Adaptive Broad Learning Network for Lightweight Multi-Dimensional Spectrum Prediction

Multi-dimensional spectrum prediction is pivotal for enabling spectrum sharing and dynamic spectrum access (DSA), which are critical for mitigating spectrum resource scarcity and enhancing wireless communication system performance. Traditional spectrum prediction techniques predominantly employ machine learning (ML) or deep learning (DL) approaches. However, ML-based methods often require manual feature extraction, leading to inefficiencies, while DL-based techniques typically demand computationally intensive graphics processing units (GPUs). To address these limitations, this paper proposes a lightweight multi-dimensional spectrum prediction model based on an adaptive broad learning network (ABLN). The model utilizes a sliding window approach for preprocessing historical spectrum data and establishes an input layer containing mapped feature nodes and enhancement nodes by randomly generating weights and biases. The weight parameters of the network are determined by solving the pseudo-inverse problem, and the network structure is extended by incremental learning method without necessitating retraining, effectively reducing the model's computational complexity. Furthermore, an adaptive node increment module is employed to quickly obtain the optimal hyperparameters. Experimental results on real-world spectrum datasets demonstrate that the proposed ABLN method consistently outperforms benchmark algorithms in terms of both prediction accuracy and computational efficiency across multi-step prediction tasks. Additionally, the model exhibits robust stability and adaptability in diverse scenarios.