You are on page 1of 10

Artificial Neural Network (ANN)

Modeling of Reservoir Operation


at Kainji Hydropower Dam, Nigeria

B. F. Sule(&), A. A. Mohammed, and A. W. Salami

Department of Water Resources and Environmental Engineering,


University of Ilorin, Ilorin, Nigeria
bfsuleiman@gmail.com, awsalami2009@gmail.com,
rasaq.muhammed@yahoo.com

Abstract. Hydropower dams are one of the many kinds of modifications that
human beings make to natural waterways, but they can have especially profound
ramifications for riverine ecosystems and the animals that live in them. The
operation of hydropower reservoirs in Nigeria does not take into cognizance the
ecological situation of the reservoir as it may affect aquatic life that depend on it.
This study is sought to model energy generation at Kainji hydropower reservoir,
Nigeria considering ecological flow release (EFR) scenarios. Two neural net-
work (NN) approaches were used to optimize the energy generation at the
station using the current operating rule and incorporation of EFR scenarios. The
study revealed that the correlation coefficients (r) for energy generation without
considering EFR using multilayer perceptron neural network (MLPNN) and
radial basis neural network (RBFNN) were 0.94 and 0.84, respectively. Average
annual energy generation at the station using MLPNN and RBFNN approaches,
without considering EFR were 223930 and 222926 MWh, respectively. The two
NN modeling approaches were found to perform well in simulating energy
generation. The appropriate reservoir operation rule stipulated that the EFR that
will not negatively affect energy generation should range between 10 to 25% of
reservoir storage.

Keywords: Energy generation  Kainji  Neural network  Reservoir operation

1 Introduction

Optimization of reservoir operation is vital to reservoir benefits and has been abun-
dantly investigated in the past. Conventional hydropower reservoir operation practices
is predominantly guided by maximization of socio-economic interests without due
consideration to well-being of aquatic life [1].
In [2] ANN was applied to optimize Karoon 5 hydropower reservoir operation in
southwest, Iran. A stochastic dynamic programming (SDP) simulation model in the
form of a computer code was developed for optimum operating rule of the hydropower
system. ANN was used to simulate the model. A study which modelled the hydropower
reservoir variables for energy generation using ANN for two hydropower dams along

© Springer Nature Switzerland AG 2019


K. Arai et al. (Eds.): SAI 2018, AISC 858, pp. 189–198, 2019.
https://doi.org/10.1007/978-3-030-01174-1_15
190 B. F. Sule et al.

the River Niger (Kainji and Jebba dams) in Nigeria was reported in [3] while author in
[4] used chaotic GA to optimize hydropower generation with ecological consideration.
Author in [5] evaluated the impact of climate change on runoff in the Kainji Lake
basin using ANN. Hydro-meteorological data used in this study included minimum and
maximum temperature, evaporation, precipitation, runoff and water level were sub-
jected to ANN. The model results revealed a positive relationship between the actual
and forecasted runoff for all the selected locations and their correlation coefficients were
0.62, 0.57, 0.55 and 0.57 for Lokoja, Kaiji, Baro and Idah, respectively. The study
revealed that climate change had positive impact on the reservoir inflow at Kainji dam
and subsequently assured more water for hydropower generation.
In [6], author modelled the reservoir inflow for hydropower dams in Nigeria using
ANN while in [7] applied ANN to model the management of hydropower reservoirs
along river Niger, Nigeria. The model was trained with monthly inflow, outflow
(release), storage and evaporation loses data of Jebba and Kainji hydropower reser-
voirs. The trained model yielded 95% and 97% of good forecast of training and testing
set for Jebba and 69% and 75% respectively for Kainji reservoir. In [8], author used
neural network based model for forecasting reservoir storages for hydropower dams
operation. The networks were created and trained with monthly reservoir operation
data. The trained networks yielded 95 and 97% of goodness of fit respectively for
training and testing of data at Jebba, 69 and 75% at Kainji and 98 and 97% at Shiroro.
The optimal operation policies of Konar reservoir in India using ANN was studied by
[9]. Five different ANN models were developed. Out the developed models, three
simulated the final storage while two simulated optimal release. Results revealed that
ANN model with current period storage and inflow as well as previous period inflow as
input, yielded good results close to stochastic model.
In most of the relevant literatures reviewed ecological flow releases (EFR) in
hydropower reservoir operations were not considered. This paper looks at how con-
straints on ecological releases affect reservoir operations policies and subsequently the
energy output from a hydropower project. Artificial neural network was used to obtain
operation policies that maximizes energy generation.

2 Study Area

The Kainji dam is located in New Bussa, Borgu Local Government Area of Niger
State, Nigeria. Kainji hydropower reservoir is fed by many tributaries such as Malando,
Danzaki and Sokoto-Rima rivers. It lies at an altitude of 108 m above sea level,
between Yelwa (latitude 10°53′N: longitude 4°45′E) and Kainji (latitude 9°50′N:
longitude 4°35′E). The reservoir that resulted from Kainji dam was built between 1964
and 1968 and commenced operation in 1968 for the purpose of generating electricity
[10]. The maximum water surface elevation is 141.9 mean above sea level (masl).
Kainji hydropower reservoir has an installed capacity of 760 MW. Kainji reservoir is
characterized by prolonged high temperature, low rainfall and low relative humidity; it
exhibits evaporation values that are in excess of rainfall [11]. Figure 1 shows the
location of the dam on a map of Nigeria.
Artificial Neural Network (ANN) Modeling of Reservoir Operation 191

Fig. 1. Location of study area.

3 Reservoir Operation Optimization

3.1 Objective Function


The annual hydropower energy generation E (MWh) is used as the objective function
in this study. This is presented in (1) according to [12].

X
12
Z ¼ max Et ð1Þ
t¼1

where:
Z = total annual energy generation (MWh)
E = monthly energy generation (MWh)

But energy can be estimated using (2).

E ¼ 2:73Rt Ht e ð2Þ

Substituting (1) in (2), resulting in the objective shown in (3):

X
12
Z ¼ max 2:73 Rt Ht e ð3Þ
t¼11

where:
Ht = reservoir level at time t, (m)
Rt = turbine release for period, t (Mm3)
ɛ = efficiency of the hydropower plant.
192 B. F. Sule et al.

3.2 Model Constraints


The constraints are:

3.2.1 Reservoir Mass Balance Equation


The mass balance between the inflows and outflow is given in (4).

St þ 1 ¼ St þ It  Rt  Lt  G t ¼ 1 to 12 ð4Þ

where:

St+1 = final storage (initial storage of the next season) for period t + 1, (Mm3)
St = initial storage for period t, (Mm3)
It = reservoir inflow for period t, (Mm3)
Rt = turbine release for period, t (Mm3)
Lt = evaporation losses from the reservoir period t, (Mm3)
G = excess release from the reservoir for period, t (Mm3)
t = time (month).

3.2.2 Reservoir Storage and Water Level Constraints


The water stored in the reservoir St (Mm3) should always be bounded by dead storage,
St min (Mm3) and reservoir maximum storage capacity St max (Mm3) is presented in (5)
while the corresponding reservoir water level, Ht (m) is restrained by the dead water
level Hmin (m) and the maximum water level Hmax (m) as shown in (6).

St min  St  St max t ¼ 1 to 12 ð5Þ

Ht min  Ht  Ht max t ¼ 1 to 12 ð6Þ

where:

S t min = minimum reservoir capacity at any time t, (Mm3)


S t max = maximum reservoir capacity at any time t, (Mm3)
St = reservoir storage at time t, (Mm3)
H t min = minimum reservoir level at time t, (m)
H t max = maximum reservoir level at time t, (m)
Ht = reservoir water level at time t, (m).

3.2.3 Turbine Release Constraints


The release through the turbines, Rt (m3/s) for energy generation should never exceed
the turbine capacity and should not be negative as presented in (7).
Artificial Neural Network (ANN) Modeling of Reservoir Operation 193

Rt min  Rt  Rt max t = 1 to 12 ð7Þ

where:

Rt min = minimum turbine release at any time t, (Mm3)


Rt max = maximum turbine release at any time t, (Mm3)
Rt = turbine release at any time t, (Mm3).

3.2.4 Plant Output Constraints


In order to maximize the plant efficiency, the real output of power plant Nt
(MW) should be between the firm power output Nf (MW) and the installed capacity Ni
(MW) at all times. It is mathematically written as in (8).

Nf  Nt  Ni t = 1 to 12 ð8Þ

where:

Nt = real output power at any time t, (MW)


Nf = output power at any time t, (MW)
Ni = installed capacity at any time t, (MW).

3.2.5 Ecological Flow Constraints


In order to protect the fish habitats in the river channel, a time varying ecological flow
requirement, EFRt (m3/s) must be maintained as shown in (9).

Rt  EFRt t = 1 to 12 ð9Þ

where:
Rt = Release at any time t, (Mm3)
EFRt = Ecological flow requirement at any time t, (Mm3).

For the EFRt, 5%, 10%, 15%, 20% and 25% of the reservoir storage were used as
EFR scenarios.
For effective management of Kainji hydropower reservoir operation there is need
for sufficient understanding of the relationship that exists between reservoir operation,
ecological flow and energy generation. In solving the formulated model for the
hydropower reservoir operation, MLPNN and RBFNN approaches in SPSS 16.0 were
used. The networks were trained with the following monthly historical data of Kainji
hydropower reservoir operation: inflow, turbine release, storage, reservoir elevation,
tailrace water level, generating head, plant use coefficient and evaporation as input
variables while energy generation was the output variable. Scenarios of ecological flow
release (EFR) as a percentage of reservoir storage were introduced in the model and
resulting energy generation was simulated. The data were split into training/calibration
and validation set. About 80% of the data was used for model calibration while about
194 B. F. Sule et al.

20% was used for model validation. The training set was used to train the network
whereas the validation set was used to monitor the network performance at regular
stages during the training. The training stopped when the error on the validation set
reached the minimum. The performance of the networks was evaluated using RMSE,
MRE and correlation coefficient as presented in equations below:
P 
ypi  ypi ðyoi  yoi Þ
r ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 ffi ð10Þ
ypi  ypi ðyoi  yoi Þ2

" #1
1X n  2 2
RMSE ¼ ypi  yoi ð11Þ
n i¼1

n  
1X  ypi  yoi 
MRE ¼ ð12Þ
n i¼1  yoi 

where:

ypi = predicted parameter


yoi = observed parameter
ypi = mean predited parameter
yoi = mean observed parameter
nP = total number of observations
= summation

4 Results and Discussion

4.1 Reservoir Operation Optimization


The results of reservoir operation optimization for energy generation considering the
current operational procedures and also incorporating scenarios of various EFRs are
presented. Results of the performance evaluation of the two NN approaches for the
energy generation are shown in Table 1. Summary of actual and modelled annual
energy generation with scenarios of EFR is shown in Table 2. A typical graph of actual
and modelled energy generation with EFR as 10% of reservoir storage is shown in
Fig. 2. The results of a typical MLPNN and RBFNN model architectures for the energy
generation using SPSS 16.0 software version are presented in Figs. 3 and 4.

4.2 Discussion of Results


The reservoir operation optimization for energy generation considering the current
operational procedures and also incorporating scenarios of various EFRs were mod-
elled using MLPNN and RBFNN modeling approaches. The percentages of data used
in model calibration and validation were over 70% and 20%, respectively for all the
Artificial Neural Network (ANN) Modeling of Reservoir Operation 195

Table I. Training and validation results for MLPNN and RBFNN models
EFR as % of Model Training Validation Training Validation
reservoir performance (MLPNN) (MLPNN) (RBFNN) (RBFNN)
storage parameter
0 RMSE 3.817 3.486 7.289 5.385
MRE 0.086 0.161 0.312 0.263
r 0.94 0.84
5 RMSE 3.997 3.813 7.186 2.135
MRE 0.095 0.193 0.288 0.84
r 0.94 0.84
10 RMSE 3.997 3.813 7.186 4.799
MRE 0.095 0.193 0.288 0.344
r 0.94 0.83
15 RMSE 3.997 3.813 7.186 4.799
MRE 0.095 0.193 0.288 0.344
r 0.94 0.83
20 RMSE 3.997 3.813 7.664 4.856
MRE 0.095 0.193 0.345 0.344
r 0.94 0.81
25 RMSE 3.997 3.813 7.186 4.799
MRE 0.095 0.193 0.285 0.344
r 0.94 0.83

Table II. Average annual actual and model energy generation with EFR scenarios
EFR as Actual MLPNN RBFNN % Diff. % Diff.
% of energy energy output energy output MLPNN to RBFNN To
storage output (MWh) (MWh) actual actual
(MWh)
0 223047 223930 222926 0.396 -0.054
5 223047 223930 222926 0.396 -0.054
10 223047 223577 224981 0.238 0.867
15 223047 223577 224981 0.238 0.867
20 223047 223577 223253 0.238 0.092
25 223047 223577 224981 0.238 0.867

scenarios. The correlation coefficients (r) for energy generation without considering
EFR using MLPNN and RBFNN were: 0.94 and 0.84 respectively. The plots of the
actual and simulated energy generation using the two NN approaches revealed a strong
relationship between the actual and simulated energy generation station. The values of
RMSE for training and validation were 3.817 and 3.486 using the MLPNN approach
while that of RBFNN were 7.289 and 5.385 respectively. Also the MRE for training
and validation were 0.086 and 0.161 using MLPNN while that of the RBFNN were
0.312 and 0.263 respectively. Average annual energy generation at the station using
MLPNN and RBFNN approaches without considering EFR was 223930 and 222926
196 B. F. Sule et al.

600000
Energy generation (MWh)
Actual energy generation (MWh)
MLPNN energy generation (MWh)
500000 RBFNN energy generation (MWh)

400000
300000
200000
100000
0

273
290
307
324
341
358
375
392
409
426
443
460
477
494
1
18
35

120
52
69
86
103

137
154
171
188
205
222
239
256
Time (month)

Fig. 2. Actual and modelled energy when EFR is 10% of storage.

Fig. 3. MLPNN model architecture for EFR = 0.

MWh, respectively. MLPNN modeling approach will increase the energy generation by
0.396% while RBFNN will decrease the energy generation by 0.054%. Network
information showed 8 hidden layers for the MLPNN (Fig. 3) while RBFNN had 5
hidden layers (Fig. 4).
The correlation coefficients (r), RMSE and MRE for training and validation con-
sidering all the EFR scenarios using MLPNN were found to be the same as 0.94, 3.997,
3.813, 0.095 0.193. The RMSE and MRE for training and validation considering all
5%, 10% and 25% EFR scenarios using RBFNN approach were found to be the same
Artificial Neural Network (ANN) Modeling of Reservoir Operation 197

Fig. 4. RBFNN model architecture for EFR = 0.

as 7.186 and 4.799 for the training and validation respectively. Correlation coefficients
(r) for all the scenarios varied from 0.81 to 0.84. The plots of the actual and simulated
energy generation using the two NN approaches also indicated strong relationship
between the actual and simulated energy generation at the station. Average annual
energy generation at the station using MLPNN approach with 10% to 25% EFR
scenarios were found to be the same and increased the energy generation by 0.238%.
While average annual energy generation at the station using RBFNN approach with
10%, 15% and 25% EFR scenarios were found to be the same and increased the energy
generation by 0.867%. Average annual energy generation at the station using MLPNN
approach with 5% EFR scenario increased the energy generation by 0.092% while
RBFNN decreased it by 0.054%. The results of the RMSE and MRE considering all the
EFR scenarios for energy generation at the station are comparable with RMSE values
of 0.33 to 23.6 and MRE values of 0.15 to 15.23 obtained in similar studies as reported
in [13] and [14].

5 Conclusion

This study is on the assessment and modeling the impact of hydropower reservoir
operations on energy output at the Kainji Dam, Nigeria using artificial neural network
(ANN) models. Two ANN modeling approaches: MLPNN and RBFNN were adopted
in this study. The two modeling approaches were found to perform well in simulating
energy generation while the consideration of EFR as part of the reservoir operating rule
will not have negative impact on the energy generation. This is so because the two
198 B. F. Sule et al.

ANN models showed similar energy output when compared to the conventional rules at
the power station.

References
1. Chen, Q., Chen, D., Li, R., Ma, J., Blanckaert, K.: Adapting the operation of two cascaded
reservoirs for ecological flow requirement of a de-watered river channel due to diversion-
type hydropower stations. Ecol. Model. 252, 266–272 (2013)
2. Haddad, O.B., Alimohammadi, S.: Evaluation of artificial neural network in optimization
models of hydropower reservoir operation. In: Ninth International Water Technology
Conference, IWTC9, Sharm El-Sheikh, Egypt, pp. 985–998 (2005)
3. Abdulkadir, T.S., Salami, A.W., Anwar, A.R., Kareem, A.G.: Modelling of hydropower
reservoir variables for energy generation: neural network approach. Ethiop. J. Environ. Stud.
Manag. 6(3), 310–316 (2013)
4. Huang, X., Fang, G., Gao, Y., Dong, Q.: Chaotic optimal operation of hydropower station
with ecology consideration. Energy Power Eng. 2, 182–189 (2013)
5. Salami, A.W., Mohammed, A.A., Okeola, O.G.: Evaluation of climate change impact on
runoff in the kainji lake basin using artificial neural network model (ANN). Malays. J. Civil
Eng. 26(1), 35–50 (2014)
6. Salami, A.W., Mohammed, A.A., Adeyemo, J.A., Olanlokun, O.K.: Modeling of reservoir
inflow for hydropower dams using artificial neural network. Niger. J. Technol. (NIJOTECH)
34(1), 28–36 (2015)
7. Abdulkadir, T.S. Sule, B.F., Salami, A.W.: Application of artificial neural network model to
the management of hydropower reservoirs along river Niger, Nigeria. Ann. Fac. Eng.
Hunedoara – Int. J. Eng. 10, 419–424 (2012)
8. Abdulkadir, T.S., Salami, A.W., Sule, B.F., Adeyemo, J.A.: Neural network based model for
forecasting reservoir storage for hydropower dam operation. Int. J. Eng. Res. Gen. Sci. 3(5),
639–647 (2015)
9. Shaikh, S.A.: Optimal operation of single reservoir using artificial neural network. Int.
J. Civil Eng. Technol. (IJCIET) 6(6), 124–132 (2015)
10. Ifabiyi, I.P.: Contributions of reservoir elements to monthly electricity generation in the
Jebba hydropower reservoir, Nigeria. Ozean J. Appl. Sci. 4(3), 251–264 (2011)
11. Abam, T.K.S.: Modification of Niger delta physical ecology: the role of dams and reservoirs.
In: Hydrology and Aquatic Ecology, Proceedings of Workshop Organized by IAHS, UK,
pp. 19–29 (2001)
12. Salami, A.W.: Operational performance of water management models for a hydropower
reservoir system under reservoir inflow forecast. A Ph.D. thesis Submitted to the Department
of Civil Engineering, University of Ilorin, Ilorin, Nigeria (2007, unpublished)
13. Giri, A., Singh, N.B.: Comparison of artificial neural network algorithm for water quality
prediction of river Ganga. Environ. Res. J. 8(2), 55–63 (2014)
14. Chokananporn, W., Tansakul, A.: Artificial neural network model for estimating the surface
area of fresh guava. Asian J. Food Agro-Ind. 1(03), 129–136 (2008)

You might also like