urban water demand forecasting using artificial neural networks: a case study of bangkok. by victor...
Post on 31-Dec-2015
217 Views
Preview:
TRANSCRIPT
URBAN WATER DEMAND URBAN WATER DEMAND FORECASTING FORECASTING USING ARTIFICIAL NEURAL USING ARTIFICIAL NEURAL NETWORKS: NETWORKS: A CASE STUDY OF BANGKOK.A CASE STUDY OF BANGKOK.BY
VICTOR SHINDEVICTOR SHINDE
CONTENTSCONTENTS
Need for water demand forecasting Description of ANN Study area description Results of the study Conclusions
INTRODUCTIONINTRODUCTION
Need for water demand forecastingNeed for water demand forecasting
• Water is a finite resource.
• Expanding the capacity of a water distribution system.
• Improving the reliability of supply.
• Effecting demand management instruments.
• Procurement of investment.
DESCRIPTION OF ANNDESCRIPTION OF ANN
Why use ANN?
• Accounts for non linearity between inputs and outputs
• Uses a universal function to convert inputs to output,
for all types of problems.
• More realistic forecasts.
DESCRIPTION OF ANNDESCRIPTION OF ANN A network which mimics the human brain.
Input layer Hidden Layer Output layer
INPUT
OUTPUT
Connection links each having some weight ‘w’
w1j
w2j
w3j
x1
x2
x3
j
yj = f(x1w1j+x2w2j+x3w3j)
f = Transfer function = (1/(1+e-t)
DESCRIPTION OF ANNDESCRIPTION OF ANN Network Training
Input layer Hidden Layer Output layer
INPUT
Output Desired Output
Compares
Cost function E = 2)( do YY
STUDY AREA DESCRIPTIONSTUDY AREA DESCRIPTION
Main study area
Metropolitan Waterworks Authority (MWA) responsibility area – Bangkok Metropolis, Nontaburi & Samut Prakarn
Secondary study areas
Hanoi (Vietnam) and Chiang Mai (Thailand)
STUDY AREA DESCRIPTIONSTUDY AREA DESCRIPTION
MWA responsibility area (2007 statistics)• Population : 7.86 Million
• Population served: 7.36 Million (93.6%)
• Average Daily production : 5.52 MCM
• Non Revenue Water : 30.32 % Secondary study areasSecondary study areas
Hanoi Chiang Mai
Temperature (0C)Maximum 32 31Minimum 14 19Rainfall (mm) 1682 1081Population (Million) 3.4 1.66GDP (USD) 40 Billion 3.3 Billion
Overall objectiveOverall objectiveTo develop ANN models to forecast the water demand for MWA – Bangkok.
Specific objectivesSpecific objectives● To forecastforecast the short term and long term water demand for MWA.
● To identifyidentify the factors most crucial in determining the short term and long term water demands for MWA.
● To comparecompare the factors influencing long term demand for Bangkok, Hanoi and Chiang Mai
OBJECTIVES OF THE STUDYOBJECTIVES OF THE STUDY
Demand Short term (ST) demand – Daily demand, 1,2 & 3 days lead
Long term (LT) demand – Monthly demand, 1,2 & 6 months lead
Data usedST Demand – Historical demand (sales), Rainfall, RH, Mean Temp
LT Demand – Historical demand (sales), Population, GPP, Household
connections, Education status, Rainfall, RH, Max TempFor comparing the cities (Bangkok, Hanoi and Chiang Mai) – Same as LT Demand, only production in lieu of sales & Mean Temp instead of Max Temp
Software used – ANN NeuroSolutions
SCOPE OF THE STUDYSCOPE OF THE STUDY
Methodology for both ST and LT demand Methodology for both ST and LT demand modelsmodels
SCOPE OF THE STUDYSCOPE OF THE STUDY
Data Collection and Analysis
Input Selection
Model Training & Testing for 1st set
Sensitivity analysis
Omission of least sensitive variables
Training & Testing for 2nd set
• Correlation Matrix• Pruning & Construction
Architectures• MLP• GFF• RBFTransfer functions• Hyperbolic tan• SigmoidLearning Rules• Backward Descend• Conjugate Gradient
150 ST Models – 15 sets88 LT Models – 6 sets60 Comparison models
RESULTSRESULTS
Short term demand, 1 day lead
Input Selection : Zhang et al. (2006), Msiza et al. (2007)
Water Sales Max Temp Min Temp Rainfall RH Evaporation Mean TempWater Sales 1 0.38 0.55 0.39 0.31 0.36 0.52Max Temp 0.38 1 0.78 -0.06 0.19 0.94 0.91Min Temp 0.55 0.78 1 0.42 0.7 0.79 0.96Rainfall 0.39 -0.06 0.42 1 0.85 -0.16 0.27RH 0.31 0.19 0.7 0.85 1 0.17 0.54Evaporation 0.36 0.94 0.79 -0.16 0.17 1 0.88Mean Temp 0.52 0.91 0.96 0.27 0.54 0.88 1
Selected Variables: Mean Temperature, Rainfall and RH
RESULTSRESULTSObserved vs. Predicted Demand for Training Set
2.9
3.1
3.3
3.5
3.7
3.9
1 11 21 31 41 51 61
Exemplars
Dem
and
(MC
M)
Observed
Predicted
Observed vs. Predicted Demand for Testing Set
3.1
3.3
3.5
3.7
3.9
4.1
1 11 21 31 41 51 61
Exemplars
Dem
and
(MC
M)
Observed
Predicted
RESULTSRESULTS
AARE =
1001
1
xO
DO
N
N
i i
ii
RMSE =
2
12
1
))(1
( i
N
ii DO
N
(Threshold static)x = (n/N) x 100
Model Architecture Hidden PE's Transfer Learning AARE RMSELayers function Rule % MCM 0.50% 1% 2% 3% 4% 5%
ST-1A(1) MLP 1 13 tanh BD 1.19 0.054 30.30 56.06 78.79 92.42 96.97 100ST-1A(2) MLP 1 11 Sigmoid BD 1.19 0.054 36.36 54.55 78.79 92.42 96.97 100ST-1A(3) MLP 1 17 tanh CG 1.18 0.052 25.76 54.55 81.82 92.42 96.97 100ST-1A(4) MLP 1 17 Sigmoid CG 1.15 0.05 28.79 54.55 81.82 95.45 100 100ST-1A(5) MLP 2 20 &15 tanh BD 1.15 0.051 30.30 59.09 80.30 93.94 98.48 100ST-1A(6) MLP 2 21 &15 Sigmoid BD 1.19 0.055 28.79 56.06 78.79 92.42 96.97 100ST-1A(7) MLP 2 12&15 tanh CG 1.19 0.054 30.30 57.58 80.30 92.42 96.97 100ST-1A(8) MLP 2 12&15 Sigmoid CG 1.2 0.055 33.33 54.55 78.79 90.91 96.97 100ST-1A(9) MLP 3 20,10&5 tanh BD 1.12 0.05 30.30 59.09 81.82 92.42 100 100ST-1A(10) MLP 3 17,12 &8 tanh CG 1.13 0.049 27.27 53.03 80.30 93.94 100 100ST-1A(11) MLP 3 17,12 &9 Sigmoid CG 1.15 0.052 34.85 57.58 78.79 90.91 96.97 100
Threshold static
Zhang et al. (2008)Adamowski (2008)Ghiassi et al. (2007)Jain et al. (2000)
Observed vs. Predicted Demand for ST-1B(16)
3.15
3.25
3.35
3.45
3.55
3.65
1 11 21 31 41 51 61
Testing Exemplars
Dem
and
(MCM
)
Observed
Predicted
Model Architecture Hidden PE's Transfer Learning AARE RMSELayers function Rule % MCM 0.50% 1% 2% 3% 4% 5%
ST-1B(12) MLP 1 18 tanh BD 1.23 0.056 31.82 54.55 78.79 90.91 96.97 98.48ST-1B(13) MLP 1 16 Sigmoid BD 1.17 0.055 34.85 60.61 80.30 92.42 96.97 98.48ST-1B(14) MLP 1 14 tanh CG 1.26 0.057 27.27 53.03 80.30 90.91 96.97 100ST-1B(15) MLP 1 13 Sigmoid CG 1.2 0.054 28.79 54.55 77.27 90.91 96.97 100ST-1B(16) MLP 2 12&8 tanh BD 1.17 0.053 33.33 59.09 78.79 90.91 96.97 100ST-1B(17) MLP 2 12&8 Sigmoid BD 1.22 0.056 31.82 57.58 78.79 90.91 96.97 98.48ST-1B(18) MLP 2 12&8 tanh CG 1.23 0.056 30.30 54.55 80.30 92.42 96.97 98.48ST-1B(19) MLP 2 12&8 Sigmoid CG 1.21 0.056 33.33 56.06 78.79 93.94 96.97 100
Threshold static
SA for ST-1A models
0.00
0.02
0.04
0.06
0.08
0.10
0.12
HWD
Mea
nTe
mp
Rain
fall RH
Stan
dard
Dev
iatio
n(M
CM)
Input Variables: HWD, Mean Temp Input Variables: HWD, Mean Temp & Rainfall& Rainfall
RESULTSRESULTSObserved vs. Predicted Demand for Testing Set
3.15
3.25
3.35
3.45
3.55
3.65
3.75
3.85
1 11 21 31 41 51 61
Exemplars
Dem
and
(MCM
)
Observed
Predicted
Model Architecture Hidden PE's Transfer Learning AARE RMSELayers function Rule % MCM 0.50% 1% 2% 3% 4% 5%
ST-1C(20) MLP 1 42 tanh BD 1.31 0.058 23.96 48.96 78.13 91.67 96.88 98.96ST-1C(21) MLP 1 42 Sigmoid BD 1.27 0.057 29.17 55.21 78.13 91.67 96.88 100ST-1C(22) MLP 1 40 tanh CG 1.32 0.058 21.88 48.96 80.21 91.67 96.88 100ST-1C(23) MLP 1 42 Sigmoid CG 1.34 0.059 26.04 46.88 77.08 91.67 96.88 98.96ST-1C(24) MLP 2 15&12 tanh BD 1.29 0.058 28.13 50.00 77.08 92.71 96.88 100ST-1C(25) MLP 2 15&12 Sigmoid BD 1.26 0.056 25.00 56.25 78.13 91.67 97.92 100ST-1C(26) MLP 2 15&12 tanh CG 1.29 0.058 27.08 50.00 78.13 91.67 96.88 100ST-1C(27) MLP 2 15&12 Sigmoid CG 1.28 0.058 29.17 52.08 77.08 91.67 96.88 100
Threshold static
RESULTSRESULTSInput Variables: Only HWDInput Variables: Only HWD
Model Architecture Hidden PE's Transfer Learning AARE RMSELayers function Rule % MCM 0.50% 1% 2% 3% 4% 5%
ST-1D(29) MLP 1 24 tanh BD 1.33 0.058 23.40 46.81 76.60 91.49 96.81 98.94ST-1D(30) MLP 1 28 Sigmoid BD 1.24 0.056 25.53 52.13 77.66 89.36 97.87 100ST-1D(31) MLP 1 28 tanh CG 1.32 0.057 23.40 47.87 77.66 91.49 96.81 100ST-1D(32) MLP 1 28 Sigmoid CG 1.35 0.059 24.47 46.81 76.60 92.55 96.81 98.94ST-1D(33) MLP 2 15&11 tanh BD 1.32 0.059 27.66 50.00 75.53 93.62 96.81 100ST-1D(34) MLP 2 16 &12 Sigmoid BD 1.26 0.057 30.85 54.26 77.66 89.36 96.81 100ST-1D(35) MLP 2 16 &12 tanh CG 1.35 0.06 23.40 47.87 77.66 91.49 96.81 98.94ST-1D(36) MLP 2 15&12 Sigmoid CG 1.36 0.06 24.47 47.87 76.60 91.49 96.81 98.94
Threshold static
Input Variables: HWD -1, HWD -2.Input Variables: HWD -1, HWD -2.
RESULTSRESULTS
Model Architecture Hidden PE's Transfer Learning AARE RMSELayers function Rule % MCM 0.50% 1% 2% 3% 4% 5%
ST-1E(37) MLP 1 21 tanh BD 1.28 0.056 23.08 48.35 79.12 93.41 97.80 98.90ST-1E(38) MLP 1 21 Sigmoid BD 1.24 0.055 27.47 56.04 75.82 90.11 97.80 98.90ST-1E(39) MLP 1 19 tanh CG 1.29 0.057 28.57 48.35 79.12 92.31 97.80 98.90ST-1E(40) MLP 1 21 Sigmoid CG 1.26 0.055 25.27 48.35 81.32 93.41 97.80 100ST-1E(41) MLP 2 16 &12 tanh BD 1.28 0.057 26.37 47.25 78.02 92.31 97.80 98.90ST-1E(42) MLP 2 16 &12 Sigmoid BD 1.22 0.055 29.67 53.85 79.12 90.11 96.70 100
Threshold static
Model Architecture Hidden PE's Transfer Learning AARE RMSELayers function Rule % MCM 0.50% 1% 2% 3% 4% 5%
ST-1F(43) MLP 1 10 tanh BD 1.27 0.056 23.26 50.00 79.07 93.02 97.67 100ST-1F(44) MLP 1 9 Sigmoid BD 1.26 0.055 24.42 55.81 77.91 93.02 98.84 100ST-1F(45) MLP 2 16&11 tanh BD 1.31 0.057 19.77 47.67 76.74 91.86 97.67 100ST-1F(46) MLP 2 18&12 Sigmoid BD 1.22 0.055 30.23 53.49 81.40 91.86 97.67 100
Threshold static
Input Variables: HWD -1, HWD -2 & HWD -3.Input Variables: HWD -1, HWD -2 & HWD -3.
Input Variables: HWD -1 through HWD -7Input Variables: HWD -1 through HWD -7
Master model for seven consecutive day Master model for seven consecutive day forecastforecastInput variables: HWD, Rainfall, Mean Temperature & Input variables: HWD, Rainfall, Mean Temperature & RHRHModel Architecture Hidden PE's Transfer Learning
Layers function Rule D+1 D+2 D+3 D+4 D+5 D+6 D+7 AverageAARE 1.32 1.66 1.87 2.00 1.97 2.02 1.96 1.83RMSE 0.058 0.076 0.083 0.084 0.086 0.086 0.081 0.08AARE 1.29 1.64 1.87 2.03 1.98 2.05 2.01 1.84RMSE 0.059 0.076 0.084 0.086 0.087 0.087 0.083 0.080AARE 1.31 1.65 1.88 2.04 2.02 2.08 2.04 1.86RMSE 0.058 0.076 0.084 0.085 0.088 0.088 0.085 0.080AARE 1.32 1.65 1.89 2.03 2.00 2.07 2.04 1.86RMSE 0.059 0.076 0.085 0.085 0.088 0.087 0.085 0.081AARE 1.17 1.52 1.80 1.83 1.82 1.83 1.85 1.69RMSE 0.053 0.068 0.078 0.076 0.076 0.077 0.076 0.072AARE 1.30 1.61 1.84 1.97 1.95 2.03 1.99 1.81RMSE 0.059 0.075 0.082 0.083 0.085 0.086 0.082 0.079AARE 1.33 1.63 1.87 2.06 2.01 2.03 1.96 1.84RMSE 0.058 0.075 0.083 0.086 0.088 0.086 0.081 0.080AARE 1.32 1.64 1.87 2.01 1.99 2.04 2.03 1.84RMSE 0.060 0.076 0.084 0.084 0.087 0.086 0.084 0.080
BDSigmoid
BDtanh
CGSigmoid
CGtanh
BDSigmoid
BDtanh
CGSigmoid
CGtanh
MM-4
MM-3
81MLPMM-2
MLP
MLP
1
1
MM-8
MM-7
MM-6
MM-5
MLP
MLP
MLP
MLP
8
14
2
2
2
2
8 & 7
10 & 7
8 & 7
9 & 7
Error for seven consecutively forecasted seven days
MM-1 1MLP 10
Date Observed Forecasted ARE Avg ARE RMSE Avg RMSEDemand Demand
MCM MCM % % MCM MCM
28-Jun-08 3.502 3.508 0.17 0.00629-Jun-08 3.481 3.489 0.22 0.00830-Jun-08 3.433 3.475 1.22 0.0421-Jul-08 3.410 3.434 0.72 0.0252-Jul-08 3.391 3.431 1.16 0.0393-Jul-08 3.413 3.426 0.37 0.0134-Jul-08 3.471 3.431 1.17 0.041
13-Aug-08 3.406 3.333 2.15 0.07314-Aug-08 3.416 3.334 2.39 0.08215-Aug-08 3.425 3.332 2.72 0.09316-Aug-08 3.410 3.358 1.51 0.05117-Aug-08 3.388 3.358 0.90 0.03018-Aug-08 3.371 3.349 0.64 0.02219-Aug-08 3.391 3.352 1.16 0.039
20-Jul-08 3.450 3.320 3.79 0.13121-Jul-08 3.432 3.323 3.20 0.11022-Jul-08 3.450 3.321 3.74 0.12923-Jul-08 3.425 3.359 1.92 0.06624-Jul-08 3.460 3.357 2.99 0.10425-Jul-08 3.510 3.345 4.70 0.16526-Jul-08 3.452 3.347 3.04 0.105
0.72 0.025
TEST SAMPLE - 1
TEST SAMPLE - 2
TEST SAMPLE - 2
0.1163.34
0.0561.64
Best fit models for Short term Demand
RESULTSRESULTS
Lead Period Input variables Architecture Accuracy• MLP - 3 layers
HWD, Rainfall, • tanh transfer functionMean Temp, RH • 20, 10 & 5 PE's
• Back Descend Rule• MLP - 2 layers• Sigmoid transfer function
HWD-1, HWD-2 • 16 & 10 PE's• Back Descend Rule• MLP - 2 layers• tanh transfer function
HWD-1 • 13 & 6 PE's• Back Descend Rule• MLP - 3 layers
HWD, Rainfall, • tanh transfer functionMean Temp, RH • 8, 7 & 7 PE's
• Backward Descend Rule
7 day consecutive
98.88%
98.53%
98.35%
98.51%
1 day
2 day
3 day
Lead Period Input variables Architecture AccuracyPopulation, GPP, Education • MLP - 3 layersstatus, Household connections • Sigmoid transfer functionHWD, Rainfall, Max Temp • 20, 10 & 8 PE's
• Conjugate Gradient RulePopulation, GPP, Education • GFF - 1 layerstatus, Household connections • tanh transfer functionHWD, Rainfall, RH • 11 PE's
• Conjugate Gradient RulePopulation, GPP, Education • MLP - 1 layerstatus, Household connections • tanh transfer functionHWD, Rainfall, Max Temp • 10 PE's
• Backward Descend Rule
1 month
2 month
6 month
98.88%
98.53%
98.35%
RESULTSRESULTS
Best fit models for Long term Demand
Factors influencing ST & LT Demand
Sensitivity Analysis
• Standardize all data
xx
Y = x & σ are the mean and standard deviation
• Increases and decreases the input variables between the standardized -1 and +1
• Thus a standardized value of ‘zero’ represents the mean of the sample
• Presents the trend of change in the demand.
RESULTSRESULTS
Sensitivity indices for ST-Demand
0.00
0.02
0.04
0.06
0.08
0.10
0.12
HWD
Mea
nTe
mp
Rainf
all RH
Stan
dard
Dev
iatio
n(M
CM)
Sensitivity indices for LT Demand
0.0
0.5
1.0
1.5
2.0
2.5
HWD
GPP
Popu
latio
n
Educ
ation
Stat
us
Hous
ehol
dCo
nnec
tions
Max
Tem
p
Rain
fall RHSt
anda
rd D
evia
tion
(MCM
)
Factors influencing ST & LT Demand
RESULTSRESULTS
20 models prepared for each city using MLP & GFF Input VariablesInput Variables: Population, GPP, Household : Population, GPP, Household
connections, Education status, HWD, Rainfall, Max connections, Education status, HWD, Rainfall, Max Temperature, RHTemperature, RH
Best fit model results, AARE
• Bangkok – 1.06%
• Hanoi – 2.18 %
• Chiang Mai – 1.26%
Sensitivity analysis to determine influencing variables
Demand models for Bangkok, Hanoi & Chiang Mai
RESULTSRESULTS
Demand models for Bangkok, Hanoi & Chiang Mai
Sensitivity indices of Input parameters for best models of the three study areas
01234567
HW
D
GP
P
Pop
ulat
ion
Edu
catio
n
Hou
seho
lds
Mea
n T
emp
Rai
nfal
l
RH
Per
cen
tag
e ch
ang
e in
Dem
and
Bangkok
Hanoi
Chiang Mai
RESULTSRESULTS
CONCLUSIONS
ANN can provide the MWA with a powerful instrument to forecast the demands. Forecasting accuracy will be over 98% for both ST & LT. Advantages for MWA• Schedule pumping operations• Reduce detention time to improve water quality• Monthly revenues can be estimated upto 6 months in advance• Diversions, Basin transfers can be planned in dry years
Factors Influencing MWA sales demandST Demand : Historical water demandLT Demand : Education status & Household connections
Comparison of factors influencing production demands of Bangkok, Hanoi and Chiang MaiBangkok : HH connections, GPP and EducationHanoi : Education status, Mean Temperature and PopulationChiang Mai : HH connections, Mean Temperature & Rainfall(This information could prove vital for goverments, international agencies and funding organizations)
top related