DAILY PREDICTION OF PARTICULATE MATTERS USING NEURAL NETWORKS AND ENSEMBLING METHODS

Final results of the investigation project.

Transcript

1. DAILY PREDICTION OF PARTICULATE MATTERS USING NEURAL NETWORKS AND ENSEMBLING METHODS By Diego Triana Supervisor: prof. dr hab. inż. Stanisław Osowski 1
2. PROBLEM DESCRIPTION Measurements of PM in a given place Raw Data Model 1 MLP Model 2 RBF Model 3 SVM Training Testing FORECAST Can we improve the forecast performance by combining the models? Performance 2
3. METHODOLOGY • PM2.5 Chinese • PM10 Polish • Data Analysis ▪ Statistical Analysis ▪ Processing input variables • Generation of Diagnostic Features • Feature Selection • PCA • Stepwise • Design of Predictors ▪ MLP ❖ Levenberg-Marquardt back- propagation algorithm ▪ RBF ❖ Levenberg-Marquardt algorithm ▪ SVM ❖ Bayesian optimization (C , σ and ε) ▪ RF ❖ LSBoost • Measure of Performance • Comparison of Results • Ensemble Methods ▪ Dynamic Integration ▪ Bagging ▪ Adaboost • Optimization of Ensembles ▪ Number of Predictors ▪ Loss Function 3
4. DATASETS (RAW DATA) 4
7. DIAGNOSTIC FEATURES 7
9. Daily Chinese Dataset 9
10. Daily Polish Dataset 10
13. PERFORMANCE OF SINGLE PREDICTORS PM2.5 Chinese Data PM10 Polish Data 13
14. REPEATABILITY OF SINGLE PREDICTORS PM10 (Poland) PM2.5 (China) 14
15. BAGGING • Bags of 80% from whole training dataset • 150 Predictors 15
16. ADABOOST Slopes: PM10 |PL | MLP | γ =3.5 PM10 |PL | RF | γ =2.0 PM2.5 |CH| MLP | γ =4.5 PM2.5 |CH| RF | γ =2.5 T=150 Predictors 16
17. MAE (AVERAGE 20 TIMES RUNS) 17
18. MAPE (AVERAGE 20 TIMES RUNS) 18
19. RMSD (AVERAGE 20 TIMES RUNS) 19
20. Q&A • THANK YOU 20 Instytut Elektrotechniki Teoretycznej i Systemów Informacyjno-Pomiarowych Zakład Elektrotechniki Teoretycznej i Informatyki Stosowanej