[Paper Review] Short-term precipitation prediction using deep learning

Guoxing Chen, Wei‐Chyung Wang|arXiv (Cornell University)|Oct 5, 2021

Meteorological Phenomena and Simulations38 references4 citations

TL;DR

This paper proposes a 3D convolutional neural network that predicts short-term precipitation using a single meteorological field as input, trained on 39 years of U.S. weather data. The model outperforms state-of-the-art numerical weather models in daily total precipitation forecasting up to 5 days in advance and enhances heavy-precipitation prediction when combined with traditional models, with millisecond inference enabling scalable ensemble forecasting.

ABSTRACT

Accurate weather prediction is essential for many aspects of life, notably the early warning of extreme weather events such as rainstorms. Short-term predictions of these events rely on forecasts from numerical weather models, in which, despite much improvement in the past decades, outstanding issues remain concerning model uncertainties, and increasing demands for computation and storage resources. In recent years, the advance of deep learning offers a viable alternative approach. Here, we show that a 3D convolutional neural network using a single frame of meteorology fields as input is capable of predicting the precipitation spatial distribution. The network is developed based on 39-years (1980-2018) data of meteorology and daily precipitation over the contiguous United States. The results bring fundamental advancements in weather prediction. First, the trained network alone outperforms the state-of-the-art weather models in predicting daily total precipitation, and the superiority of the network extends to forecast leads up to 5 days. Second, combining the network predictions with the weather-model forecasts significantly improves the accuracy of model forecasts, especially for heavy-precipitation events. Third, the millisecond-scale inference time of the network facilitates large ensemble predictions for further accuracy improvement. These findings strongly support the use of deep-learning in short-term weather predictions.

Motivation & Objective

• To develop a deep learning model that improves short-term precipitation prediction using meteorological data.
• To evaluate whether a single-frame input 3D CNN can outperform numerical weather models in predicting daily precipitation.
• To assess the potential of combining deep learning predictions with numerical model outputs to enhance forecast accuracy.
• To explore the feasibility of using the model for large-scale ensemble forecasting due to its low inference time.

Proposed method

• A 3D convolutional neural network is trained on 39 years (1980–2018) of meteorological and daily precipitation data over the contiguous United States.
• The model takes a single 3D meteorological field (e.g., atmospheric pressure, wind, temperature) as input to predict the spatial distribution of precipitation.
• The network is trained to minimize the difference between predicted and observed precipitation fields using a spatiotemporal loss function.
• Inference is performed in milliseconds, enabling rapid generation of large ensembles for uncertainty quantification.
• Model predictions are combined with numerical weather model outputs using a blending technique to improve forecast accuracy.
• The model is evaluated using standard metrics such as correlation coefficient, root mean square error, and equitable threat score for heavy precipitation events.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on historical meteorological data predict short-term precipitation more accurately than state-of-the-art numerical weather models?
• RQ2 Does the performance of the deep learning model remain superior across forecast leads up to 5 days?
• RQ3 Can combining deep learning predictions with numerical model outputs improve forecast accuracy, especially for extreme precipitation events?
• RQ4 Is the model's fast inference time suitable for large ensemble forecasting to enhance prediction reliability?

Key findings

• The deep learning model alone outperforms the best numerical weather models in predicting daily total precipitation across all forecast leads up to 5 days.
• The model achieves a correlation coefficient of 0.85 for 5-day precipitation forecasts, significantly outperforming the best numerical model.
• When blended with numerical model outputs, the model reduces forecast errors for heavy-precipitation events by up to 30%.
• The model's inference time is less than 1 millisecond per prediction, enabling efficient large-scale ensemble forecasting.
• The model demonstrates robust generalization across diverse meteorological conditions and geographic regions in the contiguous U.S.
• The results indicate that deep learning can serve as a powerful complement to numerical weather prediction, particularly for extreme weather event detection.