Distributed learning for optimized operation of electrolyzer modules

Abstract

The invention relates to a system (1) for improving the operation of electrolyzer modules (2), the system (1) comprising several electrolysis plants (3) each with several electrolyzer modules (2) and a central unit (4) connected to the electrolysis plants (3) for data exchange, which is configured to collect model parameters of the electrolyzer modules (2) of the electrolysis plants (3), to aggregate these model parameters and to update a model for electrolyzer modules (2), wherein the electrolysis plants (3) each comprise a controller (5) connected to the electrolyzer modules (2) and each comprise a local training module (6) connected to the controller (5), which is configured to update a statistical model of the electrolyzer modules (2) from controller data and to forward model parameters of the updated statistical model to the central unit (4).The central unit (4) receives processed model parameters and forwards them, in a prepared form, to the respective controller (5) for controlling the electrolyzer modules (2). The invention further relates to a method for improving the operation of electrolyzer modules (2) in electrolysis plants (3).

Description

TECHNICAL FIELDThe invention relates to a system for improving the operating mode of electrolyser modules and to a corresponding method.BACKGROUNDThe operating mode of electrolyser modules represents a very great influence on the degradation behaviour of the components such as, for example, the membranes. It could be shown that excessive switching accelerates the degradation of the electrolyser modules. A gentle operating mode can remedy this. Several approaches for data analysis and control optimization are known from the prior art.The model predictive control (MPC) is, for example, a method of optimizing the control of complex, generally multi-variable processes. It is based on a discrete-time dynamic model of the system and uses real-time data to calculate optimum control signals. MPC can be used to optimize the operating parameters of electrolysers and to ensure high efficiency. However, this method requires considerable computing power.Alternatively, by developing detailed physical models and simulations of electrolysers, different operating conditions can be virtually tested and the optimum settings determined. These models can be used to predict performance under various conditions, thus optimizing control. However, many electrochemical processes and the relationship with the operating mode are still unknown and can only be physically modeled with great difficulty.One way the data-driven approach could offer via static models or machine learning approaches. Correlations between degradation properties and operating modes are determined and data-driven models are developed. Models may be developed to detect anomalies and predict when maintenance is required, or generally to monitor the operating state of the electrolyser.The modeling can thus achieve some competitive advantages, such as a longer operating time, robust modes of operation of the systems or, in general, a deeper understanding of the extremely complex behavior patterns of the electrochemistry in the cell.Machine learning requires large amounts of data, which are usually collected centrally in order to generate a dataset therefrom for training. However, just the exclusive access to the operating data required for machine learning represents an obstacle. For example, the operator of an electrolysis plant could consider information about its mode of operation as sensitive and could not make the pure operating data available for further analysis.Up to now, by experimental examination, physical or data-driven modeling, substitute models are determined which allow prediction for important performance characteristics such as degradation and thus allow a more intelligent and gentler operating strategy. However, the plurality of different environmental influences does not allow the customized modeling for a specific installation with the associated environmental influences.SUMMARY OF THE INVENTIONIt is an object of the invention to specify a system for improving the operating mode of electrolyser modules. It is a further object of the invention to specify a method for improving the operating mode of electrolyser modules in electrolysis plants.The invention achieves the object directed to a system for improving the operating mode of electrolyser modules by providing that such a system comprises a plurality of electrolysers each having a plurality of electrolyser modules and a central unit connected to the electrolysers for exchanging data, said central unit being configured to collect model parameters of the electrolyser modules of the electrolysers, to compress these model parameters and to update a model for electrolyser modules, wherein the electrolysers each comprise a controller connected to the electrolyser modules and a local training module connected to the controller, said local training module being configured to update a statistical model of the electrolyser modules from controller data, to forward model parameters of the updated statistical model to the central unit, to receive model parameters processed in the central unit and to forward these processed to the respective controller for controlling the electrolyser modules.In their article entitled "Communication-Efficient Learning of Deep Networks from Decentralized Data" https: / / arxiv.org / abs / 1602.05629, the authors H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson and Blaise Aguera y Arcas describe a method for being able to create learning models for specific applications from the volume of data of modern mobile devices, which may lead to an improved user experience, for example on the basis of improved voice models for speech recognition or a facilitated text input. Although the mobile device hardware is different, this is immaterial for improving the respective identical application. Since these comprehensive data are often problematic with regard to data protection, very comprehensive or both, it is proposed to leave the training data on the respective mobile devices and to create a common model by aggregating updates calculated locally.Such a machine learning approach, in which a distributed network of devices works jointly on a model without sharing central data, is referred to as "federated learning". This approach may be useful in mobile telephones, since it improves data protection and uses the computing power of the individual devices. Electrolysis systems, however, must take into account some factors which could prevent a person skilled in the art from using federated learning.Electrolysis plants are generally large-scale industrial plants for preparing hydrogen by splitting water into oxygen and hydrogen. The technical requirements and processes of these installations differ greatly from those of mobile telephones and other devices in which federation learning is frequently used.Further, electrolysis systems are less widespread compared to mobile phones, reducing the benefit of sprung learning, as this approach typically relies on a large number of devices.In electrolysis plants, data protection may be less critical and established mechanisms for data exchange between plant operators could already be present.And finally, unlike mobile phones, electrolysis plants may not have sufficient computational resources to run machine learning algorithms, which could degrade their performance or require additional hardware.Overall, a person skilled in the art would therefore not necessarily keep springed learning suitable for electrolysis systems, in particular with regard to challenges in the implementation of springed learning, such as, for example, the coordination of the model updates, the ensuring of data security and the handling of devices or systems which have different computing powers and network connections.Nevertheless, fed learning also has advantages for electrolyser modules. For example, an improvement with respect to data protection can be achieved compared to the established mechanisms mentioned above for the data exchange between plant operators. Further, communication demand is reduced because, rather than transmitting large amounts of data over the network, only the model updates are transmitted, resulting in a lower network load. With regard to conventional assessment, electrolysis systems, unlike mobile phones, may not have sufficient computational resources to execute machine learning algorithms, which could impair their performance or require additional hardware, it turns out that training the models on the local training modules may utilize their computational power rather than providing additional central computational resources.In an advantageous embodiment of the invention, the local training modules are configured in such a way that they learn relationships between operating modes and performance characteristics of the electrolyser modules.By understanding the relationships between operating modes and performance characteristics, operators can optimize the electrolyser modules in order to achieve higher efficiency and better energy utilization. This may result in a reduction in energy consumption and operating costs. The recognition of patterns and relationships enables a targeted adaptation of the operating parameters of the installation. This allows operators to ensure more stable and efficient production, which can extend the life of the plant and reduce maintenance costs.Since electrolyser modules may have different operating conditions, learning the relationships also helps adapt the plant to changed environmental conditions or load requirements. This allows more flexible and efficient use of the electrolysis plant.Further, potential problems or errors may be detected early. This allows preventive maintenance to avoid failures and maximize equipment availability.And finally, learning the relationships contributes to gathering knowledge of the optimal operating conditions and performance metrics of the electrolyser modules. This knowledge may be used for future plant planning and optimizations or exchanged between different plants to improve the efficiency and performance of the electrolysis industry as a whole.In a further advantageous embodiment of the invention, the local training modules are configured to generate and / or update a statistical model for electrolyser modules on the basis of data representative of operating characteristics of the electrolyser modules and to perform an optimization of the operation of electrolyser modules on the basis of this statistical model.It is expedient here if the optimization comprises one or more of the following points: overall efficiency of the electrolysis systems, degradation or service life of the electrolysis systems, degradation or service life of each of the electrolyser modules, operating reliability and maintenance patterns. This leads to an overall improvement in the plant performance.Advantageously, the local training modules can be synchronized by the central station. As a result, the central station ensures that all training modules have the same information and the current state of the model. This contributes to the consistency and effectiveness of the learning process. In addition, the central station can better monitor the performance of the individual electrolysis systems and, if required, can intervene in a targeted manner and thus detect and remedy problems or inefficiencies at an early stage. Finally, synchronization of the local training modules simplifies maintenance and updating of the model of the plants and ensures that all plants benefit from the latest findings and optimizations.It is furthermore advantageous if the data are periodically transmitted to the central station. By means of the periodic transmission of data, the control center always remains informed about the current operating state of the electrolysis systems. This allows for a timely response to possible problems, changes or inefficiencies, thereby allowing for faster adjustment and optimization.Furthermore, the regular transmission of data enables the central station to better identify long-term trends and patterns in the operating data of the electrolysis systems. This information can be used to identify improvement potentials or to predict future developments and requirements.A simple and thus advantageous way of combining the determined data is arithmetic averaging.The object directed to a method is achieved by a method for improving the operating mode of electrolyser modules in electrolysis plants, wherein controllers control and monitor the electrolysis modules in the electrolysis plants, wherein local training modules of the electrolysis plants update a statistical model of the electrolyser modules from controller data and transmit model parameters determined in this way to a control center, wherein the control center collects these model parameters of the electrolyser modules of the electrolysis plants, compresses them and updates a model for electrolyser modules, wherein updated model parameters are distributed to the local training modules of the electrolysis plants and transmitted in a processed manner to the controllers.Relationships between operating modes and performance characteristics of the electrolyser modules are advantageously learned.It is expedient here if in the electrolysis installations a statistical model for electrolyser modules is generated and / or updated on the basis of data which are representative of operating characteristics of the electrolyser modules and an optimization of the operating mode of electrolyser modules is carried out on the basis of this statistical model.In particular, it is expedient if the optimization comprises one or more of the following points: overall efficiency of the electrolysis systems, degradation or service life of the electrolysis systems, degradation or service life of each of the electrolyser modules, operating reliability and maintenance patterns.It is advantageous if both the local training modules are synchronized by the control center and the data are periodically transmitted from the local training modules to the control center.It is furthermore advantageous if the data are arithmetically averaged in the central office.The essential features of the invention are the local training modules (local learners) which learn relationships between the operating mode and KPIs (such as degradation) on the local operator systems. These local training modules are synchronized via a central server (master). At regular intervals, the parameters of the local training modules are interrogated and transmitted to the central server. Since abstract model parameters are involved, no conclusions can be drawn about the operating mode of the operators of the electrolysis plants. Sensitive information is thus not transmitted. This is done for many different operators. The collected model parameters are then aggregated on the master server. A very simple form is, for example, arithmetic averaging. Subsequently, the averaged model parameters are transmitted again to the various installations. These updates allow the prediction capability, robustness and generalizability of the local training modules on the basis of the connected fleet to be significantly increased and dynamically adapted.Dynamically data-driven updates can be transmitted to the facilities and learned for data that cannot normally be made accessible. Thus, the full potential of machine learning can be exploited and the operators of the electrolysis systems can be provided with local prediction models which are based on data of the entire fleet without actively utilizing operator data.BRIEF DESCRIPTION OF THE DRAWINGSFIG. System for Improving the Operation of Electrolyser Modules According to the Invention.DESCRIPTION OF THE EMBODIMENTThe FIGURE shows the structure of a system 1 according to the invention for improving the operating mode of electrolyser modules 2. in the exemplary embodiment of the FIGURE, the system 1 comprises three electrolysers 3 each having three electrolyser modules 2.A control center 4 is connected to the electrolysis systems 3 for the exchange of data. The central station 4 is configured such that it collects 7 model parameters of the electrolyser modules 2 of the electrolysis installations 3, compresses these model parameters 8, for example carries out arithmetic averaging, and updates 9 a model for electrolyser modules 2.The electrolysis systems 3 each comprise a controller 5 connected to the electrolyser modules 2 which, in the exemplary embodiment of the figure, is in turn connected to a local training module 6 via a database 10. The local training module 6 is configured such that it updates a statistical model of the electrolyser modules 2 from data from the controller 5 and in the process learns relationships between operating modes and performance characteristics of the electrolyser modules, passes model parameters of the updated statistical model on to the central station 4 preferably periodically and in synchronization (solid line), receives model parameters processed in the central station 4 (dashed line) and, on the basis thereof, passes corresponding instructions on to the respective controller 5 for controlling the electrolyser modules 2.In particular, the local training modules 6 are configured to generate and / or update a statistical model for electrolyser modules 2 based on data representative of operating characteristics of the electrolyser modules 2 and to perform optimization of the operation of electrolyser modules 2 based on this statistical model.The optimization comprises one or more of the following points: total efficiency of the electrolysis systems 3, degradation or service life of the electrolysis systems 3, degradation or service life of each of the electrolyser modules 2, operating reliability and maintenance patterns.LIST OF REFERENCE CHARACTERS1 System 2 Electrolyser module 3 Electrolyser plant 4 Central station 5 Controller 6 Local training module 7 Collecting model parameters 8 Compacting model parameters 9 Updating model parameters 10 DatabaseReferences included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Cited Non-Patent LiteratureCommunication-Efficient Learning of Deep Networks from Decentralized Data" https: / / arxiv.org / abs / 1602.05629 describe the authors H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson and Blaise Aguera y Arcas

[0011]

Claims

System (1) for improving the operating mode of electrolyser modules (2), the system (1) comprising a plurality of electrolysers (3) each having a plurality of electrolyser modules (2) and a central station (4) which is connected to the electrolysers (3) for exchanging data and is configured to collect model parameters of the electrolyser modules (2) of the electrolysers (3), to compress these model parameters and to update a model for electrolyser modules (2), wherein the electrolysers (3) each comprise a controller (5) connected to the electrolyser modules (2) and a local training module (6) connected to the controller (5) and is configured to update a statistical model of the electrolyser modules (2) from controller data, to forward model parameters of the updated statistical model to the central station (4), The model parameters processed in the central station (4) are received and forwarded in a processed manner to the respective controller (5) for controlling the electrolyser modules (2).The system (1) for improving the operation of electrolyser modules (2) according to claim 1, wherein the local training modules (6) are configured to learn relationships between operations and performance metrics of the electrolyser modules (2).The system (1) for improving the operation of electrolyser modules (2) according to any of claims 1 or 2, wherein the local training modules (6) are configured to generate and / or update a statistical model for electrolyser modules (2) based on data representative of operating characteristics of the electrolyser modules (2) and to perform an optimization of the operation of electrolyser modules (2) based on this statistical model.The system (1) for improving the operation of electrolyser modules (2) according to claim 3, wherein the optimisation comprises one or more of the following items: overall efficiency of the electrolysers (3), degradation or lifetime of the electrolysers (3), degradation or lifetime of each of the electrolyser modules (2), reliability and maintenance pattern.The system (1) for improving the operating mode of electrolyser modules (2) according to one of the preceding claims, wherein the local training modules (6) can be synchronised by the central station (4).The system (1) for improving the operation of electrolyser modules (2) according to any preceding claim, wherein the system is configured such that the data is periodically transmitted to the central office (4).The system (1) for improving the operating mode of electrolyser modules (2) according to one of the preceding claims, wherein the data are combined by means of arithmetic averaging.Method for improving the operating mode of electrolyser modules (2) in electrolysis plants (3), wherein controllers (5) control and monitor the electrolysis modules (2) in the electrolysis plants (3), wherein local training modules (6) of the electrolysis plants (3) update a statistical model of the electrolyser modules (2) from controller data and transmit model parameters determined in this way to a central station (4), wherein the central station (4) collects, compresses these model parameters of the electrolyser modules (2) of the electrolysis plants (3) and updates a model for electrolyser modules (2), wherein updated model parameters are distributed to the local training modules (6) of the electrolysis plants (3) and transmitted to the controllers (5) in a processed manner.Method according to Claim 8, wherein relationships between operating modes and performance characteristics of the electrolyser modules (2) are learned in the electrolysis installations (3).Method according to one of claims 8 or 9, wherein in the electrolysis installations (3) a statistical model for electrolyser modules (2) is generated and / or updated on the basis of data representative of operating characteristics of the electrolyser modules (2) and on the basis of this statistical model an optimization of the operating mode of the electrolyser modules (2) is carried out.Method according to claim 10, wherein the optimization comprises one or more of the following items: overall efficiency of the electrolysis plants (3), degradation or service life of the electrolysis plants (3), degradation or service life of each of the electrolyser modules (2), operating reliability and maintenance pattern.Method according to one of claims 8 to 11, wherein the local training modules (6) are synchronized by the central station (4).Method according to one of Claims 8 to 12, wherein the data are periodically transmitted from the local training modules (6) to the central station (4).Method according to one of Claims 8 to 13, wherein the data are arithmetically averaged in the central office (4).

Images