Tropical Cyclone Intensity Probabilistic Forecasting System Based on Deep Learning
Table 2
Comparison of PTCIF and baseline models on deterministic, probabilistic, and interval performance metrics. The entire test phase is covered as well as all prediction times.
6 h
12 h
18 h
24 h
MAE
CRPS
PICP
MWP
MAE
CRPS
PICP
MWP
MAE
CRPS
PICP
MWP
MAE
CRPS
PICP
MWP
2015
MC dropout
5.732
4.832
0.285
0.179
3.370
2.698
0.439
0.164
4.488
3.699
0.329
0.171
5.295
4.468
0.358
0.189
Deep ensemble
4.887
3.559
0.949
0.733
4.729
3.438
0.964
0.711
6.437
4.652
0.823
0.756
5.960
4.318
0.885
0.787
QR
7.378
—
0.331
0.571
7.229
—
0.318
0.568
8.544
—
0.146
0.142
8.363
—
0.157
0.525
Bootstrap
2.027
1.583
0.383
0.101
2.847
2.369
0.295
0.101
3.859
3.315
0.257
0.109
5.242
4.370
0.292
0.188
PTCIF
1.741
2.537
1.000
0.941
2.888
2.880
0.994
0.944
3.881
2.998
0.955
0.763
4.732
3.523
0.902
0.776
2016
MC dropout
6.197
5.309
0.280
0.186
3.708
2.981
0.415
0.176
4.884
4.028
0.333
0.183
5.719
4.814
0.290
0.201
Deep ensemble
5.401
3.742
0.900
0.671
5.080
3.591
0.902
0.659
7.450
5.284
0.716
0.722
6.799
4.809
0.796
0.705
QR
7.753
—
0.297
0.504
7.733
—
0.270
0.501
9.213
—
0.078
0.126
9.296
—
0.133
0.447
Bootstrap
2.146
1.679
0.456
0.116
3.307
2.790
0.321
0.118
4.164
3.621
0.258
0.116
6.621
5.590
0.268
0.223
PTCIF
1.981
2.640
0.998
1.010
3.211
3.092
0.975
1.011
4.303
3.352
0.912
0.822
5.279
3.935
0.865
0.850
2017
MC dropout
3.273
2.465
0.517
0.260
3.708
2.981
0.415
0.176
3.823
3.233
0.358
0.189
4.348
3.685
0.357
0.215
Deep ensemble
3.724
2.668
0.936
0.642
3.713
2.632
0.923
0.625
5.490
3.836
0.757
0.684
4.919
3.483
0.827
0.695
QR
5.361
—
0.381
0.490
5.277
—
0.382
0.491
6.474
—
0.103
0.123
6.375
—
0.129
0.424
Bootstrap
1.771
1.390
0.443
0.108
2.564
2.143
0.344
0.109
3.272
2.766
0.333
0.135
4.466
3.713
0.366
0.217
PTCIF
1.633
2.539
1.000
1.169
2.612
2.847
0.993
1.182
3.289
2.756
0.969
0.964
4.272
3.288
0.919
0.992
2018
MC dropout
2.806
2.182
0.495
0.191
2.859
2.340
0.406
0.163
4.138
3.390
0.396
0.204
4.776
3.950
0.376
0.230
Deep ensemble
4.801
3.368
0.907
0.641
4.595
3.248
0.917
0.631
6.291
4.525
0.754
0.682
5.979
4.213
0.823
0.684
QR
6.771
—
0.239
0.484
6.668
—
0.249
0.485
8.224
—
0.066
0.132
8.057
—
0.099
0.433
Bootstrap
1.760
1.371
0.432
0.100
2.700
2.253
0.324
0.102
3.508
3.002
0.267
0.114
5.051
4.251
0.309
0.186
PTCIF
1.562
2.513
0.999
1.075
2.735
2.865
0.993
1.085
3.496
2.845
0.963
0.883
4.391
3.328
0.926
0.904
The best performance is labeled in bold, and the CRPS metric is labeled as — since quantile regression can only predict the upper and lower bounds.
