Investigating the Impact of Time Series Structure in Performance of Transformer-Based Model for River Streamflow Forecasting

River discharge forecasting plays a pivotal role in water resource management and environmental planning. Understanding the long-term dependence or changes in these processes is crucial for accurate predictions. Deep-learning methodologies have garnered significant scientific interest and are progressively becoming more prevalent across water-resources-related endeavors. Transformer models, a novel architecture that aims to track relationships in sequential data through attention mechanism, have increasing popularity last years. Through comprehensive experiments and analysis on real-world river discharge datasets, we aim to elucidate the impact of long-term dependence detection, as facilitated by the climacogram and Hurst coefficient, on the predictive capabilities of a transformer-based model. Insights from this investigation are anticipated to contribute to the advancement of river discharge forecasting methodologies, enhancing our understanding of long-term dependencies in these environmental processes.