation and mass conservation, the reaction exothermic intensity, energy balance residual and material balance residual are calculated and added to the operating condition input dataset to construct a physical feature vector containing the exothermic intensity index, energy residual and material residual. Using physical feature vectors as input, an improved echo state network model is established. Multi-timescale operating condition modeling is achieved through the state interaction between the fast and slow reservoirs. The memory weights are dynamically adjusted using the cross-layer migration matrix driven by energy balance error, and the temperature and pressure prediction results are output. Based on the temperature and pressure prediction results, and combined with the upper and lower limits of the allowable temperature, pressure and rate of change of the equipment, safety constraints are constructed and a set of safety constraints is generated. Based on the safety constraint set and temperature and pressure prediction results, a multi-objective self-evolution waveform generation mechanism is adopted to parameterize the feeding waveform, generate a candidate waveform set, and optimize the candidate waveforms through morphological variation operator, constraint correction operator and smoothness preservation operator, calculate the comprehensive cost function, and select the optimal feeding waveform. The optimal feeding waveform is input into the control execution unit to control the variable frequency pump or electric valve to perform dynamic quantitative feeding operation. A trust propagation diagram of the cognitive decision layer, sensor layer and actuator layer is constructed. The comprehensive trust score is calculated and gating control is performed. When the comprehensive trust score is lower than the threshold, the safety interlock is triggered and the echo state network model parameters are updated based on the feedback data.

2. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The real-time operating data includes the temperature, pressure, liquid level, stirring speed, feed flow rate, jacket temperature, and feed temperature inside the reactor.

3. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The preprocessing of real-time operating data includes synchronous sampling, noise reduction filtering and standardization of real-time operating data, removal of abnormal data and interpolation compensation for missing data.

4. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The construction of the physical feature vector, which includes the exothermic intensity index, energy residual, and material residual, includes: Read temperature, pressure, liquid level, feed flow rate, jacket temperature and feed temperature from the working condition input dataset by timestamp; The temperature change rate and liquid level change rate are determined based on the difference between samples at adjacent time points and the sampling time interval. The total amount of material in the vessel is determined using the liquid level, the equivalent cross-sectional area of ​​the vessel body, and the material density. The feed mass flow rate is determined using the feed flow rate and the feed density. The material accumulation rate is determined by the change in the total amount of material at adjacent time points and the sampling time interval. Based on the principle of heat conservation, the reaction exothermic intensity is obtained by subtracting the enthalpy flow carried by the feed from the sum of the heat change of the contents of the vessel and the heat exchange between the vessel and the jacket. The heat change is determined by the total amount of material in the vessel, the specific heat at constant pressure, and the rate of temperature change. The heat exchange is determined by the heat transfer coefficient, the heat transfer area, and the difference between the temperature inside the vessel and the temperature in the jacket. The enthalpy flow is determined by the feed mass flow rate, the specific heat at constant pressure of the feed, and the difference between the feed temperature and the temperature inside the vessel. Based on the principle of mass conservation, the material balance residual is obtained by the difference between the material accumulation rate and the feed mass flow rate. Based on the principle of heat conservation, the energy balance residual is determined by the difference between the sum of the heat change of the contents of the vessel and the heat transfer, minus the heat carried by the feed, and the intensity of the exothermic reaction. Align the reaction exothermic intensity, energy balance residual, and material balance residual with their corresponding timestamps to form a physical feature vector that includes the exothermic intensity index, energy residual, and material residual.

5. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The output temperature and pressure prediction results include: Read the working condition input dataset and physical feature vector according to the timestamp, and concatenate the two to form a comprehensive feature vector; An improved echo state network model is established, which consists of a fast dynamic expression layer, a slow time-domain memory layer, and a cross-layer memory transfer-readout fusion layer. Initial values ​​are set for the state dimension, leakage rate, spectral radius, and inter-layer coupling weight of each layer. The comprehensive feature vector is input into the fast dynamic expression layer. An impact suppression buffer is set to limit the amplitude and slope of high-frequency mutations. A residual sensitive gate is set to trigger or suppress the state update of the fast dynamic expression layer according to the energy balance residual, and the fast dynamic state is output. The comprehensive feature vector and the fast dynamic state are input into the slow time domain memory layer, multi-resolution sliding aggregation is performed to form trend features, and the memory retention time is controlled based on the material balance residual scheduling adaptive forgetting factor. Change point detection is set to refresh part of the memory window when a sudden change in the working condition is detected, and the slow time domain memory state is output. Within the cross-layer memory migration-readout fusion layer, cross-layer migration weights are calculated based on energy balance residuals and material balance residuals. Memory migration is performed between the fast dynamic state and the slow time-domain memory state to form a fusion state. Cross-layer consistency verification and rollback mechanisms are executed, and the temperature and pressure prediction values ​​for the next moment are output. Starting with the predicted temperature and pressure values, multi-step rolling predictions are performed, and the predicted temperature and pressure results for multiple future time steps are generated step by step according to the prediction step size.

6. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The generated set of security constraints includes: Obtain the temperature and pressure prediction results and uncertainty range, and read the sampling period and prediction step size; Based on the safe operating range of the equipment, set the lower temperature limit, upper temperature limit, lower pressure limit, and upper pressure limit. Check the temperature and pressure values ​​at each predicted time point by point. If the predicted value exceeds the set upper and lower limits, record the time as the boundary risk point and generate the corresponding temperature and pressure boundary constraints. Based on the upper limits of the allowable temperature change rate and pressure change rate, calculate the temperature change rate and pressure change rate between adjacent prediction times, determine whether they are within the allowable range, and if they exceed the allowable rate, record the time as the rate risk point and generate the corresponding rate constraint conditions. Based on the uncertainty information of the prediction results, set the temperature uncertainty amplification factor and the pressure uncertainty amplification factor, add the corresponding safety margin to the predicted temperature and pressure values ​​and then perform boundary checks again. Set thresholds and check the energy balance residuals and material balance residuals at each time step. If the two residuals exceed the thresholds, they are marked as physical anomalies, and robustness constraints and physical constraints are generated. The temperature and pressure boundary constraints, rate constraints, robustness constraints, and physical constraints are integrated to form a set of safety constraints.

7. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The calculation of the comprehensive cost function and the selection of the optimal feeding waveform include: The temperature and pressure prediction results, uncertainty information, and safety constraint set are obtained as input conditions for waveform generation and verification. A waveform representation method for a multi-objective self-evolutionary waveform generation mechanism is established. A morphological syntax library consisting of segmented rising segments, plateau segments, falling segments, S-shaped segments, and micro-pulse segments is used to parameterize the feeding waveform. A feeding quota scheduler is set up to allocate and recover the feeding quota of each segment so that the total feeding amount meets the allowable deviation of the target feeding amount. Perform a two-channel evolution process of self-evolution operators on the candidate waveform set: The shape channel is used to add, delete, and rearrange shape segments, and switch segment types; the parameter channel is used to adjust the amplitude, duration, and slope of each segment. The morphological mutation operator, constraint correction operator, and smoothness preservation operator are applied in sequence, and a safety margin embedding operator is introduced to automatically reduce the local slope or insert a buffer segment when the prediction is close to the boundary. A falsification backoff operator is introduced to eliminate candidate waveforms that do not meet the set of safety constraints in advance, and an intra-segment micro-pulse shaping operator is introduced to refine the micro-pulse duty cycle and interval; For each candidate waveform after evolution, the improved echo state network model is called to perform look-ahead evaluation to obtain the temperature and pressure response trajectory within the prediction step. The model is checked time by time according to the set of safety constraints. Candidate waveforms that violate the constraints are directly eliminated, and the comprehensive cost is calculated for candidate waveforms that meet the constraints. Among the candidate waveforms that meet the safety constraints, the one with the lowest overall cost is selected as the optimal feeding waveform. When there are ties for the best, the one with the lower peak temperature and the smoother flow rate change is selected first, and the optimal feeding waveform is output.

8. The dynamic quantitative feeding control method based on real-time operating condition feedback of the reactor according to claim 1, characterized in that, The construction of a trust propagation graph comprising the cognitive decision-making layer, sensor layer, and actuator layer, the calculation of a comprehensive trust score, and the implementation of gating control include: The optimal feeding waveform is received and sent to the control execution unit to establish a trust propagation graph. The trust propagation graph consists of nodes of the cognitive decision layer, each measuring point node of the sensor layer, and each execution mechanism node of the actuator layer. Directed relation edges are set between the nodes to transmit and attenuate the trust degree. During the execution of the optimal feeding waveform, real-time feedback data is collected, and the predicted temperature and pressure values, energy balance residuals and material balance residuals at the corresponding times are read. Based on the consistency between the actual values ​​and the predicted values, and the closeness between the residuals and the thresholds, the observation consistency index of each node is generated and written into the trust propagation graph. Update the relation edge weights and node trust levels in the trust propagation graph based on the observation consistency index, calculate the current comprehensive trust score, and set a first gate threshold and a second interlocking threshold, with the second interlocking threshold being less than the first gate threshold; Gating control is executed based on the comprehensive trust score, the first gating threshold, and the second interlocking threshold. When the comprehensive trust score is lower than the first gate threshold and not lower than the second interlocking threshold, it switches to conservative feeding mode to limit the upper limit of feeding flow and the rate of flow change. When the overall trust score is lower than the second interlock threshold, the safety interlock is triggered, the feeding is immediately stopped, and the cooling device and the pressure relief device are activated in a preset sequence. The control commands after gate control or interlocking are sent to the variable frequency pump and electric valve, and the corresponding node and constraint trigger information are recorded. Based on real-time feedback data and gating execution results, the improved echo state network model is updated online, including recursive correction of readout parameters and adjustment of cross-layer memory migration configuration.

9. A dynamic quantitative feeding control system based on real-time operating condition feedback of a reactor, executing the dynamic quantitative feeding control method based on real-time operating condition feedback of a reactor as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition module is used to collect real-time operating data of the reactor and preprocess it to generate an operating condition input dataset. The physical feature construction module is used to calculate the reaction exothermic intensity, energy residual, and material residual based on the operating condition input dataset, and form a physical feature vector. The modeling and prediction module is used to build an improved echo state network model, realize multi-timescale modeling through state interaction, and output temperature and pressure prediction results. The safety constraint construction module is used to set upper and lower limits for temperature, pressure and rate of change based on temperature and pressure prediction results and the equipment's allowable range, and generate a set of safety constraints. The waveform optimization module is used to generate candidate feeding waveforms and optimize them through morphological variation, constraint correction and smoothness preservation to select the optimal feeding waveform; The trust control module is used to send out the optimal feeding waveform, construct a trust propagation graph, calculate the comprehensive trust score and execute gating control. When the score is lower than the threshold, a safety interlock is triggered.

Images