This document describes using an artificial neural network (ANN) model to predict groundwater levels 30 days in the future near a public well field in Montville Township, New Jersey. The ANN model uses inputs like daily pumping rates, precipitation, and temperature. Analysis of historical data showed climatic factors influence water levels over short periods. The ANN was trained on data from 1999-2001 and accurately predicted water levels in testing and validation data, outperforming a linear regression model. A sensitivity analysis found initial water level and precipitation were the most important predictors of future water levels. The conclusions state ANNs can accurately predict water levels for areas with limited data and do not require expensive aquifer tests.

1. A NEURAL NETWORK MODEL FOR
PREDICTING AQUIFER WATER
LEVEL ELEVATIONS
MODELING AND SIMULATION FOR AGRICULTURAL WATER
MANAGEMENT
(AG60170)
Land and Water Resources Engineering
Agricultural and Food Engineering Department
Shyam Mohan Chaudhary
17AG62R13
Janaki Ballav Mohapatra
17AG62R03

2. CONTENTS
 Artificial Neural Network (ANN)
 ANN vs Numerical Model
 Objectives
 Study Area
 Data Inputs
 Climatic Influence on water levels
 ANN performance
 Sensitivity Analysis
 Conclusions

3. Artificial Neural Network (ANN)
☻Artificial neural network (ANN) technology is a compelling
alternative modeling and prediction tool.
☻It learns the system behavior of interest by processing
representative data patterns through a mathematical
structure analogous to the human brain.

4. ☻ Three layers- Input, Hidden and
Output.
☻ Hidden layers can be more than
one.
☻ Each layer consists of nodes.
☻ Connections relay information
between layers.
☻ Each neuron has some weights.
☻ Output and Hidden layers have
additional bias nodes.
Artificial Neural Network (ANN) (Contd.)

5. ☻All inputs to a node are weighted, combined and then
processed through a transfer function (tanh or sigmoid) that
controls the strength of the output of that node.
☻Generally, data points are divided into three stages: Training,
Verification and Validation.
☻During training, data patterns are processed through the
ANN and the connection weights are adaptively adjusted by
using an algorithm until a minimum error was achieved.
Artificial Neural Network (ANN) (Contd.)

6. Artificial Neural Network (ANN) (Contd.)
Where,
xi = Input variable for the ith node
wjb = Bias
wji = connection weight between ith node in the input layer and jth node
in the hidden layer

7. ☻A variety of factors are considered while selecting the most
appropriate ANN model:
– Functional form of the ANN transfer functions
– Number of hidden layers
– Appropriate set of input variables
– Method used to minimize the objective function
Artificial Neural Network (ANN) (Contd.)

8. ☻ANN do not requires explicit characterization and
quantification of physical properties and conditions.
☻Does not require accurate representation of governing
physical laws.
☻The simplifying physical and mathematical assumptions of a
numerical model and imperfect characterization of the real
world system limit simulation and prediction accuracy.
ANN Vs Numerical Model

9. OBJECTIVES
 To use ANN to predict groundwater levels 30 days into the
future near a public supply well field
 To assess the predictive performance of ANN against linear
regression

10. STUDY AREA
☻Montville Township, New
Jersey
☻Area : 48.9 km2
☻Population: 20000
☻Three high capacity
production wells installed
by the Montville Water
And Sewer Department.

11. Discontinuous layer
which results in
direct hydraulic
connection between
two aquifers
Highly prolific and
Township’s only
public drinking
water source
STUDY AREA (Contd.)

12. DATA INPUTS
 Mean daily pumping rate of production well 1 over 30-d
period
 Mean daily pumping rate of production well 2 over 30-d
period.
 Mean daily pumping rate of production well 3 over 30-d
period.
 Cumulative mean pumping rates of three production wells
over 30-d period.

13. DATA INPUTS (Contd.)
 Total precipitation over 30-d period.
 Mean daily temperature over 30-d period.
 Initial water level measurement at monitoring well at the
beginning of 30-d period.

14. CLIMATIC INFLUENCE ON WATER LEVELS
☻Inorder to justify input variable selections for the ANN, a brief
analysis of a 3-year data period spanning from January 1999 to
December 2001 was first provided.
☻The data set consists of precipitation and temperature
measurements at a nearby climate station, water level
measurements in the two monitoring wells, and recorded
pumping extractions of the three production wells.
☻This analysis supports the assumption that climate conditions
do affect the potentiometric surface of the semiconfined aquifer
over relatively short time periods.

15. CLIMATIC INFLUENCE ON WATER LEVELS
(Contd.)
Total Monthly Ground Water Extraction vs. Mean Monthly Temperature
(January 1999 - December 2001)

16. CLIMATIC INFLUENCE ON WATER LEVELS
(Contd.)
Depth to water vs. Total Monthly Precipitation
(January 1999 to December 2001)

17. CLIMATIC INFLUENCE ON WATER LEVELS
(Contd.)
Depth to Water vs. Mean Monthly Temperature
(January 1999 to December 2001)

18. DATA COLLECTION
☻ The frequency of recorded water level measurements was about
three times daily, and the Montville Township Water and Sewer
Department also recorded total daily ground water extractions for
each production well.
☻ The pumping extractions and daily mean temperatures were
averaged, while daily precipitation was summed over 30-d period.
☻ In order to generate sufficient data sets, each consecutive 30-d
period was offset by 1 d.
☻ The number of patterns, each of which represents a distinct set of
input-output variables for a given stress period, indicates how
many 30-d input-output sets were used for each phase of ANN
development and assessment.

19. ANN PERFORMANCE
Water level elevations for testing data in the Indian Lane monitoring well

20. ANN PERFORMANCE (Contd.)
Water level elevations for validation data in the Cooks Lane monitoring well

21. ANN Vs Linear Regression (LR)
Comparison of ANN with LR at Indian Lane
monitoring well
The LR has a
tendency to smooth
its predictions with a
higher tendency to
either overpredict or
under predict water
levels as compared to
the ANN.

22. ANN Vs Linear Regression (LR) (Contd.)
These statistics demonstrate that the ANNs learned to accurately predict
relatively large dynamic water level changes, reproducing rising and falling
elevations in response to variable pumping and climate conditions.

23. SENSITIVITY ANALYSIS
☻Statistics used in sensitivity analysis for training period data
☻The error associated with each input variable is the root
mean square error (RMSE) if the particular variable is
eliminated.
☻The rank corresponding to the importance of input variable
as an accurate predictor.
☻Ratio for a given predictor variable:
☻

24. SENSITIVITY ANALYSIS (Contd.)

25. CONCLUSIONS
☻Using daily data spanning ~5 months, ANNs accurately
predicted rising and falling potentiometric surface elevations
30 d into the future under variable pumping and
climate conditions.
☻ANN technology can be used in data scarce areas and it does
not require expensive aquifer pumping tests which provide
average aquifer properties.
☻Data collection systems combined with ANN technology can
produce an accurate real-time prediction and management tool
for wellfields and other water resources systems.