Control method for large-scale water electrolysis hydrogen production cluster for wind-solar coupling

Abstract

The present invention relates to the technical field of hydrogen production by electrolysis of water, and discloses a control method for a large-scale hydrogen production cluster by electrolysis of water for wind-solar coupling, including: predicting the total output power of a photovoltaic power generation system and a wind power generation system; if the actual hydrogen production of an electrolyzer cannot reach the expected hydrogen production within a set time, refitting the parameters through the production data of the electrolyzer, and obtaining multiple electrolyzer power allocation sets through a constrained optimization model; clustering the production data of the electrolyzer to obtain multiple state categories; selecting the best electrolyzer power allocation set through a clustering transfer model, and distributing it to each electrolyzer as the power of each electrolyzer at the next moment. When the present invention distributes power, it fully considers the characteristics of each electrolyzer, dynamically corrects the power online, makes full use of energy, and produces synergistically between electrolyzers, so that the hydrogen production can be maximized under the premise of ensuring the operating rate.

Description

Technical Field

[0001] The present invention relates to the technical field of water electrolysis hydrogen production, and in particular to a control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling. Background Art

[0002] The electrolyzer cluster control cannot adapt to the power fluctuations of pure wind and solar energy, resulting in the hydrogen production not reaching the optimal value. Especially under non-full power, the control strategy of a single electrolyzer becomes particularly important, which is related to the increase in hydrogen production, especially when the hydrogen production, start and stop characteristics of each electrolyzer are not exactly the same. The unreasonable energy allocation in the existing technology leads to insufficient energy utilization. When the status of the electrolyzer changes, the allocation is not adjusted in time, and the information between the electrolyzers is not interoperable. It may be encountered that an electrolyzer may have a decrease in efficiency due to equipment and other reasons. The existing technology usually does not adjust the corresponding part of the power to other electrolyzers, resulting in low hydrogen production.

[0003] Existing technologies generally allow wind and solar power production capacity to cooperate with the power grid to adapt to the full-power operation of the electrolyzer, rather than adjusting the electrolyzer hydrogen production strategy to adapt to the electricity generated by wind and solar power. Each valve needs to be adjusted manually and continuously, and cannot adapt well to the needs of multiple scenarios.

[0004] The existing technology is unreasonable in terms of energy management allocation, and the equipment cluster control strategy cannot adapt to variable power conditions, and cannot increase the overall output of hydrogen. When using artificial neural networks (ANN) to predict the power of wind and solar power generation in the existing technology, wind power generation prediction models and photovoltaic power generation prediction models are generally used, both of which rely on meteorological data (wind speed, temperature, etc.).

[0005] In the prior art, a simple linear constrained optimization solution allocation method is used based on the predicted output power. This allocation method considers fewer dimensions (for example, some only rely on the known rated power of each electrolyzer), resulting in a low startup coverage rate and the hydrogen production is not the highest. It may also lead to the waste of some electricity generated by wind and light.

[0006] In the prior art, when it comes to cluster control of equipment, the PID control mode is generally used, which cannot control the high hydrogen production under different power, different current density, different temperature, etc.; and the existing control strategy only supports a single electrolyzer, and the information between electrolyzers is not communicated; when the efficiency of an electrolyzer decreases, resources will be wasted, and the output power cannot be adjusted to other electrolyzers well. In the case of power fluctuations, the traditional electrical control logic (such as segmentation strategy) cannot adapt to the power fluctuation scenario, and it is difficult to finely control the electrolyzer, resulting in low hydrogen production. Summary of the invention

[0007] To solve the above technical problems, the present invention provides a control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling, and the energy generated by wind and light is all used for hydrogen production, producing more hydrogen for downstream generation and use. The present invention first distributes the predicted output power, gives a relatively coarse granularity, and can allocate different powers to different electrolyzers (for example, the power to a certain electrolyzer is 4000w, and the duration is 0.5 hours), and then controls the allocated multiple electrolyzers. At this time, it is necessary to finely control multiple electrolyzers to achieve global optimization to maximize the amount of hydrogen production. Artificial intelligence process control technology (Al process control, AIPC) is used for the entire process to solve the complex system control problems such as large lag, strong coupling, difference and nonlinearity that traditional electrical control technology cannot adapt to the operating conditions of equipment, and provide a strong optimization, high performance and high flexibility intelligent industrial control overall solution. AlPC mines equipment operating characteristics through equipment condition modeling and real-time operation data analysis, predicts the future process performance of equipment, optimizes control strategy and instruction generation, and replaces the traditional electrical segmented control strategy, thereby achieving better equipment control effect in a complex production environment.

[0008] The entire link process reflects energy allocation (rough allocation) and equipment cluster scheduling control based on the approximate allocation results (fine control, internal electrolyzer group self-adjustment control strategy). A single electrolyzer can adapt to power fluctuation scenarios, resulting in better hydrogen production.

[0009] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0010] A control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling, comprising the following steps:

[0011] Step 1: predict energy power, which includes the total output power of photovoltaic power generation system and wind power generation system; input photovoltaic power generation meteorological data, photovoltaic power generation system measurement data, photovoltaic power generation system basic information, wind power generation meteorological data, and wind power generation system fan data into the pre-trained BP neural network to predict the total output power of photovoltaic power generation system and wind power generation system;

[0012] Step 2: The hydrogen production H of the nth electrolyzer n The power P of the nth electrolytic cell n The relationship between them is:

[0013]

[0014] A n , B n , Cn are all fitting parameters, 1≤n≤N, N is the total number of electrolytic cells;

[0015] If the actual hydrogen production of the electrolyzer cannot reach the expected hydrogen production within the set time, the A n , B n , C n Fitting is performed; the optimization objectives are to maximize hydrogen production, maximize the number of electrolyzer startups, minimize the time required for the electrolyzer to jump directly from the idle state to the production state, and minimize the number of electrolyzer start-up and stop times, and to use power P n Power constraints must be met As the constraint condition, the solution of equation 1 is solved by the constraint optimization model to obtain multiple electrolytic cell power allocation sets Each electrolyzer power allocation set includes a power allocated to each electrolyzer;

[0016] The electrolyzer production data is clustered to obtain M categories, each category contains multiple electrolyzer production data, and the M categories are sorted in order from low to high according to the average value of hydrogen production in the production data of each category, and named as state 1 to state M in sequence;

[0017] The state of the electrolyzer at the current moment is obtained according to the current hydrogen production of the electrolyzer, and the state of the electrolyzer at the next moment is obtained according to the hydrogen production corresponding to the power of each electrolyzer in the electrolyzer power allocation set; based on the principle of maximizing the sum of the state values ​​of each electrolyzer at the next moment and minimizing the state span between the state of the electrolyzer at the current moment and the state of the next moment, the best electrolyzer power allocation set is selected as the power allocation to each electrolyzer at the next moment.

[0018] Furthermore, it also includes the control process of each valve during the production of the electrolytic cell, specifically including:

[0019] For the nth electrolytic cell, 1≤n≤N, the original electrolytic cell production data is enhanced in reverse order to obtain enhanced electrolytic cell production data; based on the enhanced electrolytic cell production data, it is input into the improved preference reinforcement learning model to output the action direction of each valve at the next moment;

[0020] The production data of the electrolyzer at the current moment, the opening of each valve at the current moment, the movement direction of each valve at the next moment, and multiple opening changes ΔK within the allowable opening range of each valve are input into the pre-trained LSTM model, and multiple hydrogen production corresponding to the multiple opening changes ΔK at the next moment of the electrolyzer are output.

[0021] The power P of the electrolytic cell at the next moment n ′ Substitute into formula 1 to obtain the hydrogen production H ′n ;

[0022] will be closest to H ′ n Hydrogen production The corresponding opening change ΔK of each valve at the next moment and the action direction of each valve at the next moment are used as action instructions to control each valve of the nth electrolytic cell.

[0023] Furthermore, the valves include a hydrogen side regulating valve, an oxygen side regulating valve, and a cooling water regulating valve.

[0024] Furthermore, the original electrolytic cell production data is enhanced in reverse order to obtain enhanced electrolytic cell production data, specifically including:

[0025] The original production data combined with the valve opening is {T1, T2, …, T k}, where T1 is the production data generated when running at minimum power, k is the number of original production data, T n The production data of full power operation is reversed and spliced ​​after the original production data to obtain enhanced production data {T1, T2, …, T k-1 ,T k ,T k-1 ,…,T2,T1}.

[0026] Furthermore, the reward of the preference reinforcement learning model for the valve action is set as follows:

[0027] Expert knowledge in valve control includes:

[0028] (1) When the liquid level on the hydrogen side Less than or equal to the oxygen side liquid level When the hydrogen side regulating valve The opening will increase; when the liquid level on the hydrogen side Greater than oxygen side liquid level When the hydrogen side regulating valve The opening is reduced;

[0029] (2) Oxygen side pressure Greater than or equal to the pressure setting value P set , oxygen side regulating valve Increase, oxygen side pressure Less than the pressure setting value P set , oxygen side regulating valve reduce;

[0030] (3) When the oxygen separator inlet temperature The set value is greater than or equal to the temperature set value T set , cooling water regulating valve The opening increases; when the oxygen separator inlet temperature Less than the temperature setting value T set , cooling water regulating valve The opening is reduced;

[0031] If the valve action is consistent with the expert knowledge of valve control, give the valve action a reward R action Set to 1, if the valve action does not conform to the expert knowledge of valve control, the reward R of the valve action action Set to 0;

[0032] The enhanced production data are clustered using the k-means clustering method to obtain multiple categories. The categories are sorted from low to high according to the average value of hydrogen production in the enhanced production data of each category, and each category is taken as a state to form a category state transition diagram. The probability value from state i to state j in the category state transition diagram is

[0033]

[0034] numbers(state(i)→state(j));

[0035] state(i) represents state i, state(j) represents state j, numbers(state(i)→state(j)) represents the number of production data pairs in the enhanced production data that meet the condition that two adjacent enhanced production data are transferred from state i to state j; numbers(state(i)) represents the number of enhanced production data in state i;

[0036] The preferred reinforcement learning model rewards the valve action Among them, w1+w2=1, w1 and w2 are artificially set as positive parameter values.

[0037] Furthermore, the pre-trained LSTM model outputs the opening change of each valve at the next moment and the corresponding hydrogen production of the electrolyzer, including

[0038] The movement directions of the valve include increasing opening, decreasing opening, and unchanged opening. An LSTM model is formed for each movement direction.

[0039] Look back step in LSTM model steps To set up dynamics:

[0040]

[0041] N step1 is the intermediate variable, is a rounding function, abs(·) is an absolute value function, max(·) means taking the maximum value; min(·) means taking the minimum value, which cannot be 0; diff(·) means a first-order difference function; K represents the valve opening, and diff(K) represents the difference between the valve opening at the next moment and the valve opening at the previous moment.

[0042]

[0043] N step2 is an intermediate variable, a represents the step amplitude of the response valve index parameter, b represents the corresponding index parameter; mean represents the mean value;

[0044] Look back at the step length steps for:

[0045] Look steps =min(N step1 ,N step2 );

[0046] Use the LSTM model corresponding to the valve's movement direction to output the valve opening change value and hydrogen production.

[0047] Furthermore, the production data of the electrolyzer include: oxygen side liquid level, hydrogen side liquid level, oxygen separator inlet temperature, hydrogen separator inlet temperature, hydrogen in oxygen, alkali liquid flow, alkali liquid temperature, alkali liquid tank level, oxygen separator pressure, rectifier cabinet current, rectifier cabinet voltage and hydrogen production.

[0048] Furthermore, the PV power generation meteorological data include direct light irradiance, global light irradiance, relative humidity, temperature, wind speed, wind direction and air pressure;

[0049] The measurement data of photovoltaic power generation system include scattered radiation intensity, temperature, wind direction, wind speed, air pressure, photovoltaic output power;

[0050] The basic information of photovoltaic power generation system includes installed capacity, photovoltaic panel size, number of photovoltaic panels, and photovoltaic panel angle;

[0051] Wind power meteorological data include: wind speed, relative humidity, temperature, air pressure, air density;

[0052] Wind turbine data of wind power generation system includes: wind speed, deflection angle and power.

[0053] Furthermore, the constrained optimization model includes a genetic algorithm, a particle swarm algorithm or a Pareto frontier method.

[0054] Compared with the prior art, the beneficial technical effects of the present invention are:

[0055] 1. When distributing power, the characteristics of each electrolyzer are fully considered, and the power is dynamically corrected online to make full use of energy. The electrolyzers can produce in coordination to maximize the hydrogen production while ensuring the operating rate.

[0056] 2. Multiple electrolytic cells can be flexibly controlled, breaking the traditional PID control mode, and realizing stable process control of equipment clusters under variable power based on the combination of reinforcement learning and LSTM. The reward and punishment mechanism in the improved preference reinforcement learning model is determined based on the state diagram transition, breaking the traditional pure single expert experience knowledge drive.

[0057] 3. When selecting the step size of the LSTM model, break the original fixed value and use the step size for dynamic control to perform personalized control on different valves.

[0058] 4. In the equipment cluster control, when the power fluctuates, it can make self-inference decisions, avoiding the traditional coarse segmented electrical control strategy, and can achieve fine control of the electrolyzer, resulting in higher hydrogen production. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of energy dispatch decision making for the present invention;

[0060] Figure 2 It is a schematic diagram of the reverse order enhancement processing of production data according to the present invention;

[0061] Figure 3 A schematic diagram of generating action instructions for the present invention;

[0062] Figure 4 Class state transition diagram generated for the present invention. DETAILED DESCRIPTION

[0063] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0064] 1. Wind and solar power generation power prediction model

[0065] The input data of the output power prediction model of the photovoltaic power generation system includes photovoltaic power generation meteorological data, photovoltaic power generation system measurement data, and photovoltaic power generation system basic information.

[0066] Photovoltaic power generation meteorological data include: direct light irradiance, global light irradiance, relative humidity, temperature, wind speed, wind direction and air pressure.

[0067] The measurement data of photovoltaic power generation system include: diffuse radiation intensity, temperature, wind direction, wind speed, air pressure and photovoltaic output power. The temperature, wind speed, wind direction and air pressure in the photovoltaic power generation system measurement data are different from those in the photovoltaic power generation meteorological data. The temperature, wind speed, wind direction and air pressure in the photovoltaic power generation meteorological data are provided by the meteorological department, while the temperature, wind speed, wind direction and air pressure in the photovoltaic power generation system measurement data are collected locally by photovoltaic equipment.

[0068] The basic information of photovoltaic power generation system includes: installed capacity, size of photovoltaic panels, number of photovoltaic panels, and angle of photovoltaic panels.

[0069] The input data of the output power prediction model of the wind power generation system includes wind power generation meteorological data and wind power generation system fan data.

[0070] Wind power generation meteorological data include: wind speed, relative humidity, temperature, air pressure and air density.

[0071] The wind turbine data of the wind power generation system includes: wind speed, deflection angle and power. The wind speed in the wind turbine data of the wind power generation system is different from the wind speed in the wind power generation meteorological data. The wind speed in the wind power generation meteorological data is provided by the meteorological department, and the wind speed in the wind turbine data of the wind power generation system is obtained by the wind turbine self-test.

[0072] Power prediction is the precursor node of subsequent energy allocation and cluster scheduling, and its accuracy is related to the operation status of the electrolyzer and the amount of hydrogen produced. Traditional wind and solar power generation power prediction generally has two models, one is wind power prediction, and the other is photovoltaic power prediction. There is no intersection between the two models. The power prediction model in the present invention adopts BP neural network, and only one model is used to predict the output power of photovoltaic power generation system and wind power generation system, avoiding the repeated use of meteorological data.

[0073] 2. Energy dispatch decision

[0074] Energy distribution needs to be combined with the conditions of each electrolyzer so as to be reasonably allocated.

[0075] H n =f n (P n ), H n Refers to the hydrogen production of the nth electrolyzer, P n refers to the power of the nth electrolytic cell, f n (·) refers to the function of the relationship between the hydrogen production and power of the nth electrolyzer. Since hydrogen is only produced in the production state, it is required Combined with expert experience and knowledge, hydrogen production and power show a positive correlation. The present invention adopts a quadratic fitting method, which is as follows:

[0076]

[0077] A n , B n , C n are the three parameters to be fitted.

[0078] Here, we can use the least squares method to calculate A based on the historical production data of the electrolytic cell. n , B n , C n Perform the fitting.

[0079] It is also necessary to establish the constraints and optimization objectives of formula 1:

[0080] δ n Indicates whether the electrolyzer is turned on or not.

[0081] It represents the working state of the nth electrolytic cell at time t, which is specifically expressed as:

[0082]

[0083] In the idle state, no load is required and no hydrogen is generated. The entire device is in idle mode. In the cold standby state, a load is required but no hydrogen is generated, which is equivalent to a cold start process. In the production state, the input power range of the electrolyzer is from the minimum power P min Run to full power P max At this time, the electrolyzer needs to be in the power operation range Run inside.

[0084] Note that the power operating ranges of the electrolyzers may be different. The normal hydrogen production power operating range of the nth electrolyzer is At lower current densities, the concentration of hydrogen in oxygen will reach the lower limit of the dangerous limit; while higher current densities than the rated current will damage the electrolyzer stack material. Therefore, these operating limits cannot be violated as a measure to avoid risks. Therefore, a minimum load (such as 10%) is generally required, and the load limit constitutes a hard constraint on the optimization model to avoid risks.

[0085] The time required to go from idle to production is longer than the time required to go from cold standby to production and varies from electrolyzer to electrolyzer.

[0086] It represents the time required for the nth electrolytic cell to change from the cold standby state at the current time t to the production state. As each electrolytic cell is used, this time factor is mainly considered.

[0087] Indicates the time required for the nth electrolyzer to go from idle state to production state, which is equivalent to the cold start time. The time required for each electrolyzer to go from idle state to production state will be updated accordingly as the history accumulates. This variable is a relatively fixed value.

[0088] Following the principle of using as many electrolyzers as possible and avoiding the waste of energy by cold starting electrolyzers that are not powered on, electrolyzers that can produce more hydrogen at a slightly lower power are preferred. The optimization goal is

[0089] maximize;

[0090] ∑δ n maximize;

[0091] minimize;

[0092] ∑SS n minimize;

[0093] SS n Indicates the number of starts and stops of the nth electrolytic cell.

[0094] The constraints are:

[0095] Power Constraints:

[0096] Under this constraint, since f n (P n ) function is a continuous function, and a general constraint optimization model can be used. There are many electrolytic cell power allocation sets that meet multiple objectives. The present invention adopts a cluster transfer model of multi-objective dynamic tasks to select an electrolytic cell power allocation set and perform power allocation on each electrolytic cell.

[0097] For the nth electrolyzer, 1≤n≤N, the overall process of energy dispatch decision includes the following steps:

[0098] S21: Determine whether the current production data of the electrolytic cell meets condition 1; if so, proceed to steps S22, S23 and S24; if not, proceed to step S25.

[0099] S22: Use the least squares method to fit the quadratic function to refit the parameter a of formula 1 n , B n , C n , and obtain multiple electrolyzer power allocation sets through a general constrained optimization model;

[0100] S23: clustering the electrolyzer production data within a set time period from the current moment to the front using a clustering method (such as the k-means method) to obtain M categories, each category containing multiple electrolyzer production data, and sorting the M categories in descending order according to the average value of hydrogen production in the production data of each category, and naming them state 1 to state M in sequence.

[0101] S24: Select an electrolyzer power allocation set from multiple electrolyzer power allocation sets through a clustering transfer model: obtain the current state of the electrolyzer according to the current hydrogen production of the electrolyzer, and obtain the state that the electrolyzer is expected to reach according to the hydrogen production corresponding to the power of each electrolyzer in the electrolyzer power allocation set; based on the principle of maximizing the sum of the states that each electrolyzer is expected to reach and minimizing the state span between the current state of the electrolyzer and the expected state, select the best electrolyzer power allocation set as the power allocation to each electrolyzer at the next moment.

[0102] Each piece of electrolyzer production data refers to the production data of the electrolyzer at a moment. The production data of the electrolyzer at a moment include the oxygen side liquid level, the hydrogen side liquid level, the oxygen separator inlet temperature, the hydrogen separator inlet temperature, the hydrogen in oxygen, the alkali liquid flow, the alkali liquid temperature, the alkali liquid tank level, the oxygen separator pressure, the sorting cabinet current, the rectifier cabinet voltage and the hydrogen production.

[0103] S25: Determine whether condition 2 is met. If so, it is necessary to shut down some electrolyzers. The principle adopted is to select electrolyzer equipment with a smaller hydrogen production under full power; if not, proceed to step S26;

[0104] S26: A plurality of electrolyzer power allocation sets are obtained through a general constrained optimization model; the current state of the electrolyzer is then determined through production data and a pre-trained clustering transfer model; and then the expected state of the electrolyzer is obtained based on the hydrogen production corresponding to the power of each electrolyzer in the electrolyzer power allocation set. The optimal electrolyzer power allocation set is selected based on the principle of minimizing the state span between the current state of the electrolyzer and the expected state.

[0105] Condition 1: The electrolyzer fails to reach the expected target value within the set time. For example, if the P value is not reached after running at the specified power for 30 minutes, n The corresponding hydrogen production H n .

[0106] Extreme weather may cause large fluctuations between the predicted energy power in the future (such as 30 minutes later) and the current power value. Considering that the drying tower of the hydrogen purification device is a seasoning type, the regeneration process cannot be shut down. If wind and light stop generating electricity, energy is needed to operate some electrolyzers to maintain purification until the desiccant regeneration is completed and then shut down.

[0107] Therefore, condition 2 is: the predicted energy power cannot meet the minimum power value of the current operation of multiple electrolyzers.

[0108] Common constrained optimization models include genetic algorithms, particle swarm algorithms, Pareto frontier methods, etc., which can obtain multiple solutions.

[0109] 3. Equipment cluster process control model

[0110] By connecting the reinforcement learning model in series with LSTM, the valve opening can be flexibly adjusted. The traditional control strategy of running towards a stable optimal state has been broken. Now it can run from the optimal to the suboptimal state, avoiding the problem of low hydrogen production caused by the unrefined operation of the segmented strategy in the traditional electrical program. When the operating power of each device is given, the operating strategy for a single device will change. For example, when the power changes from 80% to 70%, how to control the electrolyzer becomes a key issue. Because the working conditions of each device are different, each device has its own control strategy.

[0111] 3.1 Data Preprocessing

[0112] Generally, the data generated by the electrolytic cell is generated from low power to high power, or even to full power. There is a lack of data on the transition from high power to low power. The present invention enhances the production data of the electrolytic cell:

[0113] The low-power data to full-power production data are processed in reverse order, and the generated data is spliced ​​to the full-power production data to achieve full power coverage from low power to full power and then to low power.

[0114] like:

[0115] The original production data that incorporates the valve opening is in the form of {T1, T2, …, T k}, where T1 is the production data generated when running at minimum power, T n The production data of full power operation is reversed and spliced ​​after the original production data to obtain enhanced production data {T1, T2, …, T k-1 ,T k ,T k-1 ,…,T2,T1}.

[0116] 3.2 Model Training

[0117] Reinforcement learning can solve the problem of sequential decision-making in an uncertain environment and give the direction of each valve movement. The valve movement direction includes increasing, decreasing, and unchanged opening.

[0118] The training data of the LSTM model includes the production data of the electrolyzer at the current moment, the opening of each valve at the current moment, the movement direction of each valve at the next moment, and multiple opening changes ΔK within the allowable opening range of each valve. The model outputs multiple hydrogen production amounts of the electrolyzer at the next moment corresponding to the multiple opening changes ΔK.

[0119] For example, for the pressure regulating valve (oxygen side regulating valve), try to give the hydrogen production within the allowable opening range of the valve. The specific input data are: the production data at the current moment (including oxygen side liquid level, hydrogen side liquid level, oxygen separator inlet temperature, hydrogen separator inlet temperature, hydrogen in oxygen, alkali liquid flow, alkali liquid temperature, alkali liquid tank level, oxygen separator pressure, sorting cabinet current, rectifier cabinet voltage and other data), the opening of the pressure regulating valve at the current moment, the action direction of each valve at the next moment (obtained through the improved preference reinforcement learning model), and multiple opening change values ​​within the allowable opening range of the valve (one opening change value is input each time); the output is the hydrogen production under this opening change value.

[0120] 3.3 Control instruction selection model process

[0121] The instructions for coordinated control of each valve are generated by the improved preference reinforcement learning model to generate the valve action direction, and then the LSTM model generates the valve opening change value, and the final set is combined with H n Function comparison, giving control instructions. The lookback step of the LSTM model is dynamic, and the characteristics of each valve can be extracted from the data.

[0122] 3.4 Preference Reinforcement Learning Model

[0123] The reward and punishment values ​​used in the preference reinforcement learning model are obtained based on the probability of state diagram transition and expert knowledge of basic valve control.

[0124] Expert knowledge of basic valve control, as follows:

[0125] (1) When the liquid level on the hydrogen side Less than or equal to the oxygen side liquid level When the hydrogen side regulating valve The opening will increase; when the liquid level on the hydrogen side Greater than oxygen side liquid level When the hydrogen side regulating valve The opening is reduced;

[0126] (2) When the pressure regulating valve is opened too small, the flow of the medium inside the pipeline is greatly hindered, and the internal pressure drops slowly. To make the pressure drop faster, the pressure regulating valve is opened too large. At this time, the flow speed of the medium inside the pipeline is accelerated, and the pressure drops quickly. Greater than or equal to the pressure setting value P set , oxygen side regulating valve Increase, oxygen side pressure Less than the pressure setting value P set , oxygen side regulating valve reduce;

[0127] (3) When the oxygen separator inlet temperature The set value is greater than or equal to T set , cooling water regulating valve The opening increases; when the oxygen separator inlet temperature Less than the set value T set , cooling water regulating valve The opening is reduced.

[0128] The above expert knowledge is only an example. If it meets this expert knowledge, the valve action will be rewarded R. action Set to 1, if it does not meet the expert knowledge, R action Set to 0.

[0129] Secondly, we need to combine the probability value Pr of the state diagram transition value :

[0130] Traditional clustering methods (such as k-means) are used to cluster multiple enhanced production data including hydrogen production. Considering the actual operation stages of the equipment such as cold start, stable operation, and shutdown, the number of categories is set to be greater than 3*2=6 as much as possible, such as 8. Then, the mean values ​​of hydrogen production in multiple production data of each category are arranged in ascending order to form a cluster as shown in the following figure. Figure 4 The category state transition diagram shown.

[0131] Probability value

[0132] The entire reward mechanism is set up as Among them, w1+w2=1, w1 and w2 are artificially set as positive parameter values.

[0133] For production data used for training, we can know the state of production data and the state span between multiple states. For new production data, we can know the state of new production data through the clustering transition model of historical enhanced data. The probability value is still the transition probability value under historical enhanced data. Therefore, in the actual production process, we can know the probability value.

[0134] After the neural network is trained through reinforcement learning, the set of action directions of each valve is output.

[0135] 3.5LSTM Model

[0136] The movement directions of the valve include increasing the opening, decreasing the opening and keeping the opening unchanged. The present invention forms three LSTM models for the three movement directions.

[0137] The look-back step size in the LSTM model needs to be dynamically given according to the response time of different valve controls. For example, the oxygen-side regulating valve looks back at the historical production data of the previous two moments, and the hydrogen-side regulating valve looks back at the historical production data of the previous five moments.

[0138] The step length for reviewing historical production data for each valve is as follows:

[0139]

[0140] is an upward rounding function, abs(·) is an absolute value function, max(·) means taking the maximum value; min(·) means taking the minimum value, which cannot be 0; diff(·) means a first-order difference function, which can be understood as the value at the next moment minus the value at the previous moment; K represents the valve opening, and diff(K) represents the difference between the valve opening at the next moment and the valve opening at the previous moment. For example, the valve opening data at each moment can be collected, such as the valve opening K of the oxygen side is 0-100 degrees, and the valve opening change value at adjacent moments, max(abs(diff(K)))=20, min(abs(diff(K)))=5, then N step1 =4.

[0141]

[0142] N step2 is an intermediate variable, a represents the step amplitude of the corresponding valve index parameter, b represents the corresponding index parameter; mean represents the mean value.

[0143] Example: If the indicator parameter corresponding to the pressure regulating valve (oxygen side regulating valve) is pressure, the pressure value is displayed, the step amplitude of increasing the pressure (boosting) is b = 0.0002Mpa, and mean (diff (b)) = 0.0002586 is calculated, then N step2 =2.

[0144] Finalized Lookback Step Length steps for:

[0145] Look steps =min(N step1 ,N step2 );

[0146] If the neural network trained by reinforcement learning gives the direction of valve movement as increasing opening, then the LSTM model with increasing opening is used to convert Look steps As the lookback step of the LSTM model, the hydrogen production corresponding to multiple opening change values ​​is output.

[0147] 3.6 and Hn Function comparison

[0148] The multiple opening change values ​​and hydrogen production obtained by the LSTM model are compared with the hydrogen production H calculated by formula 1. n By comparison, the hydrogen production is closest to H n Finally, the valve movement direction and opening change value are sent to the device for regulation.

[0149] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0150] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling, characterized in that: The following steps are involved: Step 1: predict energy power, which includes the total output power of photovoltaic power generation system and wind power generation system; input photovoltaic power generation meteorological data, photovoltaic power generation system measurement data, photovoltaic power generation system basic information, wind power generation meteorological data, and wind power generation system fan data into the pre-trained BP neural network to predict the total output power of photovoltaic power generation system and wind power generation system; Step 2: The hydrogen production H of the nth electrolyzer n The power P of the nth electrolytic cell n The relationship between them is: Formula 1: A n , B n , C n are all fitting parameters, 1≤n≤N, N is the total number of electrolytic cells; If the actual hydrogen production of the electrolyzer cannot reach the expected hydrogen production within the set time, the A n , B n , C n Fitting is performed; the optimization objectives are to maximize hydrogen production, maximize the number of electrolyzer startups, minimize the time required for the electrolyzer to jump directly from the idle state to the production state, and minimize the number of electrolyzer start-up and stop times, and to use power P n Power constraints must be met As the constraint condition, the solution of equation 1 is solved by the constraint optimization model to obtain multiple electrolytic cell power allocation sets Each electrolyzer power allocation set includes a power allocated to each electrolyzer; The electrolyzer production data is clustered to obtain M categories, each category contains multiple electrolyzer production data, and the M categories are sorted in order from low to high according to the average value of hydrogen production in the production data of each category, and named as state 1 to state M in sequence; The current state of the electrolyzer is obtained according to the current hydrogen production of the electrolyzer, and the state of the electrolyzer at the next moment is obtained according to the hydrogen production corresponding to the power of each electrolyzer in the electrolyzer power allocation set; Based on the principle that the sum of the state values ​​of each electrolytic cell at the next moment is the largest and the state span between the state of the electrolytic cell at the current moment and the state at the next moment is the smallest, the best electrolytic cell power allocation set is selected as the power allocation to each electrolytic cell at the next moment; It also includes the control process of each valve during electrolytic cell production, including: For the nth electrolytic cell, 1≤n≤N, the original electrolytic cell production data is enhanced in reverse order to obtain enhanced electrolytic cell production data; based on the enhanced electrolytic cell production data, it is input into the improved preference reinforcement learning model to output the action direction of each valve at the next moment; The production data of the electrolyzer at the current moment, the opening of each valve at the current moment, the movement direction of each valve at the next moment, and multiple opening changes ΔK within the allowable opening range of each valve are input into the pre-trained LSTM model, and multiple hydrogen production corresponding to the multiple opening changes ΔK at the next moment of the electrolyzer are output. The power P of the electrolytic cell at the next moment n ′ Substitute into formula 1 to obtain the hydrogen production H ′ n ; will be closest to H ′ n Hydrogen production The corresponding opening change ΔK of each valve at the next moment and the action direction of each valve at the next moment are used as action instructions to control each valve of the nth electrolytic cell.

2. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized in that: The valves include a hydrogen side regulating valve, an oxygen side regulating valve, and a cooling water regulating valve.

3. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized in that: The reverse order enhancement of the original electrolytic cell production data to obtain enhanced electrolytic cell production data specifically includes: The original production data combined with the valve opening is recorded as {T1, T2, …, T k }, where T1 is the production data generated when running at minimum power, k is the number of original production data, T n The production data of full power operation is reversed and spliced ​​after the original production data to obtain enhanced production data {T1, T2, …, T k-1 ,T k ,T k-1 ,…,T2,T1}.

4. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized in that: The reward of the preference reinforcement learning model for valve action is set as follows: Expert knowledge in valve control includes: (1) When the liquid level on the hydrogen side Less than or equal to the oxygen side liquid level When the hydrogen side regulating valve The opening will increase; when the hydrogen side liquid level Greater than oxygen side liquid level When the hydrogen side regulating valve The opening is reduced; (2) Oxygen side pressure Greater than or equal to the pressure setting value P set , oxygen side regulating valve Increase, oxygen side pressure Less than the pressure setting value P set , oxygen side regulating valve reduce; (3) When the oxygen separator inlet temperature The set value is greater than or equal to the temperature set value T set , cooling water regulating valve The opening increases; when the oxygen separator inlet temperature Less than the temperature setting value T set , cooling water regulating valve The opening is reduced; If the valve action is consistent with the expert knowledge of valve control, give the valve action a reward R action Set to 1, if the valve action does not conform to the expert knowledge of valve control, the reward R of the valve action action Set to 0; The enhanced production data are clustered using the k-means clustering method to obtain multiple categories. The categories are sorted from low to high according to the average value of hydrogen production in the enhanced production data of each category, and each category is taken as a state to form a category state transition diagram. The probability value from state i to state j in the category state transition diagram is numbers(state(i)→state(j)); state(i) represents state i, state(j) represents state j, numbers(state(i)→state(j)) represents the number of production data pairs in the enhanced production data that meet the condition that two adjacent enhanced production data are transferred from state i to state j; numbers(state(i)) represents the number of enhanced production data in state i; The preferred reinforcement learning model rewards the valve action Among them, w1+w2=1, w1 and w2 are artificially set as positive parameter values.

5. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized in that: The pre-trained LSTM model outputs the opening change of each valve at the next moment and the corresponding hydrogen production of the electrolyzer, including The movement directions of the valve include increasing opening, decreasing opening, and unchanged opening. An LSTM model is formed for each movement direction. Look back step in LSTM model steps To set up dynamics: N step1 is the intermediate variable, is an upward rounding function, abs(·) is an absolute value function, max(·) means taking the maximum value; min(·) means taking the minimum value, which cannot be 0; diff(·) means a first-order difference function; K means valve opening, and diff(K) means the difference between the valve opening at the next moment and the valve opening at the previous moment; N step2 is an intermediate variable, a represents the step amplitude of the response valve index parameter, b represents the corresponding index parameter; mean represents the mean value; Look back at the step length steps for: Look steps =min(N step1 ,N step2 ); Use the LSTM model corresponding to the valve's movement direction to output the valve opening change value and hydrogen production.

6. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized in that: The production data of the electrolyzer include: oxygen side liquid level, hydrogen side liquid level, oxygen separator inlet temperature, hydrogen separator inlet temperature, hydrogen in oxygen, alkali liquid flow, alkali liquid temperature, alkali liquid tank level, oxygen separator pressure, rectifier cabinet current, rectifier cabinet voltage and hydrogen production.

7. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized by: Photovoltaic power generation meteorological data include direct light irradiance, global light irradiance, relative humidity, temperature, wind speed, wind direction and air pressure; The measurement data of photovoltaic power generation system include scattered radiation intensity, temperature, wind direction, wind speed, air pressure, photovoltaic output power; The basic information of photovoltaic power generation system includes installed capacity, photovoltaic panel size, number of photovoltaic panels, and photovoltaic panel angle; Wind power meteorological data include: wind speed, relative humidity, temperature, air pressure, air density; Wind turbine data of wind power generation system includes: wind speed, deflection angle and power.

8. The control method for a large-scale water electrolysis hydrogen production cluster for wind-solar coupling according to claim 1 is characterized by: The constrained optimization model includes a genetic algorithm, a particle swarm algorithm or a Pareto frontier method.

Images