Optimization of supervised learning models for modeling of mean monthly flows

Modeling of mean monthly flow is of particular importance for long-term planning of processes relying on water abstraction, such as reservoir operations. Advantage of data-driven models for these applications is the ability to predict monthly streamflow based on the combined hydrological and climatological input data. Methodology and recommendations for implementation of supervised learning (SL), from choice of input variables and optimization of model parameters in dependence of dataset size to final model evaluation, remains generally undefined. The main objective of this paper is to model mean monthly flow by SL models, while optimization algorithms (genetic algorithm-GA and simulated annealing-SA) are used to optimize and automate the choice of parameters and the reliable set of input variables. Detailed analysis of accuracy and amount of time needed to build three supervised learning models (ANN, SVM and NNM) for modeling of mean monthly flow is given in the paper. The 40-years input dataset has been shown as long enough for building models of satisfying quality, and was used in the further analysis where GA and SA were used first for optimization of model parameters, and later, for simultaneous optimization of both model parameters and input variables for SL models. In the analysis, time series was always split in building, calibration and verification part, while optimization was done on the building and calibration part. Data outside the particular time series was used for additional verification. Optimization of the model parameters by the exhaustive search indicated that the most accurate models were ANN and SVM, overperforming NNM across all data subsets, revealing that models need to be built with external variables. Optimization using GA and SA with SVM produced obvious movement toward optimal values, especially when the choice of parameters and input variables was optimized with GA-SVM.