Prediction Horizon

The prediction horizon specifies the exact future time step or interval for which channel state is forecast. It is typically expressed in milliseconds or as a number of discrete time slots. A horizon that is too short fails to provide sufficient time for a proactive spectrum handoff, while an excessively long horizon introduces high uncertainty and computational overhead. The optimal window is bounded by the channel holding time and the minimum handoff execution latency of the cognitive radio.

Prediction architectures are categorized by their output granularity:

• Single-step prediction: Forecasts occupancy only for the immediate next time slot. Computationally lightweight but limits proactive planning.
• Multi-step prediction: Forecasts a sequence of future states over the entire horizon. This is essential for target channel reservation and scheduling transmission bursts within a spectrum availability window. Architectures like Encoder-Decoder LSTM are specifically designed for this task.

The length of the prediction horizon directly dictates the viable handoff strategy. A horizon shorter than the handoff execution time forces a reactive spectrum handoff, resulting in higher latency and potential packet loss. A sufficiently long horizon enables proactive spectrum handoff, where the secondary user can seamlessly transition to a pre-identified idle channel before the primary user arrives, minimizing the forced termination probability.

Prediction confidence degrades as the horizon extends. Advanced models address this by providing a predictive distribution rather than a point estimate. Gaussian Process Regression naturally outputs a variance estimate for each future time step. Bayesian neural networks and Stein Variational Gradient Descent (SVGD) approximate the posterior distribution over parameters, allowing the handoff decision engine to weigh predictions by their associated uncertainty.

The choice of predictive model is heavily influenced by the required horizon:

• Short horizons (< 50 ms): Simple ARIMA models or linear regression on recent Channel State Information may suffice.
• Medium horizons (50-200 ms): LSTM Spectrum Predictors capture temporal dependencies effectively.
• Long horizons (> 200 ms): Partially Observable MDPs (POMDPs) and Deep Q-Networks learn policies that account for long-term cumulative reward, integrating prediction and decision-making.

A fixed prediction horizon may become suboptimal if primary user traffic patterns change over time. Concept drift adaptation mechanisms continuously monitor the accuracy of forecasts at the target horizon. If a change point detection algorithm identifies a statistical shift in the Primary User Activity Model, the system can dynamically adjust the horizon length or trigger model retraining to maintain prediction reliability.