cs.DCFeb 24, 2026

Autonomous Edge-Deployed AI Agents for Electric Vehicle Charging Infrastructure Management

Authors: Mohammed Cherifi

Organizations: Hyperion Consulting, Paris, France

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Public EV charging infrastructure suffers from significant failure rates -- with field studies reporting up to 27.5% of DC fast chargers non-functional -- and multi-day mean time to resolution, imposing billions in annual economic burden. Cloud-centric architectures cannot achieve the latency, reliability, and bandwidth characteristics required for autonomous operation. We present Auralink SDC (Software-Defined Charging), an architecture deploying domain-specialized AI agents at the network edge for autonomous charging infrastructure management. Key contributions include: (1) Confidence-Calibrated Autonomous Resolution (CCAR), enabling autonomous remediation with formal false-positive bounds; (2) Adaptive Retrieval-Augmented Reasoning (ARA), combining dense and sparse retrieval with dynamic context allocation; (3) Auralink Edge Runtime, achieving sub-50ms TTFT on commodity hardware under PREEMPT_RT constraints; and (4) Hierarchical Multi-Agent Orchestration (HMAO). Implementation uses AuralinkLM models fine-tuned via QLoRA on a domain corpus spanning OCPP 1.6/2.0.1, ISO 15118, and operational incident histories. Evaluation on 18,000 labeled incidents in a controlled environment establishes 78% autonomous incident resolution, 87.6% diagnostic accuracy, and 28-48ms TTFT latency (P50). This work presents architecture and implementation patterns for edge-deployed industrial AI systems with safety-critical constraints.

Figures & tables

TaskGeneralAdaptedΔ\Delta
General knowledge89.2%87.8%-1.4
OCPP protocol44.7%92.1%+47.4
Fault diagnosis38.1%87.6%+49.5
Remediation41.8%84.3%+42.5
Safety assessment52.3%93.2%+40.9
ISO 1511835.2%89.7%+54.5
Table 1: Baseline LLM Performance on EV Charging Tasks. “General” is Mistral Large 3 base (zero-shot, no domain context); “Adapted” is AuralinkLM 14B after QLoRA fine-tuning. Each category evaluated on 200–500 held-out examples using exact-match accuracy. Expert judgment applied to open-ended diagnostic responses.
ApproachEdgeLLMAutoOff.Open
Rule CSMS✗✗✗Part.Prop.
Cloud AI✗✓Part.✗Prop.
Edge GW✓✗✗✓Mix
Acad. RAGPart.✓Part.✗✓
Ours✓✓✓✓✓
Table 2: Comparison with Existing Approaches
Figure 1: Three-tier Auralink SDC architecture
Figure 2: Detailed Auralink SDC system architecture showing component interactions across all three tiers. The edge tier contains the core AI capabilities (DiagnosticEngine, ARA, CCAR) enabling autonomous operation. Latency figures represent P50 inference times on target hardware.
DomainNFunctions
Station Mgmt28Registration, provisioning
Session23Auth, metering, billing
Grid Integration20DSO protocols, flex markets
AI/ML Pipeline35Training, eval, distribution
Analytics25BI, dashboards, reporting
Operations39Monitoring, alerting, SRE
Table 3: Cloud Tier Service Architecture
ComponentSpecification
Configuration A: High Performance
ProcessorAMD Ryzen AI Max+ 395 (Strix Halo)
NPUAMD XDNA 2, 50 TOPS
Memoryup to 128GB LPDDR5X-8000
Storage1TB NVMe Gen4
Configuration B: NVIDIA DGX Spark [ 26 ]
Table 4: Auralink Edge Runtime Hardware Configurations
ModelAdapt.Quant.SpeedStruct.
Mistral94.297.192.496.8
Llama 3.391.894.389.791.2
Qwen 2.592.493.888.293.4
Phi-488.791.295.189.3
Gemma 287.392.487.888.7
Table 5: Model Family Technical Comparison
ModelParamsTierP50P99
675B675B MoECloud285ms †\dagger890ms
14B14BEdge28ms78ms
0.5B0.5BDevice11ms34ms
Table 6: Model Deployment Configuration. P50 and P99 denote the 50th and 99th percentile latency (TTFT), respectively.
CategoryCountSource
OCPP 1.68,432Spec + synthetic Q&A
OCPP 2.0.19,847Spec + synthetic Q&A
Fault Diagnosis41,234Historical incidents
Remediation34,567Service manuals
Power Electronics24,891Engineering docs
Session Management18,432Operational logs
Table 7: Training Corpus Composition
ParameterValue
Rank ( rr )64
Alpha ( α\alpha )32
Dropout0.05
Target Modulesq,k,v,o,gate,up,down
Learning Rate2×10−42{\times}10^{-4} (cosine)
Warmup Steps100
Table 8: QLoRA Hyperparameters
Figure 3: QLoRA fine-tuning pipeline: curated domain-specific training corpus processed through three-stage curriculum learning, producing domain-adapted model subsequently quantized to Q4_K_M (8.2 GB) for edge deployment.
PrecisionSizeAcc.P50
FP1628.0 GB89.4%142ms
INT8 (Q8_0)14.0 GB89.1%68ms
Q5_K_M9.8 GB88.4%42ms
Q4_K_M8.2 GB87.6%28ms
Q4_K_S7.6 GB86.9%26ms
IQ4_XS7.2 GB85.8%24ms
Table 9: Quantization Impact: AuralinkLM 14B
PlatformCPUNPUSpeedup
AMD XDNA 248ms28ms1.71 ×\times
Intel AI Boost62ms42ms1.48 ×\times
Qualcomm Hexagon85ms52ms1.63 ×\times
Table 10: NPU Offloading Performance
MetricP99Max
Interrupt latency38 \SIUnitSymbolMicros38\text{\,}\mathrm{\SIUnitSymbolMicro s}52 \SIUnitSymbolMicros52\text{\,}\mathrm{\SIUnitSymbolMicro s}
Scheduling latency72 \SIUnitSymbolMicros72\text{\,}\mathrm{\SIUnitSymbolMicro s}98 \SIUnitSymbolMicros98\text{\,}\mathrm{\SIUnitSymbolMicro s}
Timer jitter8 \SIUnitSymbolMicros8\text{\,}\mathrm{\SIUnitSymbolMicro s}15 \SIUnitSymbolMicros15\text{\,}\mathrm{\SIUnitSymbolMicro s}
Context switch6 \SIUnitSymbolMicros6\text{\,}\mathrm{\SIUnitSymbolMicro s}11 \SIUnitSymbolMicros11\text{\,}\mathrm{\SIUnitSymbolMicro s}
Inference start124 \SIUnitSymbolMicros124\text{\,}\mathrm{\SIUnitSymbolMicro s}187 \SIUnitSymbolMicros187\text{\,}\mathrm{\SIUnitSymbolMicro s}
Table 11: Real-Time Timing Metrics
Figure 4: CCAR decision flow (simplified representation). Actions above τauto=0.90\tau_{auto}=0.90 execute autonomously without notification; actions between 0.850.85 and 0.900.90 execute with operator notification; between τassist=0.70\tau_{assist}=0.70 and 0.850.85 require human confirmation; below τassist\tau_{assist} escalate to manual resolution. See Table 12 for the full five-tier decision matrix.
ConfidenceAction
C≥0.90C\geq 0.90Execute autonomously, log only
0.85≤C<0.900.85\leq C<0.90Execute with operator notification
0.70≤C<0.850.70\leq C<0.85Recommend, await approval (timeout: 4h)
0.50≤C<0.700.50\leq C<0.70Escalate with full context
C<0.50C<0.50Human expert required
Table 12: CCAR Decision Thresholds
Document TypeDocsChunks
OCPP Specifications1818,432
OEM Service Manuals34758,234
Technical Bulletins1,23442,847
ISO/IEC Standards5214,123
Historical Resolutions2,84772,345
Grid Codes288,432
Table 13: ARA Knowledge Base Composition
MetricValue
Recall@595.2%
Precision@591.4%
MRR (Mean Reciprocal Rank)0.924
NDCG@100.912
Retrieval Latency (P50)18ms
Retrieval Latency (P99)42ms
Table 14: ARA Retrieval Metrics
Figure 5: HMAO agent hierarchy with intent routing. Dashed lines indicate inter-agent collaboration for complex multi-domain incidents. Model sizes (0.5B/14B) indicate AuralinkLM deployment tier.
AgentResponsibilityModel
AutoOpsFault diagnosis14B
TechSupportGuided repair14B
OpsMgmtFleet analytics14B
DriverAssistSession support0.5B
CodeGenAPI integration14B
Table 15: Specialized Agent Configuration
ComponentSpecification
Cloud Infrastructure
GPU ClusterNVIDIA H100/H200 (distributed)
ModelAuralinkLM 675B (675B MoE, 4+1 active experts)
Context128K tokens (of 256K max)
Edge Deployment (Primary)
CPUAMD Ryzen AI Max+ 395 (Strix Halo)
Table 16: Evaluation Hardware Specifications
Incident TypeFraction
Communication30.0%
Authorization18.0%
Power Elec.15.0%
Firmware/SW13.0%
Mechanical10.0%
Payment Processing8.0%
Table 17: Test Incident Distribution (Controlled Testing Corpus)
SystemAutoAcc.P50MTTROff.
Rule-Based0%44.7%5ms96hYes
Cloud (A)28.4%72.3%420ms28hNo
Cloud (B)31.2%74.8%380ms24hNo
Non-FT22.1%52.3%32ms42hYes
Ours78% †\dagger87.6% †\dagger28ms4–8h †\daggerYes
Table 18: Comprehensive Baseline Comparison
ParameterValue (simulation)
Simulated stations1,000-charger reference
Operator profiles23 (modeled)
Country profiles14 (EU + UK + CH + NO)
Testing periodControlled testing
Simulated sessionsRepresentative sample
Diagnostic scenariosRepresentative corpus
Table 19: Simulation and Testing Parameters
SystemAuto%MTTRFP%FN%
Traditional0%96h42%–
Cloud-AI a31.2%24h11%8%
Human Expert–68h6%4%
Auralink SDC78% †\dagger4–8h †\dagger3.8%5.2%
Table 20: Autonomous Resolution: System Comparison
CategoryAcc.Prec.Rec.F1F_{1}
Comms91.8%90.2%93.4%0.918
Auth90.4%92.1%88.7%0.904
Power Elec.85.2%82.8%87.9%0.853
Firmware89.3%91.2%87.4%0.893
Mechanical79.8%77.2%82.6%0.798
Payment88.7%89.4%87.9%0.886
Table 21: Diagnostic Accuracy by Fault Category
Failure Mode% of Failures
Low confidence (escalated)50.0%
Hardware replacement req.28.0%
Multi-fault complexity11.0%
Novel fault pattern6.0%
False negative (missed)5.0%
Total failures100%
Table 22: Failure Mode Distribution (Controlled Testing)
ConfigurationAcc.Auto%Δ\Delta Acc
Full system87.6% †\dagger78% †\dagger—
−- Domain fine-tuning52.3%22.1%−- 35.3
−- ARA retrieval79.2%61.4%−- 8.4
−- CCAR calibration87.1%42.8%−- 0.5
−- Curriculum learning84.3%72.1%−- 3.3
−- Multi-agent routing85.8%74.2%−- 1.8
Table 23: Ablation Study: Component Contributions
ConfigurationP50P95P99
Cloud (675B)285ms †\dagger520ms890ms
Edge GPU (14B)28ms52ms78ms
Agent FW (0.5B)11ms24ms34ms
Full pipeline67ms124ms187ms
Table 24: Inference Latency (TTFT for model components; end-to-end for full pipeline)
HorizonPrec.Rec.F1F_{1}FPR
24h94.0% †\dagger84.2%0.8881.2%
48h94.7%79.8%0.8662.1%
72h92.4%75.1%0.8293.4%
7d85.2%68.7%0.7615.8%
Table 25: Predictive Maintenance by Horizon
CategoryTrad.Auralink
SW licensing$129K$0
Service$165K$126K
Field service$900K$375K
Downtime$219K$72K
Infrastructure—$90K
Total$1,413K$663K
Table 26: Three-Year TCO Analysis
Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Platform0.5B14BTDP
AMD Strix Halo1423845–120W
NVIDIA DGX Spark38595140W
NVIDIA RTX 5090512 ‡\ddagger128 ‡\ddagger575W
NVIDIA Jetson Thor521240–130W
Apple M4 Max12834∼\sim 50W
Qualcomm X Elite872123W
Appendix
Table 27: AuralinkLM Inference by Platform (tokens/sec)
More in this topic

Electric Vehicle Charging

Electric vehicle charging research studies how to forecast, plan, control, and secure charging while coordinating vehicles, stations, and power grids.

See all 38 papers

Explore similar work

CardsList

  1. Autonomous Incident Resolution at Hyperscale: An Agentic AI Architecture for Network Operations

    Jun 8, 2026Arun MalikMulti-Agent Orchestration

    Microsoft Azure Networking

  2. Climate-resilient electric vehicle charging infrastructure for sustainable cities: An interpretable causal-ensemble framework for preventive maintenance and low-carbon mobility

    Jul 23, 2026Cande Lian, Wentao Zeng, Jiabin Wu +2Stacked Ensemble LearningEnsemble Learning

    School of Management, Foshan University, Foshan, China · School of Management, Foshan University, Foshan 528000, China · School of Mathematics, Foshan University, Foshan 528000, China +5

  3. A Multi-Objective AutoML-based Efficient Intrusion Detection System for EV Charging Networks

    Aug 3, 2026Li YangNetwork Intrusion DetectionElectric Vehicle Charging

    Ontario Tech University, Oshawa, Ontario, Canada · Western University, London, Ontario, Canada