SOM-Based Decision Support System for Reservoir Operation Management

This paper presents a decision support system (DSS) that can be used by reservoir managers for modeling and visualizing the complex relationships between key variables such as rainfall, streamflow, reservoir level, and water releases. The proposed approach uses a self-organizing map (SOM), a competitive-learning artificial neural network frequently used for vector quantization and clustering. The map uses an unsupervised learning approach and is trained with observed (i.e., historical) input data. This novel nonparametric approach can identify and characterize the different operation conditions and operation policies implicit in the training data. A SOM provides a set of operation profiles, each with specific values of the uncontrollable, state, and controllable variables. The topology-preserving character of SOMs allow for an intuitive and helpful interpretation of the computed prototypes and their relationships. The knowledge provided by these operation profiles can be used in many different ways, some of which are illustrated in the paper. The proposed approach has been tested on the Guadalmellato River reservoir in southern Spain. In this application, intense as well as prolonged rainfall episodes can be identified and characterized, as well as the dry season. The map clearly delineates specific SOM regions for hydroelectricity generation and flow-balancing discharges. The map provides insights into the correlation between these controllable variables and the uncontrollable rainfall and streamflow variables mediated by the state variables (e.g., reservoir level). The trained SOM can be used to study past periods (e.g., of heavy rains) and understand and revise the course of action followed. It can also be used to suggest future operation decisions based on the forecasted values of rainfall, streamflow, and reservoir-level variables.