A Temporal Fusion Transformer Deep Learning Model for Long-Term Streamflow Forecasting: A Case Study in the Funil Reservoir, Southeast Brazil

Abstract

Water reservoirs play a critical role in water resource management systems, serving various purposes such as water supply, hydropower generation, and flood control. Accurate long-term streamflow predictions are essential for the efficient operation and planning of reservoirs, enabling water managers to anticipate changes in water availability, optimize reservoir storage, and make informed decisions about water allocation and infrastructure management. However, the increasing variability and uncertainty in hydrological processes due to climate change and anthropogenic activities necessitate the development of robust and precise prediction models. Temporal Fusion Transformer (TFT) models have emerged as a promising approach for hydrological forecasting, leveraging deep learning, time series analysis, and attention mechanisms to capture complex temporal dependencies and provide accurate predictions. This study employs TFT as a surrogate model to simulate the streamflow upstream of the Funil reservoir. A comparison was performed among the models Seasonal Naive, AutoARIMA, Theta method, and Deep ARIMA. The TFT model has the lowest MAE (70.88 m$^3$/s) and RMSE (121.66 m$^3$/s) of all models, which indicates that it is the most accurate one. The TFT model also has the highest NSE (0.43) and coefficient of determination (0.79), which indicates that it is the most promising model for capturing the actual streamflow patterns.
The TFT model effectively captures the intricate spatiotemporal patterns and dependencies in the streamflow data and accurately predicts streamflow upstream of the Funil reservoir, capturing seasonal patterns and long-term trends. This can help water managers make informed decisions about reservoir operations and management.