Forecasting-Driven Stable Successor Matching for UAV-Assisted Continuous Edge Services ∗
Houyi Qi∗ , Minghui Liwang∗ , Yuhan Su† , Xianbin Wang‡
arXiv:2605.02180v1 [cs.NI] 4 May 2026
Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai, China † Sch. of Electronic Science and Engineering, Xiamen University, Xiamen, China ‡ Dept. of Electrical and Computer Engineering, Western University, Ontario, Canada Email: {[email protected], [email protected], [email protected], [email protected]} Abstract—Continuous and reliable service support is crucial for emerging latency-sensitive and computation-intensive applications in UAV-assisted edge networks (UENs) due to operational dynamics and environmental uncertainty. Although conventional designs can improve coverage and computing efficiency, they often rely on instantaneous resource optimization or reactive handover, rendering ongoing services vulnerable to non-negligible interruptions when the serving UAV degrades due to mobility, energy depletion, or channel dynamics. To avoid such postfailure recovery, a promising approach is to prepare a successor UAV in advance, i.e., a standby UAV that reserves minimal resources and synchronizes service context for possible takeover. Thus, we consider a dynamic UEN architecture where each mobile user carries an ongoing computing mission requiring persistent service support, while UAVs provide wireless access and computing services under time-varying network dynamics and stringent onboard energy constraints. To facilitate proactive and continuous service provisioning, we propose a forecastingdriven proactive reservation-based continuous service scheduling framework, termed Fresco. In Fresco, an LSTM-based module is first used to predict short-term disruption risks of ongoing missions from historical network observations. Guided by these predictions, an online risk-aware successor matching scheme selects suitable standby UAVs for high-risk missions under delay, resource, and energy constraints, while incorporating minimal communication/computation reservation and lightweight servicecontext synchronization for efficient takeover preparation. Experiments show that Fresco significantly reduces service interruptions and improves mission continuity over reactive and nonpredictive baselines, with only modest reservation overhead. Index Terms—UAV-enabled edge networks, continuous services, successor reservation, service continuity, stable matching
I. I NTRODUCTION Low-altitude aerial-assisted edge computing has emerged as a compelling paradigm for extending intelligent services to dynamic or infrastructure-constrained environments. Among various low-altitude intelligent platforms, unmanned aerial vehicles (UAVs) are the most prevalent due to their high mobility, deployment flexibility, and favorable air-to-ground communication characteristics [1]. Recent studies have demonstrated their effectiveness in real-time video analytics, emergency response, and distributed artificial intelligence (AI) [2], all of which impose increasingly stringent requirements on communication latency and service continuity. Nevertheless, within highly dynamic UAV-assisted edge networks (UENs), supporting continuous services remains challenging, mainly due to the aggregated impact of node mobility, constrained onboard energy, stochastic channel fluctuations, and topology reconfiguration on operational stability and service continuity. Purely reactive re-association or post-failure re-optimization often leads to non-negligible
interruptions and degraded user experience [3], [4]. For instance, existing works mainly focus on instantaneous communication/computation optimization, including UAV deployment/association [5], mobility-aware trajectory design [6], and resilient coordination for coverage and connectivity [7]. Recent efforts further explore seamless handover, proactive relocation, and mobility-prediction-assisted service migration [3], [4], [8]–[11]. Nevertheless, existing studies lack proactive designs preparing takeover-ready successor UAVs, i.e., standby UAVs that reserve minimal resources and synchronize service context in advance to take over ongoing missions when their serving UAVs become unreliable1 . Without such pre-reserved successor UAVs, service disruption may occur during re-association in dynamic operational environments. Hence, successor reservation with minimal resource guarantees and lightweight context synchronization remains critical, yet largely underexplored under coupled resource-energy dynamics. Motivated by the above gap, we propose a forecastingdriven reservation-based continuous service scheduling framework, termed Fresco, for dynamic UENs. In Fresco, each mission is initially associated with a primary UAV. For missions exhibiting elevated disruption risk, a successor UAV is proactively reserved prior to severe service degradation. The reserved successor maintains lightweight communication and computation footprints while progressively synchronizing service context, thereby ensuring seamless and low-latency takeover once the primary UAV becomes unreliable. Main contributions are summarized as follows. • To mitigate service-continuity degradation caused by UAV mobility, energy depletion, and channel dynamics over UENs, we formulate proactive successor preparation for serving continuous and smooth edge services as an online reservation problem constrained by takeover delay, available resources, and residual-energy constraints. • We then design Fresco, which couples LSTM-assisted highrisk mission identification with constraint-aware successor matching, enabling takeover-ready UAV successors through lightweight resource reservation and progressive servicecontext synchronization. • We theoretically prove the key properties of Fresco, including feasibility, individual rationality, and constrained stability under the adopted blocking-pair definition. Simulation results further demonstrate that Fresco significantly improves service continuity and takeover effectiveness over representative base1 In dynamic UENs, a mission constitutes a continuous edge-computing process requiring sustained connectivity and computation, yet its serving UAV may deteriorate due to mobility, energy depletion, channel variation, or resource congestion.
Fig. 1. Framework of Fresco over UENs, where LSTM-based short-term disruption prediction, matching-based successor reservation, and progressive context synchronization are jointly designed to support proactive and continuous edge services.
lines with only modest reservation overhead. II. S YSTEM OVERVIEW AND M ODELING A. Overview We consider a UEN architecture composed of a set of mobile users (MUs) and a set of UAVs, where UAVs provide wireless access and edge computing services under time-varying channel conditions, MU mobility, and battery constraints. Let U = {u1 , . . . , u|U | } denote the MU set, and let N denote the overall UAV set. Moreover, the system evolves over the discrete time horizon T = {1, . . . , T } with timeslot length τ , and each MU is assumed to maintain one ongoing computing mission. At each timeslot t, the UAV set is partitioned into two disjoint role sets, i.e., N = N srv (t) ∪ N cand (t) and N srv (t) ∩ N cand (t) = ∅, where N srv (t) denotes the set of UAVs currently serving ongoing missions and N cand (t) denotes the set of candidate successor UAVs. This mutually exclusive role partition is introduced for tractability, such that each UAV operates in only one role within a timeslot. As illustrated in Fig. 1, Fresco supports continuity-oriented edge services by proactively reserving successor UAVs for high-risk ongoing missions with limited preparation overhead. Each mission is associated with at most one serving UAV, while a high-risk mission can reserve at most one candidate successor UAV. Rather than relying on post-failure recovery, the reserved successor does not immediately replace the current serving UAV; instead, it maintains lightweight communication/computation reservation and progressively synchronizes the service context to prepare for possible takeover. Specifically, the left part of Fig. 1 shows the UAV-assisted service scene, the middle part shows the LSTM-based risk prediction module, and the right part shows the successor matching result under communication, computation, synchronizationlink, and battery constraints. Once a successor is reserved, the candidate UAV starts reservation and context synchronization, while the current serving UAV continues service whenever possible. Therefore, Fresco follows a prediction–reservation– synchronization–takeover workflow, and proactive continuity support in the considered UEN is modeled as a resourceconstrained successor reservation problem between high-risk current serving pairs and candidate UAVs. B. Basic Modeling At timeslot t, mission ui is characterized by the tuple {Direm (t), Cirem (t), Timax , ζi (t)}, where Direm (t) and Cirem (t) denote the remaining input data size and remaining computation workload, respectively, Timax is the maximum tolerable service completion delay after takeover, and ζi (t) indicates the service-context size to be synchronized for successor takeover. The current serving UAV nsj (t) ∈ N srv (t) is characterized by {Ejs (t), qsj (t)}, where Ejs (t) denotes its residual battery energy and qsj (t) denotes its position. While
the candidate UAV nck (t) ∈ N cand (t) is defined by {Bkava (t), Fkava (t), Ekc (t), qck (t), Vkmax }, where Bkava (t) and Fkava (t) describe the available communication and computation resources for successor reservation, respectively. Besides, Ekc (t) denotes the residual battery energy, qck (t) shows the position, and Vkmax is the maximum flying speed. To explicitly capture the reservation relationship, define the current serving association variable xi,j (t) ∈ {0, 1}, where xi,j (t) = 1 indicates that mission ui is currently served by UAV nsj (t) at timeslot t. Further define the successor reservation variable zi,j,k (t) ∈ {0, 1}, where zi,j,k (t) = 1 denotes that candidate UAV nck (t) is reserved as the successor of mission ui currently served by nsj (t). Hence, P nc (t)∈N cand (t) zi,j,k (t) ≤ xi,j (t). k
III. K EY M ODELING , P ROBLEM D ESCRIPTION , AND M ETHODOLOGY D ESIGN A SSOCIATED WITH F RESCO A. Communication and Takeover Model For mission ui currently served by UAV nsj (t), the primary service allocation is assumed to be pre-determined by an underlying edge scheduler and is therefore treated as a given state input. If candidate UAV nck (t) is reserved as its successor, acc it allocates takeover access bandwidth B̃i,k (t) and computation resource F̃i,k (t) to mission ui . The corresponding post cand acc tk (t) . (t) log2 1 + γi,k (t) = B̃i,k takeover access rate is ri,k Since the service context is stored at the current serving UAV, context synchronization is performed from nsj (t) to nck (t) with syn syn syn synchronization rate Ri,j,k (t) = Bi,j,k (t) log2 1 + γj,k (t) . Let si,j,k (t) ∈ [0, 1] denote the context synchronization ratio of mission ui between serving UAV nsj (t) and candidate UAV nck (t) at the beginning of timeslot t. Since takeover is typically triggered during service degradation rather than after complete service collapse, we assume that the current serving UAV can still support residual synchronization within a short takeover window. Therefore, if nck (t) takes over mission ui , the posttakeover completion delay is modeled as (1 − si,j,k (t))ζi (t) Direm (t) Cirem (t) + tk + . (1) dtk syn i,j,k (t) = Ri,j,k (t) ri,k (t) F̃i,k (t) Accordingly, the context synchronization ratio evolves ( ) as syn Ri,j,k (t)τ si,j,k (t + 1) = min 1, si,j,k (t) + zi,j,k (t) . (2) ζi (t) The local reservation process is subject to the communication/computation budgets and residual-energy constraints of candidate UAVs, together with the synchronization-link capacity constraints between serving and candidate UAVs. These constraints are explicitly incorporated into the online reservation problem in Sec. III-C. B. Utility and Feasibility Characterization Before characterizing the reservation utility, we define the high-risk mission set at n timeslot t as o (H) R U (t) = ui ∈ U | ξˆ (t) ≥ ξ th , (3) i
(H) where ξˆi (t) denotes the predicted disruption probability of mission ui within the future horizon H, and ξ th is the risk threshold. For each high-risk mission currently under service, i.e., ui ∈ U R (t), at most one candidate successor UAV can be reserved. For each candidate triplet (ui , nsj (t), nck (t)), we define a minimum reserved-resource template, denoted
syn,min acc,min min by Bi,j,k (t), B̃i,k (t), F̃i,k (t) , which represents the minimum synchronization, takeover-access, and computation resources required to support feasible successor takeover. Specifically, this template is obtained by minimizing the weighted reservation cost while satisfying the post-takeover delay requirement dtk ≤ Timax as well as the i,j,k (t) synchronization-link, access, and computation feasibility constraints. If no such template exists, then candidate UAV nck (t) is excluded from the feasible reservation space of the current serving pair (ui , nsj (t)). Accordingly, let Ci,j (t) denote the feasible candidate-successor set of current serving pair (ui , nsj (t)), which contains all candidate UAVs satisfying the minimum synchronization, access, and computation requirements, together with the battery-safety condition min Ekc (t) − eres i,j,k (t) ≥ Ek . The mission-side utility of current serving pair (ui , nsj (t)) toward candidate successor nck (t) is defined as dtk,min i,j,k (t) + β2 si,j,k (t) Timax # syn,min syn Bi,j,k (t) log2 1 + γj,k (t) τ res − β4 Ci,j,k (t), + β3 ζi (t) !
"
(H)
U Ui,j (nck (t)) = ξˆi
(4)
(t) − η2
syn,min Bi,j,k (t) link Bj,k (t) + ϵ
acc,min min B̃i,k (t) (t) F̃i,k eres i,j,k (t) − η4 ava − η5 c , − η3 ava Bk (t) + ϵ Fk (t) + ϵ Ek (t) + ϵ
zi,j,k (t)wi,j,k (t).
ui ∈U R (t) nsj (t)∈N srv (t) nc (t)∈N cand (t) k
(7)
The considered time-wise successor reservation problem is then formulated as follows: P : max W(t) Z(t) X s.t.
(8) zi,j,k (t) ≤ xi,j (t),
(8a)
nc (t)∈N cand (t) k
X
acc,min zi,j,k (t)B̃i,k (t) ≤ Bkava (t), (8b)
ui ∈U R (t) nsj (t)∈N srv (t)
X
X
min zi,j,k (t)F̃i,k (t) ≤ Fkava (t),
(8c)
ui ∈U R (t) nsj (t)∈N srv (t)
where dtk,min i,j,k (t) denotes the post-takeover delay evaluated under the minimum reserved-resource template, and acc,min syn,min min B̃i,k (t) Bi,j,k (t) F̃i,k (t) res Ci,j,k (t) = ν1 link + ν2 ava + ν3 ava , Bk (t) + ϵ Fk (t) + ϵ Bj,k (t) + ϵ with ϵ > 0 being a small positive constant. Similarly, the UAVside utility of candidate UAV nck (t) when accepting current serving pair (ui , nsj (t)) is defined as (H)
W(t) =
X
(t) β1 1 −
UkN (ui , nsj (t)) = η1 ξˆi
maximize the overall reservation benefit under communication, computation, synchronization-link, and battery constraints. For each candidate triplet (ui , nsj (t), nck (t)), define the pairwise U reservation welfare as wi,j,k (t) = Ui,j (nck (t))+UkN (ui , nsj (t)). Let Z(t) denote the set of all binary successor-reservation decisions, where zi,j,k (t) ∈ {0, 1} for ui ∈ U R (t), nsj (t) ∈ N srv (t), and nck (t) ∈ N cand (t). Accordingly, the total reservation welfare X at timeslot X t is defined X as
(5)
where η1 , η2 , η3 , η4 , η5 ≥ 0 are weighting coefficients. For each current serving pair (ui , nsj (t)), we retain only those candidate UAVs that are both feasible and individually acceptable, and define the corresponding preference set as U Pi,j (t) = nck (t) ∈ Ci,j (t) Ui,j (nck (t)) ≥ 0 . (6) Since matched triplets reserve the minimum takeover template by design, a reserved successor is regarded as takeover-ready once its synchronization ratio satisfies si,j,k (t) ≥ smin , where smin is the minimum synchronization threshold. In addition, (H) successor takeover is triggered when ξˆi (t) ≥ ξ alarm , or when the current serving UAV experiences severe battery or link degradation, where ξ alarm > ξ th is the alarm threshold. C. Problem Formulation Although the considered system evolves over time, the successor reservation decision is made in a slot-by-slot manner due to time-varying resource states, uncertain future disruptions, calling for lightweight real-time implementation. The cross-timeslot effect is captured implicitly via synchronizationstate evolution in (2), dynamic battery/resource states, and finite-horizon disruption-risk prediction. Accordingly, at timeslot t, we study an online successor reservation problem for high-risk ongoing missions. The objective is to determine which candidate UAVs should be proactively reserved as successors for the current high-risk serving pairs, so as to
X
link syn,min (t) ≤ Bj,k (t), zi,j,k (t)Bi,j,k
(8d)
ui ∈U R (t)
Ekc (t) −
X
X
min zi,j,k (t)eres , (8e) i,j,k (t) ≥ Ek
ui ∈U R (t) nsj (t)∈N srv (t)
zi,j,k (t) = 0, ∀nck (t) ∈ / Pi,j (t).
(8f)
Constraint (8a) guarantees that each current serving pair reserves at most one candidate successor. Constraints (8b) and (8c) limit the access communication and computation resources of candidate UAVs, respectively. Constraint (8d) limits the synchronization-link capacity between serving UAVs and candidate UAVs. Constraint (8e) ensures that each candidate UAV satisfies the minimum residual-energy requirement after reservation. Constraint (8f) enforces that successor reservation can only be performed within feasible and individually acceptable candidate sets. Due to binary reservation decisions and resource constraints, P is combinatorial and costly to solve optimally in each timeslot. Therefore, Fresco seeks a feasible, individually rational, and stable reservation outcome through LSTM-assisted high-risk mission identification and constraintaware successor matching. D. Methodology Design Based on the above feasibility characterization and problem formulation, we design Fresco to construct an online successor reservation outcome in dynamic timeslots. Fresco consists of two tightly coupled components: (i) LSTM-assisted high-risk mission identification, which predicts which ongoing missions are likely to become unsustainable in the near future; and (ii) constraint-aware successor matching, which determines admissible candidate successor UAVs for the identified high-risk current serving pairs under local communication, computation, synchronization-link, and battery constraints. To support reservation, Fresco first identifies high-risk ongoing missions through an LSTM-based predictor. For mission ui , the predictor observation takes a length-L historical sequence Xi (t) = fi (t − L + 1), . . . , fi (t) as input, where fi (·) may include the residual battery of the current serving
UAV, the channel quality from the MU to the serving UAV, the relative distance, the current service load, and the recent (H) mission-state evolution. Let yi (t) ∈ {0, 1} denote the (H) disruption label, where yi (t) = 1 means that, without proactive successor preparation, the current serving path of mission ui will become unsustainable within the future window [t + 1, t + H]. The LSTM predictor outputs (H) ξˆi (t) = ΦLSTM Xi (t) , (9) which represents the predicted probability that mission ui will be disrupted within the next H timeslots. Based on the predicted risks, Fresco identifies the high-risk mission set according to (3). Fresco then performs successor matching for the high-risk current serving pairs. Since each current serving pair can reserve at most one candidate successor, while each candidate UAV may simultaneously accept multiple serving pairs subject to local communication, computation, synchronizationlink, and battery constraints, the considered process is essentially a resource-constrained many-to-one matching process. For notational simplicity, let µt (i, j) ∈ N cand (t) ∪ {∅} denote the matching result of current serving pair (ui , nsj (t)) at timeslot t. If µt (i, j) = nck (t), then candidate UAV nck (t) is reserved as the successor of mission ui currently served by nsj (t). Correspondingly, let µt (k) denote the set of current serving pairs tentatively accepted by candidate UAV nck (t). For candidate UAV nck (t), given a local proposal pool Sk (t), the tentative accepted set is determined by the following local exact choice rule: X Chk Sk (t) = arg max UkN (ui , nsj (t)) (10) Ak ⊆Sk (t)
(ui ,nsj (t))∈Ak
subject to local communication, computation, synchronizationlink, and battery constraints. Since the empty set is always feasible, a candidate UAV can always reject all proposals if no feasible accepted subset is beneficial. At each timeslot, Fresco first predicts disruption risks and identifies the high-risk mission set. Then, for each high-risk current serving pair (ui , nsj (t)), it computes the minimum reserved-resource template by solving the following auxiliary weighted-cost minimization problem: syn acc min ν1 Bi,j,k (t) + ν2 B̃i,k (t) + ν3 F̃i,k (t), (11) syn acc (t),F̃ Bi,j,k (t),B̃i,k i,k (t)
syn max link while holding dtk , 0 ≤ Bi,j,k (t) ≤ Bj,k (t), i,j,k (t) ≤ Ti acc ava ava 0 ≤ B̃i,k (t) ≤ Bk (t), and 0 ≤ F̃i,k (t) ≤ Fk (t). Based on the obtained minimum templates, Fresco constructs the feasible candidate set Ci,j (t) and the preference set Pi,j (t), and then iteratively performs proposal–rejection updates until no further feasible proposal exists. After the matching terminates, the matched triplets execute minimum resource reservation and context synchronization. If the current serving UAV enters an alarm state and the reserved successor has reached the takeover-ready condition, successor takeover is triggered. Accordingly, the pseudo-code of Fresco is summarized in Alg. 1. The overall procedure consists of four stages: highrisk mission identification, minimum-template-based candidate construction, constraint-aware proposal-rejection matching, and takeover preparation/update. Specifically, Fresco first predicts disruption risks and identifies the high-risk ongoing missions. It then computes the minimum reserved-resource template for each high-risk current serving pair, based on which the feasible candidate set and preference set are con-
Algorithm 1: Proposed Fresco Collect recent observation sequences and predict the disruption risks of all ongoing missions using (9); R 2 Identify the high-risk mission set U (t) according to (3); s R 3 For all current serving pairs (ui , nj (t)) satisfying ui ∈ U (t) and xi,j (t) = 1, compute the minimum reserved-resource template, feasible candidate set Ci,j (t), and preference set Pi,j (t); 4 Initialize µt (i, j) ← ∅ for all current serving pairs, and initialize µt (k) ← ∅ for all candidate UAVs; 5 while there exists an unmatched current serving pair with a non-empty preference set do 6 for each unmatched current serving pair (ui , nsj (t)) do 7 Propose to the most preferred remaining candidate UAV in Pi,j (t); 1
8 9 10 11 12 13
14 15
for each candidate UAV nck (t) ∈ N cand (t) do Merge newly received proposals with the current tentative accepted set to form Sk (t); Update the tentative accepted set by µt (k) ← Chk (Sk (t)); Reject all proposals not retained in µt (k); for each rejected current serving pair do Remove the corresponding rejecting UAV from its preference set; for each matched triplet (ui , nsj (t), nck (t)) do Set zi,j,k (t) ← 1, allocate the minimum reserved resources, and update si,j,k (t + 1) according to (2);
for each matched current serving pair do (H) if ξ̂i (t) ≥ ξ alarm or the current serving UAV enters a severe degradation state then 18 if si,j,k (t) ≥ smin then 19 Trigger successor takeover;
16
17
20 21
else Maintain the current serving path if sustainable; otherwise invoke reactive fallback;
structed. After that, serving pairs and candidate UAVs iteratively interact through a proposal-rejection process governed by the local exact choice rule. Finally, the matched triplets perform resource reservation and context synchronization, and successor takeover is triggered once the alarm condition and takeover-ready condition are both satisfied. E. Property Analysis We next briefly discuss the convergence and stability-related properties of Fresco. For notational simplicity, denote a current serving pair by πi,j (t) ≜ (ui , nsj (t)). The convergence of Alg. 1 is immediate. Since each current serving pair can propose only to candidate UAVs in its finite preference set Pi,j (t), and each rejected candidate UAV is permanently removed from that set, every pair can propose to each candidate UAV at most once. Hence, the proposal–rejection process will terminate after a finite number of rounds. Theorem 1 (Feasibility and individual rationality). The final matching result in Alg. 1 is feasible and individually rational. Proof. On the serving-pair side, each current serving pair proposes only to candidate UAVs in Pi,j (t); by construction, these candidates are feasible and yield non-negative mission-side utility. On the candidate-UAV side, the tentative accepted set is always determined by the local exact choice rule in (10), which selects a feasible subset under local communication, computation, synchronization-link, and battery constraints. Since the
empty set is feasible, no candidate UAV is forced to accept a harmful set of serving pairs. Therefore, the final matching is feasible and individually rational for both sides. Definition 1 (Blocking pair). Under a given matching, a current serving pair πi,j (t) and a candidate UAV nck (t) form a blocking pair if πi,j (t) strictly prefers nck (t) to its final match, and nck (t) can obtain no smaller utility through a feasible re-selection of its accepted set that includes πi,j (t) under local communication, computation, synchronization-link, and battery constraints. Theorem 2 (Constrained Stability). The final matching produced by Alg. 1 admits no blocking pair under Definition 12 . Proof. Assume that a blocking pair πi,j (t), nck (t) exists after Alg. 1 terminates. Since πi,j (t) strictly prefers nck (t) to its final match and proposes according to its preference order, it must have proposed to nck (t) during the algorithm. If this proposal is finally retained, then πi,j (t) is matched with nck (t), contradicting the assumption that they form a blocking pair. Otherwise, πi,j (t) must have been rejected by nck (t). According to the local exact choice rule in (10), the retained accepted set yields no smaller UAV-side utility than any feasible re-selection containing πi,j (t) under the local resource constraints. Hence, nck (t) cannot improve or maintain its utility by forming a feasible blocking pair with πi,j (t), which contradicts Definition 1. Therefore, no blocking pair exists, and the final matching is stable. IV. E VALUATION We conduct comprehensive evaluations to validate the effectiveness of Fresco, implemented by Python 3.10 with a 12th Gen Intel Core i9-12900H processor. A. Simulation Settings We consider a dynamic UEN scenario with 100 m × 100 m service area, evolving over T = 80 timeslots with interval τ = 1.0. MUs move with maximum speed 2.0 m/s, while UAVs fly with maximum speed 10.0 m/s. To examine different network sizes, we consider four jointly increasing system-scale settings, namely, S1 = (48, 12), S2 = (72, 18), S3 = (96, 24), and S4 = (120, 30), where each ordered pair denotes the numbers of MUs and UAVs, respectively. Other key parameters are set according to representative existing literature as follows [3], [5]–[7]: the initial remaining input data size, computation workload, and service-context size are generated as Direm (0) ∈ [5, 15] Mb, Cirem (0) ∈ [8, 20] Gcycles, and ζi (0) ∈ [2, 6] Mb, respectively; the maximum tolerable post-takeover completion delay is set as Timax ∈ [10, 18] s; the initially available communication and computation resources of each candidate UAV are generated as Bkava (0) ∈ [10, 20] MHz and Fkava (0) ∈ [10, 25] Gcycles/s, respectively; and the residual available onboard energy budget is set as Ekc (0) ∈ [60, 120] kJ. Unless otherwise stated, the default parameters are set as ξ th = 0.08, ξ alarm = 0.16, and smin = 0.30, while the prediction horizon and historical observation length are set to H = 4 and L = 6, respectively. For the LSTM predictor, the input dimension and hidden dimension are set to 5 and 32, respectively, and the training parameters are set to 20 epochs, learning rate 0.001, 2 Here, stability refers to resource-constrained pairwise stability with respect
to the local feasible re-selection rule in (10), rather than classical unconstrained many-to-one stability.
Fig. 2. Performance comparisons in terms of: (a) SCR, (b) AID, and (c) TCR under different system-scale settings.
and batch size 128. All reported results are averaged over 50 independent runs, and the error bars indicate the standard deviation. B. Benchmark Methods and Evaluation Metrics We compare our Fresco with the following representative baselines: (i) Reactive, a purely reactive recovery benchmark motivated by existing handover and service migration studies in UAV systems [4], [10], which does not reserve any successor UAV in advance and only triggers service recovery after the current serving UAV becomes unreliable; (ii) BestChannel, a greedy reservation benchmark motivated by robust channelaware UAV communication designs [3], which always selects the feasible candidate UAV with the best instantaneous channel quality; (iii) BestResource, a greedy benchmark motivated by resource-aware and association-aware multi-UAV edge services designs [5], [6], which always selects the feasible candidate UAV with the largest currently available communication/computation resource budget; (iv) Random, a random reservation benchmark that chooses one feasible candidate UAV uniformly at random; and (v) FrescoNoPred, an ablationstyle reservation scheme that replaces the LSTM-based shortterm disruption prediction in Fresco with the current observed risk indicator, thereby highlighting the benefit of predictive disruption awareness. We evaluate these methods using the following metrics: (i) Session continuity rate (SCR), which measures the fraction of missions that remain interruption-free throughout the simulation; (ii) Average interruption duration (AID), which measures the average interruption duration of missions that suffer service disruption; (iii) Timely completion rate (TCR), which measures the fraction of missions completed within their deadline constraints; (iv) Takeover-ready ratio (TRR), which measures the ratio of alarm events for which the reserved successor is already ready for takeover; (v) Alarmconditioned successful takeover rate (ACSTR), which measures the ratio of alarm events that are successfully handled by reserved-successor takeover; (vi) Fallback ratio (FR), which measures the ratio of alarm events that still resort to reactive fallback; (vii) Average decision time (ADT), which measures the average runtime per decision round; (viii) Preparation energy overhead (PEO), which measures the average energy consumed by successor reservation and context synchronization; and (ix) Average social welfare per task (ASW), which measures the average system-level welfare contributed by each task. These baselines and metrics allow us to evaluate whether Fresco improves continuity, takeover reliability, and utility with acceptable complexity and preparation overhead. C. Performance Evaluations We first evaluate the proposed framework from the perspective of service continuity and completion quality by
Fig. 3. Performance comparisons in terms of: (a) TRR, (b) ACSTR, and (c) FR under different system-scale settings.
Fig. 4. Performance comparisons in terms of: (a) ADT, (b) PEO, and (c) ASW under different system-scale settings.
considering SCR, AID, and TCR in Fig. 2. As shown in Fig. 2(a), Fresco consistently achieves the highest SCR across all scale settings, while FrescoNoPred generally ranks second. This indicates that LSTM-assisted disruption prediction helps identify high-risk missions earlier and reserve more suitable successors before severe service degradation occurs. Fig. 2(b) reports AID, where a smaller value is preferred. Fresco again performs the best and yields the shortest interruption duration under all system scales, whereas Reactive produces the largest AID because recovery starts only after disruption has already occurred. Similar observations can be made from Fig. 2(c), where Fresco consistently attains the highest TCR, showing that proactive successor preparation not only improves continuity but also increases the probability of deadline-aware mission completion. Overall, Fig. 2 verifies that the proposed Fresco design effectively improves both interruption resilience and completion quality in dynamic UEN. We next examine successor takeover effectiveness through TRR, ACSTR, and FR in Fig. 3. From Fig. 3(a), Fresco consistently achieves the highest TRR, indicating that the reserved successor is more likely to be takeover-ready when alarm events arise. Fig. 3(b) further shows that Fresco attains the highest ACSTR across all system scales, demonstrating that prediction-guided reservation can more effectively convert advance preparation into successful takeover. In contrast, the heuristic baselines cannot jointly account for future disruption risk, synchronization progress, and takeover feasibility, and therefore exhibit weaker takeover performance. Fig. 3(c) reports FR, where a smaller value is preferred. Fresco consistently achieves the lowest FR, while Reactive performs the worst because it lacks proactive preparation and relies heavily on post-degradation fallback. Therefore, Fig. 3 confirms that the proposed framework improves takeover support in a direct and effective manner. Finally, we evaluate decision efficiency, preparation overhead, and overall system-level gain through ADT, PEO, and ASW in Fig. 4. As shown in Fig. 4(a), ADT increases with the system scale for all methods, which is expected because larger settings involve more serving pairs, richer candidate interactions, and heavier decision complexity. Reactive incurs the smallest ADT, while FrescoNoPred and Fresco introduce higher decision overhead due to candidate construction, utility evaluation, and reservation-oriented matching. Nevertheless, this additional overhead remains acceptable given the substan-
tial continuity and takeover gains observed above. Fig. 4(b) shows that reservation-based methods incur extra preparation energy compared with purely reactive recovery, which is consistent with the design philosophy of Fresco, namely, spending moderate preparation cost to hedge against future service disruption. Importantly, the overhead of FrescoNoPred and Fresco remains well controlled across different scales. Finally, Fig. 4(c) shows that Fresco consistently achieves the highest ASW and outperforms all benchmark methods. This indicates that the proposed predictive reservation mechanism can better translate proactive successor preparation into tangible system-level benefit. In summary, the proposed Fresco achieves the best overall tradeoff among service continuity, takeover effectiveness, and system-level utility. V. C ONCLUSION This paper investigated proactive continuous service support in dynamic UEN and proposed Fresco, a forecastingdriven successor reservation framework for continuity-oriented edge services. By combining LSTM-based short-term disruption prediction, risk-aware successor matching, lightweight resource reservation, and progressive service-context synchronization, Fresco enables takeover preparation before severe service degradation occurs. Experiments results showed that Fresco significantly improves service continuity and takeover effectiveness over reactive and non-predictive baselines with only modest reservation overhead. R EFERENCES [1] X. Xia, S. M. M. Fattah, and M. A. Babar, “A survey on uav-enabled edge computing: Resource management perspective,” ACM Comput. Surveys, vol. 56, no. 3, pp. 1–36, 2023. [2] K. Wang, G. Wu, X. Fu, X. Wang, K. Liu, X. Lu, C. Ge, W. Zhai, and Z.-J. Zha, “Skyfind: A large-scale benchmark unveiling referring expression comprehension for uav,” IEEE Trans. Pattern Anal. Mach. Intell., early access, pp. 1–1, 2026. [3] J. Wang, X. Fang, X. Wang, L. Yan, J. Wu, and B. Yin, “Enhancing a2g robustness in energy-constrained multi-uav networks: Madrl for trajectory control and resource allocation,” IEEE Trans. Wireless Commun., vol. 25, pp. 11286–11302, 2026. [4] Z. Ye, P. N. Ji, and T. Wang, “Seamless service handover in uav-based mobile edge computing,” in IEEE GLOBECOM 2023, pp. 1113–1118, 2023. [5] Y. Gao, P. Wu, X. Yuan, Y. Hu, X. Cao, and A. Schmeink, “Joint uav 3d deployment and ground device association optimizing for multi-uavaided mec heterogeneous network,” IEEE Trans. Mobile Comput., early access, pp. 1–1, 2026. [6] L. Liu, H. Miao, C. Qu, Z. Wang, H. Zhang, and Z. Li, “Multi-uav path planning for mobile edge computing with high-density mobile devices,” IEEE Trans. Mobile Comput., early access, pp. 1–1, 2026. [7] Y. Wang, J. Farooq, and J. Chen, “Dynamic multi-modal uav control for optimized coverage and backhaul connectivity in spatially unstructured and dispersed user environments,” IEEE Trans. Mobile Comput., vol. 25, no. 2, 2026. [8] O. Bekkouche, S. Kianpisheh, and T. Taleb, “Toward proactive service relocation for uavs in mec,” in IEEE GLOBECOM 2021, pp. 1–7, 2021. [9] Z. Chen, S. Huang, G. Min, Z. Ning, J. Li, and Y. Zhang, “Mobilityaware seamless service migration and resource allocation in multi-edge iov systems,” IEEE Trans. Mobile Comput., vol. 24, no. 7, pp. 6315– 6332, 2025. [10] Y. Shi, C. Yi, R. Wang, Q. Wu, B. Chen, and J. Cai, “Service migration or task rerouting: A two-timescale online resource optimization for mec,” IEEE Trans. Wireless Commun., vol. 23, no. 2, pp. 1503–1519, 2024. [11] C. Bernad, A. Dedinec, K. Gilly, S. Filiposka, and A. Mishev, “Towards smart 6g: Mobility prediction for dynamic edge services migration,” Expert Syst. Appl., vol. 297, pp. 129348–129360, 2026.