Problem

The exponential growth in demand for wireless services, driven by advancements like 6G and the Internet of Things, creates significant challenges for mobile network operators and regulators in efficiently managing spectrum resources. Regulators specifically struggle with direct observation of current spectrum demand, making it difficult to plan new spectrum releases, identify areas of under or over-supply, and adapt policies effectively. This necessitates accurate modeling of spectrum demand or reliable proxies to facilitate dynamic spectrum planning and allocation.

Method

This study proposes an AI-enabled, data-driven methodology to estimate spatial spectrum demand, focusing on validating demand proxies and developing generalizable machine learning models. The approach divides five major Canadian cities (Montreal, Ottawa, Toronto, Calgary, Vancouver) into 1.5 km x 1.5 km grid tiles. Three proxies for spectrum demand are defined: the Deployed Bandwidth Proxy (from self-reported site license data), the Active Users Proxy (from crowdsourced data representing user activity), and a Combined Proxy that empirically weights and integrates both to mitigate individual limitations. Various spatial features, including demographic, economic, physical, and activity-based data, are aggregated to these grid tiles as inputs for the models. Two machine learning models, a baseline linear regression, and an XGBoost model, are employed. Cross-validation is performed using k-means clustering to address spatial autocorrelation, and spatial lag features are incorporated to capture neighborhood influences.

Results

Validation against real-world mobile network traffic data from Ottawa revealed that the Combined Proxy achieved the highest correlation with an R2 value of 0.85, outperforming the individual Deployed Bandwidth (R2=0.72) and Active Users (R2=0.64) proxies. In the predictive modeling phase, the XGBoost model using the Combined Proxy achieved the highest predictive accuracy with an R2 of 0.89, along with the lowest Normalized RMSE and MAE, demonstrating superior performance over individual proxies and a baseline model (R2=0.54). Feature importance analysis indicated that the number of small businesses, road segment counts, and daytime population were the most significant predictors across all models, highlighting their strong links to network demand. The models consistently demonstrated generalizability across the five Canadian cities.

Implications

The validated AI-enabled, data-driven approach for spectrum demand estimation offers significant implications for spectrum regulators. By providing accurate and generalizable forecasts of spatial spectrum demand, this framework supports more dynamic and efficient spectrum planning. It enables regulators to make data-driven decisions regarding resource allocation, identify areas of specific demand pressure, and implement targeted policy adjustments. This ultimately helps ensure adequate spectrum availability to meet the evolving and rapidly growing demands of future wireless networks.