1) The document describes research on developing intelligent real-time water level forecast models for pumping stations using artificial neural networks. 2) Researchers analyzed water level and rainfall data from a pumping station to select input factors for different neural network models, including BPNN, Elman NN, and NARX. 3) The results showed that the NARX model provided the most accurate water level forecasts for lead times of 10, 30, and 50 minutes compared to observed data.

1. 3rd ANNOUNCEMENT
PAWEES
2013
The 12th Conference of
International Society of
Paddy and Water
Environment
Engineering
Intelligent Real-time Water Level
Forecast Models for Pumping Stations
Department of Bioenvironmental Systems Engineering, National Taiwan University
Fi-John Chang, Ying-Ray Lu
Department of Bioenvironmental System Engineering, National Taiwan University, Taipei, Taiwan, ROC
Advisor: Distinguished Professor Fi-John Chang (changfj@ntu.edu.tw)

2. Outline
2
Study Motivation
Methodology
Applications
Results and Discussions
Conclusions

3. Motivation
• Urbanization leads to a reduction in the time of rainfall
concentration.
• Climate Change causes fast rising peak flows.
→ Urban flood control is a crucial task, particularly in developed cities.
3
達第一次警戒
達第一次警戒
達第一次警戒
達第一次警戒
達第一次警戒
達第一次警戒
達第一次警戒
達第一次警戒

4. Keelung River
Layout of Pumping Station
4
Yu-Chung Pumping station
Structure ChartRacking MachineFront Pool Pumps
Sewer
Center console

5. Materials
• Study Area
• Yu-Cheng Pumping Station
• Select Events
• 13 events of typhoons &
heavy rainfall in 2004-2013
5
• Data Collection
• Water level at the
pumping station
• Water levels of sewer
outlets (YC2-YC12)
• Rainfall (R1-R6, Average
Rainfall)
Year 2013 2012 2012 2010 2009 2008 2008 2006 2005 2005 2004 2004 2004
Event 511 Saola 612 Megi Parma Jangmi Sinlaku 910 Talim Haitang Nanmadol Nockten Haima
Number
of data
85 221 113 145 320 307 197 148 140 143 65 150 148
Mean water
level (m)
1.79 2.07 2.57 2.13 2.07 2.05 2.25 2.08 2.25 2.17 2.5 2.23 3.12
Standard
deviation (m)
0.37 0.31 0.55 0.09 0.14 0.39 0.28 0.26 0.19 0.18 0.24 0.48 1.04

6. 6
Model Construction
Data Collection
Data Analysis
Input Selection
Forecast Models
Pearson's Correlation Coefficient
Rainfall
(6 Stations)
Water Level at Pumping Station
(1 Station)
Sewer Water Level
(11 Stations)
Gamma Test
(Key Factors Assessment)
BPNN
Elman NN
NARX
Static Neural Network
Dynamic Neural Network

7. Data Analysis
7
Rainfall vs. Water level at the pumping station
• Rainfall vs. water level
• Rainfall vs. RECOVERED water level
• Accumulated rainfall vs. RECOVERED water
level
Water level vs. Water level
• Water levels of sewer outlets and the water level
at the pumping station
• Pearson's Correlation Coefficient

8. 0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0 10 20 30 40 50 60 70
CorrelationCoefficient
Time Step Difference (min)
Average Rainfall
R1
R2
R3
R4
R5
R6
• Rainfall vs. water level at the front pool
Correlation Coefficient Analysis
8
Time
Lag
Average
Rainfall
R1 R2 R3 R4 R5 R6
0 min 0.38 0.27 0.46 0.29 0.34 0.40 0.34
10 min 0.45 0.30 0.52 0.36 0.41 0.46 0.40
20 min 0.50 0.33 0.56 0.41 0.46 0.52 0.44
30 min 0.52 0.34 0.59 0.42 0.47 0.55 0.45
40 min 0.51 0.35 0.59 0.41 0.46 0.55 0.44
50 min 0.50 0.34 0.59 0.40 0.44 0.53 0.42
60 min 0.48 0.34 0.57 0.39 0.43 0.51 0.40
70 min 0.47 0.34 0.56 0.39 0.41 0.50 0.39

9. • RECOVER the water level of the pumping station
- Estimate the increased water levels based on the number of
running pumps and the actual water storage area of the
pumping station.
- Next, recover the front pool water level hydrograph.
9
Correlation Coefficient Analysis
 Estimate the actual water storage area
 Calculate the effect of the starting water level for pumps
 Calculate the corresponding number of running pumps each time
Sewer Area (m2) Fore Bay Area (m2) Flood Storage Area (m2)
163,008 1,650 164,658
Quantity Capacity (cms) Increased water level (m/10min)
7 26.3 0.096
4 12.5 0.046
0
2
4
6
8
10
12
14
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 100103106109112
Rainfall(mm)
Numbersofrunningpumps
Waterlevel(m)

10. • Rainfall and RECOVERED water level at the pumping station -
Correlation Coefficient Analysis
Correlation Coefficient Analysis
10
Time
Lag
Average
Rainfall
R1 R2 R3 R4 R5 R6
0 min 0.43 0.31 0.49 0.36 0.40 0.43 0.39
10 min 0.50 0.35 0.54 0.42 0.46 0.49 0.45
20 min 0.55 0.38 0.58 0.47 0.51 0.54 0.50
30 min 0.58 0.39 0.62 0.50 0.54 0.57 0.53
40 min 0.59 0.40 0.63 0.50 0.54 0.58 0.53
50 min 0.59 0.40 0.63 0.50 0.53 0.58 0.52
60 min 0.58 0.41 0.62 0.50 0.52 0.57 0.51
70 min 0.58 0.41 0.61 0.50 0.52 0.57 0.51
Cumulative
Time
Average
Rainfall
R1 R2 R3 R4 R5 R6
10 min 0.43 0.31 0.49 0.36 0.40 0.43 0.39
20 min 0.49 0.38 0.55 0.42 0.46 0.49 0.46
30 min 0.55 0.43 0.60 0.48 0.52 0.53 0.51
40 min 0.59 0.46 0.64 0.52 0.56 0.57 0.55
50 min 0.62 0.49 0.68 0.55 0.59 0.61 0.58
60 min 0.65 0.51 0.70 0.58 0.62 0.63 0.60
70 min 0.66 0.53 0.73 0.60 0.63 0.65 0.62
• Accumulated Rainfall and RECOVERED water level at the pumping
station - Correlation Coefficient Analysis

11. • Water levels of sewer outlets and the water level at the pumping
station
Correlation Coefficient Analysis
11
Before 2005
Time Lag YC2 YC3 YC4 YC5 YC6 YC7 YC8 YC9 YC10 YC11 YC12
0 min 0.82 0.59 0.95 -0.06 -0.16 0.78 0.87 0.94 0.98 0.89 0.51
10 min 0.81 0.58 0.95 -0.06 -0.17 0.82 0.87 0.93 0.97 0.89 0.52
20 min 0.8 0.57 0.94 -0.07 -0.17 0.85 0.86 0.92 0.95 0.88 0.53
30 min 0.78 0.56 0.92 -0.08 -0.17 0.86 0.85 0.91 0.93 0.86 0.54
40 min 0.76 0.54 0.9 -0.08 -0.18 0.85 0.83 0.89 0.9 0.85 0.55
50 min 0.74 0.51 0.88 -0.08 -0.18 0.84 0.82 0.87 0.88 0.83 0.57
60 min 0.72 0.48 0.86 -0.09 -0.19 0.82 0.8 0.85 0.85 0.81 0.58
70 min 0.69 0.45 0.84 -0.08 -0.2 0.8 0.79 0.83 0.82 0.8 0.59
Time Lag YC2 YC3 YC4 YC5 YC6 YC7 YC8 YC9 YC10 YC11 YC12
0 min 0.01 -0.14 0.03 -0.13 0.02 0.09 0.30 0.05 0.28 0.05 -0.08
10 min 0.01 -0.14 0.03 -0.12 0.03 0.10 0.30 0.05 0.27 0.05 -0.08
20 min 0.02 -0.13 0.03 -0.12 0.03 0.10 0.30 0.06 0.26 0.06 -0.07
30 min 0.03 -0.13 0.04 -0.11 0.04 0.11 0.30 0.06 0.25 0.06 -0.07
40 min 0.04 -0.12 0.04 -0.10 0.04 0.11 0.29 0.06 0.24 0.06 -0.06
50 min 0.04 -0.11 0.05 -0.10 0.05 0.12 0.29 0.07 0.23 0.07 -0.06
60 min 0.05 -0.11 0.05 -0.09 0.06 0.12 0.29 0.07 0.23 0.07 -0.06
70 min 0.06 -0.10 0.06 -0.08 0.07 0.12 0.29 0.08 0.22 0.08 -0.05
After 2005
- Water levels of sewer outlets can not be used as input factors

12. Input Selection
• Gamma Test (GT)：
• The GT (Agalbjörn et.al, 1997; Koncar, 1997) estimates
the noise level (Γ value) present in a data set.
• The GT can produce the estimation directly from the data
without assuming any parametric form of the equations
that govern the system. The only requirement is that the
system is smooth.
12

13. 100%
90% 88%
69%
63%
44%
25%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0
2
4
6
8
10
12
14
R2 Average
Rainfall
R5 R1 R6 R3 R4
Score
Frequency
F(d<d10) F(d>d90) 1-F(d>d90)/F(d<d10)
Input Selection
13
• Gamma Test：
• Select 6 rainfall station and average rainfall. (7 factors)
• Blue Bar: the occurrence frequency of a factor in the best results.
• Red Bar: the occurrence frequency of a factor in the worst results.
• Green Line: the score of the factors. (1-worst/best))
→ Select the factors (R2, AR, R5) higher than 80% to be input factors.

14. Model
• Water Level Forecast Model Construction:
• Artificial Neural Network:
• Back-Propagation Neural Network (BPNN)
Model Construction
14
Water level (t) ANN
ModelRainfall (t)
Water level
(t+n)
‧‧‧
1
n
1
1
R1 (t)
2
3
4
R2 (t)
R3 (t)
H (t)
H (t+1)
Input layer Hidden layer Output layer
n=10-60min

15. Model
• Water Level Forecast Model Construction:
• Artificial Neural Network:
• Elman Neural Network (Elman NN)
Model Construction
15
Water level (t) ANN
ModelRainfall (t)
Water level
(t+n)
Input layer Hidden layer Output layer
‧‧‧
1
n
1
1
R1 (t)
2
3
4
R2 (t)
R3 (t)
H (t)
H (t+1)
‧‧‧
1w1 (t)
wn (t)

16. Model
• Water Level Forecast Model Construction:
• Artificial Neural Network:
• The nonlinear autoregressive network with exogenous inputs
(NARX)
Model Construction
16
Water level (t) ANN
ModelRainfall (t)
Water level
(t+n)
Input layer Hidden layer Output layer
‧‧‧1
n
1
1
R1 (t)
2
3
4
R2 (t)
R3 (t)
H (t)
H (t+1)
1H1 (t+n)
H1 (t+1)
←

17. Results and Discussion
17
200 400 600 800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Waterlevel(m)
Training
100 200 300 400 500 600
Time (10min)
Validation
100 200 300 400 500
Testing
Obeserved
Predicted
200 400 600 800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Waterlevel(m)
Training
100 200 300 400 500 600
Time (10min)
Validation
100 200 300 400 500
Testing
Obeserved
Predicted
200 400 600 800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Waterlevel(m)
Training
100 200 300 400 500 600
Time (10min)
Validation
100 200 300 400 500
Testing
Obeserved
Predicted
NARX
10-min-ahead
30-min-ahead
50-min-ahead

18. Results and Discussion
18
• Compare BPNN and NARX, t+6 (60-min-ahead)
200 400 600 800
1
2
3
4
5
6
Waterlevel(m)
Training
200 400 600
Validation
100 200 300 400 500
Testing
Obeserved
Predicted
200 400 600 800
1
2
3
4
5
6
Waterlevel(m)
200 400 600 100 200 300 400 500
Time (-10min)
BPNN
NARX

19. 20
Results and Discussion
0.09
0.16
0.19
0.22
0.24 0.24
0.10
0.16
0.19
0.23
0.25
0.26
0.09
0.14
0.19
0.21
0.23 0.23
0.00
0.05
0.10
0.15
0.20
0.25
0.30
t+1 t+2 t+3 t+4 t+5 t+6
RMSE(m)
BPNN ELMAN NARX
BPNN ELMAN NARX
Time
lag
RMSE
(m)
CE Gbench
RMSE
(m)
CE Gbench
RMSE
(m)
CE Gbench
10 min 0.095 0.93 0.02 0.095 0.93 0.01 0.087 0.94 0.18
20 min 0.156 0.80 0.06 0.155 0.80 0.07 0.145 0.83 0.19
30 min 0.190 0.70 0.17 0.189 0.70 0.18 0.188 0.70 0.19
40 min 0.219 0.59 0.19 0.229 0.55 0.11 0.214 0.61 0.21
50 min 0.237 0.52 0.20 0.251 0.46 0.11 0.227 0.55 0.25
60 min 0.245 0.48 0.23 0.262 0.41 0.12 0.228 0.55 0.32

20. 21
Results and Discussion
0.94
0.83
0.70
0.61
0.55 0.55
0.40
0.50
0.60
0.70
0.80
0.90
1.00
t+1 t+2 t+3 t+4 t+5 t+6
CE
BPNN ELM NARX
BPNN ELMAN NARX
Time
lag
RMSE
(m)
CE Gbench
RMSE
(m)
CE Gbench
RMSE
(m)
CE Gbench
10 min 0.095 0.93 0.02 0.095 0.93 0.01 0.087 0.94 0.18
20 min 0.156 0.80 0.06 0.155 0.80 0.07 0.145 0.83 0.19
30 min 0.190 0.70 0.17 0.189 0.70 0.18 0.188 0.70 0.19
40 min 0.219 0.59 0.19 0.229 0.55 0.11 0.214 0.61 0.21
50 min 0.237 0.52 0.20 0.251 0.46 0.11 0.227 0.55 0.25
60 min 0.245 0.48 0.23 0.262 0.41 0.12 0.228 0.55 0.32

21. 22
Results and Discussion
BPNN ELMAN NARX
Time
lag
RMSE
(m)
CE Gbench
RMSE
(m)
CE Gbench
RMSE
(m)
CE Gbench
10 min 0.095 0.93 0.02 0.095 0.93 0.01 0.087 0.94 0.18
20 min 0.156 0.80 0.06 0.155 0.80 0.07 0.145 0.83 0.19
30 min 0.190 0.70 0.17 0.189 0.70 0.18 0.188 0.70 0.19
40 min 0.219 0.59 0.19 0.229 0.55 0.11 0.214 0.61 0.21
50 min 0.237 0.52 0.20 0.251 0.46 0.11 0.227 0.55 0.25
60 min 0.245 0.48 0.23 0.262 0.41 0.12 0.228 0.55 0.32
0.02
0.06
0.17
0.19
0.20
0.23
0.01
0.07
0.18
0.11 0.11
0.12
0.18
0.19 0.19
0.21
0.25
0.32
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
t+1 t+2 t+3 t+4 t+5 t+6
Gbench
BPNN ELM NARX
 
 






 N
t
ii
N
t
ii
dd
yd
1
2
1
1
2
1Gbench =

22. Conclusions
23
Intelligent real-time water level forecast models are
developed to forecast the 10-60 min-ahead front pool
water levels by utilizing the current rainfall and water
level.
The results indicate that all of the forecasts are good,
which can well capture the trend of the flooding
hydrograph.
The NARX network produces the best performance in
terms of RMSE, CE and G-bench values.

23. 24
Department of Bioenvironmental Systems Engineering, National Taiwan University
3rdANNOUNCEMENT
PAWEES2013
The12thConferenceof
InternationalSocietyofPaddyandWaterEnvironmentEngineering
THANK YOU FOR
YOUR ATTENTION