Primary production in lakes and reservoirs develops as a result of complex reactions and interactions. Artificial neural networks (ANN) emerges as an approach in quantification of primary productivity in reservoirs. Almost all of the past ANN applications employed input data matrices whose vectors represent either water quality parameters or environmental characteristics. Most of the time, the components of input matrices are determined using expert opinion that implies possible factors that affect output vector. Major disadvantage of this approach is the possibility of ending-up with an input matrix that may have high correlations between some of its components. In this paper, an automata networks (AN) based preprocessing technique was developed to select suitable and appropriate constituents of input matrix to eliminate redundancy and to enhance calculation efficiency. The proposed technique specifically provides an apriori rough behavioral modeling through identification of minimal AN interaction topology. Predictive ANN models of primary production levels were developed for a reservoir following AN based pre-modeling step. The achieved levels of model precisions and performances were acceptable: the calculated root mean square error values (RMSE) were low; a correlation coefficient ( R) as high as 0.83 was achieved with an ANN model of a specific structure.
Read full abstract