Coupling dynamical and statistical downscaling for high-resolution rainfall forecasting: case study of the Red River Delta, Vietnam

Progress in Earth and Planetary Science

Table 3 Distinctive ANN models considered in the preliminary test

Model	Architecture^a	Transfer	Training data	Training set dimension
Model	Architecture^a	function	Training data	Training set dimension
M1n	16-10-5-3-1	LS^b	slp, t2, hgt, pblh	4 NVs^d × 4 grids
M1a	4-4-4-4-1	LS	slp, t2, hgt, pblh	4 AVs^e
M1s	4-4-4-4-1	LS	slp, t2, hgt, pblh	4 SVs^f
M2n	20-12-8-5-1	LS	Rd1, slp, t2, hgt, pblh	5 NVs × 4 grids
M2a	5-5-5-5-1	LS	Rd1, slp, t2, hgt, pblh	5 AVs
M2s	5-5-5-5-1	LS	Rd1, slp, t2, hgt, pblh	5 SVs
M2d	20-12-8-5-1	HTS^c	Rd1, slp, t2, hgt, pblh	5 NVs × 4 grids
M2e	20-20-10-5-1	HTS	Rd1, slp, t2, hgt, pblh	5 NVs × 4 grids
M3n	28-20-10-7-1	HTS	slp, tk, hgt, pblh, u10, v10, grdflx	7 NVs × 4 grids
M4n	32-25-20-10-1	HTS	Rd1, tk, hgt, slp, grdflx, psfc, pblh, q2	8 NV × 4 grids
M4a	8-6-5-5-1	HTS	Rd1, tk, hgt, slp, grdflx, psfc, pblh, q2	8 AVs
M4as	16-10-5-3-1	HTS	Rd1, tk, hgt, slp, grdflx, psfc, pblh, q2	8 (AVs + SVs)
M5n	32-25-20-10-1	HTS	Same as M4n but for rainfall events

^aModel architecture indicates the number of neuron in each layer of a 5-layer MLP network, wherein first number is neurons of input, the following three numbers are neurons of hidden layers, and the last one is neurons of the output layer
^bLS logistic sigmoid
^cHTS hyperbolic tangent sigmoid
^d,e,fNV, AV, and SV correspond to normal variable, averaged variable, and standard deviation variable, respectively; in which:
-NVs are directly extracted from the 4-predictor grids, so the actual number of variable are multiplied to 4;
-AVs are the new features created by computing the mean of each variable in the 4-predictor grids;
-SVs are the new features created by computing the standard deviation of each variable in the 4-predictor grids