1. Hanoi, January 29 th 2015 Rodolfo Soncini-Sessa DEI – Politecnico di Milano IMRR Project 11 – Problem simplification INTEGRATED AND SUSTAINABLE WATER MANAGEMENT OF RED-THAI BINH RIVER SYSTEM IN A CHANGING CLIMATE

2. The Design Problem (It)(It) scenario (I t, θ)

3. 4th December 2013 Gestione delle Risorse Naturali, Politecnico di Milano Universal function approximators Artificial Neural Networks with some particular features can be used as universal function approximators, i.e. as policies. Multi-layer Perceptron u 1,t u q,t θ = [γ 1 1,1, …., γ 1 m,n, …, β L 1, …, β L q ]

4. 4th December 2013 SOLVING APPROACH:  ANN to describe the control law ;  GA to find the optimal ANN parameterization. ALGORITHM: Gestione delle Risorse Naturali, Politecnico di Milano Run a system simulation for each individual Selection, crossover and mutation new population initial population time series of historical inflow objectives

5. J1J1 J2J2 Initial (random) population

6. J1J1 J2J2 selection of the “best” solutions according to the Pareto dominance criterion

7. J1J1 J2J2 survival of the fittest

8. J1J1 J2J2 generation of a new population

9. J1J1 J2J2 selection of the “best” solutions according to the Pareto dominance criterion

10. J1J1 J2J2 survival of the fittest

11. J1J1 J2J2 iterating….

12. J1J1 J2J2

13. J1J1 J2J2

14. J1J1 J2J2 final approximation of the Pareto front

15. System simplification

16. Time per policy evaluation over 39 years for the RTBR system: 0.53 sec. Dim θ = (2 x N input + N output ) x N neur N neur ≥ N input + N output N input = 4+2 N output = 4 N neur = 10 Dim θ = 160 Num policies = 10 160 4 reservoirs N input = 3+2 N output = 3 N neur = 9 Dim θ = 117 Num policies = 10 117 3 reservoirs N input = 1+2 N output = 1 N neur = 5 Dim θ = 35 Num policies = 10 35 1 reservoir Too large! Might be feasible Running time: 29 days Num evaluations about 5.5 10 6 SDP 250 seconds = 470 policy evaluations SDP is surely faster How to reduce the number of reservoirs to 3 only? We will see tomorrow.

17. Four reservoirs are too many

18. The Red-Thai Binh River basin

19. Thac Ba policy design SDP

20. Other reservoirs’ policy design GA

21. Other design conditions

22. Storages thresholds Minumum and maximum (unconstrained) release legal constrains Jun 15th - Jun 25th, early flood-period Before June 25th, if Hoa Binh reservoir level is at 107 m and the Hoa Binh inflow is predicted to be higher than 4,000 m3/s, Son La reservoir is allowed to use the storage between 200 m and 205 m to prevent Hoa Binh reservoir level from exceeding 107 m. WHAT TO DO? Son La

23. Only reservoirs safety conditions will be a priori imposed

24. Thanks for your attention XIN CẢM ƠN