Generative AI Agent Empowered Power Allocation for HAP Propulsion and Communication Systems

Abstract: High altitude platforms (HAPs) are emerging as a key enabler for 6G coverage, yet limited energy must be split between propulsion and communications. Most prior HAP studies ignore propulsion power or rely on surrogates that miss hull-propeller interference, leading to misestimated communication power budgets and degraded beamforming. More importantly, HAP power allocation is intrinsically a multi-system, multidisciplinary problem in which aerodynamics, propulsion-system efficiency, and communication-system performance (quality of service (QoS) and energy efficiency (EE)) are tightly coupled.To address these challenges, this paper designs an interactive generative AI-empowered HAP power allocation agent.By interacting with the AI agent, we develop an accurate propulsion power consumption model that takes into account the aerodynamic interference between the HAP's hull and the propeller. Assisted by the AI agent, we further formulate a HAP beamforming problem to improve user QoS and enhance the EE of the HAP communication system.This paper also proposes a QoS-enhanced energy-efficient (Q3E) beamforming algorithm to solve the formulated problem.Simulation results demonstrate the accuracy of the propulsion-power model and the effectiveness of the Q3E algorithm.

• The paper introduces a generative AI framework that fuses CFD-validated propulsion models with a joint QoS and energy-efficient communications power allocation system.
• It employs a retrieval-augmented generation pipeline and an ANN-based beamforming optimizer to achieve precise power distribution under strict energy constraints.
• The methodology reduces propulsion modeling deviation by over 84% and improves energy efficiency by up to 45% compared to traditional models.

Generative AI Agent Empowered Power Allocation for HAP Propulsion and Communication Systems

Introduction and Context

High Altitude Platforms (HAPs) play a pivotal role in the architecture of 6G and beyond, promising wide-area, low-latency connectivity with rapid deployment capabilities and resilience to disruptions. However, a central systems-level bottleneck arises from the strict energy budget that must be allocated between diverse onboard subsystems, especially propulsion and communications. Historically, HAP research has largely overlooked detailed propulsion power modeling—often neglecting the aerodynamic complexities and hull–propeller interference effects—resulting in overly optimistic communication power budgets and ineffective resource allocation. Power allocation for HAPs is thus a profoundly multidisciplinary problem, intrinsically coupling aerodynamics, propulsion-system efficiency, communications quality of service (QoS), and overall energy efficiency (EE).

This paper presents a generative AI agent-based framework to fundamentally address these modeling and optimization challenges. The approach employs a retrieval-augmented generation (RAG) pipeline with LLM reasoning, enabling robust knowledge integration from heterogeneous technical domains and interactive design synthesis. The research delivers three principal contributions: (1) a validated, AI-assisted propulsion power model accounting for hull–propeller aerodynamic interference, (2) a communications power allocation framework with joint QoS and EE objectives, and (3) the design of a constrained, unsupervised neural-network-based (ANN) beamforming optimizer.

Generative AI Agent Framework for Cross-Disciplinary Integration

The generative AI agent is architected as a research-stage offline assistant that integrates and validates cross-domain knowledge through a perception–brain–action paradigm. In this setting:

• Perception: Raw technical documents (aerodynamics, MIMO communications, energy management) are semantically parsed, chunked, and embedded with FAISS indexing, enabling efficient retrieval.
• Brain: The LLM (DeepSeek-R1) synthesizes and validates system models, enforces conformance with physical laws/constraints (via RAG), and supports advanced problem decomposition—e.g., chain-of-thought reasoning and lexicographic solver policies.
• Action: Offline execution of CFD-based aerodynamic analysis and numerical optimization for communication system design.

The AI agent supports structured, engineer-in-the-loop workflows with strong validator gates, mitigating LLM hallucination and knowledge bias. All AI-inferred models and parameters are subject to pre/post-admission checks and empirical validation, ensuring research-grade rigor. The agent is strictly used for research-staging and parameter/model preparation; safety-critical onboard computation remains infeasible for current and foreseeable LLM infrastructure.

Physics-Informed HAP Propulsion Power Modeling

Propulsion power modeling is established via physics-driven formulation, validated with CFD, and systematically improved using AI-assisted parameterization. Classical models for HAP propulsion power $P_{\text{prop}}$ based on isolated drag and propeller analysis significantly underestimate true power requirements due to the neglect of hull-propeller aerodynamic interference (as shown by large simulation/model deviations, especially at high airspeeds).

Aerodynamic Drag and Propeller Efficiency:

• Drag is formulated as $T = 0.5 \rho V_0^2 C_D \Omega^{2/3}$, with $C_D$ incorporating geometry, tail effects, and Reynolds number dependence.
• Propeller efficiency is highly sensitive to local flow, hull proximity, and pressure fields—factors not tractably captured in closed-form. Direct CFD analysis, performed over an extensive airspeed grid, yields $\eta_p(V_0)$ profiles substantially lower than isolated-propeller predictions.
• A regression-based inverse-power surrogate, $\hat{\eta}_p(V_0) = -0.2 V_0^{-9/20} + 0.73$, learned from CFD samples, achieves tight empirical fit (RMSE $< 0.5\%$).

Composite Power Model:

Substituting the above into $P_{\text{prop}} = T V_0 / (\hat{\eta}_p \eta_m)$ realizes a comprehensive, HAP-geometry-specific propulsion model. The maximum model deviation versus CFD is reduced to $<6\%$ for the derived model, compared to $>20\%$ for prior benchmarks, achieving on average an 84.3% reduction in deviation and a maximum error drop from 8367.3 W to 85.32 W at 25 m/s.

Analytical models are further validated against turbulence closure schemes (SST $k$-$T = 0.5 \rho V_0^2 C_D \Omega^{2/3}$0, realizable $T = 0.5 \rho V_0^2 C_D \Omega^{2/3}$1-$T = 0.5 \rho V_0^2 C_D \Omega^{2/3}$2), with $T = 0.5 \rho V_0^2 C_D \Omega^{2/3}$3 result variability, and operating point uncertainties conservatively capped at 3.4%.

Joint Power Allocation and Communications Optimization

Given a rigorous propulsion power model, the feasible budget for RF transmit power in the communications subsystem is tightly constrained and accurately determined. The communications power model covers PA inefficiency, static (LO, BB, RF) consumption, and beamforming vector norms, enabling system-level total power compliance.

The beamforming problem is formulated as a lexicographically prioritized multi-objective optimization: maximize the cardinality of QoS-satisfied users, then maximize system EE. System modeling follows UPA-based massive MIMO with a Rician fading channel, zero-forcing beamforming, and per-user statistical CSI.

Optimization Problem:

$T = 0.5 \rho V_0^2 C_D \Omega^{2/3}$4

subject to ergodic rate constraints, per-user minimum powers, and total transmit power feasibility. The strict concavity of the log-rate curve motivates the separation of user selection (greedy cardinality maximization under power knapsack) and power allocation (EE maximization).

ANN-Based QoS-Enhanced Energy-Efficient (Q3E) Beamforming Algorithm

To address the non-convexity and integrate constraint handling efficiently, the Q3E algorithm couples a dedicated feedforward neural network (with four hidden layers), in-the-loop capped-simplex feasibility projection, and barrier-augmented EE loss function. Key technical elements:

• Constraint-Projected Training: At each epoch and mini-batch, the network output is projected to ensure per-user minimum power feasibility and total power constraint, guaranteeing hard constraint satisfaction at each step.
• Barrier Loss: Allows unconstrained gradient-based learning while penalizing infeasible slack, removing the reliance on labeled datasets.
• Lexicographic Policy Implementation: The Q3E optimizer first selects the maximal set of QoS-admissible users, then allocates surplus power for EE improvement among non-QoS users.

Comparative baselines include SLSQP, PPO, ACOR, and max-sum-rate algorithms.

Numerical Results

Simulation covers a flight scenario at 20 km altitude, UPA (12x12), 9 users, heterogeneous QoS levels (high/medium/low), and validated channel/environmental settings. Key quantitative findings:

• The derived propulsion power model yields an 84.3% average reduction in deviation over prevailing alternatives, enabling a more accurate and less conservative communication power allocation.
• Q3E Beamforming Performance: Under identical communication power constraints, Q3E surpasses all baselines in both user QoS satisfaction and EE metrics. For budgets between 90–150 W, Q3E improves EE by 35–45% compared to QoS-satisfaction alone, and maintains a 33.3% higher maximum user satisfaction ratio and up to 45% higher EE over the full power range.

Feasibility ablation studies confirm that the in-loop projection is critical; omitting constraint projection causes feasibility collapse under tight budgets.

Implications and Future Directions

This work demonstrates that leveraging generative AI agents for parameterized knowledge integration and physics-informed modeling substantially advances multi-system HAP design. The methodology enables accurate, reliable propulsion modeling and cross-domain optimization not previously tractable with manually-curated or single-discipline approaches.

Practical impacts include:

• Tighter, more reliable system power budgeting for both propulsion and communications, increasing HAP operational flexibility.
• Enabling the deployment of advanced, energy-efficient beamforming under realistic platform constraints, improving QoS for terrestrial users.
• Providing a generalizable methodology for agent-assisted multi-domain optimization applicable to other energy-constrained aerial or spaceborne platforms.

Future research will extend to cooperative, multi-HAP networks with dynamic CSI, trajectory optimization, and further inclusion of test-time adaptation, domain generalization, and federated learning to improve robustness and system-level scalability.

Conclusion

The paper rigorously addresses the coupled propulsion and communication power allocation in HAP systems with a generative AI-empowered multidisciplinary framework. The validated propulsion model and the developed ANN-based constrained beamforming algorithm effectuate significant improvements in both modeling fidelity and operational performance. The research establishes a new standard for system-level energy-aware HAP design, forming the basis for future multi-agent, cooperative, and adaptive high-altitude networking systems (2604.09015).