1150317 meeting

收尾工作 v2.0

發佈於 2026-03-15 更新於 2026-03-17 約 3655 字預計閱讀 8 分鐘次閱讀

本次實驗原先欲複測 Weather2K 、 NASA GES DISC MERRA-2 在與上次實驗相同資料範圍下，但不進行時間重塑的實驗結果。但因使用國網中心時，發現已超越單一 GPU 之 vRAM 限制（過大），故本次將對所有實驗進行參數調整，縮小 Residual 層數並重測所有實驗，已達可用之實驗組與對照組。

本次實驗順帶重構以往所有用於實驗之程式碼（SSSDS4 + AFRK 程式碼未更動），同時將兩資料集實驗範圍統一在半年區間，時序預測設為其 10 % ，約 18 天；未觀測地點為整體地點之 20 % 。

在研究為何原 SSSD 程式碼無法直接於多 GPU 上執行運算時，發現其中多步接直接指定計算裝置為 CDUA ，即

1
.device("cuda")

此舉造成使用 SSSD 程式碼時，被限定在僅可使用 NVIDIA GPU ，從而無法使用 CPU 、 MPS 、 TPU 等裝置運行，且硬性編碼 .device("cuda") 會造成只能於第一個 GPU ，即 "cuda:0" 上運行，使多 GPU 間無法進行資料傳遞。因整體進行改寫曠日廢時，故目前暫無打算進行改寫，僅以同上所述，縮減部分模型層數以進行實驗。

實驗介紹

本次進行以下實驗，詳細實驗設定參考下節。

控制變因	實驗 1	實驗 2	實驗 3	實驗 4
迭代次數	4,000	4,000	4,000	4,000
訓練策略	$SSSD^{S4 + AFRK}$	$SSSD^{S4}$	$SSSD^{S4 + AFRK}$	$SSSD^{S4}$
時間重塑	false	false	true	true
併入時間步 $p$	—	—	8 / 24	8 / 24
Input Channels	4 / 1	4 / 1	32 / 24	32 / 24
S4 Max Seq. Length	1,176 / 4,608	1,176 / 4,608	147 / 192	147 / 192
Missing $k$	88 / 144	88 / 144	11 / 6	11 / 6

實驗命名依上表不同可區分如下：

實驗 1
- Weather2k-1var-S4+AFRK
- MERRA2-1var-S4+AFRK
實驗 2
- Weather2k-1var-S4
- MERRA2-1var-S4
實驗 3
- Weather2k-S4+AFRK
- MERRA2-S4+AFRK
實驗 4
- Weather2k-S4
- MERRA2-S4

實驗設定

本次實驗使用設定如下所示，僅輸入出通道（channels）、時間序列長度（s4 max sequence length）、缺失值（missing $k$）、是否啟用 AFRK（enable spatial training）及目錄與路徑會依實驗內容不同而有更動。以下呈現 Weather2k-S4+AFRK 設定。

model.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
wavenet:
  # WaveNet model parameters
  input_channels:  32  # Number of input channels
  output_channels: 32  # Number of output channels
  residual_layers: 20  # Number of residual layers
  residual_channels: 20  # Number of channels in residual blocks
  skip_channels: 20  # Number of channels in skip connections

  # Diffusion step embedding dimensions
  diffusion_step_embed_dim_input:  64  # Input dimension
  diffusion_step_embed_dim_hidden: 128  # Middle dimension
  diffusion_step_embed_dim_output: 128  # Output dimension

  # Structured State Spaces sequence model (S4) configurations
  s4_max_sequence_length: 166  # Maximum sequence length
  s4_state_dim: 128  # State dimension
  s4_dropout: 0.2  # Dropout rate
  s4_bidirectional: true  # Whether to use bidirectional layers
  s4_use_layer_norm: true  # Whether to use layer normalization

diffusion:
  # Diffusion model parameters
  T: 100  # Number of diffusion steps
  beta_0: 0.0001  # Initial beta value
  beta_T: 0.01  # Final beta value

AFRK:
# AutoFRK model parameters, for more details, please refer to https://pypi.org/project/autoFRK/
  method: "fast"  # autoFRK method to use (e.g., "fast")
  tps_method: "rectangular"  # autoFRK's TPS method to use (e.g., "rectangular")

training.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
# Training configuration
batch_size: 100  # Batch size
output_directory: "/home/u6025091/SSSD_CP/results/Weather2k-S4+AFRK"  # Output directory for checkpoints and logs
ckpt_iter: "max"  # Checkpoint mode (max or min)
iters_per_ckpt: 500  # Checkpoint frequency (number of epochs)
iters_per_logging: 200  # Log frequency (number of iterations)
n_iters: 4000  # Maximum number of iterations
learning_rate: 0.0005  # Learning rate

# Additional training settings
only_generate_missing: true  # Generate missing values only
use_model: 2  # Model to use for training
masking: "forecast"  # Masking strategy for missing values
missing_k: 18  # Number of missing values

# Data paths
data:
  train_path: "/home/u6025091/SSSD_CP/datasets/Weather2k/data_train_known_real.npy"  # Path to training data

# autoFRK config
enable_spatial_training: true  # Enable spatial training step
location_path: "/home/u6025091/SSSD_CP/datasets/Weather2k/stations_known_locations.npy"  # Path to known locations

inference.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
# Inference configuration
batch_size: 100  # Batch size for inference
output_directory: "/home/u6025091/SSSD_CP/results/Weather2k-S4+AFRK/inference"  # Output directory for inference results
ckpt_path: "/home/u6025091/SSSD_CP/results/Weather2k-S4+AFRK"  # Path to checkpoint for inference
trials: 1 # Replications

# Additional training settings
only_generate_missing: true  # Generate missing values only
use_model: 2  # Model to use for training
masking: "forecast"  # Masking strategy for missing values  # inference mode need to fix in the code, or using other masking strategy will be failed
missing_k: 18  # Number of missing values

# Data paths
data:
  test_path: "/home/u6025091/SSSD_CP/datasets/Weather2k/data_train_known_real.npy"  # Path to test data

# autoFRK config
enable_spatial_inference: true  # Enable spatial prediction step
enable_spatial_normalization: true  # Enable spatial normalization before and after autoFRK
known_location_path: "/home/u6025091/SSSD_CP/datasets/Weather2k/stations_known_locations.npy"  # Path to known locations
unknown_location_path: "/home/u6025091/SSSD_CP/datasets/Weather2k/stations_unknown_locations.npy"  # Path to unknown locations

選取的實驗地點如下所示。

https://raw.githubusercontent.com/Josh-test-lab/website-assets-repository/refs/heads/main/posts/1150317%20meeting/Weather2k.png — Weather2k 地點分布。

https://raw.githubusercontent.com/Josh-test-lab/website-assets-repository/refs/heads/main/posts/1150317%20meeting/MERRA2.png — MERRA-2 地點分布。

下節呈現各實驗結果。

實驗 1

Weather2k-1var-S4+AFRK

訓練花費時間：35h 25m 14.82s（註：此實驗非使用單一機器，而是與 Weather2k-1var-S4 、 Weather2k-S4 、 Weather2k-S4+AFRK 共用機器進行訓練）
推論階段 AFRK 之 MRTS 基底數：382

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	1.494639e+01	1.160173e+01	2.828929e+01	7.756372e+01	7.868970e+01	7.307184e+01	8.156561e+00	4.327126e+00	2.343335e+01
RMSPE	3.866056e+00	3.406131e+00	5.318767e+00	8.807027e+00	8.870722e+00	8.548207e+00	2.855969e+00	2.080175e+00	4.840800e+00
MSPE%	1.380822e+10	3.299012e+09	5.573267e+10	2.246783e+10	2.608424e+10	8.040864e+09	1.286923e+10	8.283242e+08	6.090407e+10
RMSPE%	1.175084e+05	5.743703e+04	2.360777e+05	1.498927e+05	1.615062e+05	8.967086e+04	1.134426e+05	2.878062e+04	2.467875e+05
MAPE	2.332151e+00	2.120980e+00	3.174576e+00	5.441194e+00	5.450313e+00	5.404813e+00	1.995026e+00	1.759968e+00	2.932743e+00
MAPE%	7.822800e+08	5.627048e+08	1.658233e+09	5.047446e+08	5.480241e+08	3.320892e+08	8.123742e+08	5.642966e+08	1.802031e+09

MERRA2-1var-S4+AFRK

訓練花費時間：20h 24m 12.59s
推論階段 AFRK 之 MRTS 基底數：301

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	160.204191	160.390768	159.457883	1628.647842	1631.785612	1616.096765	0.975361	0.841929	1.509089
RMSPE	12.657179	12.664548	12.627663	40.356509	40.395366	40.200706	0.987604	0.917567	1.228450
MSPE%	0.611571	0.612622	0.607367	6.222452	6.237316	6.162993	0.003162	0.002715	0.004950
RMSPE%	0.782030	0.782702	0.779338	2.494484	2.497462	2.482538	0.056233	0.052108	0.070357
MAPE	3.875349	3.801459	4.170910	34.341577	34.369929	34.228172	0.571782	0.486805	0.911689
MAPE%	0.014226	0.013981	0.015206	0.128176	0.128351	0.127477	0.001869	0.001579	0.003032

實驗 2

Weather2k-1var-S4

訓練花費時間：10h 3m 41.84s（註：此實驗非使用單一機器，而是與 Weather2k-1var-S4+AFRK 共用機器進行訓練）
推論階段 AFRK 之 MRTS 基底數：382

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	1.160546e+01	8.335466e+00	2.465048e+01	8.308004e+01	8.508390e+01	7.508604e+01	3.855207e+00	1.334625e-02	1.918156e+01
RMSPE	3.406679e+00	2.887121e+00	4.964925e+00	9.114825e+00	9.224093e+00	8.665220e+00	1.963468e+00	1.155260e-01	4.379676e+00
MSPE%	1.214263e+10	2.615445e+09	5.014950e+10	2.295537e+10	2.672051e+10	7.935082e+09	1.097017e+10	1.642926e+06	5.472697e+10
RMSPE%	1.101936e+05	5.114142e+04	2.239408e+05	1.515103e+05	1.634641e+05	8.907908e+04	1.047386e+05	1.281767e+03	2.339380e+05
MAPE	1.034689e+00	6.119608e-01	2.721080e+00	5.516763e+00	5.541266e+00	5.419014e+00	5.486807e-01	7.745780e-02	2.428533e+00
MAPE%	3.535854e+08	7.255939e+07	1.474684e+09	4.961357e+08	5.373860e+08	3.315754e+08	3.381282e+08	2.215650e+07	1.598635e+09

MERRA2-1var-S4

訓練花費時間：16h 48m 57.28s
推論階段 AFRK 之 MRTS 基底數：301

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	153.309263	153.324779	153.247198	1564.356985	1566.560730	1555.542007	0.304088	0.082327	1.191135
RMSPE	12.381812	12.382438	12.379305	39.551953	39.579802	39.440360	0.551442	0.286926	1.091391
MSPE%	0.587130	0.587618	0.585178	5.992495	6.004252	5.945470	0.001006	0.000272	0.003941
RMSPE%	0.766244	0.766563	0.764969	2.447957	2.450357	2.438333	0.031721	0.016505	0.062781
MAPE	3.651232	3.550918	4.052489	33.980246	33.997542	33.911060	0.362544	0.249477	0.814813
MAPE%	0.013523	0.013196	0.014834	0.127042	0.127182	0.126480	0.001214	0.000836	0.002728

實驗 3

Weather2k-S4+AFRK

訓練花費時間：（註：此實驗非使用單一機器，而是與 Weather2k-1var-S4+AFRK 共用機器進行訓練）
推論階段 AFRK 之 MRTS 基底數：

MERRA2-S4+AFRK

訓練花費時間：
推論階段 AFRK 之 MRTS 基底數：

實驗 4

Weather2k-S4

訓練花費時間：4h 31m 17.40s（註：此實驗非使用單一機器，而是與 Weather2k-1var-S4+AFRK 共用機器進行訓練）
推論階段 AFRK 之 MRTS 基底數：178

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	2.729977e+01	2.688188e+01	2.896683e+01	2.397620e+02	2.747339e+02	1.002486e+02	4.261693e+00	6.368540e-03	2.123748e+01
RMSPE	5.224918e+00	5.184774e+00	5.382084e+00	1.548425e+01	1.657510e+01	1.001242e+01	2.064387e+00	7.980313e-02	4.608414e+00
MSPE%	1.143239e+10	2.267235e+09	4.799498e+10	2.030899e+10	2.313886e+10	9.019811e+09	1.046987e+10	4.046914e+06	5.222121e+10
RMSPE%	1.069224e+05	4.761549e+04	2.190776e+05	1.425096e+05	1.521146e+05	9.497268e+04	1.023224e+05	2.011694e+03	2.285196e+05
MAPE	1.437579e+00	1.047657e+00	2.993092e+00	9.437545e+00	1.014651e+01	6.609263e+00	5.701122e-01	6.103477e-02	2.600977e+00
MAPE%	3.696454e+08	7.675813e+07	1.538062e+09	4.665294e+08	4.950431e+08	3.527794e+08	3.591399e+08	3.140192e+07	1.666586e+09

MERRA2-S4

訓練花費時間：2h 14m 26.41s
推論階段 AFRK 之 MRTS 基底數：111

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	183.060926	194.459146	137.468045	1864.618866	1987.368928	1373.618619	0.723318	0.047242	3.427622
RMSPE	13.530001	13.944861	11.724677	43.181233	44.579916	37.062361	0.850481	0.217352	1.851384
MSPE%	0.680973	0.722399	0.515269	6.939079	7.383059	5.163158	0.002383	0.000159	0.011281
RMSPE%	0.825211	0.849941	0.717822	2.634213	2.717179	2.272258	0.048819	0.012604	0.106213
MAPE	3.821658	3.698350	4.314889	35.220395	36.220564	31.219723	0.416976	0.171845	1.397498
MAPE%	0.013882	0.013486	0.015466	0.129019	0.132507	0.115069	0.001397	0.000580	0.004666

Bonus

前述實驗皆以半年時長作為實驗時間序列，但似乎下修模型 Residual 層數使得模型學習到的側爭效果成效不彰，尤其以 MERRA-2 實驗尤為糟糕。故於本次實驗間額外做兩個實驗，同樣以 MERRA-2 資料集作為測試資料（因為該資料集較大，同樣時間範圍內容易達到 vRAM 上限，以下實驗訓練過程中 vRAM 皆佔有 30 GB 以上），分別測試在降低輸入 batch_size 的情況下，下列實驗參數的表現：

實驗 5 - MERRA2-1var-S4+AFRK-low-batch

本實驗的訓練集與 MERRA2-1var-S4+AFRK 相同，但降低部分參數，同時提升模型層數到與前次 meeting 接近的參數設定，因 vRAM 關係無法完全一致，具體參數如下：

model.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
wavenet:
  # WaveNet model parameters
  input_channels:  1  # Number of input channels
  output_channels: 1  # Number of output channels
  residual_layers: 20  # Number of residual layers
  residual_channels: 36  # Number of channels in residual blocks
  skip_channels: 36  # Number of channels in skip connections

  # Diffusion step embedding dimensions
  diffusion_step_embed_dim_input:  128  # Input dimension
  diffusion_step_embed_dim_hidden: 256  # Middle dimension
  diffusion_step_embed_dim_output: 256  # Output dimension

  # Structured State Spaces sequence model (S4) configurations
  s4_max_sequence_length: 3984  # Maximum sequence length
  s4_state_dim: 128  # State dimension
  s4_dropout: 0.2  # Dropout rate
  s4_bidirectional: true  # Whether to use bidirectional layers
  s4_use_layer_norm: true  # Whether to use layer normalization

diffusion:
  # Diffusion model parameters
  T: 100  # Number of diffusion steps
  beta_0: 0.0001  # Initial beta value
  beta_T: 0.01  # Final beta value

AFRK:
# AutoFRK model parameters, for more details, please refer to https://pypi.org/project/autoFRK/
  method: "fast"  # autoFRK method to use (e.g., "fast")
  tps_method: "rectangular"  # autoFRK's TPS method to use (e.g., "rectangular")

training.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
# Training configuration
batch_size: 60  # Batch size
output_directory: "/home/u6025091/SSSD_CP/results/MERRA2-1var-S4+AFRK-lowbatch"  # Output directory for checkpoints and logs
ckpt_iter: "max"  # Checkpoint mode (max or min)
iters_per_ckpt: 500  # Checkpoint frequency (number of epochs)
iters_per_logging: 200  # Log frequency (number of iterations)
n_iters: 4000  # Maximum number of iterations
learning_rate: 0.0005  # Learning rate

# Additional training settings
only_generate_missing: true  # Generate missing values only
use_model: 2  # Model to use for training
masking: "forecast"  # Masking strategy for missing values
missing_k: 432  # Number of missing values

# Data paths
data:
  train_path: "/home/u6025091/SSSD_CP/datasets/MERRA2-1var/data_train_known_real.npy"  # Path to training data

# autoFRK config
enable_spatial_training: true  # Enable spatial training step
location_path: "/home/u6025091/SSSD_CP/datasets/MERRA2-1var/stations_known_locations.npy"  # Path to known locations

訓練花費時間：35h 11m 12.33s
推論階段 AFRK 之 MRTS 基底數：301

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	164.784290	164.888406	164.367824	1668.971215	1671.274799	1659.756880	1.679684	1.545304	2.217204
RMSPE	12.836833	12.840888	12.820602	40.853044	40.881228	40.740114	1.296026	1.243102	1.489028
MSPE%	0.632449	0.633265	0.629184	6.413790	6.426202	6.364142	0.005557	0.005116	0.007321
RMSPE%	0.795267	0.795780	0.793211	2.532546	2.534995	2.522725	0.074542	0.071523	0.085561
MAPE	4.193186	4.136441	4.420163	34.907394	34.918514	34.862917	0.862729	0.798626	1.119142
MAPE%	0.015369	0.015185	0.016103	0.130691	0.130813	0.130204	0.002864	0.002647	0.003731

實驗 6 - MERRA2-1var-S4+AFRK-low-day

本實驗同樣降低部分參數，但提升模型層數到與前次 meeting 更接近的參數設定，且減小訓練集的時間序列至 3 個月，具體參數如下：

model.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
wavenet:
  # WaveNet model parameters
  input_channels:  1  # Number of input channels
  output_channels: 1  # Number of output channels
  residual_layers: 28  # Number of residual layers
  residual_channels: 40  # Number of channels in residual blocks
  skip_channels: 40  # Number of channels in skip connections

  # Diffusion step embedding dimensions
  diffusion_step_embed_dim_input:  128  # Input dimension
  diffusion_step_embed_dim_hidden: 256  # Middle dimension
  diffusion_step_embed_dim_output: 256  # Output dimension

  # Structured State Spaces sequence model (S4) configurations
  s4_max_sequence_length: 1992  # Maximum sequence length
  s4_state_dim: 128  # State dimension
  s4_dropout: 0.2  # Dropout rate
  s4_bidirectional: true  # Whether to use bidirectional layers
  s4_use_layer_norm: true  # Whether to use layer normalization

diffusion:
  # Diffusion model parameters
  T: 100  # Number of diffusion steps
  beta_0: 0.0001  # Initial beta value
  beta_T: 0.01  # Final beta value

AFRK:
# AutoFRK model parameters, for more details, please refer to https://pypi.org/project/autoFRK/
  method: "fast"  # autoFRK method to use (e.g., "fast")
  tps_method: "rectangular"  # autoFRK's TPS method to use (e.g., "rectangular")

training.yaml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
# Training configuration
batch_size: 80  # Batch size
output_directory: "/home/u6025091/SSSD_CP/results/MERRA2-1var-S4+AFRK-final-low-days"  # Output directory for checkpoints and logs
ckpt_iter: "max"  # Checkpoint mode (max or min)
iters_per_ckpt: 500  # Checkpoint frequency (number of epochs)
iters_per_logging: 200  # Log frequency (number of iterations)
n_iters: 4000  # Maximum number of iterations
learning_rate: 0.0005  # Learning rate

# Additional training settings
only_generate_missing: true  # Generate missing values only
use_model: 2  # Model to use for training
masking: "forecast"  # Masking strategy for missing values
missing_k: 216  # Number of missing values

# Data paths
data:
  train_path: "/home/u6025091/SSSD_CP/datasets/MERRA2-1var-final-low-days/data_train_known_real.npy"  # Path to training data

# autoFRK config
enable_spatial_training: true  # Enable spatial training step
location_path: "/home/u6025091/SSSD_CP/datasets/MERRA2-1var-final-low-days/stations_known_locations.npy"  # Path to known locations

訓練花費時間：24h 25m 30.20s
推論階段 AFRK 之 MRTS 基底數：301

Metric	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
MSPE	1.346637	1.077605	2.422764	10.404323	10.459679	10.182895	0.364479	0.060272	1.581304
RMSPE	1.160447	1.038078	1.556523	3.225573	3.234143	3.191065	0.603721	0.245504	1.257499
MSPE%	0.004831	0.003865	0.008696	0.037348	0.037534	0.036602	0.001305	0.000214	0.005670
RMSPE%	0.069507	0.062171	0.093250	0.193256	0.193738	0.191316	0.036130	0.014640	0.075297
MAPE	0.517593	0.406171	0.963280	2.524476	2.528199	2.509585	0.299979	0.176072	0.795608
MAPE%	0.001852	0.001454	0.003446	0.009074	0.009087	0.009024	0.001069	0.000626	0.002841

結論

目前而言，採實驗 6 之參數設定的 MERRA2-1var-S4+AFRK-low-day 實驗在各項評估指標上表現最佳，尤其在未來時間點的預測表現上，MSPE、RMSPE、MAPE 等指標均有顯著降低，且在過去時間點的預測表現也有不錯的成效。故下次實驗應以此參數設定為基礎，進行更多次的實驗以驗證其穩定性與可靠性。

目前問題為，是否需同步降低 Weather2k 之資料集時間序列，仍需教授不吝指導。

以上部分實驗空缺，因實驗尚未完成，敬請見諒！

https://raw.githubusercontent.com/Josh-test-lab/website-assets-repository/refs/heads/main/posts/1150317%20meeting/not%20yet%20finish.png — 實驗進行中。

實驗快照

本次實驗共使用 4 臺國網中心機器進行，以下呈現其中 3 臺的實驗快照。

https://raw.githubusercontent.com/Josh-test-lab/website-assets-repository/refs/heads/main/posts/1150317%20meeting/Experimental%20snapshot.png — 實驗快照。

論文暫寫

參考資料

Zhu X, Xiong Y, Wu M, et al. Weather2K: A Multivariate Spatio-Temporal Benchmark Dataset for Meteorological Forecasting Based on Real-Time Observation Data from Ground Weather Stations[C]//International Conference on Artificial Intelligence and Statistics. PMLR, 2023: 2704-2722.
Juan Lopez Alcaraz 、 Nils Strodthoff（2022）。Diffusion-based time series imputation and forecasting with structured state space models。Transactions on Machine Learning Research。參考自 https://openreview.net/forum?id=hHiIbk7ApW
SSSD（2022）。GitHub。參考自 https://github.com/AI4HealthUOL/SSSD

目錄

1150317 meeting

收尾工作 v2.0

實驗介紹

實驗設定

實驗 1

Weather2k-1var-S4+AFRK

MERRA2-1var-S4+AFRK

實驗 2

Weather2k-1var-S4

MERRA2-1var-S4

實驗 3

Weather2k-S4+AFRK

MERRA2-S4+AFRK

實驗 4

Weather2k-S4

MERRA2-S4

Bonus

實驗 5 - MERRA2-1var-S4+AFRK-low-batch

實驗 6 - MERRA2-1var-S4+AFRK-low-day

結論

實驗快照

論文暫寫

參考資料

相關文章