Handover Design

In this section, we provide a detailed explanation of our proposed handover scheme. First, we present a comprehensive overview of the proposed handover signaling procedure (IV-A). Then we delve into the two key designs: the synchronized algorithm between RAN and UPF (IV-B) and the computation overhead mitigation for UPF (IV-C). After that we further optimize the access satellite selection scheme to reduce the handover latency (IV-D). Finally, we introduce how we deal with the inaccurate prediction caused by user mobility and deviations in satellite trajectory prediction. (IV-E).

A. Detailed Handover Procedure

As shown in Fig. 4, similar with the conventional handover procedure, the proposed handover process begins with the handover decision. Handover is triggered by the source SgNB, which also selects the target S-gNB that be switched to. The following handover procedure can be summarized as below.

Step 1. The source S-gNB informs the target S-gNB of the handover decision; Step 2. The target S-gNB prepares for the handover such as pre-allocating channel resources. Then, the target S-gNB confirms the handover with the source S-gNB. After the source S-gNB receives the confirmation, the handover preparation is completed. It should be noted that the link between the core network and UE becomes disconnected from this point. Step 3. The third step can be further divided into two simultaneous operations. One is that the source S-gNB informs the UE of the handover decision and then UE disconnects with the source S-gNB and builds a novel RRC connection with the target S-gNB, corresponding to Step 3.a.1 and 3.a.2 in Fig. 4. The other one is that the source S-gNB transmits the relevant synchronized information with the target S-gNB, including the data to be transmitted and sequence number (SN) (Step 3.b). Step 4-9. The difference between our proposed handover scheme and the standard handover procedure is that the steps 4 to 9 (i.e., control signaling delivery between RAN and the core network) are avoided in our proposed handover scheme, which incurs a large delay latency in mobile satellite network. Instead, we propose a synchronized algorithm at the core network side which can synchronize with RAN even without interacting with RAN.

Step 10. At last, the source S-gNB is notified to release the resources. Once this is completed, the handover process is accomplished.

From the above description, the key innovation behind our proposed handover scheme is the synchronized algorithm between RAN and the core network which leverages the predictable trajectory of LEO satellites. However, it is non-trivial to achieve this synchronization yet without control signaling interaction between RAN and the core network. For example, limited prediction frequency may lead to unacceptable delay in handover. Meanwhile, the prediction operation for large quantity of ground UEs poses a huge computating pressure on the UPF.

In the following subsections, we will provide a detailed description of the synchronized algorithm, as well as how to mitigate the computation overhead, by leveraging two inherent features in LEO satellite networks—predictable trajectory and unique spatial distribution.

如图4所示，与常规切换流程相似，我们提出的切换过程始于切换决策。切换由源卫星基站（source S-gNB）触发，该基站同时也会选择将要切换到的目标卫星基站（target S-gNB）。后续的切换流程可概括如下：

步骤1. 源S-gNB将切换决策通知目标S-gNB。

步骤2. 目标S-gNB为切换做准备，例如预分配信道资源。随后，目标S-gNB向源S-gNB确认切换。源S-gNB收到确认后，切换准备阶段完成。值得注意的是，从此刻起，核心网与用户终端（UE）之间的链路变为断开状态。

步骤3. 第三步可进一步分解为两个同时进行的操作。其一是源S-gNB通知UE切换决策，随后UE断开与源S-gNB的连接，并与目标S-gNB建立新的无线资源控制（RRC）连接，对应于图4中的步骤3.a.1和3.a.2。另一操作是源S-gNB将相关的同步信息传输给目标S-gNB，其中包括待传数据和序列号（SN）（步骤3.b）。

步骤4-9. 我们提出的切换方案与标准切换流程的不同之处在于，我们避免了步骤4至9（即无线接入网与核心网之间的控制信令交互）， 这些步骤在卫星移动网络中会引发巨大的延迟。作为替代，我们在核心网侧提出了一种同步算法，该算法即便在不与无线接入网交互的情况下也能实现同步。

步骤10. 最后，通知源S-gNB释放资源。一旦完成，整个切换过程即告结束。

From the above description, the key innovation behind our proposed handover scheme is the synchronized algorithm between RAN and the core network which leverages the predictable trajectory of LEO satellites. However, it is non-trivial to achieve this synchronization yet without control signaling interaction between RAN and the core network. For example, limited prediction frequency may lead to unacceptable delay in handover. Meanwhile, the prediction operation for large quantity of ground UEs poses a huge computating pressure on the UPF.

In the following subsections, we will provide a detailed description of the synchronized algorithm, as well as how to mitigate the computation overhead, by leveraging two inherent features in LEO satellite networks—predictable trajectory and unique spatial distribution.

根据以上描述，我们所提出的切换方案其核心创新在于无线接入网与核心网之间的同步算法，该算法利用了LEO卫星轨迹的可预测性。然而，在无线接入网与核心网之间没有控制信令交互的情况下实现这种同步并非易事。例如，有限的预测频率可能导致不可接受的切换延迟。同时，对大量地面用户终端进行预测操作会给UPF带来巨大的计算压力。

在接下来的子章节中，我们将利用LEO卫星网络的两个内在特性 —— 可预测的轨迹和独特的空间分布，来详细阐述该同步算法，并说明如何缓解计算开销。

B. Synchronized Algorithm between RAN and Core Network

The most fundamental issue of achieving synchronization without control signaling interaction is the prediction of UEs’ access satellites at the core network side, whose performance is affected by two factors: relative position and weather [13], [19], [20], which can be obtained at the core network. The former can be acquired based on the UEs’ location and predictable trajectory of the satellites, while the latter can be obtained from Internet. By leveraging these information and UEs’ access strategies, the UEs’ access satellites can be accurately determined at the core network.

After obtaining UEs’ access satellites, the next issue is the asynchronization problem caused by the coarse-grained prediction. Predicting the users’ access satellites at one time point and at a fixed time interval ∆t is a straightforward approach. However, this mechanism can not bypass the asynchronization problem between the RAN and core network, resulting in additional latency during handover. Since ∆t is typically in the order of hundreds of milliseconds due to the computing complexity of prediction, the additional latency can severely impede the entire handover process.

Alternatively, we investigate the case of predicting the users’ access satellites at two time points simultaneously. We find that this mechanism can solve the above asynchronization problem. Detailed description is shown as follow. Assume that t 0 and t 1 = t 0 + ∆t are two consecutive time points at which the UPF performs predictions. We use U to refer to all UEs served by the UPF, and use A t to represent the set of access satellites for all users U at time t. By reasonably selecting ∆t, we ensure that at most one handover is triggered for each UE u ∈ U between t 0 and t 1 . As a result, by comparing A t 0 [u] and A t 1 [u]—the access satellites of user u at time t 0 and t 1 , we can determine whether a handover will be triggered during t 0 to t 1 . If a handover will be triggered, the UPF needs to predict the accurate handover triggering time to minimize the asynchronization duration between RAN and the core network. To avoid any additional delay, the predicted handover triggering time t p should lie between the handover trigger and the handover completion in the RAN. To achieve this aim, we can employ the simple yet effective binary search method to pinpoint this time point.

It should be noted that the above results are only discussed in common cases. In some special cases when users register, deregister, or travel fast, the corresponding access satellites change accordingly, which should be taken into account in the synchronized algorithm design.

Consequently, we propose a synchronized algorithm to ensure the synchronization between the RAN and the core network for UPF, as shown in Alg. 1. Specifically, we define T as the last time point when the periodic update is completed, and T p as the prediction set of handover triggering time for each UE. The computation results including A T , A T +∆t and T p are stored in the table R. In the subsequent discussion, we will refer to A T [u] as the access satellite of u, and A T +∆t [u] as the next-access satellite of u. The proposed synchronized algorithm considers two update cases according to the reason which incurs the update.

• Periodic update caused by satellite movement: Due to the fast travelling speed of LEO satellites, the synchronized algorithm iteratively performs a periodic update with a periodicity of ∆t. Below, we provide a step-by-step description of the periodic update for T + ∆t as an example. First, we obtain the access satellites of all UEs at T + ∆t—A T +∆t from R, and then predict the access satellites at T +2∆t—A T +2∆t according to the predicted trajectory of LEO satellites. According to the difference between A T +∆t and A T +2∆t , we employ binary search algorithm to calculate the set of predicted handover triggering time T p . Specifically, for each UE u with distinct values of A T +∆t [u] and A T +2∆t [u], we compute their associated access satellites at the intermediate time instant, i.e., T + 1.5∆t, to halve the prediction error for access time. This process is repeated iteratively until the error is less than the time required for the handover procedure at RAN. To avoid the premature update, we update the table R when the time reaches T + ∆t.

• Update caused by UE: When the UE u undergoes a localization change, which may be caused by the registration, deregistration, or movement, the algorithm will also perform an update for UE u. At this point, our proposed synchronized algorithm updates A T , AT +∆t and T p in the table R.

Next, we describe how we put the synchronized algorithm into the mobile satellite network system to achieve synchronization between RAN and the core newtork.

Actually, main modifications take place in UPF. As mentioned above, UPF maintains the table R and updates it through the synchronized algorithm. As shown in Fig. 5, R consists of user information corresponding to U, access information corresponding to A T , and next-access information corresponding to A T +∆t and T p . Notably, if there is no handover for user u between T and T +∆t, the corresponding next-access satellite remains the same as the access satellite, and t p = T p [u] = 0. When there is a packet that needs to be transmitted to UE, UPF can obtain related information such as TEID sat from table R and then feed them into the protocol header. In addition, in the UE registration/deregistration procedure, UPF triggers the synchronized algorithm to update R.

To maintain the synchronization, a few modifications are made at S-gNB. In particular, the UEs’ location information is required to be fed into the GTP Extension Header of uplink data, in order to notify the UPF of changes in UEs’ location.

Finally, we explore the appropriate value set for the update interval ∆t, which should ensure that two consecutive handovers for any stationary UE do not occur within ∆t. A feasible solution is to set ∆t to be less than constellation’s minimum service time, which is influenced both by the satellite constellation configuration and the perhaps additional limitation when selecting access satellites (to avoid excessive frequent handovers). Based on the analysis of existing satellite constellation configurations and actual measurements [1], [15], [21], [22], we set the update interval ∆t to 5 seconds. Consequently, by iterating the binary search process 9 times, we can attain a prediction accuracy as precise as 10 milliseconds, meeting the specified requirements.

在无控制信令交互的情况下实现接入网（RAN）与核心网同步的最根本问题在于：核心网需要预测用户设备（UE）所接入的卫星。该预测过程的准确性受两个因素影响，即相对位置与气象条件［13］、［19］、［20］，而这两者均可由核心网侧获取。其中，相对位置可通过 UE 的位置信息与可预测的卫星轨迹推导获得，而气象条件可通过互联网获取。借助这些信息及 UE 的接入策略，核心网能够较为准确地判断 UE 所接入的卫星。

获取 UE 的接入卫星后，下一项挑战是由粗粒度预测引起的异步问题。一种直观的预测方法是在某一时间点以固定时间间隔 ∆t 对用户接入卫星进行预测。然而，该机制无法消除 RAN 与核心网之间的异步问题，导致切换过程中出现额外时延。由于 ∆t 的典型量级通常为数百毫秒，这是由预测计算复杂性决定的，这种附加时延会严重影响整个切换过程。

为此，我们提出了一种替代机制：在两个时间点同时预测用户的接入卫星。实践表明，该机制可有效解决上述异步问题。具体描述如下。设 t0t0​ 和 t1=t0+Δtt1​=t0​+Δt 为 UPF 进行预测的两个连续时间点，令 UU 表示当前由 UPF 服务的所有 UE，AtAt​ 表示在时刻 tt 时，所有 UE 的接入卫星集合。通过合理设置 ∆t，可以确保任一 UE u∈Uu∈U 在 t0t0​ 到 t1t1​ 之间最多触发一次切换。因此，比较 At0[u]At0​​[u] 与 At1[u]At1​​[u] ——即用户在两个时间点的接入卫星，可判断该 UE 是否会在此区间内发生切换。若发生切换，则 UPF 需进一步预测准确的切换触发时刻 tptp​，以最小化 RAN 与核心网之间的异步时间。为避免附加延迟，tptp​ 应位于 RAN 切换过程中的切换触发与切换完成之间。为此，我们可采用简单且高效的二分搜索方法精确定位 tptp​。

需注意，以上分析基于常规情形。在用户注册、注销或高速移动等特殊情况下，其接入卫星也会发生变化，此类情况亦应纳入同步算法设计的考虑范围。

因此，我们提出了一种用于保障 RAN 与核心网（特别是 UPF）之间同步的同步算法，如算法 1 所示。具体地，我们定义 TT 为上次周期性更新完成的时间点，TpTp​ 为每个 UE 的预测切换触发时间集合。计算结果，包括 ATAT​、AT+ΔtAT+Δt​ 和 TpTp​，将被存储于表 RR 中。在后续描述中，我们将 AT[u]AT​[u] 视为 UE u 当前的接入卫星，而 AT+Δt[u]AT+Δt​[u] 为其下一阶段可能接入的卫星。

该同步算法根据触发更新的原因分为两类更新情况：

• 由卫星移动引起的周期性更新：由于 LEO 卫星高速运行，同步算法需以 ∆t 为周期不断迭代更新。以下以 T+ΔtT+Δt 为例说明周期性更新流程。首先，从表 RR 中获取所有 UE 在 T+ΔtT+Δt 时刻的接入卫星 AT+ΔtAT+Δt​，再根据卫星预测轨迹推测其在 T+2ΔtT+2Δt 时刻的接入卫星 AT+2ΔtAT+2Δt​。若发现某个 UE 在两个时间点接入卫星发生变化，则应用二分搜索法计算其预测的切换触发时间 tptp​。具体而言，针对满足 AT+Δt[u]≠AT+2Δt[u]AT+Δt​[u]=AT+2Δt​[u] 的用户 u，我们计算其在中间时间点 T+1.5ΔtT+1.5Δt 时的接入卫星，以将预测误差减半。此过程持续迭代，直到预测误差小于 RAN 侧切换所需时间为止。为避免过早更新，仅在时间达到 T+ΔtT+Δt 时更新表 RR

• 由 UE 引发的更新：当 UE u 的位置信息发生变化（如注册、注销或移动）时，算法将执行针对 u 的更新操作。此时，同步算法更新表 RR 中关于 u 的 ATAT​、AT+ΔtAT+Δt​ 与 TpTp​

接下来，我们说明如何将同步算法集成进移动卫星网络系统，以实现 RAN 与核心网之间的同步。

实际上，主要修改集中在 UPF 模块。如前所述，UPF 需维护表 RR 并使用同步算法进行更新。图 5 所示的表 RR 包含三类信息：UE 相关信息（U）、当前接入卫星（ATAT​）、下一阶段接入卫星与预测切换时间（AT+ΔtAT+Δt​ 与 TpTp​）。需要指出的是，若在 TT 到 T+ΔtT+Δt 之间，某个用户未触发切换，则其下一阶段接入卫星保持不变，且 tp=Tp[u]=0tp​=Tp​[u]=0。当 UPF 需向某 UE 发送数据包时，可从表 RR 中获取相关信息（如 TEID_sat），并写入协议头部。此外，在 UE 注册/注销过程中，UPF 会调用同步算法以更新表 RR。

为维持同步，S-gNB 也需进行一定修改。具体而言，UE 的位置信息需写入上行数据的 GTP 扩展头中，以通知 UPF 用户位置的变化。

最后，我们探讨更新周期 ∆t 的合理取值。其应满足任一静止 UE 在 ∆t 时间内不发生两次连续切换。一个可行方案是将 ∆t 设置为星座最小服务时间以下，该值受卫星星座配置与接入策略（如避免过度频繁切换）影响。基于对现有卫星星座配置的分析与实测数据［1］、［15］、［21］、［22］，我们将 ∆t 设置为 5 秒。据此，通过 9 次二分搜索迭代，可将预测精度提高至 10 毫秒，满足系统对切换时延控制的精度需求。

看起来很复杂, 用人话概括一下

一个基于预测的、无线接入网（RAN）与核心网（Core Network）之间的去耦合同步方案

从 “被动响应” 变成了 “主动预测”，从而消灭了为了同步信息而来回通信所浪费的时间

1. 核心思想: 核心网的UPF不再被动等待RAN的切换通知，而是利用其可获得的全局信息（卫星轨道、用户位置），主动预测切换的发生
2. 两步式精准预测:
   1. 切换检测: UPF通过计算未来两个时间点（如 t 和 t+∆t）用户所连接的卫星是否不同, 来判断此时间段内是否会发生切换
   2. if sat(t) == sat(t+∆t): 没发生切换, 直接传递即可
   3. else: 发生了切换, 现在我们想找到切换到 sat(t+∆t) 的具体时间, 如 t + 0.3∆t (使用二分!)
3. 预先配置，即时生效:
   1. UPF将预测出的精确切换时间点和新的卫星信息预先存储
   2. 当数据包到达时，UPF根据当前时间，直接查询预配置的结果，将数据发往正确的卫星路径

算法核心

注意每次必须要先做前置工作!!!

因为变化因素有以下两种:

1. Periodic update caused by satellite movement
   1. 时间区间更新
   2. 实行对比: ok or 二分查找
2. Update caused by UE
   1. 所属“接入卫星”更新
   2. 时间区间更新
   3. 实行对比: ok or 二分查找

C. Optimization for Prediction Algorithm

According to the previous discussions, each UPF needs to predict the access satellites for all UEs it serves every ∆t. A directly prediction approach, which considers all satellites in the constellation for every UE to predict the access satellite, would impose significant computational pressure and make it difficult to complete within the specified ∆t.

To address this challenge, we introduce the fast access satellite prediction algorithm. First, we take into account the access strategies of UEs to reduce the number of users that need to be considered. We categorize the existing access strategies into two types: consistent and flexible [19], [20]. The former maintains a connection until the user leaves the service coverage area, while the latter may handover even when UE is within its access satellite’s coverage area. For UEs with consistent access strategies, the algorithm simply checks if the previously connected satellite can still provide service, significantly reducing the number of UEs required consideration.

Additionally, we reduce the number of satellites in each UE’s access satellite prediction, according to satellites’ spatial distribution. We divide the Earth into several continuous rectangular blocks based on the satellite’s service radius of the satellites, and assign satellites into corresponding blocks according to their position. As a result, for each UE, only nearby satellites required computation, which greatly reduces computation complexity.

The pseudo code of the proposed algorithm is shown in Alg. 2. Specifically, the algorithm can be divided into 4 steps as following:

Step 1: Based on the ephemeris, we predict S t ′ , the satellites’ position at time t ′ .

Step 2: To different selection strategies, we employ distinct approaches to construct the UE candidate set U C , which consists of UEs that may undergo handover. For UEs with consistent strategy, we only add those whose access satellite is not available at t ′ into the set U C for further processing. For UEs with flexible strategy, we add all UEs to the set U C .

Step 3: We divide the Earth into several rectangular blocks with the satellite service radius as side length. All satellites are allocated to correspondingly blocks according to their position at time t ′ .

Step 4: For each user in U C , we calculate the access satellite it should be connected to at time t ′ based on its access strategy.

根据前述讨论，每个用户平面功能（UPF）实体需要在每个时间间隔 ∆t 内，为其服务的所有用户终端（UEs）预测接入卫星。一种直接的预测方法是，为每个UE遍历星座中的所有卫星来预测其接入卫星，但这会带来巨大的计算压力，并使其难以在规定的 ∆t 时间内完成。

为应对此挑战，我们引入了一种快速接入卫星预测算法：

首先，我们考虑用户终端的接入策略，以减少需要被纳入考量的用户数量。我们将现有的接入策略分为两类：持续性策略（consistent）和灵活性策略（flexible）[19], [20]。前者会保持连接直至用户离开当前卫星的服务覆盖区，而后者即使用户仍在原接入卫星的覆盖范围内也可能进行切换。

对于采用持续性接入策略的用户，该算法仅需检查其先前连接的卫星是否仍能提供服务，这显著减少了需要进一步处理的用户数量。

此外，我们根据卫星的空间分布特性，减少了为每个UE进行接入卫星预测时所需考虑的卫星数量。我们基于卫星的服务半径，将地球表面划分为若干个连续的矩形区块，并根据卫星的位置将其分配到相应的区块中。如此一来，对于每个UE，仅需对邻近区块的卫星进行计算，从而极大地降低了计算复杂度。

该算法的伪代码如算法2所示。具体而言，该算法可分为以下4个步骤：

步骤1： 基于星历（ephemeris），我们预测出卫星在 t' 时刻的位置 S_t'

步骤2： 针对不同的选择策略，我们采用不同的方法来构建可能发生切换的用户候选集 U_C

• 对于采用持续性策略的用户，我们仅将那些在 t' 时刻其原接入卫星已无法提供服务的用户加入集合 U_C 以进行后续处理
• 对于采用灵活性策略的用户，我们将所有用户都加入集合 U_C。

步骤3： 我们以卫星服务半径为边长，将地球表面划分为若干矩形区块。所有卫星根据其在 t' 时刻的位置被分配到相应的区块中

步骤4： 对于候选集 U_C 中的每一个用户，我们根据其接入策略，计算出在 t' 时刻它应该连接的接入卫星

形象化理解

任务是：每隔5秒钟，就要预测出哪些 顾客(UE) 需要更换 外卖小哥(Sat)，以保证送餐不中断

如果每5秒都为这100万个顾客，把天空中几千个外卖小哥的情况全部计算一遍，电脑肯定会当场冒烟死机...

所以! 必须想办法偷懒，只去计算那些最有可能需要换人的顾客!

1. 候选集 (\(U_C\))
   1. 为了偷懒而创建的“重点观察名单”
   2. 我们不会去分析全部100万个顾客，而是先通过一些非常简单的方法，快速筛选出一个小名单
2. “持续性策略” 和 “灵活性策略” 对应于用户的个性:
   1. 持续性策略:
      1. “不爱折腾”
      2. 只要当前为他服务的外卖小哥A还没飞出服务范围，他就会一直用小哥A，哪怕天上有个信号更好、速度更快的小哥B飞过，他也不会主动换人
   2. 灵活性策略:
      1. “非常精明”
      2. 他会实时评估天空中所有能为他服务的外卖小哥。虽然小哥A还在为他服务，但只要他发现旁边飞过的小哥B信号更强、离得更近，他就会立刻、主动地要求换成B
3. 不同策略的考察名单逻辑:
   1. 对"持续性策略"
      1. 你只需要问一个问题：“在5秒后，他现在的小哥A还在不在服务区？”
      2. 如果还在，那就ok，肯定不会换
      3. 如果不在，那他肯定换了，将其加入“候选集”
   2. 对"灵活性策略"
      1. 稳妥起见, 全部纳入考察名单"候选集"

D. Optimization on Access Satellite Selection

To further reduce the overall handover latency, we focus on minimizing the latency between source S-gNB and target S-gNB, which is the time taken for the handover process. Based on the preliminary experiments shown in Fig. 3, we can observe a sudden increase in latency in 30-40 ms. The main reason is that the satellite switches in contrary directions (i.e., handover from a northern-direction satellite to a southerndirection satellite, or vice versa) [23], [24]. To this end, we make an additional constraint that handovers can happen only between satellites in the similar-direction ones, which is expected to reduce the propagation delay between satellites by at most more than 200ms.

为了进一步降低整体切换延迟，我们着重于最小化源S-gNB与目标S-gNB之间的延迟，即切换过程本身所耗费的时间。根据图3所示的初步实验，我们可以观察到延迟在30-40毫秒时出现了一次突增。其主要原因是切换发生在行进方向相反的卫星之间（例如，从一个向北飞行的卫星切换到一个向南飞行的卫星，反之亦然）[23], [24]。为此，我们 施加一个额外约束，即切换只能在行进方向相似的卫星之间进行。预计这一约束最多可将卫星间的传播延迟降低超过200ms。

E. Inaccurate Prediction

User mobility and the deviations in satellite trajectory prediction may cause inaccurate predictions of UEs’ access satellites, resulting in what we refer to as ‘abnormal handovers’, where the switching target is not the predicted satellite. This will negatively affect the availability of our proposed scheme. Therefore, we discuss the impact of these two causes and propose solutions.

用户移动性以及卫星轨道预测的偏差，可能会导致对用户终端（UE）接入卫星的预测不准确 ，从而引发我们称之为“异常切换”的情况 —— 即实际的切换目标并非原先预测的卫星。这将对我们所提方案的可用性产生负面影响。因此，我们在此讨论这两种成因的影响并提出解决方案。

On one hand, in the context of users, despite their significantly lower speed compared to LEO satellites, there exist some cases where user mobility results in inaccurate prediction at the core network and leads to an abnormal handover. Such cases may occur when user move at high speeds and is located at the overlapping region of the satellites’ service coverage, as illustrated in Fig. 6. It is crucial to highlight that these scenarios are infrequent, even for UEs who are moving at high speeds. Consequently, it can be regarded that user mobility has little impact on the user experience during the handover process. In V-B, we will delve into the probability of abnormal handover using experimental data.

The perturbations in satellite motion may result in errors of several kilometers in trajectory prediction based on dailyupdated ephemeris data [25], consequently causing abnormal handovers with a probability of approximately 10 −3 . To tackle this problem, one solution is to employ short-term trajectory prediction based on minute-level updated satellite ephemeris data [26], whose prediction precision reaches 10 centimeters.

To deal with abnormal handovers, the S-gNB triggers the standard 5G NTN handover procedure, transmitting control signaling to the core network and ultimately updating the UE information at the UPF.

一方面，在用户层面，尽管用户的移动速度远低于LEO卫星，但在 某些情况下，用户移动性仍会导致核心网的预测不准确，并引发异常切换。 如图6所示，这类情况可能发生在用户高速移动且恰好位于多个卫星服务覆盖的重叠区域时。必须强调的是，即使对于高速移动的UE，这些场景也并不常见。因此，可以认为用户移动性对切换过程中的用户体验影响甚微。在章节V-B中，我们将利用实验数据深入探讨异常切换发生的概率。

另一方面，卫星运动中的 摄动（perturbations）可能导致基于每日更新的星历数据所做的轨道预测产生数公里的误差 [25]，从而以大约 10−3 的概率引发异常切换。为解决此问题，一种方案是采用基于分钟级更新的卫星星历数据进行短期轨道预测[26]，其预测精度可达10厘米。

为了处理已发生的异常切换，S-gNB会触发标准的5G NTN（非地面网络）切换流程，向核心网传输控制信令，并最终在UPF处更新该UE的信息。

保底措施

用户移动性以及卫星轨道预测的偏差，可能会导致对用户终端（UE）接入卫星的预测不准确

此时, 原来固有的信令传递机制作为“兜底” 🚀 🚥

1. 用户移动性: special judge "用户高速移动且恰好位于多个卫星服务覆盖的重叠区域"
2. 卫星轨道预测的偏差: 摄动（perturbations）可能导致基于每日更新的星历数据所做的轨道预测产生数公里的误差