A Novel Big Data Modeling Method for Improving Driving Range Estimation of EVs

In this paper, we address a big-data analysis method for estimating the driving range of an electric vehicle (EV), allowing drivers to overcome range anxiety. First, we present an estimating approach to project the life of battery pack for 1600 cycles (i.e., 8 years/160 000 km) based on the data collected from a cycle-life test. This approach has the merit of simplicity. In addition, it considers several critical issues that occur inside battery packs, such as the dependence of internal resistance and the state-of-health. Subsequently, we describe our work on driving pattern analysis of an EV, using a machine-learning approach, namely growing hierarchical self-organizing maps, to cluster the collected EV big data. This paper contains the analysis of energy consumption and driving range estimation for EVs, including powertrain simulation and driving behavior analysis. The experimental results, including both simulating battery degradation and analysis of driving behaviors, demonstrate a feasible solution for improving driving range estimation by the EV big data.

SECTION I. Introduction

The growth of electric vehicles (EV) starts changing the way that people transit. However, one of the major issues obstructing the development of EV usage is still the range anxiety. So far, most of the production EV models have touched the road only a few years and it still takes time to enhance driver’s confidence. Unlike traditional internal combustion engine (ICE) vehicles, for instance, drivers need to be concerned about-the distance travelled and ensure that their cars can reach a charging station before electric power is out. The operating data of a car can be easily uploaded to the cloud computing environment which provides a good opportunity for online inspection with the development of internet of vehicle. In order to formulate a feasible model to estimate the long-term performance of an EV, three domain knowledge including information technology, EV powertrain design and battery management are required to consider every respect of the system. In this work, we attempted to-transform the driving log into a recognized pattern and applied the data mining approach to analyze driver’s driving patterns. Moreover, we also collected the cycle-life test data of the lithium battery module of a production EV for over 1600 full charge/discharge cycles to model the energy performance for long time operation.

“Range anxiety”, the fear of losing power and seeing EV car shut down in the middle of a long-distance drive, is one of the biggest factors preventing more widespread adoption of electric vehicles. A quick way to overcome range anxiety is through the wide deployment of battery charging stations in a country or increasing the EV battery capacity [1]. However, the cost issue would severely retard market growth. As a result, the demand speeds up novel research work related to the use of battery data on long trip optimal energy planning [2], power grid information on deployment of charging stations [3], [4], and surely the GPS (Global Positioning System) data for vehicle mass estimation to optimize the deployment of charging station. Some approaches have been proposed to solve the problem; for instance, an algorithm for the residual range estimation was proposed for lead-acid battery in [5] and [6] applying Bayes estimators to simulate the driving range for vehicle purchase decisions in California. Yuhe, et al. introduced an estimation method, in which nine factors are considered, including vehicle current location, remaining battery energy, road network topology, road grade, road link travel speed, acceleration and deceleration, wind speed, status of on-board electric devices and driver’s driving style [7]. However, this approach missed the long-term performance of an EV battery. Hayes et al. provided a good point of view from the production EV models, Nissan Leaf and Tesla Roster, and also considered the impact of HVAC (heating, ventilation, and air conditioning) loads and battery [8].

Another emerging issue is that a large number of EVs connected to the power grid for fast charging simultaneously may pose a huge threat to the quality and stability of the overall power system [9], [10]. One of the critical issues is how to manage the battery pack to get better cost-performance ratio and leverage the investment of infrastructure. However, most of the studies have been focused on recording driving data for the analysis of energy efficiency [11], [12] and routing problem [13]. The big data collected from EVs creates an unprecedented opportunity for developing novel ways for transportation and information exchange with more vehicles connecting to the EV data cloud. In this work, we implemented an approach for modeling the long-term performance of a battery pack (2000 pcs of 5.3Ah battery cells were assembled as a 106Ah100S pack) and extracting driving patterns of a production EV. The goal of this work is to develop an approach offering a sensible energy management scheme for automaker to estimate the driving range more accurately, which is able to help drivers overcome range anxiety in an effective manner.

SECTION II. EV-Battery Big Data Modeling and Driving Pattern Recognition

In this work, we focus on two major issues, battery SOH estimation and driving pattern recognition, which influence the driving range of an EV. Unlike traditional ICE vehicles, the big data associated with EV operations such as condition of the battery, driving conditions, etc. can be used to analyze charging patterns of electric vehicle owners [14]. The huge amount of data generated by electric vehicles is not only useful for the drivers but also able to benefit engineers and related vendors. As more and more EVs connect to the grid and Internet of vehicles, the detailed data of an EV, such as State-of-Health (SOH) and error log, will soon be able to be recorded and transmitted on a massive scale and, in some cases, to adapt and react to changes in the environment automatically as the cloud computing model for managing the real-time streams of smart grid data [15] and big data application for business trend [16].

As suggested by Rahimi-Eichi, five types of big data were considered for estimating the driving range including route information, weather data, driving behavior data, electrical vehicle modeling data, and battery modeling data [17]. The proposed framework provided a global thinking of using public big data for EV application. However, in a production EV system, the battery pack must be well controlled in a safe and stable condition, e.g. using liquid cooling. Actually, the temperature performance has been considered when choosing the battery cell and can be solved by a simple sensor. The EV modeling data is decided when the powertrain system is designed, which could be transferred as a fixed parameter. The most key issues will be how to model the battery for long-term operation and extract driving patterns. Some researchers have proposed different methods to estimate the SOH, for instance, genetic algorithm (GA) [18], dual extended Kalman filter (DEKF) [19] and recurrent neural networks (RNNs) [20]. However, most of their work only used single cell as testing sample without considering the mechanical resistance and balancing condition of a large format module. In our work, we used a production battery pack to learn the ageing model for estimating the usable energy in the life of an EV.

Driving habits and conditions directly affects the driving range and energy consumption of an EV. The results in [21] showed that most drivers do not want to reduce range of the EVs but improved driving skill with constant velocity after 5 months of accustomization. In the study of [22], they performed an analysis on the effect of range anxiety quantifying the State of Charge (SOC) where users appear reluctant to start an electric journey, and the effect SOC has on drive efficiency demonstrating how users modify their driving style to conserve energy when SOC reduces.

In this work, we applied EV big data to analyze driving patterns and their energy efficiency of an EV by utilizing machine learning approaches. The proposed pattern recognition technique normalizes the speed-energy consumption ratio according to the power match rule for each trip to align the driving pattern in different driving conditions. In order to understand the driving behavior of each car, in this work, an unsupervised clustering approach is applied to collectively formulate its driving behavior. The driving behavior reflects the driving habit as noted in [21] and [22]. In this work, we combine the SOH ageing model and driving pattern analysis for estimating the driving range more accurately to release range anxiety.

SECTION III. Attempted Solution and System Framework

In this work, we combine EV battery modeling method and driving pattern analysis to improve the accuracy of range estimation for automaker to deal with the range anxiety issue. In order to verify the factors which may affect the performance of an EV, it is necessary to collect related data using a systematic approach to estimate the driving range of a trip. Fig. 1 illustrates the detailed description of the proposed framework.

FIGURE 1.

Research framework.

We briefly describe the scenario of our framework as follows:

• Step 1 (Modeling Battery Cycle-Life Performance):

  The first step of our work is to collect data of parameters which are critical to the performance of a production battery pack. Our work started with the process of extracting internal resistance and voltage of an operating battery. In order to estimate the usable energy of a battery module, we perform experiments to formulate the relationship between discharging capacity and its energy. Subsequently, we simulate an SOH ageing model of a battery module by analyzing the performance data of cycle-life test.

• Step 2 (Analyzing Driver’s Driving Patterns):

  The driving data was collected and screened out from the EV cloud platform. As mentioned previously, for recognition of driving patterns, the driving data of each trip needs to be transformed as a consistent vector. The driving data is reorganized as a driving pattern, consisted of a speed-energy vector based on the power match rule of EV powertrain. Then, we utilize an unsupervised clustering approach, called growing hierarchical self-organizing maps (GHSOM), to cluster driving patterns for analyzing the driving behaviors of each car.

• Step 3 (Estimating Driving Range):

  We performed a cycle-life test for evaluating the trend of internal resistance growth and energy loss. An SOH ageing model is built to simulate the battery performance for the life of a production EV at this stage. We analyzed the energy consumption factor of each driving pattern clustered by GHSOM for automakers to adjust the estimation of power consumption according to different driving conditions as the result. In the EV cloud, the system calculates the remaining driving range and provides appropriate advises through the user interface by means of comparing the status of SOH and the planning routes uploaded by drivers.

A. Collecting Battery Data

The consistency of battery cells and modules decides the final performance of an EV battery pack. Since there are hundreds to thousands cells in a pack, the first step for an EV maker is to make sure all selected cells can match the required criteria. However, it is almost impossible to evaluate an assembled pack after installing to the vehicle platform because of cost and safety issues. Three key factors, including consistency of impedance, distribution of open circuit voltage and distribution of capacity, are used to measure the quality of fresh lithium-ion batteries. In practice, a pack maker groups cells with same capacity and impedance grade when building a new battery module. Once the pack has been assembled into an EV, it needs a sophisticated approach to evaluate the operating voltage of a pack by (1):

$$\begin{equation} E_{\left ({ I,SOC,T }\right )}={OCV}_{\!\left ({SOC, T }\right )}-I\left ({ R_{\left ({ SOC,T }\right )}\!+\!\eta _{r\left ({ I, SOC, T }\right )} }\right )\quad \end{equation}$$ where I is the current, $SOC$ is the state of charge, $T$ is the temperature, $OCV$ is the open circuit voltage, $IR$ is the internal resistance of battery cell, $\eta$ is the resistance caused by polarization, and $E_{(I,SOC,T)}$ is the discharge voltage. The resistance is contributed by electronic, electrolyte, and solid electrolyte interfaces, and a cell maker can trace the long-term performance and consistence of production by monitoring the Direct Current Resistance (DCR) which is defined as the total of ohmic $R_{\mathrm {\Omega }}=\left ({ V_{1}-V_{0} }\right )/I$ and polarization resistance $R_{p}=\left ({ V_{2}-V_{1} }\right )/I$ . One method to determine DCR uses pulses, sometimes referred to as the hybrid pulse power characterization (HPPC) test [23]. The DCR is stated as (2) and pulse characterization test method is illustrated in Fig. 2. $$\begin{equation} DCR=R_{\mathrm {\Omega }}+R_{p}=\left ({ V_{2}-V_{0} }\right )/I \end{equation}$$

FIGURE 2.

Pulse characterization test for DCR.

Basically, a battery management system (BMS) monitors the status for each series in a battery pack. For example, Fig. 3 (a) is the thermal image of an EV pack after discharging, and the BMS records both temperature and voltage distributions as shown in Fig. 3 (b). Temperature performance is also a critical variable of EV battery system, and an example of discharge capacity at various temperatures of a Nickel Manganese Cobalt Oxide (NMC) lithium-ion battery is shown in Fig. 4. Basically, the resistance of a lithium-ion battery is inversely proportional to the temperature which limits the discharge capacity in lower temperature.

FIGURE 3.

(a) Thermal image of an EV battery pack. (b) Temperature and voltage distribution of an EV battery pack.

FIGURE 4.

Temperature performance of a NMC lithium-ion battery.

In this work, we assume that the temperature is well controlled in room temperature (23°C±2°C) and only consider the voltage changes of a pack. Let $V^{\,y}=\left \{{ {v}_{1}^{\,y}, {v}_{2}^{\mathrm {\,y}},\ldots ,{v}_{\mathrm {s}}^{\mathrm {\,y}} }\right \}$ be the open circuit voltage set of each series of the EV $y$ and $\mathrm {\forall y }V_{i}^{\,y}\ne \emptyset$ . The pack OCV is:

$$\begin{equation} {v}_{pack}^{\,y}=\sum \nolimits _{i=1}^{s} {v}_{i}^{\,y} \end{equation}$$ and the voltage range of a pack is $$\begin{equation} {v}_{range}^{\,y}=\max _{i}{v}_{i}^{\,y}-\min _{i}{v}_{i}^{\,y} \end{equation}$$

It is known that bad balancing condition of a battery pack will cause energy loss and safety issues. Since the balancing condition of a pack is almost impossible to be equal, the way to control the voltage within a reasonable range becomes the matter for EV control strategy. A BMS usually sets two criteria to avoid over discharging including ${UVP}_{Cell}$ (cell under voltage protection) and ${UVP}_{pack}$ (pack under voltage protection). In practice, a BMS triggers the balancing process when the ${v}_{range}^{\,y}$ is higher than the balancing voltage ${v}_{balancing}$ . Ideally, an EV gets maximum output capacity when the balancing condition is flat, but the consistence among different series gets worse as time goes by. The total output energy might be limited because one of the series may touch the ${UVP}_{Cell}$ prior to the pack-discharging to ${UVP}_{pack}$ , and we can calculate the energy loss by accumulating the remaining energy of other series as:

$$\begin{equation} \varepsilon =\sum \nolimits _{1}^{i} {e_{v_{i}}^{i}-{UVP}_{cell}} \end{equation}$$ where $\varepsilon$ is the energy loss caused by imbalance, and $e^{i}$ is the remaining energy of the series $i$ .

B. Estimation of SOC and Energy

By observing the decay of energy and DCIR of battery, we can evaluate the SOH and estimate the driving range of an EV. Firstly, we formulate the battery discharging curve for estimating the remaining capacity of battery at different voltages. For example, the learning curve of initial capacity of a 106Ah-2series EV battery module is shown in Fig. 5.

FIGURE 5.

Discharging OCV of a 106Ah2S Li-ion battery module (8.2V to 6.0V @0.5C discharge).

In general cases, the percentage of capacity is translated as SOC from 100% to 0%, and the mapped voltage is consistent. The fitting model is a polynomial in $x$ of order $k$ as:

$$\begin{equation} f\left ({ x }\right )=q_{k}x^{k}+\cdots +q_{2}x^{2}+q_{1}x+q_{0} \end{equation}$$ where $x$ is the number of discharging cycle. The relation among discharging power, output voltage and output current is as (7), the accumulated capacity $Q$ is as (8), and the battery efficiency $\eta _{b}$ for charge and discharge is as (9). $$\begin{align} P_{discharge}=&V_{i}I \\ Q=&\sum {-I\Delta t} \\ \eta _{b}=&\begin{cases} \displaystyle \frac {\left |{ V_{i}I }\right |+I^{2}r}{\left |{ V_{i}I }\right |};charge \\[8pt] \displaystyle \frac {\left |{ V_{i}I }\right |}{\left |{ V_{i}I }\right |+I^{2}r};dischare \\ \end{cases} \end{align}$$

C. Collecting EV System Data

Ideally, we hope all the electrical energy could be fully transformed into kinetic power in an EV powertrain system, but there exists electrical and mechanical resistance in the power flow from battery to wheels, which decreases the power efficiency. Moreover, it is difficult to anticipate recording the entire situation regarding detailed operating behaviors of a driver (e.g. pressing the accelerator). Thus, a sensible way for investigation is to infer the driver’s operating behaviors by observing the energy consumption patterns at different speed situation on each trip using machine learning techniques. Some studies have focused on this topic. Verwer, et al. used the RTI+ algorithm from grammatical inference to learn the timed syntactic pattern, and showed how to use computational mechanics and a form of semi-supervised classification to construct a real-time automaton classifier for driving behavior [24]. Hai, et al. applied the mountain clustering and the fuzzy c-means (FCM) clustering to recognize both speed zone and hilly zone, and applied the probability mass function (PMF) approach to transfer driving pattern for hybrid EV [25]. In this work, we try to simplify the model to an applicable and feasible way for estimating the driving range more efficiently.

The transfer of electrical energy into mechanical power (i.e., battery to wheels) in a powertrain system is shown in Fig. 6. We assume that all EVs of a selected model have consistent performance in same environmental condition, so the efficiency of motor, motor controller and transmission would be regarded as a constant in common powertrain design. However, the wind resistance, rolling resistance and breaking energy are main factors which are affected by driver’s behaviors. The detailed description of pattern collection of accumulative energy is stated as follows.

FIGURE 6.

A simplified model for the transfer of power in an electric vehicle system.

Most of the energy loss is caused by wind resistance, breaking and heat generated from power transfer performance. The wind resistance ($R_{wind})$ , rolling resistance ($R_{rolling})$ and braking energies are defined as:

$$\begin{align} R_{wind}=&{\frac {1}{2}\times \rho \times C}_{d}\times A\times {v}_{e}^{2} \\ R_{rolling}=&{w}\times f_{r}\times {v}_{e} \\ E_{k}=&\frac {1}{2}\times m\times {v}_{e}^{2} \end{align}$$ where $C_{d}$ is the drag coefficient, $A$ is the frontal area, ${v}_{e}$ is the vehicle speed, $w$ is the weight of car, $a$ is the acceleration, $f_{r}$ is the coefficient of rolling resistance, and $m$ is the mass. Since the speed is the only variable, we can estimate the energy consumption by sampling the driving speed and transfer powertrain data as a reorganized pattern.

Based on the development of an actual EV, the rated power (kW) at speed $v \quad \mathrm {p}_{\mathrm {v}_{\mathrm {e}}}$ and maximal grade climbing power (kW) $\mathrm {p}_{\mathrm {v}_{\mathrm {a}}}$ of matching and optimizing EV powertrain [26], [27] can be calculated as:

$$\begin{align} p_{v_{e}}\ge&\frac {v_{e}}{3600\cdot \eta _{t}}\left ({ m\cdot g\cdot f_{r}+\frac {C_{d}\cdot A\cdot {v}_{e}^{2}}{21.5} }\right ) \\ p_{v_{a}}\ge&\frac {v_{e}}{3600\cdot \eta _{t}}\Biggl ({\vphantom {\frac {C_{d}\cdot A\cdot {v}_{e}^{2}}{21.5}} m\cdot g\cdot f_{r}\cdot \cos \alpha _{max}+m\cdot g\cdot \sin \alpha _{max}}\notag \\&\qquad \qquad \qquad ~~{+\,\frac {C_{d}\cdot A\cdot {v}_{e}^{2}}{21.5} }\Biggr )\quad \end{align}$$ where $\eta _{t}$ is the transmitting efficiency and $\alpha _{max}=\tan ^{-1}{(i_{max})}$ is the maximal grade climbing radian. Thus, an energy consumption profile of a single trip $k$ could be formed by recording the running time as a vector in each cycle. We define an observation vector of energy consumption at different speed as $\boldsymbol {w}^{\boldsymbol {k}}=\left \{{ {w}_{1}^{k},\cdots ,{w}_{v}^{k} }\right \}$ where $$\begin{equation} {w}_{v\le V}^{k}=(\max {\left ({ p_{v_{e}},p_{v_{a}} }\right )+p_{v_{0}})\times t} \end{equation}$$ is the energy required, $p_{v_{0}}$ is the static energy of auxiliaries, and $V$ is the maximum speed of a vehicle. The total energy consumption of a single trip $\boldsymbol {W}^{\boldsymbol {k}}$ is as (16), and the normalized energy distribution of a single trip at different speeds could be represented as (17). $$\begin{align} \boldsymbol {W}^{\boldsymbol {k}}=&\sum \nolimits _{0}^{\boldsymbol {V}} {w}_{v}^{k} \\ \hat {\boldsymbol {w}}_{\boldsymbol {v}}^{\boldsymbol {k}}=&\left \{{ \frac {w_{1}^{k}}{\boldsymbol {W}^{\boldsymbol {k}}},\cdots ,\frac {w_{v}^{k}}{\boldsymbol {W}^{\boldsymbol {k}}} }\right \} \end{align}$$

In this work, the maximum speed of the tested EV is limited in 130km/h, and we have an observation $\boldsymbol {w}^{\boldsymbol {k}}$ of 130-dimensional vector $\left \{{ \mathrm {w}_{1}^{\mathrm {k}},\cdots ,\mathrm {w}_{130}^{\mathrm {k}} }\right \}$ . For example, the total energy consumption of a selected trip is 19.998kWh, and the collected energy consumption vector is $\left \{{ \mathrm {0.065, 0.011, \cdots , 0} }\right \}$ . Then, the driving pattern of the selected trip is $\left \{{ \frac {0.065}{19.998}, \frac {0.011}{19.998}, \cdots , \frac {0}{19.998} }\right \}=\left \{{ \mathrm {0.0055, 0.0009, \cdots ,0} }\right \}$ . Fig. 7 demonstrates a real sample of the speed and energy distribution at different speeds in a single trip by transferring the log data from an EV. We extracted the speed for every 0.1 second and calculated the percentage of each km/h from 0 to the speed limit. Then, the energy consumption of speed of each km/h can be calculated and transformed to an energy distribution as a driving pattern.

FIGURE 7.

Speed and energy distribution of a single trip.

D. Analyzing Operating Data

Just as the initial attempt to improve the range estimation for solving range anxiety, a driver’s behavior directly affects the running mileage of an EV with same energy. We analyzed driving pattern by modelling the energy consumption of each trip and using unsupervised method to cluster operating data collected from an EV. Then, the result would help automakers to predict the driving distance and rearrange the control process for better experience. Here, we applied Growing Hierarchical Self-Organizing Maps (GHSOM) [28], a neural network model modified from basic SOM, as our learning approach. The major advantage of GHSOM is its hierarchical structure of expandable maps. A map could expand its size during training to achieve a better result. Any neuron in the map could even expand to a lower level map when necessary. A schematic drawing of the structure of a typical GHSOM is depicted in Fig. 8. Furthermore, our previous work of GHSOM on text mining and retrieval, in which the clustering algorithm and its application were described in more details [29].

FIGURE 8.

A typical structure of GHSOM [28].

We briefly summarized the GHSOM training algorithm as follows:

• Step 1 (Initialization Step):

  Layer 0 contains a single neuron. The synaptic weight of this neuron, $s_{0}$ , is initialized to the average value of the input vectors of a single trip:

  $$\begin{align} s_{0}=\frac {1}{\boldsymbol {N}}\sum \limits _{1\le k\le {N}}\! \boldsymbol {w}^{k} \end{align}$$ where $\boldsymbol {w}^{k}$ is the $k_{th}$ training vector and $\boldsymbol {N}$ is the number of training vectors. The mean quantization error ${mqe}_{0}$ is calculated as follows: $$\begin{equation*} {mqe}_{0}=\frac {1}{N}\sum \limits _{1\le k\le \boldsymbol {N}} \| \boldsymbol {w}^{k}-s_{0} \| \tag{19}\end{equation*}$$

• Step 2 (MAP Growing Step):

  Construct a small SOM map, e.g. containing $2\times 2$ neurons, below layer 0. This is layer 1. Train this layer by SOM algorithm in $\lambda$ steps. Find the neuron which each training vector is labelled to. A training vector is labelled to a neuron if the synaptic weight of this neuron is closest to this vector. Calculate the mean quantization error of neuron $n$ as follows:

  $$\begin{equation*} {mqe}_{n}=\frac {1}{\boldsymbol {N}}\sum \limits _{k\in X_{n}} \| \boldsymbol {w}^{k}-s_{n} \| \tag{20}\end{equation*}$$ where $X_{n}$ is the set of training vectors that label to neuron $n$ and $s_{n}$ is the synaptic weight vector of neuron $n$ .

  The mean quantization error of this map, denoted by ${MQE}_{m}$ , is the average of the mean quantization error of every neuron in this map. If ${MQE}_{m}$ exceeds a fixed percentage of ${mqe}_{0}$ , i.e. ${MQE}_{m} \ge \tau _{m}\times {mqe}_{0}$ , a new row or a new column of neurons will be inserted to this SOM. This new row or column is added in the neighborhood of the error neuron $e$ with the highest ${mqe}_{e}$ . Parameters $\tau _{m}$ and $\tau _{\mu }$ are set to control the width and depth of a training map. A larger $\tau _{m}$ accepts more tolerance which produces less clustered groups and vice versa. The insertion of neurons is depicted in Fig. 9. In this figure, a new row is added below error neuron $e$ since its most dissimilar neuron, $d$ , lies below it. The insertion process proceeds until ${MQE}_{m} \ge \tau _{m}\times {mqe}_{0}$ .

• Step 3 (Hierarchy Expansion Step):

  All neurons in the maps of a layer are examined to determine if they need further expansion to new maps in the next layer. At the transition from one layer to the next, the number of input vectors used for training a particular map decreases to the subset of vectors mapped onto the respective upper-layer unit. Furthermore, each map in the hierarchy explains a particular set of characteristics of its input data [28]. A neuron with a large mean quantization error, i.e. an error greater than a percentage of ${mqe}_{0}$ , will expand to a new SOM in the next layer. The percentage is denoted by $\tau _{\mu }$ , and a large $\mathrm {\tau }_{\mathrm {\mu }}$ accepts more tolerance which produces less layers and vice versa. When neuron $i$ is selected to be expanded, the expanded SOM is trained using those vectors labelled to $i$ . The $\tau _{\mu }$ limits the depth of map, and the global termination criterion is set as ${mqe}_{i}<\tau _{u}\cdot {mqe}_{0}$ . Each new SOM could grow in the way described in Step 2.

• Step 4 (Repeat Step 3 Until No Neuron in All Maps of a Layer Needs Expansion):

  GHSOM complies with the LabelSOM [30] technique to label each neuron with some important keywords. In this work, we define an observation vector of energy consumption, $\boldsymbol {w}^{\boldsymbol {k}}$ and identify the trip of those input vectors labelled to this neuron. Thus, a neuron in the hierarchy will be labelled by a set of trips as well as a set of driving pattern.

FIGURE 9.

The insertion of neurons.

SECTION IV. Results and Discussion

A. Data Source

We collected two levels of data to test our framework, including battery and vehicle levels. First, a 106Ah2S battery module of a production EV was evaluated to model the aging trend and long-term performance of the EV battery pack. Table 1 denotes the basic information on vehicle type, cell, module, pack, and test environment. Moreover, we collected the operating log of two EVs for a year and reorganized the trip data by the process mentioned in section 3.3.

TABLE 1 Vehicle type, battery specifications and test environment.

1) Battery Test Data:

We used 1C charging and discharging rate to accelerate the cycle-life test process. In order to control the consistency of a battery pack, the capacity range of cells is limited in 30mAh (±15mAh), and the maximum variance of capacity among 50 modules is only 0.57% (30mAh/5.3Ah). The DCIR range of modules is limited in 0.3m$\Omega$ ($\pm 0.15\text{m}\Omega )$ . In [31] and [32], the authors presented a performance analysis for lithium storage technologies in an electric vehicle system. These studies provide a guideline for lab test before designing a high voltage and complex system. Moreover, many automakers follow the battery test manual for plug-in hybrid electric vehicles to test the selected cell or module before assembling as a pack [23], which provide a comprehensive procedure for EV pack test. In this work, we focus on the SOH simulation of battery module which has completed the necessary test in battery test manual [23]. We performed an experiment for 9 months, and over 23,000,000 test records for 1635 cycles were collected as shown in Fig. 10. Three different charge/discharge C-rates (including 1C, 0.5C and 0.2C) were collected, and the test temperature was well controlled in suitable temperature (15°C to 35°C). The purpose is to verify the cycle-life performance of battery pack based on 8 year operation.

FIGURE 10.

Cycle-life test data of 106Ah2S battery module for 1635 cycles.

2) EV Operating Data:

We collected 787 records (sampling rate = 1/10 second) from a pure EVs in Taiwan over one year. After screening out noisy data, there are 211 valid trips for the tested EV, and the average data length of a single trip is around 800 Kbytes. The raw data were transformed into recognized pattern as shown in Fig. 7. The big data generated by the internet of vehicle will grow tremendously when more and more EVs connect to the data cloud.

B. Estimation of Long-Term Battery Performance

An EV is almost impossible to do a full-charge/discharge in every cycle. In practice, an EV maker can only use the accumulative discharging energy to estimate the long-term performance of pack. This result is a baseline to estimate the SOH and remaining capacity of faded battery pack after long-term operation. Although the performance of different kinds of Li-ion batteries varies significantly, an automaker would apply this approach to evaluate and decide the suitable battery cell that matches the overall requirement of the powertrain design.

1) Estimation of DCIR:

As shown in (2) and Fig. 2, we rate the DCIR by checking the 0.5C discharged at 7.3V, and the fitting curve is shown in Fig. 11. According to the cycle-test data, the average discharging energy of each cycle is 571W, and we can find the relation between DCIR and number of cycling is

$$\begin{align*} \mathrm {r}=&{6.188\times {10}^{-10}x^{3}-2.575{\times 10}^{-7}x}^{2} \\&+\,5.618{\times 10}^{-5}x+3.104 \tag{21}\end{align*}$$

FIGURE 11.

Fitting curve of DCIR for a 106Ah2S battery module.

2) Estimation of Energy Decay:

Because of the percentage of capacity maps to the OCV consistently, a BMS can check the relative SOC% by seeking the preinstalled mapping table. However, the OCV table can only check the percentage of SOC but remaining capacity of each cycle. In this work, we applied the nonlinear fitting model to fit OCV-capacity curve. The comparison of OCV and fitting curve is shown in Fig. 12(a), and the nonlinear fitting result is shown in Fig. 12(b).

FIGURE 12.

(a) OCV curve and Fitting curve. (b) Result in seven polynomial fits.

The capacity fade and the resistance growth do not depend on the same conditions, and this makes the ageing comprehension a difficult task. In [33], Barré, et al. presented a review of the battery ageing mechanisms, and their consequences, occurring during a battery life. However, the loss of active material or self-discharge rate is hard to detect while an EV is in use. Whatever the factor, the usable energy will be diminished, which causes the bad cycle-life of a battery. The error range depends on the balancing condition of capacity and temperature. Under normal conditions, most of the drivers charge the pack before exhausting the energy or limp mode, that is, the pack theoretically has better calendar life than the result of cycle-life test.

Fig. 12(a) illustrates the first discharging cycle of a fresh battery module, and we fitted all discharging curves in Fig. 10 to simulate the degrading rate of the battery. Because of the BMS calculates the SOC by evaluating the OCV of a battery, we have to adjust the discharging curve of cycle-life test to suit the OCV curve. The estimated discharging capacity and its OCV of 1635 cycles could be reformatted as shown in Fig. 13. These curves also denote the recovery capacity for each cycle. For a 106Ah100S pack, the estimated energy of the first cycle is 34.5kWh, the average energy of a cycle is 28.55kWh and the end of life (EOL) energy is 27.6 kWh when the capacity is less than 80% of initial capacity (or after 1600 cycles). By combining the estimated results of DCIR and capacity, an automaker could estimate the long-term performance of the battery pack. For instance, the estimated DCIR at 1000 cycle is around 3.5m$\Omega$ , and the estimated discharging energy is around 30.015kWh.

FIGURE 13.

Reformatted OCV-capacity curves of cycle-life test data.

Moreover, according to (9), we summarized the estimated energy efficiency of each cycle as shown in Fig. 14. The percentage of capacity efficiency is worked out from the ratio of discharging capacity to rated capacity. Once the DCIR has increased, the discharging capacity will be narrowed. For example, the initial energy is 35.85kWh, the end-of-life energy is 29.2kWh, and the estimated total discharging energy is around 52040kWh for 1600 cycles. As shown in Fig. 4, the resistance of a lithium-ion battery is inversely proportional to the temperature, and thus a BMS can evaluate the usable capacity by checking the SOH and battery temperature for every operating cycle. The discharging efficiency at 1000 cycle is 87% and the battery temperature is 0°C, and the estimating energy is around 27.45kWh (35.85kW*0.87*0.88).

FIGURE 14.

Estimation of discharging efficiency for long-term battery performance.

C. Learning Driver’s Driving Patterns and Range Estimation

We used the GHSOM algorithm developed by Rauber’s team [34] to train the energy dataset vectors. The GHSOM initially contains a $2\times 2$ map in layer 1. Parameters $\tau _{m}$ and $\tau _{u}$ are set to 0.18 and 0.02, and detail power train parameters are listed in Table 2. The clustered result of the tested EV is shown in Fig. 15. The central map is the distributions of 4 groups, and the driving curves of each data node in each clustered group are also presented. The learning process stopped at the layer 1 due to no neuron in all maps needs expansion.

TABLE 2 Parameters of EV power train.

FIGURE 15.

Clustered result of driving patterns.

We normalized the clustered patterns of each neuron as a driving pattern, and calculated the energy efficiency of each group (neuron) by comparing the driving range and its energy consumption as shown in Fig. 16. The EV cloud platform could implement a pattern classifier by learning the clustered results from GHSOM, and the classifier will select a mapped distribution of energy for energy estimation next time. Moreover, the platform will update the GHSOM and retrain the classifier when next time the EV uploads a new driving pattern. For instance, the distribution of efficiency gap between speed and energy of neuron 2 is shown in Fig. 17. The percentage of driving time at the speed from 60km/h to 70km/h is almost same as the percentage of energy consumption but get worse from 70km/h to 90km/h. According to the driving pattern, we can adjust the power consumption of the speed as

$$\begin{equation*} p_{v_{e}^{d}}\ge \frac {v_{e}}{3600\cdot \eta _{t}}\left ({ m\cdot g\cdot f_{r}+\frac {C_{d}\cdot A\cdot {v}_{e}^{2}}{21.5} }\right )\cdot \eta _{v}^{d} \tag{22}\end{equation*}$$ where $\eta _{v}^{d}=(1+\mathrm {Efficient~Gap)}$ is the driving efficiency of driver d at different speeds. Moreover, the driving range can be computed by mapping the data in Fig. 13 to simulate the driving range for long-term battery degradation.

FIGURE 16.

Curves of normalized driving patterns.

FIGURE 17.

Curves of normalized driving patterns.

D. Comparison of Road Test and Estimation Model

In order to verify the proposed approach, we compared the power match rule among (13) [27], [35], our estimation model as (22) and road test results. The test route is shown in Fig. 18, and the estimated range of a round trip is 60.2 km. We fully charged the battery pack to SOC 100% and recorded the energy consumption of road test in next charge, and the comparison results are listed in Table 3.

TABLE 3 Comparison of two estimation models and road test.

FIGURE 18.

Analysis of road test.

E. Discussion of Experimental Results

We applied real test data collected from a production EV to evaluate the remaining driving range with respect to the energy consumption model of each driver. With the varied driving patterns, the performance of an EV model would be quite different. Many of the public information including 3D maps, real-time traffic report and climate could help to improve the accuracy of range estimation. However, the battery performance and driving behavior are mainly affected by the user. For instance, if the battery pack of an EV is always operated with high charging/discharging C-rate, the aging speed of the battery will be quicker than other battery packs with normal charging/discharging C-rate while at the same battery cycle shorter driving range of that EV can be expected.

The learning process of GHSOM stopped at the layer 1. Two possible reasons might explain this situation; 1) no enough patterns to be expanded to deeper layers, and 2) driver’s driving behavior is very regular. Along with collecting more driving data from the EV cloud, we hope to see as many driving behaviors as possible to extract the potential relationships among driving behavior, energy efficiency, vehicle performance and environmental factors. Different from ICE vehicles, the motor controller can control the output power for better driving experiences. For instance, an EV can extend the driving range by limiting unnecessary acceleration. An EV can also estimate the driving range more precisely and adjust the control policy dynamically if an EV cloud system could analyze the driving modes of different users such as high speed, urban, heavy load or mountain mode.

1) Comparison of Battery Degradation Between Cell and Module:

The capacity degradation and the resistance growth may not depend on the same conditions, and this makes the ageing comprehension a difficult task. However, the loss of active material or self-discharge rate is hard to detect while an EV is in use. In practice, the BMS would estimate the DCIR and recovery of capacity when charging or regular maintenance. As shown in Fig. 19, we reviewed the correlation between capacity degradation and resistance growth of in cell level with different voltage ranges. The result showed that both the cell level and the module level have same trend. However, it is not supposed to calculate the internal resistance of a module from a single cell directly because both the welding resistance and inconsistence among cells in a module need to be evaluated. In order to improve the accuracy, we completed a module level test to simulate the long-term performance of pack in section 4.2.

FIGURE 19.

Cycle-life test for battery cells with different voltage ranges.

2) Estimation of Driving Range:

We would estimate the driving range by following steps according to the experimental results:

• Step 1 (Checking Route):

  A driver could arrange the route for a trip by GPS navigator or internet map, e.g. Google maps, and the web application programming interface (API) will feedback an estimated route with current traffic condition and the speed limit of each segment. Then, the estimated route will be classified by the model classifier to choose a proper distribution of energy efficiency.

• Step 2 (Estimating Battery Degradation):

  The EV could evaluate the energy retention by checking the DCIR recorded in previous trips. For instance, the average DCIR of last 3 trips is 3.5m$\Omega$ , and the estimated cycle of battery is 1000 referring to Fig. 11 and (22). So the remaining energy is around 87.6%, or 31.4kWh, of the initial energy.

• Step 3 (Arranging Driving Distance):

  Assuming the HVAC is 0.5kWh, we can simulate the energy consumption as (21), and the estimated range is calculated by (15) and (16). The EV system then can be mapped with the traffic report information, e.g. average speed at each section of the route, and the possible driving range can be calculated based on current battery performance. For instance, if the driver is on the highway with a constant speed of 100 km/h, the estimated driving range will be $(31.4kWh\times 90\% )\div (24.15kW\times 1.07)\times 100km/h=116.22km/h$ when remaining energy is at 90% SOC.

SECTION V. Conclusions

With this concept, we can cope with the range anxiety when more and more driving data are processed. For instance, when more and more EVs are selling to the market, we can create an energy heat map for every section on the map of each EV model. Every time when a driver is routing a trip, the EV cloud computing system could estimate the driving range by combining driver’s pattern and battery aging model collected from EV automaker.

An EV-battery big data modeling method for improving driving range estimation was proposed in this paper. This method provides a feasible process of battery life estimation in a production EV, taking into account the dependence of internal resistance and SOH. Moreover, the GHSOM approach was applied to cluster driving patterns collected from EV cloud platform for modeling the energy consumption of EVs. The experimental results lead to the following conclusions:

• Comparing to the specification of a single cell, the cycle-life test result of a large format module provides an intuitive and accurate simulation of SOH for a production EV.

• The essential driving patterns clustered by GHSOM carried out on an adjusted energy consumption model showed that the machine learning approach can offer gradual driving range estimation.

• By combining the estimating process of SOH and driving pattern extracting approach, an automaker can design a predicable model based on the EV big data to project the driving range in the useful life of EV battery.

For future research, the approach will be extended to the grid level for finding the relationships among battery, vehicle, charging/swapping stations and O2O (online to offline) applications.