Abstract
This study evaluates three Machine Learning (ML) models-Temporal Kolmogorov-Arnold Networks (TKAN), Long Short-Term Memory (LSTM), and Temporal Convolutional Networks (TCN)-focusing on their capabilities to improve prediction accuracy and efficiency in streamflow forecasting. We adopt a data-centric approach, utilizing large, validated datasets to train the models, and apply SHapley Additive exPlanations (SHAP) to enhance the interpretability and reliability of the ML models. The results show that TKAN outperforms LSTM but slightly lags behind TCN in streamflow forecasting. TKAN demonstrated strong alignment with observed statistical parameters, achieving a Mean Absolute Error (MAE) of 5.799 m³/s and a Nash-Sutcliffe Efficiency (NSE) of 0.958, compared to MAE and NSE values of 8.865 m³/s and 0.942 for LSTM, and 5.706 m³/s and 0.961 for TCN, respectively. Multi-step forecasting revealed TKAN's robust performance up to a three-day forecast horizon, with a slight decline in accuracy as the forecast period extended. Uncertainty analysis indicated reasonable variance levels, with a mean 3-day forecast uncertainty of 35.02% at a 95% confidence level for TKAN, compared to 39.95% for LSTM and 28.46% for TCN. For a 7-day forecast, TKAN showed a mean uncertainty of 40.97%, compared to 45.01% for LSTM and 36.22% for TCN. By enhancing model transparency and improving datasets, this study significantly advances the integration of machine learning into hydrological forecasting, offering robust methods for developing adaptive water management systems in response to changing climate conditions.