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
Neelakantan and Pundarikanthan address the problem of optimization for reservoir operation. Theoretically the modeling framework is well known. However, this simulation can be very time consuming. This article focuses on “developing a planning model for reservoir operation that uses a simulation-optimization approach”. The study implemented a backpropagation neural network to approximate the simulation model developed. This methodology was applied for an urban water distribution reservoir in Chennai, India. The particular study case seems to be very relevant since it has a high demand and serious hydrologic conditions which makes shortfalls frequent.
The entire framework comprises four stages. The first stage is the training of a backpropagation neural network, which is used to simulate reservoir operation. The second stage links the neural network model as a submodel with Hooke and Jeeves nonlinear optimization programming model, in order to identify operation policies. In the third stage, good operation policies are selected to be refined using the conventional simulation-optimization model.
Two scenarios were studied. Scenario 1 considered the existing reservoirs and the optimal operation was compared with the standard operation policy. Two proposed reservoirs were considered in Scenario 2. With respect of overall deficit index, the addition of proposed reservoirs don’t improve the system performance, although it reduces losses due spills.
Discussion
The article presented an interesting application of neural network in water resources systems. The study case seems to be relevant and highly representative of others systems in developing countries. The conclusion that addition of two reservoirs doesn’t improve the overall performance of the system seems in a first view very surprising. This conclusion is more relevant for developing countries that face difficult financial conditions.
Marcio,
ResponderExcluirI've been trying to ask myself the same thing about how this neural network methodology can be improved upon. It has been almost a decade since the article was written.
What I've noted for this article is this "training" of the neural network. To me, this says that in 2000, we were able to guide the model towards sensible solutions. Now by 2009, there must be ongoing research out there that is looking into how to make this "training" more automated.
Will something like this be still used nowadays? Most likely. I think that the neural network concept seems quite new and still needs to be explored. What interests me are these hidden layers in the neural network and how they exactly work.