Paper Title
LEVERAGING LSTM MODELS FOR IMPROVED RIVER DISCHARGE PREDICTION: SPATIAL VARIABILITY AND IMPLICATIONS FOR FLOOD FORECASTING AND WATER MANAGEMENT

Abstract
Understanding spatial variability in river discharge prediction is critical for robust hydrological modeling, particularly in regions where complex dynamics impact flood forecasting and water resource management. This study evaluates the predictive performance of an LSTM (Long Short-Term Memory) model in daily discharge prediction across multiple hydrometric stations at varying distances from the river mouth. We employed two evaluation metrics, R², and MAPE to quantify predictive accuracy, while nonlinearity in the discharge series was characterized using the Hurst Exponent (HE) method. Results reveal an intriguing spatial pattern: stations located both near the river source (within 10 km) and far downstream (over 100 km) exhibit high R² values (above 0.98), indicating strong model performance, despite moderate to high nonlinearity (HE range: 0.70–0.79). This pattern is also evident in two gauges on the same river—Kämmerzell (172 km from the mouth) and Bad_Hersfeld (119.8 km from the mouth)—where the station farther downstream shows higher accuracy, with R² values of 0.9943 and 0.9931, and MAPE values of 0.0394 and 0.1384, respectively. Conversely, stations positioned at intermediate distances (10–100 km) show a more complex relationship, with R² values spanning 0.90 to 0.99, suggesting that these mid-range regions may contain spatially variable and nonlinearly interacting hydrological processes. Our analysis suggests that predictive accuracy in discharge may not solely depend on nonlinearity levels but also on spatial dynamics unique to specific distances along the river. High accuracy at upstream and far downstream locations could be attributed to more consistent discharge patterns captured by the LSTM model, potentially due to dominant hydrological drivers. In contrast, stations at intermediate distances may experience variability from tributary confluences and localized runoff that complicate predictive accuracy. These findings highlight the importance of incorporating spatial context in hydrological model evaluations and provide a basis for refining AI-based approaches in predictive hydrology, with implications for flood resilience and adaptive water management strategies. This study demonstrates the potential of artificial intelligence, specifically LSTM models, to enhance the accuracy and robustness of hydrological simulations, contributing to more reliable flood forecasting and water resource management in complex river systems. Keywords - LSTM, Discharge, Modeling, Time Series, Hurst Exponent, Denoising, Spatial Variability