Research on typical operating conditions of hydrogen production system with off-grid wind power considering the characteristics of proton exchange membrane electrolysis cell

0 Introduction

The increasing proportion of the installed new energy capacity and the reverse distribution characteristics of resources and loads in China has resulted in serious abandonment of wind and sunlight [1].The use of wind power generation,photovoltaic (PV) power generation,and renewable energy-based electrolytic hydrogen production technology can not only realize the production of“green hydrogen”,but also fully absorb renewable energy,thus“green hydrogen”production technology will become a key research field and investment direction of hydrogen energy development in the future [2,3].Currently,common technologies for water electrolysis hydrogen production include alkaline electrolysis water hydrogen production,proton exchange membrane electrolysis,and solid oxide electrolysis water hydrogen production [4-6].Among them,proton exchange membrane electrolysis water hydrogen production technology has high electrolytic efficiency,high purity,and fast cell response,making it suitable for the application in electrolytic hydrogen production [7-9].

Amidst recent advancements in PEM electrolysis technology for hydrogen production,the operational capacity of PEM electrolysis cells has progressively shifted from kilowatts to megawatts.The cost of hydrogen production via PEM electrolysis has declined annually.Notably,major universities and enterprises are strategically investing in research on PEM electrolysis for hydrogen production.At present,the pioneering ‘green hydrogen’ demonstration project in China is successfully underway on Dachen Island,Zhejiang,representing a significant advancement in the practical implementation of green hydrogen technology.Clarifying the fluctuating operating conditions of PEM electrolysis cells is important for the further advancement of PEM electrolysis technology and stable operation of novel power systems.Although research on the operational conditions of PEM electrolysis cells is underway,a mature analytical approach for these conditions is yet to be established.Many scholars have comprehensively studied typical operating scenarios for battery energy storage systems (BESSs) in new energy applications.In [10],the characteristics and temporal variations in the operational conditions of energy storage batteries were thoroughly considered.By employing clustering analysis,distinctive curves that depict the operational conditions of energy storage batteries have been extracted.In [11,12],the operational condition curves of energy storage systems under different modes were obtained by combining factor and clustering analyses.However,compared with battery energy storage systems,PEM electrolysis cells demonstrate faster response times to fluctuations and broader power operating ranges.The operational conditions and characteristics during actual operation differ significantly from those of BESSs.Consequently,it is essential to investigate operational scenario generation methods for wind-driven PEM electrolysis systems considering the unique characteristics of PEM electrolysis cells.

In typical wind power output scenario research,there are currently two predominant avenues of study:scenario generation and scenario reduction.Scenario generation methods are commonly employed to address problems related to insufficient raw data and scenarios characterized by strong randomness,with a focus on predicting scenarios.Drawing upon the abundant original wind power output data,this study explores the fluctuating operational conditions of PEM electrolysis cells,thus adopting a scenario reduction approach.In the research,clustering analysis is commonly adopted to reduce scenarios.Reference [13] proposed a typical output scenario extraction method based on deep embedding clustering and the traditional K-means approach.Reference [14] employed an improved fuzzy C-means clustering algorithm to aggregate the annual power-response curve of energy storage,thereby generating typical power-response curves.Both the above methods require predefining the number of clusters,are sensitive to noise,and lack robustness [15].In [16],a Gaussian mixture model based on the Bayesian information criterion was proposed.The original output curve of offshore wind power was classified,and the characteristic curve was extracted.However,this clustering method assumes that sample data follow a normal distribution,which may be unsuitable for certain complex wind power output scenarios [17].Reference [18] proposed a scene reduction method based on agglomerative hierarchical clustering,characterizing the wind-solar power characteristics of clustered power stations.However,the method has relatively high computational complexity and may not be sufficiently efficient for handling large-scale data [19].With the ongoing advancements in neural network technology,neural network clustering methods have garnered the attention of numerous scholars.Among these,the self-organizing map (SOM) neural network clustering method stands out for its ability to capture complex nonlinear relationships in wind power output scenarios through nonlinear mapping.It facilitates adaptive learning for dynamically changing data,efficiently processes large-scale datasets,and exhibits considerable robustness,making it well-suited for addressing the fluctuations and noise inherent in wind power output data.

In summary,based on the operational characteristics of PEM electrolysis cells,this study utilized the SOM neural network clustering method to cluster wind power output scenarios.The objective is to obtain the typical operating conditions for off-grid wind power PEM hydrogen production.The remainder of this paper is organized as follows.The influence of the fluctuating input on the PEM electrolysis cell and its operating characteristics are discussed in detail in Sections 1 and 2.Sections 3 and 4 introduce the typical working condition acquisition method.Section 5 verifies the feasibility of this method by using an example simulation.Finally,the conclusion is summarized in Section 6.

1 Operation characteristics of PEM electrolytic hydrogen production device

The operating characteristics of PEM electrolytic hydrogen production systems can be summarized as follows [20-22]:

1) Underload and overload characteristics:Presently,the minimum operating power limit of PEM electrolysis cells is 5% of their rated power.In other words,hydrogen permeation intensifies when the input power of the electrolysis cells is less than 5%,and an increase in the hydrogen concentration in oxygen can easily cause explosions.Additionally,the maximum power limit at which the PEM electrolysis cells can operate for a short time is 150% of their rated power.Therefore,5%～30% of the rated power of PEM electrolysis cells is generally considered as the underload operating power range,and 100%～150% of the rated power is considered as the overload operating power range.

2) Variable load characteristics:Compared with alkaline electrolysis water hydrogen production technology,proton exchange membrane electrolysis water hydrogen production technology has a wider power application range,varying from 5% to 150% of the rated power.Moreover,PEM electrolytic hydrogen production technology responds quickly to changes in input power,and the climb rate can reach 10%–15% of the rated value per second,which is more suitable for the production of hydrogen based on fluctuating power sources.

3) Start-stop characteristic:Generally,the suitable temperature for PEM electrolysis cell operation is approximately 60±5 oC,and the starting process of PEM electrolysis cell can be divided into two methods:cold start and hot start.When the electrolysis cell is initially started,the entire device needs to be heated to fulfill the requirements of the normal operating temperature,which is called the cold-start process.Compared to alkaline electrolysis water hydrogen production technology,the PEM electrolysis cell has a faster start-up speed.The cold start process can be completed in a few minutes,whereas the hot start-up speed can be completed in milliseconds.

2 Influence of power fluctuation on PEM electrolytic hydrogen production system

Wind power output exhibits randomness,volatility,and intermittency.When an electrolytic hydrogen production system operates under off-grid conditions,the electrolysis cell input retains most of its original wind power output characteristics.The factors that influence the fluctuating power of electrolytic hydrogen production systems primarily include sustained overload,sustained underload,sudden increase,sudden decrease,and frequent start-stop.

The composition of the PEM electrolysis cell is shown in Fig.1.The membrane electrode assembly (MEA) is the core component of the electrolysis cell and is the main site for electrochemical reactions.Therefore,studying the impact of power fluctuations on the electrolytic hydrogen production device needs to be considered from the perspective of the membrane electrode assembly.

Fig.1 PEM electrolysis cell structure

The MEA is primarily composed of a diffusion layer,catalyst layer,and proton exchange membrane.Power fluctuations can cause chemical effects such as catalyst particle shedding,accelerated degradation of proton exchange membranes,increased corrosion of the diffusion layer,and mechanical effects such as local catalyst overheating,membrane electrode distortion,and membrane damage [23-25].The effect of power fluctuations on the electrolytic hydrogen production device is depicted in Fig.2.Different types of power fluctuations have different mechanisms of impact on electrolysis cells;therefore,it is necessary to select different indicators to characterize wind power output.

Fig.2 Effect of power fluctuation on the PEM electrolysis cell

Fig.3 Structure of the SOM neural network

3 Selection of indicators for characterizing wind power output characteristics

Based on the operational characteristics of the PEM electrolysis cell,wind power output characteristics can be divided into three categories:underload and overload characteristics,fluctuation characteristics,and start-stop characteristics.Appropriate characteristic indicators were selected for each type to characterize the fluctuation of wind power output.

1) Underload and overload characteristics:The sustained overload and sustained underload in the wind power output curve can be considered the underload and overload characteristics of the wind power output,which characterize the data distribution and trend changes of the wind power daily output curve.Based on the underload and overload characteristics of the PEM electrolysis cell during operation,the proportions of each sample within the underload operating power range (rated power 5%～30%) and overload operating power range (rated power 100%～150%) were collected and used as characteristic indicators to describe the wind power output distribution.

2) Fluctuation characteristics:The sudden increases and decreases in the wind power output curve can be considered as the fluctuation characteristics of the wind power output,which characterize the short-term and high-frequency fluctuations of the wind power output.Accordingly,the first-order difference between adjacent sampling points in the wind power output data was calculated,and the mean and maximum values of the first-order difference were used as indicators to characterize the fluctuations in wind power output.

3) Start-stop characteristics:When the wind power output is below or above the safe operating limit of the electrolysis cell,it causes the electrolytic hydrogen production system to shut down.This start-stop situation caused by power fluctuations in the electrolysis cell can be considered the start-stop characteristics of the wind power output,which characterize the start-stop situation of the PEM electrolysis cell within a sample cycle.The number of start-stop times of the electrolysis cell in unit time was counted and used as a characteristic indicator to characterize the impact of wind power output on the start-stop of the electrolysis cell.

4 Generation of typical operating conditions for PEM electrolysis cells

Scenario analysis is a common method for studying uncertainty in renewable energy outputs.In this study,the SOM neural network clustering algorithm was used to cluster the characteristics of the annual wind power output sample set and extract typical daily output scenarios.Compared with other clustering methods,the main feature of the SOM neural network algorithm is the gradual optimization of the network structure by utilizing the mutual competition between neurons and mapping high-dimensional feature vectors in space in a low-dimensional and discrete form,thereby achieving the goal of dimensionality reduction.Therefore,the method is suitable for clustering high-dimensional feature vectors.Additionally,during the calculation process,the method updates the neighboring nodes around the winning node,and the clustering results are less affected by initialization and incorrect data.It can self-organize and adaptively complete clustering based on sample characteristics.

An SOM comprises two layers of neurons:an input layer and a competing layer (output layer).Based on the different characteristic similarities,the intra-day wind power output scenario set can be divided into several categories in the output layer.

The clustering process is as follows:

1) Initialize the network structure

Input layer:The number of neurons n in the input layer is determined by the number of characteristic indicators that describe the wind power output.

Competition layer:Input-layer neurons are mapped to output-layer neurons using n-dimensional connections.

2) Initialize the network parameters

Initialize the neighborhood radius:The initial neighborhood value is σ0;

Initialize the number of iterations:The initial number of iterations is N;

Initialize the learning rate:The initial learning rate is α0;

Initialize the neuron weight:The initial weight is .

3) Find the winning neuron

Traverse all nodes in the competition layer,select the node with the smallest distance between the input and weight vectors as the winning neuron,and use the Euclidean distance as the similarity discrimination function.

represents an n-dimensional sample vector in the input layer,and represents an n-dimensional weight vector that corresponds to a node in the output layer.

1) Neighborhood update amplitude calculation

The nodes contained in the superior neighborhood are determined based on the neighborhood radius,and the respective update amplitude is calculated using the neighborhood function.

where here,σ signifies the neighborhood radius in t iteration, h represents the neighborhood function,and S indicates the distance between the winning node and the other nodes in the neighborhood.

2) Neuron weight adjustment

According to the gradient descent method,the weights of the nodes in the superior neighborhood are updated as follows:

where W j(t)represents the weight at t iteration,η indicates the learning rate at t iteration,and X i (t )-W j(t)indicates the difference between the sample vector at t iteration and the weight vector that corresponds to a point in a superior neighborhood.

3) Iterative computations

The iteration ends after N times.

Several sets of daily wind power output scenarios representing different characteristics were obtained using the SOM neural network clustering and characteristic combinations.The actual wind power output curve can be regarded as a combination of trend and fluctuation components,where the trend components reflect the data distribution of the wind power output and exhibit strong temporal characteristics.The trend component with the highest proportion of annual occurrences was selected as the basis,and weights were distributed to the fluctuation components in each typical output scenario based on statistical probability to generate fluctuation components with annual fluctuation characteristics.The typical output curve of a wind farm can be obtained through superposition.The typical operating conditions of an electrolysis cell can be obtained by applying superposition to an off-grid electrolytic hydrogen production system.

The typical working condition generation process for an off-grid electrolytic hydrogen production system is shown in Fig.4.

Fig.4 Flow chart of typical working condition generation process of the off-grid electrolytic hydrogen production system

5 Example analysis

5.1 Capacity configuration of electrolytic hydrogen production system

In this study,an analysis is conducted using historical output data from a wind farm situated in a wind-rich region of North China.The wind farm has a rated installed capacity of 250.5 MW,with adjacent sampling points spaced at 1-minute intervals.Each output sample comprises 1440 data points for a total of 365 samples.Based on the load variation characteristics of the PWM electrolysis cells,the load variation range was set to 5%～150% of the rated power.Wind abandonment occurs during actual application.Based on the data provided by China’s New Energy Consumption Monitoring and Early Warning Center,the wind power utilization rate in North China in 2022 was 95.8%.Therefore,considering the influence of the wind abandonment factors,the rated power Pel required for the configuration of the electrolytic hydrogen production system can be obtained using equation (5).

where Pmax indicates the installed capacity of the wind farm,UW indicates the utilization rate of wind energy,and αmax indicates the maximum overload rate.

The capacity of a single electrolysis cell is limited.To fully absorb the wind power,the off-grid electrolytic hydrogen production system needs multiple electrolysis cells to run in parallel to fit the high wind power output [26].Presently,relevant applications of PWM electrolysis cells with a capacity of 10 MW or above exist in some countries.Based on the value of Pel,the electrolytic hydrogen production system is set to be composed of 16 PEM electrolysis cells with the same capacity.The rated power of a single cell is 10 MW,and the power variation range is 0.5～15 MW.

5.2 Filtering and denoising of raw output data

The original wind power output data contained numerous high-frequency random outputs caused by atmospheric turbulence.These high-frequency components do not significantly improve hydrogen production,but seriously affect the operating life of the electrolysis cell;therefore,it is necessary to filter the original output data.In practical applications,this process can be realized using energy storage devices,such as lithium battery arrays or supercapacitors.

Given the irregularity and time-varying nature of the original wind power output data,the Mallat wavelet transform algorithm can be used to filter the raw output data.Various wavelet basis functions possess diverse time-frequency characteristics,making careful selection of an appropriate wavelet basis function essential for achieving optimal denoising results in the wavelet denoising process.The wind power output sequence demonstrated both randomness and volatility.The db9 wavelet,which is characterized as a compactly supported and orthogonal wavelet basis function,exhibits excellent localization in both the time and frequency domains.Consequently,it is well-suited for capturing swift changes in wind power output sequences.Therefore,the db9 wavelet basis function was selected to decompose the original wind-power output sequence.Moreover,‘rigrsure’ is an adaptive soft thresholding method that adeptly balances the preservation of signals with the suppression of noise during the filtering process.It exhibits flexibility by autonomously adjusting the threshold parameters based on the local characteristics of the signal and level of noise.Given the intricate time-frequency variations in wind power output sequences,which may involve different noise levels in distinct time segments,the ‘rigrsure’ threshold is selected for the filtering process.The discrete wavelet decomposition process is illustrated in Fig.5.

Fig.5 Schematic of discrete wavelet decomposition

The original signal is decomposed into a low-frequency component A,which approximates the original signal,and a high-frequency component D,which represents a random disturbance,by the action of a low-pass filter and high-pass filter.After filtering the high-frequency component,the filtered wind power output signal was obtained by wavelet reconstruction.A comparison of the results before and after filtering is shown in Fig.6.The denoised signal retains key characteristics while filtering out high-frequency random outputs,thereby reducing the risk of electrolysis cell damage.

Fig.6 Comparison of wind power output curve before and after filtering

5.3 Output data clustering after filtering

Characteristic extraction was performed on the daily output samples using the three types of indicators proposed above.Each characteristic has a different number of indicators.To avoid the characteristics with a small number of indicators being overlooked by those with a large number,the three types of characteristics were clustered,and the clustering results were combined to obtain a typical output scenario set.

This study uses the Davies-Bouldin index (DBI) in conjunction with the actual requirements to determine the optimal number of clusters.The DBI represents the clustering effect by calculating the intra-and inter-cluster distances between any two clusters and taking the maximum value,which is expressed as

where k indicates the number of clusters, denotes the average distance between each sample in the cluster and the center of the cluster,and dij indicates the distance between the centers of the two clusters.The smaller the DBI,the smaller the intra cluster distance,and the larger the inter cluster distance.Thus,the clustering accuracy increases with a decrease in the DBI.

The DBI of each output characteristic under different cluster numbers can be calculated using equation (6),and the results are shown in Fig.7.

Fig.7 Change of DB index with cluster number

Based on the above calculation results,the DBI generally exhibits a decreasing trend with an increase in k value,and the rate of decrease is different.Combined with the actual application requirements,an excessive number of clusters should be avoided to simplify the scenario analysis process.Therefore,the k value when the slope of DBI tends to slow down significantly is typically selected as the number of clusters.

The corresponding clustering numbers of the underload and overload characteristics,fluctuation characteristics,and start-stop characteristics are 5,3,and 4 respectively.The statistics for each characteristic scenario set were calculated,and the results are listed in Tables 1-3.

Table 1 presents the probability of different types of underload and overload characteristics that occur throughout the year as well as the proportion of underload and overload within a day.It can be observed that from the underload and overload characteristics I to V,the proportion of the underload part decreases gradually,while the proportion of the overload part increases gradually.Generally,the overload time in a day is less than the underload time;therefore,the underload has a more significant impact on the electrolysis cell in practical applications.

Table 1 Statistical results of underload and overload characteristics

Table 2 presents the probabilities of different types of fluctuation characteristics that appear throughout the year and the fluctuations between adjacent sampling points.The absolute value of the first-order difference between adjacent sampling points increases progressively from fluctuation characteristic I to III,and there is no significant difference between the rising (positive difference) and falling (negative difference) amplitudes in the same category.

Table 2 Statistical results of fluctuation characteristics

Table 3 presents the probability and frequency of different categories of start-stop characteristics throughout the year.It can be observed that the starting and stopping times gradually increase from start-stop characteristic I to IV.According to the statistical results presented in Table 1,the main reason for the start-stop action of the electrolysis cell is the input power being lower than the minimum safe operating limit.

Table 3 Statistical results of the start-stop characteristics

5.4 Typical operating conditions of electrolytic hydrogen production system

Various characteristic output scenarios can be obtained by combining the clustering results according to the characteristics.Some scenarios contain a small number of samples,indicating that these scenarios have a low probability of appearing throughout the year and are not suitable for analysis as typical output scenarios.Therefore,90% sample coverage can be set to ensure that a typical output scenario contains more than 90% of the sample characteristics of the year.Based on the calculations,the characteristic scenarios with a sample size of no less than five accounts for 92% of the total annual samples.Therefore,characteristic scenarios with sample numbers of no less than five were considered as typical output scenarios.Additionally,the clustering center of these scenarios was selected as the typical input power curve of the PEM electrolytic hydrogen production system,as shown in Fig.8.The naming method for various curves is sequentially arranged based on the corresponding category numbers of underload and overload characteristics,fluctuation characteristics,and start-stop characteristics.

Fig.8 Typical input power curve of PEM electrolytic hydrogen production system

The trend and fluctuation components can be obtained based on the statistical probability and they are superimposed to obtain a typical output curve.

The operating conditions of the electrolysis cell in offgrid operation mode are closely related to the wind power output.When multiple cells with equal capacities are operated in parallel,the input power is evenly distributed in each cell.Based on the operating characteristics of the PEM electrolysis cells,typical working conditions that correspond to the typical output curve can be generated,and the results are depicted in Fig.9.

Fig.9 Typical working conditions of PEM electrolysis cell

The working conditions that cause performance degradation of the electrolysis cell are marked in the figure,including underload,overload,sudden increase,sudden decrease,and frequent start-stop cycles.Based on the load changing characteristics of PEM electrolysis cells,the rate of change of an electrolysis cell with a rated capacity of 10 MW can reach 1 MW/s.The maximum change rate of this working condition curve is 88.93 kW/s,thus this operating condition conforms to the variable load characteristics of PEM electrolysis cells.

This typical operating condition combines the characteristics of the annual output fluctuations of wind farms.Its application to a PEM electrolysis hydrogen production device for subsequent experimental research can effectively reduce time and economic costs and has good reference significance for studying material aging and life decay of PEM electrolysis hydrogen production systems under fluctuating power input.

6 Conclusion

This study considered the PEM electrolytic water hydrogen production system as the research object and proposed an extraction method for the typical working conditions of an off-grid wind power hydrogen production system considering the operating characteristics of PEM electrolysis cells,which is of great significance for the study of the dynamic response characteristics and performance attenuation mechanism of electrolysis cells with fluctuating power inputs.The conclusions are as follows:

1) This article summarized the operating characteristics of a PEM electrolysis water hydrogen production system and analyzed the performance degradation mechanism of the electrolysis cell under different fluctuation inputs from the chemical and mechanical aspects.

2) This study proposed a method for extracting the typical operating conditions of an off-grid wind power hydrogen production system based on the factors that affect the PEM electrolysis cell caused by wind power fluctuations.By extracting the characteristics from daily wind power output samples throughout the year and clustering them using the SOM neural network algorithm,a typical working condition curve of the electrolysis cell was generated based on the probability distribution in the time sequence.

3) The feasibility of the proposed method was verified by analyzing the actual historical output data of a wind farm in North China.In the future,the performance degradation law of PEM electrolytic hydrogen production systems under long-term operation and corresponding improvement measures can be studied based on the typical working condition curve.

Acknowledgments

This study was supported by the National Key Research and Development Program of China (Program Number 2021YFB4000100),the Beijing Postdoctoral Research Foundation (Grant Number 2023-ZZ-63).

Declaration of Competing Interest

The authors have no conflicts of interest to declare.