EnergyTwin: 6G Energy & Latency Optimization

• EnergyTwin is a specialized digital twin framework that integrates physical sensing and AI-driven control for real-time energy management and latency optimization in 6G smart city deployments.
• Its hierarchical architecture spans device, edge, and core tiers, ensuring secure and efficient resource allocation through PUF-based attestation and federated learning.
• The framework leverages compressive sensing and renewable-aware scheduling to reduce telemetry volume and energy consumption, achieving significant energy reductions and sub-millisecond latency compliance.

An EnergyTwin is a specialized digital twin framework designed for real-time monitoring, modeling, co-optimization, and control of energy and latency in massive-scale 6G-enabled smart city deployments. It consists of a hierarchical architecture that integrates physical asset states, edge intelligence, compressive telemetry, AI-federated learning resource allocation, renewable-aware scheduling, and robust security mechanisms. The EnergyTwin maintains continuous synchronization between the physical and digital domains to achieve substantial energy reductions while sustaining strict SLA-driven latency guarantees, with demonstrated scalability to tens of thousands of heterogeneous devices per square kilometer (Abouaomar et al., 2 Nov 2025).

1. Multi-Tier System Architecture and Core Components

EnergyTwin’s architecture comprises a three-tier hierarchy:

Device Tier:

• Up to 50,000 IoT devices per km², partitioned across mMTC (60%), eMBB (30%), and URLLC (10%) slices.
• Each device $d_i$ is equipped with a Physical Unclonable Function (PUF) for hardware-rooted attestation, and continuously reports local state vectors (e.g., battery, traffic) into the local CyberTwin instance.

Edge Tier:

• 100 gNodeBs per deployment, each collocated with Multi-access Edge Computing (MEC) resources $(C_j, R_j, B_j)$ denoting CPU, RAM, and bandwidth capacities.
• Every gNodeB $g_j$ runs a local CyberTwin $\mathcal T_j$ that:
  1. Gathers compressive-sensing-based telemetry from all attached devices.
  2. Maintains slice-specific resource and renewable-energy models.
  3. Orchestrates hybrid scheduling (centralized DNN + federated learning) for resource allocation.
  4. Runs local FL clients for optimization of non-latency-critical slices.

Core Tier:

• Central orchestrator responsible for real-time centralized AI scheduling for URLLC/RTS slices and a federated aggregator (Byzantine robust: Krum) for learning model aggregation on non-critical slices.

• Integrated renewable prediction (ARIMA(2,1,2)-based solar forecasting) module supplies per-gNodeB forecasts, influencing local slice scheduling.

This design enables fine-grained, real-time control with vertical separation of concerns and strong security attestation using PUFs (99.7% attack detection accuracy).

2. Joint Energy–Latency Optimization

Resource allocation is cast as a constrained multi-objective optimization:

$$\min_{\boldsymbol\alpha,\,\boldsymbol\tau} \; J(\boldsymbol\alpha,\boldsymbol\tau) = \lambda\,E_\mathrm{total}(t) + (1-\lambda)\,L_\mathrm{total}(t)$$

subject to per-slice SLA latency constraints

$$L_s(t) \le L_s^{\max}, \quad L_{LSS}^{\max}=1\,\mathrm{ms},\; L_{RTS}^{\max}=5\,\mathrm{ms},$$

as well as MEC computational, memory, and bandwidth bounds: $$\sum_{s}\alpha_{s,j}^{(c)}\leq C_j, \quad \sum_{s}\alpha_{s,j}^{(r)}\leq R_j, \quad \sum_{s}\alpha_{s,j}^{(b)}\leq B_j$$ and minimal scheduling intervals $\tau_i \geq \tau_{i}^{\min}$.

Here, $E_\mathrm{total}(t)$ models the aggregate site-level computation, communication, and solar offset, while $L_\mathrm{total}(t)$ sums per-slice latencies. The optimization is solved subject to hardware and SLA constraints at each scheduling epoch.

3. Hybrid AI/Federated Learning Scheduler

The scheduler dynamically selects between centralized and federated scheduling:

• URLLC/RTS Slices:

Centralized DNN with 128 hidden units on the orchestrator directly outputs allocation vectors $\boldsymbol\alpha$ for strict sub-millisecond targets. If latency enforcement fails, a fallback heuristic is applied.

• Non-Real-Time Slices (NRTS):

Distributed FL clients (64 hidden units per client) perform local model updates; Byzantine-robust Krum aggregation computes the global model, and predictions are applied to new network states. Only NRTS slices are handled via FL for energy efficiency.

• Security:

All requests undergo PUF-based verification; failed verifications trigger quarantine actions on resource allocations.

This hybrid workflow eliminates centralized bottlenecks for non-critical workloads, maximizes URLLC performance, and reduces over-the-air signaling overhead.

4. Compressive Sensing and Telemetry Reduction

To overcome the overhead of high-frequency state telemetry, EnergyTwin implements compressive-sensing-based digital twinning:

$$\mathbf{y} = \boldsymbol\Phi\,\mathbf{x}\,, \quad \boldsymbol\Phi\in \mathbb{R}^{m\times n},\, m = 0.3n$$

with reconstruction via

$$\hat{\mathbf{x}} = \arg\min_{\mathbf{z}}\lVert \mathbf{z} \rVert_1 \quad \text{s.t.} \quad \boldsymbol\Phi\,\mathbf{z} = \mathbf{y}$$

Priority-based subsampling further downscales low-priority device traffic, transmitting only a subsampled vector.

This achieves a 70% reduction in uplink telemetry volume, accelerates FL convergence, and preserves sufficient state accuracy for reliable optimization and SLA guarantees.

5. Renewable-Aware Resource Scheduling

EnergyTwin incorporates renewable forecasts into resource allocation using a quadratic dissatisfaction metric: $$D_\mathrm{energy}(t) = \sum_{s} w_s \left( \frac{E_s^\mathrm{actual}(t) - E_s^\mathrm{target}(t)}{E_s^\mathrm{target}(t)} \right)^2$$ and applies conditional logic: $$\mathrm{Action}(s, t) = \begin{cases} \text{AllocateRenewable}, & I_\mathrm{solar}(t)>\theta,\, s=\mathrm{NRTS} \ \text{DelayAllocation}, & I_\mathrm{solar}(t)\leq\theta,\, s=\mathrm{NRTS} \ \text{ImmediateAllocation}, & s\in\{\mathrm{LSS},\mathrm{RTS}\} \end{cases}$$ where $\theta = 700\,\mathrm{W/m^2}$. Non-critical loads are thus deferred or concentrated in solar-rich epochs, further curbing grid energy draw.

6. Performance Metrics and Evaluation

Extensive NS-3 hybrid simulations in realistic 6G smart-city scenarios yield:

Metric	Value/Result
NRTS energy reduction	52.3% vs. Diffusion-RL baseline
System end-use power (DL baseline)	5,100 W
System end-use power (EnergyTwin)	2,450 W
Real-time solar usage (NRTS, peak)	68% supplied by solar
URLLC (LSS) 99th-pct. latency	0.89 ms (target < 0.9 ms)
SLA compliance (URLLC)	99.2%
CPU overhead across MEC	< 25%
Scalability	Up to 50,000 devices/km²
PUF-based security	99.74% detection accuracy
FL convergence	95% accuracy in 78 rounds
Communication overhead reduction	45% (selective FL + comp. sensing)

These results confirm both orders-of-magnitude scaling and robust compliance with stringent 6G requirements for latency, energy, and security.

7. Architectural Significance and Extensibility

By vertically integrating physical sensing, compressive digital modeling, AI-driven orchestration, and adaptive renewable scheduling, the EnergyTwin framework establishes a demonstrably scalable, energy-proportional, and latency-assured resource management paradigm for next-generation smart cities. Its modularity supports extension to emerging use cases, including dynamic network slicing, real-time FL defense, security event response, and ML-based anomaly detection. All critical design and performance claims are validated through reproducible simulation and analytic benchmarks (Abouaomar et al., 2 Nov 2025).

---

EnergyTwin thus exemplifies state-of-the-art cyber-physical energy–latency co-optimization for urban-scale, mission-critical wireless infrastructure under operational and environmental constraints.