Photovoltaic off-grid hydrogen production system and control method thereof

Abstract

The invention discloses a photovoltaic off-grid hydrogen production system and a control method thereof, and the method comprises the following steps: obtaining original data of an environment sensor, carrying out synchronous collection through a distributed data collection system, and obtaining standardized environment data through an STL algorithm, Z-score standardization and wavelet noise reduction processing; carrying out photovoltaic output power prediction by adopting an integrated learning method, and carrying out MPPT control optimization; an electrolytic cell dynamic model is constructed, multi-objective optimization and energy storage system state estimation are carried out, and energy storage optimization scheduling is achieved; and finally, realizing overall optimization of the system through a multi-time-scale coordination control framework. The system operation efficiency is improved, the system reliability is enhanced, the operation cost is reduced, and the system is particularly suitable for an off-grid hydrogen production scene with large photovoltaic energy volatility.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrogen production, and in particular to a photovoltaic off-grid hydrogen production system and a control method thereof. Background Art

[0002] With the advancement of global energy transformation and carbon neutrality goals, hydrogen production from renewable energy has become an important technical path to solve the problems of clean energy and hydrogen supply. Photovoltaic hydrogen production systems convert solar energy into electricity and then produce hydrogen through water electrolysis, thus realizing the transformation from renewable energy to clean hydrogen energy. They have the advantages of zero carbon emissions, a wide range of raw material sources, and a clean process. Especially in remote areas, islands and other areas where power grids are difficult to cover, photovoltaic off-grid hydrogen production systems can produce hydrogen from solar energy resources on site, providing clean energy for the local area, while avoiding infrastructure investment and energy losses caused by long-distance power and hydrogen transmission, and have important engineering application value and economic benefits.

[0003] At present, the research on photovoltaic off-grid hydrogen production system mainly focuses on system configuration construction and basic control strategy. In terms of system configuration, the direct coupling of photovoltaic array, DC / DC converter and electrolyzer is generally adopted, or the energy storage unit is introduced to form a hybrid system of photovoltaic, energy storage and electrolyzer. In terms of control strategy, the traditional MPPT (maximum power point tracking) control and simple proportional-integral (PI) control are mainly used to adjust the duty cycle of DC / DC converter to achieve the adjustment of photovoltaic output power. For systems containing energy storage units, rule-based control strategies are mostly used to determine the charging and discharging operations of energy storage units according to preset logical judgment conditions. In terms of electrolyzer control, the main focus is on basic current control and temperature control to maintain the stable operation of the electrolyzer.

[0004] However, the existing photovoltaic off-grid hydrogen production system still faces some specific technical challenges in practical applications: First, the traditional MPPT algorithms such as the perturbation and observation method (P&O) and the conductivity increment method are prone to oscillation and mistracking when the light changes rapidly, affecting the system efficiency; second, the PI controller based on fixed parameters is difficult to cope with the strong nonlinear characteristics of the system, especially when the electrolyzer impedance changes dynamically with the working state, the control performance is significantly reduced; third, the existing rules based on energy storage control strategy lack optimization considerations for the long-term operating efficiency of the system and the life of the equipment, which often leads to excessive use of energy storage units or unreasonable load distribution; fourth, the coordination control between the various parts of the system is insufficient, such as the photovoltaic output fluctuation is directly transmitted to the electrolyzer through the DC / DC converter, which easily leads to frequent changes in the working point of the electrolyzer, affecting the hydrogen production efficiency and equipment life; fifth, there is a lack of accurate estimation and prediction capabilities for the system operating status, especially when the energy storage unit ages and the electrolyzer performance degrades, the system is difficult to adjust the control strategy in time; sixth, the existing control methods do not take into account uncertain factors such as weather changes and equipment failures, and the robustness and reliability of the system need to be improved. In addition, there are the intermittent and instability of photovoltaic power generation, energy management and storage issues in the hydrogen production process, and system efficiency optimization. How to stably output electricity suitable for electrolytic hydrogen production under different lighting conditions is also an important technical problem.

[0005] These technical problems seriously restrict the engineering application and promotion of photovoltaic off-grid hydrogen production systems. Summary of the invention

[0006] The purpose of the invention is to provide a photovoltaic off-grid hydrogen production system and a control method thereof, in order to solve at least one technical problem existing in the prior art.

[0007] Technical solution: A photovoltaic off-grid hydrogen production control method, comprising the following steps:

[0008] Step S1, obtaining the original data of the environmental sensor, using a distributed data acquisition system for synchronous acquisition, and obtaining a time-stamped original environmental data set containing solar radiation intensity, ambient temperature, photovoltaic panel surface temperature and wind speed; performing anomaly detection, data cleaning, standardization, spatiotemporal difference, feature extraction and dimensionality reduction on the original environmental data set in sequence, and obtaining a feature data set after dimensionality reduction;

[0009] Step S2, acquiring a feature data set and pre-stored historical photovoltaic output data, using an integrated learning method combined with time series analysis to predict photovoltaic output power, and obtaining predicted photovoltaic output power; based on the predicted photovoltaic output power and current photovoltaic operating point data, applying an extreme value optimization algorithm to perform MPPT control optimization, and obtaining an optimized photovoltaic operating point voltage; according to the predicted photovoltaic output power and pre-stored photovoltaic array topology information, applying a graph theory algorithm to reconstruct and optimize the photovoltaic array, and obtaining an optimized photovoltaic array topology;

[0010] Step S3, obtaining the pre-stored historical operation data of the electrolyzer and the specification parameters of the electrolyzer, and applying the system identification technology to construct a dynamic model of the electrolyzer; based on the constructed dynamic model and the predicted photovoltaic output power, applying a multi-objective optimization algorithm to optimize the efficiency of the electrolyzer, and obtaining an optimized electrolyzer working point sequence; according to the electrolyzer dynamic model and the optimized working point sequence, constructing an electrolyzer controller based on model predictive control;

[0011] Step S4, acquiring the voltage, current and temperature data of the energy storage system collected in real time, applying the Kalman filter algorithm to estimate the state of the energy storage system, and obtaining the battery state of charge SOC and health state SOH; based on the state estimation results and the pre-stored historical usage data, applying the semi-Markov degradation model to predict the life of the energy storage system, and obtaining the battery remaining life distribution; according to the predicted photovoltaic output power, the optimized electrolyzer operating point sequence and the battery status information, applying random dynamic programming to optimize the scheduling of the energy storage system, and obtaining the optimal scheduling strategy of the energy storage system.

[0012] According to another aspect of the present application, a photovoltaic off-grid hydrogen production system is also provided, comprising:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one processor; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the photovoltaic off-grid hydrogen production method of any of the above technical solutions.

[0016] Beneficial effects include high efficiency of electricity utilization and high economical system operation. It can also be applied to large-scale hydrogen production, which improves the stability of the system and the flexibility of hydrogen production. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the present invention.

[0018] Figure 2 The present invention discloses a photovoltaic off-grid hydrogen production system.

[0019] Figure 3 This is a structural diagram of a photovoltaic off-grid hydrogen production system control method disclosed in an embodiment of the present invention.

[0020] Figure 4 This is a flow chart of an upper-level instruction control module in the control method disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, a photovoltaic off-grid hydrogen production control method is provided, comprising the following steps:

[0022] Step S1, obtaining the original data of the environmental sensor, using a distributed data acquisition system for synchronous acquisition, and obtaining a time-stamped original environmental data set containing solar radiation intensity, ambient temperature, photovoltaic panel surface temperature and wind speed; performing anomaly detection, data cleaning, standardization, spatiotemporal difference, feature extraction and dimensionality reduction on the original environmental data set in sequence, and obtaining a feature data set after dimensionality reduction;

[0023] Step S2, acquiring a feature data set and pre-stored historical photovoltaic output data, using an integrated learning method combined with time series analysis to predict photovoltaic output power, and obtaining predicted photovoltaic output power; based on the predicted photovoltaic output power and current photovoltaic operating point data, applying an extreme value optimization algorithm to perform MPPT control optimization, and obtaining an optimized photovoltaic operating point voltage; according to the predicted photovoltaic output power and pre-stored photovoltaic array topology information, applying a graph theory algorithm to reconstruct and optimize the photovoltaic array, and obtaining an optimized photovoltaic array topology;

[0024] Step S3, obtaining the pre-stored historical operation data of the electrolyzer and the specification parameters of the electrolyzer, and applying the system identification technology to construct a dynamic model of the electrolyzer; based on the constructed dynamic model and the predicted photovoltaic output power, applying a multi-objective optimization algorithm to optimize the efficiency of the electrolyzer, and obtaining an optimized electrolyzer working point sequence; according to the electrolyzer dynamic model and the optimized working point sequence, constructing an electrolyzer controller based on model predictive control;

[0025] Step S4, acquiring the voltage, current and temperature data of the energy storage system collected in real time, applying the Kalman filter algorithm to estimate the state of the energy storage system, and obtaining the battery state of charge SOC and health state SOH; based on the state estimation results and the pre-stored historical usage data, applying the semi-Markov degradation model to predict the life of the energy storage system, and obtaining the battery remaining life distribution; according to the predicted photovoltaic output power, the optimized electrolyzer operating point sequence and the battery status information, applying random dynamic programming to optimize the scheduling of the energy storage system, and obtaining the optimal scheduling strategy of the energy storage system.

[0026] By combining data-driven and intelligent optimization, a complete photovoltaic off-grid hydrogen production system solution was constructed. First, at the data level, high-quality environmental data acquisition and feature extraction were achieved through distributed data collection and multi-level data processing; second, at the control level, the coordinated operation of each subsystem of the system was achieved through multi-objective optimization and multi-time scale coordinated control; third, at the fault handling level, the reliable operation of the system was guaranteed through deep learning fault diagnosis and adaptive fault-tolerant control. The overall solution has achieved remarkable results in improving system efficiency, enhancing system reliability, and reducing operating costs. It is particularly suitable for off-grid hydrogen production scenarios with large photovoltaic energy volatility and frequent changes in load demand.

[0027] According to one aspect of the present application, step S1 specifically comprises:

[0028] Step S11, obtaining the original data of the environmental sensor, including solar radiation intensity, ambient temperature, photovoltaic panel surface temperature and wind speed; using a distributed data acquisition system to synchronously acquire these data; adding a timestamp to the acquired data; combining the data with the timestamp into a complete original environmental data set;

[0029] Step S12, obtaining the original environmental data set obtained in step S11; applying the seasonal trend decomposition algorithm to the data set to decompose each data series into seasonal, trend and residual components; applying the median absolute deviation method to the decomposed residual components to identify outliers; using the local regression method to interpolate and replace the identified outliers; recombining the processed data to obtain a cleaned environmental data set;

[0030] Step S13, obtaining the cleaned environmental data set obtained in step S12; applying the Z-score standardization method to each variable in the data set to calculate the standardized score of each data point; applying wavelet transform to the standardized data to perform signal denoising, specifically including selecting Daubechies wavelet basis function, performing multi-scale decomposition, applying soft threshold denoising and reconstructing the signal; combining the standardized and denoised data to obtain a standardized and denoised environmental data set.

[0031] Step S14, obtaining the standardized and denoised environmental data set obtained in step S13; applying the Kriging interpolation method to the data set for spatiotemporal interpolation, first constructing a spatiotemporal variogram model, and then using a sliding time window for local Kriging interpolation; applying Kalman filtering to the interpolation result for time series smoothing; combining the interpolated and smoothed data to obtain a high-resolution spatiotemporal environmental data set;

[0032] Step S15, obtaining the high-resolution spatiotemporal environmental data set obtained in step S14; applying the empirical mode decomposition method to the data set to extract the intrinsic mode function IMF; calculating the Hilbert-Huang transform for each IMF to obtain instantaneous frequency and amplitude information; inputting the obtained frequency and amplitude information into the multidimensional scaling analysis algorithm for nonlinear dimensionality reduction; and outputting the feature data set after dimensionality reduction.

[0033] By adopting a distributed data acquisition system for synchronous environmental data acquisition and combining the seasonal trend decomposition algorithm for anomaly detection and processing, high-quality acquisition and preprocessing of environmental data are achieved. Specifically, firstly, the temporal consistency and integrity of the data are ensured through distributed acquisition; secondly, the STL algorithm is applied to decompose the data into seasonal, trend and residual components, and the MAD method is used to identify outliers, and the LOESS method is used to repair outliers, effectively removing the noise and outliers in the data; then, the data of different dimensions are unified to the same scale through Z-score standardization, and the multi-resolution analysis characteristics of wavelet transform are used to reduce signal noise, thereby improving the data quality; finally, the Kriging interpolation method is used for spatiotemporal interpolation to obtain a high-resolution environmental data set. This series of processing not only improves the reliability and accuracy of the data, but also realizes the dimensionality reduction and feature extraction of the data through empirical mode decomposition and multidimensional scaling analysis, providing high-quality input data for the subsequent photovoltaic output power prediction. In the photovoltaic off-grid hydrogen production system, this high-quality environmental data processing method greatly improves the system's perception and adaptability to environmental changes, and provides a reliable data basis for the efficient operation of the system.

[0034] According to one aspect of the present application, step S2 is specifically:

[0035] Step S21, obtaining the reduced feature data set obtained in step S1 and the pre-stored historical photovoltaic output data; using the XGBoost algorithm to build a regression model, input the feature data set, and output a preliminary prediction result; using the Prophet algorithm to build a time series prediction model, input the historical photovoltaic output data, and output the time series prediction result; aligning the output results of the XGBoost and Prophet models through the dynamic time warping DTW algorithm; integrating the aligned prediction results using the Bayesian model averaging method to obtain the final predicted photovoltaic output power;

[0036] Step S22, obtaining the predicted photovoltaic output power obtained in step S21 and the current photovoltaic working point voltage and current data collected in real time; constructing an MPPT algorithm model based on particle swarm optimization PSO, and inputting the predicted photovoltaic output power and current working point data into the algorithm model; introducing adaptive inertia weights in the PSO algorithm, and dynamically adjusting the weights according to the current number of iterations t and the maximum number of iterations T_max; applying the Lévy flight strategy in the particle update process to enhance the local search capability; applying constraint processing technology to the results of each iteration to ensure that the search space is within the effective range; iterating the optimization process until convergence or reaching the maximum number of iterations, and outputting the optimized photovoltaic working point voltage;

[0037] Step S23, obtaining the predicted photovoltaic output power obtained in step S21 and the pre-stored photovoltaic array topology information; constructing a graph model of the photovoltaic array, representing each photovoltaic panel as a node in the graph, and representing the connection relationship as an edge; applying a spectral clustering algorithm to the constructed graph model to divide the photovoltaic panels into several groups; applying a minimum spanning tree algorithm to the photovoltaic panels in each group to optimize the connection relationship within the group; using a graph coloring algorithm to assign a color to each photovoltaic panel to solve potential shadow occlusion problems; applying an A* algorithm to calculate the optimal path for array reconstruction; updating the photovoltaic array topology information according to the optimization results to obtain an optimized photovoltaic array topology.

[0038] The photovoltaic output power is predicted by combining the ensemble learning method with time series analysis, and based on this, the MPPT control and photovoltaic array reconstruction are optimized, achieving a significant improvement in system efficiency. The specific technical effects are reflected in: first, the integrated prediction method of XGBoost and Prophet models is used, the prediction results are aligned in combination with the dynamic time warping algorithm, and the Bayesian model averaging method is used for integration, which significantly improves the accuracy of photovoltaic output power prediction; secondly, in the MPPT control optimization, by introducing the particle swarm algorithm with adaptive inertia weight, combined with the Lévy flight strategy to enhance the local search capability, faster and more accurate maximum power point tracking is achieved; finally, in the optimization of photovoltaic array reconstruction, by constructing a graph model of the photovoltaic array, using the spectral clustering algorithm for grouping, and using the minimum spanning tree algorithm to optimize the connection relationship, the impact of local shadows on system efficiency is effectively solved. This multi-level optimization strategy has significant advantages in photovoltaic off-grid hydrogen production systems, which not only improves the power generation efficiency of the system, but also enhances the system's adaptability to environmental changes, especially the ability to operate stably under complex meteorological conditions.

[0039] According to one aspect of the present application, step S3 is specifically:

[0040] Step S31, obtaining pre-stored historical operation data and specification parameters of the electrolytic cell; applying the subspace identification method to process the historical operation data and constructing a preliminary state space model; using the maximum likelihood estimation method to optimize the parameters of the state space model; evaluating models of different orders by the Akaike information criterion and selecting the optimal model order; using the Bootstrap method to resample and fit the selected model multiple times to evaluate the uncertainty of the model parameters; comprehensively considering the model performance and uncertainty, determining the final electrolytic cell dynamic model;

[0041] Step S32, obtaining the dynamic model of the electrolyzer obtained in step S31 and the predicted photovoltaic output power obtained in step S2; constructing a multi-objective optimization problem, wherein the objective function includes maximizing hydrogen production and minimizing energy consumption; using the NSGA-III algorithm to solve the multi-objective optimization problem, and obtaining a series of non-dominated solutions; using a fuzzy decision method to sort and select the non-dominated solutions, and determine the optimal solution; substituting the optimal solution into a dynamic programming algorithm, considering the time series characteristics, and optimizing the timing control strategy; outputting an optimized electrolyzer operating point sequence, including current, voltage, and temperature;

[0042] Step S33, obtain the dynamic model of the electrolytic cell obtained in step S31 and the optimized electrolytic cell operating point sequence obtained in step S32; define the prediction horizon N and control horizon M of the model predictive control MPC; construct a state predictor based on the dynamic model of the electrolytic cell; design a quadratic programming QP problem, and integrate the optimization objective function, system constraints and prediction model into the optimization problem; apply the moving time domain strategy, solve the QP problem in each control cycle, and obtain the optimal control sequence; introduce the Laguerre function to parameterize the control sequence to reduce the computational complexity; integrate the optimized control strategy into the MPC controller to form a complete electrolytic cell MPC controller.

[0043] The dynamic model of the electrolyzer is constructed through system identification technology, and multi-objective optimization and model predictive control are carried out based on it, so as to realize the efficient operation of the electrolyzer system. Specifically, firstly, an accurate dynamic model of the electrolyzer is constructed by combining the subspace identification method with the maximum likelihood estimation, which can accurately describe the dynamic characteristics of the electrolyzer under different working conditions; secondly, the multi-objective optimization algorithm is used to simultaneously consider the two objectives of hydrogen production and energy consumption, and the NSGA-III algorithm is applied to solve it, and the Pareto optimal solution set is obtained. The optimal solution is selected by the fuzzy decision method, and the optimization of the electrolyzer operation efficiency is realized; finally, the model predictive controller is designed based on the optimization results, and the precise control of the electrolyzer is realized through rolling optimization and real-time feedback. This model-based optimization control method is particularly important in photovoltaic off-grid hydrogen production systems, because it can dynamically adjust the operating parameters of the electrolyzer according to the fluctuation characteristics of photovoltaic power generation, ensuring that the system maintains an efficient and stable hydrogen preparation process under changing power generation conditions, while minimizing energy loss.

[0044] According to one aspect of the present application, step S4 is specifically:

[0045] Step S41, obtaining the voltage, current and temperature data of the energy storage system collected in real time; constructing the Thevenin equivalent circuit model of the battery; designing a Sigma point Kalman filter (UKF), inputting the voltage and current data into the filter for state estimation; introducing an adaptive attenuation factor, and dynamically adjusting the filter parameters according to the estimation error; for the strong nonlinear region in the SOC estimation, switching to the particle filter algorithm for processing; combining the results of the UKF and particle filter, and outputting the final battery SOC and SOH estimation values;

[0046] Step S42, obtaining the battery SOC and SOH estimation values ​​obtained in step S41, and the pre-stored battery history usage data; constructing a multi-state semi-Markov model of battery degradation, defining degradation states and state transition rules; using the expectation maximization EM algorithm to estimate the parameters of the semi-Markov model; applying the Monte Carlo simulation method to generate multiple groups of degradation trajectories based on the estimated model parameters; using the generated degradation trajectories to calculate the probability distribution of the remaining battery life; applying the Kaplan-Meier estimator to correct the prediction results to improve the prediction accuracy; outputting the corrected battery remaining life distribution;

[0047] Step S43, obtain the predicted photovoltaic output power obtained in step S2, the optimized electrolyzer operating point sequence obtained in step S3, and the battery SOC and SOH estimates obtained in step S41; define the system state space and action space, and use the charging and discharging power of the energy storage system as the action variable; construct a Markov decision process MDP model considering uncertainty; define an objective function based on maximizing the expected total return; apply the temporal difference learning algorithm to solve the MDP model and obtain a preliminary optimal strategy; introduce a risk sensitivity factor to adjust the strategy to balance benefits and risks; output the optimal scheduling strategy for the energy storage system considering risks.

[0048] The state of the energy storage system is estimated through the Kalman filter algorithm, the life is predicted by combining the semi-Markov degradation model, and the random dynamic programming is applied for optimization scheduling, so as to achieve efficient management of the energy storage system. Specifically, firstly, the high-precision estimation of the SOC and SOH of the energy storage system is achieved by combining the Sigma point Kalman filter with the particle filter algorithm; secondly, the multi-state semi-Markov model is used to describe the battery degradation process, and the remaining battery life is predicted by the Monte Carlo simulation method, which improves the system reliability; finally, the energy storage scheduling is optimized based on the time series deep reinforcement learning algorithm, and the risk sensitivity factor is introduced to achieve effective control of the system operation risk. This comprehensive optimization method plays a key role in the photovoltaic off-grid hydrogen production system, which can effectively balance the power generation fluctuations, improve the energy utilization efficiency of the system, extend the life of the energy storage system, and ensure the safe and reliable operation of the system.

[0049] According to one aspect of the present application, it also includes:

[0050] Step S5, obtaining the optimization results and controller parameters of each subsystem obtained in steps S2, S3 and S4, and constructing a multi-time scale coordinated control framework; based on the system model parameters and controller parameters, applying Lyapunov stability theory to perform system stability analysis, and obtaining the system stability region and stability margin; according to the system model and real-time fault diagnosis results, constructing a data-driven adaptive fault-tolerant control strategy, and obtaining an adaptive fault-tolerant controller.

[0051] According to one aspect of the present application, step S5 is specifically:

[0052] Step S51, obtain the optimization results and controller parameters of each subsystem obtained in steps S2, S3 and S4; construct a three-layer control architecture based on time scale separation, including day-ahead scheduling, hourly coordination and real-time balance; design a day-ahead scheduling algorithm based on model predictive control at the upper layer to handle the optimization problem at the 24-hour scale; implement rolling time domain optimization at the middle layer, update the optimization results once an hour, and coordinate the operation of each subsystem; design a robust H∞ controller at the bottom layer to achieve millisecond-level real-time power balance; integrate the three-layer control strategies to form a complete multi-time scale coordinated control framework;

[0053] Step S52, obtaining system model parameters and controller parameters obtained in step S51; constructing a complete state space expression of the system, including the dynamic characteristics of the photovoltaic system, the electrolyzer and the energy storage system; designing a composite Lyapunov function, considering multiple operation modes of the system; applying the linear matrix inequality (LMI) technology to transform the stability problem into a convex optimization problem; solving the LMI problem to obtain a parameter range that satisfies the stability condition; using the Lyapunov exponent analysis method to evaluate the transient stability of the system under disturbance; and combining the stability analysis results to determine the stability region and stability margin of the system;

[0054] Step S53, obtain system model parameters and real-time fault diagnosis results; apply local linear embedding (LLE) algorithm to process historical fault data and construct fault feature space; construct an online fault identifier based on extreme learning machine (ELM) to classify the current system state in real time; construct multiple alternative controller models according to different fault types; design a fuzzy reasoning system to calculate the weight of each alternative controller according to the fault identification result and system state; realize smooth switching of controllers to ensure continuous and stable operation of the system under fault conditions; introduce a sliding mode observer to estimate and compensate for unmodeled dynamics and external disturbances; integrate fault identification, controller switching and disturbance compensation modules to form a complete adaptive fault-tolerant controller.

[0055] By designing a multi-time scale coordinated control framework and combining Lyapunov stability theory and data-driven adaptive fault-tolerant control strategy, the global optimal control of the system is achieved. The specific effects are as follows: first, the three-layer control architecture is adopted to achieve all-round optimization from day-ahead scheduling to real-time control; second, through Lyapunov stability analysis, the stable operation of the system under various disturbances is ensured; finally, through data-driven adaptive fault-tolerant control strategy, the robustness and reliability of the system are improved. This multi-level coordinated control method is of great significance in photovoltaic off-grid hydrogen production systems. It can achieve coordinated operation among various subsystems of the system, ensuring that the system can maintain stable operation in the face of external disturbances and internal faults, while maintaining high energy conversion efficiency.

[0056] According to one aspect of the present application, step S15 is specifically:

[0057] Step S151, obtaining the high-resolution spatiotemporal environmental data set obtained in step S14; applying white noise auxiliary decomposition to each time series in the data set to generate multiple copies with different noise levels; performing empirical mode decomposition on each noise auxiliary copy to obtain a series of IMF sets; calculating the mean of the IMFs corresponding to all noise auxiliary copies as the final IMF; applying singular spectrum analysis to the final IMF to remove residual noise; and outputting the IMF set;

[0058] Step S152, obtaining the improved IMF set outputted in step S151; applying Hilbert transform to each IMF to calculate the analytical signal; calculating the instantaneous frequency based on the analytical signal, using a differential algorithm to improve the frequency estimation accuracy; calculating the instantaneous amplitude, using an adaptive window smoothing method to reduce the endpoint effect; constructing a Hilbert spectrum to represent the time-frequency-energy distribution; applying an adaptive threshold method to extract significant features in the Hilbert spectrum; outputting the instantaneous frequency, instantaneous amplitude and significant feature data of the IMF;

[0059] Step S153, obtain the IMF instantaneous frequency, instantaneous amplitude and significant feature data output by step S152; construct a high-dimensional feature space with time, frequency, amplitude and significant features as dimensions; apply the t-SNE algorithm to initialize the embedding space, and use the PCA result as the initial point distribution; introduce an adaptive perplexity parameter (perplexity parameter) and dynamically adjust it according to the local data density; use the Barnes-Hut approximation to optimize the computational complexity; apply the early exaggeration technique to accelerate convergence; iterate the optimization until the stopping criterion is reached; output the data point coordinates in the low-dimensional embedding space as the final reduced dimensionality feature data set.

[0060] By integrating the empirical mode decomposition method with the t-SNE algorithm for data dimensionality reduction and feature extraction, efficient processing and feature expression of environmental data are achieved. Specifically, white noise-assisted decomposition is first used to generate multiple copies, and the IMF set is obtained through EMD decomposition. The singular spectrum analysis is used to remove the residual noise, which improves the accuracy of decomposition; secondly, the Hilbert transform is applied to each IMF to calculate the instantaneous frequency and amplitude, and the differential algorithm and adaptive window smoothing method are used to improve the frequency estimation accuracy; finally, the t-SNE algorithm is used for nonlinear dimensionality reduction, and the adaptive perplexity parameter and Barnes-Hut approximation are introduced to optimize the computational complexity. This data processing method has significant advantages in photovoltaic off-grid hydrogen production systems. It can effectively capture the multi-scale characteristics and nonlinear relationships in environmental data, provide more accurate feature expressions for the system's predictive control, and reduce the computational burden through dimensionality reduction processing, thereby improving the system's real-time response capability.

[0061] According to one aspect of the present application, step S22 is specifically:

[0062] Step S221, obtaining the predicted photovoltaic output power and current photovoltaic working point data of step S21; initializing the particle swarm of the HQPSO algorithm, each particle is represented as a quantum bit string; defining a fitness function, and calculating the fitness value of each particle according to the predicted power and the current working point; applying quantum measurement operations, mapping quantum state particles to the classical search space; outputting the initial particle swarm and its fitness value;

[0063] Step S222, obtaining the initial particle swarm and fitness value of step S221; constructing an adaptive quantum revolving door, dynamically adjusting the rotation angle according to the current number of iterations and the particle fitness; applying the quantum revolving door operation to each particle to update its quantum state; introducing quantum entanglement operation to enhance information exchange between particles; applying the Lévy flight strategy, generating a flight step length according to the Lévy distribution, and updating the particle position; outputting the updated particle swarm;

[0064] Step S223, obtain the particle swarm updated in step S222; apply constraint processing technology and use adaptive penalty function method to handle out-of-bounds particles; re-evaluate the fitness value of each particle; update the individual optimal solution and the global optimal solution; apply the elite retention strategy to ensure that the optimal solution is not lost during the iteration process; check the convergence condition or the maximum number of iterations; if the stop condition is not met, return to step S222 to continue iterating; after the stop condition is met, output the optimized photovoltaic operating point voltage.

[0065] The MPPT control optimization was performed by hybrid quantum-particle swarm optimization algorithm, and efficient tracking control of the photovoltaic system was achieved. The specific effects are as follows: first, the global search capability of the algorithm is enhanced through quantum bit encoding and quantum rotating gate operation; second, the introduction of adaptive quantum rotating gate and quantum entanglement operation improves the efficiency of information exchange between particles; third, the particle position is updated through the Lévy flight strategy, which enhances the local search capability; finally, the adaptive penalty function method is used to handle out-of-bounds particles to ensure the effectiveness of the search space. In the photovoltaic off-grid hydrogen production system, this optimization algorithm can quickly and accurately track the maximum power point, especially in the case of rapid changes in light and temperature, showing stronger adaptability and higher tracking efficiency, effectively improving the energy conversion efficiency of the system.

[0066] According to one aspect of the present application, step S31 can also be: obtaining pre-stored historical operation data and specification parameters of the electrolytic cell; applying the recursive subspace identification method (RSID) to process the historical operation data and construct a preliminary nonlinear state space model; using the enhanced maximum likelihood estimation (EMLE) method to optimize the parameters of the state space model; applying Hankel singular value decomposition (HSVD) to select and reduce the order of the model; using the adaptive Bootstrap method to evaluate the uncertainty of the model parameters; comprehensively considering the model performance and uncertainty, and determining the final electrolytic cell dynamic model.

[0067] Specifically, step S311, obtain the historical operation data and specification parameters of the electrolytic cell; preprocess the data, including removing outliers and standardizing the data; construct a data matrix, organize the input and output data into the form of a Hankel matrix; apply singular value decomposition (SVD) to calculate the observability and controllability matrix of the system; based on the SVD results, preliminarily estimate the order of the system; output the preprocessed data matrix and the preliminary system order estimate.

[0068] Step S312, obtain the preprocessed data matrix and preliminary order estimate of step S311; apply the recursive subspace identification algorithm, and use the sliding time window to gradually update the state space model; introduce the forgetting factor to enhance the adaptability of the algorithm to time-varying systems; use regularization technology to improve the stability and generalization ability of the model; update the model parameters in each recursive step; output the preliminary nonlinear state space model.

[0069] Step S313, obtain the preliminary nonlinear state space model of step S312; apply the enhanced maximum likelihood estimation method to construct an optimization problem considering parameter constraints; use the particle swarm algorithm to solve the optimization problem and obtain the optimal estimate of the model parameters; apply the adaptive step size strategy to improve the convergence speed and accuracy of the algorithm; iterate the optimization process until the convergence conditions are met; output the optimized model parameters.

[0070] Step S314, obtaining the model parameters optimized in step S313; applying the Hankel singular value decomposition method to analyze the singular value distribution of the model; using the information criterion (such as AIC) to determine the optimal model order; based on the determined order, performing model reduction processing; applying the adaptive Bootstrap method to evaluate the uncertainty of the model parameters through multiple resampling and fitting; combining the model performance and uncertainty analysis results to determine the final electrolytic cell dynamic model; outputting the final electrolytic cell dynamic model and its uncertainty estimate.

[0071] A high-precision dynamic model of the electrolyzer was constructed by recursive subspace identification method and enhanced maximum likelihood estimation. The specific effects are as follows: first, through data preprocessing and Hankel matrix construction, high-quality basic data is provided for model identification; second, the recursive subspace identification algorithm is combined with the sliding time window to realize the dynamic update of the model parameters; then, the model parameters are optimized by enhanced maximum likelihood estimation and particle swarm algorithm; finally, Hankel singular value decomposition is applied to reduce the model order, and the model uncertainty is evaluated by the adaptive Bootstrap method. In the photovoltaic off-grid hydrogen production system, this high-precision electrolyzer model is crucial to the efficient operation of the system. It can not only accurately predict the dynamic response of the electrolyzer, but also provide a reliable model basis for control optimization, thereby realizing efficient and safe operation of the electrolyzer.

[0072] According to one aspect of the present application, step S32 may also be:

[0073] Step S32, obtaining the dynamic model of the electrolyzer obtained in step S31 and the predicted photovoltaic output power obtained in step S2; defining a multi-objective optimization problem, wherein the objective functions include maximizing hydrogen production, minimizing energy consumption, and minimizing system stress; applying a multi-objective differential evolution algorithm (IMODEA) to solve the multi-objective optimization problem, and obtaining a series of non-dominated solutions; using an adaptive fuzzy decision method to sort and select the non-dominated solutions, and determine the optimal solution; substituting the optimal solution into a timing dynamic programming algorithm, considering the time series characteristics, and optimizing the timing control strategy; outputting an optimized electrolyzer operating point sequence, including parameters such as current, voltage, and temperature.

[0074] Specifically, step S321, obtain the dynamic model of the electrolyzer and predict the photovoltaic output power; construct the objective function of the multi-objective optimization problem, including the hydrogen production function, the energy consumption function and the system stress function; define the constraints of the problem, including the electrolyzer operation restrictions and the safety margin; initialize the population of the IMODEA algorithm, each individual represents a set of possible electrolyzer operation parameters; evaluate the fitness of the initial population, and calculate the value of each objective function; output the initial population and its fitness value.

[0075] Step S322, obtain the initial population and fitness value of step S321; apply differential evolution operations, including adaptive crossover and mutation strategies; introduce density-based selection mechanism to maintain population diversity; use fast non-dominated sorting algorithm to stratify the population; apply crowding distance calculation to distinguish individuals in the same non-dominated layer; select elite individuals to form the next generation population; repeat the evolution process until the stopping criterion is reached; output the final non-dominated solution set.

[0076] Step S323, obtain the non-dominated solution set of step S322; construct an adaptive fuzzy decision system and define a fuzzy rule set; apply fuzzy rules to each non-dominated solution and calculate its comprehensive score; sort the solutions according to the score; select the solution with the highest comprehensive score as the optimal solution; input the optimal solution into the time series dynamic programming algorithm; consider the time correlation and system dynamic characteristics, optimize the control sequence; output the final optimized electrolytic cell operating point sequence.

[0077] The electrolyzer operation is optimized through a multi-objective differential evolution algorithm, achieving efficient and energy-saving operation of the system. Specifically, firstly, by constructing a multi-objective optimization problem, the three objectives of hydrogen production, energy consumption and system stress are considered at the same time; secondly, differential evolution operations and density-based selection mechanisms are used to maintain population diversity; then, the non-dominated solutions are sorted and selected through an adaptive fuzzy decision system; finally, the optimal solution is substituted into the timing dynamic programming algorithm to optimize the timing control strategy. This optimization method is of great significance in photovoltaic off-grid hydrogen production systems. It can dynamically adjust the operating parameters of the electrolyzer according to the fluctuation characteristics of photovoltaic power generation, while ensuring hydrogen production efficiency, minimizing energy consumption and extending equipment life.

[0078] Specifically, step S41 can also be: obtaining the voltage, current and temperature data of the energy storage system collected in real time; constructing an adaptive unscented Kalman filter (IAUKF) model, including the nonlinear dynamic characteristics of the battery; designing a state-dependent noise covariance estimator to dynamically adjust process noise and observation noise; introducing a multi-model mechanism to handle the behavioral differences of the battery under different working conditions; implementing a fault detection module based on residual statistics to improve the reliability of state estimation; and combining the IAUKF and fault detection results to output the final battery SOC and SOH estimates.

[0079] Specifically, step S411, obtain real-time voltage, current and temperature data of the energy storage system; construct the nonlinear state space equation of the battery based on the Thevenin equivalent circuit model; design a Sigma point sampling strategy to generate a set of Sigma points; update each Sigma point in time to predict the state at the next moment; calculate the mean and covariance of the predicted state; output the predicted state estimate and covariance matrix.

[0080] Step S412, obtain the predicted state estimate and covariance matrix of step S411; based on the current measurement value, update the Sigma point; calculate the mean and covariance of the measurement prediction; calculate the cross-covariance between the state and the measurement; apply the Kalman gain to update the state estimate; design a state-dependent noise covariance estimator, dynamically adjust the process noise and observation noise covariance according to the innovation sequence; output the updated state estimate and covariance matrix.

[0081] Step S413, obtaining the updated state estimation and covariance matrix of step S412; implementing a multi-model mechanism, including models such as normal charging, normal discharging, overcharging and over-discharging; calculating the likelihood of each model; using an interacting multi-model (IMM) algorithm to fuse the estimation results of multiple models; applying an adaptive factor to dynamically adjust the fusion weight according to the model likelihood; and outputting the fused state estimation result.

[0082] Step S414, obtain the fusion state estimation result of step S413; construct a fault detection module based on residual statistics; calculate the statistical characteristics of the innovation sequence; apply the generalized likelihood ratio test (GLRT) method to detect anomalies; if anomalies are detected, trigger the fault diagnosis logic; adjust the state estimation strategy according to the diagnosis result; combine the normal estimation result and the fault handling result to output the final battery SOC and SOH estimation values.

[0083] The state of the energy storage system is estimated through an adaptive unscented Kalman filter model, achieving high-precision battery state monitoring. The specific effects are as follows: First, by constructing the Thevenin equivalent circuit model and designing the Sigma point sampling strategy, the accuracy of state prediction is improved; second, through the state-dependent noise covariance estimator, the process noise and observation noise are dynamically adjusted; third, the behavioral differences of the battery under different working conditions are processed through a multi-model mechanism; finally, the reliability of state estimation is improved through a fault detection module based on residual statistics. This high-precision state estimation method plays a key role in photovoltaic off-grid hydrogen production systems. It can monitor the working status of the energy storage system in real time and accurately, and provide reliable data support for the safe operation and optimal control of the system.

[0084] Specifically, step S43 can also be: obtaining the predicted photovoltaic output power obtained in step S2, the optimized electrolyzer operating point sequence obtained in step S3, and the battery SOC and SOH estimates obtained in step S41; constructing a Markov decision process (MDP) model considering uncertainty; applying the time series deep reinforcement learning algorithm (ITDRL) to solve the MDP model; introducing risk-sensitive strategy optimization to balance system benefits and operational risks; realizing an exploration-utilization balance mechanism based on Thompson sampling; integrating a model predictive control (MPC) framework to improve control accuracy; and outputting the optimal scheduling strategy for the energy storage system considering risks.

[0085] Step S431, obtain the predicted photovoltaic output power, optimize the electrolyzer working point sequence and the battery state estimation value; construct the MDP model, define the state space, action space and reward function; introduce uncertainty modeling, and use probability distribution to represent the randomness of photovoltaic output and load demand; design a state transition probability estimator based on historical data; initialize the value function and the policy network; output the initialized MDP model and neural network parameters.

[0086] Step S432, obtain the initialized MDP model and network parameters of step S431; implement the time series deep reinforcement learning algorithm, including the dual Q network structure and priority experience replay; design the recurrent neural network (RNN) layer to capture the time series characteristics; apply the gradient clipping technology to prevent gradient explosion; implement the exploration strategy based on Thompson sampling to balance exploration and utilization; regularly update the target network to improve learning stability; output the updated Q network and policy network.

[0087] Step S433, obtain the Q network and policy network updated in step S432; construct a risk-sensitive policy optimization module, and introduce conditional value at risk (CVaR) as a risk metric; design a multi-objective optimization problem, while considering expected return and risk control; apply the policy gradient method to solve the risk-sensitive optimization problem; implement an adaptive learning rate adjustment mechanism to improve convergence performance; output the risk-adjusted policy network.

[0088] Step S434, obtain the risk adjustment strategy network of step S433; integrate the model predictive control framework, and use the system dynamic model for short-term prediction; construct a rolling optimization problem, and use the reinforcement learning strategy as a reference trajectory; apply a real-time optimization solver, such as the interior point method, to solve the MPC problem; implement a feedback correction mechanism to adjust the control output according to the actual system state; combine the reinforcement learning strategy and the MPC output to generate the final energy storage system scheduling strategy; output the optimal scheduling strategy for the energy storage system that takes risk and precise control into account.

[0089] Through the time-series deep reinforcement learning algorithm and risk-sensitive strategy optimization, the intelligent dispatching control of the energy storage system is realized. The specific effects are reflected in: first, a Markov decision process model considering uncertainty is constructed, and the dynamic characteristics of the system are fully described by defining a reasonable state space, action space and reward function; second, the learning efficiency and strategy quality are improved through the time-series deep reinforcement learning algorithm, including the dual Q network structure and the priority experience replay mechanism; then, the conditional value at risk (CVaR) is introduced as a risk metric, and the effective control of the system operation risk is achieved through risk-sensitive strategy optimization; finally, the integrated model predictive control framework is used to improve the control accuracy through rolling optimization and feedback correction mechanism. This comprehensive optimization method has significant advantages in photovoltaic off-grid hydrogen production systems. It can intelligently adjust the charging and discharging behavior of the energy storage system according to the randomness of photovoltaic power generation and the changes in load demand, which not only ensures the energy balance of the system, but also realizes the effective control of risks, and significantly improves the operation reliability and economy of the system.

[0090] According to one aspect of the present application, step S5 may also be:

[0091] Step S51, obtain the optimization results and controller parameters of each subsystem obtained in steps S2, S3 and S4; construct a multi-layer adaptive coordinated control framework based on time scale separation, including long-term planning, medium-term scheduling and real-time control; realize demand forecasting and resource allocation based on deep learning in the long-term planning layer; apply the distributed model predictive control algorithm in the medium-term scheduling layer; design an adaptive robust controller in the real-time control layer; implement an inter-layer coordination mechanism to ensure the consistency of control strategies at each layer; integrate multi-layer control strategies to form a complete multi-time scale coordinated control framework.

[0092] Specifically,

[0093] Step S511, obtain the historical operation data and optimization results of each subsystem; build a long-term demand forecasting model, apply the long short-term memory network (LSTM) combined with the attention mechanism; realize multi-step forecasting, considering seasonality and long-term trends; design a resource allocation algorithm based on the forecast results, and optimize the long-term system configuration; output the long-term demand forecasting results and resource allocation plan.

[0094] Step S512, obtain the long-term prediction and allocation plan of step S511; construct a distributed model predictive control (DMPC) framework and divide the system into multiple subsystems; design a local MPC controller for each subsystem; implement a subsystem coordination algorithm based on the alternating direction method of multipliers (ADMM); introduce an adaptive weight update mechanism to improve the convergence speed; design a distributed optimization solver, such as the primal-dual interior point method; and output a medium-term scheduling strategy.

[0095] Step S513, obtain the medium-term scheduling strategy and real-time system status of step S512; design an adaptive robust controller, taking into account system uncertainty and external disturbances; construct a control architecture based on L1 adaptive control theory; implement a fast parameter estimation algorithm to dynamically update controller parameters; design an anti-disturbance observer to estimate and compensate for unmodeled dynamics; introduce sliding mode control elements to improve system robustness; and output real-time control instructions.

[0096] Step S514, obtain the control outputs of steps S511, S512 and S513; design an inter-layer coordination mechanism to ensure the consistency of long-term planning, medium-term scheduling and real-time control; implement a decision fusion algorithm based on fuzzy logic to coordinate control objectives of different time scales; build a feedback compensation mechanism to adjust the long-term strategy according to the short-term control effect; design an adaptive learning module to continuously optimize the coordination strategy; integrate multi-layer control outputs to generate the final coordinated control instructions; output a complete multi-time scale coordinated control framework.

[0097] By constructing a multi-layer adaptive coordinated control framework, the global optimization control of the system is realized. The specific effects are as follows: First, in the long-term planning layer, accurate demand forecasting is achieved through the LSTM network combined with the attention mechanism, providing a basis for optimal resource allocation; second, in the medium-term scheduling layer, the coordinated optimization between subsystems is achieved through the distributed model predictive control framework, and the convergence speed is improved through the ADMM algorithm and the adaptive weight update mechanism; then, in the real-time control layer, the anti-interference ability of the system is improved by designing an adaptive robust controller, combining L1 adaptive control theory and sliding mode control elements; finally, the consistency of control strategies at different time scales is ensured through the inter-layer coordination mechanism, and the unified coordination of control objectives is achieved through the fuzzy logic decision fusion and feedback compensation mechanism. This multi-level coordinated control method plays a core role in the photovoltaic off-grid hydrogen production system, and can achieve all-round optimization from long-term planning to real-time control, significantly improving the overall operation efficiency and reliability of the system.

[0098] In another embodiment of the present application, step S53 may also be:

[0099] Step S53, obtain system model parameters and real-time fault diagnosis results; design a multimodal fault diagnosis system based on deep learning; construct a hierarchical adaptive fault-tolerant control architecture; implement an online controller reconstruction algorithm based on reinforcement learning; design a fuzzy reasoning system for smooth controller switching; introduce a distributed estimator network to achieve robust state estimation; integrate fault diagnosis, controller reconstruction and state estimation modules to form a complete adaptive fault-tolerant control system.

[0100] Specifically, step S531, obtain multi-source sensor data and system operating parameters; build a multimodal deep learning model, such as a hybrid structure of a convolutional neural network (CNN) and a recurrent neural network (RNN); realize automatic feature extraction and fusion; design a fault pattern recognition algorithm based on an attention mechanism; introduce an active learning strategy to continuously update the fault model; output real-time fault diagnosis results, including fault type, location and severity.

[0101] Step S532, obtain the fault diagnosis result and system model of step S531; construct a hierarchical adaptive fault-tolerant control architecture, including a fault detection and isolation (FDI) layer, a control reconstruction layer and an execution layer; implement a fast fault detection algorithm based on a sliding mode observer in the FDI layer; design a fault isolation logic based on an adaptive threshold; implement virtual actuator and virtual sensor technology in the control reconstruction layer; design a reconstruction control law based on Lyapunov stability theory; and output a preliminary fault-tolerant control strategy.

[0102] Step S533, obtain the preliminary fault-tolerant control strategy and real-time system status of step S532; implement an online controller reconstruction algorithm based on deep deterministic policy gradient (DDPG); construct an action-evaluation network structure to realize control strategy learning in continuous action space; introduce an experience replay mechanism to improve learning efficiency; design a model-based imagination update strategy to enhance adaptability to rare faults; implement a progressive learning framework to gradually improve controller performance; output the optimized adaptive control strategy.

[0103] Step S534, obtain the adaptive control strategy of step S533 and the fault diagnosis result of step S531; design a fuzzy reasoning system, define the fault degree and controller switching rules; implement an adaptive fuzzy membership function, and dynamically adjust the switching strategy; build a smooth switching mechanism to avoid system disturbances caused by controller switching; design an advance switching logic based on state prediction to improve the system response speed; output the control instructions after smooth switching.

[0104] Step S535, obtain local state estimation results of multiple subsystems; build a distributed estimator network to realize information exchange and collaboration between subsystems; design a distributed Kalman filtering algorithm based on consistency theory; introduce an event trigger mechanism to reduce the communication burden; implement adaptive consistency gain adjustment to improve the estimation convergence speed; design a robust information fusion algorithm to handle abnormal data and communication loss; output a globally consistent state estimation result.

[0105] Step S536, obtain the fault diagnosis result of step S531, the adaptive control strategy of step S533, the smooth switching control instruction of step S534 and the state estimation result of step S535; design an integrated decision-making module to coordinate the outputs of fault diagnosis, control reconstruction and state estimation; implement a confidence-based decision fusion algorithm to comprehensively consider multi-source information; build an adaptive weight allocation mechanism to dynamically adjust the influence of each module; design safety supervision logic to ensure that the decision output meets the system safety constraints; integrate the outputs of all modules to generate the final fault-tolerant control instruction; output a complete adaptive fault-tolerant control system and its control strategy.

[0106] By designing a multimodal fault diagnosis system based on deep learning and a hierarchical adaptive fault-tolerant control architecture, the reliable operation of the system is achieved. The specific effects are as follows: first, by constructing a multimodal deep learning model and combining the hybrid structure of CNN and RNN, the automatic extraction and fusion of fault features are realized; second, through the hierarchical adaptive fault-tolerant control architecture, including fast fault detection based on sliding mode observer and control reconstruction based on virtual actuator, the fault adaptability of the system is improved; then, through the online controller reconstruction algorithm based on DDPG, the dynamic optimization of the control strategy is realized; third, through the fuzzy inference system and smooth switching mechanism, the smoothness of controller switching is ensured; finally, through the distributed estimator network and robust information fusion algorithm, reliable state estimation is achieved. This comprehensive fault diagnosis and fault-tolerant control method is of great significance in photovoltaic off-grid hydrogen production systems. It can detect and diagnose system faults in a timely manner, and maintain the stable operation of the system through adaptive fault-tolerant control strategies, which significantly improves the reliability and robustness of the system.

[0107] In another embodiment of the present application, step S433 may also be:

[0108] Step S433, obtain the Q network and policy network updated in step S432; construct a risk-sensitive policy optimization module, and introduce conditional value at risk (CVaR) as a risk metric; design a multi-objective optimization problem, while considering expected return and risk control; apply the policy gradient method to solve the risk-sensitive optimization problem; implement an adaptive learning rate adjustment mechanism to improve convergence performance; output the risk-adjusted policy network.

[0109] Specifically, step S4331, obtain the Q network and strategy network parameters of step S432; construct a risk measurement model to calculate the conditional value at risk (CVaR) under a given strategy; define a risk aversion factor α to balance expected return and risk; construct a composite objective function J(θ) = (1-α)E[R] - α·CVaR, where θ is a strategy parameter and R is a cumulative reward; and output a risk-aware objective function.

[0110] Step S4332, obtain the risk perception objective function of step S4331; calculate the gradient ∇θJ(θ) of the objective function with respect to the policy parameters; apply the policy gradient method, including the variance reduction technique:

[0111] 1. Use the state value function as a baseline to reduce the variance of the gradient estimate;

[0112] 2. Introduce generalized advantage estimation (GAE) to balance bias and variance;

[0113] 3. Apply importance sampling and use off-strategy data to improve sampling efficiency;

[0114] Output the optimized gradient estimate ∇θJ'(θ).

[0115] Step S4333, obtain the optimized gradient estimate of step S4332; implement the adaptive learning rate adjustment mechanism:

[0116] 1. Initialize the Adam optimizer and set the initial learning rate η0 and momentum parameters β1, β2;

[0117] 2. Calculate the exponential moving average of the gradient mt and the exponential moving average of the squared gradient vt;

[0118] 3. Apply bias correction to obtain m*t and v*t;

[0119] 4. Update learning rate: ηt = η0 * sqrt(1-β2 t ) / (1-β1 t );

[0120] 5. Update policy parameters: θt+1 = θt - ηt * m*t / (sqrtv*t + ε);

[0121] Output the updated policy network parameters θt+1.

[0122] Step S4334, obtain the policy network parameters updated in step S4333; perform policy evaluation:

[0123] 1. Use the updated strategy to perform Monte Carlo sampling to generate trajectory samples;

[0124] 2. Calculate the cumulative reward and CvaR of the sample trajectory;

[0125] 3. Evaluate the performance metrics of the new strategy, including average return and risk measures;

[0126] If the performance improvement is not significant or the maximum number of iterations is reached, the optimization is stopped; otherwise, return to step S4331 to continue the optimization.

[0127] Output the final risk-adjusted strategy network parameters θ*.

[0128] In another embodiment of the present application, step S531 may also be:

[0129] Step S531, obtain multi-source sensor data and system operating parameters; build a multimodal deep learning model, such as a hybrid structure of a convolutional neural network (CNN) and a recurrent neural network (RNN); realize automatic feature extraction and fusion; design a fault pattern recognition algorithm based on an attention mechanism; introduce an active learning strategy to continuously update the fault model; output real-time fault diagnosis results, including fault type, location, and severity.

[0130] Step S5311, obtain multi-source sensor data, including time series data and spectrum data; apply sliding window segmentation to the time series data to generate fixed-length data segments; perform time-frequency transform on the spectrum data, such as short-time Fourier transform (STFT); standardize all input data; and output the preprocessed multimodal data set.

[0131] Step S5312, obtain the preprocessed data set of step S5311; construct a CNN module:

[0132] 1. Design multi-scale convolutional layers and use different kernel sizes to capture features of different scales;

[0133] 2. Introduce dilated convolution to increase the receptive field and capture long-range dependencies;

[0134] 3. Apply batch normalization and ReLU activation function to improve training stability and nonlinear expression ability;

[0135] 4. Use global average pooling layer to reduce the number of parameters;

[0136] Output the spatial feature representation extracted by CNN.

[0137] Step S5313, obtain the preprocessed data set of step S5311; construct the RNN module:

[0138] 1. Use gated recurrent units (GRU) to alleviate the gradient vanishing problem;

[0139] 2. Implement a bidirectional GRU structure to capture bidirectional temporal dependencies;

[0140] 3. Application layer normalization to improve training speed and generalization ability;

[0141] 4. Introducing residual connections to facilitate deep network training;

[0142] Output the time series feature representation extracted by RNN.

[0143] Step S5314: Obtain the CNN features of step S5312 and the RNN features of step S5313; design a feature fusion module:

[0144] 1. Use the self-attention mechanism to calculate the correlation weights between features;

[0145] 2. Apply multi-head attention to extract feature relationships from different subspaces;

[0146] 3. Implement feature gating mechanism to dynamically adjust the importance of different modal features;

[0147] 4. Use a feedforward neural network to further integrate the fused features;

[0148] Output the fused multimodal feature representation.

[0149] Step S5315, obtaining the fusion features of step S5314; implementing the fault mode recognition module:

[0150] 1. Design a multi-task learning structure to simultaneously predict fault type, location, and severity;

[0151] 2. Use soft attention mechanism to assign attention weights to different tasks;

[0152] 3. Apply label smoothing technology to improve the model's robustness to noise;

[0153] 4. Introduce the focal loss function to solve the problem of category imbalance;

[0154] Output fault diagnosis results, including probability distribution of fault type, location and severity.

[0155] Step S5316: Obtain the diagnosis result of step S5315 and the newly collected unlabeled data; implement the active learning update strategy:

[0156] 1. Use model uncertainty (e.g., prediction entropy) as a sampling criterion;

[0157] 2. Select samples with the highest uncertainty for manual labeling;

[0158] 3. Use the newly annotated samples for incremental learning and update model parameters;

[0159] This update process is performed regularly to continuously improve the diagnostic capabilities and adaptability of the model.

[0160] Output the updated fault diagnosis model.

[0161] According to another aspect of the present application, a photovoltaic off-grid hydrogen production system is also provided, comprising:

[0162] at least one processor; and,

[0163] a memory communicatively connected to at least one processor; wherein,

[0164] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the photovoltaic off-grid hydrogen production method of any of the above technical solutions.

[0165] In another embodiment of the present application, other hardware structures for photovoltaic off-grid hydrogen production are described. Figure 2 As shown, the specific structure is as follows: photovoltaic cell, DC bus, energy storage battery, hydrogen production electrolyzer, DC-DC converter 1, DC-DC converter 2 and DC-DC converter 3. DC-DC converter 1 is connected between the photovoltaic cell and the DC bus to provide input energy for the system. DC-DC converter 2 is connected between the energy storage battery and the DC bus to stabilize the DC bus, and DC-DC converter 3 is connected between the hydrogen production electrolyzer and the DC bus to transfer electric energy for hydrogen production.

[0166] The DC-DC converter 1 is a step-down DC / DC converter or a step-up DC / DC converter. The DC-DC converter 2 is a step-down DC / DC converter. The DC-DC converter 3 is a bidirectional DC / DC converter.

[0167] Photovoltaic cells are used to generate electricity; DC busbars are used to collect the DC power output by photovoltaic cells and transmit the DC power to hydrogen production electrolyzers and energy storage batteries;

[0168] The hydrogen production electrolyzer is used to produce hydrogen by electrolysis; the energy storage battery is used to provide electricity to the hydrogen production electrolyzer or store electricity from photovoltaic cells; the DC bus is used to converge the DC power output by photovoltaic cells and transmit it to the hydrogen production electrolyzer for power supply. Compared with the AC bus structure, the DC network structure is simpler and has fewer power conversion links, so the power utilization efficiency is high and the system operation economy is high.

[0169] Figure 3This is a structural block diagram of the control method of the photovoltaic off-grid hydrogen production system of this application. The operating environment of the photohydrogen storage system is complex, and the diversity of photovoltaic changes in the system makes the system control more difficult. This embodiment sets a hierarchical control strategy. The upper layer makes corresponding instructions by collecting the power and voltage status of each part of the system. The lower layer controls the coordinated control of the power energy between each device to achieve stable operation of the system. Due to the different characteristics of each device, different controls are required for different systems. Since the photovoltaic side is easily affected by the environment, maximum power tracking control is required, which is implemented by the disturbance observation method; and on the load side, due to the volatility of the photohydrogen power generation system, voltage and power stabilization control is required to achieve stable output voltage and power of each device; the actual working voltage of the electrolyzer is lower than the bus voltage, and the power current loop is used to control the step-down circuit to achieve stability; the energy storage system adopts double closed-loop control to achieve rapid response.

[0170] like Figure 4 As shown, by collecting the solar irradiance, the output power P of the photovoltaic module is obtained through further judgment. PV , and obtain the total system input power.

[0171] Comparison of photovoltaic output power P PV The power limit P set by the system max If P PV <P max , the photovoltaic cell works in MPPT mode, and the variable step size perturbation observation method is used for maximum power point tracking. The step size of the perturbation voltage is the absolute value of the ratio of the power to the voltage change a·|dP / dV|, where a is the proportional coefficient; if P PV >P max , the photovoltaic cell works in constant power mode, the voltage movement direction is constant and positive, and the voltage adjustment step is k(P PV -P max ), where k is the step size coefficient, and this process is iterated continuously until the photovoltaic output power is equal to the power upper limit set by the system.

[0172] Comparison of photovoltaic output power P PV With electrolyzer installed capacity X e , and make decisions on the working status of the electrolyzer and battery based on the remaining power of the battery:

[0173] 1) If P PV >X e , the electrolyzer operates at maximum power, i.e. P E =P Emax If the battery's state of charge is greater than 95%, it does not need to be charged; otherwise, the battery is charged to absorb the system's generated power, and the amount of charge is the system's remaining energy P PV -P E and the maximum battery charging power P BmaxThe smaller value in .

[0174] 2) If 20%P E.e <P PV <X e , P E.e is the rated power of the hydrogen production electrolyzer, 20%P E.e is the minimum power for safe operation of the hydrogen production electrolyzer, and the system input power is all used to produce hydrogen, that is, P E =P PV .

[0175] 3) If P PV <X e And P PV <20%P E.e , that is, the system input power is small. If the battery power is less than 5%, the working power of both is 0; if the battery power is higher than 5%, it is necessary to judge 20%P E.e -P PV -P Bmax If the value is less than 0, the battery is discharged and the electrolyzer maintains the minimum power operation; if the value is greater than 0, it means that even if the battery is discharged at the maximum discharge power, it still cannot reach the minimum working power of the electrolyzer, and the electrolyzer is shut down.

[0176] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A photovoltaic off-grid hydrogen production control method, characterized in that: The steps include: Step S1, obtaining the original data of the environmental sensor, using a distributed data acquisition system for synchronous acquisition, and obtaining a time-stamped original environmental data set containing solar radiation intensity, ambient temperature, photovoltaic panel surface temperature and wind speed; performing anomaly detection, data cleaning, standardization, spatiotemporal difference, feature extraction and dimensionality reduction on the original environmental data set in sequence, and obtaining a feature data set after dimensionality reduction; Step S2, acquiring a feature data set and pre-stored historical photovoltaic output data, using an integrated learning method combined with time series analysis to predict photovoltaic output power, and obtaining predicted photovoltaic output power; based on the predicted photovoltaic output power and current photovoltaic operating point data, applying an extreme value optimization algorithm to perform MPPT control optimization, and obtaining an optimized photovoltaic operating point voltage; according to the predicted photovoltaic output power and pre-stored photovoltaic array topology information, applying a graph theory algorithm to reconstruct and optimize the photovoltaic array, and obtaining an optimized photovoltaic array topology; Step S3, obtaining the pre-stored historical operation data of the electrolyzer and the specification parameters of the electrolyzer, and applying the system identification technology to construct a dynamic model of the electrolyzer; based on the constructed dynamic model and the predicted photovoltaic output power, applying a multi-objective optimization algorithm to optimize the efficiency of the electrolyzer, and obtaining an optimized electrolyzer working point sequence; according to the electrolyzer dynamic model and the optimized working point sequence, constructing an electrolyzer controller based on model predictive control; Step S4, obtaining the voltage, current and temperature data of the energy storage system collected in real time, applying the Kalman filter algorithm to estimate the state of the energy storage system, and obtaining the battery state of charge SOC and health state SOH; Based on the state estimation results and pre-stored historical usage data, the semi-Markov degradation model is used to predict the life of the energy storage system and obtain the remaining battery life distribution; according to the predicted photovoltaic output power, the optimized electrolyzer operating point sequence and battery status information, random dynamic programming is used to optimize the scheduling of the energy storage system and obtain the optimal scheduling strategy of the energy storage system.

2. The photovoltaic off-grid hydrogen production control method according to claim 1, characterized in that: Also includes: Step S5, obtaining the optimization results and controller parameters of each subsystem obtained in steps S2, S3 and S4, and constructing a multi-time scale coordinated control framework; Based on the system model parameters and controller parameters, the Lyapunov stability theory is applied to analyze the system stability and obtain the system stability region and stability margin. According to the system model and real-time fault diagnosis results, a data-driven adaptive fault-tolerant control strategy is constructed to obtain an adaptive fault-tolerant controller.

3. The photovoltaic off-grid hydrogen production control method according to claim 2, characterized in that: Step S1 is specifically as follows: Step S11, obtaining the original data of the environmental sensor, including solar radiation intensity, ambient temperature, photovoltaic panel surface temperature and wind speed; using a distributed data acquisition system to synchronously acquire these data; adding a timestamp to the acquired data; combining the data with the timestamp into a complete original environmental data set; Step S12, obtaining the original environmental data set obtained in step S11; applying the seasonal trend decomposition algorithm to the data set to decompose each data series into seasonal, trend and residual components; applying the median absolute deviation method to the decomposed residual components to identify outliers; using the local regression method to interpolate and replace the identified outliers; recombining the processed data to obtain a cleaned environmental data set; Step S13, obtaining the cleaned environmental data set obtained in step S12; applying the Z-score standardization method to each variable in the data set to calculate the standardized score of each data point; applying wavelet transform to the standardized data to perform signal denoising, specifically including selecting Daubechies wavelet basis function, performing multi-scale decomposition, applying soft threshold denoising and reconstructing the signal; combining the standardized and denoised data to obtain a standardized and denoised environmental data set; Step S14, obtaining the standardized and denoised environmental data set obtained in step S13; applying the Kriging interpolation method to the data set for spatiotemporal interpolation, first constructing a spatiotemporal variogram model, and then using a sliding time window for local Kriging interpolation; applying Kalman filtering to the interpolation result for time series smoothing; combining the interpolated and smoothed data to obtain a high-resolution spatiotemporal environmental data set; Step S15, obtaining the high-resolution spatiotemporal environmental data set obtained in step S14; applying the empirical mode decomposition method to the data set to extract the intrinsic mode function IMF; calculating the Hilbert-Huang transform for each IMF to obtain instantaneous frequency and amplitude information; inputting the obtained frequency and amplitude information into the multidimensional scaling analysis algorithm for nonlinear dimensionality reduction; and outputting the feature data set after dimensionality reduction.

4. The photovoltaic off-grid hydrogen production control method according to claim 3, characterized in that: Step S2 is specifically as follows: Step S21, obtaining the reduced feature data set obtained in step S1 and the pre-stored historical photovoltaic output data; using the XGBoost algorithm to build a regression model, input the feature data set, and output a preliminary prediction result; using the Prophet algorithm to build a time series prediction model, input the historical photovoltaic output data, and output the time series prediction result; aligning the output results of the XGBoost and Prophet models through the dynamic time warping DTW algorithm; integrating the aligned prediction results using the Bayesian model averaging method to obtain the final predicted photovoltaic output power; Step S22, obtaining the predicted photovoltaic output power obtained in step S21 and the current photovoltaic working point voltage and current data collected in real time; constructing an MPPT algorithm model based on particle swarm optimization PSO, and inputting the predicted photovoltaic output power and current working point data into the algorithm model; introducing adaptive inertia weights in the PSO algorithm, and dynamically adjusting the weights according to the current number of iterations t and the maximum number of iterations T_max; applying the Lévy flight strategy in the particle update process to enhance the local search capability; applying constraint processing technology to the results of each iteration to ensure that the search space is within the effective range; iterating the optimization process until convergence or reaching the maximum number of iterations, and outputting the optimized photovoltaic working point voltage; Step S23, obtaining the predicted photovoltaic output power obtained in step S21 and the pre-stored photovoltaic array topology information; constructing a graph model of the photovoltaic array, representing each photovoltaic panel as a node in the graph, and representing the connection relationship as an edge; applying a spectral clustering algorithm to the constructed graph model to divide the photovoltaic panels into several groups; The minimum spanning tree algorithm is applied to the photovoltaic panels in each group to optimize the connection relationship within the group; the graph coloring algorithm is used to assign colors to each photovoltaic panel to solve the potential shadow occlusion problem; the A* algorithm is applied to calculate the optimal path for array reconstruction; the photovoltaic array topology information is updated according to the optimization results to obtain the optimized photovoltaic array topology.

5. The photovoltaic off-grid hydrogen production control method according to claim 4, characterized in that: Step S3 is specifically as follows: Step S31, obtaining pre-stored historical operation data and specification parameters of the electrolytic cell; applying the subspace identification method to process the historical operation data and constructing a preliminary state space model; using the maximum likelihood estimation method to optimize the parameters of the state space model; evaluating models of different orders by the Akaike information criterion and selecting the optimal model order; using the Bootstrap method to resample and fit the selected model multiple times to evaluate the uncertainty of the model parameters; comprehensively considering the model performance and uncertainty, determining the final electrolytic cell dynamic model; Step S32, obtaining the dynamic model of the electrolyzer obtained in step S31 and the predicted photovoltaic output power obtained in step S2; constructing a multi-objective optimization problem, where the objective function includes maximizing hydrogen production and minimizing energy consumption; using the NSGA-III algorithm to solve the multi-objective optimization problem and obtain a series of non-dominated solutions; Use fuzzy decision-making methods to sort and select non-dominated solutions and determine the optimal solution; substitute the optimal solution into the dynamic programming algorithm, consider the time series characteristics, and optimize the timing control strategy; output the optimized electrolytic cell operating point sequence, including current, voltage and temperature; Step S33, obtaining the electrolytic cell dynamic model obtained in step S31 and the optimized electrolytic cell operating point sequence obtained in step S32; defining the prediction horizon N and the control horizon M of the model predictive control MPC; A state predictor is constructed based on the dynamic model of the electrolyzer. A quadratic programming QP problem is designed, and the optimization objective function, system constraints and prediction model are integrated into the optimization problem. The moving time domain strategy is applied to solve the QP problem in each control cycle to obtain the optimal control sequence. The Laguerre function is introduced to parameterize the control sequence to reduce the computational complexity. The optimized control strategy is integrated into the MPC controller to form a complete electrolyzer MPC controller.

6. The photovoltaic off-grid hydrogen production control method according to claim 5, characterized in that: Step S4 is specifically as follows: Step S41, acquiring the voltage, current and temperature data of the energy storage system collected in real time; constructing the Thevenin equivalent circuit model of the battery; designing a Sigma point Kalman filter (UKF), inputting the voltage and current data into the filter for state estimation; introducing an adaptive attenuation factor, and dynamically adjusting the filter parameters according to the estimation error; For the strong nonlinear region in SOC estimation, switch to the particle filter algorithm for processing; Combining the results of UKF and particle filtering, the final battery SOC and SOH estimation values ​​are output; Step S42, obtaining the battery SOC and SOH estimation values ​​obtained in step S41, and the pre-stored battery history usage data; constructing a multi-state semi-Markov model of battery degradation, defining degradation states and state transition rules; using the expectation maximization EM algorithm to estimate the parameters of the semi-Markov model; applying the Monte Carlo simulation method to generate multiple groups of degradation trajectories based on the estimated model parameters; using the generated degradation trajectories to calculate the probability distribution of the remaining battery life; applying the Kaplan-Meier estimator to correct the prediction results to improve the prediction accuracy; outputting the corrected battery remaining life distribution; Step S43, obtain the predicted photovoltaic output power obtained in step S2, the optimized electrolyzer operating point sequence obtained in step S3, and the battery SOC and SOH estimates obtained in step S41; define the system state space and action space, and use the charging and discharging power of the energy storage system as the action variable; construct a Markov decision process MDP model considering uncertainty; define an objective function based on maximizing the expected total return; apply the temporal difference learning algorithm to solve the MDP model and obtain a preliminary optimal strategy; introduce a risk sensitivity factor to adjust the strategy to balance benefits and risks; output the optimal scheduling strategy for the energy storage system considering risks.

7. The photovoltaic off-grid hydrogen production control method according to claim 6, characterized in that: Step S5 is specifically as follows: Step S51, obtain the optimization results and controller parameters of each subsystem obtained in steps S2, S3 and S4; construct a three-layer control architecture based on time scale separation, including day-ahead scheduling, hourly coordination and real-time balance; design a day-ahead scheduling algorithm based on model predictive control at the upper layer to handle the optimization problem at the 24-hour scale; implement rolling time domain optimization at the middle layer, update the optimization results once an hour, and coordinate the operation of each subsystem; design a robust H∞ controller at the bottom layer to achieve millisecond-level real-time power balance; integrate the three-layer control strategies to form a complete multi-time scale coordinated control framework; Step S52, obtaining system model parameters and controller parameters obtained in step S51; constructing a complete state space expression of the system, including the dynamic characteristics of the photovoltaic system, the electrolyzer and the energy storage system; designing a composite Lyapunov function, considering multiple operation modes of the system; applying the linear matrix inequality (LMI) technique to transform the stability problem into a convex optimization problem; solving the LMI problem to obtain a parameter range that satisfies the stability condition; using the Lyapunov exponent analysis method to evaluate the transient stability of the system under disturbance; Comprehensive stability analysis results to determine the stability region and stability margin of the system; Step S53, obtaining system model parameters and real-time fault diagnosis results; The local linear embedding (LLE) algorithm is used to process historical fault data and construct the fault feature space. Build an online fault identifier based on extreme learning machine (ELM) to classify the current system status in real time; Build multiple alternative controller models according to different fault types; Design a fuzzy inference system to calculate the weight of each candidate controller based on the fault identification results and system status; Realize smooth switching of controllers to ensure continuous and stable operation of the system under fault conditions; introduce sliding mode observer to estimate and compensate for unmodeled dynamics and external disturbances; integrate fault identification, controller switching and disturbance compensation modules to form a complete adaptive fault-tolerant controller.

8. The photovoltaic off-grid hydrogen production control method according to claim 3, characterized in that: Step S15 is specifically as follows: Step S151, obtaining the high-resolution spatiotemporal environmental data set obtained in step S14; applying white noise-assisted decomposition to each time series in the data set to generate multiple copies with different noise levels; Perform empirical mode decomposition on each noise auxiliary copy to obtain a series of IMF sets; calculate the mean of the IMFs corresponding to all noise auxiliary copies as the final IMF; apply singular spectrum analysis to the final IMF to remove residual noise; output the IMF set; Step S152, obtaining the improved IMF set outputted in step S151; applying Hilbert transform to each IMF to calculate the analytical signal; calculating the instantaneous frequency based on the analytical signal, using a differential algorithm to improve the frequency estimation accuracy; calculating the instantaneous amplitude, using an adaptive window smoothing method to reduce the endpoint effect; Construct the Hilbert spectrum to represent the time-frequency-energy distribution; apply the adaptive threshold method to extract the significant features in the Hilbert spectrum; Output the instantaneous frequency, instantaneous amplitude and significant characteristic data of IMF; Step S153, obtain the IMF instantaneous frequency, instantaneous amplitude and significant feature data output by step S152; construct a high-dimensional feature space with time, frequency, amplitude and significant features as dimensions; apply the t-SNE algorithm to initialize the embedding space, and use the PCA result as the initial point distribution; introduce an adaptive perplexity parameter (perplexity parameter) and dynamically adjust it according to the local data density; use the Barnes-Hut approximation to optimize the computational complexity; apply the early exaggeration technique to accelerate convergence; iterate the optimization until the stopping criterion is reached; output the data point coordinates in the low-dimensional embedding space as the final reduced dimensionality feature data set.

9. The photovoltaic off-grid hydrogen production control method according to claim 4, characterized in that: Step S22 is specifically as follows: Step S221, obtaining the predicted photovoltaic output power and current photovoltaic working point data of step S21; initializing the particle swarm of the HQPSO algorithm, each particle is represented as a quantum bit string; defining a fitness function, and calculating the fitness value of each particle according to the predicted power and the current working point; applying quantum measurement operations, mapping quantum state particles to the classical search space; outputting the initial particle swarm and its fitness value; Step S222, obtaining the initial particle swarm and fitness value of step S221; constructing an adaptive quantum revolving door, dynamically adjusting the rotation angle according to the current number of iterations and the particle fitness; applying the quantum revolving door operation to each particle to update its quantum state; introducing quantum entanglement operation to enhance information exchange between particles; applying the Lévy flight strategy, generating a flight step length according to the Lévy distribution, and updating the particle position; outputting the updated particle swarm; Step S223, obtain the particle swarm updated in step S222; apply constraint processing technology and use adaptive penalty function method to handle out-of-bounds particles; re-evaluate the fitness value of each particle; update the individual optimal solution and the global optimal solution; apply the elite retention strategy to ensure that the optimal solution is not lost during the iteration process; check the convergence condition or the maximum number of iterations; if the stop condition is not met, return to step S222 to continue iterating; after the stop condition is met, output the optimized photovoltaic operating point voltage.

10. A photovoltaic off-grid hydrogen production system, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the photovoltaic off-grid hydrogen production method according to any one of claims 1 to 9.

Images