Deep Tower Networks for Efficient Temperature Forecasting from Multiple Data Sources.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Meteorology

Background:

• Handling diverse, simultaneous data streams presents significant challenges in data science.
• Current methods often rely on single data modalities or late fusion, limiting potential insights.
• Accurate short-term temperature forecasting is crucial for various applications.

Purpose of the Study:

• To introduce and evaluate a novel deep learning method, tower networks, for multi-modal data processing.
• To assess the effectiveness of tower networks in short-term temperature forecasting.
• To compare tower networks against established meteorological baselines and other deep learning architectures.

Main Methods:

• Development of a novel deep learning architecture: tower networks.
• Application of tower networks to short-term temperature forecasting using multiple data streams.
• Comparative analysis against meteorological baselines, statistical approaches, Convolutional Neural Networks (CNNs), and Convolutional Long Short-Term Memory (convLSTM) networks.

Main Results:

• Tower networks demonstrated strong performance compared to meteorological baselines and statistical methods.
• The method achieved an 11% lower root mean squared forecasting error than the yr.no operational service.
• Tower networks exhibited competitive performance, improved robustness, and comparable memory/training efficiency against CNN and convLSTM models.

Conclusions:

• Tower networks offer a promising solution for multi-modal data fusion in complex forecasting tasks.
• The proposed method significantly enhances short-term temperature forecasting accuracy.
• Tower networks represent a robust and efficient deep learning alternative for time-series forecasting.