How Deep Learning Could Predict Weather Events
2 likes1,434 views
This document discusses using deep learning models to predict weather events from numerical model data and satellite images. It describes three models: DeepRain predicts precipitation from weather radar images using convolutional LSTMs; DeepTC predicts tropical cyclone trajectories from numerical model data like temperature and pressure using convolutional LSTMs; and GlobeNet predicts typhoon tracks from satellite images and an autoencoder. The models show improved prediction accuracy over traditional methods, demonstrating the potential of deep learning for weather forecasting.
How Deep Learning Could Predict Weather Events
- 1. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI How Deep Learning Could Predict Weather Events Seongchan Kim, Ph.D. / Seunkyun Hong On behalf of Sa-Kwang Song, Ph.D. Research Data Platform Center GTC 2018
- 2. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 2 Outline • Introduction: Vision vs. Meteorology tasks • Deep Learning for Weather Prediction • DeepRain: Precipitation Prediction of & Next Step of Images • DeepTC: Tropical Cyclone Trajectory Prediction using Numerical Model Data • GlobeNet: Tropical Cyclone Trajectory Prediction using Satellite Images & AutoEncoder • Conclusion
- 3. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 3 Outline • Introduction: Vision vs. Meteorology tasks • Deep Learning for Weather Prediction • DeepRain: Precipitation Prediction of & Next Step of Images • DeepTC: Tropical Cyclone Trajectory Prediction using Numerical Model Data • GlobeNet: Tropical Cyclone Trajectory Prediction using Satellite Images & AutoEncoder • Conclusion
- 4. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 4 Vision vs. Meteorology Tasks Typhoon Classification Typhoon Location Weather/Climate Event Detection Or Segmentation https://chaosmail.github.io/deeplearning/2016/10/22/intro-to-deep-learning-for-computer-vision/
- 5. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 5 Similarities vs Differences Phrabat 2017(CI 2017) • Similarities • Tasks are similar: • Classification, Localization, Detection, Segmentation • Clustering, Regression • Representation Learning • Differences • Unique attributes of Weather Data • Multi-channel/Multi-variate • Different Spatio-temporal scales • Double precision floating point • Underlying statistics are likely different • Large amount of data compared to general Vision tasks
- 6. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 6 Challenge: Multi-Variate Data COMS Satellite Himawari 8 satellite WRF Result
- 7. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 7 Outline • Introduction: Vision vs. Meteorology tasks • Deep Learning for Weather Prediction • DeepRain: Prediction of Precipitation & Next Step of Images • DeepTC: GPU-Accelerated Trajectory Prediction for Tropical Cyclone using Traditional Numerical Model Data • GlobeNet: Typhoon Track Prediction & Autoencoder • Conclusion
- 8. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI Numerical Model Results Satellite Radar BAIPAS Image/Sensor Data Prediction Model based on DL Convolutional Neural Networks Recurrent Neural Networks Sensors Typhoon Track Surge LevelFlood Level 8 Prediction of Weather Events Bigdata & AI based Prediction and Analysis Platform Disaster Prediction
- 9. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 9 DeepRain: Precipitation Prediction Weather Radar Center http://radar.kma.go.kr/lecture/radar/dataflow.doMethods used to measure precipitation • Weather Radar Data • refers to data represented by a radar image that is composed using the moving speed, direction, and strength of a signal transmitted by a radar transmitter into the atmosphere and received after it has collided with water vapor or the like.
- 10. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 10 DeepRain: Precipitation Prediction 6 mins * 15 = 90 mins
- 11. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI Shenzhen Meteorological Bureau-Alibab, “Short-Term Quantitative Precipitation Forecasting,” https://tianchi.aliyun.com/competition/information.htm?spm=5176.100067.5678.2.jsxLyX&raceId=231596. 11 DeepRain: Radar Data • Research-use data by Shenzhen Meteorological Agency • Modeling specific areas in Shenzhen as a grid pattern(101*101km2) • Radar reflection values: 101*101 numerical values (dBZ) of representing each cell • Precipitation amount: Total amount of rainfall in target site (50*50 area from center) • Normalization, Anonymization 101*101km2
- 12. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI (101, 101, 4) Conv LSTM h 0 Conv LSTM Conv LSTM …… 𝐹𝑖𝑛𝑎𝑙 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 _X _X_𝑖𝑠𝑡𝑎𝑡𝑒 _𝐿𝑆𝑇𝑀_𝑂 1 2 _O _O _O _𝐿𝑆𝑇𝑀_𝑂 _𝐿𝑆𝑇𝑀_𝑂 _𝐿𝑆𝑇𝑀_𝑆 _𝐿𝑆𝑇𝑀_𝑆 _𝐿𝑆𝑇𝑀_𝑆 _X15 𝐼𝑛𝑝𝑢𝑡 𝐼𝑛𝑝𝑢𝑡 𝐼𝑛𝑝𝑢𝑡 (101, 101, 4) (101, 101, 4) 12 DeepRain: Precipitation Prediction http://radar.kma.go.kr/lecture/radar/dataflow.do Height 1 Height 2 Height 3 Height 4
- 13. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 13 DeepRain: Experimental Results 1 3 5 7 9 11 13 15 17 19 21 23 25 27 0 5 10 15 20 25 30 Epoch Loss(RMSE) FC-LSTM(GDO,0.001) FC-LSTM(Adam,0.001) convLSTM(Adam, 0.001) convLSTM(Adam,0.001,2-stacked) • convLSTM shows better learning performance than FC-LSTM • Test result with Testset • Epoch 5 • With two-stacked we achieved 23.0% performance increase than LR. • Furthermore, it is lower 21.7% than FC-LSTM • Because FC-LSTM lost spatial information Model RMSE Drop Ratio Linear Regression 14.69 - DeepRain: FC-LSTM 14.46 1.6% DeepRain: convLSTM 11.51 21.6% DeepRain: convLSTM(2-stacked) 11.31 23.0%
- 14. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 14 Seq-to-Seq DeepRain: Prediction of Next Seq. Image The last N input images as input sequence Prediction output sequence tt-1t-2t-3t-4t-5t-6t-7t-8t-9 t+1 t+2 t+3 t+4 t+5
- 15. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 15 Seq-to-Seq DeepRain: Conv2Deconv RNN/LSTM • Model: Sequence-to-Sequence Model • Input → encoder • decoder Output
- 16. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 16 DeepRain: Deep Learning Models • Input and state transformation models 1) Vanilla fully-connected models • Vanilla FullyConnected-RNN • Vanilla FullyConnected-LSTM 2) Convolution to Deconvolution (Conv2Deconv) models • Convoluted RNN’s states (Conv2Deconv-ConvRNN) • Convoluted LSTM’s states and memory cells(Conv2Deconv-ConvLSTM) 3) Full-size convolution(Input images are fed directly to the convolutional networks) • Convoluted RNN’s states (Fullsize-ConvRNN) • Convoluted LSTM’s states and memory cells (Fullsize-ConvLSTM)
- 17. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 17 DeepRain: Results
- 18. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 18 DeepRain: Conv2Deconv-ConvLSTM → Note: predicted images are blurred - Testing: mean-cost=13.16, mean-rmse=22.42 - Testing: best-cost=6.71, best-rmse=10.73 The last 10 input images Model Predicted images Ground truth images
- 19. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 19 Outline • Introduction: Vision vs. Meteorology tasks • Deep Learning for Weather Prediction • DeepRain: Prediction of Precipitation & Next Step of Images • DeepTC: Tropical Cyclone Trajectory Prediction using Traditional Numerical Model Data • GlobeNet: Typhoon Track Prediction & Autoencoder • Conclusion
- 20. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI • Today, meteorologists rely on numerical models to predict wind speed, precipitation, air pressure and other factors that indicate the path and intensity of a hurricane over its lifetime. • WRF (Weather Research and Forecasting model), MPAS (Model for Prediction Across Scales), UM (Unified Model), and CAM5 (Community Atmosphere Model ver 5.0) 20 Motivation Bolaben, Wind (UVW) Bolaben, DBz Bolaben, Flow Visualization (Vapor)
- 21. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI • Predicting typhoon track predict using data from a numerical model (WRF) simulation results with Deep Neural Network • ConvLSTM: Convolutional LSTM (shi et al. 2015) that learn spatial features of input data • Ensemble-like techniques: learning from five differently conditioned WRF results Ensemble Forecasting 21 Purpose
- 22. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI • 10 typhoons drifting in close proximity to the Korea peninsula No . ID (YYNN) Name Period Duration Simulation Durations Num. of Simulation 1 0215 루사(RUSA) 2002.08.23. 09:00 ~ 2002.09.01. 15:00 9 days 3 hours 4 days 6 hours 25 2 0314 매미(MAEMI) 2003.09.06. 15:00 ~ 2003.09.14. 06:00 3 days 15 hours 3 days 18 hours 15 3 0415 메기(MEGI) 2004.08.16. 15:00 ~ 2004.08.20. 18:00 4 days 3 hours 3 days 18 hours 15 4 0603 에위니아(EWINIAR) 2006.07.01. 03:00 ~ 2006.07.10. 22:00 9 days 19 hours 6 days 18 hours 27 5 1004 뎬무(DIANMU) 2010.08.08. 21:00 ~ 2010.08.12. 15:00 3 days 18 hours 3 days 6 hours 13 6 1214 덴빈(TEMBIN) 2012.08.19. 09:00 ~ 2012.08.31. 00:00 11 days 15 hours 7 days 18 hours 35 7 1215 볼라벤(BOLAVEN) 2012.08.20. 15:00 ~ 2012.08.29. 06:00 8 days 15 hours 6 days 18 hours 27 8 1216 산바(SANBA) 2012.09.11. 09:00 ~ 2012.09.18. 09:00 7 days 4 days 18 hours 19 9 1509 찬홈(CHAN-HOM) 2015.06.30. 21:00 ~ 2015.07.13. 06:00 12 days 9 hours 5 days 18 hours 23 10 1618 차바(CHABA) 2016.09.28. 03:00 ~ 2016.10.02. 15:00 5 days 12 hours 3 days 6 hours 13 Sum 212 22 Data
- 23. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI Name Description Dimension Data Dim. T perturbation potential temperature (5, 29, 192, 192) 4 dimension (time, height, width, length) P perturbation pressure QVAPOR Water vapor mixing ratio SST Skin sea surface temperature (5, 1, 192, 192) OLR TOA outgoing long wave • Selected variables • 5 variables deemed to be the most significant for cyclone tracking • four-dimensional • Time, height, width, and length • represented in 3-D spatial grids 23 Data
- 24. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI • Simulation Setting • 10 typhoons, 1060 predictions • For example, Rusa, 4 days 6 hours • Simulation every 6 hour cyclically • Five different initial conditions (models) • 212 * 5 = 1060 • Whole WRF file size: 2.2 TB • Data split # of instances TF Record Size Training set 600 150 Gb Validation set 200 50 Gb Testing set 200 50 Gb 24 Data
- 25. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI WRF data Conv2D LSTM h0 Conv2D LSTM Conv2D LSTM …… _X _X _𝐻 _𝑖𝑠𝑡𝑎𝑡𝑒 _𝐿𝑆𝑇𝑀_𝑂 0 1 _𝐿𝑆𝑇𝑀_𝑂 _𝐿𝑆𝑇𝑀_𝑂 _𝐿𝑆𝑇𝑀_𝑆 _𝐿𝑆𝑇𝑀_𝑆 _𝐿𝑆𝑇𝑀_𝑆 _𝐻 _𝐻 _X4 Prediction every 6 hours for next 24 hours 00:00 06:00 24:00 Best track (Ground Truth) (lat0, long0) _O0 _O1 _O4 (lat1, long2) (lat4, long4) Time 𝑘𝑒𝑟𝑛𝑒𝑙 = 3,3 𝑓𝑖𝑙𝑡𝑒𝑟𝑠: 12 𝑠ℎ𝑎𝑝𝑒 = (239,279) 𝑐ℎ𝑎𝑛𝑛𝑒𝑙 = 89 25 Model Prediction (lat0, long0) (lat1, long2) (lat4, long4)
- 26. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI • Learning 0 5 10 15 20 25 30 35 40 45 50 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 RMSE Epoch 26 Model RMSE FC-LSTM 147.24 DeepTC: ConvLSTM 13.88 Result - K40m(x2) reduced 54% of training time against CPU, - While P100(x2) reduced 56% of training time against K40m for 100 epoch • Test Result
- 27. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 27 Outline • Introduction: Vision vs. Meteorology tasks • Deep Learning for Weather Prediction • DeepRain: Precipitation Prediction of & Next Step of Images • DeepTC: Tropical Cyclone Prediction using Numerical Model Data • GlobeNet: Typhoon Track Prediction using Satellite Images & AutoEncoder • Conclusion
- 28. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 28 GlobeNet: Typhoon Track Prediction Observation Data COMS-1 MI (KMA) Himawari-8 images (16ch) Prediction of Typhoon Track Deep Learning Prediction Models HIMAWARI-8 AHI (JMA) Convolutional Neural Nets (CNNs) Long short Term Memory (LSTM) Index Conf. Lat Long hPa NSWE 1 0.9999 27.9748 125.9390 972 N 2 0.9361 13.3122 83.7975 1013 NS 3 0 0 0 0 X 4 0 0 0 0 X 5 0 0 0 0 X 6 0 0 0 0 X Prediction Results
- 29. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 29 GlobeNet: Data • Input: Satellite Data • COMS-1 Satellite: 4 channel(IR1, IR2, SWIR, WV) • Period: 2011.04~2016.12 • Label: Best Track Data • RSMC-Tokyo Best Track • 6 years: 2011~2016 (152 typhoon cases) Typhoon Track COMS-1
- 30. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 30 GlobeNet: Typhoon Location Detection Input Satellite Images Typhoon Location Tracking Conv2D based Model Inception based Model
- 31. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 31 GlobeNet: Experiments
- 32. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 32 GlobeNet: Test Error Distance(km) - Average Error(distance in km): 74.53 km - Eye of typhoon: 30-65km in diameter 30-65km
- 33. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 33 GlobeNet: Autoencoder • To build a pre-trained model • COMS-1 • 4 channel/IR1, IR2, SWIR, WV • Size: 16.55 TB (2011.04~2017, 187,083 scenes) • HIMAWARI-8 AHI • On-going • Size: hundreds of TBs • Conv-Deconv Neural Network Encoder (Convolution Step) Decoder (Deconvolution Step) Latent Vector
- 34. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 34 GlobeNet: Encoding Step Convolution Step Layer 1 Layer2 Layer3 Layer4 Layer5 Input
- 35. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 35 GlobeNet: Decoding Step Deconvolution Step Layer4 Layer3 Layer2 Layer1Layer5 Prediction
- 36. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 36 GlobeNet: Autoencoder Input Images Output Images Conv-Deconv Model Skip-connection Model
- 37. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 37 GlobeNet with Skip-connection Decoder Output Images
- 38. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 38 Recent Convolutional AE without Skip-connection Output Images Input Images
- 39. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 39 Application of Skip-connection Conv AE: PSIque • Complex Deep Seq2Seq Autoencoder model based on Memory Network structure • Unified model structure compatible with various RNNCells (e.g. LSTM, GRU, ConvLSTM) • Encoder(Conv)->Encoder(LSTM)->Decoder(LSTM)->Decoder(DeConv) ⊕ SkipConx Encoder (Conv): 5-layers ConvNet for Feature Extraction Encoder (LSTM): Encoding Spatiotemporal Changes Through Time Decoder (LSTM): Decoding Spatiotemporal Changes Through Time Decoder (DeConv): 5-layers DeConvNet for Image Restoration SkipConx: Selective, Symmetric bypassing skip connection from Encoder to Decoder Latent Info (Output/State): Condensed Tensor representing Spatiotemporal Transition
- 40. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 40 Outline • Introduction: Vision vs. Meteorology tasks • Deep Learning for Weather Prediction • GlobeNet: Typhoon Track Prediction & Autoencoder • DeepRain: Prediction of Precipitation & Next Step of Images • DeepTC: GPU-Accelerated Trajectory Prediction for Tropical Cyclone using Traditional Numerical Model Data • Conclusion
- 41. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI 41 Conclusion • Deep Learning can be applied effectively to understand meteorological phenomenon • DeepRain does using radar reflectivity data • DeepTC does using numerical model data • GlobeNet predicts meteorological phenomenon by analyzing satellite images • Distributed Deep Learning Platform is necessary. • Number of associated challenges. • Open to collaboration!!!
- 42. NVIDIA GPU Technology Conference 2018Copyright© 2018 Sa-Kwang Song, KISTI