To build a robust ML model, data from the first two years (2020-2022) was used for ML model training, while the third year (2023) was used for forecast validation. Regression models like Random Forest Regressor (RF), Lasso Regression, SVM Regressor, and XgBoost Regressor were combined to improve the accuracy. A Voting Regressor, an ensemble method, integrated predictions from the above four models for a unified and reliable approach. The findings highlight VOD's importance in bridging data gaps and improving situational awareness for airline operations. By analyzing historical weather observations and utilizing ML models, VOD accurately predicts temperature and pressure at specific locations even without direct METAR data.
The VOD methodology generates simulated temperature and pressure data valid for up to 1 hour. Initial validation results show that the MSE values for both temperature (ranging from 0.52 to 1.24) and pressure (ranging from 0.15 to 0.47) predictions varied significantly among the ten METAR locations. Some stations had lower MSE values, indicating higher prediction accuracy, while others had higher MSE values, suggesting less accurate forecasts. Understanding the factors contributing to these differences could lead to the development of more reliable weather forecasting systems overall. By blending historical observations and forecast models, VOD provides airline operators with accurate Virtual Observations, enabling informed decision-making and risk mitigation without direct METAR observations. As a result, VOD not only improves METAR data availability but also enhances airline operations safety and efficiency.
Keywords: Virtual Observations on Demand (VOD), METAR Data, Airline Operations, Caribbean Region, Tomorrow.io, ML Technology, Risk mitigation

