Prediction of Spectrum Occupancy for Dynamic Spectrum Access Using Recurrent Neural Network

Spectrum handoff upon collision detection is considered an insufficient mechanism in cognitive radio, motivating the development of spectrum occupancy prediction methods for radiocommunication systems. Prediction of transmission opportunities within the bandwidth of primary users represents a promising direction that requires novel approaches supported by advancements in Machine Learning (ML) and Deep Learning (DL). Beyond performance, DL model complexity and computational cost are key parameters, both during training and inference. While numerous modern architectures have been proposed for prediction tasks, a gap still exists between complex and efficient models. We propose a simple yet effective method using a Recurrent Neural Network (RNN), specifically Long Short-Term Memory (LSTM), where the input feature is directly represented as time differences between transmission events. Raw input signals are reduced and processed through a fast signal-processing path, allowing the LSTM to infer from an activity history buffer on demand. The model is implemented using the standard toolset with the Keras framework. The proposed approach is evaluated using synthetic data simulating ON–OFF activity patterns, under assumptions tailored to a representative channel scenario. Results demonstrate that the time-differential LSTM approach can effectively predict and identify transmission window opportunities. LSTM provides learned immunity to the random nature of primary user activity, offering a tradeoff between model complexity and prediction capability.