Assignment #8 - Neural Network / Reservoir Optimization

Review Article
1) Article Reference
Neelakantan TR, Pundarikanthan NV (2000) “Neural network-based simulation-optimization model for reservoir operation,” JOURNAL OF WATER RESOURCES PLANNING AND MANAGEMENT, 126(2) pp. 57-64

2) Summary
This paper presents an application of neural network-based simulation optimization model for reservoir operation. A case study is presented for Chennai city, in India. The general framework of the study presented by the author follows: 1) Develop a conventional simulation model to obtain results from different operation policies; 2) Train a back propagation neural network model using those results; 3) Link the neural network model as a sub model with direct search non-linear programming optimization models; 4)Find the optimal policy and near optimal policies using neural network –optimization model; 5)Refine the optimal policies obtained using conventional simulation optimization model and determine the optimal policy.
The optimization is done using the Hookes and Jeeves direct search method. For this case it was considered better to have small water supply restrictions along the time from having one big drought. Then the optimization goal was to minimize the minimum sum of the deficits. For that a deficit index was defined. The optimization procedure started at an initial point and progress using Hookes and Jeeves method until an optimal solution. Several initial points were used as this algorithm can get stacked in local minimums.
A neural network simulation model was used in order to enhance the speed of the process. For this case the authors used a back propagation approach. The network was trained using pairs of input and output vectors.
The decision set was composed of 18 decision variables. Based in inflow pattern the year was divided into six time periods. Four out of five supply levels were considered at each time. The results are presented in two different scenarios. For each scenario, the authors presents the results for 4 different policies including the storage for each time period, the release and the deficit. The inclusion of two additional reservoirs is also analyzed.
As the major result, the authors concluded that the neural network based simulation-optimization model performed satisfactory. Also the authors mention that reservoir operation problems considering several and more complicated networks could be handled by this method.

3) Discussion
The paper presents a very interesting approach to solve a very common reservoir operation problem. I believe several attempts have been made to find methodologies for optimal operation rules for reservoir. I think this is an emerging issue as water demand is increasing significantly in large metropolitan areas and the reservoirs systems have to be used optimally.