Method and computer program product for determining degradation of an energy system
By employing a cluster-based data-driven approach and Bayesian inference, and utilizing degradation reference variables and cluster information, the problem of quantifying the degradation state of PEM electrolyzer systems was solved, enabling efficient and accurate degradation assessment and maintenance optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SIEMENS ENERGY GLOBAL GMBH & CO KG
- Filing Date
- 2024-09-04
- Publication Date
- 2026-05-29
Smart Images
Figure CN122122341A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for determining, evaluating, or quantifying the degradation of an energy system, such as an electrolyzer system. More specifically, the system may involve a PEM (polymer electrolyte membrane or proton exchange membrane) electrolyzer system. These PEM electrolyzer systems may consist of multiple electrolyzer stacks and / or modules.
[0002] In other words, the proposed scheme aims to improve the health estimation or predictive maintenance of PEM electrolyzers or water electrolyzers. This is supported in particular by the cluster-based performance model, which will be further outlined here. Background Technology
[0003] Electrolysis, particularly PEM water electrolysis (PEMWE), is becoming increasingly important due to its ability to synthesize green hydrogen, a key energy source for future industrial and civilian applications. In the face of climate change, producing green hydrogen through renewable energy capacity offers numerous opportunities to achieve climate goals across many industrial sectors, and correspondingly replaces traditional energy carriers.
[0004] In PEM electrolysis, water is electrochemically split into oxygen and hydrogen (H2) using renewable energy sources such as wind, solar, or photovoltaic power as a stopgap measure. Hydrogen, or hydrogen carriers, clearly has great potential as a storage medium, green fuel, or related fuel component for a wide range of industrial processes. A particular advantage of PEM electrolysis is its outstanding partial load capacity, for example, superior to alkaline solutions. This property is especially a prerequisite for utilizing intermittent renewable energy sources (both spatially and temporally).
[0005] Furthermore, PEM electrolysis is characterized by its relatively simple design, allowing operation at high power and high current density, and providing favorable purity and output of the produced hydrogen.
[0006] The PEM electrolyzer includes a membrane with associated catalyst layers on both sides, which are in contact with so-called gas diffusion layers or gas diffusion electrodes. On the outside of these (porous) gas diffusion electrodes, contact plates or bipolar plates for associated electrical contact with the cell also abut and contact the gas diffusion electrodes.
[0007] The gas diffusion electrode is also configured to take into account the mass transfer required during the expected operation of the electrolyzer. Furthermore, the gas diffusion electrode establishes the necessary conductivity to electrically contact the bipolar plates and the catalyst layer, thereby establishing the desired electrochemical reaction.
[0008] Hydrogen production via water electrolysis requires water as a reactant, where it is decomposed into oxygen and hydrogen. The relevant chemical reactions are as follows:
[0009] These reactions are spatially separated by the aforementioned ion-conducting membrane, which needs to be in electrical contact with both the cathode catalyst and the anode catalyst, respectively.
[0010] Quantifying and tracking the degradation status of water electrolyzer energy systems is essential for optimizing their operation and planning often complex and costly maintenance activities. This is especially true for electrolyzers operating at customer sites, where maintenance is even more challenging and service providers rely on customer cooperation. Therefore, regularly conducting standardized tests is often impractical.
[0011] Furthermore, measurements related to degradation vary significantly with specific operating conditions, such as operating temperature and operating current.
[0012] This invention attempts to address the aforementioned problems by introducing empirical models and degradation reference variables. Furthermore, as described below, a cluster-based approach that incorporates the degradation history of multiple related energy systems is part of this invention.
[0013] Currently, the degradation status of water electrolyzers is quantified by conducting "polarization profile" tests. This is a common practice, and it is also mentioned in unified protocols, such as those used for testing water electrolyzers operating at low temperatures.
[0014] However, this standard approach has three main drawbacks. First, it does not provide the necessary continuous quantification of degradation. Instead, such testing is performed only occasionally, such as at a test bench or customer site where the energy system or fleet is located. Sometimes it is even performed only on demand, i.e., only when the customer requests or agrees to perform such testing, and is therefore usually very rare.
[0015] The term “swarm” should preferably refer to a group of pre-existing, similar energy systems, such as a group with quantified or predefined aging behaviors.
[0016] As a second drawback, the associated energy or electrolyzer systems require adherence to rather complex and time-consuming (predefined) procedures to provide stable and reliable degradation results. For example, performing these procedures reliably typically takes several days. Therefore, frequently scheduling such tests is impractical.
[0017] Third, it is often difficult to ensure that tests are conducted under strictly comparable and reproducible conditions. Small deviations in temperature, such as only 1°C, can cause huge errors in measurements. Summary of the Invention
[0018] Therefore, the object of the present invention is to provide means for solving and / or alleviating the above-mentioned problems.
[0019] The objective is achieved through the subject matter of the independent claims. Advantageous embodiments are the subject matter of the dependent claims.
[0020] One aspect of the present invention relates to a method for determining or assessing the degradation of an energy system (e.g., an electrolyzer system).
[0021] The method includes providing an empirical (data) model of a degradation reference variable, wherein the reference variable is adapted to estimate the degradation of a given system among, for example, a group of similar energy system / electrolyzer systems, and wherein, in the model, at least one unknown model parameter, such as one of a plurality of unknown model parameters, is fitted. The term "similar" should be interpreted broadly, for example, to include systems that may differ from the reference system, but in the context of this invention, modeling or comparing data from the two mentioned systems still makes sense.
[0022] The method also includes providing cluster information from the group of systems to the model, wherein fitted model parameters from multiple systems are aggregated to form cluster-based parameters, wherein each parameter in the fitted model parameters includes an evolution over time or time intervals. The cluster information may include or relate to current data or knowledge that is available to date. In other words, this step can be understood as calculating the reference variable.
[0023] The method also includes matching or fitting the model to a target energy system (e.g., an electrolyzer system that actually needs to be quantified for degradation) by using cluster-based parameters as a prior probability distribution and through statistical inference, particularly Bayesian inference, wherein degradation variables or degradation states of the target system are calculated.
[0024] In other words, the present invention uses data-driven modeling techniques, while known solutions (e.g., polarization curve testing) rely on and require physical testing (in the field).
[0025] The key advantage of this invention lies in providing more efficient, continuous, and reliable degradation quantification, particularly independent of and unrelated to given operating conditions. The proposed model specifically allows for the calculation of degradation variables under substantially the same operating conditions, independent of actual operating circumstances. Therefore, the provided model offers robust health estimation or predictive maintenance capabilities, having proven highly reliable and providing very accurate degradation assessments. This, in turn, largely renders field maintenance testing redundant.
[0026] As another advantage, the proposed solution saves time and cost associated with physical testing. This, in turn, increases the availability and profitability of electrolysis facilities, and thus improves the efficiency of hydrogen production.
[0027] In one embodiment, the system is an electrolyzer system, and the degradation reference variable is a reference voltage. Accordingly, groups of (similar) systems can involve sets of (similar) electrolyzer systems. According to this embodiment, the advantages of the invention are particularly applicable to electrolyzer systems.
[0028] In one embodiment, the model maps cell voltage to the most relevant operating conditions of the electrolyzer system, such as operating current and temperature. This embodiment advantageously allows for consideration of relevant operating conditions, which is beneficial for general modeling and degradation estimation.
[0029] In one embodiment, the degradation reference variable is the concentration of an impurity gas, such as the hydrogen concentration in oxygen, or the oxygen concentration in hydrogen. Furthermore, according to an alternative embodiment, the degradation reference variable may be the fluoride release rate and / or the hydrogen production rate.
[0030] Essentially, all degradation parameters affected by the operating conditions of the energy system under discussion (e.g., current, temperature, pressure, time since startup, etc.) can be addressed using the method of this invention by transforming them to reference conditions. Therefore, all such parameters are equally suitable as degradation reference variables.
[0031] In one embodiment, the degradation reference variable is fitted to multiple operating criteria or conditions, including operating time (e.g., the number of hours after system startup), operating temperature, and / or current or current density. Therefore, the operating time preferably does not refer to the total time measuring aging or degradation, nor is it the total number of operating hours of the system. As another example, the operating temperatures of several systems in this group may differ by several degrees (°C), where one system may reach an operating temperature of 57°C, while another system reaches, for example, 59°C.
[0032] In one embodiment, the logarithm of the operating time since the last startup of the corresponding energy system is parameterized in the fitted deterioration reference variable. This feature advantageously allows transient effects to be corrected to a common reference time or steady-state time, such as 100 hours (h), and still allows for another, preferably much slower, time dependence, i.e., for example, across many intervals within these 100-h intervals.
[0033] In one embodiment, the empirical model is a linear model comprising several, for example three or more, preferably five, unknown parameters, and a reference variable is used to fit these unknown model parameters. This particular embodiment has proven advantageous and provides a good balance between accuracy or complexity on one hand and practicality and computational time on the other.
[0034] In one embodiment, the degradation reference variable is fitted over multiple time intervals, preferably for each time interval. Therefore, cluster information or aggregated cluster-based parameters are also provided to each model interval. The mentioned time interval or interval duration can be a “sampling” time interval, which is preferably chosen to be not too long to allow for an appropriate time resolution (e.g., comparable to the time since a given system started), but also long enough to ensure a sufficiently large database.
[0035] To avoid ambiguity, it is not necessarily required to fit a reference variable for every time interval of its dynamic development. That is, the proposed inventive concept can still function if some data is missing, such as data being lacking in a particular interval. In any case, some absence or loss of data is permissible under the concept of the present invention.
[0036] The interval can be, for example, one day or more. If the chosen interval is too short, there is too little time-series data to fit. However, if the interval is too long, degradation may occur simultaneously, thus reducing the accuracy of the model. Therefore, given the time-varying degradation dynamics of the PEM electrolyzer system under operating conditions, it is appropriate to fit a model analytically over multiple time intervals.
[0037] In one embodiment, the data reliability of the fitted parameters is considered and recorded (preferably for each interval) according to the data distribution. This data reliability consideration can advantageously validate, clean, and / or correct for obvious outlier values in the data inventory.
[0038] In one embodiment, regression analysis with probability estimation, such as Gaussian process regression, is used to aggregate the fitted model parameters. The relevant kernel function, particularly the radial basis function kernel, can be used as input to the regression analysis. This approach has proven to be a highly advantageous, reliable, and efficient projection of statistical data.
[0039] In one embodiment, Gaussian process regression is used for aggregation, and kernel functions, particularly radial basis function kernels, are used as inputs to the regression analysis. Therefore, the advantages of relevant kernel functions can be utilized in the proposed scheme.
[0040] In one embodiment, a model is fitted or a matching is performed via Bayesian inference, for example, by using data from the target system to combine information from aggregated cluster parameters and operational data from the target system.
[0041] Typically, a key advantage of Bayesian inference schemes is that the probabilities of hypotheses can be updated as more evidence or information becomes available over time. In this case, the aggregated cluster parameters are relevant to the hypothesis, while the "operational data" of the target system can be considered "evidence."
[0042] In one embodiment, the model is configured to estimate degradation variables of a target electrolyzer system (i.e., an electrolyzer system for which degradation quantification is required) at a daily resolution. According to this embodiment, the resolution can be customized based on operational needs (e.g., the frequency of condition monitoring).
[0043] In one embodiment, an empirical model of a degraded reference variable is obtained from a feature selection method, wherein the fit quality across multiple datasets with different features is taken into account.
[0044] Another aspect of the invention relates to a computer program product comprising executable program instructions configured to perform the following steps when executed: providing an empirical model of a degradation reference variable for an energy system, wherein the reference variable is adapted to estimate the degradation of a given system among a group of similar energy systems, and wherein at least one unknown model parameter is fitted; providing cluster information from the group of systems to the model, wherein fitted model parameters for multiple systems are aggregated to form cluster-based parameters, wherein each of the fitted model parameters contains an evolution over time; and matching the model to a target energy system via statistical inference, particularly Bayesian inference, using the cluster-based parameters as a prior probability distribution, wherein a degradation variable for the target system is calculated.
[0045] Another aspect of the invention relates to a computer-readable medium comprising executable program instructions configured to, when executed, accordingly perform the mentioned steps, namely: providing an empirical model of a degradation reference variable for an energy system, wherein the reference variable is adapted to estimate the degradation of a given system among a group of similar energy systems, and wherein at least one unknown model parameter is fitted; providing cluster information from the group of systems to the model, wherein fitted model parameters for multiple systems are aggregated to form cluster-based parameters, wherein each of the fitted model parameters contains an evolution over time; and matching the model to a target energy system via statistical inference, particularly Bayesian inference, using the cluster-based parameters as a prior probability distribution, wherein a degradation variable for the target system is calculated.
[0046] Another aspect of the invention relates to an apparatus for determining the degradation of an electrolyzer system (PEMWE), the apparatus comprising a modeling unit or a computing unit, preferably including a processor, the unit being configured to implement an empirical model of a degradation reference variable, wherein at least one unknown model parameter is fitted in the model.
[0047] The device also includes a reading module coupled to a modeling unit for recording cluster information from a group of similar electrolyzer systems, wherein fitted model parameters of multiple electrolyzer systems are aggregated to form cluster-based parameters; and a (degradation) testing module configured to use the cluster-based parameters as a prior probability distribution and to calculate or quantify degradation variables of the electrolyzer systems by matching the electrolyzer systems to the model via statistical inference, particularly Bayesian inference.
[0048] The various modules or units mentioned in this application are broadly understood as entities capable of acquiring, obtaining, receiving, or retrieving general data and / or instructions through a user interface and / or programming code and / or executable programs, or any combination thereof. In particular, the modeling unit and the various modules mentioned below may be adapted to run programming code and executable programs and pass results for further processing.
[0049] Therefore, the different modules and units or portions thereof mentioned in this application may each include at least a central processing unit, a CPU, and / or at least one graphics processing unit, a GPU, and / or at least one field-programmable gate array, an FPGA, and / or at least one application-specific integrated circuit, an ASIC, and / or any combination thereof. Each of them may also include working memory operatively connected to at least one CPU and / or non-transitory memory operatively connected to at least one CPU and / or working memory.
[0050] The computer program product or computer-readable medium mentioned in this application may refer to a computer program or medium that constitutes or includes (volatile and non-volatile) memory cards, USB flash drives, CD-ROMs, DVDs, or files downloaded from or downloadable from a server or network. Such a product may be provided via a wireless communication network or by transmitting corresponding information from a given computer program, computer program product, etc. The computer program product may be, include, or be included in a (non-transitory) computer-readable storage medium storing applications, programs, program modules, scripts, source code, program code, object code, bytecode, compiled code, interpreted code, machine code, executable instructions, etc.
[0051] All parts of the system of this invention can be implemented in hardware and / or software, wired and / or wireless, and any combination thereof. Any part of the system may include an interface to an intranet or the Internet, to cloud computing services, to a remote server, etc.
[0052] In particular, each module and unit of the device may be implemented in part and / or entirely in a local computer, in a system of multiple computers, and / or in part and / or entirely in a remote system, especially in an edge computing platform or a cloud computing platform.
[0053] In cloud computing-based systems, a large number of devices are connected to the cloud computing system via the Internet. These devices can be located in remote facilities connected to the cloud computing system. For example, the devices can include equipment, sensors, actuators, robots, and / or machinery in industrial installations, or components thereof. The devices can be home devices, office devices, or industrial devices.
[0054] Cloud computing systems enable remote configuration, monitoring, control, and maintenance of connected devices. Furthermore, cloud computing systems facilitate the storage of large amounts of data periodically collected from devices, data analysis, and the provision of insights (e.g., parameters or relevant key performance indicators) and alerts to device operators, field engineers, or owners via graphical user interfaces (e.g., web applications).
[0055] A cloud computing system can also include multiple servers or processors (also known as “cloud infrastructure”) that are geographically distributed and interconnected via a network.
[0056] The advantages and embodiments associated with the described methods for determining degradation and / or the described computer program product or computer-readable medium are equally effective or applicable to the described apparatus, and vice versa.
[0057] Furthermore, the features and advantageous embodiments will become apparent from the following description of exemplary embodiments in conjunction with the accompanying drawings. Attached Figure Description
[0058] Figure 1 Exemplary and qualitative polarization (I / V) curves of energy systems, particularly electrolyzer systems, are shown using the cell voltage (reference voltage) as a function of current.
[0059] Figure 2 The evolution of the reference voltage Uref with time t is qualitatively illustrated using an exemplary curve.
[0060] and Figure 2 resemblance, Figure 3 Using the slot voltage as a function of the operating current, a quantitative exemplary polarization curve is shown, but multiple datasets are presented.
[0061] Figure 4 An exemplary process of modeling parameters over time, based on the data stock of multiple energy systems, is shown in the empirical model according to the present invention.
[0062] Figure 5 It shows the relationship with Figure 4 The curves shown are similar, but with added cluster knowledge, which is aggregated, cluster-based parameter information.
[0063] Figure 6The Bayesian inference scheme for matching the model to a specific energy system is illustrated through three different sub-images.
[0064] Figure 7 An exemplary dataset of predicted degradation of a target energy system is shown, which is parsed according to three different operating conditions.
[0065] Figure 8 The apparatus and method steps of the present invention, along with further information regarding the introduced data-driven degradation estimation method, are illustrated in the form of a basic flowchart. Detailed Implementation
[0066] In the accompanying drawings, similar elements, elements of the same kind, and elements with the same function may have the same reference numerals. The drawings are not necessarily depicted to scale and may be enlarged or reduced to allow for a better understanding of the principles illustrated. Rather, the drawings described are to be understood in a broad sense and as a qualitative basis that allows those skilled in the art to apply the presented teachings in a general manner.
[0067] The term "and / or" as used in this application means that each of the listed elements can be used alone or in combination with two or more other elements.
[0068] Figure 1 A voltage plot (similar to the polarization curve of an electrolyzer) is shown, particularly the reference voltage Uref as a function of the relevant operating current. The curves shown should preferably and generally suitably indicate the deterioration behavior of an energy system (see reference numerals 1 and 1'), especially an electrolyzer system such as a PEM electrolyzer system. The plot includes a point cloud of measured data points D and a relevant (best-fit) regression line M indicating the model fit. However, in practice, modeling the relationship between cell voltage and relevant operating conditions remains a challenge, and is particularly tricky with limited data coverage. Reference Figure 1 This problem is represented by data point D', which is significantly higher than the model line M and was not taken into account in the model fitting.
[0069] Figure 2 and Figure 3 The diagram shows the evolution, particularly the increase, of the reference cell voltage Uref of electrolyzer system 1 with time t as the electrochemical overpotential increases. Here, the symbol t represents the (overall) degradation time of the system in question.
[0070] In the current condition of the electrolytic cell, the increase in overpotential is due to any kind of degradation that makes the cell actually require more energy than thermodynamically expected. Therefore, Figure 2 The measurement of Uref in the data increased slightly over time.
[0071] Pay attention to the change in tank voltage with current or current density, such as Figure 3 As shown, it is clear that at the beginning of the lifetime (BoL) of the relevant tank or system, the voltage does not exceed a value of about 1.85V, while, for example, in a significantly degraded state of the tank, the voltage or potential increases significantly to 2.2V within the same current range.
[0072] In order to solve the problems addressed by the present invention, particularly as described above... Figure 1 To address the described problem, a method is proposed for determining the degradation of energy systems 1, 1' (e.g., an electrolyzer system). Specifically, this method provides or introduces an empirical model M for a degradation reference variable Uref, wherein the reference variable Uref is adapted to estimate the degradation of a given system 1 among a group of similar energy systems 1', and wherein at least one unknown model parameter Cn is fitted.
[0073] As can be seen from the following equation, multiple parameters Cn, such as C1, C2, C3, C4, and C5 (C1 and C3 are preferably greater than zero, while C4 is preferably less than zero), are part of the exemplary model according to the present invention: U = reversible voltage (U0) + overpotential. As mentioned above, the overpotential takes into account the influence of the current I, I max It is the maximum current, T is the temperature, and t is the running time.
[0074] More specifically, the voltage can be represented by the following linear model: U = U0 + C1I + C2IT + C3ln(t') + C4ln(1 - I / I) max )+C5.
[0075] The coefficients or parameters Cn are fitted separately, and preferably fitted for each time interval. Without departing from the concept of the invention, any other number of (unknown) parameters may be used, such as three parameters Cn, four parameters Cn, or even more, such as six, seven, eight or more parameters Cn.
[0076] Essentially, all degradation-related indicators affected by system operating conditions (such as current, temperature, pressure, or time since startup) can be handled using this method; they are suitable and can be used as degradation reference variables. For example, impurity gas concentration, fluoride release rate, and / or hydrogen production rate are, in principle, effective degradation indicators.
[0077] The proposed method is particularly suitable and advantageous for correcting for degradation-related variables with two interfering timescales: long-term (deterioration over many years) and short- or medium-term. The short- or medium-term timescale may involve other effects, such as reversible degradation due to catalyst oxidation that lasts only a few days. Short- or medium-term effects often interfere with the convenient observation of long-term degradation.
[0078] In addition to reversible degradation due to catalyst oxidation (which is already covered by the "ln(t)" term), water contamination can also cause this short- or medium-term effect. For example, water contaminants can impede the electrolysis process and result in a fairly high voltage after each system startup, which stabilizes after several days.
[0079] Without loss of generality, this invention can be implemented using any other (nonlinear) model. When using or fitting even more complex terms and parameters, it is still possible to improve the accuracy of the model and degradation assessment at the cost of more complex schemes, which may consume more computational effort.
[0080] The proposed modeling scheme is derived from or obtained from feature selection methods, which takes into account the fitting accuracy across multiple datasets with different features.
[0081] Figure 4 The graph illustrates several exemplary processes of any of the parameters Cn (e.g., representing C1) as examples, plotted as a function of time t (the notation in the graph preferably refers to days). The numbers on the x-axis specifically represent the number of days since the installation of the given energy system or electrolyzer system 1. The plots shown largely overlap each other, the parameters exhibit a slight upward trend, and peak and outlier data are also shown separately.
[0082] This invention incorporates cluster knowledge, such as that of groups or clusters of (similar) electrolyzer systems, to achieve the most accurate degradation assessment possible. In particular, cluster knowledge allows for the incorporation of multiple operating conditions into degradation prediction.
[0083] Therefore, as Figure 5 As shown, the degraded reference variable Uref is fitted specifically for each of multiple time intervals (not explicitly shown), and the cluster information is also provided to the model M on an interval-by-interval or for each interval.
[0084] In addition, based on the data distribution, the data reliability of the fitted parameters is considered and recorded, wherein the aforementioned outlier data values are preferably cleared and deleted throughout the distribution.
[0085] In fact, the way the cluster knowledge is incorporated into the model is that the scheme of the present invention provides cluster information F from a group of systems 1, 1' to the model M, wherein the fitted model parameters Cnm (each parameter Cn includes the evolution over time t) of multiple systems in these systems 1, 1' are aggregated to form cluster-based parameters Cn'.
[0086] Index n represents different coefficients or parameters, while the index m mentioned above can be associated with different (similar) energy systems.
[0087] and Figure 4 compared to, Figure 5 The regression line F of the cluster information is shown, and is therefore obtained by aggregating the parameters Cnm. The symbol CI also indicates the confidence interval. The Gaussian process model describes the probability distribution of a function that may fit a set of points, and it can provide a probability prediction by providing the mean and standard deviation as outputs in the form of the confidence interval CI. The interval CI can then be calculated, for example, using the standard deviation.
[0088] The regression type used for aggregation is preferably Gaussian process regression, and a kernel function, particularly a radial basis function kernel, is used as the input to the process regression. In addition to the kernel, there may be another input or hyperparameter that takes noise into account and controls the smoothness of the function F.
[0089] In statistical modeling, regression analysis is typically a set of statistical procedures used to estimate the relationship between a dependent variable and one or more independent variables. The most common form of regression analysis is linear regression, where a line is found that best fits the data according to specific mathematical criteria. Regression analysis is broadly used for prediction and forecasting purposes. More specifically, in statistics, Gaussian process regression is a nonparametric probabilistic technique used to interpolate data (such as the parameter Cnm in this case) with estimations of uncertainty.
[0090] Figure 6 A basic schematic diagram of the principle of so-called Bayesian inference applied to the proposed degradation estimation scheme is also shown. This method requires matching the model M with the target energy system 1 via statistical inference in the third step by using the cluster-based parameter F as the prior probability distribution Pr of the target system and the data.
[0091] exist Figure 6 At the bottom, the probability distribution P, as a function of the exemplary parameter C1, is qualitatively shown. Within the plot, three partial probability distributions are shown: the prior probability distribution Pr, the posterior probability distribution Pos, and the data likelihood L. Figure 6 The upper right portion shows that the likelihood L is taken from, for example, the operating data of target energy system 1 on a given date.
[0092] Bayesian inference typically derives the posterior probability from two preconditions: the prior probability and a likelihood function derived from a statistical model of the observed data (see above). Then, Bayesian inference calculates the posterior probability based on known Bayesian theors, in the current case: Pos(C ǀ data)~L(data ǀ C) Pr(C).
[0093] Figure 6The upper left part also shows that, through this Bayesian scheme, the data points for parameter C to date (see vertical dashed line) can be reliably inferred or calculated.
[0094] Finally, the degradation variable U or Uref can be calculated by reasoning about each relevant parameter in the above modeling equation (see the attached figure above).
[0095] pass Figure 7 The results of the latter calculation (i.e., given the reference variable U of the energy system or electrolyzer module) are shown for three different operating conditions. Each operating condition is represented by a line, and these lines are... Figure 7 The plotted content is arranged vertically. It should be understood in this application that the reference voltage is considered as a degradation variable only as an example. As mentioned above, other degradation indicators can also be treated similarly according to the present invention.
[0096] At a very low position, the voltage under one operating condition is shown, which means an operating temperature of 57°C, continuous operation for three days after system startup, and a current of 1500A. This voltage fit actually yields a mean square value, or root mean square deviation or error (RMSE), of only 0.0054V, which impressively demonstrates the accuracy of the proposed degradation estimate.
[0097] The other two point clouds (or lines, depending on the specific context) showed similar behavior, among which, Figure 7 The midpoint cloud represents the operating temperature of 59°C and the operating time of three days at a current of 4950A. In this case, the RMSE is even slightly smaller, at 0.0052V.
[0098] For the third estimated operating conditions, namely a temperature of 57°C, one day after startup, and a current of 7500A, the model remains accurate and shows a root mean square error of 0.0278V.
[0099] Figure 8 The provided method summarizes the data process by providing simplified diagrams of different process steps i), ii), and iii) through flowcharts and classifications.
[0100] Using terminology different from the invention description above, the method includes constructing an empirical voltage model in step i) that maps the relationship between voltage and the most relevant operating conditions (e.g., current and temperature). The model shown currently has a linear exemplary form, has multiple, for example, five unknown parameters, and preferably has an average fitting error, such as RMSE, of less than 1 mV. This simple yet accurate model is achieved through a "feature selection across multiple datasets" approach.
[0101] In the second fundamental methodological step ii), swarm knowledge of clusters or groups of similar electrolyzers, along with information about how the unknown parameters evolve over time, is input into the model. As mentioned above, this may mean fitting an empirical voltage model for each time interval (e.g., daily) for all similar electrolyzers or systems. Furthermore, the reliability of the fitted parameters is recorded for each interval and according to the data distribution. The fitted parameters are then aggregated with good reliability to form swarm knowledge, i.e., knowledge about how the unknown parameters actually evolve.
[0102] In iii), the empirical model is fitted to the target electrolyzer (which requires actual degradation quantification). This is accomplished using the Bayesian approach mentioned above, which uses the swarm knowledge from the second step as a prior probability distribution input to the Bayesian approach.
[0103] exist Figure 8 On the right side, device 10 is also depicted. Device 10 is adapted to determine the degradation of electrolyzer system 1. Device 10 includes a modeling unit 11 for implementing an empirical model M of the degradation reference variable Uref, wherein at least one unknown model parameter Cn is fitted in model M.
[0104] The device 10 also includes a reading module 12 coupled to the modeling unit 11 for recording cluster information from a group of similar electrolyzer systems 1, 1', wherein fitted model parameters Cn' of multiple electrolyzer systems 1, 1' are aggregated to form cluster-based parameters Cn', as described above. The device 10 also includes a testing module 13 configured to use the cluster-based parameters Cn' as a prior probability distribution Pr and to calculate the degradation variable U of electrolyzer system 1 by matching electrolyzer system 1 with model M via described statistical inference, particularly Bayesian inference.
[0105] Furthermore, as indicated by reference CP in the attached figure, a computer program product CP is proposed, comprising executable program instructions I, wherein the computer program product CP is configured to run the method steps of the present invention when executed, namely (i) providing an empirical model M of a degradation reference variable Uref of an energy system, wherein the reference variable Uref is adapted to estimate the degradation of a given system 1 among a group of similar energy systems 1, 1', and wherein at least one unknown model parameter Cn is fitted; (ii) providing cluster information from the group of systems 1, 1' to the model M, wherein the fitted model parameters Cnm of multiple systems 1, 1' are aggregated to form a cluster-based parameter Cn', wherein each of the fitted model parameters contains an evolution over time; and (iii) matching the model M with a target energy system 1 via statistical inference, particularly Bayesian inference, using the cluster-based parameter Cn' as a prior probability distribution, wherein the degradation variable U of the target system 1 is calculated.
[0106] As another part of the present invention, a computer-readable medium M is proposed, which includes executable program instructions I, wherein the computer-readable medium M is configured to perform the above-described steps of the present invention when executed.
Claims
1. A method for determining the degradation of an electrolyzer system (1, 1'), said electrolyzer system being, for example, an electrolyzer system, the method comprising the following steps: - (i) Provides an empirical model (M) for a degradation reference variable (Uref), wherein the reference variable (Uref) is adapted to estimate the degradation of a given system (1) among a group of similar electrolyzer systems (1, 1'), and wherein at least one unknown model parameter (Cn) is fitted. - (ii) Provide the cluster information (F) from the group of systems (1, 1') to the model (M), wherein the fitted model parameters (Cnm) of multiple systems (1, 1') are aggregated to form cluster-based parameters (Cn'), wherein each parameter (Cn) in the fitted model parameters contains an evolution over time (t). - (iii) The model (M) is matched with the target electrolyzer system (1) by using cluster-based parameters as a prior probability distribution (Pr) via statistical inference, particularly Bayesian inference, wherein the degradation variable (U) of the target system (1) is calculated.
2. The method according to claim 1, wherein, The degradation reference variable (Uref) is the reference voltage.
3. The method according to claim 2, wherein, The model (M) maps the cell voltage to the most relevant operating conditions of the electrolyzer system (1, 1').
4. The method according to claim 1, wherein, The degradation reference variables (Uref) are the concentration of impurity gases, the fluoride release rate, and / or the hydrogen production rate.
5. The method according to any one of the preceding claims, wherein, The degradation reference variable (Uref) is fitted to multiple operating criteria, including operating time (t), operating temperature (T), and / or current (I).
6. The method according to any one of the preceding claims, wherein, The logarithm of the running time (t') since the last startup of the corresponding electrolyzer system is parameterized in the fitted deterioration reference variable (Uref).
7. The method according to any one of the preceding claims, wherein, The empirical model (M) is a linear model, which includes several, preferably five, unknown parameters (C1, C2, C3), and the reference variable (Uref) fits these unknown model parameters (Cn).
8. The method according to any one of the preceding claims, wherein, The degradation reference variable (Uref) is fitted over multiple time intervals.
9. The method according to any one of the preceding claims, wherein, The reliability of the fitted parameters should be assessed and recorded based on the data distribution.
10. The method according to any one of the preceding claims, wherein, Use regression analysis, especially Gaussian process regression, to aggregate the fitted model parameters (Cn).
11. The method according to claim 10, wherein, The aggregation was performed using Gaussian process regression, and kernel functions, particularly radial basis function kernels, were used as inputs to the regression analysis.
12. The method according to any one of the preceding claims, wherein, The model (M) is fitted by Bayesian inference using data from the target system as the so-called data likelihood (L).
13. The method according to any one of the preceding claims, wherein, The model (M) is configured to estimate the degradation variable (U) of the target electrolyzer system (1) at a daily resolution, the target electrolyzer system being the electrolyzer system for which degradation quantification is required.
14. A computer program product (CP) comprising executable program instructions (I) configured to perform the following steps when executed: - (i) Provide an empirical model (M) of the degradation reference variable (Uref) for the electrolyzer system (1, 1'), where, The reference variable (Uref) is suitable for estimating the degradation of a given system (1) among a group of similar electrolyzer systems (1, 1'), and wherein at least one unknown model parameter (Cn) is fitted. - (ii) providing cluster information from the group of systems (1, 1') to the model (M), wherein the fitted model parameters (Cnm) of multiple systems (1, 1') are aggregated to form cluster-based parameters (Cn'), wherein each parameter in the fitted model parameters contains an evolution over time. - (iii) The model (M) is matched with the target electrolyzer system (1) by using cluster-based parameters (Cn') as a prior probability distribution via statistical inference, particularly Bayesian inference, wherein the degradation variable (U) of the target system (1) is calculated.
15. A computer-readable medium (M) comprising executable program instructions (I), said computer-readable medium being configured to perform the following steps when executed: - (i) Provide an empirical model of the degradation reference variable (Uref) for the electrolyzer system (1, 1'), where, The reference variable (Uref) is suitable for estimating the degradation of a given system (1) among a group of similar electrolyzer systems (1, 1'), and wherein at least one unknown model parameter (Cn) is fitted. - (ii) Provide the model with cluster information from the group of systems, wherein the fitted model parameters (Cnm) of multiple systems (1, 1') are aggregated to form cluster-based parameters (Cn'), wherein each parameter (Cn) in the fitted model parameters contains an evolution over time. - (iii) The model (M) is matched with the target electrolyzer system (1) by using cluster-based parameters (Cn') as a prior probability distribution (Pr) via statistical inference, particularly Bayesian inference, wherein the degradation variable (U) of the target system (1) is calculated.
16. An apparatus (10) for determining the degradation of an electrolyzer system (1), the apparatus comprising: - A modeling unit (11) for implementing an empirical model (M) of the deteriorating reference variable (Uref), wherein at least one unknown model parameter (Cn) is fitted in the model (M). - A reading module (12), coupled to the modeling unit (11), for recording cluster information from a group of similar electrolyzer systems (1, 1'), wherein fitted model parameters (Cn') of multiple electrolyzer systems (1, 1') are aggregated to form cluster-based parameters (Cn'), and - Test module (13), which is configured to use the cluster-based parameters (Cn') as a prior probability distribution (Pr) and to calculate the degradation variable (U) of the electrolyzer system (1) by matching the model (M) with the electrolyzer system (1) through statistical inference, particularly Bayesian inference.