For electric vehicles, a precise, route-specific range prediction is mandatory to increase the amount of destinations reachable with high reliability
A calculation concept developed at the IAV GmbH comprises a data-based model for range prediction, which takes interactive effects between environmental, route, vehicle and driver-specific influences into account. The driver behaviour was considered through lazy learning of the relationships between driver characteristics and environment, route and vehicle, followed by a subsequent prediction step. The predicted energy demand can afterwards be used as an input for a subsequent operation strategy, which could increase the vehicles range. The concept presented in this article could improve the user acceptance of future battery electric vehicles. It was honoured with the IAV Talent Award in the category “Hermann Appel” in 2018.
PRECISE & ROBUST METHOD
Local emission-free, individual mobility is made possible by the constantly growing product range of electric vehicles and is inspiring more and more users. The small capacity of current energy storage systems (batteries) in combination with the fragmentary infrastructure for charging facilities requires a precise and robust method for range prediction. Such a method could promote the long-lasting confidence of future users in this key technology [1].
This article describes a concept for energy demand prediction, which allows the processing of individual, route-specific influence variables in a data-based models. This information is also used for an operation strategy, which increases the vehicle’s range, if required. Thus, by increasing the reliability of reaching certain destinations, the trust in electric mobility could be improved. Further on, the concept is suited to be implemented in embedded systems [2].
PROCESSING OF LOCALLY DISCRETISED FEATURES
The prediction of propulsion energy demand for traveling a certain route is based on the processing of influence and target features in data-based models. This is basically motivated by depicting correlated, non-linear effects between the relevant impact parameters. Since this does not apply for auxiliaries, the energy demand of the latter is calculated with physical, equation-based models. The concept uses locally discretised information, which requires the introduction of several segment types, taking into account different road topologies. These are defined in terms of events in the respective middle of each segment. (1) shows the algorithm, which is developed for the segmentation of a recorded route, consisting of GPS and compass data points. Additionally, different types of segments are displayed.
The algorithm starts with the identification of events within time series of GPS coordinates and driving direction data, along the previously chosen route. The positions of the events could be retrieved from meta data, accessible from map services like Open-StreetMap [3]. The orientation as an additional information has to be known. This enables the differentiation between segments, defined by the same event position, but passed from different directions.
Subsequently, the segment limitations are identified in the local middle between two events. This is followed by the check of the resulting segment halves exceeding a maximum length threshold. If this is the case, an additional segment of the event type, containing a straight-on road part is inserted together with the respective limitation. The previous step is again repeated with the updated segmentation. The resulted segmentation is shown in (2) for an exemplary route.
The propulsion energy demand as a target value as well as different input values are extracted segment-wise from the available database. These are organised according to their origin in the categories route, environment, vehicle and driver. Route-specific values, for example, are speed limit, level increase, curve radius and number of curves. Variables like time values and weather characteristics are related to the environment. The vehicle-specific category is mapped through, for instance, mass, air resistance and maximum recuperation power. Driver-specific parameters are, for example, the mean and maximum velocity, the positive and negative acceleration parts, the mean frequency of the velocity profile as well as the number of mechanical brake operations.
ADAPTIVE DRIVER CHARACTERISATION
Apart from parameters describing the driver’s behaviour, all mentioned impact parameters could be contemporary determined with already existing information systems or change rather slowly. To fulfil the requirement of an adaptive concept for the prediction of a route-specific energy demand, the driver’s behaviour has to be projected in dependency of interactive parameters. The resulting concept uses lazy learning to derive the dependency of driver characteristics to external impacts as shown in (3) [4].
Lazy learning simply means storing different combinations of driver parameters with simultaneously occurred environment-, route-, and vehicle-specific parameters. In this case, it is distinguished between different segment types. The previously stored information is consequently used within the k-nearest neighbours algorithm to predict future driver parameters [5]. More precisely, the nearest-neighbours to the known route-, environment- and vehicle-specific parameters are identified in a database. The underlying driver parameters are used together with distance-metrics for the computation of new, predictive driver parameters. The used parameters are previously selected by a correlation analysis.
PREDICTION OF THE ENERGY DEMAND WITH DATA-BASED MODELS
For the creation of data-based prediction models, the normalised input features are initially investigated regarding their impact on the traction’s energy demand for each segment type. This is performed with the Neighbourhood Component Feature Selection [6]. Subsequently, the available database is randomly divided into a set for training, validation and testing at the ratio of 70:15:15. Based on an error metric for the test data set, an optimal topology for each artificial neural network, as well as optimal parameters for support vector machines are determined [7].
The overall energy demand required to drive a certain route is computed by adding all segment-wise predicted values for the traction and the auxiliaries. The operation strategy, as shown in (4), is controlled with its difference to the stored energy (state of charge) at the time of the prediction.
OPERATION STRATEGY FOR NEEDS-ORIENTED RANGE EXTENSION
Several actions are defined to reduce the energy demand of a vehicle. These are ordered by the driver’s perceptibility regarding reduction of comfort and vehicle dynamics [8]. The deactivation of safety-relevant auxiliaries is not allowed. Basically, the prediction of the energy demand for the remaining route is repeatedly performed after a new mode has been activated within the operation strategy. If a certain action results in a negative energy balance, further steps are performed. This process is repeated as long as the predicted energy demand to drive the remaining route is below the currently available energy amount or all possible steps within the operation strategy have been performed.
The first action is turning-off the seat heating, which has a comparably low impact, especially due to its only seasonal usage period and the supplementation by the vehicle’s heating. The adaption of the interior temperature up to a maximum difference of 3 K to the initially wished temperature is the second step of action. This difference corresponds to the minimal temperature change perceptible by humans of around 3 K [9]. Increasing the proportion of recirculated air up to a value of 100 % is the third action. Thus, the proportion of the ambient air, which has to be conditioned for the vehicle’s interior, can be lowered. In dependency of the interior and ambient temperature, it is required to ensure that the windows do not mist due to a high humidity. So, the air in the interior is dehumidified by the air-conditioning system as required.
The final action – and thus the one with the highest impact to the driving experience – is the gradual limitation of the maximum driving velocity of the vehicle. This measure is done depending on the current GPS position within or outside built-up areas and according to the speed limit. This enables to save more energy in comparison to the limitation of the acceleration [8]. The defined value of 70 km/h outside of built-up areas is motivated by the type-limited maximum velocity of 60 km/h on German highways, supplemented by an additional buffer. The assumption of a maximum speed of 40 km/h within built-up areas is based on a publication of the Federal Statistical Office [10]. The daily average distance is hereby quantified as 39 km, travelled in 1:19 h of time. This computes to an average velocity of 33 km/h, which results in the previous value.
INVESTIGATION OF THE INTRODUCED CONCEPT THROUGH SIMULATION
The presented concept has been implemented and analysed in the Matlab software environment. (5) shows the simulation results for an exemplary driving scenario. Here, the required range is equal to the whole driven distance along the route minus the already driven proportion. At the beginning of this driving scenario, the predicted range is below the required range of 18.8 km. Thus, the energy demand is reduced by the Operation Strategy (OS) until the predicted range is above the required range.
In particular, at the beginning of the scenario a high saving potential could be achieved due to the degradation of the thermal auxiliaries. Since comparably low velocities are predicted within the simulated driving scenario, the degradation of the drive has a low impact on the overall energy demand compared to the thermal auxiliaries. The proposed concept for an energy management system at overall vehicle system level is visualised in (6).
SUMMARY & OUTLOOK
In this article, a concept for data-based range prediction has been introduced. It allows the consideration of interactive effects between impact factors like environment, route, vehicle and driver. All features used are gained by locally discretising routes into segments. Different segment types were defined. This enables the mapping of different road-topology specific characteristics. The driver’s behaviour in the data-based models is taken into account by lazy learning of combinations between driver parameters and the environment-, route- and vehicle-specific parameters.
This is followed by the computation of predictive driver parameters using the k-nearest neighbours algorithm. For every segment type artificial neural networks, as well as support vector machines with optimised topologies have been trained to predict the energy demand of the propulsion system. In combination with a physical model for the auxiliaries, the gained information is used for a downstream operation strategy. This strategy decreases the energy demand of the vehicle in dependence of the situation and is prioritised according to subjective perception in the vehicle.
The extension of the concept to enable route-unspecific range prediction is proposed for future works. Additionally, the dynamic integration of new data sources as well as the implementation of iterative training processes for data-based models on a backend would save computational effort in the vehicle itself.
REFERENCES
[1] Plötz, P.; Gnann, T.; Kühn, A.; Wietschel, A.: Markthochlaufszenarien für Elektrofahrzeuge. Studie im Auftrag der acatech – Deutsche Akademie der Technikwissenschaften und der Arbeitsgruppe 7 der Nationalen Plattform Elektromobilität (NPE), 2013
[2] Simonis, c.: Routenspezifische Reichweitenprädiktion und betriebsstrategie für Elektrofahrzeuge, Munich, Technical University, master’s thesis, 2018
[3] OpenStreetMap Foundation: Open Data Commons Open Database license/creative commons by-SA. Online: www.openstreetmap.org/copyright, access: July 4, 2019
[4] Simonis, c.; Sennefelder, R.: Route-specific Driver characterization for Data-based Range Prediction of battery Electric Vehicles. 14th International conference on Ecological Vehicles and Renewable Energie (Ever), Monaco, 2019
[5] Ertel, W.: Grundkurs künstliche Intelligenz. Wiesbaden: Vieweg+Teubner-Verlag, 2009
[6] yang, W.; Wang, K.; Zuo, W.: Neighborhood component Feature Selection for High-Dimensional Data. In: Journal of computers 7 (2012), No. 1, pp. 161-168
[7] Vapnik, V. N.: An overview of statistical learning theory. In: IEEE Transactions on Neural Networks 10 (1999), No. 5, pp. 988-999
[8] basler, A.: Eine modulare Funktionsarchitektur zur Umsetzung einer gesamtheitlichen betriebsstrategie für Elektrofahrzeuge. Dissertation, Karlsruher Institut für Technologie, KIT Scientific Publishing, volume 42, 2015
[9] Schmidt, M.: Ein selbstadaptierender, dynamischer Energiemanagemen-tansatz für das elektrische Kraftfahrzeugbordnetz. Shaker-Verlag, Düren, 2008
[10] Hütter, A.: Verkehr auf einen blick. German Federal Statistical Office, Wiesbaden, 2013
AUTHOR
CHRISTOPH SIMONIS, M. SC. was master student at the department of mechanical engineering of the Technical University of Munich and is employee in the Powertrain Mechatronics/Drivetrain department of the IAV GmbH in the office of Munich (Germany).
THANKS
The author would like to thank Roman Sennefelder (IAV) and Matthias Stein-sträter (TU Munich) for their excellent collaboration during the development of the proposed concept within this master’s thesis. Further, he would like to thank Iris Schmitz-Wilke for the graphical illustration of the title figure for this article. Last but not the least, gratitude should be expressed to IAV GmbH for the opportunity and promotion of this work in the presented subject.
By Christoph Simonis
Source: https://autotechreview.com
CUT COTS OF THE FLEET WITH OUR AUDIT PROGRAM
The audit is a key tool to know the overall status and provide the analysis, the assessment, the advice, the suggestions and the actions to take in order to cut costs and increase the efficiency and efficacy of the fleet. We propose the following fleet management audit.
