Autonomous neural models applied to medium-term water inflow forecasting

Abstract

In the Brazilian National Interconnected System (NIS), the hydro plants make up the most of the generating park, being of great importance to study the behavior of future inflows to the energy operation planning. Due to the cyclic characteristic of wet and dry periods in Brazil, the behavior of the water inflow time-series is non-stationary, thereby hampering the direct use of classical models of time series analysis. The large number of water inflow time-series to be analyzed jeopardizes the individual treatment of each series by a specialist. The use of artificial neural networks (ANN) allows direct modeling aspects of seasonal and non-linear series of water inflows, but requires specialist intervention to specify the modes, i.e., input selection, structure definition and complexity control. This paper presents autonomous neural models for forecasting monthly water inflows. Chaos Theory is used for initial input space representation, with automatic clustering techniques been applied for autonomous identification of regions of the reconstructed attractor to be modeled. For the modeling task, Bayesian Inference applied to multilayer perceptrons (BIAMLPs) training and specification is used. The models are compared against another ANN proposals and classical models.