Keywords

1 Introduction

1.1 Background

The importance of sustainable energy solutions in the automotive industry has led global automakers to promote electric vehicles (EVs) to meet environmental regulations. Despite technological advancements, challenges persist, hindering a smooth transition to EVs. While fast charging is available, EV charging times still exceed those of conventional internal combustion engine vehicles. Additionally, the growth of EV charging infrastructure has not kept pace with EV adoption, leaving drivers uncertain about charging availability when needed. Moreover, drivers experience ongoing fatigue from monitoring the Distance To Empty (DTE) and the distance to charging stations or destinations. DTE is determined by multiplying the vehicle’s total energy by its power efficiency, which decreases as the vehicle’s battery charge level (State of Charge (SoC)) nears 0%. Consequently, drivers may experience heightened anxiety once the SoC drops below a certain threshold. Accurately predicting DTE is thus a critical challenge for EVs. However, designing precise DTE prediction logic is challenging due to its complexity, influenced by factors such as driving style, external conditions, air conditioning usage, battery health, and driving mode. Furthermore, predicting remaining range becomes even more challenging due to unknown future driving conditions. (see Fig. 1).

Fig. 1.
figure 1

Sharp rise in anxiety once the SoC has crossed a certain point and Various factors affecting Distance To Empty (DTE).

Full size image

In addition, current remaining range prediction methods use calibration values because it is difficult to define the effects of various factors in a fixed formula. To determine energy efficiency for future routes, official efficiency values for each road type are calibrated and used as constants. Similarly, driving mode weights are fixed through calibration. Setting these values requires extensive data collection to match actual vehicle specs and standard values for general drivers, a process that demands significant resources from automakers. However, these calibration values are averages and do not account for individual driving styles or conditions, limiting the accuracy of the actual Distance to Empty (DTE).

Recent advances in artificial intelligence are impacting the automotive industry, with significant research being conducted using machine learning [2]. Remaining range prediction in electric vehicles is a critical task, and various methods have been proposed. Some studies estimate energy consumption based on past driving records [1, 6], which is accurate for familiar routes but less so for new ones. Modeling approaches also predict energy consumption using vehicle dynamics equations [3]. Recently, machine learning methods have emerged [4, 5], though they heavily depend on data.

This paper introduces an improved DTE algorithm that works with minimal data, leveraging past, present, and future vehicle data and machine learning. The algorithm can classifies the driver’s style from past data and predicts DTE using future driving environment information, like road gradient, and the current vehicle state. This approach addresses the uncertainties of existing DTE methods by learning and adapting to the current driver’s style in real-time, enhancing DTE accuracy with less data.

2 Proposed Algorithm

2.1 Proposed Algorithm Structure

The integrated algorithm consists of a driving style assessment network and a remaining range prediction network (see Fig. 2).

Fig. 2.
figure 2

Schematic diagram of the proposed algorithm.

Full size image

The driving style assessment network leverages machine learning techniques to analyze driving data and determine driving styles. Simultaneously, the remaining range prediction network considers various factors and incorporates the assessed driving style to enhance prediction accuracy. In practical driving scenarios, diverse drivers operate vehicles in various situations, making it challenging to determine specific driving styles. This study aims to assess driving style based solely on data obtainable through the vehicle’s Controller Area Network (CAN) communication.

From this, we propose a remaining range prediction algorithm that incorporates the evaluated driving style. The integration of these networks aims to provide a comprehensive solution for predicting remaining range by considering both driving style and relevant factors in real-time EV driving scenarios.

2.2 Data Acquisition

The creation of a high-quality dataset is crucial for enhancing machine learning performance. In this study, real-world driving data from a HYUNDAI electric vehicle was collected to build the dataset. Twelve factors were obtained through CAN (Controller Area Network) communication, including vehicle speed, road gradient, accelerator position sensor (APS), brake position sensor (BPS), steering wheel angle, steering wheel angular velocity, outside temperature, odometer, total power consumption, air conditioning power consumption, and Distance to Empty (DTE) value. Diverse scales among selected factors can adversely affect the learning process. To mitigate this, normalization was applied, ensuring that data characteristics of each factor were represented on a similar scale. Additionally, synchronization for the timestamp was performed. The processed datasets were then divided into a randomly selected test set and a training set for network training and performance evaluation. The test set-to-training set ratio was set at 0.1:0.9 for this study. The total driving data acquired amounted to approximately 5,000 s and 40 km, comprising short driving sessions from different days using the company’s test car. During testing, half of the driving sessions were conducted under normal conditions, while the remaining sessions involved diverse driving behaviors, including rapid acceleration/deceleration maneuvers and adjustments to the air conditioner. From the collected data, five features were utilized to assess driver behavior, including APS, BPS, steering wheel manipulation, and longitudinal vehicle velocity. These feature data were segmented into 500-s intervals and categorized into three label classes (Bad, Normal, Good) to classify driver styles (see Fig. 3).

Fig. 3.
figure 3

Labeled collection data.

Full size image

The remaining features were input into the second network to predict the vehicle’s power efficiency, considering the current vehicle conditions. These feature data were divided into 100-s intervals, and clustering was employed to determine the class label based on the actual travel distance divided by the total energy consumption of the corresponding driving data.

2.3 Network Layer

The order and outline of the layers used to optimize the performance of the proposed algorithm are as follows. (see Table 1).

Table 1. Layer Descriptions
Full size table

3 Conclusion

The power efficiency classification results of the network are depicted in Fig. 4.

Fig. 4.
figure 4

Results of the test data set.

Full size image

The distance calculated by multiplying the network’s final output, the predicted power efficiency, by the vehicle’s total energy, proved to be more effective in determining the practical remaining range value compared to the existing method. This success can be attributed to the algorithm’s effective handling of shorter distances, unlike existing Distance To Empty (DTE) logic, which relies on power efficiency calculated from previous driving data. This paper presents a practical approach to predict the remaining range, alleviating anxiety for EV drivers. The proposed method demonstrates effectiveness even with limited data and holds promise for real-world electric vehicles with further development. Moreover, it is expected to be particularly effective in scenarios where drivers change frequently, such as shared vehicles or corporate fleets. Future studies will consider additional factors and the impact of regenerative braking on electric vehicles.