1-5 December 2019
Africa/Johannesburg timezone
Note: Intel Keynote starts at 18:00 today (Monday)

The adoption of Deep Learning in Weather Forecast

3 Dec 2019, 14:30


Talk Earth Systems Modelling HPC Applications


Dr Yania Molina Souto (LNCC)


Technological development and the internet of things (IoT) increased the data assimilation sources in meteorology through satellites, sensors, weather stations, solar panels, cell phones, traffic lights, to name a few whose number grows daily. This begins to build an ideal scenario for Artificial Intelligence where the demand for data is high. This combined information generates spatio-temporal maps of temperature, rainfall, air movement, etc., with high precision in regions with larger data sources. Applying AI techniques in conjunction with a physical understanding of the environment can substantially improve prediction skill for multiple types of high-impact weather events.

In 2017, the American Meteorological Society (AMS) published a paper [1] with a broad summary indicating how modern AI techniques are helping to improve insights and make decisions in weather prediction. Among the most used techniques are Support Vector Machines (SVM), regression trees, k-means for radar image segmentation and traditional neural networks (ANN). These techniques lack temporal and spatio-temporal analysis, typical of meteorological phenomena. Today, most of the temporal analysis done on meteorological data is through statistical algorithms, such as autoregressive methods. However, the AMS recognizes that the novel Deep Learning techniques could soon be the cause of new improvements and says, “In the future, convolutional neural networks operating in a deep learning framework may reduce the need for feature engineering even further” .

Complementarily, recurrent neural networks, designed for analysis of natural language processing, are known by their results in numerical problems like temperature prediction, among others. In order to improve the temporary forecasts of spatio-temporal phenomena, such as rain and temperature, hybrid architectures have emerged that build the temporary forecast coding the spatial pattern of the neighborhood.

In [7], Shi et al. formulate precipitation nowcasting as a spatio-temporal sequence forecasting problem in which both the input and the prediction target are spatio-temporal sequences. They extend the fully connected LSTM (FC-LSTM) to have convolutional structures in both the input-to-state and state-to-state transitions, they propose the convolutional LSTM (ConvLSTM) and use it to build an end-to-end trainable model for the precipitation nowcasting problem. Experiments show that the ConvLSTM network captures spatio-temporal correlations better and consistently outperforms FC-LSTM and the state-of-the-art operational ROVER algorithm for precipitation nowcasting.

In [8], Souto et al. use a ConvLSTM architecture as a spatio-temporal ensemble approach. The channels in the convolution operator are used to input different physical weather models. In this way the ConvLSTM encodes the spatial information which are subsequently learned by the recurrent structures of the network. The results show that ConvLSTM achieves superior improvements when compared to traditionally used ensemble techniq ues such as BMA [9].

As a matter of fact, there are a plethora of opportunities to be investigated extending the initial results we have achieved in adopting Deep Neural networks to weather prediction. Linear and causal convolution operators (the latter also known as temporal convolution), for instance, have resulted in deep networks architectures that use convolutions to encode and decode time and space with greater precision. Raissi and colleagues [10] investigate the integration of physical laws described as a set of partial differential equations to the training process. By means of such integration, the training process is bound to obey the physical laws, an approach that has been dubbed as model teaching . Another area of interest is multimodal machine learning (MML) [11]. In MML, data from different representations are complementarily used in building models, including: images; textual data; quantitative dataetc... This can be extremely interesting in weather forecast as more data is captured from satellite images and sensors data to weather bulletins and predictive physical models.

1. Amy McGovern, Kimberly L. Elmore, David John Gagne II, Sue Ellen Haupt, Christopher D. Karstens, Ryan Lagerquist, Travis Smith, and John K. Williams. Using Artificial Intelligence to Improve Real-Time Decision-Making for High-Impact Weather. AMS. 2017
2. Qi Fu, Dan Niu, Zengliang Zang, Junhao Huang, and Li Diao. Multi-Stations’ Weather Prediction Based on Hybrid Model Using 1D CNN and Bi-LSTM. Chinese Control Conference (CCC), 3771-3775. 2019
3. Anthony Wimmers, Christopher VThomas Bolton and Laure Zanna . Applications of Deep Learning to Ocean Data Inference and Subgrid Parameterization. Journal of Advances in Modeling Earth Systems 11:1, 376-399. 2019
4. Anthony Wimmers and Christopher Velden , and Joshua H. Cossuth . Using Deep Learning to Estimate Tropical Cyclone Intensity from Satellite Passive Microwave Imagery. Monthly Weather Review 147:6, 2261-2282. 2019
5. S. Scher. Toward Data-Driven Weather and Climate Forecasting: Approximating a Simple General Circulation Model With Deep Learning. Geophysical Research Letters 45:22, 12,616-12,622. 2018
6. Brian S. Freeman, Graham Taylor, Bahram Gharabaghi and Jesse Thé. Forecasting air quality time series using deep learning. Journal of the Air & Waste Management Association, Volume 08, Issue 8 , Pages 866-886, 2018
7. Xingjian Shi, Zhourong Chen, and Hao Wang. Convolutional LSTM Network : A Machine Learning Approach for Precipitation Nowcasting. arXiv, pages 1–11, 2015.
8. Yania Molina Souto, Fábio Porto , Ana Maria de Carvalho Moura , Eduardo Bezerra : A Spatiotemporal Ensemble Approach to Rainfall Forecasting. IJCNN , pages 1-8. 2018
9. Raftery, A. E., Gneiting, T., Balabdaoui, F., and Polakowski, M.: Using Bayesian model averaging to calibrate forecast ensembles, Mon. Weather Rev., 133, 1155–1174, 2005.
10.Raissi, M., Predikaris, P., Karniadakis, G.E., Physics-informed neural network: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, pp.686-707, 2019.
11.Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency, MultiModal Machine Learning: A Survey and Taxonomy, https://arxiv.org/pdf/1705.09406.pdf

Primary author

Dr Yania Molina Souto (LNCC)


Dr Fabio Porto (LNCC)

Presentation Materials

There are no materials yet.