Forecasting High Resolution Precipitation Events With Logistic Echo State Networks

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Moncada Morales, Lizda Nazdira, Bonas, Matthew, Castruccio, Stefano, Crippa, Paola
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• Subjects: echo state networks; forecasting; logistic regression; time series; urban precipitation; WRF
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000291

Accurately predicting hydroclimate events is crucial for understanding the impacts of climate change and effectively managing water resources, particularly for flood mitigation and timely warnings. Despite recent advances in machine learning, forecasting precipitation events continues to be challenging due to inherent data imbalances and the intricate dynamics governing these occurrences, rendering them difficult to model accurately. Echo State Networks (ESNs) offer a promising solution; their ability to model complex processes has been demonstrated throughout the field of environmental science. For this work, we propose a novel adaptation of ESNs, termed BinESN, to binary classification problems of precipitation occurrence. In particular, we extend the ESN to a generalized linear model framework, we leverage the ESN's ability to recognize complex dynamics while maintaining interpretability of the predicted output. Through simulation studies and an application to numerically simulated precipitation, we show that BinESN produces more accurate forecasts of sparse events in both short‐ and long‐range scenarios compared to other common machine learning approaches. Specifically, the proposed BinESN outperforms other reference methods by over 10% in terms of its area under the receiver operating characteristic curve. Plain Language Summary Predicting hourly precipitation data is critical in weather forecasting, and even more so in urban environments where severe precipitation could lead to severe consequences spanning from human loss to infrastructure damage. In this work, we introduce a novel machine learning approach which is able to capture the precipitation dynamics via a special type of neural network able to account for data in time, while also recognizing that hourly precipitation is a sparse event, that is, it has a large probability of not occurring. Our results show that this approach leads to improved forecasting against other traditional approaches. Key Points A new model based on a generalized linear model extension of an echo state network is developed The new model allows to capture the dynamics of sparse time series such as precipitation occurrence and more generally non‐Gaussian data The new model outperforms traditional approaches in forecasting hourly precipitation occurrence in regional model simulations