Range calculation system and methods for range calculation

Abstract

The invention relates to a range calculation system for zero-emission commercial vehicles, comprising: - a database (7) for storing data on the energy consumption and / or range of a fleet (10) of zero-emission commercial vehicles, - a machine learning model configured to predict the energy consumption and / or range of a zero-emission commercial vehicle for a given route, - a training module (3) configured to train the machine learning model using the data from the database (7), - an evaluation module (4) for evaluating the machine learning model, - a quality module (6) for determining deviations between the predicted energy consumption and / or range on the one hand and the actual energy consumption and / or range measured in the commercial vehicle on the other hand, in order to determine whether or not retraining of the machine learning model is required, wherein the range calculation system is configured, if retraining of the machine learning model is required, to determine whether the data collected from the fleet (10) of electrically powered vehicles are sufficiently diverse to retrain the machine learning model, and otherwise to perform a simulation of journeys and / or routes in a simulation module (8) to enrich the data.

Description

[0001] The invention relates to a range calculation system according to the preamble of claim 1 and a method for range calculation according to the preamble of claim 6.

[0002] Implementing a route-based range and energy forecasting system for electric commercial vehicles (BEVs; FCeVs) presents the challenge of vehicle variant diversity, which has a significant impact on vehicle energy consumption and thus on the accuracy of range predictions. Customers have countless operating profiles (load, long-haul, distribution, topography, climate, etc.), different vehicle bodies with varying energy consumption (refrigerated bodies, refuse collection, cement mixers, crane trucks, tank trucks, etc.), and different driver behavior regarding energy consumption (speed selection, use of regenerative braking, etc.).

[0003] It will be virtually impossible to generate a sufficiently accurate energy and range prediction for every scenario, especially if each variant requires manual engineering adjustments. Even the application of machine learning methods alone cannot solve the problem as long as a static prediction model without a self-learning function is used.

[0004] CN 112860782 A describes a method for estimating the range of a pure electric vehicle based on big data analytics. The method includes: capturing and preprocessing the original driving data of an electric vehicle; defining a target variable for the driving consumption of an electric vehicle under a unit mileage; analyzing the correlation between the variables in the original data and the driving consumption, and analyzing the correlation weight between each correlation variable and the driving consumption; performing dimensionality reduction processing on the obtained original variables; selecting a predictive model for the big data analysis of driving consumption in a suitable mode according to the traffic conditions on the current route, and predicting the consumption value on a future driving route.Calculating the distance traveled on the destination route and the distance traveled on the route prior to the trip, in combination with the current value of the remaining battery capacity and the predicted driving consumption; and performing an analysis of variance of the predicted and actual values ​​of the battery capacity consumed by the electric vehicle on the route to perform a secondary correction of the result.

[0005] The invention is based on the objective of providing a novel range calculation system and a novel method for range calculation.

[0006] The problem is solved according to the invention by a range calculation system with the features of claim 1 and a method for range calculation with the features of claim 6.

[0007] Advantageous embodiments of the invention are the subject of the dependent claims.

[0008] A range calculation system for zero-emission commercial vehicles is proposed, including: - a database for storing data on the energy consumption and / or range of a fleet of zero-emission commercial vehicles, - a machine learning model configured to predict the energy consumption and / or range of a zero-emission commercial vehicle for a given route, - a training module configured to train the machine learning model using data from the database, - an evaluation module for evaluating the machine learning model, - a quality module for determining deviations between the predicted energy consumption and / or range on the one hand and the actual energy consumption and / or range measured in the commercial vehicle on the other hand, in order to determine whether or not retraining of the machine learning model is required.

[0009] According to the invention, the range calculation system is configured to determine, when retraining of the machine learning model is required, whether the data collected from the fleet of electrically powered vehicles is sufficiently diverse to retrain the machine learning model and, if not, to perform a simulation of journeys and / or routes in a simulation module to enrich the data.

[0010] The solution according to the invention enables improved range prediction quality for a vehicle fleet, including special applications, while simultaneously reducing development and operating costs. Improved, more precise range prediction allows for better utilization of the battery and transport capacity of commercial vehicles, thus improving the customer's overall costs.

[0011] Exemplary embodiments of the invention are explained in more detail below with reference to a drawing.

[0012] This shows: Fig. 1: A schematic view of a self-learning range calculation system for especially zero-emission commercial vehicles.

[0013] The only Fig. Figure 1 is a schematic view of a self-learning range calculation system for especially zero-emission commercial vehicles.

[0014] First, a machine learning model is trained in step 3 using data from a fleet of 10 electrically powered vehicles, acquired in step 1. This training is then evaluated in step 4. The process involves steps such as data acquisition (1), data preprocessing (2), model training (3), and model evaluation (4).

[0015] In step 5, the trained machine learning model is used as a live model and provides predictions about the energy consumption and / or range of electric vehicles for a given route.

[0016] In a quality control step 6, deviations between the predicted energy consumption and / or range on the one hand and the actually measured energy consumption and / or range on the other are determined to ascertain whether or not retraining of the machine learning model is necessary. Other parameters, such as a predicted speed and / or the mechanical energy at a wheel of the vehicle, can be considered to more accurately identify the deviations in the prediction and determine where a correction might be required. Possible causes of deviations could include, for example, a particularly low ambient temperature and the associated change in the traction battery's behavior, leading to a significant reduction in the vehicle's range.If such a case was not sufficiently represented in the training dataset, the prediction model cannot provide a good prediction. As a measure, the fleet data can then be searched for such cases with low outside temperatures. If a case is found, the prediction model can be retrained, whereby the identified case must be more strongly represented in the training data or can be specifically used in a subsequent training session. If the case is not sufficiently represented in fleet data pool 7, i.e., the training data is not sufficiently diverse, a simulation (step 8) can be used to generate the missing data, utilizing existing data and physical knowledge.

[0017] If retraining of the machine learning model is required, it is determined whether the data collected from the fleet of 10 electric vehicles is sufficiently diverse to retrain the machine learning model. If the data collected from the fleet of 10 electric vehicles is not sufficiently diverse, then in step 8 a simulation of journeys and / or routes is performed to enrich the data, and the machine learning model is retrained as described above.

[0018] The present invention solves the problem of range prediction by automating a learning function. In a quality module 6, an automated comparison is performed between the predicted energy consumption and the measured actual energy consumption of each vehicle in a fleet 10. Notable deviations are collected, corresponding category data is retrieved from a database 7 containing streaming data from the fleet 10, prepared for training, and a range model is trained (steps 1 to 4). The quality module 6 can determine whether there is sufficient data in the database 7 for retraining the range model, which is designed as a neural network, to identify any notable deviations. If not, a simulation module 8 is activated, which simulates further journeys for the underrepresented data to artificially enrich the data set.The learning cycle shown achieves continuous quality assurance while simultaneously reducing development effort.

[0019] The device aims to provide self-learning range prediction, particularly for zero-emission commercial vehicles (e.g., electric trucks). To achieve this, a data acquisition module 1 extracts all available driving data (vehicle sensor data) from the database 7 and enriches it with external data such as route and weather data for each trip. The data pool receives real-time sensor data streamed from the vehicles.

[0020] The classic process for training neural networks takes place. For this, the data is preprocessed in a preprocessing module 2 (cleaning, normalization, and transformation of the data) and then trained by a training module 3 on a modular neural network, a machine learning model. The chosen machine learning model is modular, meaning that parts of the network can be adapted depending on the vehicle model and configuration. This model features a classic neural network that receives input in the form of the preprocessed data and predicts energy consumption, and a pre-processing module that contains the vehicle configuration. Thus, different neurons are added to the model for each different vehicle type. The trained model is then evaluated by an evaluation module 4.If selected metrics of the trained network exceed the chosen thresholds, the model is deployed in the real-world environment in step 5 to provide active range predictions. Otherwise, the training process is restarted. The information from the evaluation is used to further improve the model.

[0021] Driver 9, Fleet Manager 9, or Dispatcher 9 plan transport orders and use the range prediction, which provides energy consumption data for a given route (state of charge upon arrival, energy consumption in kWh, drivable route segment, recommendation for a charging station or gas station) and supports feasible route planning. The route planning defines the routes to be driven, the corresponding vehicle, and, if applicable, the driver 9, which the model can process. When the planned route is then driven with the corresponding vehicle, all vehicle measurement data is saved, in particular data regarding the vehicle's energy consumption (energy consumption of the drive system and all auxiliary consumers, changes in the battery's state of charge, etc.).

[0022] Next, quality module 6 becomes active, retrieves the tour planning data from range module 5 and the corresponding vehicle data, and compares the predicted energy consumption with the actual measured energy consumption. This comparison includes not only the absolute energy consumption and the battery's state of charge (SOC), but also all component consumptions (auxiliary consumers, driving consumption, and auxiliary consumption) and intermediate calculation steps (predicted speed compared to measured speed, mechanical energy at the wheels, electrical energy of the drive system, and electrical energy of auxiliary consumers). This reveals where the greatest deviations in the prediction occurred and at which calculation step a correction should be made, if necessary.If actual consumption and forecasted consumption do not largely match, quality module 6 checks whether there are other similar cases, thus requiring retraining, or whether it was a one-off anomaly. In the first case, it checks whether the dataset is (now) diverse enough to perform this training, or whether simulated data is needed to represent these rare cases (corner cases).

[0023] For this purpose, the characteristics of the incorrectly predicted journeys are fed to simulation module 8, which simulates a sufficient number of journeys and attempts to reproduce the rarely occurring cases found. The data pool is thus expanded with artificially generated data.

[0024] The loop of data acquisition (1), preprocessing (2), model training (3), and model evaluation (4) is executed again to incorporate the new data into the neural network. During retraining, the upstream part of the neural network is relearned, while the base model remains largely unchanged. If the newly trained network surpasses the metrics of the current live model, the latter is replaced by the new model. Reference symbol list Step 1, data collection, data collection module 2-step preprocessing, preprocessing module 3-step model training, training module 4-step model evaluation, evaluation module Step 5 6. Quality control step, quality module 7 Database 8-step simulation module 9 drivers, fleet manager, dispatcher 10 Fleet, vehicle fleet QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] CN 112860782 A

[0004]

Claims

[1] Range calculation system for zero-emission commercial vehicles, comprising: - a database (7) for storing data on the energy consumption and / or range of a fleet (10) of zero-emission commercial vehicles, - a machine learning model configured to predict the energy consumption and / or range of a zero-emission commercial vehicle for a given route, - a training module (3) configured to train the machine learning model using the data from the database (7), - an evaluation module (4) for evaluating the machine learning model, - a quality module (6) for determining deviations between the predicted energy consumption and / or range on the one hand and the actual energy consumption and / or range measured in the commercial vehicle on the other hand, in order to determine whether or not retraining of the machine learning model is required, characterized by , that the range calculation system is configured to determine, when retraining of the machine learning model is required, whether the data collected from the fleet (10) of electrically powered vehicles are sufficiently diverse to retrain the machine learning model, and otherwise to perform a simulation of journeys and / or routes in a simulation module (8) to enrich the data. [2] Range calculation system according to claim 1, characterized by, that the quality module (6) is configured to take into account a predicted speed and / or the mechanical energy at a wheel of the vehicle in order to determine the deviations. [3] Range calculation system according to claim 1 or 2, characterized by , that the machine learning model is designed as a modular model in which parts of a network can be adapted depending on the vehicle model and vehicle configuration, wherein the model has a neural network configured to predict energy consumption and a pre-processing part containing a vehicle configuration, with different neurons being able to be added to the model for each different vehicle type. [4] Range calculation system according to claim 3, characterized by, that during retraining the upstream part of the neural network is retrained and the base model remains at least largely unchanged, whereby a current live model of the neural network is replaced by the newly trained neural network if its metrics exceed the metrics of the current live model. [5] Range calculation system according to one of the preceding claims, characterized by , that the data stored in the database (7) regarding the energy consumption of the vehicle include energy consumption of a drive and auxiliary consumers as well as changes in the state of charge of at least one battery of the commercial vehicle. [6] Method for calculating the range of zero-emission commercial vehicles, wherein: - Data on the energy consumption and / or range of a fleet (10) of zero-emission commercial vehicles are stored in a database (7), - a prediction of the energy consumption and / or range of a zero-emission commercial vehicle for a given route is made using a machine learning model, - the machine learning model is trained using the data from the database (7) and then evaluated, - Deviations between the predicted energy consumption and / or range on the one hand and the actual energy consumption and / or range measured in the commercial vehicle on the other hand are determined in order to ascertain whether or not retraining of the machine learning model is necessary. characterized by, that when retraining of the machine learning model is required, it will be determined whether the data collected from the fleet (10) of electrically powered vehicles are sufficiently diverse to retrain the machine learning model, and otherwise be put into a simulation of journeys and / or routes to enrich the data.

Images