Stochastic generation and forecasting of monthly hydrometeorological data based on non-traditional neural network

The benefits of well-informed water management systems are related to the forecasting skills of hydrological variables. These benefits can be reflected in reducing economic and social losses to come. Therefore, the optimal design of water management projects frequently involves finding the methods or techniques that generate long sequences of hydrological data. These sequences considered as time series can be used to analyze and optimize the performance of the project designed. In order to cover these requirements, this work presents a new model of the stochastic process applied in problems that involve phenomena of stochastic behavior and periodic characteristics. Two components were used, the first one, a type of recurrent neural network relatively recent introduced in the literature and conceptually simple called ESN (echo state network) as the deterministic component, an interesting feature of ESN is that from certain algebraic properties, training only the output of the network is often sufficient to achieve excellent performance in practical applications. The second part of the model incorporates the uncertainty associated with hydrological processes, the model is finally called ESN-RNN. This model was calibrated with time series of monthly discharge data from four different river basins of MOPEX data set. The performance of ESN-RNN is compared with two feedforward neural networks ANN-1, ANN-2 (with one and two past months respectively) and the Thomas-Fiering model. The results show that the ESN-RNN model provides a promising alternative for simulation purposes, with interesting potential in the context of hydrometeorological resources.