Performance change prediction for device subject to intermittent loading

Abstract

To provide a computer-implemented method, a program, and a system that generate a prediction model configured to output a prediction change in the performance index of an electrochemical device that is subject to intermittent loading, at least a portion of which is generated by one or more renewable energy sources.SOLUTION: A method for generating a predictive model that includes projecting the expected intermittent loading onto multiple basement functions to obtain the successor of each linear combination thereof and applying the successor of each linear combination to multiple fundamental changes comprises: obtaining multiple time series representing the actual intermittent loading of each, at least partially generated by one or more renewable energy sources; machine learning multiple basement functions representing the basic intermittent loading of each; and determining multiple basic changes for each basement function, including the basic changes in the performance index of the electrochemical device.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to the field of computer programs and systems, and more particularly to methods, systems, and programs for generating predictive models configured to output predicted changes in performance indicators of electrochemical devices subjected to loads that are expected to be intermittent. [Background technology]

[0002] As part of the energy transition, the use of electrochemical devices to store and / or release electrical energy intermittently is becoming crucial. Against this background, the demand for green hydrogen is growing at a remarkable pace. In practice, a key process of green hydrogen production is to feed electrolyzers with electricity generated from renewable energy sources. For example, water electrolyzers are used to produce hydrogen in combination with renewable energy sources. Electrolyzers usually operate with a constant and stable input, while renewable energy sources are characterized by a fluctuating nature. The application of intermittent loads with such fluctuating profiles to electrochemical devices can significantly alter their operation and, in particular, accelerate the degradation of said devices, resulting in reduced performance. The rate of degradation can have a significant impact on the operability of the project. However, experimental tests to predict degradation can be quite long-term or are not representative of all renewable profiles.

[0003] There is a group of white-box models that integrate the physicochemical mechanisms that cause the aging of electrochemical devices. On the other hand, a group of black-box models tries to predict the performance of the device without any knowledge of the inner workings of the device. Such black-box models are sometimes based on machine learning approaches. In general, the physical phenomena in electrochemical devices are controlled by complex multiphysics, making it difficult to provide fast and simple real-time white-box modeling. On the other hand, a huge amount of training data is required to obtain an accurate black-box model.

[0004] In this context, there remains a need for improved solutions for predicting changes in performance indicators of electrochemical devices subjected to loads that are expected to be intermittent. Summary of the Invention

[0005] Accordingly, there is provided a computer-implemented method for generating a predictive model configured to output a predicted change in a performance metric of an electrochemical device subjected to a load that is generated at least in part by one or more renewable energy sources and that is assumed to be intermittent. The method includes obtaining a plurality of time series, each representing a real intermittent load generated at least in part by one or more renewable energy sources and configured to load the electrochemical device. The method further includes machine learning a plurality of basis functions, each representing a respective underlying intermittent load. The plurality of basis functions forms a projection space of the plurality of time series, whereby the plurality of basis functions are configured to approximate each time series by a successor of a respective linear combination applied to the plurality of basis functions. The method further includes determining a plurality of basis changes, including a respective underlying change in the performance metric of the electrochemical device, for each basis function when the electrochemical device is subjected to a respective underlying intermittent load represented by the basis function. The predictive model includes projecting the expected intermittent load onto a plurality of basis functions, thereby obtaining respective successors of linear combinations that can be applied to the plurality of basis functions to approximate the load that is assumed to be intermittent, and applying the respective successors of the linear combinations to a plurality of basis changes, thereby outputting predicted changes.

[0006] The method comprises one or more of the following elements: - determining each basis change of a performance index of the electrochemical device, said step comprising, for each basis function: o obtaining at least one physical instance of an electrochemical device; o subjecting said at least one physical instance of an electrochemical device to at least one actual instance of said respective elementary intermittent loads represented by basis functions; o performing at least one measurement of a performance index for at least one physical instance of the electrochemical device; and Calculate each underlying change based on at least one measurement comprising; - the at least one measurement of the performance indicator in the at least one physical instance of the electrochemical device comprises a first measurement and a second measurement, the first measurement being performed at the beginning of the at least one actual instance of the respective basic intermittent load and the second measurement being performed at the end of the at least one actual instance of the respective basic intermittent load, and the calculation of each basic change consists in calculating a difference between a value obtained from the second measurement and a value obtained from the first measurement; - for at least one basis function, the at least one actual instance of each underlying intermittent load represented by the basis function is comprised of a plurality of actual instances of the respective underlying intermittent load; each time series includes a respective classification label, and the machine learning of the plurality of basis functions includes, for each time series, minimizing a classification error with respect to the classification label and / or a distance metric between the basis function and the provided time series; -The distance metric is dynamic time warping and the classification error is cross-entropy loss; -Multiple basis functions are composed of multiple shapelets; -each shapelet includes a length parameter, and the machine learning of the multiple basis functions further includes determining a length of each shapelet to minimize an error; - the machine learning of the plurality of basis functions further comprises selecting a subset of a predetermined size from the determined plurality of basis functions, the selected subset having a classification error in a first neighborhood of the minimized classification error and / or a distance metric in a second neighborhood of the minimized distance metric; - the performance index represents the efficiency in a constant mode, the temperature signal, the pressure signal, the ion concentration signal, or the gas concentration signal, and / or The electrochemical device comprises at least one of an electrolyzer connected to an intermittent power source, an electrochemical energy storage device connected to an intermittent power source, and an electrochemical energy storage device connected to an intermittent load.

[0007] There is further provided a computer-implemented method of using a predictive model generated according to the above-described method for a respective electrochemical device, the method including obtaining a load assumed to be intermittent, at least in part generated by one or more renewable energy sources, and applying the predictive model to the assumed intermittent load, thereby outputting a predicted change in a performance metric of the respective electrochemical device when the respective electrochemical device is exposed to the assumed intermittent load. Applying the predictive model includes projecting the assumed intermittent load onto a plurality of basis functions, thereby obtaining respective successors of linear combinations applicable to the plurality of basis functions to approximate the assumed intermittent load, and applying the respective successors of the linear combinations to the plurality of basis changes, thereby outputting the predicted changes.

[0008] Further provided is a computer program comprising instructions for carrying out any of the methods described above.

[0009] Furthermore, a computer-readable recording medium having this computer program recorded thereon is provided.

[0010] Further provided is a system including a processor coupled to a memory and a graphical user interface, the memory having a computer program recorded thereon. [Brief description of the drawings]

[0011] [Figure 1] An example of a system is shown. [Diagram 2-7] An example of this method is given below. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Non-limiting examples are described below with reference to the accompanying drawings.

[0013] A computer-implemented method for generating a predictive model is proposed. The predictive model is configured to output a predicted evolution of a performance indicator of an electrochemical device, the electrochemical device being subjected to a load that is assumed to be intermittent. The intermittent load is generated at least in part by one or more renewable energy sources. The method includes obtaining a plurality of time series. Each of the time series represents a respective actual intermittent load that is generated (i.e., formed) at least in part by one or more renewable energy sources. Each actual intermittent load is configured to be targeted to the electrochemical device. The method further includes a machine learning approach for identifying a plurality of basis functions. Each basis function represents a respective underlying intermittent load. The plurality of basis functions forms a projection space of the plurality of time series. The plurality of basis functions is thereby configured to approximate each time series by a respective successor of a linear combination applied to the plurality of basis functions. The method further includes determining a plurality of underlying changes. The plurality of basis changes comprises, for each basis function, a respective basis change in a performance index of the electrochemical device when the electrochemical device is subjected to a respective basis intermittent load represented by that basis function. The predictive model comprises obtaining respective successors of linear combinations by projecting the assumed intermittent load onto the plurality of basis functions. The respective successors of the linear combinations are applied to the plurality of basis functions to approximate the assumed intermittent load. The predictive model further includes applying the respective successors of the linear combinations to the plurality of basis changes thereby outputting predicted changes.

[0014] The method constitutes an improved solution for predicting the evolution of performance indicators of electrochemical devices. The method provides a simple tool for predicting the evolution of performance indicators of such devices. Such prediction is particularly important for electrochemical devices subjected to intermittent loads, at least in part generated by renewable energy sources. For example, in the case of renewable energy resources such as wind and solar resources, they are not always available (due to inevitable variations in the position of the sun and weather conditions) and are largely unpredictable, resulting in intermittent loads. The one or more renewable energy sources generating at least a portion of the assumed intermittent load and / or the one or more renewable energy sources generating at least a portion (e.g., each) of the respective actual intermittent loads represented by the multiple time series may consist of any one or any combination of solar energy sources, wind energy sources, and hydroelectric energy sources. Supplying such loads to a device, e.g. an electrochemical device, causes a much greater degradation in the performance of the device (e.g., compared to a constant load).

[0015] The change or modification of a performance index goes hand in hand with predicting said performance index after defining an initial condition for said performance index. Such initial condition for a performance index, e.g., the efficiency of an electrochemical device, is generally known for the electrochemical device, e.g., by the manufacturer. Thus, the expressions "prediction of the change in a performance index" and "prediction of a performance index" are used interchangeably herein.

[0016] The method does not require complex hardware and / or measurement techniques and can rely (e.g. solely) on processing data obtained simply from the device and calculations based on said data. The tool is based on the important observation that the change in performance for a given (intermittent) load can be obtained as the successor of a linear combination of the underlying changes (of said performance indicators). The underlying changes represent the influence of the underlying load (represented by the respective basis functions) on the performance change. The method obtains a number of said underlying loads by performing machine learning among a number of provided time series and determines the respective underlying changes, for example by performing a respective experiment. The experiment consists in subjecting the electrochemical device to the respective underlying intermittent loads and determining the respective underlying changes by observing / measuring the respective changes in the performance indicators.

[0017] The tools provided by the method are further beneficial in providing an efficient online / offline task separation. Having generated a predictive model in offline mode, the method is ready to receive the predicted loads online, project them onto multiple basis functions to obtain the successors of the respective linear combinations, and apply the obtained linear combinations to the multiple basis changes already determined. This makes the online phase simply a projection of the provided online loads and application of the respective linear combinations to the set of determined values, which is significantly faster and less computationally expensive than the offline phase. This improves the real-time performance of the method.

[0018] Moreover, in the context of the proposed method, it is confirmed that the decomposition of the prediction load via said projection, succeeded by a linear combination of (a reduced number of) basis functions, does not result in significant loss of relevant information about the task at hand, which allows the method to predict the evolution of performance metrics with high accuracy.

[0019] By "electrochemical device" is meant a device that includes at least one electrochemical cell. An electrochemical cell, as known per se, is a device that can generate electrical energy from a chemical reaction or use electrical energy to cause a chemical reaction. In an example, the electrochemical device may include at least one electrolyzer (configured / intended to be) connected to an intermittent power source (and configured to consume the power received from said power source to carry out an electrolysis process), an electrochemical energy storage device (configured / intended to be) connected to an intermittent power source (and configured to convert the power of the power source into the chemical potential energy of said energy storage device), and an electrochemical energy storage device (configured / intended to be) connected to an intermittent load (and configured to receive energy from said load or to release energy to said load). Hereby, an "intermittent load" means intermittently releasing or receiving power, in other words, an intermittent energy change, positive or negative. On the other hand, an "intermittent power source" is an intermittent power supply, in other words, a specific form of an intermittent load. In other words, an intermittent load may refer to an electrochemical device that operates in either a "sink" mode (e.g., an electrolyzer connected to a renewable power source) or a "send" mode (e.g., a battery connected to an engine operating intermittently).

[0020] Hereby, the expression "electrochemical device" refers to a class of electrochemical devices of the same type and using the same operating parameters. In particular, the method is capable of generating a predictive model of a class of electrochemical devices, which is applicable to all devices of a class (e.g., same model / reference from the same brand). This provides a universal predictive capability of the method.

[0021] "Electrochemical performance index" refers to any index used in the art to evaluate the performance of the electrochemical device. For example, the performance index can be the output voltage or current of the electrochemical device. Alternatively, the performance index can be the efficiency of the device. The efficiency of the device can be any efficiency used in the art. For example, the efficiency of the device can be the ratio of the output energy of the device to the input energy of the device.

[0022] As other examples, the performance index can represent any of the efficiency in constant mode, the temperature signal, the pressure signal, the ion concentration signal, or the gas concentration signal. "Efficiency in constant mode" means the ratio of the efficiency of the device (i.e., when the device is subjected to intermittent load) to the efficiency for constant load. As mentioned above, when subjected to intermittent load, the efficiency may decrease over time compared to the efficiency for constant load. The pressure and temperature signals may be pressure / temperature signals from the device, for example, pressure / temperature signals from a reservoir in the device (e.g., a reservoir / pipe configured to collect / transport hydrogen (H2) and oxygen (O2) generated during electrolysis). The pressure or temperature signal may be obtained using any type of known pressure / temperature sensor. The gas concentration may be a signal representing the concentration of an element in the device, for example, the O2 concentration in the H2 reservoir / pipe and / or the H2 concentration in the O2 reservoir / pipe of the electrolyzer. The concentration signal may be obtained using any type of known device / method, for example, optical sensor, mass spectrometry, gas chromatography, etc. As is known, H2 and oxygen O2 obtained by electrolysis in the electrolyzer are stored in hydrogen and oxygen vessels / reservoirs (HV and OV, respectively), where they are captured and separated from the electrolyte. The pressure between the two vessels (i.e., anode circuit and cathode circuit) may be different. During the electrolysis process, a certain percentage (e.g., vol.%) of H2 may migrate to the O2 reservoir over a barrier installed between the vessels (e.g., a gas separator), and vice versa a certain percentage of O2 may migrate to the H2 reservoir. In particular, the O2 vol.% in the H2 reservoir can be connected to an alarm for safety reasons, so that the oxygen concentration does not approach the lower explosive limit of the gas mixture. A similar alarm system can also be set for the pressure and / or temperature of each vessel, warning of potential overpressure / temperature in the reservoir. The gas composition / pressure / temperature in each reservoir may fluctuate depending on the load. Such fluctuations do not allow the device to function in a continuous mode, but force a transient behavior. The device operates in a transient mode, not in a continuous mode. Known control methods for system dynamics are not as fast as load variations.Therefore, the control of such devices is a difficult problem. Especially, devices under intermittent loads are prone to experience the explosion limit of O2 in HV, overpressure, overtemperature, etc. Therefore, it is more important to have accurate and fast predictive models for electrochemical devices exposed to this kind of load.

[0023] "Change in a performance metric of an electrochemical device" refers to the behavior of that metric over time while the device is functioning. A progression in a performance metric of an electrochemical device may represent a decrease in efficiency and / or capacity over time (i.e., between an initial time and a certain time), i.e., degradation. "Predictive model configured to output a predicted change in a performance metric of an electrochemical device" refers to a model that accepts data regarding an electrochemical device as input and outputs a number of values ​​that represent a change in a performance metric of the device.

[0024] As is known in the field of electrical engineering, renewable energy sources often generate loads that operate only for a small portion of a 24-hour period. One or more renewable energy sources that partially generate an intermittent load may consist of (e.g., consist of) such renewable energy sources. An intermittent load may be a load that has an oscillatory behavior (e.g., due to periodic fluctuations of said one or more renewable energy sources) and may disappear during certain periods of time. The concept of an intermittent load is sometimes defined inversely to a continuous load, which is usually defined as a load that operates continuously for a long period of time. In other words, a continuous load is a load that continues for a long period of time under the same conditions. On the other hand, an intermittent load is a load that lasts for a short period of time and then stops either for a long period of time or for a very short period of time. Applications of intermittent loads include electrochemical devices that operate in a "receiving" mode (e.g., an electrolyzer connected to natural energy) or in a "feeding" mode (e.g., a battery connected to a motor with an intermittent load). The intermittent load is at least partially generated by one or more renewable energy sources, such as one or more of a solar energy source, a wind energy source, a geothermal energy source, and a marine energy source.

[0025] The method consists of obtaining a number of time series representative of respective real intermittent loads. The real intermittent loads may be real observed loads. In other words, the intermittent loads are records of the energy production of an electrochemical device, for example a photovoltaic or wind power system. These records are often created by a so-called SCADA (Supervisory Control And Data Acquisition) system. Furthermore, the real intermittent loads may be historical loads, i.e. loads measured over a period of time in the past and stored in a database. In other words, the time series representative of the real intermittent loads are the history of said intermittent loads. Alternatively, the real intermittent loads may be generated loads (i.e. each time series may be artificially generated). For example, the time series may be constructed by generating wind power from a wind profile and characteristics of a wind turbine. Thus, the multiple time series may be composed of real observed load time series and / or historical load time series and / or generated time series. Alternatively, the multiple time series may consist exclusively (i.e., for 100% of the time series) or substantially exclusively (e.g., for 90% or more of the time series) of actually observed loads and / or historical loads.

[0026] Each intermittent load is generated in part by one or more renewable energy sources and is configured for application to an electrochemical device. By a load being "configured for application to an electrochemical device," it is meant that the load is applicable, i.e., is in a physical configuration, to an electrochemical device.

[0027] The method then comprises machine learning a plurality of basis functions representing each underlying intermittent load. The plurality of basis functions provides a finite-dimensional projection space of the plurality of time series. The method may machine learn the plurality of basis functions using known methods to obtain the projection space of the time series. By way of example, the finite-dimensional projection space may be an optimal projection space. By optimal projection space, we mean a projection space in which, for a certain dimensionality of the projection space (i.e., for a given number of basis functions), the projection error on the projection space is minimized. The projection error may follow any known metric. In particular, the optimal projection space may have a dimension that is smaller than that required to accurately define the provided plurality of time series. In this aspect, the optimal projection space provides a dimensionally reduced space for describing the time series (thereby reducing the complexity). Such a dimensionally reduced space is valuable enough to describe the provided time series and the expected intermittent loads with the basis functions sufficiently accurately to obtain an accurate and fast forecasting model. In an example, the number of basis functions may be less than 50, 25, 15. The projection space is a linear projection space whereby the multiple basis functions are constructed to approximate each time series by a respective succession of linear combinations applied to the multiple basis functions.

[0028] In an example, the multiple basis functions may be composed of multiple shapelets. In other words, the method may identify (using machine learning) multiple basis functions to obtain a shapelet decomposition / learning of the intermittent load. Alternatively, the multiple basis functions may be composed of multiple Fourier functions (i.e., providing a Fourier decomposition) or multiple wavelet functions (i.e., providing a wavelet decomposition). Shapelet decomposition provides an extension to Fourier decomposition and / or wavelet decomposition. Just as Fourier decomposition decomposes a signal by its (eigen) frequency, wavelet decomposition decomposes a signal by its frequency and order. Wavelet decomposition identifies the dominant frequencies and further provides information on the localization of the dominant perturbations and tunes the signal. On the other hand, shapelet learning can decompose a signal by frequency, order and pattern. In other words, shapelet learning identifies representative patterns of a signal and directly identifies the patterns. Furthermore, shapelet learning can simultaneously perform a decomposition of the signal based on a classification label. Taking such labels into account when decomposing the signal is particularly important in the application to predicting performance changes of the target equipment of renewable energy sources. Such energy sources exhibit high variability with respect to weather conditions (e.g., wind, clouds, sun, etc.) and / or time of day or year (day / night, winter / summer, etc.). For example, the output profile (and its pattern) of a solar panel on a sunny summer day may be significantly different from its output profile on a winter night or on a rainy day. It is therefore important to consider additional parameters to resolve patterns between similar data of comparable operating conditions of the electrochemical device. Using shapelet learning, the metadata of the provided time series can be used, for example, to improve the machine learning step. At the beginning of the learning process, shapelet learning can start with a portion of the signal at hand. However, as the learning is iterated (e.g., over epochs), a mix or average between said portions of the signal is obtained to have the best compromise.In other words, shapelet learning takes a signal, cuts it into smaller pieces, modifies the smaller pieces to reduce a dynamic time warping (DTW) metric, and keeps the best pieces according to said metric. In some examples, shapelet learning can use a genetic algorithm that uses DTW as the metric of choice. In examples, linear combinations of shapelets are determined for different time intervals of the considered time series. The time intervals are determined by an optimization algorithm, in particular as known in the field of general additive models (GAMs).

[0029] The method then determines a plurality of basis changes for each of the basis functions, each basis change consisting of a respective basis change in a performance index of the electrochemical device when the electrochemical device is exposed to a respective basis intermittent load represented by the basis function. In other words, the method can include a test program in which the electrochemical device is exposed to a plurality of (fundamental) intermittent loads. Examples of such test programs are further described below.

[0030] As mentioned above, the method is a method for generating a predictive model. The (generated) predictive model is a function that takes as input an assumed intermittent load and outputs a predicted change in a performance index of the electrochemical device when subjected to said assumed intermittent load. To that end, the function is constructed (in a first functional block) from a projection of the assumed intermittent load onto a number of basis functions. Such projection results in respective successors of linear combinations that can be applied to the number of basis functions to approximate the assumed intermittent load. The projection function is directly output by the machine learning step. The predictive model is further constructed by applying respective successors of linear combinations (further output by the machine learning step) to a number of basis changes, thereby outputting a predicted change. The respective successors of linear combinations include, for each linear combination and basis function, respective scalar coefficients. The respective scalar coefficients are coefficients applied to the respective basis changes corresponding to the basis functions.

[0031] In an example (e.g., the test program described above), determining the basis changes of each of the performance indicators of the electrochemical device comprises obtaining at least one physical instance of the electrochemical device for each basis function. By "physical instance" is meant an instance of an electrochemical device of a particular class (described above) in the real world. An instance of an electrochemical device (of said class) is defined by a set of specifications that define its behavior in terms of changes in performance when subjected to intermittent loads. The set of specifications may consist of any technical characteristics (e.g., nominal power, voltage response, materials, etc.) that control and define the behavior of the device in function. A physical instance is a real-world piece of hardware that meets all the specifications and forms a working version of the electrochemical device.

[0032] The determining may then include subjecting at least one physical instance of the electrochemical device to at least one actual instance of the respective basis intermittent load represented by the basis functions. The determining may further include performing at least one measurement of a performance index on the at least one physical instance of the electrochemical device and calculating a respective basis change based on the at least one measurement. For example, the basis measurement may be a measurement of maximum capacity or power of the electrochemical device and the basis change may be a basis degradation caused to the physical instance of the device by the actual instance of the respective actual intermittent load.

[0033] In an embodiment, the at least one measurement of the performance indicator of the at least one physical instance of the electrochemical device may include a first measurement and a second measurement. The first measurement may be performed at the beginning of at least one actual instance of the respective basic intermittent load. The second measurement may be performed at the end of at least one actual instance of the respective actual intermittent load. In other words, the method may provide at least two measurements at the beginning and end of each load. The calculation of each basis change may consist of calculating a difference between a value obtained from the second measurement and a value obtained from the first measurement. Such difference may indicate a change (i.e., a variation) in the performance indicator corresponding to the basic intermittent load represented by the basis function.

[0034] In an embodiment, for at least one basis function, at least one actual instance of each fundamental intermittent load represented by the basis function may be composed of multiple actual instances of each fundamental intermittent load. In other words, such an embodiment provides a repetitive pattern of intermittent load to the device. Applying repetitive intermittent loads allows the generated model to predict the evolution of the performance indicators over time. In particular, in various applications and under many load conditions, the behavior of the electrochemical device can be considered linear, i.e., without hysteresis. This improves the method of performing parallel predictions with multiple models instead of performing predictions at separate points in time.

[0035] In an embodiment, each time series may include a respective classification label. And, the machine learning of the multiple basis functions may include, for each time series, minimizing a classification error with respect to the classification label and / or a distance metric between the basis function and the provided time series. Each classification label may represent a binary classification (i.e., having two label values) or a multi-class classification. The distance (or similarity) metric represents a similarity between the time series and the basis function. The classification metric evaluates whether the fitted classification of the time series corresponds to a known classification according to the classification label.

[0036] In particular, when the basis functions consist of multiple shapelets as described above, the resulting shapelet is the one that best represents the classification. In such a case, the disparity metric can be a median value that represents "how well" the shapelet represents the class. This metric is the lens through which each class is viewed by the shapelet. In such a combination, and for classification fitting to time series, the method associates a certain number of shapelets to each class. Then, a metric can be used to evaluate the distance between the input time series and each shapelet of each class. The class that is closest to the time series is then associated with the time series. The goal of shapelet learning is to find the best shapelet that can best fit multiple time series to known classes.

[0037] In an embodiment, the distance metric may be dynamic time warping as known in the art. The classification error may be any classification metric known in the art. For example, the classification error may be a cross-entropy loss, such as an L2-penalized cross-entropy loss.

[0038] As mentioned above, according to some examples, the basis functions are composed of shapelets. In such examples, each shapelet may include a length parameter, and the machine learning of the basis functions may further include determining the length of each shapelet so as to minimize the error. The method may include an optimization step to obtain an optimal length of each shapelet. The optimization process may optimize the length so as to minimize a combination of a distance metric and / or a classification error. The optimization step may follow any known optimization method in the art. In examples, the optimization process may be configured according to a neural network or a genetic algorithm, for example, the implementation of Gendis (Vandewiele et al., "GENDIS: Genetic Discovery of Shapelets", Sensors 2021, 21(4), 1059).

[0039] Alternatively or additionally, the method can accept an input defining the respective length of each shapelet (of the plurality of shapelets). A user can provide the input in any manner, for example based on domain knowledge. Such an example can utilize an implementation such as tslearn (see https: / / tslearn.readthedocs.io / en / stable / ) that machine-learns shapelets using the provided input defining the length of the shapelets.

[0040] In an embodiment, the machine learning of the basis functions may further include selecting a subset of a predetermined size from the determined basis functions. Such an embodiment may be combined with the embodiment of the method described above, in which the machine learning of the basis functions comprises minimizing a classification error and / or a distance metric. In such a combination, the selected subset may have a classification error of a first neighborhood that minimizes the classification error and / or a distance metric of a second neighborhood that minimizes the distance metric.

[0041] Such an example concerns a clustering algorithm that regroups the extracted shapelets (as determined basis functions) into selected shapelets that best represent the shapelet set while retaining classification power. In particular, due to computational constraints, the shapelet learning process can only test a certain number of shapelets of a certain maximum length. This length can be determined using domain knowledge or using optimization, as described above. A larger number is set to be extracted in order to give the algorithm sufficient search flexibility to obtain the (desired) number of shapelets. The shapelets can then be made equal in length by padding with tail values. The resulting shapelets may then be clustered to the desired number. The remaining tail plateaus are then cut off. Finally, an inertia criterion is evaluated for the shapelets before and after clustering to verify that the clustering preserves the accuracy of the classifier.

[0042] Further proposed is a computer-implemented method of using the generated predictive model for each electrochemical device, where the model is generated according to the above-mentioned method. Such a method of use comprises obtaining an assumed intermittent load generated at least in part by one or more renewable energy sources. The method of use further comprises applying the predictive model to the assumed intermittent load. Thus, the method of use outputs a predicted change in a performance indicator of each electrochemical device when the respective electrochemical device is exposed to the assumed intermittent load. The application of the predictive model comprises projecting the assumed intermittent load onto a plurality of basis functions, thereby obtaining respective successors of linear combinations that can be applied to the plurality of basis functions to approximate the assumed intermittent load. The application of the predictive model further comprises applying the respective successors of the linear combinations to the plurality of basis changes, thereby outputting the predicted changes.

[0043] The above-mentioned methods, i.e., generating a predictive model, using the generated predictive model, and forming a dataset, are computer-implemented. This means that the steps of the method (or substantially all steps) are performed by at least one computer, or any system. Thus, the steps of the method are performed by a computer, possibly fully automatically or semi-automatically. In an embodiment, the triggering of at least some of the steps of the method is performed by a user-computer interaction. The level of user-computer interaction required depends on the balance between the foreseen level of automation and the need to realize the user's wishes. In an embodiment, this level may be user-defined and / or predefined. For example, in the method of generating a predictive model, the step of acquiring a plurality of time series may be performed upon input from a user. Upon receiving such input, the method may automatically perform the remaining steps and generate a predictive model. As another example, in the method of using, the step of acquiring a predicted intermittent load may be performed upon receiving input by a user. Upon receiving such input, the method automatically applies the predictive model to the assumed intermittent load and outputs a predicted evolution of the performance index for each electrochemical device.

[0044] A typical example of a computer implementation of the method is executing the method on a system adapted for this purpose. The system may include a processor coupled to a memory and a graphical user interface (GUI), the memory having recorded thereon a computer program comprising instructions for carrying out the method. The memory may also store a database. The memory is any hardware adapted for such storage, and may possibly consist of several physically distinct parts (e.g. one for the program and possibly one for the database).

[0045] FIG. 1 shows an example of a system, which is a client computer system, such as a user's workstation.

[0046] The client computer of this embodiment comprises a central processing unit (CPU) 1010 connected to an internal communication BUS 1000, and a random access memory (RAM) 1070 also connected to the BUS. The client computer further comprises a graphics processing unit (GPU) 1110 associated with a video random access memory 1100 connected to the BUS. The video RAM 1100 is also known in the art as a frame buffer. A mass storage controller 1020 manages access to mass storage devices such as a hard drive 1030. Mass storage devices suitable for embodying computer program instructions and data include, by way of example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks and removable disks; and all forms of non-volatile memory including magneto-optical disks. Any of the above may be supplemented by or incorporated in specially designed ASICs (application specific integrated circuits). A network adapter 1050 manages access to a network 1060. The client computer may also include a haptic device 1090 such as a cursor control device, keyboard, etc. A cursor control device is used in the client computer to allow a user to selectively place a cursor at any desired location on the display 1080. In addition, the cursor control device allows a user to select various commands and input control signals. The cursor control device includes a number of signal generating devices for inputting control signals into the system. Typically, the cursor control device is a mouse, and the buttons of the mouse are used to generate the signals. Alternatively or additionally, the client computer system may include a sensitive pad, and / or a sensitive screen.

[0047] A computer program may include instructions executable by a computer, the instructions including means for causing said system to carry out said method. The program may be recordable on any data storage medium, including the memory of the system. The program may be implemented, for example, in digital electronic circuitry, computer hardware, firmware, software, or a combination thereof. The program may be implemented as an apparatus, for example an article embodied in a machine-readable storage device for execution by a programmable processor. The method steps may be performed by a programmable processor executing a program of instructions that performs the functions of the method by manipulating input data to generate output. Thus, the processor may be programmably coupled to receive data and instructions from, and to transmit data and instructions to, a data storage system, at least one input device, and at least one output device. The application program may be implemented in a high level procedural or object oriented programming language, or in assembly or machine language as appropriate. In either case the language may be a compiled or interpreted language. The program may be a full installation program or an update program. Application of the program on the system in any case provides instructions for carrying out the method. The computer program may alternatively be stored and executed on a server in a cloud computing environment, the server being in communication with one or more clients over a network, where a processing unit executes instructions constituted by the program to thereby perform a method on the cloud computing environment.

[0048] Next, an implementation of the above method will be described.

[0049] The present embodiment relates to a methodology for predicting degradation of devices connected to upstream / downstream intermittent loads. The degradation of electrochemical systems such as electrolyzers is governed by several factors, including the voltage of the electrodes that trigger the degradation, the generation of thermal gradients due to start / stop and load changes, and the diffusion of species (e.g. gas crossover) due to start / stop and load changes. The implementation is able to capture both phenomena. The implementation imposes a cell (of the electrochemical device) to stay at a given voltage (as defined by the amplitude of the shapelets) for a given time while reflecting the load changes.

[0050] In particular, since renewable energy sources generally have periodic energy production patterns (e.g., in the case of solar-driven renewable energy resources, as a response to periodic inputs from the sun), it is crucial to be able to capture these periodic patterns as characteristics of the energy production at different scales. The present embodiment deals with a methodology consisting of identifying said patterns and characterizing and associating to each pattern an impairment. A forecasting model then predicts the temporal shape of the renewable energy production, which is the shape projected onto the identified patterns. Projection coefficients are used to combine the different impairments that characterize the patterns, giving an estimate of the expected impairment going forward.

[0051] Shapelet learning is a generalization of Fourier and wavelet decomposition to any shape. Shapelet learning is a machine learning approach for supervised classification. To implement this, the implementation uses "LearningShapelets" from the python package tslearn version 0.5.2 (see https: / / tslearn.readthedocs.io / en / stable / index.html).

[0052] In a first step, the implementation collects (i.e., obtains) a large number of data sets of intermittent loads (e.g., 10 or more sets of data) and uses machine learning / artificial intelligence algorithms to generate a base of unit patterns onto which these functions can be projected (s iEach unit pattern is itself a load profile with duration ranging from a few milliseconds to a few hours.

[0053] In a second step, the implementation runs a test program. The test program has as many test phases as the number of unit patterns obtained in the first phase. Each phase of the test program consists of applying a load corresponding to one unit pattern several times. The typical duration of each test is between 1 hour and 10000 hours. The test period can last up to a year. At the end of each phase, the degradation of the device is monitored by measuring its efficiency compared to the constant mode. In this embodiment, the degradation of the unit patterns (s) is calculated by dividing the observed degradation by the number of unit patterns in that phase. i Degradation per unit (d i ) to calculate

[0054] To speed up the testing, implementations perform multiple parallel tests on separate equipment. Such a parallel strategy is based on the assumption that the degradation of the system is not affected by the previous lifetime (i.e., there is no hysteresis). In particular for electrolysers, this is a valid assumption since a linear degradation over time is observed. Since failures introduce a change in the slope of the degradation, cells / stacks are replaced well before failure occurs.

[0055] FIG. 2A outlines the first and second steps. In step 1, the implementation collects a large dataset by accessing a library 201 of intermittent loads. Then, the implementation uses machine learning and artificial intelligence algorithms to identify the best family of eigenvectors that describe the intermittent profile obtained by the method. The implementation can use different techniques such as Fourier transform, wavelets, or shapelets to identify such eigenvectors. As a result of this step, the embodiment produces a set of identified eigenvectors (or patterns) that can be used to describe the intermittent loads. iThis set is made up of 10 eigenvectors. Next, the embodiment applies each specific pattern of the set to test the device (e.g., electrolytic cell) and obtains the specific degradation d caused by each pattern. i Determine.

[0056] In the third step, the degradation of the connected equipment can be predicted with a new intermittent load profile S(t). The new load is approximated by a linear combination of the unitary patterns generated in step 1 as follows: P(t)=S i=1.n a i s i (t)

[0057] Then, by summing the individual degradations of the unitary patterns used and weighting them by the coefficients used to transform the new weights into a linear combination of the unitary patterns, the degradation (d) can be predicted as follows: d = S i=1..n a i d i

[0058] FIG. 2B outlines the third step. In this example, a new profile P(t) is provided to the method. The profile may relate to daily, monthly or yearly power / load. The load is normalized. The implementation computes the spectral transform of the provided profile, e.g. using Fourier, wavelet or shapelet, and then predicts the degradation as described above. An embodiment may estimate the confidence level by calculating the distance between the original (i.e., provided profile) profile and the reconstructed profile (i.e. using the computed spectral transform).

[0059] This embodiment can be used in a similar manner to predict degradation of other types of devices, such as electrolyzer families connected to intermittent power sources (e.g., renewable energy, hybrid systems, etc.), electrochemical energy storage devices connected to intermittent power sources, and / or electrochemical energy storage devices connected to intermittent loads (e.g., electric motors, etc.). The embodiment can also predict other output signals, such as temperature, pressure, ion concentration, and / or gas, in a similar manner.

[0060] Next, we discuss the metrology aspects of the implementation.

[0061] As is well known, shapelet extraction is a method for extracting characteristic patterns from a collection of signals, in other words, identifying the basic shapes of signals by learning shapelets.

[0062] Figure 3 shows the base signals used to generate the training and test data. Figures 4A-B show the time series of the training and test sets, respectively. Two-thirds of the dataset is used as the training set and the remaining one-third is used as the test set.

[0063] The shapes identified by the implementation at two moments of the process are evaluated to illustrate the identification process. The implementation is performed by tslearn, a Python package that provides machine learning tools for time series analysis. The tslearn package implements a neural network that modifies its weights after each pass through the entire dataset, called an epoch, to optimize its performance. The two moments of evaluation are at different numbers of epochs, one before convergence (e.g. 3000 epochs) and the other when convergence is considered to have been achieved (e.g. 10000 epochs).

[0064] The first moment, i.e., for 3000 epochs, is shown in Figure 5. Looking at the fitness curve 510, we can see that convergence has not yet been reached. Moreover, the extracted (i.e., identified) shapelets 520 are a mixture of the base signals used to generate the training and test sets, as shown in Figure 3.

[0065] The second moment, i.e., 10000 epochs, is shown in Fig. 6. The fitness curve 610 is lower and flatter compared to 510 in Fig. 5, indicating that convergence has been achieved. Moreover, the extracted shapelets 620 almost show the pure base signals used to generate the training and test days as shown in Fig. 3.

[0066] Comparing the accuracy for both 3000 and 10000 epochs of training, there is an increase of 0.6% from the first case (Figure 5) to the second case (Figure 6). Note that accuracy measures the classification performance of the shapelets, not their closeness to the base signal. The accuracy of the shapelets' shapes is not necessary for the shapelet learning classifier, but rather depends on their difference. However, the accuracy of the identified shapes improves the classification performance.

[0067] These observations indicate that the implementation converges to the correct shapelets, in other words, the learning shapelets algorithm does indeed identify and extract the correct underlying patterns from a set of signals.

[0068] 7A-C show examples of extracted shapelets: FIG. 7A is an identified sinusoidal signal, FIG. 7B is an identified triangular signal, and FIG. 7C is an identified square signal.

[0069] In this implementation, datasets representing renewable energy production in different locations, under different conditions are collected and labeled by macroscopic conditions (day / night, winter / summer, solar / wind, climate...). For these collected datasets or examples, machine learning methods can calculate the distance or similarity between subsections or shapelets of this example, using a metric called "dynamic time warping". The length (or duration) of these shapelets is either determined by optimization techniques or by domain knowledge. The group of shapelets that are closest to all other possible subsections and that best represent the classification given by the user is kept or extracted. The classification is done by logistic regression built as a neural network, with weights optimized by stochastic gradient descent for L2-penalized cross-entropy loss.

[0070] Because experimentation with electrolytic cells requires significant time and resources and is subject to operational constraints, implementations may use approaches that do not explore all possible shapelet lengths. In such an approach, embodiments may use a clustering algorithm to regroup the "extracted shapelets" into k "selected shapelets" that best represent the shapelets. The clustering algorithm may select the k selected shapelets to retain either x% of the classification power from logistic regression or x% of the total distance between the shapelets. The value x may be determined by experimental design constraints that characterize the degradation associated with each "selected shapelet."

[0071] The experimental design can be optionally simplified to two concomitant experiments: one is a baseline scenario that serves as a reference point against which the second consists of N repetitions of each "selected shapelet", where N can be determined by the central limit theorem such that the characterized degradation is statically representative (typically 30 or more if a Gaussian distribution is assumed).

Claims

1. 1. A computer-implemented method for generating a predictive model configured to output a predicted change in a performance indicator of an electrochemical device subjected to an assumed intermittent load, the electrochemical device being at least in part generated by one or more renewable energy sources, the method comprising: - obtaining a plurality of time series, each of the plurality of time series representing a respective real intermittent load generated at least in part by one or more renewable energy sources and configured to load an electrochemical device; - machine learning a plurality of basis functions, each representative of a respective underlying intermittent load, the plurality of basis functions forming a projection space of a plurality of time series, the plurality of basis functions being configured thereby to approximate each of the time series by a succession of respective linear combinations applied to the plurality of basis functions; and determining, for each basis function, a plurality of basis changes, including each basis change of a performance index of the electrochemical device, when the electrochemical device is subjected to each basis intermittent load represented by the basis function; wherein the predictive model comprises: - projecting the assumed intermittent load onto the plurality of basis functions, thereby obtaining a succession of respective linear combinations that can be applied to the plurality of basis functions to approximate the assumed intermittent load; - applying each successive linear combination to the plurality of basis changes, thereby outputting a predicted change.

2. Determining a basis change in each of the performance indicators of the electrochemical device for each basis function - obtaining at least one physical instance of said electrochemical device; - subjecting at least one physical instance of said electrochemical device to at least one actual instance of each elementary intermittent load represented by a basis function; - performing at least one measurement of a performance index for at least one physical instance of said electrochemical device; and The method of claim 1, further comprising calculating said respective underlying changes based on at least one measured value.

3. 3. The method of claim 2, wherein the at least one measurement of a performance indicator of at least one physical instance of the electrochemical device comprises a first measurement and a second measurement, the first measurement being performed at a start of the at least one actual instance of the respective basal intermittent load and the second measurement being performed at an end of the at least one actual instance of the respective basal intermittent load, and the calculation of the respective basal change comprises calculating a difference between a value obtained from the second measurement and a value obtained from the first measurement.

4. 4. The method of claim 2 or 3, wherein for at least one basis function, the at least one actual instance of the respective fundamental intermittent load represented by the basis function includes a plurality of actual instances of the respective fundamental intermittent load.

5. Each time series contains its own classification label, and machine learning of multiple basis functions is performed for each time series. - the classification error for said classification label, and / or - a distance metric between the basis functions and the provided time series The method of any one of claims 1 to 4, comprising minimizing

6. - the distance metric is dynamic time warping; and / or - the classification error is the cross-entropy loss, The method according to claim 5.

7. The method of any one of claims 1 to 6, wherein the plurality of basis functions consists of a plurality of shapelets.

8. Each shapelet includes a length parameter, and the machine learning of the plurality of basis functions further comprises: determining the length of each shapelet to minimize the error, The method according to claim 7.

9. The machine learning of the plurality of basis functions further comprises: - selecting a subset of a predetermined size from said determined plurality of basis functions, The selected subset is the classification error in a first neighborhood of said minimized classification error, and / or - having a distance metric of a second neighbour of said minimising distance metric, 9. A method according to claim 7 or 8 and further dependent on claim 5.

10. The performance index is - efficiency in constant mode, - temperature signal, - pressure signal, - ion concentration signal, or represents the gas concentration signal, 10. A method according to any one of the preceding claims.

11. The electrochemical device comprises: - an electrolytic cell connected to an intermittent power supply; - an electrochemical energy storage device connected to an intermittent power source; and - Electrochemical energy storage device connected to an intermittent load 13. The method of any one of the preceding claims, comprising at least one of:

12. A computer-implemented method of using a predictive model generated by the method of any one of claims 1 to 11 for a respective electrochemical device, comprising the steps of: The method comprises: - obtaining a notional intermittent load, at least a portion of which is generated by one or more renewable energy sources; - applying said predictive model to an assumed intermittent load, thereby outputting a predicted change in a performance index of each electrochemical device when the respective electrochemical device is subjected to the assumed intermittent load; applying the predictive model o projecting the assumed intermittent load onto a plurality of basis functions, thereby obtaining respective successors of linear combinations that can be applied to the plurality of basis functions to approximate the assumed intermittent load; and applying each successive linear combination to a plurality of underlying transformations, thereby outputting a predicted transformation; A method comprising:

13. A computer program comprising instructions for carrying out the method according to any one of claims 1 to 11 and / or the method according to claim 12.

14. A computer-readable recording medium having the computer program according to claim 13 recorded thereon.

15. 14. A system including a processor coupled to a memory, the memory having the computer program of claim 13 recorded thereon.

Images