Method for training a machine learning model which is designed to determine the charging time for batteries of electric vehicles
A hybrid machine learning model trained on fleet and vehicle-specific data predicts electric vehicle battery charging times accurately, enhancing efficiency and reducing resource consumption.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2025-12-01
- Publication Date
- 2026-06-25
AI Technical Summary
Conventional methods for estimating the charging time of electric vehicle batteries are imprecise, leading to inefficiencies such as increased downtime and unnecessary resource consumption, and existing approaches face challenges in accurately predicting charging times due to the complexity of chemical processes within batteries and the need for vehicle-specific data.
A hybrid machine learning model is trained using initial data from a fleet of vehicles, further refined through federated learning on individual vehicles, combining cloud data with vehicle-specific data to create specialized models that accurately predict charging times.
The method provides precise charging time predictions, optimizing charging processes and reducing resource usage by allowing drivers to plan breaks and select appropriate charging stations, while conserving battery life through accurate current management.
Smart Images
Figure EP2025084876_25062026_PF_FP_ABST
Abstract
Description
[0001] R.414687
[0002] - 1 -
[0003] Description
[0004] title
[0005] Method for training a machine learning model intended for determining the charging time of electric vehicle batteries
[0006] The present invention relates to a method for training a machine learning model intended for determining the charging time of electric vehicle batteries, a method for determining the charging time of an electric vehicle battery using such a machine learning model, as well as a computing unit and a computer program for carrying it out.
[0007] Background of the invention
[0008] Modern vehicles can be electrically powered and use a battery as an energy storage device. These can be purely electric vehicles.
[0009] Disclosure of the invention
[0010] According to the invention, a method for training a machine learning model, a method for determining the charging time of an electric vehicle battery, as well as a computing unit and a computer program for carrying them out, with the features of the independent claims, are proposed. Advantageous embodiments are the subject of the dependent claims and the following description.
[0011] The invention generally deals with determining the charging time of an electric vehicle battery. Charging time can be understood in particular as the duration until the battery, starting from its current state of charge, is fully charged (i.e., to 100%) or, for example, until it reaches another R.414687
[0012] - 2 - a predetermined value, e.g., 90% or 80% charged. It is also conceivable that several different such values are determined. Furthermore, charging time can be understood not only as the duration, but also as the end time of the charging process (when fully charged or up to a predetermined value); this can be calculated accordingly based on the duration and the charging start time.
[0013] For the purposes of this text, an electric vehicle is understood to be primarily a purely electrically powered vehicle. However, it can also refer to a so-called hybrid vehicle, which is or can be electrically powered and has a battery for this purpose. A battery is understood to be specifically an electrical energy storage device, such as those typically used in electrically powered vehicles.
[0014] To make charging an electric vehicle as convenient as possible for the driver, an accurate estimate of the charging time at the start of the charging process is crucial. Unfortunately, conventional approaches are imprecise; this is the case, for example, when simple heuristic chemical-physical models or further simplified electrical-thermal models are used. A particularly accurate estimate of the charging time, such as that made possible by the proposed method, allows the driver to save costs, such as downtime charges, plan breaks more precisely, and helps in selecting the right charging power supply. Furthermore, it might even be possible to reduce the number of charging stations if the more accurate prediction allows the driver to leave the charging station sooner.
[0015] Furthermore, it can be useful if the driver can specify a target time at which the vehicle should have reached a certain charge level. An accurate estimate of the charging time under given environmental conditions allows for an optimized (especially lower – since less tolerance needs to be assumed in the calculation) charging current, which conserves battery life. R.414687
[0016] - 3 -
[0017] A purely physical model is difficult to implement because the chemical processes within a battery are highly complex over many charging cycles and are not yet fully understood. Furthermore, the computational demands on the vehicle's control unit (the electric vehicle's control unit) increase with the model's complexity. Accurately estimating charging time depends on many different factors, such as the charging current, voltage, battery condition (e.g., age), state of charge, temperature, etc. Some of these factors are vehicle-specific, such as the battery's age or condition. This makes a purely data-driven approach challenging (it would require data from batteries in many different states, charged under a wide variety of conditions). While a purely adaptive approach (learning from past charging cycles) could theoretically offer an improvement, the infrequent charging cycles would pose a problem for adequately training the model.
[0018] Against this background, an approach is proposed that uses a combination of knowledge from various sources, such as the cloud and the vehicle. Information about the battery's state from the cloud ("Battery In The Cloud") could, in itself, only solve some of the problems. However, the algorithm presented here addresses many of these issues, thereby significantly improving the accuracy of the charging time prediction.
[0019] A (particularly computer-implemented) method for training a machine learning model is proposed for determining the charging time of electric vehicle batteries. This method utilizes training data and input variables (for the machine learning model). The training data comprises one or more sets of values for each battery and electric vehicle, with each set of values including: values of quantities characteristic of the battery and / or the electric vehicle, and a charging time value for each of these characteristic quantities. Such characteristic quantities include, in particular, the current during battery charging, the voltage during battery charging, the maximum voltage during battery charging, the maximum current during battery charging, the maximum charging power during battery charging, and R.414687.
[0020] - 4 - Current battery charge status, current battery health status, current battery temperature, current ambient temperature, target battery charge status, current power consumption of the electric vehicle's auxiliary consumers, expected power consumption of the electric vehicle's auxiliary consumers, and a time-dependent charging power profile during battery charging. The training data is primarily, at least partially, collected from or measured on vehicles. The input variables include values characteristic of the battery and / or the electric vehicle.
[0021] Initial training data from the batteries of a large number of electric vehicles is now being made available. This initial training data can be collected from an entire fleet of vehicles, potentially distributed worldwide. Based on this data, a machine learning model is then trained so that it can output a charging time for the electric vehicle's battery based on the input values. This initial training process results in a pre-trained machine learning model. This initial training can take place on a computing unit such as a central server, for example, in the cloud. The pre-trained machine learning model is then deployed.
[0022] Furthermore, characteristic data (referred to here as model data for differentiation) are provided for trained individual machine learning models. These trained individual machine learning models are each obtained by a second training of the pre-trained machine learning model, based on their respective second training data. This allows each machine learning model to output a value for the charging time of the electric vehicle's battery based on input values. The second training data is obtained for each of a second set of electric vehicles. Specifically, the centrally pre-trained machine learning model can be distributed across electric vehicles. There, the pre-trained machine learning model can then be further trained on a vehicle-specific basis. This training can take place in a separate memory area and, at this stage, has no influence on the vehicle ("shadow mode") or its behavior.The resulting R.414687.
[0023] - 5 - trained individual machine learning models, or other characteristic model data such as model weights and / or learned gradients, can then be transmitted back to the central server or the like.
[0024] Furthermore, additional data can be transferred that characterizes the individual vehicle more precisely; here (for distinction), this is also referred to as vehicle-specific data or simply vehicle data. This could include, for example, histograms or clusters (e.g., centroids) of the input data from the respective second training data set. This enables the analysis and targeted provision of models. A histogram of temperature, for example, could show that a vehicle was frequently driven in high temperatures. Thus, it would be possible to obtain two specialized models, one for high temperatures and one for low temperatures. The pre-trained machine learning model is then adjusted based on the additional data provided (the vehicle-specific data, e.g., vehicle-specific data).The histograms) and the model data characteristic of the trained individual machine learning models are used to obtain one or more trained final machine learning models. The characteristic model data and / or vehicle data can be weighted accordingly, meaning the different trained individual machine learning models are weighted differently in the one or more trained final machine learning models. The one or more trained final machine learning models are then made available, particularly for use in electric vehicles.
[0025] In one embodiment, it is further provided that one or at least one of the several trained final machine learning models is validated. The one or at least one validated trained final machine learning model is then made available for use in one or more electric vehicles.
[0026] In one embodiment, one or each of the several trained final machine learning models is used for a group of electric vehicles R.414687
[0027] - 6 - obtained, where each group of electric vehicles is defined by one or more usage parameters. Such a usage parameter could be, for example, an average temperature (this could be, for example, the ambient or battery temperature) during battery charging. However, other parameters characteristic of the battery and / or the electric vehicle, which are listed herein, are also relevant. For example, if someone always charges their vehicle in a similarly temperature-controlled garage, an optimized model could be provided for this target group, which still meets certain minimum standards for other characteristic parameters.
[0028] In summary, a merger of the various approaches is proposed. This involves not only using a hybrid approach combining a chemical-physical and a data-driven model, but also incorporating additional information from sources such as the cloud. To enable the model to adapt to the vehicle, federated learning is employed. This not only improves the model's accuracy but also solves the problem of accessing personal and / or vehicle-related data, as the model and its parameters are transmitted, not the personal or vehicle-related data itself.
[0029] A trained final machine learning model obtained in this way can then be used to determine the charging time of an electric vehicle's battery. Values of parameters characteristic of the battery and / or the electric vehicle are provided as inputs for the machine learning model. A charging time value is then determined as an output of the machine learning model, based on the values of the inputs. This determined charging time value is then made available. This determination of the charging time can then be performed, for example, on a processing unit of the electric vehicle.
[0030] A computing unit according to the invention, e.g. a server or other central computing system, or a computing unit of a vehicle, is, in particular R.414687
[0031] - 7 - programmed to carry out a method according to the invention.
[0032] Implementing a method according to the invention in the form of a computer program or computer program product with program code for carrying out all method steps is also advantageous, as this incurs particularly low costs, especially if an executing control unit is already available for other tasks. Finally, a machine-readable storage medium is provided with a computer program stored on it as described above. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical, and electrical storage media, such as hard drives, flash memory, EEPROMs, DVDs, etc. Downloading a program via computer networks (Internet, intranet, etc.) is also possible. Such a download can be wired or wireless (e.g., via a WLAN network, a 3G, 4G, 5G, or 6G connection, etc.).
[0033] Further advantages and embodiments of the invention will become apparent from the description and the accompanying drawing.
[0034] The invention is schematically illustrated in the drawing using exemplary embodiments and is described below with reference to the drawing.
[0035] Brief description of the drawings
[0036] Figure 1 schematically shows an arrangement to illustrate the invention.
[0037] Figure 2 schematically shows a process of a method according to the invention in one embodiment.
[0038] Embodiment(s) of the invention R.414687
[0039] - 8 -
[0040] Figure 1 schematically shows an arrangement to illustrate the invention. Three electric vehicles 100a, 100b, 100c are shown as examples. Each electric vehicle has a battery 102a, 102b, 102c and a computing unit 104a, 104b, 104c. A central server 110 is also shown, which can, for example, represent the so-called cloud. Each electric vehicle 100a, 100b, 100c, or rather its respective computing unit, can exchange data with the central server 110, for example, via an integrated or additional radio module.
[0041] The three electric vehicles shown as examples, 100a, 100b, and 100c, are intended to represent, for example, electric vehicles used in different ways, such as those that are charged at different speeds or with different charging currents (e.g., direct or alternating current), or those that are charged with different frequencies or at different temperatures. In particular, the aforementioned parameters characteristic of the battery and / or the electric vehicle can vary depending on the electric vehicle and thus its battery.
[0042] Figure 2 schematically illustrates the sequence of a process in one embodiment. This will be explained below with reference to Figure 1.
[0043] One goal is to have a trained final machine learning model available for each electric vehicle—for example, a machine learning algorithm such as an artificial neural network—that can determine the charging time for that specific electric vehicle, i.e., when the battery will be fully charged and / or, for example, 80% charged. This estimation, i.e., determining the charging time, can be repeated shortly before charging as well as during the charging process. Various characteristic quantities (variables), as explained in detail above, serve as inputs to the model, i.e., the machine learning model. Some of these quantities may require physical preprocessing (hybrid model) before being used in the machine learning model. Optionally, it is also possible to integrate quantities from the cloud as inputs to the model (e.g., the battery state).
[0044] However, the machine learning model must be trained before use. R.414687
[0045] - 9 -
[0046] In step 200, initial training data 202 from batteries of a first large number of electric vehicles are provided. This initial training data includes values of the parameters characteristic of the battery and / or the electric vehicle, each with a corresponding charging time value for these characteristic parameters.
[0047] In step 204, a machine learning model 206 is adapted or trained in an initial training phase based on the first training data. This results in a pre-trained machine learning model 208, which is made available in step 210.
[0048] The resulting model, i.e., the pre-trained machine learning model 208, can now make generally acceptable predictions. However, there are two difficulties that diminish its performance. First, the machine learning model's learning algorithm attempts to predict the average of the training data (so-called "Empirical Risk Minimization" using, for example, "Mean Square Error"). This prevents specialization for an individual electric vehicle. Second, the training data typically does not cover the complete probability distribution of the population, as it is difficult to collect training data from many different electric vehicles or batteries.
[0049] To counteract these difficulties, the proposed algorithm uses a technique known as "federated learning." In this process, the machine learning model, i.e., the pre-trained machine learning model 208, is further trained ("fine-tuned") online in the control unit or computing unit of the individual electric vehicle.
[0050] In step 212, the pre-trained machine learning model 208 is transferred to each of a second set of electric vehicles. This second set of electric vehicles can be the same as the first set, but it can also be fewer, more, or different electric vehicles. R.414687
[0051] - 10 -
[0052] The electric vehicles themselves can (but do not necessarily have to) store corresponding secondary training data 214, which are then vehicle-specific. This data can also be generated gradually. For example, the required secondary training data can be collected in advance in the control unit over several charging cycles (e.g., regarding model input values at the start of charging and the charging time required until the battery reaches the desired charge capacity of, for example, 80% or 100%).
[0053] A special filtering process can be implemented to ensure that the collected data provides sufficient population coverage; for example, the operating point (remaining capacity) should differ between load cycles. For filtering, a counter could, for example, count how much training data already exists for a given load state, e.g., [0-50%: 5; 51-99%: 30]. To prevent the limited memory from being unevenly distributed and the stored data from becoming outdated, newer data could overwrite older data (ring buffer – once all ring buffers have been completely filled / processed at least once, training can begin). Once the relevant criteria are met and a sufficient number of secondary training data points are available, in step 216, the pre-trained machine learning model can be further trained, for example, for a previously specified number of training cycles ("epochs") using a chosen optimization algorithm (e.g., BFGS).It is also possible to optionally change only certain layers of the machine learning model, while the other layers remain unchanged.
[0054] The resulting specialized model, a trained individual machine learning model 218 for each electric vehicle, is not used directly in the electric vehicle, as it cannot usually be sufficiently verified and model overfitting would be possible due to the small amount of training data. Instead, in step 220, for example, the new model weights or the learned gradients—or, more generally, model data characteristic of the trained individual machine learning models—are sent to the original model in the cloud in step 222. R.414687
[0055] - 11 -
[0056] In step 224, characteristic model data 226 for trained individual machine learning models are provided. Additionally, characteristic vehicle data 230 for the individual vehicle, e.g., selected histograms, are included with the second training data and provided in step 228. Since the model parameters are not interpretable, this also constitutes a form of anonymization. The specialized models for many different vehicles are thus collected in the cloud. These models can now be further analyzed. For example, the models can provide valuable information about different vehicle types, vehicle markets, etc.
[0057] In the cloud, all specialized models can now be combined (e.g., by weighted averaging). The pre-trained machine learning model is thus, in step 232, adapted based on the data characteristic of the trained individual machine learning models in order to obtain one or more trained final machine learning models 234, which are then made available in step 236.
[0058] The resulting model can then undergo an automated validation process (step 238) before being deployed (step 240) and transferred back into the electric vehicle control units. This incremental process ensures continuous model improvement, benefiting even infrequently used electric vehicles.
[0059] The weights used for averaging when combining individual models can be derived, for example, from the analysis of the individual models and their selected histograms. This can incorporate knowledge about the entire fleet. This helps counteract potential bias caused by distorted data transmission (e.g., due to insufficient mobile coverage). For instance, data from vehicles that are regularly charged at unusually high temperatures, but which frequently transmit data, can be weighted less heavily in the averaging.
[0060] Furthermore, it is possible to offer specialized models for specific vehicle markets or even for vehicle users with specific charging behavior - R.414687
[0061] - 12 - General, i.e., trained, final machine learning models, each applicable to a specific group of electric vehicles. One example is the average temperature during charging. Vehicles primarily charged in garages have a different average temperature than those charged exclusively outdoors. By combining all model weights of vehicles with a specific temperature profile (all other models having a weight of 0 or very small), a model is obtained that delivers particularly accurate results in these temperature ranges.
[0062] In this way, there is a final trained machine learning model that applies to an entire group of vehicles and can then be used to determine the charging time, as mentioned at the beginning.
Claims
R.414687 - 13 - Claims 1. Method for training a machine learning model intended to determine a charging time of batteries (104a, 104b, 104c) of electric vehicles (100a, 100b, 100c) using training data and input variables, wherein the training data for each battery and electric vehicle comprise one or more sets of values, each set of values comprising: Values of parameters characteristic of the battery and / or the electric vehicle, - each a value of a charging time for the values of the quantities characteristic of the battery and / or the electric vehicle, and wherein the input quantities include values of the quantities characteristic of the battery and / or the electric vehicle, the method comprising: Providing (200) initial training data (202) from batteries of an initial variety of electric vehicles; initial training (204) of a machine learning model (206) based on the initial training data, such that the machine learning model, based on input values, outputs a value of the charging time of the battery of the electric vehicle in order to obtain a pre-trained machine learning model (208); Provision (210) of the pre-trained machine learning model (208); Providing (224) characteristic model data (226) for trained individual machine learning models, wherein the trained individual machine learning models were each obtained by second training (216) of the pre-trained machine learning model (208) based on respective second training data (214), such that the respective machine learning model, based on input values, outputs a value of the charging time of the electric vehicle's battery, wherein the R.414687 - 14 - second training data (214) have been obtained for each of a second set of electric vehicles; Providing (228) vehicle data characteristic of the individual vehicle (230) of the respective second training data; Adapting (232) the pre-trained machine learning model, based on the vehicle data and based on the model data characteristic of the trained individual machine learning models, to obtain one or more trained final machine learning models (234); and Providing (236) one or more of the trained final machine learning models.
2. The method of claim 1, wherein the model data characteristic of the trained individual machine learning models each comprise model weights and / or learned gradients.
3. Method according to claim 2, wherein in one or at least one of the several trained final machine learning models the characteristic model data of at least a part of the trained individual machine learning models and / or the vehicle data characteristic of the individual vehicle from at least a part of the second training data are taken into account in a weighted manner.
4. A method according to any of the foregoing claims, further comprising: Validating (238) one or at least one of the several trained final machine learning models; and Providing (240) one or at least one validated trained final machine learning model for use in one or more electric vehicles.
5. Method according to one of the preceding claims, wherein one or each of the several trained final machine learning models is obtained for a group of electric vehicles, wherein a group of electric vehicles is defined by one or more usage parameters. R.414687 - 15 - 6. A method for determining the charging time of an electric vehicle battery, using a machine learning model trained according to a method according to any of the preceding claims and obtained as a trained final machine learning model, comprising: Providing values of quantities characteristic of the battery and / or the electric vehicle as input variables for the machine learning model; Determining a value for the loading time as an output variable of the machine learning model, based on the values of the input variables; and Providing the value of the loading time.
7. Method according to any of the preceding claims, wherein the dimensions characteristic of the battery and / or the electric vehicle include at least one of the following dimensions: A current during battery charging, A current voltage during battery charging, A maximum voltage when charging the battery, a maximum current when charging the battery, a maximum charging power when charging the battery, a current charge status of the battery, A current battery health status, A current battery temperature, A current ambient temperature, A target battery charge status, Current performance of auxiliary consumers of the electric vehicle, An expected performance of the auxiliary consumers of the electric vehicle, A time-dependent curve of charging power during battery charging.
8. Computing unit or computing unit group configured to perform all process steps of a process according to any of the preceding claims. R.414687 - 16 - 9. Computer program that causes a computing unit to perform all the process steps of a method according to any one of claims 1 to 7 when executed on the computing unit.
10. Machine-readable storage medium with a computer program stored thereon according to claim 9.