Friday, 28 July 2017: 2:30 PM
Constellation E (Hyatt Regency Baltimore)
Water table forecasts are a significant resource in the development of water management plans for water allocation, especially in areas where groundwater is the main resource for irrigation. However, the nonlinear behaviour of the system is strongly affected by irrigation decisions and by hydro-climatological processes (such as rainfall and evapotranspiration) and physical parameters (conductivity and transmissivity) which are highly uncertain. This makes groundwater forecasts in irrigation systems using physically-based models challenging. The objective of this study is to evaluate uncertain in semi-seasonal to seasonal forecasts of groundwater-well levels for agriculture. This paper proposes an ensemble-based data-driven approach, including statistically-based precipitation forecast. The first stage uses five different techniques: Random Forest, Support Vector Machines, Artificial Neural Networks, Deep Neural Networks and Genetic Programming. These techniques are used to build models for the monthly forecasts, up to 5 months, of water table level in response to precipitation, snowmelt and evapotranspiration, in an intensively cultivated area in the Northern High Plains Aquifer in Nebraska, US. The purpose of this stage is to test the capability of data-driven models to forecast water-table depth with sufficient accuracy and to potentially overcome several limitations of the physically-based approach. The results suggests that all the data-driven models predict water table depth with high accuracy up to two months ahead. When the prediction horizon increases, genetic programming (GP) and artificial neural networks (ANN) have shown better performance than the other modelling techniques.
Secondly, we use GP and ANN to develop the probabilistic-based ensemble forecast, in order to assess the propagation of uncertainty in the precipitation input and in irrigation practices on the groundwater system. Historical records are used to fit statistical distributions to precipitation at each month of the year, and to compute precipitation quantiles. Each of them represents a statistically forecasted input scenario which governs the dynamic of the farmer’s pumping and of the groundwater systems accordingly. Pumping is represented here by means of the difference of total evapotranspiration (ET) and rainfall (R), under the assumption of conditional inference between ET and R. Analysis of the results of the ensemble procedure provides a better estimation of extreme conditions with respect to the first stage of the approach (10% of average decrease in RMSE during drought conditions), and ensure the quantification of the uncertainty propagation in groundwater forecasts associated with the high variability and the non-stationarity of the hydro-climatic variables.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner