A distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization

Abstract

The application discloses a kind of distributed photovoltaic water electrolysis hydrogen production control methods based on artificial intelligence optimization, it is related to new energy and hydrogen energy technical field, the method first real-time acquisition photovoltaic array and electrolytic cell and so on operating data and carry out pretreatment;Then the double-layer optimization model including upper mixed integer programming model and lower deep learning model is established, wherein the upper model is with system total operating cost or unit hydrogen production cost minimum as target, lower model is LSTM time series prediction model, for predicting system efficiency;Utilize historical data to train lower model and update online;In each control cycle, the predicted information of lower model is input into upper model, and the optimal control sequence is solved by rolling horizon optimization strategy and is executed down.This application also considers electrolytic cell attenuation cost, multiple safety constraints and fault-tolerant control, suitable for the distributed hydrogen production system of multiple electrolytic cells in parallel, can realize system overall efficiency optimization.

Description

Technical Field

[0001] This invention relates to the fields of new energy and hydrogen energy technology, specifically to a distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, the development and utilization of renewable energy has become a core part of national energy strategies. Among these, solar energy, with its advantages of wide distribution, zero pollution, and inexhaustibility, has become one of the most promising clean energy sources. However, photovoltaic power generation is characterized by significant intermittency and volatility, with its output power severely affected by weather, seasons, and day / night cycles, leading to substantial "curtailment" of solar power. This not only wastes energy but also hinders the further development of the photovoltaic industry.

[0003] Hydrogen energy, as a highly efficient and clean secondary energy carrier, boasts high energy density, storability, and zero carbon emissions, and is considered a crucial component of the future energy system. Utilizing surplus photovoltaic power to produce hydrogen through water electrolysis—known as "green hydrogen"—is an effective way to achieve large-scale consumption of renewable energy and enhance the flexibility of the energy system. This "electricity-hydrogen" coupling system not only solves the grid connection challenges of photovoltaic systems but also provides green fuel for transportation, industry, and other sectors, demonstrating broad application prospects.

[0004] However, combining distributed photovoltaic power generation with water electrolysis for hydrogen production still faces many technical challenges in building a stable, efficient, and economical hydrogen production system: 1. Source-load matching and volatility issues: The randomness and intermittency of photovoltaic power output make it difficult for electrolyzers to obtain a continuous and stable power supply. Operating at maximum power directly may lead to decreased efficiency or even damage to the electrolyzers during periods of insufficient sunlight; frequent start-ups and shutdowns will increase equipment wear and maintenance costs. The key to stable system operation lies in how to rationally adjust the electrolyzer's operating status according to changes in photovoltaic output, achieving either "source follows load" or "load follows source."

[0005] Multi-device coordination and optimal scheduling problem: In practical engineering, distributed hydrogen production systems with multiple electrolyzers operating in parallel are often used to improve hydrogen production efficiency and system reliability. Different electrolyzers may have different operating characteristics and health conditions. How to coordinate and schedule multiple electrolyzers to achieve the highest overall system efficiency and lowest cost while meeting hydrogen production needs is a complex optimization problem.

[0006] The balance between system economy and lifespan: The number of start-ups and shutdowns and the operating efficiency of the electrolyzer directly affect its lifespan and degradation costs. Frequent start-ups and shutdowns accelerate the aging of electrode materials and increase equipment replacement costs. Therefore, in the process of optimizing control, it is necessary to comprehensively consider hydrogen production revenue, operating energy consumption, equipment depreciation and maintenance costs, as well as the lifespan loss of the electrolyzer, to seek the optimal balance between economy and equipment reliability.

[0007] Limitations of traditional control methods: Existing hydrogen production control systems are mostly based on empirical rules or simple feedback control, making it difficult to accurately predict the changing trends of photovoltaic output, and also unable to fully consider the dynamic characteristics of the electrolyzer and the complex coupling relationships between multiple devices. When faced with complex operating conditions and uncertainties, traditional control methods often fail to achieve global optimization, resulting in low system operating efficiency and poor economic performance.

[0008] In conclusion, developing an intelligent control method that can adapt to fluctuations in photovoltaic output, achieve coordinated optimization of multiple electrolyzers, and balance economic efficiency and equipment lifespan is of great significance for improving the performance and application value of distributed photovoltaic water electrolysis hydrogen production systems. This requires technological innovation in data acquisition and processing, system modeling, optimization algorithms, and intelligent prediction to meet the demands of modern energy systems for efficient, intelligent, and reliable control. Summary of the Invention

[0009] The purpose of this invention is to provide a distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization. By integrating deep learning prediction and mixed integer programming optimization technology, intelligent collaborative scheduling of the electrolyzer group is achieved, which can effectively reduce hydrogen production costs and extend the service life of key equipment while ensuring the safe and stable operation of the system.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization, comprising the following steps: S1: By deploying various sensors and data acquisition modules at the photovoltaic array site and hydrogen production equipment, the output power of the distributed photovoltaic array, ambient irradiance, ambient temperature, operating parameters of the electrolyzer, and status information of the system energy storage unit are obtained in real time. Then, the collected raw data is preprocessed by cleaning, denoising, and normalizing to eliminate outliers and dimensional differences in the data, providing a high-quality data foundation for subsequent modeling. S2: Construct a two-layer optimization framework consisting of an upper-layer mixed integer programming model and a lower-layer deep learning model. The upper-layer model is responsible for formulating global device scheduling strategies, while the lower-layer model focuses on mining deep patterns in system operation data. The two models are coupled through an information transmission interface to form a closed-loop optimization system. S3: Use the historical operating data accumulated over a long period of time as a training set to iteratively train the lower-level deep learning model, and continuously adjust the model parameters based on newly generated online data during system operation to ensure that the model can adapt to the changes in characteristics caused by photovoltaic power output fluctuations and equipment aging, and maintain prediction accuracy. S4: Within each fixed control cycle, the predicted information about system efficiency or revenue output by the lower-level model is fed back to the upper-level model. Based on this information and combined with the current system state, the upper-level model uses a rolling time-domain optimization algorithm to solve for the globally optimal control sequence for a future period of time. S5: The optimal control commands calculated by the upper-level model, including the start-stop commands and power setpoints of each electrolytic cell, are sent to the corresponding actuators, such as power electronic converters or electrolytic cell controllers, via industrial bus or wireless communication, thereby precisely adjusting the operating status of the electrolytic cells.

[0011] Furthermore, in step S2, the upper-level hybrid integer programming model constructs a mathematical model with the minimum total system operating cost and / or the minimum unit hydrogen production cost as the core objective function. Its decision variables explicitly include the binary start-stop state variables of the electrolyzer and the continuous target power setting value. The constraints of the model at least cover the inherent start-stop logic constraints of the electrolyzer, the power ramp-up rate constraints to prevent equipment damage, the upper and lower limit constraints of the power within the rated range of the equipment, and the power balance and reverse current restriction constraints when interacting with the external power grid, so as to ensure the feasibility of the solution in engineering.

[0012] Furthermore, in step S2, the lower-level deep learning model is specifically designed as a time series prediction model. This model is configured to receive a multi-dimensional input tensor including a photovoltaic power prediction sequence, the current system state vector, and historical operating data. After feature extraction and nonlinear mapping by the neural network, it outputs the predicted value of the system's comprehensive efficiency or expected return index under different electrolytic cell operating schemes in the future period, providing key value assessment basis for upper-level optimization.

[0013] Furthermore, the lower-level deep learning model adopts a Long Short-Term Memory (LSTM) network architecture, which utilizes its internal memory units and gating mechanisms such as forget gates, input gates, and output gates to effectively handle long-term dependencies in the input sequence, accurately capture the intermittent and fluctuating characteristics of photovoltaic power output affected by weather, and the nonlinear dynamic characteristics of electrolyzer operating parameters changing over time, thereby improving the robustness and accuracy of the prediction model.

[0014] Furthermore, in step S4, the specific implementation of the rolling time-domain optimization strategy is as follows: at each discrete sampling time, a future time domain of finite length is selected as an optimization window, and the optimization sub-problems within the window are solved to obtain a complete set of control sequences; in actual execution, only the control quantity of the first control cycle in the sequence is used on the system, and when the next sampling time arrives, the actual state of the system is observed again, and the optimization time domain is rolled forward, and the optimization calculation and execution are repeated to adapt to the dynamic changes of the system.

[0015] Furthermore, in the upper-level mixed integer programming model, a quantitative index of the electrolytic cell's attenuation cost is introduced, which is mathematically correlated with its actual operating efficiency and cumulative start-stop count. At the same time, the real-time efficiency prediction value output by the lower-level model is used as a feedback signal to dynamically adjust the weight coefficient of the attenuation cost in the overall objective function, so that the upper-level optimization strategy can adaptively balance the relationship between operating costs and equipment lifespan loss according to the current health status of the equipment.

[0016] Furthermore, step S6 is included: throughout the optimization solution and execution process, a multi-level fault diagnosis and safety protection mechanism is established, and hard constraints for overvoltage protection, overcurrent protection, overheat protection of the electrolytic cell, and reverse current protection against backfeeding from the power grid are embedded into the control logic. Once the system status parameters are detected to exceed the safety threshold or an abnormal operating condition occurs, the preset graded safety shutdown procedure is immediately triggered to cut off the relevant power supply and trigger an alarm, thereby ensuring the safety of equipment and personnel.

[0017] Furthermore, it also includes step S7: designing system-level fault-tolerant control logic. When the communication link between the main controller and the field equipment is interrupted, or the system encounters an unforeseen serious fault that causes the optimization algorithm to fail to operate normally, the system can automatically detect the fault state and switch to a fault-tolerant control mode based on predefined safety rules. This mode maintains the basic operation of the system or performs an orderly shutdown according to a conservative strategy to avoid paralyzing the entire hydrogen production system due to a single point of failure.

[0018] Furthermore, the data preprocessing in step S1 not only includes conventional missing value imputation and outlier removal, but also includes normalizing the collected data to eliminate the influence of different physical dimensions. On this basis, statistical features reflecting photovoltaic power output characteristics (such as volatility and correlation) and degradation features reflecting the health status of the electrolyzer (such as resistance change trend) are extracted, providing richer input dimensions for feature engineering and model training.

[0019] Furthermore, this method is applied to a distributed hydrogen production system containing multiple electrolyzers configured in parallel. The upper-level mixed integer programming model has the ability to coordinate and schedule multiple devices, and can treat each electrolyzer as an independent decision-making unit. By establishing a large optimization space that includes the start-up and shutdown states and power allocation of all electrolyzers, the system can solve for the equipment combination operation scheme that maximizes the overall energy conversion efficiency or economic benefits of the system.

[0020] This invention provides a distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization, which has the following beneficial effects: 1. Significantly improves system economy and operational efficiency This method achieves a deep integration of global optimization and precise local control by constructing a two-layer intelligent architecture of "upper-layer optimization + lower-layer prediction". The upper-layer mixed-integer programming model can comprehensively consider start-up and shutdown costs, power ramp-up, and grid interaction constraints, making decisions with the goal of minimizing operating costs or hydrogen production costs; while the lower-layer LSTM deep learning model can accurately predict future photovoltaic output and system efficiency under different operating schemes. This combination enables the control system to move beyond just responding to the current moment and instead possess "predictive" capabilities, allowing it to adjust the operating combination and power allocation of the electrolyzer cluster in advance before solar radiation fluctuations. It effectively solves the problem of frequent start-up and shutdown losses caused by the intermittent nature of photovoltaic power in traditional control strategies. By optimizing the coordinated scheduling of electrolyzers, it maximizes the absorption of photovoltaic green electricity, thereby significantly reducing equipment degradation costs and energy consumption, and improving overall hydrogen production efficiency and economic returns.

[0021] Significantly enhances the system's adaptability to complex environments. Leveraging the powerful nonlinear fitting and temporal feature capture capabilities of Long Short-Term Memory (LSTM) networks, this method can deeply mine the complex mapping relationships between photovoltaic irradiance, temperature, and electrolyzer performance contained in historical operating data. Compared to traditional physical mechanism models, this deep learning model is more sensitive to environmental changes and can more accurately predict the actual operating efficiency of the system under different weather conditions. This allows the control system to dynamically adjust its control strategy based on prediction results even when facing complex weather changes such as cloudy or rainy weather, quickly adapting to changes in the external environment, maintaining the stability of system operation, and avoiding energy waste or equipment damage caused by inaccurate predictions.

[0022] Effectively extend the service life of core equipment and reduce maintenance costs This invention innovatively introduces a dynamic attenuation cost calculation mechanism based on operating efficiency and the number of start-stop cycles into the upper-level optimization model. The system no longer simply pursues instantaneous power matching, but instead incorporates the health status of the electrolyzer into the optimization objective. By dynamically adjusting the weight of attenuation costs based on efficiency predictions fed back from the lower-level model, the system automatically avoids high-load or frequent start-stop conditions that, while meeting power requirements, would accelerate equipment aging during the optimization process. This "soft protection" mechanism fundamentally reduces the loss of electrolyzer electrode materials and thermal stress impact, significantly extends the design life of the equipment, and lowers the user's long-term operation and maintenance costs and equipment replacement frequency.

[0023] Comprehensive protection of system operation security and reliability This method incorporates a multi-layered safety protection system. Firstly, it embeds multiple hard constraints, including overvoltage, overcurrent, overheating, and backflow prevention, into the core logic of the optimization solution, ensuring that any control command generation conforms to the equipment's safety boundaries. Secondly, it employs a robust fault-tolerance mechanism, allowing the system to seamlessly switch to a degraded operation mode based on predefined rules when the main communication link is interrupted or unpredictable sudden failures occur. This dual strategy of "online optimization + offline backup" ensures that the hydrogen production system maintains basic safe operation even under extreme conditions or partial subsystem failures, preventing safety accidents and significantly improving the system's robustness in unattended environments.

[0024] Optimize resource allocation and achieve fine-grained cluster management For large-scale distributed hydrogen production scenarios involving multiple electrolyzers operating in parallel, the upper-level mixed-integer programming model of this method possesses powerful collaborative scheduling capabilities. Like a conductor, it intelligently allocates power generation tasks based on the differences in model, health status, and current marginal efficiency of different electrolyzers. For example, it prioritizes starting high-efficiency electrolyzers operating at their optimal temperature, while temporarily suspending inefficient or cold-running equipment. This refined cluster management mode breaks through the limitations of independent control of individual devices, achieving the optimal energy efficiency solution across the entire system. It is particularly suitable for distributed renewable energy access scenarios with high volatility, possessing extremely high engineering application value and promotion potential. Attached Figure Description

[0025] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0026] Figure 1 This is the overall control flowchart of the present invention; Figure 2 This is a flowchart of the two-layer optimization model structure of the present invention; Figure 3 This is a flowchart of the lower-level model training and updating process of the present invention; Figure 4 This is a flowchart of the rolling time-domain optimization execution process of the present invention; Figure 5 This is a flowchart of the security protection and fault tolerance mechanism of the present invention. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] How to use: Step 1: Data Acquisition and Preprocessing First, the system collects real-time data on the output power of the distributed photovoltaic array, ambient irradiance, ambient temperature, operating parameters of the electrolyzer, and the status information of the system's energy storage units. Then, this raw data undergoes preprocessing, including data cleaning, normalization, and extraction of statistical features reflecting the photovoltaic output characteristics and the health status of the electrolyzer, to ensure the input quality of subsequent models.

[0030] Step 2: Constructing a two-level optimization model Construct a two-layer optimization model. The upper layer is a mixed-integer programming model, and the lower layer is a deep learning model. The upper-layer model is responsible for global decision-making, and the lower-layer model is responsible for local prediction.

[0031] Step 3: Training and online updating of the lower-level deep learning model Historical operating data is used to train the lower-level deep learning model. This lower-level model is a time series prediction model, employing a Long Short-Term Memory (LSTM) network architecture. It captures the nonlinear temporal characteristics of photovoltaic power output and system operation through its internal memory units and gating mechanisms. It is configured to receive photovoltaic power prediction sequences, current system state, and historical operating data as input, and output predicted system efficiency values ​​under different electrolyzer operating schemes over a future period. After training, the model is updated online based on the latest operating data to maintain its prediction accuracy.

[0032] Step 4: Solving the upper-level mixed-integer programming model and rolling optimization In each control cycle, the predicted information output by the lower-level model is input into the upper-level mixed-integer programming model. The upper-level model takes the minimum total system operating cost and / or the minimum unit hydrogen production cost as its objective function, and its decision variables include the start-up and shutdown status of the electrolyzer and the target power setpoint. The constraints include at least the start-up and shutdown logic constraints of the electrolyzer, the power ramp-up rate constraints, the upper and lower power limit constraints, and the grid interaction power constraints. In particular, the degradation cost of the electrolyzer is related to its operating efficiency and the number of start-ups and shutdowns, and the efficiency prediction value output by the lower-level model is used to dynamically adjust the weighting coefficient of the degradation cost.

[0033] To solve the upper-level model, a rolling time-domain optimization strategy is adopted: at each sampling time, an optimization subproblem in a finite time domain is solved, the solution of the optimization subproblem in the first control cycle is executed, and the system state is reacquired at the next sampling time, and the above optimization process is repeated.

[0034] Step 5: Issuance of control commands The optimal control command obtained from the solution is sent to the actuator to control the operating state of the electrolytic cell, thereby achieving precise regulation of the start-up and shutdown status and power of the electrolytic cell.

[0035] Step 6: Embedding Multiple Security Constraints Throughout the optimization and execution process, multiple safety constraints are embedded, including overvoltage protection, overcurrent protection, overheat protection, and grid reverse current protection for the electrolytic cell. Once a system anomaly is detected, a pre-set safety shutdown procedure is immediately triggered to ensure the safety of equipment and personnel.

[0036] Step 7: Fault-Tolerant Control When communication between the system and the main controller is interrupted or an unpredictable fault occurs, the system automatically switches to a fault-tolerant control mode based on predefined rules to maintain the basic safe operation of the system and avoid system paralysis due to sudden failures.

[0037] Example: Example 1: Basic Control Flow Based on LSTM Prediction and Rolling Optimization This embodiment demonstrates a basic application of a distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization. The method first executes step S1, which involves real-time acquisition of the distributed photovoltaic array's output power, ambient irradiance, ambient temperature, electrolyzer operating parameters, and the status information of the system's energy storage unit during the initial system startup phase. This raw data is then cleaned and normalized, and statistical features reflecting the photovoltaic output characteristics are extracted to provide a high-quality data foundation for subsequent model calculations.

[0038] In step S2, the system establishes a core two-layer optimization model. The upper layer of this model is a mixed integer programming model, designed to minimize the total operating cost of the system; the lower layer is a time series prediction model using a Long Short-Term Memory (LSTM) network architecture. This model utilizes its internal memory units and gating mechanisms to specifically capture the nonlinear time series characteristics of photovoltaic power output and system operation.

[0039] In step S3, the system fully trains the lower-level LSTM model using historical operating data and updates its model parameters online to ensure high accuracy in predicting future system efficiency. Subsequently, in step S4, the system enters the core rolling time-domain optimization phase. In each control cycle, the lower-level LSTM model receives the current photovoltaic power prediction sequence and system state, outputting the efficiency prediction values ​​under different electrolyzer operating schemes over a future period. The upper-level mixed-integer programming model uses these predictions as input, combining them with electrolyzer start-up / shutdown logic constraints and power ramp-up constraints to solve for the optimal control sequence. Specifically, by solving a finite-time-domain optimization subproblem at each sampling time, only the solution for the first control cycle is executed, and re-optimization occurs at the next time step.

[0040] Finally, step S5 sends the optimal control commands obtained from the solution regarding the electrolytic cell's start-up and shutdown status and the target power setpoint to the actuator. The entire process also incorporates multiple safety constraints, such as overvoltage, overcurrent, and overheat protection, as described in step S6, ensuring that a safe shutdown procedure can be triggered in the event of an anomaly, thereby achieving stable system operation in complex environments.

[0041] Example 2: Dynamic weighted optimization considering the cost of electrolyzer degradation This embodiment details how to consider the lifespan cost of the electrolyzer in the control method. After data preprocessing in step S1, the system constructs a two-layer optimization model as described in claims 2 and 6. The upper-layer mixed-integer programming model not only aims to minimize operating costs but also pays special attention to the long-term degradation cost of the electrolyzer.

[0042] The unique feature of the system in executing step S4 is the introduction of a dynamic adjustment mechanism. After receiving the photovoltaic power prediction sequence and historical operating data, the lower-level deep learning model (LSTM) outputs predicted system efficiency values ​​under different electrolyzer operating schemes over a future period. When constructing the objective function, the upper-level model correlates the electrolyzer's degradation cost with its operating efficiency and the number of start-ups and shutdowns. More importantly, the efficiency prediction values ​​output by the lower-level model are used to dynamically adjust the weighting coefficients of the degradation cost. For example, when sufficient sunlight and high system efficiency are predicted in the future, the algorithm appropriately relaxes the restrictions on the number of start-ups and shutdowns, prioritizing continuous operation to reduce the unit hydrogen production cost; conversely, it increases the weight of the degradation factor in the cost calculation to reduce unnecessary frequent start-ups and shutdowns.

[0043] In this way, the system can dynamically adjust the focus of optimization objectives based on the latest efficiency predictions within each control cycle. Subsequently, an optimal control command that balances economy and equipment lifespan is solved through a rolling time-domain optimization strategy and then executed. This method effectively solves the problem of neglecting equipment degradation in traditional control, significantly extending the service life of the electrolyzer.

[0044] Example 3: Parallel Cooperative Scheduling and Fault Tolerance Application of Multiple Electrolyzers This embodiment addresses a large-scale distributed hydrogen production system comprising multiple electrolyzers connected in parallel. First, following step S1, the system collects data from the entire array and each branch and performs preprocessing. In step S2, when constructing the two-layer model, the upper-layer mixed-integer programming model is specifically configured to coordinate the scheduling of multiple electrolyzers.

[0045] During the rolling optimization process in step S4, the upper-level model not only controls a single electrolyzer but also treats the start-up and shutdown states of multiple electrolyzers and the target power setpoint as unified decision variables. Based on the comprehensive efficiency predictions output by the lower-level LSTM network, the model calculates the optimal allocation scheme that maximizes the overall efficiency of the entire system. For example, during periods of fluctuating illumination, the algorithm might instruct some high-efficiency electrolyzers to operate at full capacity, while allowing inefficient or preheating electrolyzers to remain in standby mode, thereby maximizing cluster benefits.

[0046] Furthermore, this embodiment also includes the fault-tolerant mechanism described in step S7. In actual operation, if the communication between the main controller and the field equipment is suddenly interrupted, or an unpredictable hardware failure is detected, the control system will automatically and seamlessly switch to a fault-tolerant control mode based on predefined rules. In this mode, the system no longer relies on complex AI optimization calculations, but maintains the basic safe operation of the remaining healthy electrolyzers according to preset safety rules and priorities, preventing the entire hydrogen production system from being interrupted due to a single point of failure, thus greatly improving the robustness and reliability of the system.

[0047] Example 4: Safety Protection and Adaptive Adjustment under Extreme Weather This embodiment simulates the practical application of the control method under severe weather conditions. In the initial stage, the system strictly follows step S1 to collect environmental irradiance and temperature data. Due to the rapid changes in the external environment, the online update frequency of the lower-level deep learning model in step S3 is automatically increased to enhance the model's adaptability to sudden environmental changes.

[0048] When entering the control cycle of step S4, although the lower-level LSTM model tries its best to predict future photovoltaic output, uncertainties in the external environment may lead to prediction deviations. At this time, the embedded multiple safety constraints described in step S6 play a crucial role. The system monitors the voltage, current, and temperature of the electrolyzer in real time. Once abnormal voltage fluctuations are detected due to a sudden drop in sunlight or other reasons, or if the system is about to experience a reverse current risk, the safety protection module will immediately intervene.

[0049] In this scenario, the solution result of the upper-level mixed-integer programming model is forcibly corrected by safety constraints. For example, even if the optimization algorithm instructs the system to increase power, if the safety monitoring module detects an overheating trend, the system will refuse to execute the instruction and trigger a preset safety shutdown procedure to prioritize equipment safety. Once environmental conditions improve and the system returns to normal, the control flow will return to the normal "data acquisition-prediction-optimization" closed loop. This process fully demonstrates the adaptive control capability of this invention under the principle of safety priority.

[0050] Example 5: Full Life Cycle Energy Efficiency Management and Continuous Optimization This embodiment focuses on demonstrating the energy efficiency management advantages of this method during long-term operation. After system deployment, the process from S1 to S5 is continuously executed in a loop. In step S3, the lower-layer deep learning model is not static but continuously absorbs new operational data to update its network weights as the running time progresses. This makes the model's fitting of the light characteristics and electrolyzer aging characteristics of this specific geographical location increasingly accurate.

[0051] In the rolling time-domain optimization of step S4, this continuously evolving model provides increasingly accurate predictions of future efficiency for higher-level mixed-integer programming. This enables the upper-level model to formulate more forward-looking scheduling strategies. For example, by analyzing long-term changes in decay cost weights, the system can learn the optimal electrolyzer maintenance window in different seasons or time periods, thereby consciously avoiding certain high-loss periods in optimized scheduling.

[0052] Ultimately, through the optimal control commands issued in step S5, the entire distributed photovoltaic water electrolysis hydrogen production system maintains high energy conversion efficiency and low overall operating costs throughout operation, even over months or years. This method, through the deep integration of artificial intelligence technology and mathematical programming, achieves a leap from instantaneous control to long-term asset management, providing users with a truly intelligent, full-lifecycle hydrogen production solution.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization, characterized in that, Includes the following steps: S1: Real-time acquisition of the output power of the distributed photovoltaic array, ambient irradiance, ambient temperature, operating parameters of the electrolyzer, and status information of the system energy storage unit, and preprocessing of the data; S2: Establish a two-layer optimization model, which includes an upper-layer mixed integer programming model and a lower-layer deep learning model; S3: Train the lower-layer deep learning model using historical operating data and update its model parameters online; S4: In each control cycle, the prediction information output by the lower-level model is input into the upper-level model, and the optimal control sequence is solved through the rolling time-domain optimization strategy; S5: Send the optimal control command obtained from the solution to the actuator to control the operating status of the electrolytic cell.

2. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: In step S2, the upper-level mixed integer programming model takes the minimum total operating cost of the system and / or the minimum unit hydrogen production cost as the objective function. Its decision variables include the start-up and shutdown status of the electrolyzer and the target power setting value. The constraints include at least the start-up and shutdown logic constraints of the electrolyzer, the power ramp-up rate constraints, the upper and lower power limits constraints, and the power grid interaction constraints.

3. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: In step S2, the lower-level deep learning model is a time series prediction model, which is configured to receive photovoltaic power prediction sequence, current system status and historical operating data as input, and output the system efficiency prediction value under different electrolyzer operating schemes in the future period.

4. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 3, characterized in that: The underlying deep learning model adopts a Long Short-Term Memory (LSTM) network architecture, which uses its internal memory units and gating mechanisms to capture the nonlinear temporal characteristics of photovoltaic power output and system operation.

5. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: In step S4, the specific implementation of the rolling time-domain optimization strategy is as follows: at each sampling time, solve an optimization sub-problem in a finite time domain, execute only the solution of the first control cycle of the optimization sub-problem, and reacquire the system state at the next sampling time, repeating the above optimization process.

6. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 2, characterized in that: In the upper-level mixed integer programming model, the degradation cost of the electrolyzer is related to its operating efficiency and the number of start-stop cycles, and the efficiency prediction value output by the lower-level model is used to dynamically adjust the weight coefficient of the degradation cost.

7. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: It also includes step S6: during the optimization solution and execution process, multiple safety constraints are embedded, including overvoltage protection, overcurrent protection, overheat protection and grid reverse current protection of the electrolytic cell, and a preset safety shutdown procedure is triggered when a system abnormality is detected.

8. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: It also includes step S7: when communication between the system and the main controller is interrupted or an unpredictable fault occurs, automatically switch to a fault-tolerant control mode based on predefined rules to maintain the basic safe operation of the system.

9. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: The data preprocessing in step S1 includes cleaning and normalizing the collected data, as well as extracting statistical features that reflect the photovoltaic power output characteristics and the health status of the electrolyzer.

10. The distributed photovoltaic water electrolysis hydrogen production control method based on artificial intelligence optimization according to claim 1, characterized in that: This method is applied to a distributed hydrogen production system containing multiple electrolyzers connected in parallel. The upper-level mixed integer programming model can coordinate the scheduling of multiple electrolyzers to optimize the overall system efficiency.

Images