20250821 meeting

Introduction

This experiment is the same as 20250819 meeting, with the only change being a modification to the training dataset. The goal is to attempt simplifying the spatiotemporal data (#simulation01).

Configuration File

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

# ----------------------------------------------------
# 溫度變化模型設定
# ----------------------------------------------------
temperature_model:
  seasonal_amplitude: 5            # 季節變化振幅 (°C)
                                     # T_season(t) = A_season * sin(2π * t / 365)
  daily_amplitude: 5                 # 日變化振幅 (°C)
                                     # T_day(h) = A_day * sin(2π * h / 24 - π/2)
  weekly_amplitude: 0                # 每週變化振幅 (°C)
                                     # T_week(t) = A_week * sin(2π * t / cycle_days)
  noise_sigma: 1.0                   # 高斯雜訊標準差 (°C)，ε ~ N(0, σ²)
  spatial_noise_correlation: True    # 啟用空間相關噪聲強度
  temporal_autocorrelation: 0.8      # 時間自相關 (0-1) AR(1) 模型
                                     # T(t) = ρ * T(t-1) + (1 - ρ) * T_base + ε

# ----------------------------------------------------
# 調變（modulation）設定
# ----------------------------------------------------
# 僅在 spatial_noise_correlation 啟用時生效
modulation:
  spatial_field:
    corr_length_km: 10.0             # 空間相關長度（公里），控制空間隨機場的平滑程度，數值越大越平滑
    sigma_C: 0.8                     # 空間隨機場的標準差，控制波動強度
    temporal_modulation: 0.5         # 時間調變強度（0-1），控制空間隨機場隨季節變化的調幅

SSSD + autoFRK

autoFRK inference time: 2.17848 hours (CPU parallel computation, 23 cores).

TSMixer + autoFRK

TSMixer inference time: 1:32:11.011375 (CPU parallel computation, 24 cores).

autoFRK inference time: 3.103498 hours (CPU parallel computation, 23 cores).

RegressionEnsemble + autoFRK

RegressionEnsemble inference time: 0:01:08.390425 (CPU parallel computation, 12 cores).

autoFRK inference time: 2.511842 hours (CPU parallel computation, 23 cores).

RegressionEnsemble (LSTM) + autoFRK

RegressionEnsemble inference time: 0:22:01.921387 (CPU parallel computation, 12 cores).

autoFRK inference time: 3.103747 hours (CPU parallel computation, 23 cores).

Conclusion

Epilogue

Environment

• Local Operating System: Windows 11 24H2
  • Programming Language: Python 3.12.9
• Computing Platform: National Center for High-Performance Computing (NCHC) – Taiwan AI Cloud
  • Operating System: Ubuntu
  • Miniconda
  • GPU: NVIDIA Tesla V100 32GB GPU
  • CUDA 12.8 driver
  • Programming Language: Python 3.10.16 for Linux

Further Learning

• I tested this project on the GitHub repository.

References

• Global Modeling and Assimilation Office (GMAO). (2015). MERRA-2 tavg1_2d_flx_Nx: 2d,1-Hourly,Time-Averaged,Single-Level,Assimilation,Surface Flux Diagnostics (Version 5.12.4) [Dataset]. Goddard Earth Sciences Data and Information Services Center (GES DISC). Retrieved from https://doi.org/10.5067/7MCPBJ41Y0K6

• Juan Lopez Alcaraz, Nils Strodthoff (2022). Diffusion-based time series imputation and forecasting with structured state space models. Transactions on Machine Learning Research. Retrieved from https://openreview.net/forum?id=hHiIbk7ApW

• SSSD (2022). GitHub. Retrieved from https://github.com/AI4HealthUOL/SSSD

• SSSD_CP (2024). GitHub. Retrieved from https://github.com/egpivo/SSSD_CP

• Unit8 SA (n.d.). Time Series Made Easy in Python. Darts. Retrieved from https://unit8co.github.io/darts/index.html

• darts (2025). GitHub. Retrieved from https://github.com/unit8co/darts

• Tzeng, S., & Huang, H. C. (2018). Resolution Adaptive Fixed Rank Kriging. Technometrics, 60(2), 198–208. Retrieved from https://doi.org/10.1080/00401706.2017.1345701

• autoFRK (2024). GitHub. Retrieved from https://github.com/egpivo/autoFRK

• Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O. Arik, and Tomas Pfister. (2023). TSMixer: An all-MLP architecture for time series forecasting. arXiv. Retrieved from https://arxiv.org/abs/2303.06053