Table 3 The performance of the best artificial neural networks for the forecast of NO2 concentration in the air.

From: Artificial intelligence accuracy assessment in NO2 concentration forecasting of metropolises air

Function

Test set

Train set

activation

Training

Structure

R2

MSE

RMSE

MAE

R2

MSE

RMSE

MAE

Logsig-Logsig-Purelin

LM

23-26-26-1

0.741

0.235

0.485

0.355

0.879

0.125

0.354

0.252

Tansig-Tansig-Purelin

GDM

23-22-22-1

0.562

0.386

0.621

0.468

0.564

1.631

1.277

0.521

Tansig-Purelin

BQN

23-26-1

0.763

0.208

0.456

0.327

0.836

0.163

0.403

0.278

Tangsig-Purelin

CGB

23-26-1

0.742

0.232

0.481

0.334

0.877

0.122

0.35

0.25

Tansig-Tansig-Purelin

GDA

23-19-19-1

0.722

0.246

0.496

0.341

0.821

0.181

0.425

0.294

Logsig-Logsig-Purelin

CGP

23-24-24-1

0.711

0.254

0.504

0.353

0.829

0.172

0.414

0.288

  1. LM Levenberg–Marquardt, GDM gradient descent with momentum, BQN BFGS quasi newton, CGB conjugate gradient with Powell/Beale, GDA gradient descent with adaptive learning rate, CGP conjugate gradient with Polak/Ribiere.
Back to article page