基于深度学习的纯电动汽车经济驾驶系统

Deep Learning–based Eco-driving System for Battery Electric Vehicles

基于深度学习的纯电动汽车经济驾驶系统

A National Center for Sustainable Transportation Research Report
国家可持续交通中心研究报告，加州大学

EXECUTIVE SUMMARY

The uninterrupted growth in transportation activities, for both people and goods movement, has been exerting significant pressure on our socio-economics and environment. However, emerging technologies such as connected and automated vehicles (CAVs), transportation electrification, and edge computing have been stimulating increased efforts by engineers, researchers, and policymakers to tackle transportation-related problems, including those focused on energy and the environment. The eco-driving strategies based on CAV technology particularly have attracted significant interest from all over the world due to its potential to save energy and reduce tail-pipe emissions. Among all CAV based eco-driving strategies, the Eco-Approach and Departure (EAD) application at Signalized Intersections has shown the most significant promise. In this system, an equipped vehicle can take advantage of the signal phase and timing (SPaT) and geometric intersection description (GID) information from the upcoming signalized intersection and calculate the optimal speed to pass on a green light or to decelerate to a stop in the most eco-friendly manner. Speed recommendations may be provided to the driver using a driver-vehicle-interface (DVI) or to the vehicle systems that support automated longitudinal control capabilities.
人员和货物运输的运输活动的不间断增长给我们的社会经济和环境带来了巨大的压力。 然而，联网和自动驾驶汽车 (CAV)、交通电气化和边缘计算等新兴技术一直在刺激工程师、研究人员和政策制定者加大力度解决交通相关问题，包括那些关注能源和环境的问题。 基于 CAV 技术的环保驾驶策略因其节省能源和减少尾气排放的潜力而引起了世界各地的极大兴趣。 在所有基于 CAV 的环保驾驶策略中，信号交叉口的环保进场和离场 (EAD) 应用显示出最显着的前景。 在该系统中，配备的车辆可以利用即将到来的信号交叉口的信号相位和定时（SPaT）以及几何交叉口描述（GID）信息，并计算通过绿灯或减速停车的最佳速度。 最环保的方式。 可以使用驾驶员车辆接口（DVI）向驾驶员或支持自动纵向控制功能的车辆系统提供速度建议。
In this project, the research team conducted a thorough literature review of EAD algorithms, and identified a major research gaps in the following areas: (1) the balance between system optimality and computational efficiency; (2) designated algorithms for electric vehicles (e.g., consideration of regenerative braking); and (3) taking into account downstream traffic information (e.g., prediction of preceding vehicle’s state). To address these gaps, the research team proposed a deep learning–based trajectory-planning algorithm (DLTPA) for EAD application, which can be considered as an approximation of a global optimal algorithm (called a graph-based trajectory planning algorithm or GTPA) that the research team previously developed. The proposed DLTPA has two processes: offline (training) and online (implementation), and it is composed of two major modules: 1) a solution feasibility checker that identifies whether there is a feasible trajectory subject to all the system constraints, e.g., maximum acceleration or deceleration; and 2) a regressor to predict the speed of the next time step.
在本项目中，研究团队对EAD算法进行了全面的文献回顾，发现了以下几个方面的主要研究空白：（1）系统最优性和计算效率之间的平衡； (2) 电动汽车的指定算法（例如考虑再生制动）； （3）考虑下游交通信息（例如，预测前车状态）。 为了解决这些差距，研究团队提出了一种用于 EAD 应用的基于深度学习的轨迹规划算法（DLTPA），该算法可以被视为全局最优算法（称为基于图的轨迹规划算法或 GTPA）的近似 研究小组之前开发的。 所提出的 DLTPA 有两个过程：离线（训练）和在线（实现），它由两个主要模块组成：1）解决方案可行性检查器，用于识别是否存在受所有系统约束（例如最大约束）的可行轨迹。 加速或减速； 2）一个回归器来预测下一个时间步的速度。
Preliminary simulation study in microscopic traffic modeling software PTV VISSIM showed that the proposed DLTPA can achieve a great balance of energy savings vs. computational efforts when compared to the baseline scenario where no EAD was implemented and the optimal solution (in terms of energy savings) was provided by GTPA.
微观交通建模软件 PTV VISSIM 中的初步模拟研究表明，与未实施 EAD 且最优解决方案（在节能方面）为 由 GTPA 提供。

Introduction

The uninterrupted growth in transportation activities, for both people and goods movement, has exerted significant pressure on our socio-economics and environment. The transportation sector in the United States consumed approximated 27.5 quadrillion BTUs (British thermal unit) of energy in 2016, 92.2% percent of which came from petroleum [1]. In addition, the latest annual report by the U.S. Environmental Protection Agency (USEPA) estimated that surface transportation modes (such as passenger cars, trucks, buses, and motorcycles) contributed 1,556 MMT CO2eq to greenhouse gas (GHG) emissions in 2016, accounting for 28.5% of nationwide GHG emissions [2]. According to the same report, transportation just slightly surpassed the electric power industry (28.4%) and became the largest source of GHG across all U.S. economic sectors in 2016.
人员和货物运输的运输活动的不间断增长给我们的社会经济和环境带来了巨大的压力。 2016年，美国交通运输部门消耗了约27.5万亿BTU（英国热量单位）能源，其中92.2%来自石油[1]。 此外，美国环境保护署（USEPA）最新年度报告估计，2016年地面交通方式（如乘用车、卡车、公共汽车和摩托车）的温室气体（GHG）排放量为1,556 MMT CO2eq，占 占全国温室气体排放量的28.5%[2]。 根据同一份报告，2016年交通运输业略超过电力行业（28.4%），成为美国所有经济部门最大的温室气体来源。
On the other hand, emerging technologies such as connected vehicles (CV), transportation electrification, and edge computing have stimulated increased efforts from engineers, researchers and policymakers to tackle transportation-related energy and environmental problems. Good examples include the Applications for the Environment: Real-Time Information Synthesis (AERIS) Program initiated by the U.S. Department of Transportation [3], and the eCoMove project funded by the European Commission [4]. A variety of environmentallyfriendly CV applications, in particular those related to eco-driving strategies, have been proposed, developed, and validated [5]. Among all environmentally-friendly eco-driving strategies, the Eco-Approach and Departure (EAD) at Signalized Intersections application has shown significant promise [6–10]. In this system, a vehicle can take advantage of the signal phase and timing (SPaT) and geometric intersection description (GID) information from the upcoming signalized intersection and calculate the optimal speed to pass on a green light or to decelerate to a stop in the most eco-friendly manner. Speed recommendations may be provided to the driver using a driver-vehicle-interface (DVI) or to the vehicle systems that support automated longitudinal control capabilities.
另一方面，互联汽车（CV）、交通电气化和边缘计算等新兴技术刺激了工程师、研究人员和政策制定者加大力度解决与交通相关的能源和环境问题。 很好的例子包括美国交通部发起的环境应用：实时信息综合（AERIS）计划[3]，以及欧盟委员会资助的eCoMove项目[4]。 各种环保的CV应用，特别是与生态驾驶策略相关的应用，已经被提出、开发和验证[5]。 在所有环保的生态驾驶策略中，信号交叉口的生态进场和离场（EAD）应用已显示出巨大的前景[6-10]。 在该系统中，车辆可以利用来自即将到来的信号交叉口的信号相位和定时（SPaT）以及几何交叉口描述（GID）信息，并计算通过绿灯或减速停车的最佳速度。 最环保的方式。 可以使用驾驶员车辆接口（DVI）向驾驶员或支持自动纵向控制功能的车辆系统提供速度建议。
Due to the benefits of EAD-like eco-driving algorithms, numerous studies have focused on their development and testing [11–20]. However, many of these algorithms are not flexible enough to effectively handle customized powertrain characteristics, interaction with other traffic, road grade, and travel through multiple intersections [33]. This project aims to address some of these gaps, and the uniqueness of this research includes:
由于类似 EAD 的生态驾驶算法的优点，许多研究都集中在其开发和测试上 [11-20]。 然而，其中许多算法不够灵活，无法有效处理定制的动力系统特性、与其他交通的交互、道路坡度以及穿过多个交叉路口的行驶[33]。 该项目旨在解决其中一些差距，这项研究的独特性包括：

• Customized electric powertrain. Based on real world data, an electric vehicle (EV) energy consumption model is developed and integrated into a new eco-driving algorithm and the regenerative braking effect is taken into account.
• 定义电动动力总成。 基于现实世界数据，开发了电动汽车（EV）能耗模型，并将其集成到新的节能驾驶算法中，并考虑了再生制动效应。
• Prediction of downstream vehicle’s trajectory. Machine learning technique is applied to a snippet of a vehicle’s downstream trajectory (which may be obtained from an onboard sensor, such as radar) to predict its movement (e.g., stopping, acceleration, cruising). This information may help the vehicle better plan its trajectory for saving energy.
• 下游车辆轨迹预测。 机器学习技术应用于车辆下游轨迹的片段（可以从雷达等车载传感器获得）来预测其运动（例如停止、加速、巡航）。 这些信息可以帮助车辆更好地规划其轨迹以节省能源。
• Deep learning–based EAD algorithm. This algorithm can achieve a balance between solution optimality and computational efficiency.
• 基于深度学习的EAD算法。 该算法能够实现解最优性和计算效率之间的平衡。

Literature Review

In this section, we first review previous research on Eco-Approach and Departure (EAD) applications and then give a brief introduction on the powertrain model used for fuel/energy consumption estimation in this study.
在本节中，我们首先回顾了之前关于经济接近和远离（EAD）应用的研究，然后简要介绍了本研究中用于燃料/能耗估计的动力系统模型。

The State-of-the-Art on Eco-Approach and Departure

In the past decade, a variety of studies have been conducted on EAD, especially from the perspective of an isolated intersection. Mandava et al. [11] proposed a piecewise lineartrigonometric function–based vehicle trajectory planning algorithm for eco-driving along an urban arterial road. The algorithm was extensively evaluated and validated in simulations [21] and field testing [22], in the form of an advanced driver assistance system [23] and partially automated control [24]. It showed excellent real-time performance and substantial benefits in reducing fuel consumption and tailpipe emissions. However, significant efforts may be necessary to modify the algorithm to adapt it to customized powertrain models and rolling terrain. Based on the VT-Micro1 model, Rakha and Kamalanathsharma [13] developed a constant deceleration based eco-driving strategy to avoid full stops at signals. They later improved upon this, using multi-stage dynamic programming and recursive path-finding principles, as well as evaluation with an agent-based model [25]. Asadi and Vahidi [14] proposed a two-step predictive cruise control concept, aiming to reduce fuel use and trip time by using traffic signal status information. The first step is to determine the target speed based on an available green window, and the second step is to perform the optimal tracking of the target speed. Katsaros et al. [15] developed a Green Light Optimized Speed Advisory (GLOSA) system designed to minimize average fuel consumption and average stop delay at a traffic signal. By taking into account the queue discharging process, Chen et al. [16] developed an ecodriving algorithm for a vehicle approaching and leaving a signalized intersection to minimize a linear combination of emissions and travel time, without taking into account roadway grade information. Jin et al. [17] formulated the power-based optimal longitudinal connected ecodriving problem into a 0-1 Binary Mixed Integer Linear Programming (MILP), which is applicable to signalized intersections, non-signalized intersections, and freeways. The approach can take into account road grade effects and powertrain dynamics, but has relatively low computational efficiency. Li et al. [18] used the Legendre Pseudo-Spectral method and knotting technique to overcome the discrete gear ratio issue in the optimal control for eco-driving at signalized intersections. Huang and Peng [19] adopted a simplified powertrain model and applied the Sequential Convex Optimization approach to optimize vehicle speed trajectory at signalized intersections, which aimed to keep a balance between the optimality and real-time performance.
在过去的十年中，人们对 EAD 进行了各种各样的研究，特别是从孤立交叉口的角度进行了研究。 曼达瓦等人。 [11]提出了一种基于分段线性三角函数的车辆轨迹规划算法，用于城市主干道的经济驾驶。 该算法以先进的驾驶员辅助系统[23]和部分自动化控制[24]的形式在模拟[21]和现场测试[22]中进行了广泛的评估和验证。 它显示出出色的实时性能，并在降低油耗和尾气排放方面具有显着的效益。 然而，可能需要付出巨大的努力来修改算法以使其适应定制的动力系统模型和起伏的地形。 基于 VT-Micro1 模型，Rakha 和 Kamalanathsharma [13] 开发了一种基于恒定减速的经济驾驶策略，以避免在信号灯处完全停车。 他们后来对此进行了改进，使用多阶段动态规划和递归寻路原理，以及基于代理的模型进行评估[25]。 Asadi 和 Vahidi [14] 提出了一种两步预测巡航控制概念，旨在通过使用交通信号状态信息来减少燃料使用和行程时间。 第一步是根据可用的绿色窗口确定目标速度，第二步是对目标速度进行最优跟踪。 卡萨罗斯等人。 [15]开发了绿灯优化速度咨询（GLOSA）系统，旨在最大限度地减少交通信号灯处的平均油耗和平均停车延迟。 通过考虑队列卸载过程，Chen 等人。 [16] 为车辆接近和离开信号交叉口开发了一种经济驾驶算法，以最大限度地减少排放和行驶时间的线性组合，而不考虑道路坡度信息。 金等人。 [17]将基于功率的最优纵向连通经济驾驶问题表述为0-1二元混合整数线性规划（MILP），适用于信号交叉口、非信号交叉口和高速公路。 该方法可以考虑道路坡度效应和动力总成动力学，但计算效率相对较低。 李等人。 [18]利用勒让德伪谱方法和knoting技术克服了信号交叉口环保驾驶优化控制中的离散齿轮比问题。 Huang和Peng[19]采用简化的动力系统模型，并应用顺序凸优化方法来优化信号交叉口的车辆速度轨迹，旨在保持最优性和实时性之间的平衡。
When considering the application of an Eco-Approach and Departure system in a more realistic environment, many studies took a “reactive” approach to cope with the disturbance from the downstream traffic (e.g., switching to the car-following mode control if the subject vehicle was too close to its predecessor) or assumed traffic signals were running in a fixed-time mode [20, 26, 27]. To address these issues, some researchers specifically focused on tackling the queuing effects for Eco-Approach and Departure at Signalized Intersections (EADSI) by applying the shockwave theory [28] or data-driven techniques [29] to predict the queue length or, in essence, the trajectory of the subject vehicle’s predecessor. Other approaches were dedicated to dealing with uncertainties in traffic signal operation such as countdown information by improving the prediction of SPaT [30] or developing more robust eco-driving strategies [31, 32].
当考虑在更现实的环境中应用经济接近和驶离系统时，许多研究采取了“反应式”方法来应对下游交通的干扰（例如，如果目标车辆 与其前身太接近）或假设交通信号以固定时间模式运行 [20,26,27]。 为了解决这些问题，一些研究人员专门致力于通过应用冲击波理论 [28] 或数据驱动技术 [29] 来预测队列长度，从而解决信号交叉口经济接近和驶离 (EADSI) 的排队效应。 本质上，目标车辆前身的轨迹。 其他方法致力于通过改进 SPaT 的预测 [30] 或开发更强大的经济驾驶策略 [31, 32] 来处理交通信号操作中的不确定性，例如倒计时信息。

Estimation of Electric Vehicle Energy Consumption

Recently, a good deal of effort has been devoted to developing estimation models for energy consumption in electric vehicles [34–42]. Several studies established EV energy consumption estimation models at different granularity for the purpose of eco-routing applications [42], assessment of different aggregation level influence on energy consumption [36], and ecodriving on urban arterial roads [35].
最近，人们致力于开发电动汽车能耗估算模型[34-42]。 一些研究建立了不同粒度的电动汽车能耗估算模型，用于经济规划路线应用[42]、评估不同聚合水平对能耗的影响[36]以及城市主干道的经济驾驶[35]。
From the perspective of modeling methodology, the knowledge-driven (“white-box”) approach either considers the vehicle as a point-mass by applying Newton’s Laws (analytical model) or builds up a detailed physical process for each module in an electric vehicle [34, 38, 40]. This approach has the advantage of providing a direct link between vehicle or power train dynamics and the variables affecting energy consumption. However, the knowledge-driven models may be too generic without differentiating the powertrain type or overly complex in real-time implementation. For example, Wu et al. [40] used real-world measurements and established an instantaneous EV energy consumption model directly derived from the vehicle dynamics. All the parameters in their model are reduced to predetermined constants. In contrast, the datadriven (“black-box”) approach applies statistical techniques [35, 39] or machine learning algorithms [37] to the dataset collected from a vehicle test bed or real-world driving. This approach may result in very accurate estimation results based on a customized dataset or a specific scenario. However, the applicability to another situation is questionable. In addition, the physical meanings of selected variables and interpretation of such models are not justifiable. Yao et al. [39] proposed a purely statistical model in which variables include a complete list of combinations of speed and acceleration up to the third order. The meaning of some variables is vague. Further, the results are validated only with data collected in an ideal environment without considering road grade effects. Table 1 lists some related work by other researchers on the EV energy consumption estimation.
从建模方法的角度来看，知识驱动（“白盒”）方法要么通过应用牛顿定律（分析模型）将车辆视为质点，要么为电动汽车中的每个模块建立详细的物理过程 [34、38、40]。 这种方法的优点是在车辆或动力传动系动力学与影响能耗的变量之间提供直接联系。 然而，知识驱动模型可能过于通用，没有区分动力系统类型，或者在实时实现中过于复杂。 例如，吴等人。 [40]使用现实世界的测量结果并建立了直接从车辆动力学导出的瞬时电动汽车能耗模型。 他们模型中的所有参数都减少到预定常数。 相比之下，数据驱动（“黑匣子”）方法将统计技术 [35, 39] 或机器学习算法 [37] 应用于从车辆测试台或现实驾驶中收集的数据集。 这种方法可能会根据定制的数据集或特定场景产生非常准确的估计结果。 然而，对于另一种情况的适用性值得怀疑。 此外，所选变量的物理意义和此类模型的解释是不合理的。 姚等人。 [39]提出了一个纯粹的统计模型，其中变量包括直至三阶的速度和加速度组合的完整列表。 有些变量的含义是模糊的。 此外，仅使用在理想环境中收集的数据来验证结果，而不考虑道路坡度影响。 表1 列出了其他研究人员在电动汽车能耗估算方面的一些相关工作。

Vehicle Movement Prediction

Accurate and reliable prediction of vehicle speed trajectory is an important component in many Intelligent Transportation Systems (ITS) applications, particularly for safety and environmental related applications. Making such predictions is a challenging task, as the vehicle speed trajectory may be affected by various dynamic factors, e.g., signal status, maneuvers of surrounding vehicles, and perhaps interruption from pedestrians. In the literature, various approaches for vehicle speed prediction have been investigated and evaluated [43–51]. In general, the existing vehicle speed prediction strategies can be categorized into two major classes: model-based approaches and data-driven approaches. The model-based approaches predict the vehicle speed trajectory based on pre-defined model structures such as Constant Speed Model, Constant Acceleration Model, Constant Yaw Rate and Acceleration Model [43].
准确可靠的车辆速度轨迹预测是许多智能交通系统 (ITS) 应用的重要组成部分，特别是对于安全和环境相关的应用。 做出这样的预测是一项具有挑战性的任务，因为车辆速度轨迹可能受到各种动态因素的影响，例如信号状态、周围车辆的操纵以及行人的干扰。 在文献中，已经研究和评估了各种车辆速度预测方法[43-51]。 一般来说，现有的车速预测策略可以分为两大类：基于模型的方法和数据驱动的方法。 基于模型的方法根据预定义的模型结构（例如恒定速度模型、恒定加速度模型、恒定偏航率和加速度模型）来预测车辆速度轨迹[43]。
However, the underlying dynamics of human cognition, decision-making, and execution of drivers and vehicle systems are extremely complex and these simplified models may not be applicable [44]. On the other hand, data-driven approaches have recently been well investigated since they show more flexibility and applicability in representing system dynamics. Good examples of effective data-driven approaches for vehicle speed trajectory prediction include Non-Parametric Regression, Gaussian Mixture Regression and Artificial Neural Networks [45– 48]. In a report by Houenou et al. [46], the defined maneuver recognition algorithm selected the best vehicle trajectory that would minimize a cost function by comparing the current maneuver to the pre-defined trajectory set on the highways. Considering the requirement for a large sample of vehicle trajectories and the complexity of maneuver recognition in urban areas, applying this algorithm real-world urban traffic is challenging. Gaussian Mixture Regression is another promising parametric method to approximate or predict vehicle trajectories by calculating a conditional probability density function that consists of a weighted linear combination of Gaussian component densities [47]. Artificial Neural Networks have been proven to be an effective method for accurately forecasting vehicle speed and position, due to their strong capability of capturing complex and nonlinear dynamics [48– 50]. A comparative study of major parametric and non-parametric approaches for vehicle speed prediction on highways indicates that Artificial Neural Networks outperform all the other methods in terms of both predictive accuracy and applicability [48]. Some approaches (i.e., TrackT [51] and TMicroscope [52]) have been proposed to enhance tracking of RFID systems to retrieve trajectory information. These approaches could provide real time trajectory information with high accuracy and be combined with advanced predictors to improve the overall performance.
然而，人类认知、决策以及驾驶员和车辆系统执行的基本动态极其复杂，这些简化模型可能不适用[44]。 另一方面，数据驱动的方法最近得到了很好的研究，因为它们在表示系统动态方面表现出更大的灵活性和适用性。 用于车辆速度轨迹预测的有效数据驱动方法的好例子包括非参数回归、高斯混合回归和人工神经网络[45-48]。 在 Houenou 等人的一份报告中。 [46]，定义的机动识别算法通过将当前机动与高速公路上设置的预定义轨迹进行比较来选择最佳车辆轨迹，从而最小化成本函数。 考虑到对车辆轨迹的大样本要求以及城市地区机动识别的复杂性，将该算法应用于现实城市交通具有挑战性。 高斯混合回归是另一种有前途的参数方法，通过计算由高斯分量密度的加权线性组合组成的条件概率密度函数来近似或预测车辆轨迹[47]。 人工神经网络已被证明是准确预测车辆速度和位置的有效方法，因为它们具有捕获复杂和非线性动态的强大能力[48-50]。 对高速公路车速预测的主要参数和非参数方法的比较研究表明，人工神经网络在预测准确性和适用性方面优于所有其他方法[48]。 已经提出了一些方法（即 TrackT [51] 和 TMicroscope [52]）来增强 RFID 系统的跟踪以检索轨迹信息。 这些方法可以提供高精度的实时轨迹信息，并与先进的预测器相结合以提高整体性能。

Deep Learning Approach and its Applications in Transportation

This section briefly reviews some recent publications on the application of deep-learning techniques to the transportation area.
本节简要回顾了最近一些关于深度学习技术在交通领域应用的出版物。
Increasing transportation efficiency reliably and at a low cost becomes challenging as transportation infrastructure becomes more complex. Deterministic logic is often impractical because of the complexity of the challenges in modern transportation networks. Deep learning shows promise for solving transportation problems because of its ability to learn non-linear functions. Studies have shown that transportation modes can be predicted using Long Short Term Memory and Convolutional Neural Networks [53], [54]. Further studies have shown that traffic flow can be accurately predicted with deep learning [55], [56]. Deep learning also has applications to mitigating calibration challenges. One study used reinforcement learning to achieve adaptive ramp metering without calibration [57]. Moreover, deep learning can be used to predict vehicle speeds. Network wide traffic speed predictions had a high accuracy when images were used as training data [58]. Lastly, deep learning shows promise for enabling autonomous vehicles. Results from empirical evaluations showed how deep learning can be used to perform lane and vehicle detection at speeds required for real world scenarios [59].
随着交通基础设施变得更加复杂，以低成本可靠地提高交通效率变得具有挑战性。 由于现代交通网络挑战的复杂性，确定性逻辑通常不切实际。 深度学习因其学习非线性函数的能力而显示出解决交通问题的前景。 研究表明，可以使用长短期记忆和卷积神经网络来预测交通方式[53]、[54]。 进一步的研究表明，通过深度学习可以准确预测交通流量[55]，[56]。 深度学习还可以用于缓解校准挑战。 一项研究使用强化学习来实现无需校准的自适应坡道计量[57]。 此外，深度学习可用于预测车速。 当使用图像作为训练数据时，网络范围内的交通速度预测具有很高的准确性[58]。 最后，深度学习显示出实现自动驾驶汽车的前景。 实证评估的结果表明，如何使用深度学习以现实场景所需的车道和车辆速度[59]。

A Hybrid Model for Electric Vehicle Energy Consumption

Data Acquisition and Processing

We developed our models for energy consumption rate estimation based on driving data collected from a test electric vehicle (2013 NISSAN LEAF) in real world traffic.
我们根据从测试电动汽车（2013 款日产聆风）在现实交通中收集的驾驶数据开发了能源消耗率估算模型。
Data Acquisition
The field data were collected during two periods: 1) March–July 2013; and 2) July 2018. The test EV was equipped with two major data acquisition systems: the CONSULT III plus kit and GPS data logger (see Figure 1). These were used to access vehicle activities (e.g., instantaneous speed), energy related parameters (e.g., battery current and voltage), and real-time location information (i.e., latitude, longitude).
现场数据收集于两个时期：1）2013 年 3 月至 7 月； 2) 2018 年 7 月。测试 EV 配备了两个主要数据采集系统：CONSULT III plus 套件和 GPS 数据记录器（见图 1）。 这些用于访问车辆活动（例如瞬时速度）、能源相关参数（例如电池电流和电压）以及实时位置信息（即纬度、经度）。

More specifically, the CONSULT III plus kit, which is designated for professional diagnostics of all NISSAN models (including NISSAN LEAF), can retrieve high resolution (up to 100 Hz) data from the vehicle’s CAN bus, such as speed, current and voltage for each cell, A/C power, accessory power, and state of charge. On the other hand, the GPS data logger can report the latitude and longitude of the test EV in real-time. Such information can be synchronized with existing geographic information system (GIS) to acquire the network-wide index and grade information of the road link on which the vehicle is traveling.
更具体地说，CONSULT III plus 套件专为所有 NISSAN 车型（包括 NISSAN LEAF）的专业诊断而设计，可以从车辆 CAN 总线检索高分辨率（高达 100 Hz）数据，例如速度、电流和电压。 每个电池、交流电源、配件电源和充电状态。 另一方面，GPS数据记录器可以实时报告测试电动汽车的纬度和经度。 这些信息可以与现有的地理信息系统（GIS）同步，获取车辆行驶路段的全网索引和等级信息。
We selected four typical routes near Riverside, California (USA) for real-world data collection: 1) SR 91-Magnolia loop; 2) Riverside Plaza-Towngate loop; 3) Columbia-Alessandro loop; and 4) Palmyrita Avenue close to CE-CERT. Figure 2 presents the SR 91-Magnolia loop which covers a major freeway (SR 91) and arterial road (Magnolia Ave.), and a variety of traffic conditions and road grades. In total, more than 100 hours of vehicle driving data under real-world conditions were collected.
我们选择了美国加利福尼亚州里弗赛德附近的 4 条典型路线进行实际数据收集：1）SR 91-Magnolia 环线； 2）河滨广场-Towngate环线； 3）哥伦比亚-亚历山德罗环路； 4) 靠近 CE-CERT 的 Palmyrita Avenue。 图 2 显示了 SR 91-Magnolia 环路，涵盖主要高速公路 (SR 91) 和主干道 (Magnolia Ave.)，以及各种交通状况和道路等级。 总共收集了超过 100 小时的真实条件下的车辆驾驶数据。

Data Processing
Before the application of the aforementioned field test data for model development, we first combined the dataset from the CONSULT III plus kit with that from the GPS data logger. This data fusion consists of two steps:
在应用上述现场测试数据进行模型开发之前，我们首先将 CONSULT III plus 套件中的数据集与 GPS 数据记录仪中的数据集结合起来。 该数据融合包括两个步骤：

1. Frequency realignment. The frequency of raw GPS data was realigned into a 1-Hz signal, which is consistent with the data resolution from the CONSULT III plus kit and suitable for energy consumption estimation; and
2. 频率调整。 原始 GPS 数据的频率被重新调整为 1 Hz 信号，这与 CONSULT III plus 套件的数据分辨率一致，适合能耗估算；
3. Trip start time synchronization. Unlike the GPS data logger, the CONSULT III plus kit uses a relative time stamp (i.e., each run always starts from time “0”) instead of a global time clock (i.e., Coordinated Universal Time). To synchronize these two datasets, we applied a cross-correlation technique on vehicle speed that was common to both data sources.
4. 行程开始时间同步。 与 GPS 数据记录器不同，CONSULT III plus 套件使用相对时间戳（即每次运行始终从时间“0”开始）而不是全球时钟（即协调世界时）。 为了同步这两个数据集，我们对两个数据源通用的车速应用了互相关技术。

Modeling Approach

Power Flows at the Battery Terminal
The energy consumption of the electric vehicle considered in this study is specified as an integration of output power of the vehicle at the battery terminal. The equations for electric power for propulsion and regenerative braking at the battery terminal are as follows.
本研究中考虑的电动汽车的能耗被指定为车辆在电池端子处的输出功率的积分。 电池端子处用于推进和再生制动的电力方程如下。

其中，${\eta }_t$ 代表传动效率，${\eta }_m$ 代表电机驱动效率。 这两个参数通常近似为常数值。 𝑚 是电动汽车的质量； 𝑓 为滚动阻力系数； g是引力常数； 𝜌 为空气密度（kg/m3）； $C_D$ 为气动阻力系数； $A_f$是电动汽车的正面区域； 𝛿 是与 EV 质量相关的系数； 𝑣 为车辆速度（m/s）； 𝛼 为道路坡度（rad）； 𝑘 (0 < 𝑘 < 1) 为再生制动系数，表示电动机可回收的总制动能量的百分比。 再生制动系数实际上是一个复杂且随时间变化的系数，将在后面的章节中详细讨论。
Thus, the total power flows at the battery terminal can be defined as:
因此，电池端子处的总功率流可以定义为：

In the real-world test data, the accessory power and A/C power were also measured, which turned out to be non-trivial and was therefore taken into account in the power estimation for the battery terminal.
在实际测试数据中，还测量了配件功率和 A/C 功率，结果证明这并不重要，因此在电池端子的功率估算中将其考虑在内。
Regenerative Braking Factor
Regenerative braking power is one of the most distinct features of electric vehicles and plays an important role in improving drivetrain efficiency. Compared to conventional internal combustion engine (ICE) vehicles, electric vehicles with regenerative braking have an advantage in energy efficiency, especially under stop-and-go driving scenarios. As mentioned in the above section, the regenerative braking factor, k, indicates the percentage of braking energy recovered back to charge the battery pack, implying the vehicle’s recharging efficiency. The value of k is between 0 and 1 (in practice, k cannot reach 1 due to energy lost from battery internal resistance and cable resistance). Due to the complexity and time varying character of the regenerative braking factor k, it is essential to identify the factors that affect k in order to model electric vehicles’ regenerative braking energy. According to a literature review, two major approaches have been used to model regenerative braking effects: a piecewise linear function of the vehicle’s speed and a fuzzy logic model considering acceleration, jerk, and road grade as input variables (Figure 3). The first approach was derived based on the assumption that the regenerative braking factor can be represented by braking force, which is supposed to be linearly related to the vehicle’s speed (see Eq. (4) in [39]). Please note that the equation is an approximation within a certain speed range (between 0 and 38 m/s). The second approach considers a more complex regenerative braking process with the measurement data available for many factors (e.g., speed, acceleration, road grade) as shown in Figure 3.
再生制动功率是电动汽车最显着的特征之一，在提高传动系统效率方面发挥着重要作用。 与传统内燃机（ICE）汽车相比，具有再生制动功能的电动汽车在能源效率方面具有优势，尤其是在走走停停的驾驶场景下。 如上节所述，再生制动系数 k 表示制动能量回收给电池组充电的百分比，反映了车辆的充电效率。 k的值在0和1之间（实际上，由于电池内阻和电缆电阻的能量损失，k不能达到1）。 由于再生制动系数 k 的复杂性和时变特性，有必要确定影响 k 的因素，以便对电动汽车的再生制动能量进行建模。 根据文献综述，有两种主要方法用于模拟再生制动效果：车辆速度的分段线性函数和考虑加速度、加加速度和道路坡度作为输入变量的模糊逻辑模型（图 3）。 第一种方法是基于再生制动系数可以用制动力来表示的假设推导的，制动力应该与车辆的速度线性相关（参见[39]中的方程（4））。 请注意，该方程是特定速度范围（0 到 38 m/s 之间）内的近似值。 第二种方法考虑更复杂的再生制动过程，具有许多因素（例如速度、加速度、道路坡度）的可用测量数据，如图 3 所示。


Hybrid Model for EV Energy Consumption Rate Estimation
In the proposed hybrid approach, the model variables are carefully selected based on the EV physical model instead of blindly exhausting a long list of variables and their combination. Then, a multi-linear regression (MLR) model is employed to calibrate the corresponding coefficients. The proposed approach excels in real-time performance but is more adaptive to different driving conditions without significantly compromising the model accuracy, when compared to a knowledge-based approach (e.g., [48]). The latter may require a large effort to measure the related parameters and calibrate the coefficients for the electric vehicle in a specific condition.
在所提出的混合方法中，模型变量是根据电动汽车物理模型精心选择的，而不是盲目地耗尽一长串变量及其组合。 然后，采用多元线性回归（MLR）模型来校准相应的系数。 与基于知识的方法（例如[48]）相比，所提出的方法在实时性能方面表现出色，但更能适应不同的驾驶条件，而不会显着影响模型的准确性。 后者可能需要花费大量的精力来测量电动汽车在特定条件下的相关参数并校准系数。
Based on the battery power equations (Eq. (1)-(3)) and the characteristics of regenerative braking systems discussed above, two types of “hybrid” regression models are proposed for EV energy consumption rate estimation. The major difference of these two models lies in the complexity of modeling the regenerative braking effects. For simplicity, we assume the transmission efficiency 𝜂𝑡 and motor efficiency 𝜂𝑚 as constants.
基于上述电池功率方程（方程（1）-（3））和再生制动系统的特性，提出了两种类型的“混合”回归模型用于电动汽车能耗率估计。 这两个模型的主要区别在于再生制动效应建模的复杂性。 为简单起见，我们假设传动效率 ${\eta }_t$ 和电机效率 ${\eta }_m$ 为常数。
A Type I hybrid energy consumption model is formulated in Eq. (5), which simply considers the regenerative braking factor as a linear function of the vehicle’s speed, or 𝑘 ∝ 𝑣. Therefore,
I 类混合能耗模型如式（5）所示。它简单地将再生制动系数视为车辆速度的线性函数，或 𝑘 ∝ 𝑣。 所以，

For comparison, we also applied the Type I hybrid energy consumption rate model to the subsets of data partitioned according to Eq. (4), i.e., 𝑣 < 5 𝑚/𝑠 and 𝑣 ≥ 5 𝑚/𝑠, and calibrated the coefficients, respectively. However, further investigation is needed to evaluate the impacts of the speed threshold.
为了进行比较，我们还将 I 类混合能耗率模型应用于根据式（4）划分的数据子集。即𝑣 < 5 𝑚/𝑠 和 𝑣 ≥ 5 𝑚/𝑠，并分别校准系数。 然而，需要进一步调查来评估速度阈值的影响。
In the Type II model, we considered the regenerative braking factor to be related to not only the vehicle’s speed but also the other factors mentioned above, including acceleration, jerk, and road grade. Based on the fuzzy logic model provided by Maiaa et al. [60], we estimated the regeneration factor using our field driving data. Further dependency tests showed that jerk (ψ) may not be a significant indicator, so we did not include it in the model. The resulting Type II MLR model is shown in Eq. (6):
在Type II模型中，我们认为再生制动因素不仅与车辆速度有关，还与上述其他因素有关，包括加速度、加加速度和道路坡度。 基于 Maiaa 等人提供的模糊逻辑模型。 [60]，我们使用现场驾驶数据估计了再生因子。 进一步的依赖性测试表明，加加速度（ψ）可能不是一个显着指标，因此我们没有将其包含在模型中。 由此产生的 II 型 MLR 模型如公式 （6） 所示：

Table 2 and Table 3 list the calibration parameters for the Type I and Type II hybrid energy consumption rate models. In our study, we chose the Type I model to balance model complexity (such as number of variables) with computational efficiency.
表2和表3列出了I型和II型混合能耗率模型的标定参数。 在我们的研究中，我们选择 I 类模型来平衡模型复杂性（例如变量数量）和计算效率。

Machine Learning–based Vehicle Speed Forecasting

In this study, we aimed to develop a direct time series forecasting model with a second-bysecond vehicle speed trajectory detected by onboard sensors (e.g., radar) as inputs. The historical speed horizon of the input and forecasting horizon of the output are both in three time steps (i.e., 3 seconds) for training and testing the speed forecasting models. We implement a Radial Basis Function Neural Network (RBF-NN) for vehicle speed forecasting and compare its performance with other well-known nonlinear regression models like Gaussian Processes (GP) and Multi-Layer Perceptron Neural Network (MLP-NN) for different driving scenarios. The general RBF-NN based vehicle speed predictor has a feed-forward neural network framework with one hidden layer in which the nodes have radial transfer function, as shown in Figure 4. The network input is a vector containing the vehicle’s historical speed trajectory of the preceding 3 seconds, and the output is a predicted speed trajectory within a 3- second horizon.
在本研究中，我们的目标是开发一种直接时间序列预测模型，以车载传感器（例如雷达）检测到的逐秒车辆速度轨迹作为输入。 输入的历史速度范围和输出的预测范围均采用三个时间步长（即 3 秒），用于训练和测试速度预测模型。 我们实现了用于车速预测的径向基函数神经网络（RBF-NN），并将其性能与其他著名的非线性回归模型（例如高斯过程（GP）和多层感知神经网络（MLP-NN））针对不同驾驶情况进行了比较场景。 基于 RBF-NN 的通用车速预测器具有一个带有一个隐藏层的前馈神经网络框架，其中节点具有径向传递函数，如图 4 所示。网络输入是包含车辆历史速度轨迹的向量。 前 3 秒，输出是 3 秒范围内的预测速度轨迹。

The implemented RBF-NN is a three-layer feed-forward network with K hidden nodes. A radial basis function needs to be pre-defined for each hidden node to activate neurons in the hidden layer. Each hidden node contains a nonlinear activation function. Here, we chose the Gaussian function as the activation function for the RBF-NN, formulated as:
实现的 RBF-NN 是一个具有 K 个隐藏节点的三层前馈网络。 需要为每个隐藏节点预先定义径向基函数来激活隐藏层中的神经元。 每个隐藏节点都包含一个非线性激活函数。 这里，我们选择高斯函数作为 RBF-NN 的激活函数，公式为：

where 𝜑𝑗 is the activated function of node 𝑗; 𝑥̅ is the input vector for node 𝑗; 𝑤𝑘𝑗 is the output weight and 𝑏𝑘𝑗 is the constant bias; and 𝜇𝑗 and ∑𝑗 are the mean vector and covariance matrix of the 𝑗𝑡ℎ Gaussian function. The mean 𝜇𝑗 represents the center, and ∑𝑗 indicates the shape of the activation function. Finally, the output of each node at the RBF-NN’s output layer is computed as a linear combination of the outputs of the hidden nodes.
其中${\phi }_j$是节点 𝑗 的激活函数； 𝑥̅ 是节点 𝑗 的输入向量； $w_{kj}$ 是输出权重，$b_{kj}$ 是常数偏差； ${\mu }_j$和 ${\sum }_j$ 是 $j^{th}$ 高斯函数的均值向量和协方差矩阵。 均值 ${\mu }_j$ 表示中心，${\sum }_j$ 表示激活函数的形状。 最后，RBF-NN 输出层每个节点的输出被计算为隐藏节点输出的线性组合。
An advantage of an RBF-NN compared to a Gaussian Process and MLP-NN is the efficiency of training based on a two-stage procedure. The time complexity of training in a Gaussian Process for prediction increases exponentially with the sample size, which becomes quite an issue when applied to a large network in real time. An MLP-NN could have more than one hidden layer, uses an iterative technique, and works globally, while an RBF-NN has only one hidden layer, is based on a non-iterative technique, and acts as a local approximation. Besides, an RBF-NN is more robust to adversarial noise and easier to generalize than is an MLP-NN. In the first stage of RBF-NN training, the parameters of the basis function are set to model unconditional data density. The centers of our trained RBF network are determined by fitting a Gaussian mixture model with circular covariance, using the Expectation-Maximization (EM) algorithm. The second stage of training determines the weights between the hidden layer and the output layer by using the Moore-Penrose generalized pseudo-inverse, which overcomes many issues in traditional gradient algorithms such as stopping criterion, learning rate, number of epochs, and local minima. The structure of the RBF-NN was optimized by pruning the network based on 5- fold cross validation in this study. We selected an RBF-NN for real-time vehicle speed forecasting in urban driving because of its shorter training time, forecasting accuracy, and generalizability. Figure 5 illustrated an example of forecasted speed trajectory using RBF-NN, where those short segments represent the short term (3 seconds) forecasting results.
与高斯过程和 MLP-NN 相比，RBF-NN 的优点是基于两阶段过程的训练效率。 用于预测的高斯过程训练的时间复杂度随着样本大小呈指数增长，当实时应用于大型网络时，这成为一个相当大的问题。 MLP-NN 可以有多个隐藏层，使用迭代技术，并且全局工作，而 RBF-NN 只有一个隐藏层，基于非迭代技术，并充当局部近似。 此外，RBF-NN 比 MLP-NN 对对抗性噪声更稳健，并且更容易泛化。 在 RBF-NN 训练的第一阶段，设置基函数的参数来对无条件数据密度进行建模。 我们训练的 RBF 网络的中心是通过使用期望最大化 (EM) 算法拟合具有循环协方差的高斯混合模型来确定的。 第二阶段训练采用Moore-Penrose广义伪逆确定隐藏层和输出层之间的权重，克服了传统梯度算法中停止准则、学习率、epoch数、局部极小值等诸多问题 。 本研究通过基于五倍交叉验证的网络剪枝来优化 RBF-NN 的结构。 我们选择 RBF-NN 来进行城市驾驶中的实时车速预测，因为它的训练时间较短、预测准确度高且具有普适性。 图 5 展示了使用 RBF-NN 预测速度轨迹的示例，其中这些短段代表短期（3 秒）预测结果。

Deep Learning–based Eco-Approach and Departure

System Architecture

Figure 6 presents the system architecture of the proposed algorithm. As illustrated in the figure, the architecture is divided into two portions: an off-line process (upper portion) and an on-line process (lower portion). In the off-line process, vehicle dynamics (e.g., speed and acceleration) and powertrain operation (e.g., engine/motor speed, engine/motor torque, and gear ratio if any) from real-world testing are first logged via on-board diagnostics (OBD) systems. These data are then post-processed to characterize the specific powertrain model, which is used to build the components of a graph-based trajectory planning algorithm (GTPA) developed in our previous study [61]. This GPTA includes nodes (discretized states in the velocity-time-distance space), links (feasible transition from one state to another), and associated link costs (fuel consumption for the state transition at each time step). The Database from Simulation Runs module contains possible vehicle dynamics and SPaT values that are expected for the testing scenarios. The GTPA is called for the values in the database to find an associated optimal speed trajectory profile as the target vehicle approaches an intersection. Towards that end, a deep neural network model is trained using the same inputs to the GTPA as training features and the same computed paths used by the GTPA as target values. By parsing the optimal paths, the Deep Learning Trajectory Planning Algorithm (DLTPA) can learn to predict the velocity for the next time step. In other words, DLTPA is developed and trained to approximate the GTPA’s optimal performance.
图 6 展示了所提出算法的系统架构。 如图所示，该架构分为两部分：离线流程（上部） 和在线流程（下部）。 在离线过程中，首先通过车载记录来自实际测试的车辆动力学（例如速度和加速度）和动力系统操作（例如发动机/电机速度、发动机/电机扭矩和传动比（如果有）） 诊断（OBD）系统。 然后对这些数据进行后处理，以表征特定的动力系统模型，该模型用于构建我们之前的研究中开发的基于图形的轨迹规划算法（GTPA）的组件[61]。 该 GPTA 包括节点（速度-时间-距离空间中的离散状态）、链路（从一种状态到另一种状态的可行转换）以及相关的链路成本（每个时间步的状态转换的燃料消耗）。 模拟运行模块的数据库包含测试场景预期的可能的车辆动力学和 SPaT 值。 当目标车辆接近十字路口时，GTPA 会调用数据库中的值来找到相关的最佳速度轨迹曲线。 为此，使用 GTPA 的相同输入作为训练特征以及 GTPA 所使用的相同计算路径作为目标值来训练深度神经网络模型。 通过解析最佳路径，深度学习轨迹规划算法（DLTPA）可以学习预测下一个时间步的速度。 换句话说，DLTPA 的开发和训练是为了接近 GTPA 的最佳性能。

Once DLTPA is trained off-line, it will be adopted for on-line implementation. In this study, an Application Programming Interface (API) is developed for the DLTPA module in a microscopic traffic simulation environment, where vehicle and traffic states feed into DLTPA in real time and the prediction by DLTPA is used as the target velocity of the host vehicle for the next time step. The details on the creation of the training dataset for DLTPA and construction of the deep neural network are presented in the next section.
DLTPA 离线训练后，将用于在线实施。 在本研究中，在微观交通模拟环境中为 DLTPA 模块开发了一个应用程序编程接口（API），其中车辆和交通状态实时输入 DLTPA，并将 DLTPA 的预测用作主车辆的目标速度用于下一个时间步。 下一节将详细介绍 DLTPA 训练数据集的创建和深度神经网络的构建。

DLTPA Description

The DLTPA uses an MLP-NN. This type of neural network takes as input one or more parameters. It is trained on previous inputs and the corresponding expected outputs. After training, it predicts an output when given a new input. An MLP-NN contains a variable number of nodes. Each node performs an activation function on an input, such as rectified linear unit (ReLU), tangent hyperbolic, or linear. The input to the activation function is a weighted sum of all the input parameters to the node. The architecture of a node is shown in Figure 7.
DLTPA 使用 MLP-NN。 这种类型的神经网络采用一个或多个参数作为输入。 它根据先前的输入和相应的预期输出进行训练。 训练后，它会在给定新输入时预测输出。 MLP-NN 包含可变数量的节点。 每个节点对输入执行激活函数，例如修正线性单元 (ReLU)、正切双曲或线性。 激活函数的输入是节点所有输入参数的加权和。 节点架构如图7 所示。

An MLP-NN is made up of 3 types of layers of nodes. The first layer is called the input layer. It is where known parameters initially enter the MLP. In Figure 8, the nodes labeled 1, 2, and 3 make up the input layer, and X1, X2, X3 are the inputs. Following the input layer is one or more hidden layers. The nodes labeled 4, 5 in Figure 8 make up a hidden layer. The hidden layer generally has a non-linear activation function in each node, such as ReLU. The last type of layer is the output layer, which follows the last hidden layer. The output layer formats the output to be consistent with the format of the expected predictions. In Figure 8, the output layer is the node labeled 6.
MLP-NN 由 3 种类型的节点层组成。 第一层称为输入层。 这是已知参数最初进入 MLP 的地方。 在图 8 中，标记为 1、2 和 3 的节点构成输入层，X1、X2、X3 是输入。 输入层之后是一个或多个隐藏层。 图 8 中标记为 4、5 的节点构成隐藏层。 隐藏层一般在每个节点都有非线性激活函数，例如ReLU。 最后一种类型的层是输出层，位于最后一个隐藏层之后。 输出层将输出格式化为与预期预测的格式一致。 在图 8 中，输出层是标记为 6 的节点。
Each node in a layer is connected by an edge to one or more nodes in the next layer, with the exception of the output layer. A common architecture, and the one that we employed in this study, is to connect all nodes in a layer with every node in the next layer. This is commonly referred to as “densely connected,” as shown in Figure 8.
层中的每个节点都通过边连接到下一层中的一个或多个节点（输出层除外）。 我们在本研究中采用的一种常见架构是将一层中的所有节点与下一层中的每个节点连接起来。 这通常称为“密集连接”，如图 8 所示。
Each edge has an associated weight. The edge weights are the weights used to scale each input to the node when computing the weighted sum. At initialization, the weights are generally random or predefined. During training, the edge weights are adjusted so that the MLP-NN gives better predictions. Training occurs by predicting on a training unit and adjusting the edge weights based on the error between the prediction and the expected value.
每条边都有一个相关的权重。 边权重是在计算加权和时用于缩放节点的每个输入的权重。 在初始化时，权重通常是随机的或预定义的。 在训练过程中，会调整边缘权重，以便 MLP-NN 提供更好的预测。 训练是通过对训练单元进行预测并根据预测与期望值之间的误差调整边缘权重来进行的。

The DLTPA uses an MLP-NN known as a regressor. Its architecture follows the architecture described above. Consequently, we choose to use the rectified linear unit (ReLU) for the hidden layer activation functions, whose mathematical formula is given by:
DLTPA 使用称为回归器的 MLP-NN。 其架构遵循上述架构。 因此，我们选择使用修正线性单元（ReLU）作为隐藏层激活函数，其数学公式为：

Often a dropout layer is added. This is a simple method that deters the prediction model from memorizing the training data without learning a generalized function. The output layer characterizes the MLP. It is one node with activation f (x) = x.
通常会添加一个 dropout 层。 这是一种简单的方法，可以阻止预测模型在不学习广义函数的情况下记住训练数据。 输出层表征 MLP。 它是一个激活 f (x) = x 的节点。
The created dataset is partitioned to determine a neural network architecture. The first partition is a validation training set constituting 60% of all the data. The next 20% is a validation set. The last 20% is a final test set. Different numbers of hidden layers were run for 300 iterations through the validation training data and compared. Based on this, we chose to use hidden layers with 1024 nodes. The accuracy after each iteration was tested against the validation set and plotted. The results for different neural network architectures are shown in Figure 9. One hidden layer converges quickly but shows little potential for gaining accuracy from increased data. Three hidden layers may create difficulty because they require more iterations through the data, which may cause over-fitting. We chose to use an MLP-NN with 2 hidden layers. We used 250 iterations through the training data because the validation plots showed little improvement after 250 iterations. Our final architecture is provided here.
对创建的数据集进行分区以确定神经网络架构。 第一个分区是验证训练集，占所有数据的 60%。 接下来的 20% 是验证集。 最后 20% 是最终测试集。 通过验证训练数据运行不同数量的隐藏层 300 次迭代并进行比较。 基于此，我们选择使用具有 1024 个节点的隐藏层。 每次迭代后的准确性都根据验证集进行测试并绘制。 不同神经网络架构的结果如图 9 所示。一个隐藏层收敛速度很快，但从增加的数据中获得准确性的潜力很小。 三个隐藏层可能会造成困难，因为它们需要对数据进行更多迭代，这可能会导致过度拟合。 我们选择使用具有 2 个隐藏层的 MLP-NN。 我们对训练数据进行了 250 次迭代，因为验证图在 250 次迭代后几乎没有显示出任何改进。 这里提供了我们的最终架构。
Dense: nodes = 1024; activation = ReLU; Dropout: rate = 0.50
Dense: nodes = 1024, activation = ReLU; Dropout: rate = 0.50
Dense: nodes = 1, activation = linear

Simulation Study

Setup

We used PTV Vissim (http://vision-traffic.ptvgroup.com/en-us/home/) as our microscopic traffic simulation tool to model the vehicle dynamic and movement on a road network. The simulated network is a 3-mile signalized corridor with 11 signalized intersections along University Ave in Riverside, CA.
我们使用 PTV Vissim (http://vision-traffic.ptvgroup.com/en-us/home/) 作为微观交通模拟工具，对道路网络上的车辆动态和运动进行建模。 模拟网络是一条 3 英里长的信号走廊，沿加利福尼亚州里弗赛德大学大道设有 11 个信号交叉口。

Computational Time Comparison

The DLTPA was evaluated against the GTPA in terms of computational time in simulation by controlling one electric vehicle for 10 different simulation seeds. Different seeds have different times at which the controlled vehicle enters simulation. Consequently, this affects the signal phase and timing (SPaT) that the vehicle encounters in the simulation. The results are shown in Table 4. Since the DLTPA is executed at each timestep, the total computational time is amortized over the entire path. On the other hand, the computational time of the GTPA is not evenly distributed, based on our experience. A major portion of the computational time is consumed when generating a path. This advantage makes the DLTPA more applicable to real world scenarios.
通过控制一辆电动汽车使用 10 个不同的模拟种子，在模拟计算时间方面对 DLTPA 与 GTPA 进行了评估。 不同的种子受控车辆进入模拟的时间不同。 因此，这会影响车辆在模拟中遇到的信号相位和定时 (SPaT)。 结果如表 4 所示。由于 DLTPA 在每个时间步执行，因此总计算时间分摊到整个路径上。 另一方面，根据我们的经验，GTPA 的计算时间分布并不均匀。 大部分计算时间消耗在生成路径时。 这一优势使得 DLTPA 更适用于现实场景。

The computational time of the DLTPA is constant for an intersection despite the vehicle states and SPaT. This feature makes the DLTPA more attractive than the GTPA in real world implementation. Although there is one instance where the computation time is slightly off from the rest of the simulation runs, we attribute this small variation to other disturbances in the simulation runs. When we compare the results between the GTPA and DLTPA, we note that the mean absolute errors vary from 1.07 to 1.83 across different seeds. Figure 9 illustrates some example trajectory results from the GTPA and DLTPA.
无论车辆状态和 SPaT 如何，DLTPA 的计算时间对于交叉路口都是恒定的。 这一特性使得 DLTPA 在现实世界的实施中比 GTPA 更具吸引力。 尽管在一种情况下，计算时间与模拟运行的其余部分略有偏差，但我们将这种微小的变化归因于模拟运行中的其他干扰。 当我们比较 GTPA 和 DLTPA 之间的结果时，我们注意到不同种子的平均绝对误差从 1.07 到 1.83 不等。 图 9 显示了 GTPA 和 DLTPA 的一些轨迹结果示例。

Energy Consumption Results

A single electric vehicle was running along the simulated corridor in traffic for 10 different simulation runs (with different seeds) under different control strategies. For each run, the energy consumption using the VISSIM by-default control (baseline), DLTPA and GTPA were recorded in separate simulations. Results are provided in Table 5. Columns 2-4 show the energy used by each method in kilojoules/meter (kJ/meter). Columns 5-6 show the improvementin percentage of the DLTPA and GTPA compared to the baseline. In some simulation runs, the DLTPA actually performed better than the GTPA. We hypothesize that small prediction errors affect simulation events. For example, a vehicle might enter a queue before the controlled vehicle in the GTPA simulation but not in the DLTPA simulation because the controlled vehicle in the DLTPA simulation was moving slightly faster than the controlled vehicle in the GTPA simulation. On average, the DLTPA uses 0.006 kJ/meter more energy than the GTPA and provides 0.61% less improvement to the baseline energy consumption than the GTPA. Nevertheless, the DLTPA outperforms the baseline scenarios by 13.76%. Figure 10 gives an example of speed trajectories based on DLTPA vs. GTPA.
一辆电动汽车在交通中沿着模拟走廊行驶，在不同的控制策略下进行了 10 次不同的模拟运行（使用不同的种子）。 对于每次运行，在单独的模拟中记录使用 VISSIM 默认控制（基线）、DLTPA 和 GTPA 的能耗。 结果如表 5 所示。第 2-4 列显示每种方法使用的能量，单位为千焦耳/米 (kJ/m)。 第 5-6 列显示与基线相比，DLTPA 和 GTPA 的改进百分比。 在一些模拟运行中，DLTPA 实际上比 GTPA 表现更好。 我们假设小的预测误差会影响模拟事件。 例如，在 GTPA 模拟中，车辆可能会在受控车辆之前进入队列，但在 DLTPA 模拟中则不会，因为 DLTPA 模拟中的受控车辆的移动速度比 GTPA 模拟中的受控车辆稍快。 平均而言，DLTPA 比 GTPA 多使用 0.006 kJ/m 的能源，并且与 GTPA 相比，对基准能耗的改善少 0.61%。 尽管如此，DLTPA 的表现仍优于基准情景 13.76%。 图 10 给出了基于 DLTPA 与 GTPA 的速度轨迹示例。

Conclusion

The proposed DLTPA can significantly reduce the computational time to constant complexity (O(1)). Further, it addresses the issue of inflexibility to rapidly changing traffic conditions by predicting only the velocity for the next timestep. We tested this innovative Eco-Approach and Departure (EAD) algorithm against the GTPA (which was also developed in our previous work) in a simulation environment. In summary,
所提出的 DLTPA 可以显着地将计算时间减少到恒定复杂度 (O(1))。 此外，它通过仅预测下一个时间步的速度来解决快速变化的交通条件的不灵活性问题。 我们在模拟环境中针对 GTPA（这也是我们之前的工作中开发的）测试了这种创新的生态进场和离场 (EAD) 算法。 总之，

• This project developed an innovative deep learning–based Eco-Approach and Departure algorithm for electric vehicles with customized powertrain models.
• 该项目为具有定制动力系统模型的电动汽车开发了一种基于深度学习的创新经济接近和远离算法。
• The proposed DLTPA can significantly reduce the computational complexity of the trajectory planning algorithm, which is favorable for real-time implementation.
• 所提出的DLTPA可以显着降低轨迹规划算法的计算复杂度，有利于实时实现。
• The proposed DLTPA can increase the flexibility of the trajectory planning algorithm in response to a rapidly changing environment.
• 所提出的 DLTPA 可以提高轨迹规划算法的灵活性，以应对快速变化的环境。
• Based on the preliminary simulation study in VISSIM, the DLTPA can provide a 13.76% improvement above the baseline. The results from the DLTPA consume only 0.006 kJ/meter more energy per distance than the GTPA (which is considered to be an optimal solution in terms of energy efficiency) but significantly improve the computational efficiency by up to 90% (e.g., simulation run #8 in Table 4).
• 根据 VISSIM 中的初步模拟研究，DLTPA 可以比基线提高 13.76%。 DLTPA 的结果仅比 GTPA 每距离多消耗 0.006 kJ/米的能量（这被认为是能源效率方面的最佳解决方案），但计算效率显着提高了 90%（例如，模拟运行表 4 中的 #8）。