[20240506_LabSeminar_Huy]Conditional Local Convolution for Spatio-Temporal Meteorological Forecasting.pptx

Conditional Local Convolution for Spatio-Temporal Meteorological Forecasting

1. Quang-Huy Tran Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: [email protected] 2024-05-06 Conditional Local Convolution for Spatio -Temporal Meteorological Forecasting Haitao Lin et al. AAAI’36: 2022 Conference on Artificial Intelligence

2. OUTLINE MOTIVATION INTRODUCTION METHODOLOGY EXPERIMENT & RESULT CONCLUSION

3. MOTIVATION Graph Neural Network (GNN) has been superior solutions to model spatio –temporal dependency in traffic forecasting . However, there is not much work using GNN for meteorology forecastin g . Meteorology overview Challenges: irregular sampling of meteorological signals usually disables the classical (CNNs) because s ignals are usually acquired from irregularly distributed sensors, and the manifolds from which signals are sampled are usually non-planar. High temporal and spatial dependency makes it hard to model the dynamics. For instance, different landforms show totally distinct wind flow or temperature transferring patterns and extreme climate incidents like El Nino often cause non-stationarity for prediction.

4. INTRODUCTION Propose conditional local kernel. embed it in a graph-convolution-based recurrent network. convolution is performed on the local space of each node, where kernel based on the assumption: smoothness of location characterized patterns. establishing the spatio -temporal model with the proposed graph convolution which achieves state-of-the-art performance in weather forecasting tasks. conducting further analysis on learned local pattern visualization, framework choice, local space and map choice and ablation.ß

5. METHODOLOGY PROBLEM SETTING Signals observed at time t of the nodes on G are denoted by For the forecasting tasks, our goal is to learn a function for approximating the true mapping of historical observed signals to the future T signals:   Given N correlated signals located on the sphere manifold at time t, represent the signals as a (directed) graph G = (V, E, A). meaning that it records the position of N nodes, which satisfies  

6. METHODOLOGY Overall Architecture

7. METHODOLOGY Local Convolution on Sphere The geodesics and induced distance on sphere are important to both defining the neighborhood of a node Source: Google Distance function is usually called great-circle distance on sphere. K-nearest neighbors algorithm to construct the graph struture is conducted based on spherical distance

8. METHODOLOGY Local Convolution on Sphere Given a center node x i in Euclidean space: General GCN Given a center node x i in Sphere space:

9. METHODOLOGY Orientation-Preserving Local Regions Preserves the relative orientation on geographic graticules on the earth surface, which has explicitly geophysical meaning in meteorology. Logarithmic maps distort the relative position in orientation on graticules. Define cylindrical-tangent space and horizon maps to construct local spaces and to map neighbors into them:

10. METHODOLOGY Conditional Local Convolution Goal: Kernel conditional on centers for location-characterized: In the local regions of different center nodes, the meteorological patterns governed by convolution kernel differ. Smoothness of local patterns.: Patterns are broadly similar when the center nodes are close in spatial distance.

11. METHODOLOGY Conditional Local Convolution Reweighting for irregular spatial distribution: The kernel is shared by different local spaces where the neighbors’ spatial distribution is distinct. However, the nodes are discrete and irregularly distributed on the sphere. To reweight the convolution kernel for each x j , consider both their angle and distance scales:

12. METHODOLOGY Temporal Dependency Modeling GRU cell is applied after smooth conditional local kernel in the case of irregular spatial distribution

13. EXPERIMENT AND RESULT EXPERIMENT Measurement: Mean Absolute Error (MAE). Root Mean Square Error (RMSE). Mean Absolute Percentage Error (MAPE). Dataset: WeatherBench 2048 nodes on the earth sphere . four-hour wise weather forecasting tasks including temperature, cloud cover, humidity and surface wind component. Task: set input time length as 12 and forecasting length as 12 for the four datasets..

14. EXPERIMENT AND RESULT EXPERIMENT - BASELINE [1] Rozemberczki , B., Scherer, P., He, Y., Panagopoulos, G., Riedel, A., Astefanoaei , M., ... & Sarkar, R. (2021, October). Pytorch geometric temporal: Spatiotemporal signal processing with neural machine learning models. In Proceedings of the 30th ACM international conference on information & knowledge management (pp. 4564-4573).. 2061-2064).. [2] Yu, B., Yin, H., & Zhu, Z. (2017). Spatio -temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875.. [3] Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019, July). Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence (Vol. 33, No. 01, pp. 922-929) [4] Zhao, L., Song, Y., Zhang, C., Liu, Y., Wang, P., Lin, T., ... & Li, H. (2019). T- gcn : A temporal graph convolutional network for traffic prediction. IEEE transactions on intelligent transportation systems, 21(9), 3848-3858. [5] Seo , Y., Defferrard , M., Vandergheynst , P., & Bresson, X. (2018). Structured sequence modeling with graph convolutional recurrent networks. In Neural Information Processing: 25th International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, Proceedings, Part I 25 (pp. 362-373). Springer International Publishing. [6] Li, Y., Yu, R., Shahabi , C., & Liu, Y. (2017). Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926. [7] Bai, L., Yao, L., Li, C., Wang, X., & Wang, C. (2020). Adaptive graph convolutional recurrent network for traffic forecasting. Advances in neural information processing systems, 33, 17804-17815.

15. EXPERIMENT AND RESULT RESULT – Overall Perfor mance

16. EXPERIMENT AND RESULT RESULT – Advantages of Horizon Map

17. EXPERIMENT AND RESULT RESULT – Ablation study

18. CONCLUSION Proposed a local conditional convolution to capture and imitate the meteorological flows of local patterns whole sphere . Smoothness of location-characterized patterns assumptions. MLP and reweighting terms with continuous relative positions of neighbors and center as inputs are employed as convolution kernel to handle uneven distribution of nodes. Further analysis reveals two existing problems of proposal method: the over-smoothness of the learned local patterns and and instability of the training process.