Urban flood prediction using fuzzy neural

networks:
An investigation on automated network
architecture

Usman T Khan1, Rahma Khalid1, Jianxun He2 & Caterina Valeo3

1York University, Toronto, Canada
2University of Calgary, Canada
3University of Victoria, Canada
 Hayward, J., 2013. The Canadian Press

13/09/2017 2
Physical v data-driven models

-v-

13/09/2017 3
Model architecture
Input Layer Hidden Layer Output Layer
1. Input selection? BH

• Compare 3 input variable selection (IVS)

methods: PCA, PMI & CNPSA
i1 W1
2. Model structure? .
.
W2
• Search algorithm for optimum structure . QP
.
3. Model parameters and output? .
i36
• Fuzzy neural network with fuzzy BO
parameters
4. Others?
• Training methods & criteria, complexity…

13/09/2017 4
13/09/2017 5
 Correlation coefficient

0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

13/09/2017
'UTmax1'
'UTmin1'
'UTAvg1'
'UTP1'
'UQd1'
'UQp1'
'DTmax1'
'DTmin1'
'DTAvg1'
'DTP1'
'DQd1'
'DQp1'
'UTmax2'
'UTmin2'
'UTAvg2'
'UTP2'
'UQd2'
'UQp2'
'DTmax2'
'DTmin2'
'DTAvg2'
'DTP2'
'DQd2'
'DQp2'
Principal Component Analysis

'UTmax3'
'UTmin3'
'UTAvg3'
'UTP3'
'UQd3'
'UQp3'
'DTmax3'
'DTmin3'
PC3
PC1

'DTAvg3'
'DTP3'
'DQd3'
PC4
PC2

'DQp3'
Partial Mutual Information
1 variable 22 variables
800 800
Peak Flow Peak Flow
Predicted Peak flow Predicted Peak flow
700 800
X: 1039 Peak flow
700 800
Peak flow
Y: 787.4
Residuals Residuals
700 700

600 600

600 500
600 500
Discharge (m 3/s)

Discharge (m 3/s)
400 400

300 300

500 500
Discharge (m 3/s)

Discharge (m 3/s)
200 200

100 100

0 0
400 400
-100 -100
10

09

08

07

06

05

04

03

02

01

00

10

09

08

07

06

05

04

03

02

01

00
20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20
300 300

200 200

100 100

0 0
10

09

08

07

06

05

04

03

02

01

00

10

09

08

07

06

05

04

03

02

01

00
13/09/2017 7
20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20

20
Combined Neural Pathway Strength
Analysis Output Layer
Input Layer Hidden Layer
BH

i1 W1
.
. W2
. QP
.
.
i36
BO

13/09/2017 8
Combined Neural Pathway Strength
Analysis

The strength of a pathway from the input to output:

WIO = W1 x W2

• 36 different WIO values for each year

• The larger the value of the WIO the more influential the input

13/09/2017 9
Combined Neural Pathway Strength
Analysis 10

5
WIO = W1 x W2

-5

-10
Q_p(t-1) Q_d(t-1) T_min(t-1) T_max(t-1) T_mean(t-1) P(t-1)
Q_p(t-2) Q_d(t-2) T_min(t-2) T_max(t-2) T_mean(t-2) P(t-2)
Q_p(t-3) Q_d(t-3) T_min(t-3) T_max(t-3) T_mean(t-3) P(t-3)
-15
2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010
13/09/2017 10
Combined Neural Pathway Strength
Analysis

13/09/2017 11
Combined Neural Pathway Strength
Analysis

The ensemble interquartile range (EQR):

EQR = {min(|Q1|,|Q3|)/max(|Q1|,|Q3|)}.sgn(Q1).sgn(Q3)

where Q1 & Q3 are the first & third quartile of all WIO for each
input

13/09/2017 12
 13/09/2017
EQR

-1.00
-0.75
-0.50
-0.25
0.00
0.25
0.50
Q_d(t-1)

P(t-1)

T_min(t-2)

T_max(t-2)

T_mean(t-2)

T_max(t-1)

T_mean(t-1)

T_min(t-1)

Q_p(t-2)

P(t-2)

Q_p(t-3)

T_max(t-3)
EQR for 18 input parameters

Q_d(t-3)

P(t-3)

T_mean(t-3)

Q_p(t-1)

T_min(t-3)

Q_d(t-2)
13
ANN input selection
Input Layer Hidden Layer Output Layer
BH

WIH

Qd(t-1)

WHO
P(t-1) Qp(t)

Tmin(t-2)
BO

13/09/2017 14
ANN structure selection
Input Layer Hidden Layer Output Layer
BH

WIH

Qd(t-1)

WHO
P(t-1) Qp(t)

Tmin(t-2)
BO

13/09/2017 15
ANN architecture selection

13/09/2017 16
Training a Fuzzy Neural Network
Input Layer Hidden Layer Output Layer
BH

WIH

Qd(t-1)

WHO
P(t-1) Qp(t)

Tmin(t-2)
BO

13/09/2017 17
Fuzzy Neural Networks
• Fuzzy number inputs, outputs, weights & biases used for
uncertainty quantification
• Upper & lower bounds rather than deterministic output
16

14
2006
1
12 Risk

100% observations
10

captured

π(x)
8

2
0
100% observations
01/04 19/04 07/05 captured

13/09/2017 18
Model performance

13/09/2017 19
 Flow Rate (m3/s)

500

0
1000
1500
2000
2500
3000
16-May-13

13/09/2017
30-May-13

13-Jun-13

27-Jun-13

11-Jul-13

25-Jul-13

08-Aug-13

22-Aug-13
Test dataset: 2013 flood results

05-Sep-13

19-Sep-13
20
Conclusions + future work
• CNPSA efficient IVS method

• Search algorithm can replace ad hoc approach

• FNN produces intervals for risk analysis

• Real-time updating capability to improve performance

• Include complexity as a criteria for model selection

13/09/2017 21
Acknowledgements
• Dr A P Duncan (Centre for Water Systems, Exeter)
• Mr M Hoege (Dresden/Stuttgart)
• Funding: Natural Sciences & Engineering Research Council of
Canada, York University, University of Victoria

Dr Usman T Khan
utkhan.info.yorku.ca
usman.khan@lassonde.yorku.ca

13/09/2017 22
References
• Fuzzy Neural Networks:
Khan U.T., & Valeo, C. (2017). Optimising Fuzzy Neural Network Architecture for
Dissolved Oxygen Prediction and Risk Analysis. Water, 9(6), 381.
Khan, U. T., & Valeo, C. (2016). Dissolved oxygen prediction using a possibility theory
based fuzzy neural network. Hydrology & Earth System Sciences, 20(6), 2267.
• Data-driven methods for real-time flood prediction
Khan, U. T., & Valeo, C. (2016). Short-Term Peak Flow Rate Prediction and Flood Risk
Assessment Using Fuzzy Linear Regression. Journal of Environmental
Informatics, 28(2), 71-89.
• Combined Neural Pathway Strength Analysis:
Duncan, A. P. (2015). The Analysis and Application of Artificial Neural Networks for Early
Warning Systems in Hydrology and the Environment. PhD Thesis, University of
Exeter, UK.

13/09/2017 23