PCDT: A Predictive Cognitive Digital Twin Framework for Intelligent and Autonomous 6G Network Ecosystems

• Because a present-state replica cannot prevent a surge after it creates congestion, the twin must reason about future conditions and act in advance.Network DTs can synchronize telemetry, represent network state, and support safe evaluation of control decisions.Source: §I · p. 1
• PCDT combines a persistent traffic world model, uncertainty-aware horizon forecasts, proactive capacity decisions, and continuous re-synchronization after actions.The comparison with traditional reactive DTs emphasizes forecasting and capacity control before a violation occurs.Source: §I · pp. 1–2, Figure 1 · p. 1
• The control problem balances SLA violations from underallocation, wasted resources from over-allocation, and churn from frequent changes.The paper formulates a multi-objective controller to jointly minimize provisioning cost, violations, and reconfiguration churn.Source: §I · p. 2
• 43% lower MAE and 50% lower operating cost were reported for PCDT compared with baseline methods.Source: §I · p. 2
• The paper presents PCDT and its results as aligned with 6G and 3GPP properties and roadmap.Source: §I · p. 2

A. Network Digital Twins

• Most network digital-twin frameworks focus on performance optimization, while fewer studies address deeper closed-loop autonomy.Source: A. Network Digital Twins · p. 2
• Related architectures emphasize synchronization and lifecycle mirroring, and some handle synchronization, prediction, and resource allocation separately,,.The IoDT review identifies data synchronization between physical and digital spaces as a core requirement and challenge.Source: A. Network Digital Twins · p. 2
• PCDT is positioned to integrate synchronization, uncertainty-aware prediction, proactive allocation, and continuous updates from action outcomes.Source: A. Network Digital Twins · p. 2

B. Traffic Forecasting and Control

• Forecasting spans attention and graph-temporal models, but accurate forecasts alone do not guarantee good control; PCDT connects forecasts to capacity decisions and feeds outcomes back into the twin.The cited control landscape includes static provisioning, threshold heuristics, model predictive control, reinforcement learning, and intent-based automation, with trade-offs in efficiency, stability, and training needs.Source: B. Traffic Forecasting and Control · p. 2

C. Autonomous 6G Network Ecosystems

• Autonomy-oriented work includes drift-triggered re-synchronization and O-RAN monitoring, simulation, and optimization, while the cited mechanisms often keep cognition, prediction, and control separate.PCDT is presented as a closed-loop cognitive mechanism.Source: C. Autonomous 6G Network Ecosystems · pp. 2–3

A. Dataset and Preprocessing

• Traffic intensity is aggregated at 10-minute intervals in a real-world cellular traffic dataset.Traffic intensity reflects user demand and is relevant to capacity planning and service-level requirements.Source: A. Dataset and Preprocessing · p. 3
• A 25 × 25 grid region illustrates spatial variation in traffic intensity across cells.The dataset pairs time with cell_id, allowing analysis of an individual cell and neighboring-cell effects.Source: A. Dataset and Preprocessing · p. 3
• Cell 260 is the highest-traffic cell and a nonstationary urban hotspot, making it a testbed for comparing proactive and reactive allocation.Its load varies sharply across the day and rises through sustained peak windows.Source: A. Dataset and Preprocessing · p. 3
• Min-max scaling maps traffic intensity to [0, 1], with ε = 10^-9 added to the denominator to prevent division by zero for zero-variance cells.The scaling places variables on a common scale for comparison.Source: A. Dataset and Preprocessing · p. 3

B. Predictive Cognition Layer

• The feature vector combines recent load values with trend, rolling summaries, volatility, and time-of-day context.The window slope captures whether load is rising or falling, while volatility measures recent variability and signals regime instability.Source: B. Predictive Cognition Layer · p. 4
• Daily phase is encoded with sine and cosine, alongside a linear hour-of-day feature that represents absolute time-of-day effects.The cyclical encoding is continuous across the day boundary; the same technique is applied during training and inference to avoid train/serve skew.Source: B. Predictive Cognition Layer · p. 4
• The forecaster outputs future load estimates over a horizon, and a ridge-regularized linear autoregressor was consistently selected across pipeline runs.Persistence, the linear autoregressor, and a three-layer MLP were trained on the same feature representation; selection used validation RMSE, followed by one held-out test evaluation.Source: B. Predictive Cognition Layer · p. 4

C. Goal-Oriented Control Layer

• The controller sets capacity to balance SLA violations when load exceeds capacity, wasted resources when capacity exceeds load, and reconfiguration cost when capacity changes frequently.These are the costs described for the controller's total objective.Source: C. Goal-Oriented Control Layer · p. 4
• The controller uses the 95th percentile of horizon forecasts, an uncertainty-scaled safety margin, and exponential smoothing to determine capacity.The 95th percentile trades a small tail risk for lower over-provisioning; the adaptive margin widens with forecast standard deviation, while smoothing damps rapid reconfigurations.Source: C. Goal-Oriented Control Layer · p. 4
• Using max(ˆx_t) instead of the smoothed forecast costs 1.3% more capacity for the same violation rate.The smoothed choice is described as damping single-step forecast noise without loss in reliability.Source: C. Goal-Oriented Control Layer · p. 4
• During warm-up, before the lookback window is available, the algorithm sets capacity to current load plus the base margin.Once operating, it featurizes load history, forecasts the horizon, and returns a nonnegative capacity.Source: C. Goal-Oriented Control Layer · p. 4

D. Benchmark Controllers

• Static overprovisioning allocates capacity at the 98th percentile of test load as a conservative response to traffic fluctuation.This is the conservative provisioning baseline.Source: D. Benchmark Controllers · p. 4
• Reactive digital-twin control allocates the previous observed load plus a margin, mirroring traffic reactively.This baseline represents reactive mirroring for traffic variation.Source: D. Benchmark Controllers · p. 4
• The threshold heuristic increases capacity using a recent positive slope, representing handcrafted anticipation in practical traffic-load planning.The paper benchmarks PCDT against three frameworks.Source: D. Benchmark Controllers · p. 4

IV. RESULTS DISCUSSION

• PCDT records MAE of approximately 0.028 and RMSE of approximately 0.035, versus 0.049 and 0.052, respectively, for reactive control.Figure 4 compares MAE and RMSE across the frameworks.Source: §IV · p. 4, Figure 4 · p. 5
• PCDT allocates the lowest mean normalized capacity (≈0.05), compared with approximately 0.07–0.10 for baseline controllers.The distribution of normalized capacity allocation is shown in Figure 5.Source: §IV · p. 5, Figure 5 · p. 5
• PCDT ranks first in violation rate, capacity utilization, violation severity, and total cost, with a superior multi-objective trade-off.Figure 6 shows benchmark ranks, while Figure 7 presents the normalized controller trade-off.Source: §IV · p. 5, Figure 6 · p. 5, Figure 7 · p. 5
• PCDT ranks second in reconfiguration churn; static overprovisioning is most stable because it never changes allocation, but uses significantly more resources.The resource-usage trade-off is shown in Figures 6 and 7.Source: §IV · p. 5, Figure 6 · p. 5, Figure 7 · p. 5
• Reactive and threshold-based controllers align one time step after load changes (≈1.00 and ≈0.89), whereas PCDT aligns four time steps ahead.The passage reports that this alignment enables lower capacity allocation without increasing SLA violations; Figure 8 compares correlation lag.Source: §IV · p. 5, Figure 8 · p. 5

V. CONCLUSION

• The conclusion argues that digital twins should move from synchronized replicas toward entities that infer future network intent and allocate resources before service degradation.PCDT is presented as one practical realization of this transition toward AI-native autonomous 6G networks.Source: §V · p. 5
• Future work will study multiple-cell analysis, integration of multiple forecasting frameworks, and evaluation across different changing traffic distributions.Source: §V · p. 5