Web and social event signals for AI-driven mobile network demand forecasting

Background Mobile network demand is increasingly volatile due to collective human activities such as concerts, sports events, and public gatherings. Traditional capacity planning methods, largely based on historical network indicators, struggle to anticipate these transient and localized demand surges. Recent advances in social sensing and artificial intelligence suggest that web and social signals can provide early indicators of real-world events, enabling proactive, event-aware network management in future mobile network infrastructures. Methods This paper presents a combined analysis and empirical study on AI-driven event-aware demand forecasting for mobile networks. We first review methods for extracting event signals from web and social data and analyze prior evidence linking such signals to cellular traffic variations. We then introduce a forecasting-driven orchestration pipeline and evaluate it through a case study using the NetMob’23 dataset, which provides high-resolution, service-level mobile traffic traces from multiple urban areas. The empirical evaluation focuses on the city of Nantes. Several forecasting models—ranging from naïve baselines and linear regression to Random Forests and LSTM neural networks, and a PatchTST-style Transformer baseline —are compared. We further investigate the impact of event-related features based on scheduled-event contextual information and temporally aligned event indicators and introduce an asymmetric loss function designed to penalize traffic underestimation in proactive orchestration scenarios. Results Among the evaluated models, LSTM opt_std achieves the lowest RMSE (1.94 × 10 7 ) and the highest R2 (0.936), whereas the asymmetric-loss LSTM opt_cst achieves the largest relative overload reduction, reaching 71.3%, 69.1%, and 61.2% at the 70th, 80th, and 90th percentile capacity thresholds, respectively, with RMSE =2.24 × 10 7 and MAPE =24.01%. For opt_cst, removing event-related inputs increases RMSE from 2.24 × 10 7 to 2.59 × 10 7 and MAPE from 24.01% to 28.9%, corresponding to reductions of 13.4% and 16.9% when event-aware inputs are used. Event-input perturbations produce modest and non-monotonic sensitivity, with RMSE changes between −1.90% and + 0.21% and OR@80 changes between −1.37 and + 1.19 percentage points relative to the clean condition. Conclusions The study confirms that event-aware contextual features into AI-based forecasting pipelines provides a measurable anticipatory advantage for proactive mobile network orchestration. Event-aware forecasting emerges as a promising component for predictive, self-optimizing future mobile network infrastructures, bridging social sensing and automated network management.