Method and system for predicting charging behavior of electric vehicles

Abstract

In a method for predicting the charging behavior of electric vehicles, charging process data from charging processes of electric vehicles in a vehicle fleet are determined. The charging process data are each assigned to a charging process and one of the vehicles. Fleet data is generated based on the charging process data. The fleet data is processed using cluster analysis to divide the vehicles of the vehicle fleet into at least two groups that differ in terms of their charging behavior. At least for each of the groups, a prediction model is determined that is designed to predict at least one parameter of the charging process based on an input related to the circumstances of a charging process. The prediction models are further provided.

Description

[0001] The invention relates to a method for predicting the charging behavior of electric vehicles. The invention further relates to a system for predicting the charging behavior of electric vehicles.

[0002] Certain parameters of an electric vehicle charging process, such as the expected plug-in time, can be requested from a user of the electric vehicle at the start of the charging process. Furthermore, methods are known from the prior art that make it possible to predict certain parameters of the charging process. Based on the requested or predicted parameters, the charging process can be optimized. For example, the charging process can be controlled based on the expected plug-in time so that it is completed exactly at the time of unplugging. This can extend the battery life. However, predicting these parameters is very difficult due to the small number of data points for individual vehicles. For example, a private owner of an electric vehicle charges it on average only twice a week.In addition, vehicles within a fleet can exhibit very different charging behaviors. For example, some vehicles are routinely charged overnight in the same geographical location, while others charge more randomly and use more rapid charging sessions. Obtaining these parameters is not particularly reliable.

[0003] The object of the invention is to provide a method and a system for predicting a charging behavior of electric vehicles, which make it possible to reliably predict parameters of a charging process of an electric vehicle.

[0004] This object is achieved by a method having the features of claim 1 and by a system having the features of the independent device claim. Advantageous further developments are specified in the dependent claims.

[0005] In the proposed method for predicting the charging behavior of electric vehicles, charging process data from charging processes of electric vehicles in a vehicle fleet are determined. The charging process data is each assigned to a charging process and one of the vehicles. Fleet data is generated based on the charging process data. The fleet data is processed using cluster analysis to divide the vehicles of the vehicle fleet into at least two groups that differ in terms of their charging behavior. At least for each of the groups, a prediction model is determined that is configured to predict at least one parameter of the charging process based on an input related to the circumstances of a charging process. The prediction models are further provided.

[0006] With the help of the prediction models, it is possible to predict the charging behavior of individual vehicles in the vehicle fleet. For example, a time of unplugging can be predicted in order to carry out a charging process such that it is completed exactly at the time of unplugging. Each prediction model accepts at least one input related to the circumstances of the charging process to be predicted. Depending on the input, the prediction models then generate an output corresponding to at least one parameter of the charging process. For example, the time of day at which the charging process is to take place and / or a state of charge at the start of the charging process can be used as the input. The prediction model used can then output the expected time of unplugging as the output. Further exemplary inputs and outputs are mentioned below in connection with further embodiments of the method.

[0007] In order to create the predictive models, at least the charging process data is first recorded. The charging process data corresponds to information that is related to the charging processes of the vehicles in the vehicle fleet. For example, the charging process data can include when the charging process takes place and / or how long the vehicle remains plugged in, the so-called plug-in time. Further information that can be recorded as the charging process data is mentioned below in connection with further embodiments of the method. Fleet data is then generated at least from the charging process data. For example, the fleet data can simply be the entirety of the recorded data. However, the fleet data can also include information derived from the recorded data, for example aggregated variables such as mean values.The fleet data is subjected to cluster analysis to identify groups of vehicles within the fleet that exhibit similar charging behavior. Predictive models are then generated based on these groups. Specifically, at least one predictive model is created for each identified group, which can be used to predict the charging behavior of vehicles in the respective group. In other words, to generate the predictive models, data points from vehicles that exhibit similar charging behavior are used. This allows more charging events to be used to create the predictive models than with a method that uses only the charging events of a single vehicle to make a prediction for that vehicle.Since the prediction models in the proposed method can draw on more historical data, the prediction is more reliable than with known methods.

[0008] In one embodiment, at least one of the prediction models is a machine learning method that has been trained using at least a portion of the fleet data to predict the at least one parameter of the charging process based on the input. The machine learning method is trained, in particular, as supervised learning. The following machine learning methods can be used, in particular: linear regression, support vector regression, decision trees, and / or neural networks. Furthermore, ensemble methods can be used, such as averaging, stacking, blending, bagging (random forests), and / or boosting. Machine learning methods have the ability to learn from large amounts of data and recognize patterns. This enables the prediction models to make more precise predictions.

[0009] In a further embodiment, a preliminary prediction model is determined for each of the groups. Based on the preliminary prediction models and using transfer learning, a vehicle-specific prediction model is generated for each vehicle in the vehicle fleet. In this embodiment, the preliminary and / or vehicle-specific prediction models can each be machine learning methods. The preliminary prediction models are each specific to one of the groups of vehicles. The vehicle-specific prediction models are then generated from the preliminary prediction models. This can be done in particular by supervised retraining with vehicle-specific data, for example, the part of the fleet data that can be assigned to one of the vehicles in the vehicle fleet. This approach enables the vehicle-specific prediction models to learn from the extensive fleet data.The vehicle-specific prediction models are thus not entirely dependent on the few vehicle-specific data points. This allows the vehicle-specific prediction models to be better adapted to the individual characteristics of each vehicle in the fleet without overfitting and thus resulting in uncertain predictions due to a dataset that was too small for training.

[0010] In another embodiment, the cluster analysis is performed using unsupervised learning. A number of well-known classification algorithms can be used for cluster analysis, for example, a partitioning clustering method such as the k-means algorithm or affinity propagation, a hierarchical clustering method, a density-based clustering method such as DBSCAN or OPTICS, or combined methods such as spectral clustering or BIRCH. All of these methods allow the fleet data to be reliably classified to identify groups of vehicles with similar charging behavior.

[0011] In a further embodiment, the charging process data each comprise at least one of the following information: a unique identifier of the vehicle performing the charging process; the day and / or time of day on which the charging process is started and / or ended; the state of charge at the start and / or end of the charging process; the charging type of the charging process; the duration of the charging process; and / or the reason why the charging process was ended. A chassis number, for example, can be used as a unique identifier. Information related to the day can be recorded, for example, the date, the day of the week, and / or whether the day is a holiday or a working day. The charging type can be recorded, for example, whether charging is carried out with direct current or alternating current. In addition, a distinction can be made between charging with a low voltage and a high voltage.Possible reasons why the charging process was terminated include: the plug was pulled out or the charging process was terminated because the state of charge reached and / or exceeded a predetermined value. Based on the above information, the vehicles can be reliably classified into groups using cluster analysis.

[0012] In a further embodiment, trip data is determined for each vehicle in the vehicle fleet, corresponding to at least one piece of trip information related to the movement of the vehicle and / or the location of the vehicle. The fleet data is generated taking the trip data into account. In this embodiment, additional information is determined that is taken into account when generating the fleet data. For example, by adding the trip data to the fleet data. So that the trip data can be clearly assigned to one of the vehicles in the vehicle fleet, the trip data also include, in particular, a unique designation of the vehicle. Taking additional information into account allows the cluster analysis to be carried out more precisely and robustly.

[0013] In a further embodiment, the travel information is one of the following information: a geographical position of the vehicle, in particular a geographical position of the vehicle during a charging process, or a route traveled by the vehicle. In particular, based on the geographical position of the vehicle during the charging process, different scenarios can be reliably differentiated. For example, charging at home and charging at a public charging infrastructure can be differentiated. The aforementioned information can also be used to identify scenarios in which the vehicle is traveling outside of its usual operating radius. Such scenarios can, for example, be given less weight in the cluster analysis in order to obtain a more precise classification into the groups.

[0014] In a further embodiment, at least one aggregated value is determined for each vehicle in the fleet. The fleet data is generated taking into account the at least one aggregated value. In this document, an aggregated value is understood to be a value derived from a group of data points to provide summary information. For example, an average, a median, the sum, or the maximum of a group of values can serve as an aggregated value. Aggregation simplifies the fleet data, making it easier to process further.

[0015] In a further embodiment, the at least one aggregated value corresponds to one of the following pieces of information, each of which is assigned to one of the vehicles in the vehicle fleet: a number of charging processes per time interval, an average duration of the charging processes, an average plug-in time per charging process, an average state of charge at the beginning of the charging processes, an average state of charge at the end of the charging processes, or an average proportion of a specific charging type. Instead of the average and the mean, the median or another suitable statistical mean can also be used. The aforementioned aggregated values can, in particular, also be standardized or scaled in a suitable manner to achieve better comparability. A more robust cluster analysis can be carried out on the basis of the aforementioned aggregated values.

[0016] In a further embodiment, the input comprises at least one of the following pieces of information, each relating to the charging process for which the at least one parameter is to be predicted: the day and / or time of day at which the charging process takes place; the geographical position of the vehicle and / or the charging infrastructure during the charging process; and / or the state of charge at the start of the charging process. The aforementioned parameters characterize the charging process in such a way that a reliable statement about the parameter to be predicted is possible. For example, a distinction can be made between charging processes on public and private charging infrastructure based on the time of day and the location at which the charging process takes place.

[0017] In a further embodiment, the prediction models are each configured to predict at least one of the following parameters: a target state of charge, a time of unplugging, and / or whether the charging process will take place overnight. Based on the aforementioned parameters, the charging process can be optimized, for example. For example, a specific charging function can be selected taking into account the predicted time of unplugging, so that the target value for the state of charge is reached precisely at the expected time of unplugging.

[0018] In another embodiment, the charging process data is continuously recorded. The fleet data is regenerated and / or updated at regular intervals. In particular, the prediction models are retrained or regenerated using the new fleet data. As a result, the data set used to generate the prediction models is continuously expanded and updated, for example, to reflect new trends.

[0019] The invention further relates to a system for predicting the charging behavior of electric vehicles. The system comprises at least one processing unit configured to receive and process charging process data from charging processes of electric vehicles in a vehicle fleet, wherein the charging process data is each assigned to a charging process and one of the vehicles. The processing unit is also configured to generate fleet data based on the charging process data and to process the fleet data using cluster analysis to divide the vehicles in the vehicle fleet into at least two groups that differ in terms of their charging behavior.The processing unit is further configured to determine, at least for each of the groups, a prediction model configured to predict at least one parameter of the charging process based on an input related to circumstances of a charging process, and to provide the prediction models.

[0020] The processing unit can be implemented in various ways. For example, the processing unit can comprise one or more servers maintained by a system operator. Cloud computing, especially serverless computing, offers a particularly advantageous implementation option. These methods allow the processing of charging process data and fleet data to take place in a cloud, without the operator having to worry about maintaining or scaling the servers. A key advantage is cost efficiency, as billing is based on the actual computing time used.

[0021] The proposed system has the same advantages described above as the claimed method. In particular, the system can be further developed with the features of the dependent claims directed to the method. Furthermore, the method described above can be further developed with the features described in this document in connection with the system.

[0022] Embodiments of the invention are explained in more detail below with reference to the figures, in which: Fig. 1 shows a schematic representation of an architecture of a system for predicting charging behavior of electric vehicles according to an embodiment; and Fig. 2 shows a flowchart of a method for predicting charging behavior of electric vehicles according to an embodiment.

[0023] Fig. 1 shows a schematic representation of an architecture of a system 100 for predicting a charging behavior of electric vehicles according to an embodiment.

[0024] The system 100 is used to predict at least one parameter of a charging process of an electric vehicle so that this charging process can be optimized, for example. For example, based on the prediction of the system 100, a decision can be made as to when a vehicle's battery should be preconditioned, how long the charging process will take, or how much energy is required until the next charge. Fig. The system 100 described in Figure 1 is implemented in a cloud computing environment, but this embodiment is merely an example and does not imply any limitation on possible implementations. Fig. In the embodiment shown in Figure 1, the cloud computing environment thus forms a processing unit 102 of the system 100. The processing unit 102 comprises, purely by way of example, a plurality of modules 106, 110, 112, 114, 116, 118 for processing data. These modules 106, 110, 112, 114, 116, 118 can be implemented, in particular, as software or parts of a software that is executed by the processing unit 102. The processing unit 102 further comprises memory modules 104 for storing data.

[0025] A receiving module 106 of the processing unit 102 is configured to receive at least charging process data from charging processes of electric vehicles in a vehicle fleet. The charging process data are each assigned to a charging process and include, for example, when the charging process was started and / or ended, the state of charge at the start and / or end of the charging process, the duration of the charging process, and the reason why the charging process was ended. In order to be able to clearly assign the charging process to one of the vehicles in the vehicle fleet, the charging process data can additionally include a unique designation of the vehicle, for example, a chassis number or a customer number uniquely assigned to the owner.

[0026] The receiving module 106 can also be configured to receive trip data from the vehicles in the vehicle fleet, which corresponds to at least one piece of trip information related to the movement of the vehicle and / or the location of the vehicle. The trip information can be, for example, a geographical position of the vehicle during a charging process or a route traveled by the vehicle. Based on this trip information, a geographical location can be assigned to each charging process.

[0027] In particular, the receiving module 106 continuously records the charging process data and / or the travel data. The receiving module 106 is further configured to store the data received from the vehicles in one of the memory modules 104 of the processing unit 102.

[0028] The charging process data and / or the travel data are determined in particular by the vehicles themselves. For this purpose, the vehicles can each comprise a determination unit 108 arranged in the vehicle. In such an embodiment, the determination units 108 each form a part of the system 100. For the sake of clarity, Fig. 1 only an exemplary investigation unit 108 is shown.

[0029] An aggregation module 110 of the processing unit 102 is configured to generate fleet data from the data received from the vehicles. For example, the aggregation module 110 summarizes the charging process data and the trip data and stores them in one of the memory modules 104 of the processing unit 102. However, the aggregation module 110 can also be configured to determine aggregated values from the charging process data and / or the trip data, for example, a number of charging processes per time interval and vehicle. The aggregated values are generated in such a way that they can each be clearly assigned to one of the vehicles in the vehicle fleet. If the charging process data and / or the trip data are continuously recorded, the aggregation module 110 can further be configured to regenerate the fleet data at regular intervals or to update it based on the newly recorded data.

[0030] A cluster analysis module 112 is configured to perform a cluster analysis of the fleet data in order to classify the vehicles of the vehicle fleet into different groups that differ in terms of their charging behavior. Preferably, each vehicle is assigned to only one of the groups. The cluster analysis is carried out, in particular, as unsupervised learning. The cluster analysis identifies clusters within the fleet data that correspond to groups of vehicles in the vehicle fleet that each exhibit similar charging behavior. The cluster analysis module 112 is further configured to store the result of the cluster analysis in the form of data in one of the memory modules 104 of the processing unit 102. If the charging process data and / or the trip data are continuously recorded, the cluster analysis module 112 can be configured to repeat the cluster analysis at regular intervals.

[0031] A model generation module 114 is configured to determine a prediction model at least for each of the groups. Each of the prediction models is configured to predict the at least one parameter of the charging process based on an input related to the circumstances of a charging process. For example, the model generation module 114 trains one or more machine learning methods as the prediction models to predict the at least one parameter of the charging process based on the input. The model generation module 114 uses at least a portion of the fleet data as training data, for example, the portion of the fleet data assigned to the group for which the prediction model is to be created. The prediction models are each behavioral models that can predict the charging behavior of the vehicles in the groups.The model generation module 114 can also be configured to initially generate the group-specific prediction models as preliminary prediction models, which can be further processed by an optional transfer learning module 116 to generate vehicle-specific prediction models. The model generation module 114 is configured to store the prediction models in the form of data in one of the memory modules 104 of the processing unit 102. If the charging process data and / or the trip data are continuously recorded, the model generation module 114 can further be configured to redetermine the prediction models at regular intervals and / or to update them based on the new fleet data.

[0032] The optional transfer learning module 116 is configured to process the group-specific preliminary prediction models using transfer learning to generate the vehicle-specific prediction models. For example, the transfer learning module 116 is configured to retrain a preliminary prediction model specific to one of the groups using fleet data related to a single vehicle in this group to obtain a prediction model specific to that vehicle. The transfer learning module 116 is further configured to store the vehicle-specific prediction models as data in one of the memory modules 104 of the processing unit 102.If the charging process data and / or the trip data are continuously recorded, the transfer learning module 116 can also be designed to redetermine the vehicle-specific prediction models at regular intervals and / or to update them based on the new fleet data, in particular by retraining them with the updated fleet data.

[0033] Furthermore, the processing unit 102 comprises, purely by way of example, a programming interface 118 (API) via which the prediction models are provided. Alternatively or additionally, requests can be sent to the prediction models and outputs can be received via the programming interface 118. This also provides the prediction models within the meaning of this document.

[0034] By using Fig. 1, a method for predicting charging behavior of electric vehicles is carried out. This method is described below with reference to Fig. 2 is described in more detail.

[0035] Fig. 2 shows a flowchart of the method for predicting a charging behavior of electric vehicles according to an embodiment.

[0036] The method serves to generate and provide the prediction models so that the at least one parameter of a charging process of an electric vehicle can be predicted using these prediction models. The method can be implemented, for example, by the method described with reference to Fig. 1 described system 100 and is described below purely by way of example using this system 100.

[0037] The method is started in step S200. In step S202, the charging process data from charging processes of the vehicles in the vehicle fleet are determined. The charging process data is determined, for example, by the vehicles themselves and then transmitted to the processing unit 102. In the optional step S204, the trip data is determined. The trip data can also be determined by the vehicles themselves and then transmitted to the processing unit 102. The determination of the charging process data and / or the trip data by one of the vehicles itself can be carried out in particular by the determination unit 108 arranged in the vehicle. Steps S202 and / or S204 can be performed continuously in order to continuously update the charging process data and / or the trip data.

[0038] In step S206, the fleet data is generated at least based on the charging process data. If trip data was determined in step S204, the fleet data is also generated based on the trip data. The fleet data can be generated by simply aggregating the charging process data and / or the trip data. However, it is advantageous to process the charging process data and / or the trip data before generating the fleet data from them. For example, aggregated values can be generated from the individual data points of the charging process data and / or the trip data, each of which is assigned to one of the vehicles in order to facilitate further processing. For example, an average number of charging processes per time interval, for example per week, and / or an average duration of the charging processes can be determined for each vehicle in the vehicle fleet.To better compare data points, individual values of the charging process data and / or the trip data can be normalized or scaled. For example, scaled values z can be normalized according to the formula. z=xi−μσ where x i the original value, µ a mean, and σ the standard deviation. The mean µ can, for example, be the mean for one of the vehicles, one of the groups, or the entire vehicle fleet. Values scaled according to this formula have a mean of 0 and a standard deviation of 1, which makes the values particularly easy to compare. Step S206 can be repeated at regular intervals to update the fleet data based on more recent charging data and / or trip data.

[0039] In step S208, the fleet data is subjected to cluster analysis to divide the vehicles of the vehicle fleet into groups. Based on characteristics determined from the fleet data, the vehicles of the vehicle fleet are grouped such that vehicles with similar charging behavior are assigned to the same group. These groups form the basis for the prediction models created in the following steps. For example, the vehicles of the vehicle fleet can be divided into a predetermined number of groups using a k-means algorithm. The cluster analysis can be repeated at regular intervals to update the allocation of vehicles to the groups based on more recent charging process data and / or trip data.

[0040] In step S210, at least one prediction model is determined for each of the groups. In particular, the prediction models can be generated in this step as preliminary prediction models, from which the vehicle-specific prediction models are later generated in the optional step S212. The group-specific prediction models are each designed to predict at least one parameter of the charging process for a vehicle in the group based on an input related to the circumstances of a charging process. In the optional step S212, the vehicle-specific prediction models are created from the preliminary prediction models using transfer learning. This is done, for example, by retraining the group-specific preliminary prediction models with training data specific to one of the vehicles in the group.Steps S210 and / or S212 may be repeated regularly to update the prediction models based on more recent charging data and / or trip data.

[0041] In step S214, the group-specific and / or vehicle-specific prediction models are provided. The prediction models are provided, in particular, via the programming interface 118. Alternatively, the prediction models can be provided by enabling queries to the prediction models. The method then terminates in step S216.

[0042] In the case of the Fig. 1 and Fig. 2, at least the processing unit 102 forms the system 100 for predicting charging behavior of electric vehicles. Further Fig. 1 and Fig.Elements and features shown in Figure 2 and mentioned in the preceding description may be part of the system 100. Likewise, method steps described with reference to the system 100 may be part of the claimed method. List of reference symbols 100 systems 102 processing unit 104 memory module 106 Receiver module 108 Investigation Unit 110 Aggregation module 112 Cluster analysis module 114 Model generation module 116 Transfer learning module 118 Programming interface

Claims

[1] Method for predicting charging behavior of electric vehicles, in which a) charging process data from charging processes of electric vehicles in a vehicle fleet are determined, whereby the charging process data are assigned to a charging process and to one of the vehicles; b) fleet data is generated on the basis of the charging process data; (c) the fleet data are processed using cluster analysis to classify the vehicles in the fleet into at least two groups that differ in terms of their charging behaviour; d) at least for each of the groups, a prediction model is determined which is designed to predict at least one parameter of the charging process on the basis of an input related to the circumstances of a charging process; and e) the forecast models are provided. [2] The method of claim 1, wherein at least one of the predictive models is a machine learning method trained using at least a portion of the fleet data to predict the at least one parameter of the charging process based on the input. [3] Method according to claim 1 or 2, wherein a preliminary prediction model is determined for each of the groups; and based on the preliminary prediction models and using transfer learning, a vehicle-specific prediction model is generated for each vehicle of the vehicle fleet. [4] Method according to one of the preceding claims, wherein the cluster analysis is carried out as unsupervised learning. [5] Method according to one of the preceding claims, wherein the charging process data each comprise at least one of the following information: a unique designation of the vehicle carrying out the charging process; the day and / or time of day on which the charging process is started and / or ended; the state of charge at the start and / or end of the charging process; the type of charging process; the duration of the charging process; and / or the reason why the charging process was ended. [6] Method according to one of the preceding claims, wherein for each vehicle of the vehicle fleet, journey data are determined which correspond to at least one item of journey information relating to the movement of the vehicle and / or the location of the vehicle; and wherein the fleet data are generated taking into account the journey data. [7] Method according to claim 6, wherein the travel information is one of the following information: a geographical position of the vehicle, in particular a geographical position of the vehicle during a charging process, or a travel route of the vehicle. [8] Method according to one of the preceding claims, wherein at least one aggregated value is determined for each vehicle of the vehicle fleet and the fleet data are generated taking into account the at least one aggregated value. [9] The method according to claim 8, wherein the at least aggregated value corresponds to one of the following information, each associated with one of the vehicles of the vehicle fleet: a number of charging processes per time interval, an average duration of the charging processes, an average plug-in time per charging process, an average value of the state of charge at the beginning of the charging processes, an average value of the state of charge at the end of the charging processes, or an average proportion of a particular charging type. [10] Method according to one of the preceding claims, wherein the input comprises at least one of the following information, each related to the charging process for which the at least one parameter is to be predicted: the day and / or the time of day at which the charging process takes place; the geographical position of the vehicle and / or the charging infrastructure during the charging process; and / or the state of charge at the start of the charging process. [11] Method according to one of the preceding claims, wherein the prediction models are each designed to predict at least one of the following parameters: a target state of charge, a time of unplugging and / or whether the charging process takes place overnight. [12] Method according to one of the preceding claims, wherein the charging process data are continuously recorded and the fleet data are regenerated and / or updated at regular intervals. [13] System (100) for predicting a charging behavior of electric vehicles, comprising at least one processing unit (102) which is designed to to receive and process charging process data from charging processes of electric vehicles in a vehicle fleet, wherein the charging process data is assigned to a charging process and one of the vehicles; to generate fleet data based on the charging process data; to process the fleet data using cluster analysis to divide the vehicles in the fleet into at least two groups that differ in terms of their charging behaviour; to determine, at least for each of the groups, a prediction model which is designed to predict at least one parameter of the charging process based on an input related to circumstances of a charging process; and to provide the forecast models.

Images