Comparative Analysis of Artificial Intelligence Methods for Streamflow Forecasting

Abstract

Deep learning excels at managing spatial and temporal time series with variable patterns for streamflow forecasting, but traditional machine learning algorithms may struggle with complicated data, including non-linear and multidimensional complexity. Empirical heterogeneity within watersheds and limitations inherent to each estimation methodology pose challenges in effectively measuring and appraising hydrological statistical frameworks of spatial and temporal variables. This study emphasizes streamflow forecasting in the region of Johor, a coastal state in Peninsular Malaysia, utilizing a 28-year streamflow-pattern dataset from Malaysia's Department of Irrigation and Drainage for the Johor River and its tropical rainforest environment. For this dataset, wavelet transformation significantly improves the resolution of lag noise when historical streamflow data are used as lagged input variables, producing a 6% reduction in the root-mean-square error. A comparative analysis of convolutional neural networks and artificial neural networks reveals these models' distinct behavioral patterns. Convolutional neural networks exhibit lower stochasticity than artificial neural networks when dealing with complex time series data and with data transformed into a format suitable for modeling. However, convolutional neural networks may suffer from overfitting, particularly in cases in which the structure of the time series is overly simplified. Using Bayesian neural networks, we modeled network weights and biases as probability distributions to assess aleatoric and epistemic variability, employing Markov chain Monte Carlo and bootstrap resampling techniques. This modeling allowed us to quantify uncertainty, providing confidence intervals and metrics for a robust quantitative assessment of model prediction variability.