Computational tools to aid groundwater management ���Reetik Kumar Sahu�GRBEPM Seminar 30/06/2021

1

Physical System model

• Physics based
• Data driven (machine learning)

Objective criteria

• Sustainability
• Remediation
• Water /Energy /Food security

Uncertainty

• Climate variability
• QA/QC measurements
• Downscaling
• Extreme events

Optimal decision rule

• Social vs non-cooperative
• Myopic vs Long term
• Real-time vs ad-hoc decision rules

Other Decision Makers

• Shared watersheds
• Water rights and market

Risk Assessment

• Identification of critical areas
• Preparedness
• Guidance for future

Developing decision-support using computational tools for sustainability

Part 1: Machine Learning in Hydrology

3

ML provides ‘computationally-cheap’ decision-support for water regulations

• Sustainable Groundwater Management Act, 2014
• For high and medium priority basins: halt overdraft and bring groundwater basins into balanced levels of pumping and recharge
• Decision makers need to manage groundwater without

Need for practical decision tools for groundwater management

Current obstacles:

• Uncertainty in future weather leads to uncertainty in amount of available water
• Uncertainty in groundwater demands and pumping

​

Prepare for what-if scenarios:

• How will groundwater levels change under another drought scenario
• Identify new watersheds in risk

​

​

drought

?

Cumulative Precipitation (mm)

?

Groundwater level (meters)

Current simulation tools are hard to use !

High Fidelity models [MODFLOW, PARFLOW]

(Physics + observations)

Requires

• Extensive characterization
• High computing resources
• Long run time

​

Machine Learning

(Observations)

• Computationally cheap surrogate models
• Can run multiple future scenarios
• Can run on laptop

​

Data preparation and Site selection is non trivial

Salinas Basin

​

GW data across various wells

0

-20

-40

-120

-140

-80

-60

-100

Depth to water table (ft)

‘04

‘08

‘06

‘10

‘18

‘12

‘14

‘16

Year

Butte County

GW data from an observation well

Year

Depth to water table (ft)

-40

-45

-50

-55

-60

-65

-70

‘10

‘11

‘12

‘13

‘14

‘15

‘16

‘17

‘18

Groundwater level prediction using past observations

Input

Hidden layer

Output

t-n, t-(n-1),…., t-2, t-1, t

​

Past observations:

• Groundwater
• Temperature
• Precipitation
• Streamflow
• Evapotranspiration

t+1

Groundwater

level

Multi Layer Perceptron(MLP) vs Recurrent Neural Network(RNN)

Components

• Neurons
• Connections, weights, biases
• Propagation function
• Learning rule

Recurrent Neural Network(RNN) vs Long Short Term Memory Network (LSTM)

10

Components

• Cell
• Input gate
• Output gate
• Forget gate

Lots of options to choose from Machine Learning models

- 11 -

Choice of Neural network

• Multi-Layer Perceptron
• Long Short Term Memory
• Recurrent Neural Network

Neural Network Hyperparameters

• # Nodes
• # Hidden Layers
• Dropout ratio
• Lag parameter: ‘n’
• Batch-size

Traditional methods: Hand Tuning, Grid search, Random search

​

Hyperparameter optimization using ‘black box optimization’ approaches

Objective: Predict groundwater level for the next day

• Location: Butte county, CA
• Historical observations
• Groundwater level
• Temperature
• Rainfall
• Nearby streamflow

​

• Data : 2010-2018 (daily time resolution)
• Predicted parameter: Groundwater surface elevation above Mean Sea Level
• Data sources: California Data Exchange Center, California Natural Resources Agencies

Automated approaches to find the best hyperparameters

Features

• No manual tuning
• Accurate and faster
• Multi-scenario analysis for future risk

Optimal choice of input features gives robust ML models

Input features:

• Groundwater
• Temperature
• Precipitation
• Riverflow

Input features:

• Groundwater
• Temperature

Black: observed

Blue-band : ML Model uncertainty

Low uncertainty

High uncertainty

Groundwater level (ft)

Groundwater level (ft)

Machine learning: Impact of Input feature selection

• What is the minimum input features for making a good ML model?
• How long into the future can we predict
• Can better guide future data infrastructure and investment

Other projects with ML

16

• Gap filling of groundwater time series with ML: (past postdoc Berkeley Lab)
• Interpreting Deep Learning Models in Hydrology: (YSSP, 2021)
•  LSTM type ML model learns interpretable hydrological concepts from data alone.

Future guiding questions

- 17 -

Spatially distributed watershed analysis

Incorporate ‘human decisions’ without pumping data ?

How far can we predict with future climate scenarios?

*IPCC, 2013

Part 2: Dynamics of Groundwater depletion

18

Potential for managing groundwater as a common pool

Spatially distributed finite element model

Improving hydrological complexity while performing the same optimization approaches

- 20 -

Hydrological

• Incorporate physical groundwater flow constraints
• Well yield constraints
• Environmental flow
• Steady state is not pumping = Recharge

​

Optimization

• Cooperative: Managed by a single decision maker to defines the actions of other agents with foresight
• Non cooperative: Each agent acts independently without foresight
• Incorporate environmental flow constraints

​

Goal: Classify different aquifer depletion scenarios and their critical areas for improvement or intervention

- 21 -

• 9 agents, 15km by 10km unconfined aquifer area
• Constant conductivity, variable transmissivity
• Arid, semi-arid condition with just mountain recharge and no return flow from crops
• Discount rate: 3% low, 10% high
• Demand: low and high
• Strategy: cooperative and non cooperative
• Environmental flow constraint: 50% of unpumped steady state
• Pumping switched on for 6 months then switched off
• Boundary recharge condition constant
• 50/ 25 year simulations

​

Initial unpumped steady state, Groundwater is not equally distributed

- 22 -

Groundwater depletion is a very transient phenomenon over decadal timescales

- 23 -

Low Demand

High Demand

Water productivity is doubled

• Increase in pumping
• Lower outflow to downstream
• Higher storage depeltion

Non cooperation leads to inefficiency

24

• 10% less Net Present Value of revenue
• 10-20% higher pumping and higher storage depletion
• Slight increase in outflow capture, greatest change in storage depletion over 50 years

- 25 -

Environmental flow constraint in high conductivity aquifer

• Environmental flow constraint: 50% recharge should be left downstream
• Pumping behavior is severely reduced
• Myopic scenario cannot simulate this and always fails the constraint

2 Wells closest to outflow stop pumping

- 26 -

Higher discount factor leads to earlier activation of outflow constraint

3% discount rate

10% discount rate

Constraint hits in 9 years

Constraint hits in 5 years

In a low conductivity aquifer this process will take longer or not take place before well yield constraint hit !

- 27 -

Aquifer with lower specific yield and smaller thickness

• Pumping constrained not by the economics but by the well yield
• Non Cooperative case, the agents hit the constraint from 4-10^th year of the simulation

​

Pumping rates of 9 players over 50 years arranged in single vector

Future/ On-going work

28

• Incorporate predictive behavior among agents for future uncertainty
• Modeling the groundwater problem a dynamic game to understand the effects of competition of long periods

​

​

​

Multi-scale dynamics of Groundwater Depletion, Water Resources Research (2021)

​

​

- 29 -

Thank you.

Questions?