G-c3n4 photocatalytic hydrogen production adaptive control method and system based on multi-modal perception

Abstract

The application provides a g-C3N4 photocatalytic hydrogen production adaptive control method and system based on multi-modal perception, and the method specifically comprises the following steps: constructing a four-dimensional state vector of hydrogen production reaction based on reactor slurry temperature, pH, dissolved hydrogen and catalyst concentration time series data, taking light intensity and stirring speed as algebraic variables, and establishing a segmented state model covering carrier behavior, hydrogen evolution reaction and mass transfer; control intervals are divided according to light intensity, temperature and dissolved hydrogen concentration thresholds, a non-smooth control barrier function is constructed, the neighborhood width and broken line slope of the barrier function are adjusted through the reversal density of the state increment symbol sequence; a constraint control law is constructed based on the adjusted barrier function and the state model, the light source power, stirring frequency and liquid inlet flow control amount are solved, and matching amplitude limiting constraints are applied, the control law is switched or kept according to the index; the model weight is corrected by using decoupling observation variables, and the control input is output to complete the hydrogen production closed-loop control.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to an adaptive control method and system for g-C3N4 photocatalytic hydrogen production based on multimodal sensing. Background Technology

[0002] The photocatalytic hydrogen production process of graphitic carbon nitride (g-C3N4) is a kinetic system involving photon absorption, photogenerated carrier generation and recombination, surface hydrogen evolution reaction, and multiphase mass transfer. In actual reactors, key state parameters such as slurry temperature, pH, dissolved hydrogen concentration, and catalyst concentration change with the reaction progress. Photocatalytic hydrogen production systems often employ an open-loop operation mode with constant light intensity and fixed stirring speed, which cannot adapt to the changing internal state of the reactor in real time. This static operation mode easily leads to ineffective recombination of photogenerated carriers, localized supersaturation of dissolved hydrogen causing inhibitory effects, and impaired gas-liquid-solid three-phase mass transfer, resulting in low overall quantum efficiency and actual hydrogen production rate. Modern control theory can be applied to the optimization of photocatalytic reactor operation. For example, the control method based on the control barrier function (CBF) has advantages in ensuring system state safety constraints and handling boundary conditions. By establishing a state model containing the core reaction variables and combining it with CBF to design constraint control laws, the light source power and stirring rate can be adjusted. However, the photocatalytic reaction is affected by the switching of light intensity thresholds and the escape of dissolved hydrogen bubbles, exhibiting piecewise non-smooth characteristics. Smooth functions cannot adapt to this abrupt mechanism, leading to sluggish or conservative responses in the control system during interval switching. Existing control strategies lack a mechanism to suppress high-frequency oscillations when the system approaches the high hydrogen production rate boundary, and do not utilize the sign-reversal sequence characteristics of state increments to adjust the boundary neighborhood. This results in frequent jumps and crossings of the control input at the boundary, exacerbating actuator wear and disrupting flow field stability. The model fails to achieve decoupled observation of the microscopic characteristics of photocurrent and the macroscopic mass transfer of gas holdup, and does not include a mechanism for real-time correction of carrier recombination and mass transfer weights based on model prediction bias and state return values. Consequently, the obtained control input cannot approach the optimal efficiency of photocatalytic hydrogen production. Summary of the Invention

[0003] According to one aspect of the present invention, an adaptive control method for g-C3N4 photocatalytic hydrogen production based on multimodal sensing is provided, the method comprising: A four-dimensional state vector for hydrogen production reaction is constructed by acquiring time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration. Time-series data of light intensity and stirring speed are acquired as algebraic variables. Based on the state vector and algebraic variables, a segmented state model including carrier generation, recombination, hydrogen evolution reaction, and mass transfer is established. Control intervals are divided according to preset thresholds for light intensity, temperature, and dissolved hydrogen concentration. Non-smooth control barrier functions are constructed for each control interval. The non-smooth control barrier functions use a combination of absolute value terms and quadratic terms as boundary conditions. The slope of the boundary neighborhood broken line is determined based on the historical state difference sign sequence. The sign inversion density of the sign sequence is calculated and mapped to neighborhood width and piecewise slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. Based on the adjusted non-smooth control barrier function and state model, a constraint control law is constructed to obtain the control quantities of light source power, stirring frequency, and liquid inlet flow rate. A limiting constraint consistent with the subgradient of the non-smooth control barrier function is applied to the single-step increment of the control quantity. The control interval control law is selected according to the directional derivative of the non-smooth control barrier function and the inversion density. When the boundary back-and-forth crossing trend is detected, the control structure of the previous control interval is maintained and the control quantity amplitude is updated. By decoupling the observed variables photocurrent and gas holdup, and correcting the carrier recombination and mass transfer weights based on model prediction bias and state return, the control input for the next moment is obtained to complete hydrogen production control.

[0004] According to another aspect of the present invention, an adaptive control system for g-C3N4 photocatalytic hydrogen production based on multimodal sensing is provided, the system comprising the following modules: The determination module is used to acquire time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration and catalyst concentration to construct a four-dimensional state vector of hydrogen production reaction, and acquire time-series data of light intensity and stirring speed as algebraic variables. Based on the state vector and algebraic variables, a segmented state model including carrier generation, recombination, hydrogen evolution reaction and mass transfer is established. The control interval is divided according to the preset thresholds of light intensity, temperature and dissolved hydrogen concentration. A non-smooth control barrier function is constructed for each control interval. The non-smooth control barrier function adopts a combination of absolute value term and quadratic term boundary. The slope of the boundary neighborhood piecewise linear function is determined according to the historical state difference sign sequence. The update module is used to calculate the sign inversion density of the sign sequence and map it to neighborhood width and piecewise slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. Based on the adjusted non-smooth control barrier function and state model, a constraint control law is constructed to obtain the control quantities of light source power, stirring frequency and liquid inlet flow rate. The single-step increment of the control quantity is subject to a limiting constraint consistent with the subgradient of the non-smooth control barrier function. The control interval control law is selected according to the directional derivative of the non-smooth control barrier function and the inversion density. When the boundary back-and-forth crossing trend is detected, the control structure of the previous control interval is maintained and the control quantity amplitude is updated. The control module is used to decouple the observed variables from the photocurrent and the gas holdup, and correct the carrier recombination and mass transfer weights according to the model prediction bias and the state return amount, so as to obtain the control input for the next moment and complete the hydrogen production control.

[0005] This invention establishes a segmented state model encompassing carrier changes and mass transfer processes by extracting multidimensional time-series data, representing the multiphysical and chemically related processes in photocatalytic hydrogen production. By adjusting the boundary neighborhood width and piecewise slope of the non-smooth control barrier function using sign sequence and sign inversion density, high-frequency chattering during boundary crossing is eliminated. Simultaneously, subgradient amplitude limiting constraints are applied to the single-step increment of the control variable, and boundary round-trip crossing trends are used to avoid actuator jumps and losses. By decoupling the observed variables and correcting the core reaction weights based on prediction bias, model mismatch errors are compensated, improving the overall gas production efficiency and long-term operational stability of the photocatalytic system. Attached Figure Description

[0006] Figure 1 The flowchart shows the adaptive control method for g-C3N4 photocatalytic hydrogen production based on multimodal sensing. Figure 2 A schematic diagram of the absolute width variation curve of the boundary neighborhood under strong perturbation; Figure 3 This is a schematic diagram showing the variation curves of the mean square error in the prediction of photocurrent and gas content under strong perturbation. Figure 4 This is a schematic diagram comparing the steady-state deviation curves of hydrogen production under different control strategies. Detailed Implementation

[0007] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0008] One or more embodiments of the present invention provide an adaptive control method for g-C3N4 photocatalytic hydrogen production based on multimodal sensing, such as Figure 1 As shown, the method includes: S1. Obtain time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration to construct a four-dimensional state vector for hydrogen production reaction. Obtain time-series data of light intensity and stirring speed as algebraic variables. Based on the state vector and algebraic variables, establish a segmented state model that includes carrier generation, recombination, hydrogen evolution reaction, and mass transfer. Divide the control interval according to preset thresholds for light intensity, temperature, and dissolved hydrogen concentration. Construct a non-smooth control barrier function for each control interval. The non-smooth control barrier function uses a combination of absolute value terms and quadratic terms as boundary. Determine the slope of the boundary neighborhood broken line based on the historical state difference sign sequence.

[0009] A multi-channel data acquisition card is configured to read analog signals from the PT100 temperature sensor, pH meter, dissolved hydrogen electrode, and UV-Vis spectrophotometer. After high-frequency noise is filtered out using a median filter algorithm, a four-dimensional state vector is generated, containing reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration. Simultaneously, light intensity and stirring speed data are acquired as external algebraic variables using a power meter and a Hall effect speed sensor. In one embodiment, photocurrent and gas holdup are also acquired by embedding a photoelectrochemical working electrode circuit and a multiphase flow optical / ultrasonic detection probe within the reactor system.

[0010] The integral solver module of the SciPy library is used to construct a piecewise state model based on the Langmuir-Hinshelwood dynamics mechanism. This model includes the carrier generation rate, recombination rate, surface hydrogen evolution reaction rate, and gas-liquid mass transfer rate calculated according to the Lambert-Beer law.

[0011] The controller uses the where conditional judgment function of the Numpy library to identify whether the current values ​​of light intensity, temperature and dissolved hydrogen concentration cross the pre-set empirical thresholds, thereby dividing the state space into light-restricted region, mass transfer-restricted region and kinetic-restricted region.

[0012] For each defined control interval, a non-smooth control barrier function is constructed using the SymPy symbolic computation library, which includes the sum of a linear term representing the absolute value of the system state deviation and a quadratic matrix representing the state deviation. The maxima-minus normalization algorithm is used to remove the dimensions of the state variables. A fixed-length historical state sliding window of 10 steps is set. The diff function from the NumPy library is used to calculate the state increment of adjacent steps, and the sign function is called to generate a sequence of state increment signs. The length of consecutive identical signs in this sequence is counted, and the slope of the piecewise linear transition region within the non-smooth boundary is calculated in real time based on a preset linear mapping relationship.

[0013] As one possible implementation, the step of acquiring time-series data on reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration to construct a four-dimensional state vector for the hydrogen production reaction, and acquiring time-series data on light intensity and stirring speed as algebraic variables, includes: The sensor network synchronously collects signals of light intensity, slurry temperature, pH value, dissolved hydrogen concentration, catalyst concentration, and stirring speed in the reactor at a preset sampling frequency. The collected signals are low-pass filtered and normalized to eliminate sensor noise and dimensional differences. The processed slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration are combined in time series to construct a four-dimensional state vector representing the multi-physics field correlation state of the hydrogen production reaction system at the current moment. At the same time, the processed light intensity and stirring speed are extracted as algebraic variables.

[0014] An industrial IoT sensor network, consisting of thermocouples, glass electrode pH meters, microelectrode dissolved hydrogen sensors, photometers, and speed sensors arranged inside and on the surface of the reactor, synchronously and in real time collects various physical quantities at a preset sampling frequency. The collected raw time-series signals are input into a second-order Butterworth digital low-pass filter with a cutoff frequency of 2.5Hz to filter out high-frequency mechanical noise caused by electromagnetic interference and the stirring impeller.

[0015] The Min-Max linear mapping algorithm is used to normalize the filtered data, for example, mapping the actual slurry temperature and pH value to the dimensionless interval [0,1] to eliminate differences in physical dimensions. After data preprocessing, the microprocessor extracts the above four core parameters at each discrete sampling time to generate a four-dimensional state vector. Simultaneously, the normalized light intensity and stirring speed were extracted and defined as algebraic variables. Construct a digital twin dataset that includes system state and boundary conditions.

[0016] As one possible implementation, the segmented state model established based on state vectors and algebraic variables, encompassing carrier generation, recombination, hydrogen evolution reaction, and mass transfer, is used to divide the control interval according to preset thresholds for light intensity, temperature, and dissolved hydrogen concentration, including: A set of differential equations is established based on the semiconductor photocatalysis mechanism. The set of differential equations includes the time derivative of a four-dimensional state vector, the four-dimensional state vector, and a control input vector. Thresholds for light intensity, slurry temperature, and dissolved hydrogen concentration are set. The light intensity in the algebraic variables and the slurry temperature and dissolved hydrogen concentration in the current state vector are compared with the corresponding thresholds. Based on the Boolean logic combination of each variable being higher or lower than the threshold, the global state space is divided into multiple subspaces as control intervals. A corresponding system parameter matrix is ​​matched for each control interval.

[0017] The control system was established by analyzing the band structure and photoelectrochemical reaction kinetics of g-C3N4 material, and the form was as follows: The multidimensional differential equation. Classification thresholds are pre-configured in the controller's memory, preferably set as follows: light intensity threshold of 600 W / m², temperature threshold of 313.15 K, and dissolved hydrogen concentration threshold of 0.05 mmol / L.

[0018] Within the control cycle, the current state is compared with the aforementioned threshold in real time to generate a three-bit binary feature identifier. Based on Boolean logic combinations, the global state space is divided into eight orthogonal control intervals. Once the state trajectory enters a specific interval, the state machine triggers a parameter matching and switching mechanism, retrieving the appropriate local system parameter matrix and empirical coefficients from a static lookup table, thereby solving the distortion problem caused by a single global model.

[0019] As one possible implementation, constructing a non-smooth control barrier function for each control interval, wherein the non-smooth control barrier function uses a combination of absolute value terms and quadratic terms as boundaries, and determines the slope of the boundary neighborhood piecewise linear curve based on the historical state difference sign sequence, includes: Define the deviation vector from the current state of the system to the target equilibrium manifold. Construct a non-smooth control barrier function by multiplying the first weight coefficient with the absolute norm of the deviation vector and the second weight coefficient with the square of the Euclidean norm of the deviation vector. Extract the norm increment of the state deviation over a preset number of sampling periods. Use the sign function to extract the increment direction and generate a state increment sign sequence. Based on the distribution characteristics of positive and negative signs in the sign sequence, set the initial piecewise linear slope parameter of the non-smooth control barrier function in the zero neighborhood.

[0020] After entering the designated control zone, the state estimator calculates the dimensionless system deviation vector between the current state and the target equilibrium point. And construct a composite nonsmooth control barrier function with forced convergence properties. First weighting coefficient Option 1.2 is preferred, utilizing the non-differentiability of the absolute value at the origin to dominate finite-time convergence, with the second weighting coefficient... The preferred value is 0.5, which constrains response overshoot and enhances boundary convexity.

[0021] In addition, the controller sets a sliding data window with a length of 10 steps, extracts the norm increment of continuous state deviation, and inputs the sign function. Generate a binary state increment symbol sequence. Based on the sequence consistency characteristics, adjust the initial piecewise slope parameter of the barrier function in the zero-point neighborhood, such as making it smoother. If it diverges, it becomes steep. .

[0022] S2 calculates the sign inversion density of the sign sequence and maps it to neighborhood width and piecewise slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. Based on the adjusted non-smooth control barrier function and state model, a constraint control law is constructed to obtain the control quantities of light source power, stirring frequency, and liquid inlet flow rate. A limiting constraint consistent with the subgradient of the non-smooth control barrier function is applied to the single-step increment of the control quantity. The control interval control law is selected according to the directional derivative of the non-smooth control barrier function and the inversion density. When the boundary back-and-forth crossing trend is detected, the control structure of the previous control interval is maintained and the control quantity amplitude is updated.

[0023] The NumPy library's `diff` function is called again to calculate the first-order difference of the state increment symbol sequence generated in the previous step. The number of non-zero elements in the difference result is counted and divided by the total number of sliding window steps to obtain the symbol inversion density. The SciPy library's interpolation module is used to establish a mapping relationship between the inversion density and the adjustment parameters, mapping a high inversion density to a wider neighborhood and a smaller absolute slope value, and a low inversion density to a narrower neighborhood and a larger absolute slope value.

[0024] Substitute the adjusted parameters into the non-smooth control barrier function, call the CVXPY convex optimization calculation library to convert the state model into a control affine form, and construct a quadratic programming problem with the objective function of minimizing the change in control input and the constraint that the control barrier function is greater than or equal to zero.

[0025] The OSQP open-source solver was configured to calculate three physical control quantities: the pulse width modulation duty cycle of the light source power, the given frequency of the stirrer motor inverter, and the inlet flow rate of the peristaltic pump. For points where the absolute value term is not differentiable, the backpropagation function in the differential module of the PyTorch deep learning framework was called to calculate the subgradient, and the clip limiting function was used to strictly constrain the upper and lower bounds of the control single-step increment to a constant multiple of the subgradient.

[0026] In addition, the dot product function of the NumPy library is called to calculate the directional derivative between the gradient vector of the non-smooth control barrier function and the closed-loop vector field of the system. When the directional derivative alternates between positive and negative in adjacent sampling periods and the sign inversion density is greater than 0.5, it is determined that a boundary crossing trend has occurred. At this time, the control state machine is locked by the Python conditional statement, and the quadratic programming objective function structure of the control interval activated in the previous moment is forcibly maintained. Only the new state is substituted into the OSQP solver to update the control quantity amplitude, so as to prevent the physical control command from oscillating at high frequency.

[0027] As one possible implementation, the calculation of the symbol inversion density of the symbol sequence and its mapping to neighborhood width and piecewise linear slope adjustment parameters, wherein when the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased, and vice versa, the neighborhood width is decreased and the absolute value of the slope is increased, including: The number of times adjacent elements in the state increment symbol sequence have opposite signs is counted. The symbol inversion density is calculated using the ratio of the number of occurrences to the sequence length. A positive linear mapping relationship between the boundary neighborhood width and the symbol inversion density is established, as well as a negative linear mapping relationship between the absolute value of the slope of the broken line and the symbol inversion density, while ensuring that the absolute value of the slope is not lower than the preset non-negative minimum absolute value of the slope. When the state trajectory oscillates at high frequency, causing the symbol inversion density to increase, the neighborhood width is expanded and the steepness of the slope is weakened, thus smoothing the control effect and suppressing chattering.

[0028] Perform an XOR operation or a first-order discrete difference on an incremental symbol sequence of a preset length N=20, count the number of zero-crossing events m, and calculate the symbol inversion density representing the oscillation frequency. The mapping executor is based on a linear relationship. The boundary neighborhood width is calculated, with a baseline width of 0.05 and a gain of 0.2, based on an exponential decay model. Calculate the absolute value of the slope of the broken line.

[0029] When turbulent flow causes high-frequency oscillations in the state trajectory, resulting in a sharp increase in reversal density, the command neighborhood bandwidth is widened to 0.18 and the absolute value of the slope is reduced to 0.59. This topology transformation converts hard switching into a continuous sliding mode transition within a buffer band, filtering out chattering from the actuator. Figure 2 As shown, Figure 2 The curve shows the change in the absolute width of the boundary neighborhood under strong disturbances of light intensity and stirring speed over 50 seconds.

[0030] As one possible implementation, the step of constructing a constraint control law based on the adjusted non-smooth control barrier function and state model, obtaining the control quantities of light source power, stirring frequency, and liquid inlet flow rate, and applying a limiting constraint consistent with the subgradient of the non-smooth control barrier function to the single-step increment of the control quantities includes: The control law solution is transformed into a restricted quadratic programming problem. The objective function includes quadratic and linear terms of the control quantity. The constraint condition is set as the time derivative of the non-smooth control barrier function plus the value of the extended class K function is less than or equal to zero. By solving the restricted quadratic programming problem, the corresponding light source power, stirring frequency, and liquid inlet flow rate control quantities are calculated. The set of subgradients of the non-smooth control barrier function in the current state is calculated. The direction of the minimum norm of the subgradient is taken as the limiting reference direction. The minimum value between the single-step maximum physical response increment of the actuator and the subgradient norm after the positive diagonal matrix transformation used for dimension conversion and time adjustment is used to limit the single-step increment of the control quantity.

[0031] The control law is derived in real time using a constrained quadratic programming framework based on the interior-point method. The objective function is formulated as follows: The control variable u includes the light source power, stirring frequency, and liquid inlet flow rate; the constraint condition is the time Lie derivative of the non-smooth control barrier function and its extended class K function. Neither positive nor negative. For non-differentiable singularities on the boundary polygonal line, use Clark's generalized subdifferential rule to solve for the subgradient convex hull set. The minimum norm direction is extracted as the limiting reference. The maximum permissible physical increment of the actuator is obtained and compared with the subgradient norm after scaling the positive diagonal matrix. Hard limiting is implemented by taking the minimum value operator to ensure that the controlled physical object moves smoothly within the safety limit strictly according to the subgradient descent guide.

[0032] As one possible implementation, the step of selecting the control interval control law based on the directional derivative and inversion density of the non-smooth control barrier function, and maintaining the control structure of the previous control interval and updating the control magnitude when a boundary round-trip crossing trend is detected, includes: The Lie derivative of the system state along the current vector field direction is calculated, and the migration trend of the system state to the adjacent control interval is judged by combining the sign inversion density. When it is detected that the system state crosses the boundary between two adjacent control intervals multiple times and the sign inversion density is greater than the preset oscillation threshold, the switching locking mechanism is triggered. During the locking period, the system model switching to the new control interval is suspended, the system parameter matrix and control law structure of the previous control interval are used, and the control quantity amplitude in the current control law is updated iteratively only by gradient descent method until the sign inversion density is lower than the safe escape threshold.

[0033] When dealing with the multidimensional changes caused by bubble turbulence, the digital processing module calculates the Lie derivative of the current vector field and combines it with the inversion density. Assess the system's cross-regional migration trend. If the monitor detects that within a preset listening time window (e.g., 5.0 seconds), the state trajectory crosses the boundary of adjacent physical regions more than a preset number of times (e.g., 3 times), and the sign inversion density is high... If the value exceeds the warning threshold, preferably 0.75, the main controller will trigger a switching and locking mechanism.

[0034] Within a lockout delay period, such as 10 seconds, the matching and loading of the new system model is paused, and the system model parameters and quadratic optimization architecture from the previous period are reused. In the locked state, only a small exploration step, such as the learning rate, is initiated. The gradient descent operation only iteratively fine-tunes the magnitude of the control input. This degradation-preservation strategy continues to run until density is detected. Once the temperature naturally drops and stabilizes at a safe level beyond the tolerance boundary, the lock is released, and the autonomous switching and parameter optimization mode across the entire range is restored.

[0035] S3 utilizes the photocurrent and gas holdup to decouple the observed variables, and corrects the carrier recombination and mass transfer weights based on the model prediction bias and state return amount, respectively, to obtain the control input for the next moment and complete the hydrogen production control.

[0036] Real-time photocurrent data was acquired using an externally connected electrochemical workstation, and the pressure difference across the reactor was converted into gas holdup data using a differential pressure transmitter. The extended Kalman filter module from the FilterPy filtering algorithm library was invoked to configure the photocurrent and gas holdup as decoupled observation vectors for time-space equation updates and measurement updates.

[0037] The prediction deviation between the actual observed values ​​of the four-dimensional state vector and the predicted values ​​of the piecewise state model is calculated. Using the peak-finding algorithm in the SciPy library, the maximum state return amount after the state variable crosses the boundary of the control barrier function and returns to the safe set is recorded. The recursive least squares algorithm module in the Statsmodels statistical analysis library is called to fit the recombination rate constant in the carrier recombination rate formula online based on the prediction deviation, and the volumetric mass transfer coefficient in the gas-liquid mass transfer rate formula is corrected based on the maximum state return amount. Using the state model with updated weights, the CVXPY convex optimization problem is reconfigured, and the optimal control sequence for the next control cycle is generated.

[0038] Using Python's pymodbus communication library, the calculated light source power, stirring frequency, and liquid inlet flow rate control input commands are sent to the programmable logic controller, thereby completing the constant hydrogen production closed-loop control of the entire photocatalytic system.

[0039] As one possible implementation, the method of decoupling the observed variables of photocurrent and gas holdup, and correcting the carrier recombination and mass transfer weights according to model prediction bias and state return amount to obtain the control input for the next time step to complete hydrogen production control includes: Decoupled observation data of photocurrent and gas holdup are acquired. The predicted square error between the model prediction and the actual measured values ​​of the reactor photocurrent and gas holdup at the current sampling time is calculated. The maximum back-return amplitude after the state deviates from the equilibrium point is calculated. Microscopic compensation functions containing the predicted square error of photocurrent and the maximum back-return amplitude, and macroscopic compensation functions containing the predicted square error of gas holdup and the maximum back-return amplitude are constructed respectively. The model weights of the carrier recombination rate matrix in the state model are corrected using the microscopic compensation function, and the model weights of the gas-liquid mass transfer coefficient matrix are corrected using the macroscopic compensation function. The decoupled model parameters are made closer to the actual reaction kinetics. The updated model weights are substituted into the quadratic programming solver at the next sampling time to recalculate the control law, generate closed-loop drive commands, and send them to the actuator.

[0040] Photocurrent and gas holdup are used as decoupled observation variables. On the hardware side, a picoampere-level high-sensitivity electrochemical workstation collects microscopic photocurrent, while a macroscopic gas holdup is extracted simultaneously using an ultrasonic Doppler monitoring system, with an optimal observation range of 5% to 20%. The control unit calculates the squared prediction errors between the actual observed values ​​of photocurrent and gas holdup and the virtual simulation prediction values. and Simultaneously, the maximum Euclidean distance at which the system falls back after being disturbed and deviating from the safety boundary is extracted as the maximum backtracking deviation parameter. The hard constraint warning threshold is ≤0.08. For example... Figure 3 As shown, Figure 3 The curves show the variation of the prediction mean square error of photocurrent and gas holdup under strong perturbation.

[0041] Based on the above parameters, initiate the optimization iteration: construct the micro-compensation function. The model weight parameters of the carrier recombination rate matrix are updated continuously and smoothly, for example, the initially fixed recombination weight of 0.4 is increased to approach the actual 0.45; a macroscopic compensation function is constructed simultaneously. The mass transfer coefficient constant matrix weights representing the liquid-gas fusion volume flux are corrected. After completing the full-scale calibration from micro to macro, the updated parameter matrix cluster will be used in the next control cycle. The underlying computational registration stack is pushed in, and the constrained quadratic programming problem is solved again. The latest planned smooth drive instructions are sent to the LED power management device, the variable frequency liquid feed pump, and the multiphase stirring motor through the optocoupler barrier, thus forming a closed loop for the optical drive hydrogen production control.

[0042] The experimental conditions were set in a 5L multiphase photocatalytic reactor, loaded with a uniformly concentrated carbon nitride slurry. The initial slurry temperature was controlled at 298 Kelvin, the initial light intensity was set at 400 W / m², and the sampling frequency was uniformly locked at 20Hz. Three control models were set up for comparative verification. Control group 1 used a traditional global fixed gain continuous state feedback control architecture; control group 2 adopted a segmented interval partitioning and static fixed boundary non-smooth barrier function control algorithm; and experimental group 3 adopted a complete neighborhood piecewise linear slope adjustment and a composite control strategy based on decoupled dual-path weight correction of photocurrent and gas holdup. At 50s of stable operation, a transient composite extreme disturbance was used. A step command was used to instantaneously jump the light intensity to 850 W / m², and a strong excitation pulse with an amplitude of 300 rpm from the stirring motor was applied to artificially create conditions of deteriorated gas-liquid mass transfer and abrupt change in photothermal boundary. The total test monitoring time was set to 200s. The comparison of the steady-state deviation of hydrogen production after strong disturbance under different control strategies is as follows: Figure 4 As shown.

[0043] Specific experimental data show that under external abrupt shocks, control group 1 lost its stable response state. The single-step power adjustment command for the lower light source experienced a sudden change of up to 75 watts, triggering a hardware safety limit dead zone. The control system fell into a high-frequency oscillation trap, causing the symbol inversion density to surge to 0.85. It took 120 seconds to barely return to the safe envelope, and the steady-state deviation of hydrogen production remained at 12.5%. Control group 2 benefited from segmented interval matching, which improved its anti-disturbance collapse capability. However, the fixed obstacle polygonal line caused the trajectory to cross the boundary limit as many as 8 times, and the steady-state optimization recovery time consumed 65 seconds. The prediction tracking mean square error remained at 5.8%. Experimental group 3, within the time delay of detecting the symbol inversion density climbing to 0.62, performed a topology transformation to widen the absolute boundary width to 0.18 and soften the polarization slope to 0.59. The control command increment was constrained within a 42-watt no-bounce range. Only a single lossless smooth transition occurred, and the physical surge was quelled within 21 seconds, returning to the ideal target center steady-state point. Furthermore, the steady-state observation absolute tracking error was compressed to only 0.9%.

[0044] A mechanism employing symbolic sequence calculation to calculate density in real-time and adjust the control surface morphology achieves online immune filtering of high-frequency unknown flow field noise, solving the technical problem of hard breakdown and chattering wear of actuators induced by classical non-smooth obstacle strategies at multi-subspace critical surfaces. A dynamic matrix cross-weight rebalancing algorithm, decoupled from the transient characteristics of photocurrent at the micro-level and the gas holdup volume measurement at the macro-level, repairs the distortion black hole in the mechanistic parameter model of the three-phase interface, increasing the trajectory strong pull-back response speed by several times in the face of sudden severe energy field imbalances. This endows the hydrogen production system with a wide range of surge-resistant forced survival and stable endurance performance, while ensuring the equipment does not suffer irreversible mechanical, electrical, or thermal breakdown to the greatest extent possible.

[0045] One or more embodiments of the present invention provide an adaptive control system for g-C3N4 photocatalytic hydrogen production based on multimodal sensing, the system comprising: The determination module is used to acquire time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration and catalyst concentration to construct a four-dimensional state vector of hydrogen production reaction, and acquire time-series data of light intensity and stirring speed as algebraic variables. Based on the state vector and algebraic variables, a segmented state model including carrier generation, recombination, hydrogen evolution reaction and mass transfer is established. The control interval is divided according to the preset thresholds of light intensity, temperature and dissolved hydrogen concentration. A non-smooth control barrier function is constructed for each control interval. The non-smooth control barrier function adopts a combination of absolute value term and quadratic term boundary. The slope of the boundary neighborhood piecewise linear function is determined according to the historical state difference sign sequence. The update module is used to calculate the sign inversion density of the sign sequence and map it to neighborhood width and piecewise slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. Based on the adjusted non-smooth control barrier function and state model, a constraint control law is constructed to obtain the control quantities of light source power, stirring frequency and liquid inlet flow rate. The single-step increment of the control quantity is subject to a limiting constraint consistent with the subgradient of the non-smooth control barrier function. The control interval control law is selected according to the directional derivative of the non-smooth control barrier function and the inversion density. When the boundary back-and-forth crossing trend is detected, the control structure of the previous control interval is maintained and the control quantity amplitude is updated. The control module is used to decouple the observed variables from the photocurrent and the gas holdup, and correct the carrier recombination and mass transfer weights according to the model prediction bias and the state return amount, so as to obtain the control input for the next moment and complete the hydrogen production control.

[0046] As one possible implementation, the step of acquiring time-series data on reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration to construct a four-dimensional state vector for the hydrogen production reaction, and acquiring time-series data on light intensity and stirring speed as algebraic variables, includes: The sensor network synchronously collects signals of light intensity, slurry temperature, pH value, dissolved hydrogen concentration, catalyst concentration, and stirring speed in the reactor at a preset sampling frequency. The collected signals are low-pass filtered and normalized to eliminate sensor noise and dimensional differences. The processed slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration are combined in time series to construct a four-dimensional state vector representing the multi-physics field correlation state of the hydrogen production reaction system at the current moment. At the same time, the processed light intensity and stirring speed are extracted as algebraic variables.

[0047] As one possible implementation, the segmented state model established based on state vectors and algebraic variables, encompassing carrier generation, recombination, hydrogen evolution reaction, and mass transfer, is used to divide the control interval according to preset thresholds for light intensity, temperature, and dissolved hydrogen concentration, including: A set of differential equations is established based on the semiconductor photocatalysis mechanism. The set of differential equations includes the time derivative of a four-dimensional state vector, the four-dimensional state vector, and a control input vector. Thresholds for light intensity, slurry temperature, and dissolved hydrogen concentration are set. The light intensity in the algebraic variables and the slurry temperature and dissolved hydrogen concentration in the current state vector are compared with the corresponding thresholds. Based on the Boolean logic combination of each variable being higher or lower than the threshold, the global state space is divided into multiple subspaces as control intervals. A corresponding system parameter matrix is ​​matched for each control interval.

[0048] As one possible implementation, constructing a non-smooth control barrier function for each control interval, wherein the non-smooth control barrier function uses a combination of absolute value terms and quadratic terms as boundaries, and determines the slope of the boundary neighborhood piecewise linear curve based on the historical state difference sign sequence, includes: Define the deviation vector from the current state of the system to the target equilibrium manifold. Construct a non-smooth control barrier function by multiplying the first weight coefficient with the absolute norm of the deviation vector and the second weight coefficient with the square of the Euclidean norm of the deviation vector. Extract the norm increment of the state deviation over a preset number of sampling periods. Use the sign function to extract the increment direction and generate a state increment sign sequence. Based on the distribution characteristics of positive and negative signs in the sign sequence, set the initial piecewise linear slope parameter of the non-smooth control barrier function in the zero neighborhood.

[0049] As one possible implementation, the calculation of the symbol inversion density of the symbol sequence and its mapping to neighborhood width and piecewise linear slope adjustment parameters, wherein when the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased, and vice versa, the neighborhood width is decreased and the absolute value of the slope is increased, including: The number of times adjacent elements in the state increment symbol sequence have opposite signs is counted. The symbol inversion density is calculated using the ratio of the number of occurrences to the sequence length. A positive linear mapping relationship between the boundary neighborhood width and the symbol inversion density is established, as well as a negative linear mapping relationship between the absolute value of the slope of the broken line and the symbol inversion density, while ensuring that the absolute value of the slope is not lower than the preset non-negative minimum absolute value of the slope. When the state trajectory oscillates at high frequency, causing the symbol inversion density to increase, the neighborhood width is expanded and the steepness of the slope is weakened, thus smoothing the control effect and suppressing chattering.

[0050] As one possible implementation, the step of constructing a constraint control law based on the adjusted non-smooth control barrier function and state model, obtaining the control quantities of light source power, stirring frequency, and liquid inlet flow rate, and applying a limiting constraint consistent with the subgradient of the non-smooth control barrier function to the single-step increment of the control quantities includes: The control law solution is transformed into a restricted quadratic programming problem. The objective function includes quadratic and linear terms of the control quantity. The constraint condition is set as the time derivative of the non-smooth control barrier function plus the value of the extended class K function is less than or equal to zero. By solving the restricted quadratic programming problem, the corresponding light source power, stirring frequency, and liquid inlet flow rate control quantities are calculated. The set of subgradients of the non-smooth control barrier function in the current state is calculated. The direction of the minimum norm of the subgradient is taken as the limiting reference direction. The minimum value between the single-step maximum physical response increment of the actuator and the subgradient norm after the positive diagonal matrix transformation used for dimension conversion and time adjustment is used to limit the single-step increment of the control quantity.

[0051] As one possible implementation, the step of selecting the control interval control law based on the directional derivative and inversion density of the non-smooth control barrier function, and maintaining the control structure of the previous control interval and updating the control magnitude when a boundary round-trip crossing trend is detected, includes: The Lie derivative of the system state along the current vector field direction is calculated, and the migration trend of the system state to the adjacent control interval is judged by combining the sign inversion density. When it is detected that the system state crosses the boundary between two adjacent control intervals multiple times and the sign inversion density is greater than the preset oscillation threshold, the switching locking mechanism is triggered. During the locking period, the system model switching to the new control interval is suspended, the system parameter matrix and control law structure of the previous control interval are used, and the control quantity amplitude in the current control law is updated iteratively only by gradient descent method until the sign inversion density is lower than the safe escape threshold.

[0052] As one possible implementation, the method of decoupling the observed variables of photocurrent and gas holdup, and correcting the carrier recombination and mass transfer weights according to model prediction bias and state return amount to obtain the control input for the next time step to complete hydrogen production control includes: Decoupled observation data of photocurrent and gas holdup are acquired. The predicted square error between the model prediction and the actual measured values ​​of the reactor photocurrent and gas holdup at the current sampling time is calculated. The maximum back-return amplitude after the state deviates from the equilibrium point is calculated. Microscopic compensation functions containing the predicted square error of photocurrent and the maximum back-return amplitude, and macroscopic compensation functions containing the predicted square error of gas holdup and the maximum back-return amplitude are constructed respectively. The model weights of the carrier recombination rate matrix in the state model are corrected using the microscopic compensation function, and the model weights of the gas-liquid mass transfer coefficient matrix are corrected using the macroscopic compensation function. The decoupled model parameters are made closer to the actual reaction kinetics. The updated model weights are substituted into the quadratic programming solver at the next sampling time to recalculate the control law, generate closed-loop drive commands, and send them to the actuator.

[0053] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. An adaptive control method for photocatalytic hydrogen production of g-C3N4 based on multimodal sensing, characterized in that, include: A four-dimensional state vector for hydrogen production reaction is constructed by acquiring time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration. Time-series data of light intensity and stirring speed are acquired as algebraic variables. Based on the state vector and algebraic variables, a segmented state model including carrier generation, recombination, hydrogen evolution reaction, and mass transfer is established. Control intervals are divided according to preset thresholds for light intensity, temperature, and dissolved hydrogen concentration. Non-smooth control barrier functions are constructed for each control interval. The non-smooth control barrier functions use a combination of absolute value terms and quadratic terms as boundary conditions. The slope of the boundary neighborhood broken line is determined based on the historical state difference sign sequence. The sign inversion density of the sign sequence is calculated and mapped to neighborhood width and piecewise slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. Based on the adjusted non-smooth control barrier function and state model, a constraint control law is constructed to obtain the control quantities of light source power, stirring frequency, and liquid inlet flow rate. A limiting constraint consistent with the subgradient of the non-smooth control barrier function is applied to the single-step increment of the control quantity. The control interval control law is selected according to the directional derivative of the non-smooth control barrier function and the inversion density. When the boundary back-and-forth crossing trend is detected, the control structure of the previous control interval is maintained and the control quantity amplitude is updated. By decoupling the observed variables photocurrent and gas holdup, and correcting the carrier recombination and mass transfer weights based on model prediction bias and state return, the control input for the next moment is obtained to complete hydrogen production control.

2. The method according to claim 1, characterized in that, The time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration are used to construct a four-dimensional state vector for the hydrogen production reaction, and the time-series data of light intensity and stirring speed are used as algebraic variables, including: The sensor network synchronously collects signals of light intensity, slurry temperature, pH value, dissolved hydrogen concentration, catalyst concentration, and stirring speed in the reactor at a preset sampling frequency. The collected signals are low-pass filtered and normalized to eliminate sensor noise and dimensional differences. The processed slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration are combined in time series to construct a four-dimensional state vector representing the multi-physics field correlation state of the hydrogen production reaction system at the current moment. At the same time, the processed light intensity and stirring speed are extracted as algebraic variables.

3. The method according to claim 1, characterized in that, The segmented state model, based on state vectors and algebraic variables, encompasses carrier generation, recombination, hydrogen evolution reaction, and mass transfer. Control intervals are defined according to preset thresholds for light intensity, temperature, and dissolved hydrogen concentration, including: A set of differential equations is established based on the semiconductor photocatalysis mechanism. The set of differential equations includes the time derivative of a four-dimensional state vector, the four-dimensional state vector, and a control input vector. Thresholds for light intensity, slurry temperature, and dissolved hydrogen concentration are set. The light intensity in the algebraic variables and the slurry temperature and dissolved hydrogen concentration in the current state vector are compared with the corresponding thresholds. Based on the Boolean logic combination of each variable being higher or lower than the threshold, the global state space is divided into multiple subspaces as control intervals. A corresponding system parameter matrix is ​​matched for each control interval.

4. The method according to claim 1, characterized in that, The construction of a non-smooth control barrier function for each control interval, wherein the non-smooth control barrier function uses a combination of absolute value terms and quadratic terms as boundaries, and determines the slope of the boundary neighborhood piecewise linear curve based on the historical state difference sign sequence, includes: Define the deviation vector from the current state of the system to the target equilibrium manifold. Construct a non-smooth control barrier function by multiplying the first weight coefficient with the absolute norm of the deviation vector and the second weight coefficient with the square of the Euclidean norm of the deviation vector. Extract the norm increment of the state deviation over a preset number of sampling periods. Use the sign function to extract the increment direction and generate a state increment sign sequence. Based on the distribution characteristics of positive and negative signs in the sign sequence, set the initial piecewise linear slope parameter of the non-smooth control barrier function in the zero neighborhood.

5. The method according to any one of claims 1-4, characterized in that, The calculation of the symbol inversion density of the symbol sequence is mapped to neighborhood width and piecewise linear slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. This includes: The number of times adjacent elements in the state increment symbol sequence have opposite signs is counted. The symbol inversion density is calculated using the ratio of the number of occurrences to the sequence length. A positive linear mapping relationship between the boundary neighborhood width and the symbol inversion density is established, as well as a negative linear mapping relationship between the absolute value of the slope of the broken line and the symbol inversion density, while ensuring that the absolute value of the slope is not lower than the preset non-negative minimum absolute value of the slope. When the state trajectory oscillates at high frequency, causing the symbol inversion density to increase, the neighborhood width is expanded and the steepness of the slope is weakened, thus smoothing the control effect and suppressing chattering.

6. The method according to claim 5, characterized in that, The constraint control law is constructed based on the adjusted non-smooth control barrier function and state model. The control quantities of light source power, stirring frequency, and liquid inlet flow rate are obtained. A limiting constraint consistent with the subgradient of the non-smooth control barrier function is applied to the single-step increment of the control quantities, including: The control law solution is transformed into a restricted quadratic programming problem. The objective function includes quadratic and linear terms of the control quantity. The constraint condition is set as the time derivative of the non-smooth control barrier function plus the value of the extended class K function is less than or equal to zero. By solving the restricted quadratic programming problem, the corresponding light source power, stirring frequency, and liquid inlet flow rate control quantities are calculated. The set of subgradients of the non-smooth control barrier function in the current state is calculated. The direction of the minimum norm of the subgradient is taken as the limiting reference direction. The minimum value between the single-step maximum physical response increment of the actuator and the subgradient norm after the positive diagonal matrix transformation used for dimension conversion and time adjustment is used to limit the single-step increment of the control quantity.

7. The method according to claim 1, characterized in that, The step of selecting the control interval control law based on the directional derivative and inversion density of the non-smooth control barrier function, and maintaining the control structure of the previous control interval and updating the control magnitude when a boundary back-and-forth crossing trend is detected, includes: The Lie derivative of the system state along the current vector field direction is calculated, and the migration trend of the system state to the adjacent control interval is judged by combining the sign inversion density. When it is detected that the system state crosses the boundary between two adjacent control intervals multiple times and the sign inversion density is greater than the preset oscillation threshold, the switching locking mechanism is triggered. During the locking period, the system model switching to the new control interval is suspended, the system parameter matrix and control law structure of the previous control interval are used, and the control quantity amplitude in the current control law is updated iteratively only by gradient descent method until the sign inversion density is lower than the safe escape threshold.

8. The method according to claim 1 or 7, characterized in that, The process of decoupling the observed variables of photocurrent and gas holdup, and correcting the carrier recombination and mass transfer weights based on model prediction bias and state return parameters to obtain the control input for the next time step to complete hydrogen production control includes: Decoupled observation data of photocurrent and gas holdup are acquired. The predicted square error between the model prediction and the actual measured values ​​of the reactor photocurrent and gas holdup at the current sampling time is calculated. The maximum back-return amplitude after the state deviates from the equilibrium point is calculated. Microscopic compensation functions containing the predicted square error of photocurrent and the maximum back-return amplitude, and macroscopic compensation functions containing the predicted square error of gas holdup and the maximum back-return amplitude are constructed respectively. The model weights of the carrier recombination rate matrix in the state model are corrected using the microscopic compensation function, and the model weights of the gas-liquid mass transfer coefficient matrix are corrected using the macroscopic compensation function. The decoupled model parameters are made closer to the actual reaction kinetics. The updated model weights are substituted into the quadratic programming solver at the next sampling time to recalculate the control law, generate closed-loop drive commands, and send them to the actuator.

9. An adaptive control system for g-C3N4 photocatalytic hydrogen production based on multimodal sensing, characterized in that, Includes the following modules: The determination module is used to acquire time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration and catalyst concentration to construct a four-dimensional state vector of hydrogen production reaction, and acquire time-series data of light intensity and stirring speed as algebraic variables. Based on the state vector and algebraic variables, a segmented state model including carrier generation, recombination, hydrogen evolution reaction and mass transfer is established. The control interval is divided according to the preset thresholds of light intensity, temperature and dissolved hydrogen concentration. A non-smooth control barrier function is constructed for each control interval. The non-smooth control barrier function adopts a combination of absolute value term and quadratic term boundary. The slope of the boundary neighborhood piecewise linear function is determined according to the historical state difference sign sequence. The update module is used to calculate the sign inversion density of the sign sequence and map it to neighborhood width and piecewise slope adjustment parameters. When the inversion density increases, the neighborhood width is increased and the absolute value of the slope is decreased; conversely, the neighborhood width is decreased and the absolute value of the slope is increased. Based on the adjusted non-smooth control barrier function and state model, a constraint control law is constructed to obtain the control quantities of light source power, stirring frequency and liquid inlet flow rate. The single-step increment of the control quantity is subject to a limiting constraint consistent with the subgradient of the non-smooth control barrier function. The control interval control law is selected according to the directional derivative of the non-smooth control barrier function and the inversion density. When the boundary back-and-forth crossing trend is detected, the control structure of the previous control interval is maintained and the control quantity amplitude is updated. The control module is used to decouple the observed variables from the photocurrent and the gas holdup, and correct the carrier recombination and mass transfer weights according to the model prediction bias and the state return amount, so as to obtain the control input for the next moment and complete the hydrogen production control.

10. The system according to claim 9, characterized in that, The time-series data of reactor slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration are used to construct a four-dimensional state vector for the hydrogen production reaction, and the time-series data of light intensity and stirring speed are used as algebraic variables, including: The sensor network synchronously collects signals of light intensity, slurry temperature, pH value, dissolved hydrogen concentration, catalyst concentration, and stirring speed in the reactor at a preset sampling frequency. The collected signals are low-pass filtered and normalized to eliminate sensor noise and dimensional differences. The processed slurry temperature, pH value, dissolved hydrogen concentration, and catalyst concentration are combined in time series to construct a four-dimensional state vector representing the multi-physics field correlation state of the hydrogen production reaction system at the current moment. At the same time, the processed light intensity and stirring speed are extracted as algebraic variables.

Images