基于最小化原则的电动汽车高速公路节能驾驶

Highway Eco-Driving of an Electric Vehicle Based on Minimum Principle

基于最小化原则的电动汽车高速公路节能驾驶

Abstract—The digital map in the on-board navigation system on passenger vehicles nowadays gives controllers access to detailed information of a planned route. Connected automated vehicles equipped with high computing power, are able to process data and leverage high control freedom available through automation. This paper proposes a fast optimization algorithm for eco-driving of electric vehicles in highway cruising scenarios. The algorithm is based on optimal control theory and takes into account road grades as well as state constraints imposed by speed limits and safe headway to the preceding car. Using Autonomie, the proposed algorithm is evaluated by simulating two example scenarios. The first result demonstrates an energy saving potential of 4.4%.
Index Terms—eco-driving, highway cruising, minimum principle, electric vehicle, connected automated vehicle, state constraints
摘要：如今，乘用车车载导航系统中的数字地图使管制员能够获取计划路线的详细信息。 配备高计算能力的联网自动车辆能够处理数据并利用自动化提供的高控制自由度。 本文提出了一种高速公路巡航场景下电动汽车节能驾驶的快速优化算法。 该算法基于最优控制理论，并考虑了道路坡度以及速度限制和前车安全车距所施加的状态约束。 使用 Autonomie，通过模拟两个示例场景来评估所提出的算法。 第一个结果显示节能潜力为 4.4%。
索引术语——节能驾驶、高速公路巡航、最低化原则、电动汽车、联网自动驾驶汽车、状态约束

I. INTRODUCTION

Road transport is a major human activity and has a substantial impact on the environment. As the whole society becomes increasingly aware of resource sustainability and environment protection, scientists worldwide are concerned with reducing energy consumption. Without jeopardizing the essential demand of transportation and travel, actively utilizing “green” energy and improving energy efficiency are considered as two viable options。
道路运输是一项重要的人类活动，对环境产生重大影响。 随着全社会对资源可持续性和环境保护意识的日益增强，减少能源消耗成为全世界科学家关注的焦点。 在不影响交通和出行的基本需求的情况下，积极利用“绿色”能源和提高能源效率被认为是两个可行的选择。
Powertrain technology must adapt to green energy sources. A fully electrified powertrain is seen as an attractive possibility to achieve zero-emissions. The success of electric cars on the market, like Tesla Model S, Nissan LEAF and BMW i3, indicated a bright future for electric cars. Therefore, this paper focuses on the electric car。
动力总成技术必须适应绿色能源。 全电气化动力系统被视为实现零排放的一种具有吸引力的可能性。 特斯拉 Model S、日产聆风、宝马 i3 等电动汽车在市场上的成功，预示着电动汽车的光明前景。 因此本文重点研究电动汽车。
Among methods to improve energy efficiency, optimization of the driving profile, which is also called “eco-driving,” poses an effective approach considering it does not require fundamental changes in terms of hardware. Eco-driving is, in nature, an optimal control problem (OCP). As addressed in [1], applications regarding eco-driving for road vehicles are in general categorized in highways [2]–[4], urban driving [5], [6], intersection crossing [6], [7], and car-following [8], [9]. The problem can be solved using various methods such as dynamic programming (DP), analytical solutions, and Pontryagin’s Minimum Principle (PMP). DP is known for its versatility and solution robustness against nonlinearity and it is able to yield a close-to-optimum solution that suffers only from numerical inaccuracies due to discretization. Examples of its implementation are found in [10], [11]. Analytical solutions are generally bound with utilization of affine functions for control-oriented modeling. Applying the necessary conditions for the optimal control leads to bang-bang control. However, its calculation is very fast, since it reduces the problem to an analytical relation between the terminal state (normally the travel distance or time) and some parameters (e.g., co-states), see [4], [6]. PMP approaches use higher-order polynomials for system modeling. Consequently, applying the necessary conditions of PMP converts the OCP into a boundary value problem, which can be solved numerically, using a variety of techniques, e.g., multiple shooting [12] and collocation [13].
在提高能源效率的方法中，驾驶模式的优化（也称为“节能驾驶”）是一种有效的方法，因为它不需要对硬件进行根本性的改变。 节能驾驶本质上是一个最优控制问题（OCP）。 如[1]中所述，道路车辆生态驾驶的应用一般分为高速公路[2]-[4]、城市驾驶[5]、[6]、十字路口[6]、[7]和 跟车[8]、[9]。 该问题可以使用动态规划（DP）、解析解和庞特里亚金最小原理（PMP）等多种方法来解决。 DP 以其多功能性和针对非线性的解决方案鲁棒性而闻名，并且能够产生接近最优的解决方案，该解决方案仅因离散化而受到数值不准确的影响。 其实现示例可在 [10]、[11] 中找到。 分析解决方案通常与仿射函数的利用相结合，以进行面向控制的建模。 应用最优控制的必要条件会导致爆炸式控制。 然而，它的计算速度非常快，因为它将问题简化为最终状态（通常是行进距离或时间）和一些参数（例如共状态）之间的分析关系，参见[4]、[6]。 PMP 方法使用高阶多项式进行系统建模。 因此，应用 PMP 的必要条件将 OCP 转换为边值问题，可以使用多种技术（例如多重射击 [12] 和搭配 [13]）进行数值求解。
In previous research [14], [15], the authors investigated ecodriving based on PMP for passenger cars with conventional and parallel hybrid powertrains, respectively. In that work, inspired by [2], [16], the slopes along the driven route are assumed to be piecewise constant. This paper continues the discussion on optimizing vehicle driving on highways by utilizing information from digital maps. Adding to the previous work, this paper introduces a trajectory-building technique using vehicle speed as a reference axis. This technique allows building solution trajectories in advance, which greatly shortens calculation time. Moreover, the paper discusses two kinds of state constraints, speed limits and car following, and proposes methods to extend the existing algorithm to handle them. In the final section, the algorithm is evaluated through simulation to show its merits in terms of energy saving.
在之前的研究[14]、[15]中，作者分别针对传统和并联混合动力系统的乘用车研究了基于PMP的节能驾驶。 在该工作中，受 [2]、[16] 的启发，沿驾驶路线的坡度被假定为分段常数。 本文继续讨论利用数字地图信息来优化高速公路上的车辆驾驶。 在之前的工作的基础上，本文介绍了一种使用车辆速度作为参考轴的轨迹构建技术。 该技术允许提前构建求解轨迹，从而大大缩短计算时间。 此外，本文讨论了两种状态约束：速度限制和车辆跟随，并提出了扩展现有算法来处理它们的方法。 最后部分通过仿真对该算法进行评估，以展示其在节能方面的优点。

II. CONTROL-ORIENTED MODELING

The vehicle we consider in this paper is equipped with a battery-electric powertrain commonly available on the market. The configuration is displayed in Fig. 1.
我们在本文中考虑的车辆配备了市场上常见的电池电机动力系统。 配置如图1所示。

A. Powertrain Modeling

The motor receives electric power from the battery and propels the driving wheels through the gearbox and the final drive. The reduction ratio of the transmission is fixed.1 The traction force Ft at the wheels results from the motor output torque Tm multiplied by the transmission ratio γ and the efficiency η,
电机从电池接收电力，并通过变速箱和主减速器推动驱动轮。 变速器的减速比是固定的^1。 车轮上的牵引力$F_t$由电机输出扭矩 $T_m$ 乘以传动比 $\gamma$ 和效率 $\eta$ 得出，

The total transmission ratio γ and total efficiency η refer to the whole driveline and consider the gearbox, the final drive, and the wheel.
总传动比 $\gamma$ 和总效率 $\eta$ 涉及整个传动系统，并考虑变速箱、主减速器和车轮。
Neglecting the wheel slip, the rotation speed of the motor !m is proportional to the vehicle speed v,
忽略车轮滑移，电机转速 ${\omega }_m$与车速 $v$ 成正比，

As Fig. 1 shows, power flows left to the motor are electrical. Instead of explicitly describing losses/efficiencies on each component, i.e., battery, inverter, and motor, we lump them together to model energy conversion, a method widely adopted in the literature [6], [17], [18]. Thus, the consumed battery power Pb is described as a quartic piecewise function of two variables, motor torque Tm and speed !m, where all losses on the electric path are taken into account,
如图 1 所示，流向电机的能量是电力。 我们没有明确描述每个组件（即电池、逆变器和电机）的损耗/效率，而是将它们集中在一起来建模能量转换，这是文献 [6]、[17]、[18] 中广泛采用的方法。 因此，消耗的电池功率 $P_b$ 被描述为两个变量（电机扭矩 $T_m$ 和速度 ${\omega }_m$）的四次分段函数，其中考虑了电气路径上的所有损耗，

The two groups of coefficients cm;ij and cg;ij, i; j 2 f0; 1; 2g, are associated with the motor and generator operation modes, respectively. They are obtained using least squares surface fitting with the measurement data
两组系数 $c_m,ij$ 和 $c_g,ij,i,j\in [0,1,2]$，分别与电动机和发电机操作模式相关。 它们是使用最小二乘表面拟合与测量数据获得的.
During the fitting, there are two conditions to fulfill:
在拟合过程中，需要满足两个条件：

where superscripts + and − denote the right- and left-hand limits, respectively. The first condition (4a) guarantees the continuity at switching between motor and generator modes. It is easy to deduce further that cm;0j = cg;0j for j 2 f0; 1; 2g. The second condition (4b) is based on a general observation of the measurement data, and it ensures a convex union at Tm = 0.
其中上标 + 和 − 分别表示右侧和左侧极限。 第一个条件（4a）保证了电动机和发电机模式之间切换的连续性。 很容易进一步推导出 $c_{m,0j}=c_{g,0j}$ 对于 $j\in [0,1,2]$。 第二个条件 (4b) 基于对测量数据的一般观察，它确保 $T_m=0$ 时的凸并集。
Fig. 2 compares the fitting results with the original data regarding the electrical path efficiency.
图 2 将拟合结果与有关电气路径效率的原始数据进行了比较。

B. Chassis Dynamics

According to Newton’s laws of motion, the vehicle dynamics result from the forces acting on the chassis and can be expressed as [19],
根据牛顿运动定律，车辆动力学由作用在底盘上的力产生，可以表示为[19]，

where m is the total inertia, including the vehicle mass and equivalent mass of all rotating parts, and s is the traveled distance. The mechanical brake force is denoted by Fb. The total resistance force Fres sums up the wheel rolling resistance, the aerodynamic drag, and the uphill force, which can be expressed with a quadratic function of v,
其中 m 是总惯量，包括车辆质量和所有旋转部件的等效质量，s 是行驶距离。 机械制动力用 $F_b$ 表示。 总阻力 $F_{res}$ 是车轮滚动阻力、空气动力阻力和上坡力的总和，可以用 v 的二次函数表示，

where coefficients cf0;1;2g are constants determined by physical characteristics of the vehicle and the environment. In the last term, mg represents the vehicle weight (excluding inertia of rotating parts). Road grade α is a function of distance s and is assumed to be piecewise constant.
其中系数 $c_{0,1,2}$ 是由车辆和环境的物理特性确定的常数。 最后一项中，$mg$ 代表车辆重量（不包括旋转部件的惯量）。 道路坡度 α 是距离 s 的函数，并假定为分段常数。

III. OPTIMAL CONTROL PROBLEM

The OCP is to minimize the overall energy consumption in the battery over a given highway road section from start to end time, t0 to tf, which is mathematically formulated as,
OCP 旨在最大限度地减少给定高速公路路段从开始时间到结束时间 $t_0$ 到 $t_f$ 的电池总体能耗，其数学公式为：


Here, the system transition equation (7b) of state x = [v; s]⊤ and control u = [Tm; Fb]⊤ is declared in (5). The admissible ranges for state and control are X and U in (7c). In (7d) and (7e), x0 = [v0; s0]⊤ and xf = [vf; sf]⊤, respectively, define the speed and position of the vehicle at the outset and at the finish.
这里，系统转移方程(7b)的状态 $x=[v,s{]}^T$ 并控制 $u=[T_m,F_b{]}^T$ 在（5）中声明。 状态和控制的允许范围是（7c）中的 $X$ 和 $U$。 在(7d)和(7e)中，$x0=[v0,s0{]}^T$ 且 $x_f=[v_f,s_f{]}^T$ 分别定义车辆在起点和终点的速度和位置。

A. Necessary Conditions for Optimal Solution

Based on the OCP (7), the Hamiltonian can be formulated,
基于 OCP (7)，哈密顿算子可以表述为：

with an appended co-state λ. Inserting the explicit expressions for system transition (5), and decomposing state and control variables, H into expands,
带有附加的共状态 $\lambda$。 插入系统转移的显式表达式（5），并将状态变量和控制变量 $H$ 分解为展开式，

The necessary conditions from PMP are known as,
PMP 的必要条件被称为，

where the asterisk ∗ denotes the optimal variable. In light of (9a), no explicit dependency of H on t indicates a constant value, which we define as H(t) = H; 8t 2 [t0; tf]. Note that, λv can be directly derived from (8),
其中星号*表示最优变量。 根据 (9a)，$H$ 对 $t$ 没有明确的依赖性，这表明它是一个常数值，我们将其定义为 $H(t)=xxxxxx,\forall t\in [t_0,tf]$。 注意，${\lambda }_v$ 可以直接从(8)推导出来，

Owing to the assumption of piecewise constant α, based on (9b), λs is also piecewise constant,
由于分段常数 $\alpha$ 的假设，根据(9b)，${\lambda }_s$ 也是分段常数，

The last PMP condition (9c) yields the optimal Fb in a bangbang form,
最后一个 PMP 条件 (9c) 产生 bang-bang 形式的最佳 $F_b$，

where the lower and upper motor torque limits Tmin and Tmax depend on !m. The unbound optimal motor torques Tmo f+;−g are the roots to (9c) in the absence of control constraints, i.e., ∇TmH = 0, applying coefficient groups cm;ij and cg;ij, respectively. Subscripts + and − are designated to motor and generator modes. As stated in [15], [16], inserting ∇TmH = 0 in (10) delivers a quadratic equation of Tmo f+;−g without involving λv,
其中电机扭矩下限和上限 $T_{min}$ 和 $T_{max}$ 取决于 ${\omega }_m$。 在没有控制约束的情况下，分别应用系数组 $c_m,ij和c_g,ij$，无界最优电机扭矩 $T_{m+,-}^o$ 是 (9c) 的根，即 ${\Delta }_{T_m}H=0$。 下标 + 和 - 表示为电动机和发电机模式。 如[15]、[16]中所述，在(10)中插入 ${\Delta }_{T_m}H=0$可得到 $T_{m+,-}^o$的二次方程，而不涉及 ${\lambda }_v$，

It follows that T o mf+;−g becomes a function of v together with two parameters, H and λs.2
由此可见，$T_{m+,-}^o$ 成为关于 v的函数，包含两个参数 $H$ 和 ${\lambda }_s$ 的函数。
Summarizing, we have obtained the instantaneous optimal solution for control u depending on the current vehicle speed v and parameters H and λs.
综上所述，我们根据当前车速 $v$ 以及参数 $H$ 和 ${\lambda }_s$ 获得了控制 $u$ 的瞬时最优解。

B. Trajectory Building in Reference to Vehicle Speed

The end time tf of the OCP (7) is usually ambiguous in practice. Noted in [16], the Hamiltonian value H has a nonlinear monotonic relation to the end time tf. Therefore, instead of deducing H by solving a classical boundary value problem with an exact tf, we assume such an H that the vehicle runs on level ground in a steady state at the desired speed vset set by the driver.
OCP (7) 的结束时间 $t_f$ 在实践中通常是不明确的。 在[16]中指出，哈密顿量 $H$ 与结束时间 $t_f$ 具有非线性单调关系。 因此，我们不是通过求解具有精确 $t_f$ 的经典边值问题来推导 $H$，而是假设车辆以驾驶员设定的所需速度 $v_{set}$ 在稳定状态下在水平地面上行驶。
Note the optimal control policy (12) and (13), a steady state (v_ = 0) is reached when the two radicals of Tm∗ converge, meaning δ = 0, and thus,
注意最优控制策略（12）和（13），当 $T_m^{\ast }$ 的两个根式收敛时达到稳态（$\dot{v}=0$），即 $\delta =0$，因此，

Moreover, the above assumption for H implies that this equilibrium at vset should occur with no time constraint, which leads to
此外，上述对 $H$ 的假设意味着 $v_{set}$ 处的平衡应该在没有时间限制的情况下发生，即

Solving (14a) and (14b) simultaneously yields not only the mission-constant H but also λs for α = 0.
同时求解 (14a) 和 (14b) 不仅可以得到任务常数 $H$，还可以得到 $\alpha =0$ 时的 ${\lambda }_s$。
Assuming each road slope with different α is sufficiently long, there is a steady-state speed vss as a function of α that fulfills conditions (14) in place of vset. While H is invariant, λs adapts to α to fulfill conditions (14) with vss(α) in place of vset.
假设具有不同坡度 $\alpha$ 的每个道路足够长，则存在关于 $\alpha$ 的稳态速度 $v_{ss}$ 函数，其满足条件（14）代替 $v_{set}$。 虽然 $H$ 是不变的，但 ${\lambda }_s$ 会适应 $\alpha$ 以满足条件 (14)，并用 $v_{ss}(\alpha )$ 代替 $v_{set}$。
Following the conclusion in Subsection III-A, we are able to build solution trajectories in reference to vehicle speed v now that parameters H and λs are known for all road grades. For example, Fig. 3 shows the trajectories of Tm, λv, s, and t over v, associated with acceleration (Tm+ red) and deceleration (Tm− blue), respectively, for α = 0.
根据第 $III-A$ 小节中的结论，既然所有道路等级的参数 $H$ 和 ${\lambda }_s$ 都已知，我们就能够根据车辆速度 $v$ 构建求解轨迹。 例如，图 3 显示了当 $\alpha =0$ 时，$T_m、{\lambda }_v、s、t$ 对于 $v$ 的轨迹，分别与加速度（$T_{m+}$ 红色）和减速度（$T_{m-}$ 蓝色）相关。

图 3. 电机扭矩 $T_m$、共态 ${\lambda }_v$、距离 $s$ 和时间 $t$ 随车速 $v$ 变化的轨迹。参数对 {$H,{\lambda }_s$} 对应于 $v_{set}=25m/s$ 的水平地面上的稳态行驶。

C. Route Segmentation and Solution Concatenation

Nowadays, digital maps can provide detailed elevation and slope profiles for any road section. In [14], in order to comply with the constant-slope assumption for the optimization calculation, a preprocessing program is proposed to simplify real-world routes into piecewise constant slope profiles. Consequently, the route is split into a succession of segments, each with a single slope grade.
如今，数字地图可以提供任何路段的详细海拔和坡度剖面。 在[14]中，为了符合优化计算的恒定斜率假设，提出了一种预处理程序，将现实世界的路线简化为分段恒定斜率剖面。 因此，路线被分成一系列路段，每个路段都有一个坡度。
Since v is not explicitly involved in jumps of α, its corresponding co-state λv must be continuous between segments, according to Weierstrass-Erdmann corner conditions [20]. Junction points in v at segment borders can be found by leveraging this continuity condition [2], [14].
由于 $v$ 没有明确参与 $\alpha$ 的跳跃，因此根据 Weierstrass-Erdmann 角点条件 [20]，其相应的共状态 ${\lambda }_v$ 在段之间必须是连续的。 利用这种连续性条件可以找到在段边界处 $v$ 的连接点 [2]、[14]。
Now that the junction points between segments are fully defined, each segment becomes a self-contained OCP. Trajectories built for each slope in preparation, as described in Subsection III-B, are used to form the solution for each segment. The final solution is composed by concatenating the sub-arcs of each segment in succession.
现在，段之间的连接点已完全定义，每个段都成为一个独立的 OCP。 如第 III-B 小节所述，为准备中的每个斜坡构建的轨迹用于形成每个路段的解决方案。 最终的求解结果是通过将每个线段的子弧依次连接而成。

IV. STATE CONSTRAINTS IN HIGHWAY CRUISING

Highway driving scenarios in the real world are far more complex than just varying terrain. Often vehicles encounter pure state constraints in the form of h(x; t) ≤ 0 due to safety and legal reasons. According to the definition [21], a pthorder state constraint h(x; t) is differentiated p times in t until the control variable explicitly appears, h§(x; u; t). Using the indirect adjoining method [21], the Hamiltonian is extended with h§ multiplied by µ,
现实世界中的高速公路驾驶场景比简单的地形变化要复杂得多。 由于安全和法律原因，车辆经常遇到 $h(x,t)\le 0$ 形式的纯状态约束。 根据定义[21]，$p$ 阶状态约束 $h(x,t)$ 在 $t$ 中被微分 $p$ 次，直到控制变量显式出现 $h^{(p)}(x,u,t)$。 使用间接邻接方法 [21]，哈密顿量通过 $h^{(p)}$ 乘以 $\mu$ 进行扩展，

Let t1 and t2 be the entry and exit times of the state boundary. For t 2 [t1; t2] ⊂ [t0; tf], constraint is active h = 0 and multiplier µ > 0; for t 2 [t0; t1)[(t2; tf], constraint is inactive h < 0 and multiplier µ = 0. The augmented version of PMP necessary conditions (9) hold for H^ ,
设 $t_1$ 和 $t_2$ 为状态边界的进入和退出时间。 对于 $t\in [t_1,t_2]\subset [t_0,t_f]$， $h=0$ 且乘数 $\mu >0$ 约束有效； 对于$t\in [t_0,t_1)\cup (t_2,t_f]$， $h<0$ 且乘数 $\mu =0$ 约束无效。 PMP 必要条件 (9) 的增强版本对于 $\hat{H}$ 成立，

At entry time t1, discontinuity is expected in co-state λ as well as in Hamiltonian H,
在进入时间 $t_1$时刻，共状态 $\lambda$ 以及哈密顿量 $H$ 预计会出现不连续性，

where νk(t1) are non-negative multipliers to the partial derivatives. If t1 = t2, then the boundedness becomes a contact point, and t1 is called a contact time. At a contact time, (17) still holds with νk(t1) = 0 for k ≥ 2.
其中 ${\nu }^k(t_1)$ 是偏导数的非负乘数。 如果 $t_1=t_2$，则有界性成为接触点，$t_1$ 称为接触时间。 在接触时间，当 ${\nu }^k(t_1)=0$ 且 $k\ge 2$ 时，(17) 仍然成立。

A. Speed Limits

Piecewise speed limits are imposed by traffic signs mandating the highest permissible travel speed on a road section. Point-wise speed limits results from either a mandatory full stop or a sharp road curvature. For simplicity, in this paper, we concentrate only on maximum speed limits.
分段速度限制是由交通标志规定的，规定路段上允许的最高行驶速度。 逐点限速是由强制完全停车或道路急弯造成的。 为简单起见，在本文中，我们仅关注最大速度限制。
Speed limits, no matter which type, can be written as
速度限制，无论哪种类型，都可以写成

Here, vmax(s) is a piecewise constant function of distance s.
这里，$v_{max}(s)$ 是关于距离 $s$ 的分段常数函数。

where i = 0; 1; : : : ; nvlim − 1, and nvlim is the number of successive speed limit occurrences. The start position of the ith speed limit is denoted by sl;i. Speed limits are state constraints of the first order as h(1) 1 = _ v(x; u) [9].
其中 $i=0,1,...,n_{vlim}-1$ ， $n_{vlim}$ 是连续出现速度限制的次数。 第 $i$ 个速度限制的起始位置用 $s_{1,i}$表示。 速度限制是一阶状态约束，如 $h_1^{(1)}=\dot{v}(x,u)$ [9]。
In view of (17), jumps are described as,
鉴于（17），跳跃被描述为，


The first case of (21) is discussed in [9], and leads to the conclusion that λv, λs, and H are continuous if the boundary entry/exit is not at a start/end of a speed limit, s = 2 fsl;ig.
(21) 的第一种情况在[9]中讨论，并得出这样的结论：如果边界入口/出口不在速度限制的开始/结束处 $s\notin s_{1,i}$，则 ${\lambda }_v、{\lambda }_s和H$ 是连续的。
第二种和第三种情况如图 4 所示。与第一种情况 $s\notin s_{1,i}$ 一样，$H$ 连续性 (20c) 决定了 ${\lambda }_v$ 的连续性，因此 ${\nu }_1^1=0$。然而，根据 (20b)， $\partial v_{max}=\partial s$ 的狄拉克增量脉冲应导致 ${\lambda }_s$ 的有限跳跃。 由于坡度的变化和速度限制通常不会重叠，因此我们假设速度限制变化之前和之后的坡度相同。 因此，请注意第 III-B 小节中的分析，目标稳态速度 $v_{ss}$ 也保持不变。 如果斜率相关的 $v_{ss}$ 高于 $v_{max,i}$，则 $v^{\ast }$ 在开始前从 $v_{ss}$ 下沉并在 $v_{max,i}$ 结束后飙升回 $v_{ss}$ 才是合乎逻辑的。 这一观察结果也适用于逐点限速的情况。对于 $s\in [s_{1,i},s_{1,i+1}]$ 的 ${\lambda }_s$ 的精确值是在 $h_1^{(1)}=\dot{v}=0$ 的条件下计算的，因此 $\delta (v_{max,i})=0$。值得注意的是，在边界区间上，与 (14b) 不同，$\partial \delta [v_{max,i}]=\partial v=0$ 不一定成立。
Consequently, the speed limit start positions fsl;ig must be appended to the set of segmentation positions (constant slope start positions). In the segments where speed limits are active, vss and λs in Subsection III-B must be modified [14], [15].
因此，速度限制起始位置 {$s_1,i$} 必须附加到分段位置组（恒定坡度起始位置）。 在速度限制生效的路段中，必须修改第 III-B 小节中的 $v_{ss}$ 和 ${\lambda }_s$ [14]、[15]。

B. Car Following

The vehicle in front presents a distance constraint,
前方车辆存在距离限制，

Upper distance bound sb takes into account the position of the preceding car sobj and the safety headway rule σh [22],
距离上限 $s_b$ 考虑了前车位置 $s_{obj}$ 和安全车头时距规则 ${\sigma }_h$ [22]，

vobj denotes the preceding car speed. Coefficients are chosen according to [22], ch1 = 0:3 s; ch0 = 0:08 m.
$v_{obj}$表示前车速度。 根据[22]选择系数，$c_{h1}=0.3s,c_{h0}=0.08m$。
距离约束是二阶的，即 $h_2^{(1)}=v-\dot{s_b}和h_2^{(2)}=\dot{v}(x,u)-\ddot{s_b}$。 鉴于（17），阶跃被描述为，

As ν1 2; ν22 ≥ 0, jumps can occur in H, λv, and λs at entry (or contact) times of boundary h2.
当${\nu }_2^1,{\nu }_2^2\ge 0$，在边界 $h_2$ 的进入（或接触）时刻，$H、{\lambda }_v和{\lambda }_s$ 可能发生跳跃。
Since the distance constraint (22) is explicitly dependent on time, unlike the speed constraints, the point where boundedness is activated is ambiguous, with respect to distance (as well as to time or speed), in the absence of a known state trajectory. Intuitively, in order to observe state boundedness, we compare the unconstrained solution trajectory of s(t) with the distance bound sb(t). As previously mentioned, the route segmentation is performed with reference to distance points, i.e., start points of slopes and speed limits. Thus, it is more convenient to use the inverse form of sb(t), mapped onto the distance axis, i.e., tb(s). The distance constraint becomes a time constraint in reference to distance: tb(s) − t(s) ≤ 0.
由于距离约束 (22) 明确地依赖于时间，与速度约束不同，在没有已知状态轨迹的情况下，相对于距离（以及时间或速度），激活有界的点是不明确的。 直观上，为了观察状态有界性，我们将 $s(t)$ 的无约束解轨迹与距离界限 $s_b(t)$ 进行比较。 如前所述，路线分段是参考距离点（即斜坡起点和速度限制）进行的。 因此，使用 $s_b(t)$ 的逆形式映射到距离轴上，即 $t_b(s)$ 更方便。 距离约束变成关于距离的时间约束：$t_b(s)-t(s)\le 0$。
Inspired by [23], we propose a method to find positions in s where the t(s) is violated the most by the distanceunconstrained solution trajectory. Then the route, as well as the OCP, is split in two with the constrained s, v and t touching the inverse distance constraint tb(s). The two steps above are repeated until the distance constraint (22) is met for all s 2 [s0; sf].
受[23]的启发，我们提出了一种方法来查找 $s$ 中 $t(s)$ 被距离无约束的解轨迹破坏最多的位置。 然后，路线以及 OCP 被一分为二，约束 $s、v和t$ 接触反距离约束 $t_b(s)$。 重复上述两个步骤，直到所有 $s\in [s_0,s_f]$ 满足距离约束 (22)。
The recursive program goes as follows:
递归程序如下：

• Compute the distance-unconstrained optimal solution (including speed limits).3 The set of OCP boundary points is initialized: S = fs0; sfg. If the distance constraint is violated, i.e., t(s) − tb(s) < 0:
• 计算距离无约束的最优解（包括速度限制）。OCP边界点集初始化：$S=s_0,s_f$；如果违反距离约束，即 $t(s)-t_b(s)<0$：
  repeat
  重复
  Locate the position where the bound tb(s) is violated the most,
  找到最违反边界 $t_b(s)$ 的位置，
  
  where i is the ith iteration of the recursive program.
  其中 i 是递归程序的第 i 次迭代。
  Add &i into S, which becomes the jth element in S, denoted as sj. Extend the OCP with an additional boundary value condition:
  将 ${\zeta }_i$ 添加到 $S$ 中，成为 $S$ 中的第 $j$ 个元素，记为 $s_j$。 使用附加边界值条件扩展 OCP：
  
  Retain the already computed solution trajectories for s 2 [s0; sj−1] and s 2 [sj+1; sf].
  保留已计算的 $s\in [s_0,s_{j-1}]$ 和 $s\in [s_{j+1},s_f]$。
  Solve the sub-OCP for s 2 [sj−1; sj], with (25) as the terminal condition. Use the shooting method to adjust H (within this sub-OCP) so as to fulfill the fixed end time of the sub-OCP.4
  求解 $s\in [s_{j-1}，s_j]$ 的子OCP，以(25)作为终止条件。 使用shooting方式调整 $H$（子OCP内），以满足该子OCP的固定结束时间。
  Solve the sub-OCP for s 2 [sj; sj+1], with (25) as the initial condition. If the end time is fixed, use the shooting method to adjust H (within the sub-OCP); otherwise, use the H for the free end time case.5
  求解 $s\in [s_j，s_{j+1}]$的子OCP，以(25)为初始条件。 如果结束时间固定，则使用shooting方法调整 $H$（在子OCP内）； 否则，使用 $H$ 表示空闲结束时间情况。
  until min s2[s0;sf](t(s) − tb(s)) ≥ ϵtb, where ϵtb is the minimum tolerance.
  直到 $mi{n}_{s\in [s_0,s_f]}(t(s)-t_b(s))\ge {ϵ}_{tb}$，其中 ${ϵ}_{tb}$ 是最小公差。

V. SIMULATION RESULTS

In this section, two example scenarios on highway cruising are presented to showcase the speed optimization. Then, we use forward-looking vehicle models developed within Autonomie [24], a framework for vehicle fuel consumption and performance simulation, to evaluate the merits of the proposed algorithm in terms of energy saving.
在本节中，将介绍两个高速公路巡航示例场景来展示速度优化。 然后，我们使用 Autonomie [24]（车辆燃油消耗和性能模拟框架）中开发的前瞻性车辆模型来评估所提出的算法在节能方面的优点。

A. Simulation Setup

The simulation environment is set up as Fig. 5 illustrates. The proposed algorithm, called “Dyn-Opt,” is compared with another control strategy, called “Baseline,” where a nonoptimal reference speed trace is applied. For both strategies, the same vehicle plant model is used, whose configuration is shown in Table I.
仿真环境设置如 图5 所示。 所提出的称为“Dyn-Opt”的算法与另一种称为“Baseline”的控制策略进行了比较，其中应用了非最佳参考速度轨迹。 对于这两种策略，使用相同的车辆工厂模型，其配置如 表1 所示。


The Baseline strategy uses the default rule-based controller for an electric vehicle with a single-speed fixed gear transmission in Autonomie. The driver model (in Autonomie) follows the reference speed profile. Its acceleration/deceleration request is then interpreted, and the rule-based powertrain control translates the request into more detailed commands for the motor and the brake.
基线策略使用 Autonomie 中带有单速固定齿轮变速器的电动汽车的默认基于规则的控制器。 驾驶员模型（在 Autonomie 中）遵循参考速度曲线。 然后解释其加速/减速请求，并且基于规则的动力系统控制将请求转换为针对电机和制动器的更详细的命令。The Dyn-Opt strategy is an offline implementation of the eco-driving speed optimization algorithm. The optimal control policy for [Tm; Fb]⊤ is first calculated offline by solving the OCP (7). The solution trajectories [Tm; Fb]⊤ are tabulated in reference to distance s. They are then applied in Autonomie directly as control commands for full model simulation, with no feedback loop.
Dyn-Opt策略是经济驾驶速度优化算法的离线实现。 $[T_m,F_b{]}^T$ 的最优控制策略首先通过求解 OCP (7) 进行离线计算。 求解轨迹 $[T_m,F_b{]}^T$ 参照距离 $s$ 制成表格。 然后将它们直接应用在 Autonomie 中作为完整模型仿真的控制命令，无需反馈回路。

B. Uncongested Real-World Highway

The highway from Chicago O’Hare airport to Argonne National Laboratory is selected as one example real-world route for simulation test. Data of high-accuracy road slope and speed limits are pulled from HERE digital maps. To meet the assumption of constant road slopes for optimization calculation, the original slope profile needs to be simplified as piecewise constant, using the method explained in detail in [14]. The original and simplified slope profiles are displayed in Fig. 6. Note that the original slope profile is implemented in the simulation, although the road slope is approximated in the optimization calculation.
选择从芝加哥奥黑尔机场到阿贡国家实验室的高速公路作为模拟测试的真实路线示例。 高精度道路坡度和速度限制数据取自 HERE 数字地图。 为了满足优化计算中道路坡度恒定的假设，需要使用[14]中详细解释的方法将原始坡度剖面简化为分段常数。 原始和简化的坡度剖面如图 6 所示。请注意，尽管在优化计算中对道路坡度进行了近似，但在模拟中实现了原始坡度剖面。

Relying on the developed methodology, the route is divided into 57 segments, according to the changes of either speed limit or slope, as depicted in Fig. 7. The optimized speed profile v∗ is acquired applying the proposed optimization algorithm, with the desired travel speed vset = 23:5 m=s (84:6 km=h). Since the traffic is uncongested, no discussion is necessary on distance constraints, and thus a mission-wide constant H = 13:70 kW. The trace of the segment-constant λs is displayed in the lowest plot of Fig. 7.
依靠所开发的方法，根据限速或坡度的变化，将路线分为 57 段，如图 7 所示。应用所提出的优化算法获得优化的速度曲线 $v^{\ast }$，并具有所需的行程速度 $v_{set}=23.5m/s(84.6km/h)$。 由于交通不拥堵，因此无需讨论距离限制，因此任务范围内的常数 $H=13.70kW$。 段常数 ${\lambda }_s$ 的迹线显示在图 7 的最下面的图中。

Alongside the optimal speed profile, the algorithm calculates the control policy at the same time. The simulation result of the open-loop control implementation of Dyn-Opt, the offline calculation, is compared in Table II with Baseline. Baseline follows a reference speed profile vref(s) constant within speed limits and adopts linear-decay speed transitions between constant speed sections [15]. It aims at the simulation end time t∗ f of Dyn-Opt (see Fig. 5), which results in almost the same average speeds and allows a fair comparison. As Fig. 7 shows, Dyn-Opt takes advantage of the freedom in speed, and further the exchange of kinetic and gravity potential energy, to optimally plan the control policy of motor torque. It slightly improves the average motor efficiency. The overall smooth speed transitions and slight fluctuation in speed lead to the relative energy saving of 1:19 %.
除了最佳速度曲线外，该算法还同时计算控制策略。 Dyn-Opt 开环控制实现的仿真结果（离线计算）与 Baseline 进行了比较，如表 2 所示。Baseline遵循速度限制内恒定的参考速度曲线 $v^{ref}(s)$，并在恒速部分之间采用线性衰减速度过渡 [15]。 它的目标是Dyn-Opt的模拟结束时间 $t_f^{\ast }$（见图5），这会导致几乎相同的平均速度并允许公平的比较。 如图7所示，Dyn-Opt利用速度的自由度，进一步利用动能和重力势能的交换，来优化规划电机扭矩的控制策略。 它略微提高了平均电机效率。 整体速度过渡平稳，速度波动小，相对节能1.19%。

C. Car Following

A car following scenario is set up as in Fig. 8. To emulate a highway situation, the standard highway test cycle HWFET (from 20 to 743 s) is assigned to the preceding vehicle’s driving schedule vobj. The example scenario takes place on a route made of two segments with two speed limits in succession (23 and 27 m=s). The preceding vehicle is placed 200 m ahead of the start position. In light of (23a), driving schedule vobj is translated into distance constraint sb(t) and its inverse function tb(s).
跟车场景如图 8 所示。为了模拟高速公路情况，将标准高速公路测试循环 HWFET（从 20 秒到 743 秒）分配给前车的行驶时间表 $v_{obj}$。 示例场景发生在由连续两个速度限制的两个路段组成的路线上（23 和 27 m/s）。 前车放置在起始位置前方 200 m 处。 根据(23a)，行驶时间表 $v_{obj}$ 被转化为距离约束 $s_b(t)$及其反函数 $t_b(s)$。

The host vehicle starts with v0 = 20 m=s and ends with vf = 20 m=s. The desired travel speed in this case is set to vset = 25 m=s. As Subsection IV-B describes, the constrained solution for car following is derived iteratively, as Fig. 9 shows. After each iteration, the time headway curve (t − tb) moved upwards, and the OCP is extended with a new boundary value condition at one of the positions marked with vertical dotted lines. The final solution is found after three iterations so that the host vehicle is behind a safety headway throughout the mission, t(s) − tb(s) ≥ 0 for s 2 [s0; sf].
主车辆以 $v_0=20m/s$ 开始，以 $v_f=20m/s$ 结束。 在这种情况下，所需的行进速度设置为 $v_{set}=25m/s$。 正如第 IV-B 小节所描述的，汽车跟随的约束解是迭代导出的，如图 9 所示。 每次迭代后，车头时距曲线 $(t-t_b)$ 向上移动，并且 OCP 在垂直虚线标记的位置之一处以新的边界值条件扩展。 经过三次迭代后找到最终解决方案，使主车辆在整个任务中落后于安全车距，$t(s)-t_b(s)\ge 0$ for $s\in [s_0,s_f]$。

The calculated solution of control trajectory is given as open-loop control input to the Autonomie model to evaluate energy performance. For reference, the baseline follows the driving schedule of the preceding car, i.e., vref(s) = vobj(s), see Fig. 5. The host and preceding car are configured exactly the same, see Table I.
控制轨迹的计算结果作为 Autonomie 模型的开环控制输入来评估能量消耗性能。 作为参考，Baseline遵循前车的行驶时间表，即 $v^{ref}(s)=v_{obj}(s)$，见图 5。主车和前车的配置完全相同，见表 1。
The simulation results are compared and summarized in Table III. The host car runs slower than the target speed vset due to the slow car in front. The speed trace of the host car is much smoother than that of the preceding car, which allows the host motor to run more efficiently and results in an energy saving of 4:4 %.
模拟结果比较并总结在 表III 中。 由于前面的车慢，主车的行驶速度比目标速度 $v_{set}$ 慢。 主车的速度轨迹比前车更加平滑，主车电机运行效率更高，节能4.4%。

VI. CONCLUSION

In this paper, a computationally efficient algorithm is proposed for ecological highway cruising of electric vehicles. The algorithm is based on the PMP solution to the presented OCP. The important assumption of piecewise constant road slopes allows us to build solution trajectories for each slope in advance and improve calculation speed. For a real-world route, the slope is approximated into piecewise constants and divided into segments accordingly. Discussion on the discontinuity condition of co-state λs due to speed limits leads to additional segmentations at the positions of speed limit changes. Furthermore, in the presence of a preceding car, the host car is constrained with a distance boundary. The otherwise constant value H of the Hamiltonian jumps at activation of the distance boundary. The algorithm is thus extended with an iterative process to adjust H and find the constrained solution.
本文提出了一种用于电动汽车高速公路经济巡航的高效计算算法。 该算法基于所提出的 OCP 的 PMP 解决方案。 分段恒定道路坡度的重要假设使我们能够提前为每个坡度建立求解轨迹并提高计算速度。 对于现实世界的路线，坡度被近似为分段常数并相应地划分为段。 对速度限制导致的共状态 ${\lambda }_s$ 不连续条件的讨论导致在速度限制变化的位置进行额外的分割。 此外，在存在前车的情况下，主车受到距离边界的约束。 哈密顿量的其他恒定值 $H$ 在距离边界激活时跳跃。 因此，该算法通过迭代过程进行扩展，以调整 $H$ 并找到约束解。
The algorithm presented is evaluated through simulation in the form of open-loop control in two example scenarios. In the first scenario of uncongested highway cruising, the simulation result shows a relative energy saving of 1:2 % compared to a baseline strategy. In the second scenario, the host vehicle follows a slow preceding car with varying speed. Driving with the optimal control policy, the host car exhibits 4:4 % energy savings over the preceding car.
通过在两个示例场景中以开环控制的形式进行仿真来评估所提出的算法。 在第一个不拥堵的高速公路巡航场景中，仿真结果显示与Baseline策略相比相对节能 1.2%。 在第二种情况下，主车辆以不同的速度跟随一辆缓慢的前车。 采用最优控制策略行驶，主车比前车节能4.4%。
In future work, the proposed algorithm will be transfered to real-time control using a closed-loop framework, such as nonlinear model predictive control. Moreover, more traffic situations will be considered to enhance the usability and the reliability of the algorithm.
在未来的工作中，所提出的算法将转移到使用闭环框架的实时控制，例如非线性模型预测控制。 此外，将考虑更多的交通情况以增强算法的可用性和可靠性。