PREDICTION OF DISSOLVED OXYGEN IN 
HARBOURS USING ARTIFICIAL NEURAL 
NETWORKS: AN APPLICATION TO THE PORT OF 
COLOMBO 
By 
W.K.C.N. DAYANTHI 
THIS THESIS W A S S U B M I T T E D TO THE D E P A R T M E N T OF CIVIL ENGINEERING OF 
THE UNIVERSITY OF M O R A T U W A IN PARTIAL FULFILLMENT OF THE 
REQUIREMENTS FOR THE DEGREE OF M A S T E R OF ENGINEERING. 
D E P A R T M E N T OF CIVIL ENGINEERING 
UNIVERSITY OF M O R A T U W A 
SRI L A N K A 
D E C E M B E R 2001. 
0 7 4 4 4 3 
mvtrsil* of Moratuwa 
5 4 3 . 3 ; 0 0 1 
74443 
DECLARATION 
I hereby declare that the work included in the thesis in part or whole has not been 
submitted in any form for any other academic qualification of any institution. 
Miss W.K.C.N.Dayanthi 
M.Eng /C /01 /2000 
Certified by 
Dr. Mahesh Jayaweera. 
Supervisor. 
ABSTRACT 
Maintenance of high dissolved oxygen (DO) level in harbours is highly important 
as it could give rise to catastrophic effects if it is depleted affecting day- to- day 
port functions such as dredging activities and other maintenance work. The 
depletion of DO results not only in toxic gases such as methane and hydrogen 
sulfide but also in accumulation of wastes. Frequent monitoring of DO is 
therefore imperative, but makes practical difficulties due to ship movements and 
other activities. Hence, prediction of DO with an empirical model using Artificial 
Neural Networks (ANNs) was done with success with an application to the Port of 
Colombo (PoC). This model aims to lessen the frequency of monitoring DO and 
to foresee the responses of the system due to environmental changes. The 
performances of ANNs were compared with Multiple Linear Regression (MLR). 
Monthly values of 14 water quality parameters at several depths for the period of 
four years from 1997 to year 2000 were collected. The values of weather 
parameters of rainfall and wind velocity for the corresponding period were also 
collected. The neural network possessing 7 inputs and 45 hidden neurons, 
performed well giving rise to correlation coefficient (R) as 0.87 and 0.67 for 
calibration and verification respectively. The inputs are temperature, depth and 
five rainfall intensities (including values on four immediate previous days). A 
sensitivity analysis was carried out to assess the potentials of small changes in 
each input on the neural network output. MLR model with 7 input variables 
indicated R to be 0.45 for calibration after several transformations. The 
temperature was the most influential variable among the ANN inputs affecting the 
output. In conclusion, it could be inferred that the ANN model is capable of 
predicting DO in PoC considerably well compared with MLR. 
ACKNOWLEDGEMENT 
First, I would like to render my cordial thanks to my supervisor, Dr. Mahesh 
Jayaweera for advising and guiding me in all aspects. My gratitude should also go 
to Professor (Mrs.) N. Rathnayake, Director of Postgraduate Studies, for 
supporting me in numerous ways. 
Further, I would like to extend my gratitude to the Asian Development Bank 
(ADB) for the provision of financial aids. Special thanks hereby go to the Japan 
Port Consultant (Pvt.) Ltd. for giving me the necessary water quality data of the 
inner harbour of the Port of Colombo. I am also grateful to the Meteorological 
Department for providing me with meteorological data. 
It is with grateful appreciation that I acknowledge the contribution of the 
Environmental Engineering Laboratory-staff in the University of Moratuwa, 
namely Ms. Priyanka Dissanayake, Mrs. Nilanthi Gunathilake and Mr. Justin 
Silva. At last, but not least my sincere thanks go to all the others who helped me 
in numerous ways to make this a success. 
TABLE OF CONTENTS 
Page No. 
List of tables i 
List of figures iii 
List of symbols vi 
List of abbreviations vii 
Chapter 1 : Introduction 
1.1 Background 1-1 
1.2 Application to the Port of Colombo 1 -4 
1.2.1 General description of the Port of Colombo 1 -4 
1.2.2 Environmental degradation in the Port of 
Colombo 1-5 
1.2.3 Water quality parameters 1-6 
1.3 Aims and Objectives 1-10 
1.4 Methodology 1-10 
1.5 Main findings 1-11 
1.6 Presentation of chapters 1-12 
Chapter 2 : Literature Review 
2.1 Research carried out so far on Artificial Neural 
Network (ANN) in the field of ecology 2-1 
Chapter 3 : Model Formulation 
3.1 Multiple Linear Regression 3-1 
3.1.1 Introduction 3-1 
3.1.2 Multiple correlation 3-2 
3.1.3 Transformation of data 3-2 
3.1.4 Linear regression 3-5 
3.1.5 Multiple linear regression using more 
> 
than two independent variables 3-8 
3.1.6 Stepwise regression 3-9 
3.2 Artificial Neural Networks 
3.2.1 Introduction 3-9 
3.2.2 Backpropagation techniques 3-9 
3.3 Windows Neural Network ( WinNN); Software 
Package 3-20 
Chapter 4 Materials and Methods 
4.1 Data collection 4-1 
4.2 Modeling Techniques 4-1 
4.2.1 Introduction 4-1 
4.2.2 Case Study 4-4 
4.2.3 Modeling using 9 input parameters 4-6 
4.2.4 Modeling using 7 input parameters 4-7 
Chapter 5 : Results 
5.1 Summary 5-1 
5.2 Results of the case study 5-1 
5.2.1 ANN-16 5-1 
5.2.2 Sensitivity analysis on ANN-16 5-1 
5.2.3 MLR-16 5-8 
5.2.4 Sensitivity analysis on MLR-16 5-8 
5.3 Modeling using 9 input parameters 5-8 
5.3.1 ANN-9 5-8 
5.3.2 MLR-9 5-16 
5.4 Modeling using 7 input parameters 5-16 
5.4.1 ANN-7 5-16 
5.4.2 Sensitivity analysis on ANN-7 5-16 
5.4.3 MLR-7 5-18 
> 
Chapter 6: Discussion 
6.1 Correlation between observed and estimated values 6-1 
6.1.1 ANN-16 and ANN-9 6-1 
6.1.2 MLR-16 and MLR-9 6-2 
6.1.3 ANN-7 and MLR-7 6-3 
6.2 Examination of the residuals 6-5 
6.2.1 Relationship between the residuals and 
DO (Observed and Predicted) 6-5 
6:2.2 The normality of residuals 6-7 
6.3 Sensitivity analysis 
6.3.1 Sensitivity analyses on ANN-16, ANN-7 
and MLR-16 6-8 
Chapter 7: Conclusions and Further recommendations 
7.1 Conclusion and further recommendations 7-1 
Bibliography 
Appendix A A-l 
Appendix B B-l 
Appendix C C-l 
LIST OF TABLES 
Table 3.1 Transforms for several types 
of curves 
Table 5.1 Parameters of the WinNN 
control panel for each model 
Table 5.2 Statistical Evaluation 
Table 6.1 Percentage variation of residuals for 
testing data of each model 
Table A-l Input data of calibration data set of 
ANN-16 and MLR-16 
Table A-2 Input data of verification data set of 
ANN-16 and MLR-16 
Table A-3 Input data of calibration data set of 
ANN-9 and MLR-9 
Table A-4 Input data of verification data set of 
ANN-9 and MLR-9 
Table A-5 Input data of calibration data set of 
ANN-7 and MLR-7 
Table A-6 Input data of verification data set of 
ANN-7 and MLR-7 
Table B-l Correlation coefficients of the trained 
neural networks using 9 input parameters 
Table B-2 Network characteristics of the trained neural 
networks using 7 input parameters 
Table C-l Predicted and observed values of the 
calibration data set of ANN-16, ANN-9 
and ANN-7 
Table C-2 Predicted and observed values of the 
verification data setof ANN-16, ANN-9 
l 
and ANN-7 
Table C-3 Predicted and observed values of the 
calibration data set of MLR-16, MLR-9 
and MLR-7 
Table C-4 Predicted and observed values of the 
verification data set of MLR-16, MLR-9 
and MLR-7 
C-8 
C- l l 
C-18 
11 
LIST OF FIGURES 
Page No. 
Fig. 3.1 Confirmatory and exploratory fits 3-6 
Fig. 3.2 Confirmatory and exploratory fits 3-6 
Fig. 3.3 Input Matrix 3-11 
Fig. 3.4 Output Matrix 3-11 
Fig. 3.5 Schematic illustration of a three-layered 
feed-forward neural (backpropagation) network 3-13 
Fig. 3.6 Network designed for basic backpropagation 3-14 
Fig. 3.7 Generalized network design with time lags 3-19 
Fig. 4.1 Location of sampling points 4-2 
Fig. 5.1 Graph of predicted vs observed DO for 
ANN-16 (Training) 5-2 
Fig. 5.2 Graph of predicted vs observed DO for 
ANN-16 (Testing) 5-2 
Fig. 5.3 Sensitivity of temperature for ANN-16 model 5-3 
Fig. 5.4 Sensitivity of Rain-2 for ANN-16 model 5-3 
Fig. 5.5 Sensitivity of Ammonium-N for ANN-16 model 5-3 
Fig. 5.6 Sensitivity of Total-N for ANN-16 model 5-4 
Fig. 5.7 Sensitivity of COD for ANN-16 model 5-4 
Fig. 5.8 Sensitivity of SS for ANN-16 model 5-4 
Fig. 5.9 Sensitivity of Nitrate-N for ANN-16 model 5-5 
Fig. 5.10 Sensitivity of Nitrite-N for ANN-16 model 5-5 
Fig. 5.11 Sensitivity of Organic-N for ANN-16 model 5-5 
Fig. 5.12 Sensitivity of Total-P for ANN-16 model 5-6 
Fig. 5.13 Sensitivity of Wind for ANN-16 model 5-6 
Fig. 5.14 Sensitivity of Rain-1 for ANN-16 model 5-6 
Fig. 5.15 Sensitivity of Rain-3 for ANN-16 model 5-7 
Fig. 5.16 Sensitivity of Rain-4 for ANN-16 model 5-7 
Fig. 5.17 Sensitivity of Rain-5 for ANN-16 model 5-7 
Fig. 5.18 Graph of predicted vs observed DO for 5-9 
MLR-16 (Training) 
Fig. 5.19 Graph of predicted vs observed DO for 
iii , , „ 
MLR-16 (Testing) 5-9 
Fig. 5.20 Sensitivity of Temperature for MLR-16 model 5-10 
Fig. 5.21 Sensitivity of Ammonium-N for MLR-16 model 5-10 
Fig. 5.22 Sensitivity of Rain-1 for MLR-16 model 5-10 
Fig. 5.23 Sensitivity of Rain-2 for MLR-16 model 5-11 
Fig. 5.24 Sensitivity of Rain-5 for MLR-16 model • 5-11 
Fig. 5.25 Sensitivity of Total-P for MLR-16 model 5-11 
Fig. 5-26 Sensitivity of COD for MLR-16 model 5-12 
Fig. 5-27 Sensitivity of SS for MLR-16 model 5-12 
Fig. 5-28 Sensitivity of Organic-N for MLR-16 model 5-12 
Fig. 5-29 Sensitivity of Nitrate-N for MLR-16 model 5-13 
Fig. 5-30 Sensitivity of Nitrite-N for MLR-16 model 5-13 
Fig. 5-31 Sensitivity of Total-N for MLR-16 model 5-13 
Fig. 5-32 Sensitivity of Wind for MLR-16 model 5-14 
Fig. 5.33 Variation of correlation coefficients for 
training and testing data 5-14 
Fig. 5.34 Graph of predicted vs observed DO for 
ANN-9 (Training) 5-15 
Fig. 5.35 Graph of predicted vs observed DO for 
ANN-9 (Testing) 5-15 
Fig. 5.36 Graph of predicted vs observed DO for 
MLR-9 (Training) 5-17 
Fig. 5.37 Graph of predicted vs observed DO for 
MLR-9 (Testing) 5-17 
Fig. 5.38 Variation of correlation coefficients for 
training and testing data 5-18 
Fig. 5.39 Graph of predicted vs observed DO for 
ANN-7 (Training) 5-19 
Fig. 5.40 Graph of predicted vs observed DO for 
ANN-7 (Testing) 5-19 
Fig. 5.41 Sensitivity of Temperature for ANN-7 model 5-20 
Fig. 5.42 Sensitivity of Rain-1 for ANN-7 model 5-21 
Fig. 5.43 Sensitivity of Rain-2 for ANN-7 model 5-21 
Fig. 5.44 Sensitivity of Rain-3 for ANN-7 model 5-21 
iv 
Fig. 5.45 Sensitivity of Rain-4 for ANN-7 model 5-22 
Fig. 5.46 Sensitivity of Rain-5 for ANN-7 model 5-22 
Fig. 5.47 Graph of predicted vs observed DO for 
MLR-7 (Training) 5-23 
Fig. 5.48 Graph of predicted vs observed DO for 
MLR-7 (Testing) 5-23 
Fig. 5.49 Graph of predicted vs observed DO for MLR-7 
(Training omitting 50 data sets) 5-24 
Fig. 5.50 Graph of predicted vs observed DO for MLR-7 
(Training omitting 50 data sets) 5-24 
Fig. 6.1 Relationship between the residuals and the 
estimated and observed values of DO for 
training data set-(l),(2):ANN-7;(3),(4):ANN-9; 
(5),(6):ANN-16 6-10 
Fig. 6.2 Relationship between the residuals and the 
estimated and observed values of DO for 
training data set-(l),(2):MLR-7;(3),(4):MLR-9; 
(5),(6):MLR-16 6-11 
Fig. 6.3 Relationship between the residuals and the 
estimated and observed values of DO for 
testing data set-(l), (2): ANN-7;(3),(4):ANN-9; 
(5 ) , (6 ) :ANN-16 6-12 
Fig. 6.4 Relationship between the residuals and the 
estimated and observed values of DO for 
testing data set-(l), (2): MLR-7;(3),(4):MLR-9; 
(5),(6):MLR-16 6-13 
Fig. 6.5 Distribution of residuals for training 
data sets(l) MLR, 16 ; (2) ANN, 16 ; 
(3) MLR, 9 ; (4) ANN, 9 
(5) MLR, 7 (Transformed);(6) ANN,7 6-14 
Fig. 6.6 Distribution of residuals for testing 
data sets(l) MLR, 16 ; (2) ANN. 16 ; 
(3) MLR, 9 ; (4) ANN, 9 
(5) MLR, 7 (Transformed);(6) ANN,7 6-15 
v 
LIST OF SYMBOLS 
Name 
Momentum 
Learning parameter 
Change in Y for a unit change in X2 keeping Xi 
constant 
Effect of X'{ on Y' keeping X\ constant. 
Change in Y for a unit change in X2 keeping X/ 
constant 
Effect of X\ on Y[ keeping X\ constant 
Number of pairs. 
Correlation coefficient 
Determination coefficient 
Correlation of X] with Y 
Correlation of X2 with Y 
Correlation of Xj with X2 
Partial correlation of J^Gand Y with Xj Controlled. 
The Slope between low summary point and the 
middle of the curve. 
The Slope between middle summary point and the 
high summary point 
Standard deviation of X/, 
Standard deviation of X2. 
Standard deviation of Y data. 
vi 
Standard deviation of Xj data. 
Standard deviation of Y' data. 
Standard deviation of X' data. 
Dependent variable. 
Dependent variable. 
Standardized version of X. 
A'coordinate of the high summary point 
X coordinate of the low summary point 
A'coordinate of the middle summary point 
Sample mean 
Independent variable 
Standardized version of Y 
Sample mean 
Y coordinate of the high summary point 
Y coordinate of the low summary point. 
Y axis value of middle summary point. 
vn 
LIST O F ABBREVIATIONS 
Abbreviation Name 
Ammonium-N Ammonium Nitrogen 
ANN Artificial Neural Network 
ANN-16 ANN model with 16 inputs 
ANN-9 ANN model with 9 inputs 
ANN-7 ANN model with 7 inputs 
BPN Backpropagation Technique 
BQ Bandaranayake Quay 
CH 4 Methane 
COD Chemical Oxygen Demand 
DO Dissolved Oxygen 
H 2 S Hydrogen Sulphide 
JCT Jaya Container Terminal 
M Mean 
MLR Multiple Linear Regression 
MLR-16 MLR model with 16 inputs 
MLR-9 MLR model with 9 inputs 
MLR-7 MLR model with 7 inputs 
Nitrate-N Nitrate Nitrogen 
Nitrite-N Nitrite Nitrogen 
Organic-N Organic Nitrogen 
ORP Oxygen Reduction Potential 
PoC Port of Colombo 
PVQ Prince Vijaya Quay 
QEQ Queen Elisabeth Quay 
Rain-1 Rainfall intensity on the 1 s t day 
Rain-1 Rainfall intensity on the 2 n d immediate previous 
day 
I* 
r 
Vlll 
Rain-1 Rainfall intensity on the 3 r d immediate previous 
day 
Rain-1 Rainfall intensity on the 4 t h immediate previous 
day 
Rain-1 Rainfall intensity on the 5 t h immediate previous 
day 
RMS Root Mean Square 
SLPA Sri Lanka Ports Authority 
SD Standard Deviation 
SS - Suspended Solids 
Total-N Total Nitrogen 
Total-P Total Phosphorous 
WinNN Windows Neural network 
i x