Affiliation:
1. Electromagnetic Compatibility Laboratory, Swiss Federal Institute of Technology (EPFL), 1015 Lausanne, Switzerland
2. State Grid JiangSu Electric Power Co., Ltd. Suzhou Branch, Suzhou 215000, China
3. Institute for Information and Communication Technologies, University of Applied Sciences of Western Switzerland (HES-SO), 1400 Yverdon-les-Bains, Switzerland
Abstract
Lightning is directly or indirectly responsible for significant human casualties and property damage worldwide. A timely prediction of its occurrence can enable authorities and the public to take necessary precautionary actions resulting in diminishing the potential hazards caused by lightning. In this paper, based on the assumption that atmospheric phenomena behave in a continuous manner, we present a model based on residual U-nets where the network architecture leverages this inductive bias by combining information passing directly from the input to the output with the necessary required changes to the former, predicted by a neural network. Our model is trained solely on lightning data from geostationary weather satellites and can be used to predict the occurrence of future lightning. Our model has the advantage of not relying on numerical weather models, which are inherently slow due to their sequential nature, enabling it to be used for near-future prediction (nowcasting). Moreover, our model has similar performance compared to other machine learning based lightning predictors in the literature while using significantly less amount of data for training, limited to lightning data. Our model, which is trained for four different lead times of 15, 30, 45, and 60 min, outperforms the traditional persistence baseline by 4%, 12%, and 22% for lead times of 30, 45, and 60 min, respectively, and has comparable accuracy for 15 min lead time.
Funder
JiangSu Electric Power Co., Ltd. Suzhou Branch, Suzhou, China
Swiss National Science Foundation
Subject
Atmospheric Science,Environmental Science (miscellaneous)
Cited by
2 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献