A novel pyramid temporal causal network for weather prediction

Area of Science:

• Deep Learning
• Meteorological Forecasting
• Time Series Analysis

Background:

• Traditional deep learning sequence prediction models struggle to extract and utilize multi-scale temporal features from historical meteorological data for accurate weather forecasting.
• Existing methods often fail to adequately capture diverse temporal dependencies crucial for predicting future weather patterns.

Purpose of the Study:

• To propose a novel deep learning model, the Pyramid Temporal Causal Network (PTCN), designed to effectively extract and leverage multi-scale temporal features for enhanced weather prediction.
• To introduce a new multivariate loss function (MVLoss) that improves prediction accuracy, particularly for variables with small variance, outperforming traditional loss functions like Mean Square Error (MSE).

Main Methods:

• Development of the Pyramid Temporal Causal Network (PTCN), comprising stacked causal dilated blocks to process and integrate features across multiple temporal scales.
• Implementation and evaluation of a novel multivariate loss function (MVLoss) designed to better handle variables with low variance in multivariate prediction tasks.
• Extensive testing of the PTCN model and MVLoss on the Weather Forecasting Dataset 2018 (WFD2018) across various prediction tasks.

Main Results:

• The PTCN model demonstrates significant benefits from utilizing multi-scale temporal features, leading to improved weather prediction performance.
• The MVLoss function shows superior performance in fitting small variance variables and enhances the overall average prediction accuracy compared to MSE.
• Experimental results confirm the effectiveness of both the PTCN architecture and the MVLoss function in advancing meteorological forecasting capabilities.

Conclusions:

• The proposed PTCN model successfully addresses the challenge of extracting multi-scale temporal features for improved weather prediction.
• The MVLoss function offers a significant advancement in multivariate prediction, particularly for variables with small variance, and boosts overall model accuracy.
• This research contributes novel deep learning methodologies for more accurate and robust meteorological forecasting.