Dawn Lalmansingh, Philippe Tissot, Richard Hay Texas A&M University – Corpus Christi

1. Neural Network Modeling of Spring Levels Linked to a Karst Aquifer: Case Study of the Comal Springs Dawn Lalmansingh, Philippe Tissot, Richard Hay Texas A&M University – Corpus Christi

2. Introduction • Importance of modeling Comal Springs baseflow • What is neural network modeling? • Neural Network Design strategy

3. Study Area COMAL r e v i R e p u l a d a u G NEW BRAUNFELS " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " 08169000 # § ¨ ¦ 35 NEW BRAUNFELS Legend # stream gages GUADALUPE " precipitation stations Urban Areas County Boundary Drainage Area Recharge Zone Artesian Zone streams 0 0.5 1 2 3 4 Highways Miles Inputs to Model Comal County Baseflow from gage station Precipitation from Stations 1, 2, and 3 3 Comal Springs 2 1 Comal River Guadalupe River ³

4. Historical Background • Edwards Aquifer – principal source of water for ~ 2 million people • Comal Springs – largest spring system in aquifer • Declining flow over past 3 decades • Provides habitat for 4 endangered species • Survival depends on sustaining minimum flow

5. Background cont’d • Current groundwater models forecast dry spells for different periods of time. • Major and minor droughts • Lead to extinction of endangered species • Alternate water resources • Neural Network – alternative model • Accurately forecast spring levels

6. Neural Network Modeling • Started in the 60’s • Key innovation in the late 80’s: Backpropagation learning algorithms • Number of applications has grown rapidly in the 90’s especially financial applications • Growing number of publications presenting environmental applications

7. Neural Network Features • Non linear modeling capability • Generic modeling capability • Robustness to noisy data • Ability for dynamic learning • Requires availability of high density of data

8. Neural Network Forecasting of Spring Levels BaseflowHistory  (X1+b1)  (a1,ixi) Precipitation 1 History  (X3+b3) b1  (a3,ixi) H (t+i) Precipitation 2 History b3 Spring Level Forecast  (a2,ixi)  (X2+b2) Precipitation 3 History b2 Input Layer Hidden Layer Output Layer Philippe Tissot - 2000

9. Data • Daily baseflow and precipitation data spanning 50 years (01/01/1950-12/31/2000) • HYSEP – separated streamflow data into baseflow and surface runoff components • Separate into 5 sets of 10 years • Train on 1 set and test on other 4 • Compare results to persistence model

10. Neural Network Design Strategy • Find optimum number of previous days of baseflow as inputs to model. • Find optimum number of previous days of precipitation as inputs to model. • Find how many neurons provide the best neural network performance.

11. Target Error Criterion and Root Mean Square Error • Target Error Criterion – percentage of model’s results that fall within 20% of the measured baseflow values • Root Mean Square Error (RMSE) – overall error of a sampling distribution of a group with n cases in a group

16. Good Example Baseflow (cfs) Measured Baseflow Days in Test Set ANN Forecast

17. Bad Example – Too many spikes Baseflow (cfs) Measured Baseflow Days in Test Set ANN Forecast

18. Comparison of Target Error Criterion and RMSE through Neural Network Design Process

19. Conclusion • Trained & tested ANN over 50 years of data • Best model – 3 days baseflow, 5 days precipitation, and 2 neurons • Main problem – spikes • Slightly underperforming persistence model • Explore other inputs

20. Questions