Fig 1.
Temporal evolution of tourist traffic flows in Aosta Valley from 2020-2023.
The time series displays 41,061,941 vehicle passages aggregated across all 14 monitoring gates. The blue line represents daily traffic flow, while the purple line shows the 30-day moving average revealing underlying trends. The graph exhibits clear dual-season patterns with pronounced peaks during summer months (July-August, highlighted in orange) and winter skiing seasons (December-January, highlighted in light blue). The red shaded region (March 2020 – June 2021) illustrates the dramatic impact of COVID-19 restrictions, showing traffic reduction of up to 70% during lockdown periods. Annotations mark key events: the initial lockdown, recovery phase, and seasonal peaks. The data demonstrates progressive recovery and normalization of tourist activity throughout 2022-2023, with traffic volumes approaching and exceeding pre-pandemic levels by late 2023.
Table 1.
Sensor portal locations and characteristics.
Fig 2.
Adaptive Temporal Ensemble (ATE) framework pipeline.
The system architecture displays the complete data flow from multi-source inputs through the ensemble prediction process. The pipeline begins with three parallel data streams: traffic sensor data (14 monitoring gates, 41M+ passages), meteorological records (43K+ hourly observations with 673 variables), and calendar information (52K+ records with 98 features).
Fig 3.
Adaptive Temporal Ensemble (ATE) architecture.
The framework combines four base learners through a meta-learning layer that dynamically adjusts weights based on real-time performance and contextual features. The performance feedback loop enables continuous adaptation to changing data patterns.
Table 2.
Diebold–Mariano test results: ATE versus competing models (log-difference series, test period October–December 2023). DM statistic and p-value for the null of equal predictive accuracy against ATE. Negative DM statistic indicates ATE superiority.
Table 3.
Performance by regime (holiday vs. regular periods), test period October–December 2023. ATE improvements are computed relative to XGBoost.
Table 4.
Performance comparison of forecasting models.
Fig 4.
Comprehensive performance analysis of the Adaptive Temporal Ensemble.
(A) Mean Absolute Error decreases with ensemble approach across all prediction horizons. (B) Consistent high performance across all seasons with R2 scores above 0.91. (C) Dynamic weight evolution showing adaptive behavior where XGBoost receives higher weight during stable periods while other models contribute more during anomalous conditions.
Table 5.
Ablation study results.
Fig 5.
Adaptive weight distribution across different weather conditions.
During adverse weather (heavy snow, storms), the ensemble increases Random Forest weight from 20% to 35% due to its robustness to outliers and missing data, simultaneously reducing XGBoost weight from 45% to 30%. This automatic rebalancing occurs because Random Forest’s bagging mechanism provides inherent stability against data perturbations, while XGBoost’s gradient-based optimization becomes less reliable when weather disruptions create non-representative training patterns. Conversely, during clear weather periods with stable traffic patterns, XGBoost dominates (45% weight) as its ability to model complex feature interactions proves most effective for regular forecasting scenarios.
Fig 6.
Prediction performance (R2 scores) across all 14 monitoring gates.
Highway gates (g130, g131) achieve highest accuracy due to regular traffic patterns, while valley gates show more variation reflecting the complex tourist flow dynamics in mountainous terrain.