DeepSpectrum: A Localized Demand Estimation Model for Mobile Spectrum using Deep Learning

With the emergence of many new 5G and 6G use cases, the demand for spectrum continues to grow. In response, spectrum regulators worldwide are actively exploring innovative approaches to manage spectrum more efficiently. Spectrum sharing emerges as a particularly promising approach, as it allows for more intensive use of spectrum by enabling other services and users to access idle bands. Nevertheless, identifying areas where spectrum is either under- or over-supplied poses a significant challenge for regulators, given that demand insights are typically only observable to mobile operators.This paper proposes DeepSpectrum, a novel Deep Learning (DL) model that employs Multi-Task Learning, to estimate the local demand for mobile spectrum. The proposed model is trained on publicly available geospatial datasets and is used to estimate a novel demand proxy derived from crowdsourced data. The model features a combination of Convolutional Neural Networks and custom Residual Networks to extract relevant patterns from the data itself, thereby eliminating the need for traditional manual feature engineering. Overall, DeepSpectrum achieves a performance improvement of over 20% compared to alternative ML regression algorithms, demonstrating the advantage of DL for more accurate spectrum demand modeling.