Energy-storage-free off-grid photovoltaic hydrogen production system and control method

Abstract

The invention discloses a non-energy-storage off-grid photovoltaic hydrogen production system and a control method. The control method comprises five steps of environmental data acquisition and preprocessing, photovoltaic cell output power estimation, electrolytic cell power management, converter control parameter optimization and system performance evaluation and adaptive adjustment. The method comprises the following steps: firstly, processing environmental data by adopting an isolation forest algorithm, Kriging interpolation and an empirical mode decomposition method, and extracting time sequence features; and then combining with the estimated photovoltaic output power. And then the genetic algorithm and dynamic programming are utilized to realize electrolytic cell power optimization distribution. And then converter control parameters are optimized through subspace identification, pole assignment and a hybrid optimization algorithm. And finally, system performance evaluation and adaptive adjustment are realized by adopting a sliding time window method, singular spectrum analysis and a fuzzy prediction control method. The technical problems of the energy-storage-free photovoltaic hydrogen production system in the aspects of environmental data processing, power estimation, energy management, control optimization and the like are solved, and the operation efficiency and reliability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrogen production, and in particular to an off-grid photovoltaic hydrogen production system without energy storage, a control method and a control parameter optimization method. Background Art

[0002] By directly coupling photovoltaic power generation with water electrolysis to produce hydrogen, it is possible to achieve instant conversion and storage of energy, which is of great significance for promoting the development of the hydrogen energy economy and building a clean energy system.

[0003] At present, the research on off-grid photovoltaic hydrogen production systems without energy storage is mainly focused on basic control strategies and system integration. Traditional control methods mostly use simple maximum power point tracking (MPPT) algorithms, such as the perturbation observation method and the conductance increment method, to achieve the maximum power output of the photovoltaic array by adjusting the duty cycle of the DC / DC converter. In terms of electrolyzer control, constant current or constant voltage control strategies are often used to maintain the stable operation of the electrolyzer through a simple PID controller. In terms of system integration, a modular design method is mainly used to control the photovoltaic array, DC / DC converter and electrolyzer as independent units, and the coordination between the units mainly relies on simple logical judgment and threshold switching.

[0004] However, when facing the special needs of off-grid photovoltaic hydrogen production systems without energy storage, the existing technologies have some specific technical problems: First, in terms of environmental data processing, the existing anomaly detection methods are difficult to effectively handle the complex anomaly patterns in light intensity and temperature data, especially when the data presents multi-scale and non-stationary characteristics, traditional statistical methods are prone to misjudgment; second, in the process of photovoltaic model parameter identification, the existing optimization algorithms often fall into local optimality, and it is difficult to balance the relationship between global search and local refinement, resulting in insufficient model accuracy; third, the electrolyzer power management strategy lacks comprehensive consideration of multiple objectives, and it is difficult to find the optimal balance between system efficiency, hydrogen production efficiency and safety; fourth, in the process of DC / DC converter control parameter optimization, the existing methods are difficult to accurately quantify parameter sensitivity, resulting in low optimization efficiency; fifth, the system performance evaluation lacks in-depth analysis of multi-scale time series characteristics, and it is difficult to accurately predict performance trends and timely discover potential problems. The existence of these technical problems seriously restricts the operating efficiency and reliability of off-grid photovoltaic hydrogen production systems without energy storage. Summary of the invention

[0005] The purpose of the invention is to provide an off-grid photovoltaic hydrogen production system without energy storage and a control method, in order to solve at least one problem existing in the prior art.

[0006] Technical solution, according to one aspect of the present application, provides a method for controlling off-grid photovoltaic hydrogen production without energy storage, comprising the following steps:

[0007] Step S1, obtaining raw data from a light intensity sensor and a temperature sensor; using an isolation forest algorithm to detect and clean up the data; applying a Kriging interpolation algorithm to perform spatiotemporal series interpolation on the cleaned data; using an empirical mode decomposition method to perform multi-scale decomposition on the interpolated data; using a time series Shapelet algorithm to extract features from the decomposed data to obtain a light intensity feature vector and a temperature feature vector;

[0008] Step S2, reading the light intensity feature vector and the temperature feature vector, identifying the photovoltaic cell model parameters by combining the particle swarm optimization algorithm and the Levenberg-Marquardt algorithm; based on the identified model parameters and the environmental feature vector, predicting the maximum power point by using the integrated method of support vector regression and Gaussian process regression;

[0009] Step S3, historical operation data, combined with genetic algorithm and recursive least squares method to identify electrolytic cell model parameters; based on the identified model parameters and electrolytic cell specifications, a convex optimization algorithm is used to calculate the safe operating power range; according to the maximum power point and safe operating range predicted in step S2, a dynamic programming algorithm is used to generate a power allocation strategy;

[0010] Step S4, using the converter historical operation data, use the subspace identification method to identify the converter model parameters; based on the model parameters and power allocation strategy identified in step S3, use the pole configuration method to initialize the controller parameters; combine the simulated annealing algorithm and the response surface method to optimize the control parameters; output and send the optimized control parameters.

[0011] According to another aspect of the present application, there is also provided a non-storage off-grid photovoltaic hydrogen production control system, comprising:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] 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 off-grid photovoltaic hydrogen production control method without energy storage described in any of the above technical solutions.

[0015] The beneficial effect is to solve the technical problems of photovoltaic hydrogen production systems without energy storage in environmental data processing, power estimation, energy management and control optimization, and improve the system operation efficiency and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the present invention.

[0017] Figure 2This is a schematic structural diagram of an off-grid photovoltaic hydrogen production system without energy storage disclosed in an embodiment of the present invention.

[0018] Figure 3 The present invention discloses a flow chart of a hydrogen production power supply control method.

[0019] Figure 4 Schematic diagram of the control strategy for the interleaved three-level buck converter.

[0020] Figure 5 This is the voltage waveform of the photovoltaic cell in the simulation experiment results.

[0021] Figure 6 This is a partial enlarged view of the photovoltaic cell voltage and reference voltage waveform in the simulation experiment results.

[0022] Figure 7 This is the electrolytic cell power waveform in the simulation experiment results. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, according to one aspect of the present application, a method for controlling off-grid photovoltaic hydrogen production without energy storage is provided, comprising the following steps:

[0024] Step S1, obtaining raw data from a light intensity sensor and a temperature sensor; using an isolation forest algorithm to detect and clean up the data; applying a Kriging interpolation algorithm to perform spatiotemporal series interpolation on the cleaned data; using an empirical mode decomposition method to perform multi-scale decomposition on the interpolated data; using a time series Shapelet algorithm to extract features from the decomposed data to obtain a light intensity feature vector and a temperature feature vector;

[0025] Step S2, reading the light intensity feature vector and the temperature feature vector, identifying the photovoltaic cell model parameters by combining the particle swarm optimization algorithm and the Levenberg-Marquardt algorithm; based on the identified model parameters and the environmental feature vector, predicting the maximum power point by using the integrated method of support vector regression and Gaussian process regression;

[0026] Step S3, historical operation data, combined with genetic algorithm and recursive least squares method to identify electrolytic cell model parameters; based on the identified model parameters and electrolytic cell specifications, a convex optimization algorithm is used to calculate the safe operating power range; according to the maximum power point and safe operating range predicted in step S2, a dynamic programming algorithm is used to generate a power allocation strategy;

[0027] Step S4, using the converter historical operation data, use the subspace identification method to identify the converter model parameters; based on the model parameters and power allocation strategy identified in step S3, use the pole configuration method to initialize the controller parameters; combine the simulated annealing algorithm and the response surface method to optimize the control parameters; output and send the optimized control parameters.

[0028] It not only solves the technical problems faced by energy storage-free systems, such as large power fluctuations and complex control, but also achieves the optimal balance between system efficiency, hydrogen production efficiency and safety. Especially when environmental conditions change rapidly, the system can respond quickly and maintain stable operation, significantly improving overall performance and reliability.

[0029] According to one aspect of the present application, step S1 specifically comprises:

[0030] Step S11, obtaining real-time readings of the light intensity sensor and the temperature sensor, collecting data at a predetermined frequency not less than 1 Hz, and obtaining raw light intensity data L_raw and raw temperature data T_raw;

[0031] Step S12, inputting the original light intensity data L_raw and the original temperature data T_raw into the isolation forest algorithm; calculating the isolation index for each data point, the isolation index is equal to the negative power of 2, and the power is the ratio of the average path length of the data point to the normalization factor of the data set size; marking the data point with an isolation index greater than a preset threshold as abnormal and removing it from the data set; outputting the light intensity data L_clean and the temperature data T_clean after abnormality detection;

[0032] Step S13, input the light intensity data L_clean and temperature data T_clean that have passed the anomaly detection into the Kriging interpolation algorithm; construct a spatiotemporal variation function, which is a function of spatial distance and time interval, and includes undetermined parameters a, b1, b2, r1 and r2; use the variational Bayesian method to estimate the undetermined parameters and minimize the prediction error variance; apply the estimated parameters to interpolate, and output the interpolated light intensity data L_interp and temperature data T_interp;

[0033] Step S14, input the interpolated light intensity data L_interp and temperature data T_interp into the empirical mode decomposition algorithm; iteratively screen the signal to obtain a series of intrinsic mode functions; apply Hilbert-Huang transform to each intrinsic mode function to analyze its instantaneous frequency characteristics; group the intrinsic mode functions according to the frequency characteristics to obtain multi-scale components; output the light intensity data {L_1, L_2, ..., L_m} and temperature data {T_1, T_2, ..., T_m} after multi-scale decomposition;

[0034] Step S15, input the multi-scale decomposed light intensity data {L_1, L_2, ..., L_m} and temperature data {T_1, T_2, ..., T_m} into the time series Shapelet algorithm; calculate the distance between the Shapelet and the time series subsequence, and the distance is defined as the minimum value of the square root of the sum of the squares of the differences between the Shapelet and the corresponding elements of the subsequence; select the most recognizable Shapelet using the information gain criterion; construct a feature vector, and the vector elements are the distances between the selected Shapelet and the time series; output the light intensity feature vector F_L and the temperature feature vector F_T.

[0035] Through a multi-level data processing and analysis framework, high-quality collection and feature extraction of environmental data are achieved. Specifically, the isolation forest algorithm is used for anomaly detection to improve the reliability of the original data. By constructing a variogram containing both time and space dimensions and combining the variational Bayesian method to optimize parameters, the spatiotemporal continuity and accuracy of the data are significantly improved. The empirical mode decomposition method is used to decompose the data at multiple scales. Through iterative screening and Hilbert-Huang transform analysis of instantaneous frequency characteristics, adaptive decomposition of the signal is achieved, and environmental change characteristics at different scales are effectively separated. The system's adaptability to environmental changes and prediction accuracy are significantly improved, laying a data foundation for the stable operation of the photovoltaic hydrogen production system without energy storage.

[0036] According to one aspect of the present application, step S2 is specifically:

[0037] Step S21, input the light intensity feature vector F_L and the temperature feature vector F_T into the particle swarm optimization algorithm; wherein the objective function is the square sum of the difference between the measured current and the model current; update the particle position, and the position update amount is the weighted sum of the current speed, the individual optimal position and the global optimal position; input the updated particle position into the Levenberg-Marquardt algorithm for fine-tuning; calculate the Jacobian matrix and the residual vector, solve the normal equation to obtain the parameter increment; iterate the optimization until convergence, and output the photovoltaic cell model parameter vector P_pv;

[0038] Step S22, input the photovoltaic cell model parameter vector P_pv and the light intensity feature vector F_L and the temperature feature vector F_T into the support vector regression model; use the kernel function to map the input to the high-dimensional feature space and calculate the regression function value; at the same time, input the same input data into the Gaussian process regression model, calculate the mean function and covariance function, and obtain the probability prediction result; perform weighted integration on the outputs of the two models, and the weights are determined by minimizing the historical prediction error; output the predicted maximum power point MPP_pred, which includes the maximum power point voltage, current and power value.

[0039] By combining the particle swarm optimization algorithm with the Levenberg-Marquardt algorithm, and the integrated method of support vector regression and Gaussian process regression, a high-precision estimation of photovoltaic cell output power is achieved. First, the particle swarm optimization algorithm is used to perform a global search on the photovoltaic cell model parameters, which significantly improves the accuracy of model parameter identification. The integrated method of support vector regression and Gaussian process regression is used to predict the maximum power point, and a probabilistic prediction result with uncertainty estimation is obtained. Not only is the accuracy of the maximum power point prediction improved, but the uncertainty of the prediction result can also be quantified, providing a reliable basis for subsequent control decisions. Through this method of combining parameter identification and prediction, the system can accurately track the maximum power point of the photovoltaic array and effectively improve the energy utilization efficiency.

[0040] According to one aspect of the present application, step S3 is specifically:

[0041] Step S31, input the pre-stored electrolytic cell historical voltage and current data [V_hist, I_hist] into the genetic algorithm; encode the electrolytic cell model parameters to form chromosomes; define the fitness function as the inverse of the sum of squares of the difference between the measured voltage and the model voltage; perform selection, crossover and mutation operations to generate a new population; input the individuals in the population into the recursive least squares algorithm, calculate the gain matrix, and update the parameter estimate; iterate and optimize until convergence, and output the electrolytic cell model parameter vector P_el;

[0042] Step S32, input the electrolytic cell model parameter vector P_el and the pre-stored electrolytic cell specification parameter Spec_el into the convex optimization algorithm; construct the objective function as the electrolytic cell power function, add safety constraints such as maximum current, maximum temperature, etc.; apply the interior point method to solve the maximization problem to obtain the maximum safe power P_max; similarly, solve the minimization problem to obtain the minimum safe power P_min; output the safe operating power range of the electrolytic cell [P_min, P_max];

[0043] Step S33, input the maximum power point MPP_pred and the safe operating power range of the electrolyzer [P_min, P_max] into the dynamic programming algorithm; construct a state space, including circuit breaker, MPPT and constant power mode; construct an action space including switching modes and adjusting power; construct a reward function, considering system efficiency, hydrogen production efficiency and power loss; use the value iteration method to solve the optimal strategy, iteratively update the value function until convergence; output the power allocation strategy S_power, including the operating mode and the target power value.

[0044] By combining genetic algorithm, recursive least squares method and dynamic programming algorithm, efficient power management of electrolyzer is achieved. Specifically, genetic algorithm is first used to encode and optimize the parameters of electrolyzer model, which significantly improves the accuracy of electrolyzer model. Convex optimization algorithm is used to calculate the safe working power range, and the safe working range of electrolyzer is accurately determined. Finally, dynamic programming algorithm is used to generate power allocation strategy, which realizes intelligent allocation of electrolyzer power. This multi-level power management method not only ensures the safe operation of electrolyzer, but also significantly improves the overall efficiency of the system, which is particularly suitable for the characteristics of photovoltaic hydrogen production system without energy storage.

[0045] According to one aspect of the present application, step S4 is specifically:

[0046] Step S41, input the pre-stored converter historical operation data [V_in_hist, I_in_hist, V_out_hist, I_out_hist, D_hist] into the subspace identification algorithm; construct a state space model, use singular value decomposition to estimate the system order; use the projection algorithm to estimate the system matrix A, B, C, D; extract physical parameters from the system matrix, and output the converter model parameter P_conv;

[0047] Step S42, input the converter model parameter P_conv and the power allocation strategy S_power into the pole placement algorithm; calculate the characteristic equation coefficients according to the desired closed-loop poles; solve the Diophantine equation to obtain the controller parameter polynomial; extract the PID parameters from the polynomial and output the initial control parameter θ_init;

[0048] Step S43, input the initial control parameter θ_init, the converter model parameter P_conv obtained in step S41, the power allocation strategy S_power obtained in step S33 and the maximum power point MPP_pred predicted in step S22 into the simulated annealing algorithm; define the objective function, considering the adjustment time, overshoot and steady-state error; generate new parameters θ_new according to the current temperature parameters; use the response surface method to construct a second-order polynomial approximation model of the objective function; combine the simulated annealing algorithm and the response surface model to perform parameter search and optimization; output the optimized control parameter θ_opt.

[0049] The precise optimization of the control parameters of the DC / DC converter is achieved through the subspace identification method, pole placement method and hybrid optimization algorithm. The converter is modeled using the subspace identification algorithm to accurately obtain the dynamic characteristics of the converter. Subsequently, the controller parameters are initialized based on the pole placement method to achieve the preliminary configuration of the desired closed-loop characteristics. The control parameters are optimized by combining the simulated annealing algorithm and the response surface method, which significantly improves the efficiency and accuracy of parameter optimization. It not only improves the dynamic response performance of the DC / DC converter, but also enhances the stability of the system when the light intensity changes rapidly, which is of great significance for the reliable operation of the photovoltaic hydrogen production system without energy storage.

[0050] In another embodiment of the present application, it also includes:

[0051] Step S5: During operation, obtain the real-time operation data of the system and calculate the performance index using the sliding time window method; analyze the performance trend based on the historical performance index data in combination with the singular spectrum analysis and the ARIMA model; and generate an adaptive adjustment strategy using the fuzzy inference system and the model predictive control method according to the performance trend, the optimized control parameters and the power allocation strategy.

[0052] The step S5 is specifically as follows:

[0053] Step S51, obtaining the real-time operation data D_real of the system, including the photovoltaic voltage and current, the electrolyzer voltage and current, and the converter voltage and current; defining the sliding time window size W and the sliding step size S; in each time window, calculating the system efficiency, hydrogen production efficiency, power loss and response time; the system efficiency is the ratio of output power to input power, the hydrogen production efficiency is the ratio of the high calorific value of hydrogen to the input energy of the electrolyzer, the power loss is the difference between the input power and the output power, and the response time is calculated by step response analysis; outputting the performance index vector PI;

[0054] Step S52, input the pre-stored historical performance index data PI_hist into the singular spectrum analysis algorithm; construct a trajectory matrix, and perform singular value decomposition on the covariance matrix; select the eigenvectors corresponding to the main eigenvalues ​​to reconstruct the time series, and obtain the trend, cycle and noise components; input the extracted trend components into the ARIMA model, determine the difference order, autoregression order and moving average order; fit the ARIMA model, and calculate the slope of each indicator; output the performance trend vector T_perf;

[0055] Step S53, input the performance trend vector T_perf, the optimized control parameter θ_opt and the power allocation strategy S_power into the fuzzy reasoning system; construct a fuzzy rule base, and use the Takagi-Sugeno fuzzy reasoning method to calculate the output; at the same time, input the same input data into the model predictive control algorithm to predict the future system state and minimize the weighted sum of the prediction error and the control effort; combine the outputs of the fuzzy reasoning system and the model predictive control to generate the final adjustment strategy S_adjust, including the control parameter adjustment amount and the power allocation strategy adjustment amount.

[0056] Through the combination of sliding time window method, singular spectrum analysis, ARIMA model and model predictive control method, real-time evaluation and adaptive adjustment of system performance are realized. First, the sliding time window method is used to calculate performance indicators such as system efficiency, hydrogen production efficiency, power loss and response time, and the changing law of system performance is accurately grasped. Based on the fuzzy reasoning system and model predictive control method, an adaptive adjustment strategy is generated to realize the intelligent adjustment of the control strategy. Not only can the system performance degradation be discovered in time, but the control strategy can also be adjusted actively according to the performance change trend, which significantly improves the system's operational reliability and efficiency. Especially when the photovoltaic hydrogen production system without energy storage faces complex and changeable environmental conditions, this adaptive adjustment capability plays an important role in maintaining the stable operation of the system.

[0057] According to one aspect of the present application, step S12 is specifically:

[0058] Step S121, inputting the original light intensity data L_raw and the original temperature data T_raw into the multidimensional data stream feature extractor; segmenting the data using the sliding window method, and the window size is adaptively adjusted according to the time characteristics of the data; calculating the statistical features of the data in each window, including the mean, variance, skewness, kurtosis and autocorrelation coefficient; outputting the feature matrix F_raw;

[0059] Step S122, input the feature matrix F_raw into the local anomaly factor algorithm; calculate the average distance between each data point and its k nearest neighbors, where the k value is determined by an adaptive method; introduce a new density estimation function, considering the distribution characteristics of data points in different dimensions; calculate the local anomaly factor of each data point, where the local anomaly factor is equal to the ratio of the k-distance of the data point to the average k-distance of its k nearest neighbors; output the anomaly score vector S_anomaly;

[0060] Step S123, input the anomaly score vector S_anomaly into the dynamic threshold generator; use the kernel density estimation method to fit the probability distribution of the anomaly score; introduce an adaptive bandwidth selection algorithm to dynamically adjust the kernel function bandwidth according to the local density characteristics of the data; calculate the cumulative distribution function of the anomaly score distribution, and determine the dynamic threshold by the inverse function method; output the dynamic threshold T_dynamic;

[0061] Step S124, input the anomaly score vector S_anomaly and the dynamic threshold T_dynamic into the adaptive anomaly marker; compare the anomaly score of each data point with the dynamic threshold, and mark the data point with an anomaly score higher than the threshold as a potential anomaly; introduce a confidence evaluation mechanism to calculate the confidence of each potential anomaly point, and the confidence is proportional to the difference between the anomaly score and the threshold; output the anomaly labeling matrix M_anomaly, which contains anomaly labels and confidence information;

[0062] Step S125, input the anomaly labeling matrix M_anomaly, the raw light intensity data L_raw and the raw temperature data T_raw into the context-aware anomaly verifier; for each potential anomaly point, extract its temporal and spatial context information; design a new context similarity metric, taking into account the temporal continuity and spatial correlation of the data; calculate the similarity between the potential anomaly point and its context, and if the similarity is lower than a preset threshold, confirm it as a real anomaly; output the verified anomaly labeling matrix M_verified;

[0063] Step S126, input the verified abnormal marking matrix M_verified, the original light intensity data L_raw and the original temperature data T_raw into the adaptive abnormal repairer; for the confirmed abnormal points, use the local linear embedding (LLE) algorithm to repair; construct the manifold structure of the abnormal point neighborhood, considering the continuity constraint of the time series; optimize the reconstruction weights so that the repaired data points maintain the original geometric structure locally; for the data points that cannot be reliably repaired, mark them as missing values; output the repaired light intensity data L_clean and temperature data T_clean, and the missing value marking matrix M_missing;

[0064] Step S127, the repaired light intensity data L_clean and temperature data T_clean, and the missing value marking matrix M_missing are input into the data quality evaluator; various statistical indicators of the cleaned data are calculated, including the effective data ratio, data distribution characteristics and time continuity indicators; a multi-criteria decision-making method is used to comprehensively evaluate the data quality and generate a data quality report Q_report; the repaired light intensity data L_clean and temperature data T_clean, as well as the missing value marking matrix M_missing and the data quality report Q_report are output.

[0065] By constructing a multidimensional anomaly detection and repair framework, adaptive quality control of environmental data is achieved. Specifically, this step first uses the sliding window method to segment the data to provide a multidimensional feature basis for anomaly detection. The local outlier factor (LOF) algorithm makes anomaly detection more accurate. The combination of the kernel density estimation method and the adaptive bandwidth selection algorithm improves the adaptability of detection. The accuracy of anomaly detection is further improved by the context-aware anomaly verifier. Finally, the local linear embedding (LLE) algorithm is used for anomaly repair to achieve accurate repair of abnormal data. The quality and reliability of environmental data are significantly improved, especially in the case of rapid changes in light intensity and temperature data. It can accurately identify and process outliers, provide high-quality data input for subsequent power estimation and control optimization, and is of great significance to the stable operation of photovoltaic hydrogen production systems without energy storage.

[0066] According to one aspect of the present application, step S13 is specifically:

[0067] Step S131, input the restored light intensity data L_clean and temperature data T_clean, and the missing value marking matrix M_missing into the spatiotemporal correlation analyzer; use Moran's I index to calculate the spatial autocorrelation of the data, considering the temporal continuity of the data; apply wavelet transform to perform multi-scale decomposition on the data, and analyze the spatiotemporal correlation characteristics at different scales; output the spatiotemporal correlation feature matrix C_spatio_temporal;

[0068] Step S132, input the spatial correlation feature matrix C_spatio_temporal into the adaptive variogram generator; design a new non-stationary variogram model, allowing the variogram parameters to change with time and spatial position; use the local weighted regression method to estimate the variogram parameters, and the weights are based on the temporal and spatial distances of the data points; output the adaptive variogram parameter set P_variogram;

[0069] Step S133, input the adaptive variogram parameter set P_variogram, the repaired light intensity data L_clean, the temperature data T_clean and the missing value marking matrix M_missing into the multi-scale Kriging interpolator; decompose the data at multiple scales, and apply the Kriging interpolation algorithm at each scale; introduce a new nonlinear kernel function to improve the accuracy of interpolation in non-stationary areas; use an adaptive neighborhood selection strategy to dynamically adjust the interpolation neighborhood according to the local data density; output the preliminary interpolation result I_preliminary;

[0070] Step S134, input the preliminary interpolation result I_preliminary into the variational Bayesian optimizer; construct a new probabilistic graphical model, and formalize the interpolation problem into a Bayesian inference task; design a variational objective function, while considering the data fit and the smoothness of the interpolation result; use the stochastic gradient variational inference algorithm to optimize the model parameters, and introduce an adaptive learning rate adjustment mechanism; output the optimized interpolation result I_optimized;

[0071] Step S135, input the optimized interpolation result I_optimized into the uncertainty quantification module; use the Monte Carlo method to generate multiple samples of the interpolation result; calculate the confidence interval and prediction variance of each interpolation point; introduce a new uncertainty propagation model to evaluate the impact of interpolation uncertainty on subsequent analysis; output the final interpolation result I_final and uncertainty matrix U_interpolation with uncertainty estimation.

[0072] The high-precision reconstruction of environmental data was achieved by adopting a multi-scale spatiotemporal interpolation method. First, the spatial autocorrelation of the data was calculated by the Moran's I index, which provided a reliable correlation basis for subsequent interpolation. The non-stationary variogram model allows parameters to change with time and spatial position, realizing the adaptation of the interpolation model to local characteristics. In the multi-scale Kriging interpolation process, the introduction of a new nonlinear kernel function improved the interpolation accuracy in non-stationary areas, and the adaptive neighborhood selection strategy dynamically adjusted the interpolation neighborhood according to the local data density, significantly improving the local adaptability of the interpolation. The variational Bayesian optimizer formalized the interpolation problem as a Bayesian inference task by constructing a new probabilistic graphical model, while considering the data fit and the smoothness of the interpolation results, and realized the global optimization of the interpolation parameters. The Monte Carlo method was used to generate multiple samples of the interpolation results, and combined with the new uncertainty propagation model, the interpolation uncertainty was accurately quantified. This adaptive multi-scale interpolation method not only improves the spatiotemporal continuity and integrity of environmental data, but also accurately evaluates the reliability of the interpolation results, providing a reliable guarantee for the reconstruction of environmental data for photovoltaic hydrogen production systems without energy storage.

[0073] According to one aspect of the present application, step S14 is specifically:

[0074] Step S141: Input the final interpolation result I_final into the signal preprocessing module. Apply an adaptive median filter to remove residual noise, and dynamically adjust the filter window size according to the local signal characteristics. Use the spline interpolation method to resample the data to ensure uniform sampling intervals. Output the preprocessed signal.

[0075] Step S142: Input the preprocessed signal into the multi-scale singular value decomposition (MSVD) module. Build a multi-scale analysis framework and apply singular value decomposition at different time scales. Introduce a new adaptive threshold strategy to dynamically select effective components based on the singular value distribution. Output the multi-scale decomposition result.

[0076] Step S143: Input the multi-scale decomposition results into the empirical mode decomposition (EMD) algorithm. Design a new envelope estimation method and use local polynomial regression to improve the accuracy of extreme point interpolation. Introduce an adaptive stopping criterion and dynamically adjust the number of iterations based on the orthogonality between components and energy distribution. Output a preliminary IMF component set.

[0077] Step S144: Input the preliminary IMF component set into the modal alignment and optimization module. Use the dynamic time warping (DTW) algorithm to align IMFs of different scales. Design a new modal mixing metric to evaluate the correlation and independence between IMFs. Apply independent component analysis (ICA) technology to further separate the mixed modes. Output the optimized IMF component set I_optimized.

[0078] Step S145: Input the optimized IMF component set I_optimized into the Hilbert-Huang transform module. Apply Hilbert transform to each IMF to calculate the instantaneous frequency and instantaneous amplitude. Design a new frequency-time-energy distribution representation method to improve the resolution of time-frequency analysis. Output the Hilbert spectrum H_spectrum and the instantaneous frequency matrix F_instantaneous.

[0079] This step constructs a multi-level signal decomposition framework. The data is preprocessed by an adaptive median filter. The filter window size is dynamically adjusted according to the local signal characteristics. Uniform resampling is achieved by combining spline interpolation, which improves the quality of the original signal. The multi-scale singular value decomposition module realizes the preliminary decomposition of the signal by introducing an adaptive threshold strategy. The empirical mode decomposition algorithm uses local polynomial regression to improve the accuracy of extreme point interpolation. The adaptive stopping criterion dynamically adjusts the number of iterations based on the orthogonality between components and energy distribution to ensure the reliability of the decomposition results. The dynamic time warping algorithm is used to align IMFs of different scales, which significantly improves the accuracy of the decomposition. Finally, the instantaneous frequency and amplitude analysis of each IMF is realized by Hilbert-Huang transform, and the newly designed frequency-time-energy distribution representation method improves the resolution of time-frequency analysis. The ability to recognize various scale change features in environmental data is significantly improved, especially in the case of rapid changes in light intensity, and the transient features can be accurately captured, providing reliable signal feature input for the precise control of photovoltaic hydrogen production systems without energy storage.

[0080] According to one aspect of the present application, step S15 is specifically:

[0081] Step S151: Input I_optimized, H_spectrum and F_instantaneous obtained in step S14 into the multiscale feature extractor. Calculate statistical features at different time scales, including mean, variance, skewness, kurtosis and entropy. Introduce new multiscale complexity indicators, such as multiscale permutation entropy and multiscale fuzzy entropy. Output the multiscale feature matrix F_multiscale.

[0082] Step S152: Input F_multiscale obtained in step S151 into the adaptive Shapelet generator. Design a new Shapelet candidate generation strategy based on the local characteristics and global structure of the signal. Use a genetic algorithm to optimize the Shapelet length and position, and introduce a multi-objective optimization framework to balance the discrimination and interpretability. Output the initial Shapelet set S_initial.

[0083] Step S153: Input S_initial obtained in step S152 into the Shapelet evaluation and optimization module. Design a new Shapelet similarity metric, taking into account the nonlinear and non-stationary characteristics of the time series. Use local sensitive hashing (LSH) technology to accelerate similarity calculation. Apply a multi-criteria decision method to comprehensively evaluate the quality of the Shapelet, taking into account discrimination, stability, and interpretability. Output the optimized Shapelet set S_optimized.

[0084] Step S154: Input S_optimized obtained in step S153 into the time series encoder. Use the selected Shapelet to construct a symbolic representation of the time series. Design a new encoding strategy to discretize continuous similarity values ​​into a symbolic sequence. Introduce context embedding technology to capture long-range dependencies in the symbolic sequence. Output the symbolic time series representation T_symbolic.

[0085] Step S155: Input T_symbolic obtained in step S154 into the feature fusion and selection module. Combine the multi-scale features and symbolic representation of step S151 to construct a high-dimensional feature space. Apply a feature selection algorithm, such as the maximum relevance minimum redundancy (mRMR) method based on mutual information, to select the most representative and discriminative feature subset. Output the final feature vector F_final.

[0086] By establishing a multi-level feature extraction system, the deep feature expression of time series data is achieved. The statistical characteristics of the signal are fully captured, and the representativeness of the features is improved. The efficiency of feature extraction is significantly improved. The expressiveness of the features is enhanced. The redundancy of the features is significantly reduced. Not only is the expressiveness of environmental data features improved, but also the computational efficiency of the subsequent control system is improved by reducing feature redundancy, which is of great significance to the real-time control of photovoltaic hydrogen production systems without energy storage.

[0087] According to one aspect of the present application, step S21 is specifically:

[0088] Step S211: Input the feature vector F_final and the pre-stored photovoltaic cell historical operation data D_pv_hist into the data preprocessing module. Perform standardization on F_final to obtain a standardized feature vector F_norm. Extract voltage, current and temperature data from D_pv_hist to construct a training data set D_train. Output F_norm and D_train.

[0089] Step S212: Input F_norm and D_train to the initial particle generator. Use the Latin hypercube sampling method to generate a uniformly distributed initial particle group in the parameter space. Each particle represents a set of photovoltaic model parameters [I_sc, V_oc, I_m, V_m, R_s, R_sh]. Calculate the initial fitness value of each particle, and the fitness function is the mean square error between the model prediction value and the actual value. Output the initial particle group P_init and the initial fitness value set F_init.

[0090] Step S213, input P_init and F_init into the particle swarm optimization algorithm. Introduce an adaptive inertia weight strategy to dynamically adjust the inertia weight according to the convergence of the particle swarm. Design a new local search operator to perform a local refined search for the optimal particle in each iteration. Apply the crowding calculation method to maintain the diversity of the particle swarm. Iterate and update the particle position and velocity until the preset termination condition is reached. Output the optimized particle swarm P_pso and the corresponding fitness value F_pso.

[0091] Step S214, input P_pso and F_pso to the Levenberg-Marquardt algorithm optimizer. Select the particle with the best fitness as the initial parameter estimate. Construct the Jacobian matrix J and calculate the residual vector r. Design an adaptive damping factor adjustment strategy to dynamically adjust the damping factor λ according to the convergence speed of the optimization process. Iterate and solve the normal equation (J^TJ + λI)δ = J^T r to update the parameter estimate. Repeat this process until convergence or the maximum number of iterations is reached. Output the optimized parameter estimate P_lm.

[0092] Step S215: Input P_lm and D_train to the model verification module. Use P_lm to calculate the model prediction value and compare it with the actual value. Calculate multiple performance indicators, including root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). If the performance indicator does not reach the preset threshold, return to S213 for re-optimization. Otherwise, output the final photovoltaic cell model parameter vector P_pv_final and model performance report R_performance.

[0093] By building an adaptive hybrid optimization framework, high-precision identification of photovoltaic model parameters is achieved. The global nature of initial parameter estimation is ensured. The global search capability of parameter optimization is significantly improved. Accurate local optimization of parameters is achieved. The accuracy of model parameter identification is ensured. Especially in the case of rapid changes in light intensity and temperature, the output characteristics of the photovoltaic array can be accurately predicted, laying the foundation for the precise control of photovoltaic hydrogen production systems without energy storage.

[0094] According to one aspect of the present application, step S22 is specifically:

[0095] Step S221, input P_pv_final, F_final and pre-stored environmental condition data D_env into the data enhancement module. Use P_pv_final to generate simulated data under different working conditions. Use methods such as Gaussian noise and random occlusion to expand the data set. Combine the simulated data and actual data to construct an enhanced data set D_augmented. Perform feature engineering on D_augmented to extract features related to the maximum power point. Output the enhanced feature matrix X_aug and the corresponding maximum power point label Y_aug.

[0096] Step S222: Input X_aug and Y_aug to the support vector regression (SVR) training module. Use grid search and cross-validation methods to select the optimal kernel function and hyperparameters. Apply the sequential minimum optimization (SMO) algorithm to train the SVR model. Calculate the training error and validation error. If overfitting, apply L1 regularization. Output the trained SVR model M_svr and its performance index P_svr.

[0097] Step S223, input X_aug and Y_aug into the Gaussian process regression (GPR) training module. Design a composite kernel function, combining the radial basis function (RBF) kernel and the periodic kernel to capture the nonlinear and periodic characteristics of the data. Use the marginal likelihood maximization method to optimize the kernel function hyperparameters. Apply sparse approximation techniques to process large-scale data sets, such as the full set of bootstrap approximation method (FITC). Output the trained GPR model M_gpr and its performance index P_gpr.

[0098] Step S224: Input M_svr, P_svr and M_gpr, P_gpr into the model integration module. Design a dynamic weight allocation strategy to dynamically adjust the weight according to the performance of each model under different working conditions. Use the Bayesian model average (BMA) method to integrate the prediction results and uncertainty estimates of the two models. Construct the final integrated model M_ensemble. Output M_ensemble and the integrated model performance index P_ensemble.

[0099] S225. Input M_ensemble and current environmental condition data D_env_current into the real-time prediction module. Use M_ensemble to predict the maximum power point under the current conditions. Calculate the confidence interval of the prediction result. Output the predicted maximum power point MPP_pred, including voltage V_mpp, current I_mpp, power P_mpp and corresponding confidence interval CI_mpp.

[0100] By building an integrated learning framework, high-precision prediction of the maximum power point is achieved. The problem of insufficient actual data is effectively solved. The quality and representativeness of the training data are significantly improved. The generalization ability of the model is ensured. The computational efficiency is significantly improved. The prediction accuracy is further improved. A reliable decision-making basis is provided for system control. It not only improves the accuracy of the maximum power point prediction, but also quantifies the uncertainty of the prediction. It is particularly suitable for the maximum power point tracking control of the photovoltaic hydrogen production system without energy storage under complex and changeable environmental conditions, and effectively improves the energy utilization efficiency of the system.

[0101] According to one aspect of the present application, step S31 is specifically:

[0102] Step S311: Input the pre-stored electrolytic cell historical operation data D_el_hist into the data preprocessing module. Perform outlier detection and removal, and use the sliding window method to smooth the data. Normalize the voltage, current, temperature and pressure data. Construct a training data set D_el_train, which includes input features and corresponding output voltages. Output D_el_train and data statistical information S_el_stat.

[0103] Step S312, input D_el_train and S_el_stat to the initial population generator. Use prior knowledge and statistical information to set the parameter range. Apply the orthogonal experimental design method to generate the initial population to ensure the diversity and representativeness of the population. Each individual represents a set of electrolytic cell model parameters [V_rev, r, s, t, A]. Calculate the initial fitness value of each individual, and the fitness function is the inverse of the mean square error between the model predicted voltage and the actual voltage. Output the initial population G_init and the initial fitness value set F_init.

[0104] Step S313, input G_init and F_init to the genetic algorithm optimizer. Design adaptive crossover and mutation operators, and dynamically adjust the crossover rate and mutation rate according to the diversity of the population. Introduce an elite retention strategy to ensure that the best individuals will not be lost during the evolution process. Apply the island model parallelization strategy, divide the population into multiple sub-populations for parallel evolution, and migrate individuals regularly. Iterate the selection, crossover and mutation operations until the termination condition is reached. Output the optimized population G_ga and the corresponding fitness value F_ga.

[0105] Step S314: The individual with the highest fitness in G_ga is used as the initial parameter estimate P_init and input into the recursive least squares (RLS) optimizer together with D_el_train. Construct the observation matrix H and the output vector y. Initialize the parameter estimate θ and the covariance matrix P. For each data sample, calculate the prediction error e and the Kalman gain K. Update the parameter estimate θ and the covariance matrix P. Introduce the forgetting factor λ to adapt to the slow changes in the electrolyzer parameters. Repeat this process until convergence or all data are processed. Output the optimized parameter estimate P_rls.

[0106] Step S315, input P_rls and D_el_train into the model verification module. Use P_rls to calculate the model predicted voltage and compare it with the actual voltage. Calculate multiple performance indicators, including root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). If the performance indicator does not reach the preset threshold, return to S313 for re-optimization. Otherwise, output the final electrolyzer model parameter vector P_el_final and model performance report R_el_performance.

[0107] Adopting the adaptive hybrid optimization strategy, the accurate identification of electrolyzer parameters is achieved. It not only improves the accuracy of the electrolyzer model, but also adapts to the dynamic changes of electrolyzer parameters, providing accurate model support for the electrolyzer control of the photovoltaic hydrogen production system without energy storage.

[0108] According to one aspect of the present application, step S32 is specifically:

[0109] Step S321, input SP_el_final and the pre-stored electrolytic cell specification parameter Spec_el into the constraint generator. Construct the electrolytic cell voltage model according to P_el_final. Extract the maximum current I_max, maximum temperature T_max, maximum pressure P_max and other limiting parameters from Spec_el. Construct the inequality constraints g1(V,I) ≤ 0, g2(V,I) ≤ 0, ..., gn(V,I) ≤ 0, indicating that the current, temperature, pressure, etc. do not exceed the limit value. Construct the equality constraint h1(V,I) = 0, indicating the voltage-current relationship of the electrolytic cell. Output the constraint set C_constraints.

[0110] Step S322: Input C_constraints obtained in S321 into the objective function construction module. Design multiple objective functions: f1(V,I) represents hydrogen production efficiency, f2(V,I) represents system power, and f3(V,I) represents energy consumption. Combine multiple objective functions into a weighted sum objective function F(V,I) = w1f1 + w2f2 + w3*f3, where wi is the weight coefficient. Output the comprehensive objective function F(V,I) and the initial weight coefficient W_init.

[0111] Step S323, input F(V,I), W_init obtained in S322 and C_constraints obtained in S321 into the interior point method solver. Use the logarithmic barrier function to convert the inequality constraint into the penalty term of the objective function. Construct an augmented Lagrangian function. Design an adaptive step size strategy to dynamically adjust the step size according to the feasibility and optimality of the current solution. Use Newton's method to solve the KKT conditions and obtain the search direction. Iterate and update the original variables and dual variables until convergence or the maximum number of iterations is reached. Output the optimal solution X_opt, including the optimal voltage V_opt and current I_opt.

[0112] Step S324: Input X_opt obtained in S323 to the Pareto frontier generator. Generate the Pareto optimal solution set using the ε-constraint method. Solve a series of single-objective optimization problems by systematically changing the ε value. For each ε value, repeatedly call the interior point solver in S323. Collect all non-dominated solutions to form the Pareto frontier. Output the Pareto optimal solution set PS_pareto.

[0113] Step S325, input the PS_pareto obtained in S324 into the decision-making module. Use the TOPSIS method to select the best solution from the Pareto optimal solution set. Construct a normalized decision matrix and calculate the weighted normalized decision matrix. Determine the ideal solution and the negative ideal solution, and calculate the distance of each solution to the ideal solution and the negative ideal solution. Calculate the relative proximity of each solution, and select the solution with the largest relative proximity as the final solution. Output the safe operating power range of the electrolytic cell [P_min, P_max], where P_min and P_max correspond to the minimum and maximum safe operating powers, respectively.

[0114] In this embodiment, a multi-objective optimization method is used to achieve accurate calculation of the safe working area of ​​the electrolyzer. It not only ensures the safe operation of the electrolyzer, but also achieves the optimal balance of efficiency, power and energy consumption, which is particularly suitable for the characteristics of photovoltaic hydrogen production systems without energy storage.

[0115] According to one aspect of the present application, step S33 is specifically:

[0116] Step S331, input the MPP_pred obtained in S22 and [P_min, P_max] obtained in S32 into the state space construction module. Discretize the system power P into N levels and construct the state space S = {s1, s2, ..., sN}, where si represents the i-th power level. Define the action space A = {a1, a2, a3}, which respectively represent increasing power, keeping unchanged and decreasing power. Construct the state transition probability matrix P(s'|s,a), and use historical data to estimate the transition probability. Output the state space S, action space A and transition probability matrix P.

[0117] Step S332: Input S, A, and P obtained in S331 into the reward function design module. Design the immediate reward function R(s,a), taking into account the system efficiency η_system, hydrogen production efficiency η_H2, and power loss P_loss. Introduce a safety penalty term, and give a negative reward when the power exceeds the safety range. Use fuzzy logic methods to integrate multiple goals into a single reward value. Output the reward function R(s,a).

[0118] Step S333: Input S, A, P obtained in S331 and R(s,a) obtained in S332 into the value iteration algorithm module. Initialize the value function V(s) to zero. Iterate and update the value function: V(s) = max_a(R(s,a) + γΣP(s'|s,a)V(s')), where γ is the discount factor. Introduce an asynchronous update strategy to give priority to updating the value function of high reward states. Design a convergence criterion and stop the iteration when the value function changes less than a preset threshold. Output the optimal value function V*(s).

[0119] Step S334: Input V*(s) obtained in S333 to the strategy extraction module. For each state s, calculate the optimal action a* = argmax_a(R(s,a) + γΣP(s'|s,a)V*(s')). Construct a deterministic strategy π*(s) to map each state to the optimal action. Output the optimal strategy π*(s).

[0120] Step S335: Input the π*(s) obtained in S334 and the current system state s_current into the real-time decision execution module. Determine the optimal action a_optimal in the current state based on π*(s). Convert a_optimal into a specific power adjustment instruction ΔP. Consider the system response delay and design a smooth transition strategy to avoid power mutation. The output power allocation strategy S_power includes the target power P_target and the adjustment mode mode (MPPT / constant power / disconnection).

[0121] Step S336: Input S_power obtained in S335 and the real-time operation data D_real_time of the system into the online learning module. Use the temporal difference (TD) learning method to update the value function estimate based on the actual transfer and reward. Calculate the TD error δ = R + γV(s') - V(s). Use a function approximator (such as a radial basis function network) to represent the value function, and update the function approximator parameters by the gradient descent method. Update the transition probability matrix P and the reward function R regularly. Output the updated value function V_updated(s) and the strategy π_updated(s).

[0122] Step S337: Input V_updated(s), π_updated(s) and long-term performance statistics D_performance obtained in S336 into the strategy evaluation module. Calculate the average cumulative reward, stability index and security index. Use the Monte Carlo method to simulate the long-term system behavior and evaluate the robustness of the strategy. If the performance index is lower than the threshold, trigger a re-optimization signal. Otherwise, keep the current strategy. Output the strategy evaluation report R_evaluation and the optimization flag flag_optimize.

[0123] Step S338: Input R_evaluation and flag_optimize obtained in S337 to the adaptive adjustment module. If flag_optimize is true, adjust the parameters of the value iteration algorithm, such as the discount factor γ and the convergence threshold. Update the reward function weight to balance efficiency, safety, and stability. Reinitialize the process of S333-S337 to generate a new optimization strategy. If flag_optimize is false, fine-tune the current strategy to adapt to the slowly changing system dynamics. Output the adjusted dynamic programming parameter set P_adp and the update flag flag_update.

[0124] Step S339: Input P_adp and flag_update obtained in S338 and S_power obtained in S335 into the strategy integration and output module. If flag_update is true, use P_adp to update the core components of the dynamic programming algorithm. Integrate the short-term strategy S_power and the long-term optimization results to generate the final power allocation strategy S_power_final. Store S_power_final in the strategy database for subsequent decision-making reference. Output the final power allocation strategy S_power_final, including the optimized target power P_target_opt and the operating mode mode_opt.

[0125] In this embodiment, an adaptive dynamic programming framework is constructed to achieve intelligent optimization of power distribution. It not only achieves real-time optimization of power distribution, but also can adaptively adjust the control strategy according to the system operation status, significantly improving the operating efficiency and stability of the photovoltaic hydrogen production system without energy storage.

[0126] According to one aspect of the present application, step S41 is specifically:

[0127] Step S411: Input the pre-stored converter historical operation data [V_in_hist, I_in_hist, V_out_hist, I_out_hist, D_hist] into the data preprocessing module. Use median filtering to remove outliers and apply wavelet transform for noise reduction. Normalize all data to obtain a standardized data set D_norm. Construct an input matrix U and an output matrix Y, where U contains [V_in, I_in, D] and Y contains [V_out, I_out]. Output D_norm, U and Y.

[0128] Step S412: Input U and Y obtained in S411 into the singular value decomposition (SVD) module. Construct the Hankel matrix H = [U; Y] for the input and output data. Perform SVD decomposition on H: H = USV^T. Analyze the singular value distribution and use the cumulative energy contribution method to determine the system order n. Decompose U and S into Un, Σn (corresponding to the first n singular values) and Un⊥, Σn⊥ (corresponding to the remaining singular values). Output Un, Σn, Un⊥, Σn⊥ and system order n.

[0129] Step S413, input Un, Σn, Un⊥, Σn⊥ and n obtained in S412 into the state space matrix estimation module. Construct the projection matrix Π = Un(Un^TUn)^(-1)Un^T. Calculate the extended observability matrix Γ = Un(Σn)^(1 / 2). Estimate the system matrices A and C using the least squares method: [A; C] = Γ^+[Γ↓; Y], where Γ↓ means that Γ removes the last row. Similarly estimate B and D: [B; D] = [U⊥^TU; Y](V⊥S⊥^(-1))^T. Output the state space matrices A, B, C, D.

[0130] Step S414, input A, B, C, D obtained in S413 into the model verification module. Use part of the data not used for identification for cross-validation. Calculate the root mean square error (RMSE) and fit (FIT) between the model output and the actual output. If the performance index does not reach the preset threshold, return to S412 to adjust the system order. Otherwise, proceed to the next step. Output the verified state space matrix A_val, B_val, C_val, D_val and the model performance index P_model.

[0131] Step S415: Input A_val, B_val, C_val, and D_val obtained in S414 into the parameter extraction module. Use the eigenvalue decomposition method to extract the system natural frequency and damping ratio from A_val. Extract the gain coefficient from B_val and C_val. Using this information, combined with the converter topology, reversely solve the physical parameters L (inductance value), C (capacitance value), R_L (inductor equivalent series resistance), and R_C (capacitor equivalent series resistance). Output the converter model parameters P_conv = [L, C, R_L, R_C] and the identification model M_ss.

[0132] In this embodiment, by constructing an adaptive subspace identification framework, high-precision identification of the DC / DC converter model is achieved. Not only can the dynamic characteristics of the converter be accurately obtained, but the model order can also be automatically determined according to the input and output data. It is particularly suitable for online modeling of DC / DC converters in photovoltaic hydrogen production systems without energy storage, and provides an accurate mathematical model for subsequent controller design.

[0133] According to one aspect of the present application, step S42 is specifically:

[0134] Step S421, input P_conv and M_ss obtained in S41, and S_power_final obtained in S339 into the expected performance index generation module. According to the operating mode mode_opt in S_power_final, select the corresponding performance index template. Use the fuzzy rule base to map the target power P_target_opt to the expected adjustment time Ts, overshoot Mp and steady-state error Ess. Consider the parameter uncertainty in P_conv and set the allowable range for each indicator. Output the expected performance index vector PI_desired = [Ts, Mp, Ess] and its allowable range ΔPI.

[0135] Step S422, input PI_desired and ΔPI obtained in S421, and M_ss obtained in S41 into the closed-loop characteristic equation construction module. According to PI_desired, use the second-order underdamped system model to estimate the expected dominant pole position p1,2 = -ζωn ± jωn√(1-ζ^2). Considering the system order, add an appropriate number of non-dominant poles to ensure that their real part is less than 5 times the real part of the dominant pole. Construct the expected closed-loop characteristic polynomial Δ(s) = (s-p1)(s-p2)...(s-pn). Output the expected closed-loop characteristic polynomial coefficient vector a = [a0, a1, ..., an].

[0136] Step S423, input the a obtained in S422 and the M_ss obtained in S41 into the Diophantine equation solving module. Construct the system transfer function G(s) = num(s) / den(s) according to M_ss. Set the controller transfer function to C(s) = X(s) / Y(s). Construct the Diophantine equation: A(s)X(s) + B(s)Y(s) = Δ(s), where A(s) and B(s) are the denominator and numerator polynomials of G(s), respectively. Use the extended Euclidean algorithm to solve the Diophantine equation and obtain the coefficients of X(s) and Y(s). Output the controller parameter polynomials X(s) and Y(s).

[0137] Step S424: Input X(s) and Y(s) obtained in S423 into the PID parameter extraction module. Expand X(s) / Y(s) into the standard PID form: C(s) = Kp + Ki / s + Kds. Extract the proportional gain Kp, integral gain Ki and differential gain Kd. Considering the actual implementation, filter the Kd term and introduce the filter time constant Tf. Output the initial PID parameter θ_init = [Kp, Ki, Kd, ​​Tf].

[0138] Step S425: Input the θ_init obtained in S424 and the M_ss obtained in S41 to the robustness analysis module. Use the structural singular value (μ) analysis method, consider the uncertainty range of the parameters in P_conv, and calculate the robust stability margin and robust performance margin of the closed-loop system. If the margin is insufficient, return to S422 to adjust the desired pole position. Otherwise, keep the current parameters. Output the robustness analysis report R_robust and the final initial control parameter θ_init_robust.

[0139] In this embodiment, the adaptive pole configuration method is used to achieve the precise design of the controller. It not only achieves the required dynamic performance indicators, but also ensures the system stability under the condition of parameter uncertainty, which is particularly suitable for the precise control of DC / DC converters in photovoltaic hydrogen production systems without energy storage.

[0140] According to one aspect of the present application, step S43 is specifically:

[0141] Step S431, input θ_init_robust obtained in S425, M_ss and P_conv obtained in S41, and S_power_final obtained in S339 into the objective function construction module. Design the multi-objective function J(θ) = w1*J1(θ) + w2*J2(θ) + w3*J3(θ), where J1 represents the adjustment time, J2 represents the overshoot, and J3 represents the steady-state error. Dynamically adjust the weights w1, w2, and w3 according to mode_opt in S_power_final. Construct constraints to ensure that the parameters are within a reasonable range. Output the objective function J(θ), the constraint C(θ), and the initial weight W_init.

[0142] Step S432, input J(θ), C(θ), W_init obtained in S431 and θ_init_robust obtained in S425 into the simulated annealing algorithm module. Initialize temperature T = T0 and set the cooling coefficient α. At each temperature, generate a new solution θ_new =θ_current + r*T, where r is a random number between [-1,1]. Calculate the energy difference ΔE = J(θ_new) - J(θ_current). Decide whether to accept the new solution based on the Metropolis criterion. Introduce an adaptive step size strategy to dynamically adjust the perturbation range based on the acceptance rate. Repeat this process until the stopping criterion is reached. Output the simulated annealing optimization result θ_sa.

[0143] Step S433, input θ_sa obtained in S432 and J(θ) and C(θ) obtained in S431 into the response surface construction module. Design a central composite test plan around θ_sa. Perform the experiment and collect response data. Use weighted least squares method to fit the quadratic response surface model: J(θ) ≈ β0 + Σβi*θi + Σβii*θi^2 + Σβij*θi*θj. Perform variance analysis to evaluate the significance and goodness of fit of the model. If the model does not meet the requirements, add sampling points or adjust the model structure. Output the response surface model RSM and its statistical analysis report R_rsm.

[0144] Step S434: Input the RSM and R_rsm obtained in S433 and the θ_sa obtained in S432 into the sequential quadratic programming (SQP) optimization module. Use RSM as a proxy model for the objective function. Construct a quadratic programming subproblem, including linearized constraints and quadratic approximate objective functions. Use the active set method to solve the subproblem and obtain the search direction and step size. Update the current solution and repeat this process until convergence. During the optimization process, the actual objective function J(θ) is used to regularly correct the RSM prediction to ensure the accuracy of the optimization. Output the SQP optimization result θ_sqp.

[0145] Step S435: Input θ_sqp obtained in S434, M_ss and P_conv obtained in S41 into the parameter sensitivity analysis module. Use Morris screening method to calculate the basic effect of each parameter. Draw μ*-σ diagram to identify important parameters. Perform Sobol global sensitivity analysis on important parameters and calculate first-order and total effect sensitivity indexes. Based on the sensitivity results, fine-tune the sensitive parameters in θ_sqp. Output the optimized control parameter θ_opt and sensitivity analysis report R_sensitivity.

[0146] In this embodiment, by constructing a hybrid optimization framework, precise optimization of control parameters is achieved. This not only significantly improves the efficiency of control parameter optimization, but also achieves precise tuning of important parameters, which is particularly suitable for control parameter optimization of DC / DC converters in photovoltaic hydrogen production systems without energy storage.

[0147] According to one aspect of the present application, step S52 is specifically:

[0148] Step S521, input the pre-stored historical performance index data PI_hist into the data segmentation and stationarity test module. Use the sliding window method to divide the data into multiple time periods. Perform an augmented Dickey-Fuller (ADF) test on each time period to determine whether it is stationary. Perform differential processing on the non-stationary series until it passes the ADF test. Record the differential order of each series. Output the segmented stationary time series set TS_stationary and the corresponding differential order set D_order.

[0149] Step S522, input TS_stationary obtained in S521 into the multi-scale decomposition module. Apply the singular spectrum analysis (SSA) algorithm to each time series. Construct the trajectory matrix X and perform singular value decomposition X = UΣV^T. Analyze the singular value attenuation curve and use the steep slope method to select the number of principal components r. Reconstruct r principal components to obtain trend, cycle and noise components. For each component, calculate the variance contribution rate and autocorrelation function. Output the decomposed multi-scale component set MS_components and the statistical features S_features of each component.

[0150] Step S523, input the MS_components and S_features obtained in S522 into the ARIMA model identification module. For each main component (variance contribution rate is greater than the threshold), use the autocorrelation function (ACF) and partial autocorrelation function (PACF) diagrams to assist in determining the initial order (p, d, q) of the ARIMA model. Construct a set of candidate models. Use the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to select the optimal model. Perform model diagnosis to check the white noise and normality of the residuals. If the diagnosis fails, adjust the model order and refit. Output the ARIMA model parameter set P_arima and model diagnostic report R_diagnostic for each main component.

[0151] Step S524: Input P_arima and R_diagnostic obtained in S523 and MS_components obtained in S522 into the forecasting and reconstruction module. Perform multi-step ahead forecasts for each fitted ARIMA model. Calculate the forecast interval and use the Bootstrap method to estimate the forecast uncertainty. Perform a weighted combination of the forecast results of each component, with the weights based on the variance contribution rate in S_features. For the difference sequence, perform cumulative summation to restore the original scale. Calculate the merged forecast interval. Output the final forecast result TS_forecast, including point forecasts and interval forecasts.

[0152] Step S525: Input the TS_forecast and original PI_hist obtained in S524 into the trend analysis module. Use the Mann-Kendall trend test to determine the long-term trend of each performance indicator. For significant trends, use Sen's slope estimator to calculate the rate of change. Perform change point detection to identify the mutation point of the trend. Combine the SSA decomposition results to analyze the periodic pattern. Calculate the sliding mean and sliding standard deviation of each indicator to evaluate stability. Output the trend analysis report R_trend, which includes slope estimation, mutation points and stability indicators.

[0153] By constructing a multi-scale time series analysis framework, an in-depth evaluation of system performance is achieved. It can not only accurately evaluate the current performance of the system, but also reliably predict future performance trends, which is particularly suitable for performance monitoring and predictive maintenance of photovoltaic hydrogen production systems without energy storage.

[0154] According to one aspect of the present application, step S53 is specifically:

[0155] Step S531: Input R_trend obtained in S525, TS_forecast obtained in S524 and current system status data D_current into the fuzzy rule base generation module. Based on the trend information in R_trend, design adaptive fuzzy rules. Use FCM clustering algorithm to classify historical data and formulate specific fuzzy rules for each type of data. Introduce time weighting mechanism to increase the importance of recent data. Construct Takagi-Sugeno type fuzzy rules, IF (condition) THEN (linear model). Output fuzzy rule base FR_base and membership function parameters MF_params.

[0156] Step S532: Input FR_base, MF_params and D_current obtained in S531 into the fuzzy inference system. Calculate the trigger strength of each rule and synthesize the fuzzy relationship using the product inference method. Use the centroid method to defuzzify and obtain a clear output. Introduce an adaptive mechanism to dynamically adjust the membership function parameters according to the accuracy of the inference result. Output the fuzzy inference result FI_result, which contains the predicted system state and control suggestions.

[0157] Step S533, input the FI_result obtained in S532 and the Mss obtained in S41 into the model predictive control (MPC) framework construction module. Use Mss as the prediction model to construct the state equation and output equation. Define the prediction time domain Np and the control time domain Nc. Design the objective function, including tracking error, control effort and constraint penalty terms. Adjust the objective function weight according to the control suggestions in FI_result. Construct constraints, including input restrictions, output restrictions and rate of change restrictions. Output the MPC problem description, including the objective function J_mpc, constraints C_mpc and weight coefficient W_mpc.

[0158] Step S534: Input J_mpc, C_mpc and W_mpc obtained in S533 into the real-time optimization solver. Use the fast gradient method to solve the quadratic programming problem. Introduce the warm start strategy and use the solution of the previous step as the initial guess value. Implement the early termination mechanism to obtain the suboptimal solution within a limited time. Use the moving time domain method to only execute the optimal control input at the current moment. Output the optimal control sequence U_optimal and the predicted state sequence X_predict.

[0159] Step S535: Input U_optimal and X_predict obtained in S534 and FI_result obtained in S532 into the control strategy fusion module. Use the confidence weighting method to fuse the fuzzy reasoning results and MPC output. Design an adaptive weight adjustment mechanism to dynamically adjust the weights of the two methods according to the historical prediction accuracy. Consider the system response characteristics and introduce a smooth transition strategy to avoid drastic changes in the control input. Output the fused control strategy CS_fused, which contains the final control input U_final and the expected system trajectory X_expected.

[0160] Step S536: Input CS_fused and system real-time feedback data D_feedback obtained in S535 into the online learning and adaptation module. Calculate the prediction error and update the prediction model parameters. Use recursive least squares (RLS) to adjust the consequent parameters of the fuzzy rules in real time. Use the sliding time window method to regularly re-evaluate and update the fuzzy rule base. Introduce anomaly detection mechanism to identify and handle abnormal working conditions. Output the updated fuzzy rule base FR_base_updated and model parameters P_model_updated.

[0161] Step S537: Input FR_base_updated, P_model_updated and long-term performance data D_performance obtained in S536 into the performance evaluation and optimization module. Calculate control performance indicators such as average tracking error, energy efficiency and stability indicators. Use data envelopment analysis (DEA) method to evaluate the efficiency of the control strategy. Build a performance benchmark model and compare it with the historical optimal performance. If the performance degrades, trigger the global optimization process. Otherwise, perform local parameter fine-tuning. Output the performance evaluation report R_performance and optimization suggestions S_optimization.

[0162] Step S538: Input R_performance and S_optimization obtained in S537 and CS_fused obtained in S535 to the adaptive controller update module. According to S_optimization, adjust the prediction time domain and control time domain of MPC. Update the objective function weight to balance multiple control objectives. Modify the constraints to adapt to the changes in the dynamic characteristics of the system. Redesign the structure of the fuzzy rule base, such as adding or deleting rules. Use a progressive learning strategy to ensure a smooth transition of controller performance. Output the updated controller parameter set P_controller_updated and the adaptive adjustment report R_adaptation.

[0163] Step S539: Input P_controller_updated and R_adaptation obtained in S538 into the control strategy packaging and deployment module. Package the updated control strategy into an executable module. Design a rollback mechanism to allow rapid recovery to the previous stable version in an emergency. Implement online hot update function to minimize system interruption. Generate a control strategy document, including parameter settings, performance indicators and operation guides. Output the final control strategy package CS_package, including control algorithm, parameter configuration and deployment instructions.

[0164] In this embodiment, the adaptive optimization of system control is realized by establishing an intelligent fuzzy predictive control framework. It can not only realize the precise control of the system, but also has a strong adaptive learning ability, which is particularly suitable for the control of complex systems such as photovoltaic hydrogen production systems without energy storage, which are highly nonlinear and time-varying.

[0165] According to another aspect of the present application, there is also provided a non-storage off-grid photovoltaic hydrogen production control system, comprising:

[0166] at least one processor; and

[0167] a memory communicatively connected to at least one of the processors; wherein,

[0168] 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 off-grid photovoltaic hydrogen production control method without energy storage described in any of the above technical solutions.

[0169] According to one aspect of the present application, in addition to the control system, it also includes: a photovoltaic cell, an interleaved parallel three-level buck converter and a hydrogen production electrolyzer system; the input end of the interleaved parallel three-level buck converter is connected to the photovoltaic cell; the output end of the interleaved parallel three-level buck converter is connected to the hydrogen production electrolyzer.

[0170] It should be noted that photovoltaic cells are composed of multiple photovoltaic panels to form a photovoltaic array, and the photovoltaic array adopts the classic equivalent circuit form.

[0171] In this embodiment, the control method includes: the hydrogen production system can execute the control method according to light detection, time setting or manual control;

[0172] The upper and lower limits of the working power of the electrolytic cell are set to P max , P min The estimated maximum photovoltaic output power is P MPP , the electrolytic cell input power is P E ;

[0173] 1) When P MPP <Pmin When , the system is disconnected;

[0174] 2) When P MPP >P min And P E <P max When , the photovoltaic control enters the 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 voltage change a·|dP / dV|, where a is the proportional coefficient;

[0175] 3) When P E >P max When the photovoltaic control switches to the constant power mode, the voltage movement direction is constant and positive, and the voltage adjustment step is k(P E -P max ), where k is the step size coefficient, P E Power is input to the electrolyzer, and the process is iterated continuously until the electrolyzer power is equal to the upper limit of the working power of the electrolyzer.

[0176] The main goal of the converter control strategy is to track the input voltage, and the voltage error is compensated by PI to obtain the duty cycle. In addition, the average current equalization method is used to achieve parallel branch current equalization; in order to ensure the midpoint potential balance of the three-level circuit, the duty cycle offset control method is used to achieve capacitor voltage equalization.

[0177] According to one aspect of the present application, the parameter optimization method further includes: using a switching cycle average model method to perform small signal modeling analysis on the staggered parallel three-level buck converter under input voltage control. For ease of calculation, it is assumed that the converter is in an inductor current equalization state, the input voltage equalization state, and the circuit device parameters are balanced, that is, I L1 =I L2 =I L3 =I L4 , U cd1 =U cd2 , D S1 =D S2 =D S3 =D S4 , L1=L2=L3=L4.

[0178] In the hydrogen production system topology, the dynamic resistance r PV To represent the nonlinear relationship between photovoltaic voltage and current, the output voltage U o As a constant voltage load, further filter capacitor C o The voltage changes are ignored.

[0179] List the average model of KCL and KVL in the circuit:

[0180] (8)

[0181] (9)

[0182] By adding the corresponding small signal disturbance variable and eliminating the DC term and quadratic term, we can get its linearized model, and perform Laplace transform to get:

[0183] (10)

[0184] (11)

[0185] (12)

[0186] Keep the output voltage constant, that is =0, the transfer function of the duty cycle disturbance to the input voltage of the interleaved parallel three-level Buck converter can be obtained:

[0187] (12)

[0188] In order to achieve better dynamic response and steady-state characteristics in the entire working range, the working range is divided into five intervals according to the light intensity, and different PI control parameters are selected.

[0189] In the time period of 7:00~8:00, the photovoltaic voltage is an open circuit voltage because the photovoltaic power does not reach the working limit power of the electrolyzer. In the time periods of 8:00~12:00 and 13:00~17:00, as the light intensity changes, the MPPT algorithm stably tracks the maximum power point. In the time period of 12:00~13:00, the system is in the load power tracking mode, and the photovoltaic voltage deviates from the MPP. The photovoltaic voltage responds quickly and meets the cycle period requirement of less than 0.01s of the photovoltaic control algorithm. In the time period of 12:00~13:00, after a short transition, the electrolyzer power stabilizes at 4.8WM.

[0190] 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 method for controlling off-grid photovoltaic hydrogen production without energy storage, characterized in that: The steps include: Step S1, obtaining raw data from a light intensity sensor and a temperature sensor; using an isolation forest algorithm to detect and clean up the data; applying a Kriging interpolation algorithm to perform spatiotemporal series interpolation on the cleaned data; using an empirical mode decomposition method to perform multi-scale decomposition on the interpolated data; using a time series Shapelet algorithm to extract features from the decomposed data to obtain a light intensity feature vector and a temperature feature vector; Step S2, reading the light intensity feature vector and the temperature feature vector, identifying the photovoltaic cell model parameters by combining the particle swarm optimization algorithm and the Levenberg-Marquardt algorithm; based on the identified model parameters and the environmental feature vector, predicting the maximum power point by using the integrated method of support vector regression and Gaussian process regression; Step S3, historical operation data, combined with genetic algorithm and recursive least squares method to identify electrolytic cell model parameters; based on the identified model parameters and electrolytic cell specifications, using convex optimization algorithm to calculate the safe operating power range; according to the maximum power point and safe operating range, using dynamic programming algorithm to generate a power allocation strategy; Step S4, using the historical operation data of the converter and using a subspace identification method to identify converter model parameters; Based on the model parameters and power allocation strategy identified in step S3, the controller parameters are initialized using the pole placement method; the control parameters are optimized by combining the simulated annealing algorithm and the response surface method; Output and send optimized control parameters.

2. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 1 is characterized in that: Also includes: Step S5: During operation, real-time operation data of the system is obtained, and the performance index is calculated using a sliding time window method; Based on historical performance indicator data, singular spectrum analysis and ARIMA model are used to analyze performance trends; Based on the performance trends, optimized control parameters and power allocation strategy, an adaptive adjustment strategy is generated using fuzzy inference system and model predictive control methods.

3. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 2 is characterized in that: The step S1 is specifically as follows: Step S11, obtaining real-time readings of the light intensity sensor and the temperature sensor, collecting data at a predetermined frequency not less than 1 Hz, and obtaining raw light intensity data L_raw and raw temperature data T_raw; Step S12, inputting the original light intensity data L_raw and the original temperature data T_raw into the isolation forest algorithm; Calculate the isolation index for each data point. The isolation index is equal to the negative power of 2, where the power is the ratio of the average path length of the data point to the normalization factor of the data set size. Mark the data points whose isolation index is greater than the preset threshold as abnormal and remove them from the data set; output the light intensity data L_clean and temperature data T_clean after abnormal detection; Step S13, input the light intensity data L_clean and temperature data T_clean that have passed the anomaly detection into the Kriging interpolation algorithm; construct a spatiotemporal variation function, which is a function of spatial distance and time interval, and includes undetermined parameters a, b1, b2, r1 and r2; use the variational Bayesian method to estimate the undetermined parameters and minimize the prediction error variance; apply the estimated parameters to interpolate, and output the interpolated light intensity data L_interp and temperature data T_interp; Step S14, input the interpolated light intensity data L_interp and temperature data T_interp into the empirical mode decomposition algorithm; iteratively screen the signal to obtain a series of intrinsic mode functions; apply Hilbert-Huang transform to each intrinsic mode function to analyze its instantaneous frequency characteristics; group the intrinsic mode functions according to the frequency characteristics to obtain multi-scale components; output the light intensity data {L_1, L_2, ..., L_m} and temperature data {T_1, T_2, ..., T_m} after multi-scale decomposition; Step S15, input the multi-scale decomposed light intensity data {L_1, L_2, ..., L_m} and temperature data {T_1, T_2, ..., T_m} into the time series Shapelet algorithm; calculate the distance between the Shapelet and the time series subsequence, where the distance is defined as the minimum value of the square root of the sum of the squares of the differences between the Shapelet and the corresponding elements of the subsequence; select the most recognizable Shapelet using the information gain criterion; Construct a feature vector, where the vector elements are the distances between the selected Shapelet and the time series; output the light intensity feature vector F_L and the temperature feature vector F_T.

4. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 3 is characterized in that: The step S2 is specifically as follows: Step S21, input the light intensity feature vector F_L and the temperature feature vector F_T into the particle swarm optimization algorithm; wherein the objective function is the square sum of the difference between the measured current and the model current; update the particle position, and the position update amount is the weighted sum of the current speed, the individual optimal position and the global optimal position; input the updated particle position into the Levenberg-Marquardt algorithm for fine-tuning; calculate the Jacobian matrix and the residual vector, solve the normal equation to obtain the parameter increment; iterate the optimization until convergence, and output the photovoltaic cell model parameter vector P_pv; Step S22, input the photovoltaic cell model parameter vector P_pv and the light intensity feature vector F_L and the temperature feature vector F_T into the support vector regression model; use the kernel function to map the input to the high-dimensional feature space and calculate the regression function value; at the same time, input the same input data into the Gaussian process regression model, calculate the mean function and covariance function, and obtain the probability prediction result; perform weighted integration on the outputs of the two models, and the weights are determined by minimizing the historical prediction error; output the predicted maximum power point MPP_pred, which includes the maximum power point voltage, current and power value.

5. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 4 is characterized in that: The step S3 is specifically as follows: Step S31, input the pre-stored electrolytic cell historical voltage and current data [V_hist, I_hist] into the genetic algorithm; encode the electrolytic cell model parameters to form chromosomes; define the fitness function as the inverse of the sum of squares of the difference between the measured voltage and the model voltage; perform selection, crossover and mutation operations to generate a new population; input the individuals in the population into the recursive least squares algorithm, calculate the gain matrix, and update the parameter estimate; iterate and optimize until convergence, and output the electrolytic cell model parameter vector P_el; Step S32, inputting the electrolytic cell model parameter vector P_el obtained in step S31 and the pre-stored electrolytic cell specification parameter Spec_el into the convex optimization algorithm; constructing the objective function as the electrolytic cell power function, adding safety constraints such as maximum current, maximum temperature, etc.; applying the interior point method to solve the maximization problem, and obtaining the maximum safe power P_max; Similarly, solve the minimization problem to obtain the minimum safe power P_min; output the safe operating power range of the electrolyzer [P_min, P_max]; Step S33, input the maximum power point MPP_pred predicted in step S22 and the safe operating power range [P_min, P_max] of the electrolyzer obtained in step S32 into the dynamic programming algorithm; construct a state space, including circuit breaking, MPPT and constant power modes; construct an action space including switching modes and adjusting power; construct a reward function, considering system efficiency, hydrogen production efficiency and power loss; use the value iteration method to solve the optimal strategy, iteratively update the value function until convergence; output the power allocation strategy S_power, including the operating mode and the target power value.

6. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 5 is characterized in that: The step S4 is specifically as follows: Step S41, input the pre-stored converter historical operation data [V_in_hist, I_in_hist, V_out_hist, I_out_hist, D_hist] into the subspace identification algorithm; construct a state space model, use singular value decomposition to estimate the system order; use the projection algorithm to estimate the system matrix A, B, C, D; extract physical parameters from the system matrix, and output the converter model parameter P_conv; Step S42, input the converter model parameter P_conv obtained in step S41 and the power allocation strategy S_power obtained in step S33 into the pole configuration algorithm; calculate the characteristic equation coefficients according to the desired closed-loop poles; solve the Diophantine equation to obtain the controller parameter polynomial; extract the PID parameters from the polynomial and output the initial control parameter θ_init; Step S43, input the initial control parameter θ_init obtained in step S42, the converter model parameter P_conv obtained in step S41, the power allocation strategy S_power obtained in step S33 and the maximum power point MPP_pred predicted in step S22 into the simulated annealing algorithm; define the objective function, taking into account the adjustment time, overshoot and steady-state error; generate a new parameter θ_new according to the current temperature parameter; The response surface method is used to construct a second-order polynomial approximation model of the objective function; the simulated annealing algorithm and the response surface model are combined to perform parameter search and optimization; and the optimized control parameter θ_opt is output.

7. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 6 is characterized in that: The step S5 is specifically as follows: Step S51, when working, obtain the real-time operation data D_real of the system, including the photovoltaic voltage and current, the electrolyzer voltage and current, and the converter voltage and current; define the sliding time window size W and the sliding step size S; In each time window, the system efficiency, hydrogen production efficiency, power loss and response time are calculated; the system efficiency is the ratio of output power to input power, the hydrogen production efficiency is the ratio of the high calorific value of hydrogen to the input energy of the electrolyzer, the power loss is the difference between input power and output power, and the response time is calculated by step response analysis; Output performance index vector PI; Step S52, input the pre-stored historical performance index data PI_hist into the singular spectrum analysis algorithm; construct a trajectory matrix, perform singular value decomposition on the covariance matrix; select the eigenvectors corresponding to the main eigenvalues ​​to reconstruct the time series, and obtain the trend, cycle and noise components; Input the extracted trend component into the ARIMA model, determine the difference order, autoregression order and moving average order; fit the ARIMA model, calculate the slope of each indicator; output the performance trend vector T_perf; Step S53, input the performance trend vector T_perf obtained in step S52, the optimized control parameter θ_opt obtained in step S43 and the power allocation strategy S_power obtained in step S33 into the fuzzy inference system; construct a fuzzy rule base, and use the Takagi-Sugeno fuzzy inference method to calculate the output; at the same time, input the same input data into the model predictive control algorithm to predict the future system state and minimize the weighted sum of the prediction error and the control effort; combine the outputs of the fuzzy inference system and the model predictive control to generate the final adjustment strategy S_adjust, including the control parameter adjustment amount and the power allocation strategy adjustment amount.

8. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 7 is characterized in that: The step S12 is specifically as follows: Step S121: inputting the original light intensity data L_raw and the original temperature data T_raw obtained in step S11 into the multi-dimensional data stream feature extractor; The data is segmented using the sliding window method, and the window size is adaptively adjusted according to the time characteristics of the data; statistical characteristics are calculated for the data in each window, including mean, variance, skewness, kurtosis, and autocorrelation coefficient; Output feature matrix F_raw; Step S122: input the feature matrix F_raw obtained in step S121 into the local anomaly factor algorithm; calculate the average distance between each data point and its k nearest neighbors, where the k value is determined by an adaptive method; introduce a new density estimation function, taking into account the distribution characteristics of data points in different dimensions; calculate the local anomaly factor of each data point, where the local anomaly factor is equal to the ratio of the k-distance of the data point to the average k-distance of its k nearest neighbors; output the anomaly score vector S_anomaly; Step S123: input the anomaly score vector S_anomaly obtained in step S122 into the dynamic threshold generator; use the kernel density estimation method to fit the probability distribution of the anomaly score; introduce an adaptive bandwidth selection algorithm to dynamically adjust the kernel function bandwidth according to the local density characteristics of the data; calculate the cumulative distribution function of the anomaly score distribution, and determine the dynamic threshold by the inverse function method; output the dynamic threshold T_dynamic; Step S124: input the anomaly score vector S_anomaly obtained in step S122 and the dynamic threshold T_dynamic obtained in step S123 into the adaptive anomaly marker; compare the anomaly score of each data point with the dynamic threshold, and mark the data point with anomaly score higher than the threshold as a potential anomaly; introduce a confidence evaluation mechanism to calculate the confidence of each potential anomaly point, and the confidence is proportional to the difference between the anomaly score and the threshold; output the anomaly labeling matrix M_anomaly, which contains anomaly labels and confidence information; Step S125: inputting the anomaly marking matrix M_anomaly obtained in step S124, as well as the raw light intensity data L_raw and the raw temperature data T_raw into the context-aware anomaly verifier; For each potential anomaly point, extract its temporal and spatial context information; design a new context similarity metric that takes into account the temporal continuity and spatial correlation of the data; calculate the similarity between the potential anomaly point and its context, and if the similarity is lower than a preset threshold, confirm it as a real anomaly; Output the verified abnormal marking matrix M_verified; Step S126: input the verified abnormal marking matrix M_verified, the original light intensity data L_raw and the original temperature data T_raw obtained in step S125 into the adaptive abnormal repairer; for the confirmed abnormal points, use the local linear embedding (LLE) algorithm to repair; construct the manifold structure of the abnormal point neighborhood, considering the continuity constraint of the time series; optimize the reconstruction weights so that the repaired data points maintain the original geometric structure locally; for the data points that cannot be reliably repaired, mark them as missing values; output the repaired light intensity data L_clean and temperature data T_clean, and the missing value marking matrix M_missing; Step S127: inputting the restored light intensity data L_clean and temperature data T_clean obtained in step S126, and the missing value marking matrix M_missing into the data quality assessor; calculating various statistical indicators of the cleaned data, including the effective data ratio, data distribution characteristics and time continuity indicators; Use a multi-criteria decision-making method to comprehensively evaluate data quality and generate a data quality report Q_report; output the repaired light intensity data L_clean and temperature data T_clean, as well as the missing value marking matrix M_missing and the data quality report Q_report.

9. The off-grid photovoltaic hydrogen production control method without energy storage according to claim 7, characterized in that: The step S13 is specifically as follows: Step S131, input the restored light intensity data L_clean and temperature data T_clean obtained in step S12, and the missing value marking matrix M_missing into the spatiotemporal correlation analyzer; use Moran's I index to calculate the spatial autocorrelation of the data, considering the temporal continuity of the data; apply wavelet transform to perform multi-scale decomposition on the data, and analyze the spatiotemporal correlation characteristics at different scales; output the spatiotemporal correlation feature matrix C_spatio_temporal; Step S132, input the spatial correlation feature matrix C_spatio_temporal obtained in step S131 into the adaptive variogram generator; design a new non-stationary variogram model, allowing the variogram parameters to change with time and spatial position; use the local weighted regression method to estimate the variogram parameters, and the weights are based on the temporal and spatial distances of the data points; output the adaptive variogram parameter set P_variogram; Step S133, input the adaptive variogram parameter set P_variogram obtained in step S132, and the restored light intensity data L_clean, temperature data T_clean and missing value marking matrix M_missing in step S12 into the multi-scale Kriging interpolator; decompose the data at multiple scales, and apply the Kriging interpolation algorithm at each scale; introduce a new nonlinear kernel function to improve the accuracy of interpolation in non-stationary areas; use an adaptive neighborhood selection strategy to dynamically adjust the interpolation neighborhood according to the local data density; output the preliminary interpolation result I_preliminary; Step S134, input the preliminary interpolation result I_preliminary obtained in step S133 into the variational Bayesian optimizer; construct a new probabilistic graphical model, and formalize the interpolation problem into a Bayesian inference task; design a variational objective function, while considering the data fit and the smoothness of the interpolation result; use the stochastic gradient variational inference algorithm to optimize the model parameters, and introduce an adaptive learning rate adjustment mechanism; output the optimized interpolation result I_optimized; Step S135, input the optimized interpolation result I_optimized obtained in step S134 into the uncertainty quantification module; use the Monte Carlo method to generate multiple samples of the interpolation results; calculate the confidence interval and prediction variance of each interpolation point; introduce a new uncertainty propagation model to evaluate the impact of interpolation uncertainty on subsequent analysis; output the final interpolation result I_final and uncertainty matrix U_interpolation with uncertainty estimation.

10. A non-storage off-grid photovoltaic hydrogen production control system, characterized in that: include: at least one processor; as well as a memory communicatively connected to at least one of the processors; 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 off-grid photovoltaic hydrogen production control method without energy storage as described in any one of claims 1 to 9.

Images