目錄

1150116 meeting

本次改採用 Weather2K 資料集作訓練,並做不同方面的實驗。

資料集

Weather2K 資料集變數如下:

Numpy IndexLong NameShort NameUnit
0Latitudelat(°)
1Longitudelon(°)
2Altitudealt(m)
3Air pressureaphpa
4Air Temperaturet(°C)
5Maximum temperaturemxt(°C)
6Minimum temperaturemnt(°C)
7Relative humidityrh(%)
8Precipitation in 3hp3(mm)
9Wind directionwd(°)
10Wind speedws(ms-1)
11Maximum wind directionmwd(°)
12Maximum wind speedmws(ms-1)

在以下實驗中,皆隨機選取資料集中 500 個地點進行實驗,並選取 2020 年 9 月至 2021 年 9 月之資料作為訓練與測試日期。其中,已知站點占全部的 0.95 ,訓練時間點占全部的 0.9 ,缺失時間占測試時間點中的 0.3 。

實驗一

本實驗採用全部 9 個變數,並將其依時間個別拆分為 8 個變數,形成總共 $9 \times 8 = 72$ 個變數,增加各變數之間的共變性。

在以下實驗中,標示為 FRK 的表格表示在訓練過程中加入了以下程式碼,用於使訓練過程中考慮空間關係。其中,為確保 autoFRK 可順利使用,將 SSSDepsilon_theta 預測值加入微小的隨機值,讓 autoFRK 能順利預測。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
if loc is not None:
    temp = epsilon_theta.permute(1, 0, 2)
    epsilon_theta_fused = torch.zeros_like(epsilon_theta)
    for i in range(temp.shape[0]):
        success = False
        while not success:
            try:
                df = temp[i] + 1e-6 * torch.randn_like(temp[i])
                frk_model = AutoFRK(device=df.device, dtype=df.dtype)
                frk_model.forward(data=df, loc=loc, requires_grad=True)
                pred_res = frk_model.predict(newloc=loc)
                frk_pred = pred_res['pred.value']
                epsilon_theta_fused[:, i, :] = frk_pred
                success = True
            except Exception as e:
                print(f"[Warning] Processing of record {i} failed. Will retry. Error: {e}")
else:
    epsilon_theta_fused = epsilon_theta
  • 標示為 FRK-3000 表示在訓練過程中採用了上述方式進行訓練,且訓練了 3,000 次迭代。
  • 標示為 No-FRK-3000 表示在訓練過程中無使用 autoFRK 進行訓練,且訓練了 3,000 次迭代。

FRK-3000

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE1.076264e+049.718320e+032.074638e+042.651666e+042.517610e+043.013954e+044.097486e+033.178491e+031.677236e+04
RMSPE1.037432e+029.858154e+011.440361e+021.628394e+021.586698e+021.736074e+026.401161e+015.637811e+011.295081e+02
MSPE%4.264463e+123.355991e+121.654310e+121.205811e+138.965348e+125.017307e+129.671504e+119.828016e+112.315043e+11
RMSPE%2.065058e+061.831936e+061.286200e+063.472479e+062.994219e+062.239935e+069.834381e+059.913635e+054.811490e+05
MAPE3.594331e+013.361454e+015.946985e+016.423800e+016.219757e+017.843039e+012.397248e+012.152171e+015.144809e+01
MAPE%1.945626e+111.457879e+111.661445e+114.945142e+113.543819e+113.898608e+116.766003e+105.753665e+107.149523e+10

No-FRK-3000

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE1.133920e+049.039191e+033.470412e+042.658836e+042.294857e+043.913878e+044.887639e+033.154452e+033.282792e+04
RMSPE1.064857e+029.507466e+011.862904e+021.630594e+021.514879e+021.978352e+026.991165e+015.616451e+011.811848e+02
MSPE%5.032590e+125.385601e+122.013602e+121.464198e+131.582552e+136.260803e+129.670808e+119.687107e+112.167088e+11
RMSPE%2.243344e+062.320690e+061.419014e+063.826484e+063.978131e+062.502160e+069.834027e+059.842310e+054.655199e+05
MAPE3.660217e+013.291928e+017.455091e+016.447815e+015.998680e+018.969546e+012.480849e+012.146764e+016.814359e+01
MAPE%2.228934e+111.742531e+112.015055e+115.885326e+114.585876e+115.083518e+116.819993e+105.395768e+107.168597e+10

No-FRK-18000

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE1.078145e+048.978216e+032.185983e+042.655790e+042.283698e+043.292088e+044.106801e+033.114891e+031.718016e+04
RMSPE1.038338e+029.475345e+011.478507e+021.629659e+021.511191e+021.814411e+026.408433e+015.581121e+011.310731e+02
MSPE%5.164828e+126.281615e+121.840018e+121.502257e+131.817500e+135.649133e+129.942465e+111.249800e+122.284686e+11
RMSPE%2.272626e+062.506315e+061.356472e+063.875896e+064.263214e+062.376790e+069.971191e+051.117945e+064.779839e+05
MAPE3.592561e+013.253873e+016.130469e+016.400942e+015.888296e+018.374264e+012.404400e+012.139310e+015.181171e+01
MAPE%2.251511e+111.839792e+111.881160e+115.951222e+114.823095e+114.650825e+116.862479e+105.776257e+107.093789e+10

實驗二

為避免實驗一中各變數無共變性,以下僅使用「溫度」變數作為測試,並將其拆分為 8 個變數,共 $1 \times 8 = 8$ 個變數。

FRK-3000

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE15.65645625.63669812.29814529.52732650.61756726.8126009.78801115.0678696.157414
RMSPE3.9568245.0632703.5068715.4339057.1146025.1780883.1285803.8817352.481414
MSPE%0.8155362.0964420.4788851.6959284.5431571.1076580.4430631.0612930.212866
RMSPE%0.9030701.4479090.6920161.3022782.1314681.0524530.6656301.0301910.461374
MAPE2.9070603.9582032.5767044.2959166.6083073.8593232.3194672.8370052.034057
MAPE%0.1425670.3047370.0950690.2361130.5780680.1545240.1029900.1890970.069915

FRK-8500

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE15.78867532.75910313.27497029.18740053.62274029.42865510.11998323.9321796.440719
RMSPE3.9734975.7235573.6434835.4025367.3227555.4248183.1811924.8920532.537857
MSPE%0.8266602.4272070.5218681.6961854.7766951.2207070.4587841.4331930.226206
RMSPE%0.9092081.5579500.7224051.3023772.1855651.1048560.6773361.1971600.475611
MAPE2.9256294.4683052.6675184.2929456.7603294.1916002.3471493.4986032.022714
MAPE%0.1438880.3310440.0995650.2373660.5916190.1680000.1043400.2208010.070612

No-FRK-3000

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE15.86654231.03481512.43418129.70431953.10950432.70990910.01209821.6955243.855988
RMSPE3.9832835.5708903.5262135.4501677.2876275.7192583.1641904.6578451.963667
MSPE%0.8235462.3731330.4948851.7127564.7687691.3485000.4473421.3595950.133741
RMSPE%0.9074941.5404980.7034811.3087232.1837511.1612490.6688361.1660170.365706
MAPE2.9257404.5818472.4232844.3206306.7837944.4528362.3355943.6502531.564628
MAPE%0.1431660.3404650.0908980.2379600.5926390.1774590.1030610.2337760.054276

No-FRK-18000

MetricALL Locs & All TimeKnown Locs & All TimeUnknown Locs & All TimeALL Locs & FutureKnown Locs & FutureUnknown Locs & FutureALL Locs & PastKnown Locs & PastUnknown Locs & Past
MSPE15.20875426.98469713.91552427.84254051.69752826.9519979.86369116.5292688.400094
RMSPE3.8998405.1946803.7303525.2766037.1900995.1915313.1406514.0656202.898292
MSPE%0.7908322.1112850.5416401.6147764.6354211.1146000.4422411.0433820.299234
RMSPE%0.8892871.4530260.7359621.2707382.1530031.0557460.6650121.0214610.547023
MAPE2.8660634.2888172.8040054.1620846.6553263.9432672.3177463.2876012.322010
MAPE%0.1403380.3193130.1039950.2300230.5820430.1578880.1023940.2081570.081194

實驗三

本實驗未確認實驗二是否有效之對照組,與實驗二同為取「溫度」變數,但未衍生為 8 個變數,只使用 1 個變數進行實驗。

(實驗進行中)

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