PRB Utilization Prediction

PRB prediction operates at the slot and subcarrier level, forecasting demand on the smallest schedulable resource unit. A single PRB spans 12 subcarriers over one slot (1ms in 5G NR numerology 0), and a typical 20 MHz carrier contains 100 PRBs per slot. The prediction model must capture utilization patterns at this fine resolution to inform the MAC scheduler's resource allocation decisions. Key considerations include:

• Slot-level granularity: Predictions must align with the scheduling interval (1ms or shorter with mini-slots)
• Frequency-domain awareness: Different PRBs experience varying channel conditions and interference profiles
• Multi-numerology support: 5G NR's flexible subcarrier spacing (15kHz to 120kHz) changes the PRB count and slot duration, requiring adaptive prediction models

Accurate PRB prediction depends on ingesting a rich set of correlated telemetry features beyond historical utilization alone. The model learns the causal relationships between network conditions and resource demand. Critical input features include:

• Channel Quality Indicator (CQI): UE-reported metric (0-15) reflecting downlink channel conditions; higher CQI enables more efficient modulation and coding, affecting PRB consumption per bit
• RRC Connected Users: The number of active User Equipment devices in the cell directly drives aggregate PRB demand
• Buffer Status Reports (BSR): Uplink data volume pending at each UE, indicating imminent PRB requests
• Reference Signal Received Power (RSRP): Signal strength measurements that correlate with cell-edge user demand and potential handover events
• Historical PRB utilization: Lagged values at multiple time scales (5min, 15min, 1hr) to capture autocorrelation patterns

PRB utilization exhibits multiple superimposed temporal patterns that prediction models must disentangle. Effective architectures decompose the signal into its constituent components:

• Diurnal cycles: Predictable daily peaks (e.g., commuter traffic at 8am and 6pm) and troughs (3am maintenance windows)
• Weekly seasonality: Weekday vs. weekend traffic profiles, with enterprise cells showing pronounced Monday-Friday patterns
• Event-driven spikes: Concerts, sporting events, or emergencies causing sudden, localized PRB demand surges
• Trend components: Long-term growth in data consumption (30-40% year-over-year) requiring models to adapt to non-stationary baselines
• Holiday effects: National holidays causing atypical traffic patterns that differ from standard weekly cycles

LSTM and Transformer-based architectures excel here because their attention mechanisms and memory cells can learn these multi-scale dependencies without manual feature engineering.

PRB utilization in one cell is not independent of its neighbors. User mobility and inter-cell interference create strong spatial dependencies that predictive models must capture:

• Handover cascades: A congested cell offloads users to neighbors, propagating the PRB load spatially
• Interference coupling: High PRB utilization in one cell increases interference in adjacent cells, degrading their CQI and paradoxically increasing their PRB demand to maintain throughput
• Mobility trajectories: Users moving along highways or rail lines create predictable spatial-temporal PRB demand patterns across cell sequences

Graph Neural Networks (GNNs) are increasingly applied here, modeling the cellular topology as a graph where nodes are cells and edges represent neighbor relations with learned attention weights. This allows the model to condition each cell's prediction on the forecasted state of its neighbors.

The choice of prediction horizon (how far ahead to forecast) and lookback window (how much history to consider) fundamentally shapes model architecture and operational utility:

• Short horizon (10ms-1s): Enables real-time scheduling decisions in the Near-RT RIC. Requires low-latency inference and streaming feature pipelines. Suitable for per-slot PRB allocation optimization
• Medium horizon (1-60 minutes): Supports proactive load balancing and handover parameter tuning. Allows more complex model architectures with longer inference times
• Long horizon (hours-days): Informs capacity planning and energy-saving cell sleep mode decisions. Tolerates higher latency but must capture long-range dependencies

The lookback window must be long enough to capture the dominant seasonality (e.g., 24 hours for diurnal patterns) but short enough to avoid including stale, distribution-shifted data. Typical configurations use 168 hours (1 week) for weekly seasonality or 24 hours with explicit day-of-week embeddings.

PRB prediction models deployed in production face concept drift as network topology, user behavior, and traffic patterns evolve. Static models degrade within weeks. Robust systems implement:

• Online learning: Incremental weight updates from streaming telemetry using algorithms like Online Gradient Descent or Recursive Least Squares, avoiding costly full retraining
• Sliding window retraining: Periodic batch retraining on recent data windows (e.g., last 30 days) to forget obsolete patterns
• Drift detection triggers: Statistical tests (e.g., Kolmogorov-Smirnov test on prediction error distribution) that automatically flag performance degradation and trigger model refresh
• Ensemble weighting: Maintaining multiple model variants and dynamically weighting their predictions based on recent accuracy on a holdout set
• Transfer learning: Fine-tuning a global model on cell-specific data to rapidly adapt to new sites without training from scratch

The fundamental unit of resource allocation in LTE and 5G NR air interfaces. A PRB consists of 12 consecutive subcarriers in the frequency domain (180 kHz) and one slot in the time domain (0.5 ms for 5G NR). The total number of PRBs available depends on the channel bandwidth—a 100 MHz 5G carrier provides up to 273 PRBs. PRB utilization measures the percentage of these blocks actively scheduled for data transmission, making it the most direct indicator of cell load and the primary target variable for capacity forecasting models.

A critical input feature for PRB prediction models, reported by User Equipment (UE) to the gNB. CQI values range from 0 to 15 and indicate the highest modulation and coding scheme (MCS) the UE can decode with a block error rate below 10%. Key characteristics:

• Higher CQI enables more efficient PRB usage (fewer PRBs needed for same throughput)
• CQI varies rapidly due to mobility and fading
• Wideband CQI reports average channel quality across all sub-bands
• Sub-band CQI provides granular frequency-selective information
• Temporal CQI patterns are strong predictors of future PRB demand

The look-ahead window defining how far into the future PRB utilization is forecast. Horizon selection involves a fundamental trade-off:

• Short horizons (100ms–1s): Higher accuracy, suitable for real-time scheduler optimization and intra-slot resource allocation
• Medium horizons (1s–10s): Balance between accuracy and proactivity, ideal for Near-RT RIC xApp control loops and handover parameter adjustment
• Long horizons (10s–minutes): Lower accuracy but enables strategic actions like cell sleep mode activation and energy-saving carrier shutdown Multi-horizon architectures often use separate model heads to predict at multiple time scales simultaneously.

Long Short-Term Memory networks remain a workhorse architecture for PRB time-series prediction due to their ability to capture long-range temporal dependencies. Typical architecture includes:

• Input: Multivariate sequences of PRB utilization, CQI, RRC connections, and throughput
• Stacked LSTM layers: 2–4 layers with 64–256 hidden units each
• Attention mechanism: Often added to weight the importance of different historical time steps
• Output: Point forecast or probabilistic prediction (mean + variance) for uncertainty quantification LSTMs excel at learning daily/weekly periodicity in traffic patterns without explicit feature engineering, though they can struggle with very long sequences (>1000 time steps).

Modern PRB prediction models ingest diverse telemetry streams beyond historical utilization alone. Essential input features include:

• PRB utilization history: Up/downlink split, per-QCI class breakdown
• Active UE count: Number of RRC-connected users
• CQI distributions: Mean, variance, and percentile values across connected UEs
• Traffic volume metrics: PDCP layer throughput and buffer status reports
• Temporal encodings: Hour-of-day, day-of-week, holiday indicators
• Spatial context: Neighboring cell load states for inter-cell dependency modeling Feature importance analysis consistently shows that recent PRB history and CQI are the most predictive variables.

A deployment paradigm where the PRB prediction model continuously updates from streaming telemetry without full offline retraining. Key mechanisms:

• Incremental gradient updates: Model weights adjusted per batch of new data
• Sliding window retraining: Periodic full retraining on the most recent N days of data
• Ensemble weighting: Multiple model variants weighted by recent prediction accuracy
• Concept drift detection: Statistical tests (e.g., ADWIN, Page-Hinkley) trigger model refresh when data distribution shifts Online learning is essential for adapting to network upgrades, seasonal traffic changes, and new site deployments without manual intervention.