BatStation: Multipurpose Battery Systems

• BatStation is a multi-context battery infrastructure label encompassing EV swapping, telecom hubs, radar sensing, and autonomous power nodes.
• It integrates queueing models, scheduling algorithms, grid interaction, and degradation-aware control to optimize energy management and service reliability.
• It supports advanced siting, demand estimation, and sensing technologies, enabling efficient operations from EV refueling to UAV and rover missions.

BatStation is a label used in several technically distinct research contexts centered on batteries, swapping stations, or base stations. In one major line of work, it denotes a battery swapping and charging station for electric vehicles, with formal models for queueing performance, scheduling, siting, degradation-aware control, and energy-market participation. In other work, it denotes a battery-enabled 5G/6G base station that jointly manages telecom load, renewables, and EV charging, or a lightweight in-situ radar sensing framework running directly on 5G base stations. The term is also used for autonomous power hubs that charge and swap battery modules for planetary rovers and for static battery management stations in multi-UAV missions (Tan et al., 2017, Shen et al., 18 Aug 2025, Gao et al., 8 Sep 2025, Holand et al., 2024, Bouček et al., 2024).

1. Terminological scope and principal usages

In the sources considered here, “BatStation” does not designate a single canonical system. This suggests that the term functions as a family resemblance label for infrastructures in which a battery asset, a station asset, or both become the operational center of a larger cyber-physical system.

Usage	Core configuration	Representative source
EV battery swapping station	EV arrivals, battery inventory, chargers, swap service	(Tan et al., 2017)
Battery-enabled telecom hub	5G/6G base station, BESS, renewables, EVSE, dynamic pricing	(Shen et al., 18 Aug 2025)
Base-station radar sensing framework	5G uplink receiver with radar separation and zero-shot correlation	(Gao et al., 8 Sep 2025)
Rover power hub	Central hub charging and swapping battery modules for rovers	(Holand et al., 2024)
UAV battery management station	Static station visited during mission replanning	(Bouček et al., 2024)

A narrower and historically earlier usage is the battery swapping station for EVs, often discussed under the names BSS or BSCS. A broader contemporary usage extends the idea to base-station-centric energy and sensing systems, where the “station” is a telecom BS rather than an EV swap facility.

2. Queueing-theoretic EV BatStations

A battery swapping and charging station (BSCS) is an energy-refueling facility where EVs arriving with depleted batteries swap for fully charged batteries, and the swapped depleted batteries are subsequently charged to fully charged batteries. The core model in this literature is a mixed queueing network (MQN) composed of an open EV-queue and a closed battery-queue. EV arrivals are Poisson with rate $\lambda$; swapping time at each swapping island is exponential with rate $\nu$; charging time at each charger is exponential with rate $\mu$; the EV-queue capacity is $N$; the number of swapping islands is $S$; the number of chargers is $C$; and the number of batteries is $B$ (Tan et al., 2017).

The state is a triple $(n,b,j)$, where $n$ is the number of EVs in the open EV-queue, $b$ is the number of fully charged batteries in the FB-queue, and $\nu$0 is the number of depleted batteries in the DB-queue, with $\nu$1. The exact two-dimensional CTMC is constructed over $\nu$2, and the EV blocking probability is defined through PASTA as

$\nu$3

Blocking occurs exactly when an EV arrival sees the open EV-queue full, i.e., $\nu$4, regardless of FB availability in the station (Tan et al., 2017).

The asymptotic theory is organized by a charging-capacity threshold. When

$\nu$5

the station is in a charging-limiting regime; when

$\nu$6

it is in a swapping-limiting regime. These conditions are also the necessary and sufficient conditions for the positive recurrence of the relevant limit CTMC as $\nu$7 (Tan et al., 2017).

The resulting asymptotic blocking laws are unusually sharp. In charging-limiting mode,

$\nu$8

whereas in swapping-limiting mode,

$\nu$9

The interpretation is correspondingly binary: asymptotically, blocking depends either on charger capacity and charging rate, or on parking/swapping capacity and swap rate, but not both. Numerical studies with parameters such as $\mu$0, $\mu$1, $\mu$2, $\mu$3, and $\mu$4 confirm convergence to the predicted limits, and “quick saturation” with respect to $\mu$5 is reported, with $\mu$6 slightly larger than $\mu$7 typically sufficient (Tan et al., 2017).

This queueing-theoretic strand is important because it establishes a precise separation between service-capacity bottlenecks and battery-pool effects. It also makes explicit that simply adding batteries cannot drive blocking below $\mu$8 in a charging-limited station.

3. Scheduling, grid interaction, and battery valuation in EV BatStations

A second strand treats the EV BatStation as a schedulable asset. One scheduling model represents each battery through four binary states—empty, charging, fully charged, and out-of-station—with exclusivity constraints, an inventory conservation constraint, a charger-capacity limit, and a demand-coverage requirement. In the illustrative example, the station owns $\mu$9 batteries, has $N$0 chargers, uses $N$1 batteries, and assumes a minimum charging time of $N$2 hours. The example schedule maintains at least one fully charged battery available at all times, while explicitly showing queueing of empty batteries when the four chargers are occupied (Mahoor et al., 2017).

The same physical infrastructure can also be scheduled as a grid-facing storage asset. In a distribution-level model for capturing solar variability, the feeder net load is

$N$3

and the BSS is constrained so that the net load ramp seen by the utility remains within a utility-specified limit:

$N$4

The numerical study uses $N$5 batteries of $N$6 each and $N$7 AC Level-2 chargers with maximum per-battery charging power $N$8. Under pure price-based arbitrage, the operation cost is reported as $N$91{,}555.72$S$0\Delta U=1\ \mathrm{MW/h}$S$1-$S$2, a $S$3 increase in cost relative to the arbitrage case. Under $S$4 solar forecast error, the robust cost sequence reported is $S$51{,}216.14\rightarrow-$S$6967.10\rightarrow-$S$7910.49$ as the uncertainty budget increases (Hosseini et al., 2018).

A further development places degradation at the center of dispatch. In the intertemporal framework for battery valuation and management, life-cycle benefit is

$S$8

where $S$9 is the life-cycle marginal degradation cost (MDC), and the adjusted short-term MDC is $C$0. The daily problem co-optimizes energy arbitrage, non-spinning reserve, and swapping revenue, subject to SOC dynamics, power bounds, and swapping energy limits. In the case study, life-cycle revenue peaks around $C$135/\mathrm{MWh\text{-}throughput}$$. With swapping, the physical life of <a href="https://www.emergentmind.com/topics/boundary-embedding-shaping-bes" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">BES</a> ends faster for low-to-moderate MDCs, but the economic life becomes considerably longer because swaps enhance utilization and revenue (<a href="/papers/2302.14291" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Chen et al., 2023</a>).</p> <p>The most explicit operational controller in this group is BSS-MPC, a degradation-aware mixed-integer <a href="https://www.emergentmind.com/topics/tactile-reactive-multi-agent-model-predictive-controller-mpc" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">MPC</a> framework that co-optimizes arbitrage, swapping logistics, and battery health over a$$C$2-hour horizon. In the reported setup, a community contains $C$3 batteries, the station inventory is $C$4 packs, the SOC threshold is $C$5, and each MPC iteration solves in about $C$6–$C$7 seconds with warm starts. Over $C$8 days, Normalized Loss is $C$9 for BSS-MPC in high-profit mode and $B$0 in low-cf mode, versus $B$1 for the rule-based baseline; Average cf is $B$2 and $B$3, respectively, versus $B$4 for the rule-based baseline; and SOC satisfaction is $B$5 for both BSS-MPC modes with $B$6 (Li et al., 9 Oct 2025).

Taken together, these results show a progression from inventory-feasibility scheduling to coupled service-energy-degradation control. A plausible implication is that the modern EV BatStation is increasingly treated not merely as a refueling node, but as a multi-timescale storage system with service-level constraints.

4. Siting, demand estimation, and optimization algorithms

Location planning introduces a network-wide view. In the city-network BSS location model, demand loss has two forms: Type I loss, when an EV arrives at a BSS but the queue is full, and Type II loss, when the EV cannot reach any open BSS within its battery and detour constraints. The city is represented as a connected graph $B$7, path demands arrive as Poisson processes with rates $B$8, and each station is modeled as an $B$9 queue. The objective is to choose exactly $(n,b,j)$0 BSS locations to minimize the percentage of lost demand over time, with user choice among reachable stations governed by a multinomial-logit utility depending on detour distance and empirical mean waiting time (Liu et al., 2024).

To solve large instances, the paper combines large neighborhood search with Bayesian optimization for the single-station repair subproblem. The small synthetic example shows that a MILP solution with inaccurate average waiting times selects $(n,b,j)$1 and yields $(n,b,j)$2 loss under accurate simulation, whereas simulation enumeration finds $(n,b,j)$3 with $(n,b,j)$4 loss. In the Shanghai GPS taxi study, using $(n,b,j)$5 records, $(n,b,j)$6 passenger paths, and $(n,b,j)$7 processed path demands, the initial NIO BSS layout yields $(n,b,j)$8; simulated annealing reaches $(n,b,j)$9 after $n$0 hours; and LNS-BO reaches $n$1 after $n$2 hours (Liu et al., 2024).

At the operational-demand layer, a separate study constructs nine scenario-specific battery swap demand datasets from charging pile data and then optimizes daily scheduling with an LRU-enhanced genetic algorithm and a dual-factor decision system. The constructed datasets exhibit stable trend characteristics, adhere to $n$3-hour and $n$4-hour periodicity patterns, and have outlier ratios consistently below $n$5. Relative to a baseline GA, the improved algorithm achieves better fitness individuals in $n$6 of test regions under the same iterations. Relative to an immediate swap-and-charge strategy, the peak cost reduction is $n$7, peak user satisfaction reaches $n$8, and the average iteration time remains below $n$9 seconds (Li et al., 10 Apr 2025).

These methods move BatStation analysis beyond station-internal queueing. They explicitly couple siting, demand stochasticity, congestion, time-varying electricity prices, and algorithmic search. The methodological shift is from closed-form asymptotics toward simulation optimization, Bayesian search, and metaheuristics calibrated on high-frequency operational data.

5. Battery-enabled telecom hubs for EV charging and energy management

In a different literature, BatStation denotes a base-station-centric integrated Energy-Communication-Transportation Hub. The architecture equips a 5G/6G base station with a battery energy storage system, connects it to renewables and the grid, and adds EV charging. The base station load is traffic-dependent,

$b$0

the charger load is

$b$1

and grid import is constrained to be nonnegative,

$b$2

The battery point has discrete charge/discharge/idle actions, SoC bounds, and a blackout-resilience reserve constraint intended to guarantee communication during outages (Shen et al., 18 Aug 2025).

Control is split between pricing and storage scheduling. ECT-Price is a causal inference-based stratification system that distinguishes Always Charge, Incentive Charge, and No Charge sessions, and discounts charging primarily for the Incentive Charge stratum. ECT-DRL is an Actor-Critic PPO scheduler whose state includes recent windows of real-time price, weather, traffic, charging price, and current SoC, and whose reward is

$b$3

Training uses Adam with learning rate $b$4, weight decay $b$5, batch size $b$6, $b$7-day episodes, $b$8 training episodes, and $b$9 test episodes (Shen et al., 18 Aug 2025).

The empirical study uses weather from NSRDB, RTP from ENGIE Resources, telecom traffic from a public dataset, and EV charging histories from $\nu$00 stations over $\nu$01 years, or about $\nu$02 records. For ECT-Price, the reported reward at $\nu$03 discount is $\nu$04 for the proposed method, versus $\nu$05 for DR, $\nu$06 for IPS, and $\nu$07 for OR; at $\nu$08 discount, the corresponding numbers are $\nu$09, $\nu$10, $\nu$11, and $\nu$12. Time-of-day analysis indicates that Incentive Charge is most prevalent in the $\nu$13–$\nu$14 window. For ECT-DRL, average daily rewards over $\nu$15 hubs are consistently higher than the baselines, with examples including Hub1 at $\nu$16 versus $\nu$17 for DR and $\nu$18 for OR, Hub2 at $\nu$19 versus $\nu$20 for OR and $\nu$21 for DR, and Hub12 at $\nu$22 versus $\nu$23 for DR and $\nu$24 for OR (Shen et al., 18 Aug 2025).

The operational premise is that telecom sites already have a dense footprint, existing DC power infrastructure, and on-site batteries on the DC bus with automatic transfer switching. This makes the BatStation, in this sense, a base-station-centered energy hub rather than a pure swap facility.

6. Radar sensing on 5G base stations

A further specialization uses BatStation to denote a lightweight, in-situ radar sensing framework integrated into 5G base stations for radar–5G coexistence. The target problem is the superposition of incumbent radar signals and uplink 5G transmissions in shared spectrum such as CBRS. BatStation processes each uplink resource grid through three stages: radar signal separation, resource grid reshaping, and zero-shot template correlation (Gao et al., 8 Sep 2025).

The received signal model is

$\nu$25

or, on the discrete resource grid,

$\nu$26

Separation reconstructs and cancels 5G uplink using DMRS-based CSI with Hampel filtering, then reshapes the residual by re-FFT along time and max-pooling along frequency. Detection and localization are performed by zero-shot template correlation on reshaped residuals, with threshold $\nu$27 chosen as the $\nu$28th percentile under no-radar residuals to target about $\nu$29 false alarms (Gao et al., 8 Sep 2025).

The framework is explicitly lightweight: it has $\nu$30 trainable parameters and about $\nu$31 million MACs per slot. It is implemented on an SDR testbed with a USRP N321 BS at $\nu$32 sampling, center frequency $\nu$33, $\nu$34 channel, TDD periodicity of $\nu$35 slots over $\nu$36, up to five uplink slots, numerology 1 with $\nu$37, FFT size $\nu$38, $\nu$39 active subcarriers, and $\nu$40 OFDM symbols per $\nu$41 uplink slot. The user devices are OnePlus 8T phones, and the radar generator is a USRP X310 (Gao et al., 8 Sep 2025).

Performance is reported separately for PUCCH and PUSCH. Mean detection probability is $\nu$42 on PUCCH and $\nu$43 on PUSCH; classification accuracy reaches $\nu$44 on PUCCH and $\nu$45 on PUSCH at $\nu$46 SNR; median localization errors are $\nu$47–$\nu$48 in frequency and $\nu$49–$\nu$50 in time; and runtime latency is $\nu$51 on an NVIDIA A100 GPU and $\nu$52 on an Intel Xeon 6384 CPU, satisfying per-slot real-time requirements. Under heavy-traffic PUSCH, radar separation reduces residual 5G energy by $\nu$53–$\nu$54 (Gao et al., 8 Sep 2025).

In this usage, BatStation is neither an EV station nor an energy hub. It is a BS-resident sensing stack whose significance lies in spectrum sharing, hardware portability, and immediate integration into uplink scheduling.

7. Robotic and aerial BatStations

Outside terrestrial EV and telecom infrastructures, BatStation denotes centralized power hubs for autonomous agents. In the planetary-fleet architecture, the BatStation is a heterogeneous multi-agent hub that outsources power generation, charging, and swapping to miniaturized rovers. A single hub can support

$\nu$55

rovers, where $\nu$56 is hub generation, $\nu$57 is hub self-consumption, $\nu$58 is the charging bottleneck, and $\nu$59 is average rover power draw. The prototype uses a $\nu$60 hub, $\nu$61 battery modules with capacity $\nu$62, and achieves an average autonomous servicing time of $\nu$63 seconds over $\nu$64 consecutive dock-and-swaps. A standalone swapping reliability test reaches $\nu$65 electrical continuity success across $\nu$66 consecutive transfers. Optimization of passive guide rails via quadratic Bézier bumpers and added rover bumpers increases the success-region volume by $\nu$67, docking succeeds at $\nu$68 frontal pitch, and the integrated prototype advances the technology readiness level from TRL $\nu$69 to TRL $\nu$70 (Holand et al., 2024).

For UAVs, a BatStation is a static battery management station inserted into long-duration missions by a centralized planner. The station publishes its location, number of slots, swap duration, safety window, number of charged batteries, and dynamic schedule. The planner uses a linear endurance model with

$\nu$71

imposes $\nu$72, and treats a station visit as an instantaneous reset to full charge in the current model. In the city-park scenario, pruning candidate station actions to those within $\nu$73 minutes of the current waypoint reduces solution time from more than $\nu$74 hours to $\nu$75, while producing a $\nu$76-swap solution for $\nu$77 UAVs and $\nu$78 stations (Bouček et al., 2024).

Station placement for UAV fleets has also been studied jointly with trajectory design. The UBAT framework proves the joint problem NP-Hard and uses ant colony optimization to minimize mission finish time and the number of charging stations. In the reported simulations, UBAT produces multi-objective solutions within $\nu$79 and $\nu$80 of the true optimal solutions, respectively; restricting candidate cells to the convex hull of regions of interest reduces single-iteration runtime by about $\nu$81; and 2-OPT path correction reduces path length by $\nu$82 on average without increasing the number of stations (Won, 2019).

Across these robotic and aerial systems, the common design pattern is a shared hub that centralizes energy logistics and offloads mass, charging time, or planning complexity from mobile agents. This suggests that “BatStation” has become a useful label for centralized energy-service nodes even when the underlying mobility platform is not an EV.