Regional Analysis of Daily Precipitation Stochastic Model Parameters Using Artificial Neural Networks

The development and the implementation of successful water resources management tools to assess engineering and environmental problems, such as flood control, on-line reservoir operation, hydropower generation, water quality control or river ecosystem constraints, among several others, often require the analysis, simulation and prediction of rainfall data. The Markov Chain-Mixed Exponential stochastic model (MCME) is extensively used for estimation of rainfall data. In spite of this method's wide acceptability, improvements in order to estimate the Fourier coefficients of the MCME model at ungaged meteorological stations are incorporated in this paper. The performance of feed forward neural networks (CNNs) to forecast the coefficients of the MCME model at basins in southern Spain are analyzed. Historical precipitation data from 15 meteorological stations in Andalucía (Spain), each with 52-year daily precipitation records (1953–2004), are used to test the efficiency of incorporated improvements. For that purpose several CNN models, trained with the Levenberg-Marquardt algorithm, are implemented and compared. The performance of the MCME model through the weighting interpolation model was compared with neural approaches as data-driven to generate daily precipitation records in locations where observed rainfall records are not available. To assess the performance of the models during the validation phase and therefore to identify the best model, several measures of accuracy are applied, as there is not a unique and more suitable performance evaluation test.