Sensor Fusion-Based Machine Learning Algorithms for Meteorological Conditions Nowcasting in Port Scenarios

Area of Science:

• Maritime operations and environmental monitoring
• Machine learning applications in logistics
• Internet of Things (IoT) for port management

Background:

• Port operations are increasingly impacted by unpredictable weather and environmental shifts.
• Accurate short-term forecasting is essential for maritime safety and efficiency.
• Existing systems require enhanced situational awareness for critical operations.

Purpose of the Study:

• To develop a real-time, multi-target nowcasting framework for wind conditions.
• To integrate heterogeneous data sources using sensor fusion and IoT architecture.
• To evaluate machine learning models for predicting wind gust speed, sustained wind speed, and direction.

Main Methods:

• Utilized an IoT architecture (oneM2M standard) at the Port of Livorno.
• Integrated data from meteorological stations, anemometers, and vessel-mounted LiDAR.
• Employed feature-level sensor fusion and compared Random Forest, XGBoost, LSTM, TCN, Ensemble Neural Network, Transformer, and Kalman filter models.

Main Results:

• XGBoost demonstrated the highest accuracy for all wind targets (R² ≈ 0.999 in single-split, mean R² = 0.9976 in cross-validation).
• Ensemble models showed improved robustness compared to deep learning methods.
• The sensor fusion framework effectively enhanced situational awareness for critical variables like gust speed.

Conclusions:

• The proposed sensor fusion-based machine learning framework is highly effective for real-time wind nowcasting.
• XGBoost is a top-performing model for this multi-target prediction task.
• The framework has significant potential for deployment in Maritime Autonomous Surface Ship (MASS) systems and port decision-support platforms for enhanced safety and operational continuity.