In this Project several data mining techniques have been applied to the field of reservoir operation. Special interest has been given to study the general performance and applicability of Reinforcement Learning technique. As a case study Everett Jordan Dam and Lake has been used. The main objective of the study was to find a policy of operation that improves fulfilment of the actual operation criteria but does not differ drastically from the actual operation policy, so that it can be easily adopted and implemented. The first step has been the exploration of the data set's internal structure. A SelfOrganising Feature Map (SOFM) was built to identify possible groupings of the variables under study and classify the data set. The SOFM showed the existence of ten different zones of operation. According to the characteristics of the clusters a simplified classifier was created using Decision Rules. For each cluster, the action function that defines the outflow from the reservoir was approximated using Genetic Programming. With the action function it is possible to predict new, previously not observed events, with a reasonable level of uncertainty. The ten operation zones observed and their action functions have been used to find an optimal policy for the reservoir operation using Reinforcement Learning and Dynamic Programming. In this case the optimisation consisted on defining which action to perform for different states of the system. Based on the objectives of operation of the reservoir a utility function has been built to indicate the `value' of performing a given action given a certain state of the system. This function is used to discriminate between the different actions and define a desirable operation policy. The two methods of optimisation have a comparable performance and they have shown to be an improvement over the historical operation criteria. During high flood events, in conditions where the only information available is the current state of the system, Reinforcement Learning performs better that Dynamic Programming
Back to the list of MSc abstracts