Dynamic Emulation Modeling of irrigation water deficit in the Red-Thai Binh River Delta

Water deficit (WD) is the factor that most affects agricultural productivity in irrigated land. The optimal management of water resources requires estimating and forecasting its intensity in space and time. In large irrigation systems, the adoption of a physically-based (PB), distributed, dynamic model (e.g.Mike11) is a viable solution, since it can simulate the hydraulic processes that occurs in the system, including the operation of hydraulic structures according with rules based on water demand and water availability. However PB models always have a high dimensional state that prevents their adoption within optimization algorithms, even the most advanced ones. Additionally, the information they produce (the time trajectories of all the variables of the systems) is definitively larger than what is actually required (the trajectories of WDs in pre-specified areas) in the management problem. Thus input/output (I/O) lumped models would be more appropriate, but their calibration is generally prevented by the lack of sufficiently long time series of the output variables (WDs in our case). The solution we explored is to split the process into two steps: the first is the calibration of a PB model (Mike11 in our case study) in a traditional way, using the available time series; the second is the identification of an I/O model from the WD time series generated by the PB model We present a dynamic emulation modelling (DEMo) approach which leads to the identification of an emulation model, namely a simplified, computationally-efficient model built over a sample data-set produced via simulation of the original model (Mike11). The core mechanism of the procedure is a feature-ranking algorithm, based on Extremely Randomized Trees, through which the suitable input variables of the emulation model are automatically selected. The emulation model, in the form of an I/O relationship, is then identified using Artificial Neural Networks. The proposed approach is demonstrated on a real-world case study: the Red-Thai Binh River Delta in Vietnam. It is a large and complex river system, supplied by five unregulated rivers and four large multi-purpose reservoirs, the operating rules of which have to be designed. In order to ascertain the effects of their regulation on the WD in the Delta, a Mike11 model was calibrated and validated. It describes 320 rivers and canals for a total length of 4200 km, 11 irrigation districts and many structures, among which 88 sluice gates and 302 irrigation water intakes (simulated as controlled pumps). Its inputs are the four reservoirs releases, the non-regulated flows of the five rivers, the sea levels at the nine river mouths, the water demand at each one of the irrigation water intakes. The considered outputs are the daily WD in the 11 districts. The dimension of the state vector of the model is of the order of 16,000. From the high dimensional Mike11, a low dimensional (11 dimension) DEMo model was identified that mimics very well (R2=0.95) the WDs as computed by the Mike11 model. The DEMo model will be later on embedded within a large multi-objective optimal control problem to design the operating rules of the four reservoirs.