Investigation of Meta-heuristics Algorithms in ANN Streamflow Forecasting

The deterministic approach, which utilizes the gradient information in the search process, is prone to trapping at local minima, primarily due to the presence of saddle points and local minima in the non-convex objective function of an artificial neural network (ANN). This study investigated the efficacy of a hybrid model that adopted a meta-heuristic algorithm (MHA) as an optimizer to extend the training ANN method, from a gradient-based to a stochastic population-based approach for streamflow forecasting. In the latter, parameter tuning utilizing the design of experiment (DOE) technique, has become an integral element in the optimization process due to reliance on their parameters. For model convenience, a wavelet transform was employed to decompose the series into sub-series. The empirical studies of MHA performance showed that the hybrid MHA-ANN was superior for streamflow forecasting, especially with the firefly algorithm that had an average RMSE = 96.06, an improvement of approximately 17% over the gradient-based ANN (RMSE = 113.92). However, among the adopted MHAs, not all are compatible with optimizing the ANN for streamflow forecasting, thus requiring a thorough study as performance varies from case to case. Two additional statistical tests, such as the Kruskal-Wallis H test and the Mann-Whitney U test, further validated such disparity in the MHA's performance.

Automated summary

This study investigated the use of a hybrid model that combines a meta-heuristic algorithm (MHA) with an artificial neural network (ANN) for streamflow forecasting. The results showed that the hybrid MHA-ANN, especially with the firefly algorithm, outperformed the gradient-based ANN method with a 17% improvement in RMSE. However, not all MHAs were compatible with optimizing the ANN for streamflow forecasting, and additional statistical tests were used to validate the performance disparity.