Estimativa Probabilística de Caudais em Tempo Real Integrada com Transformers de Última Geração

The objectives of PREDICT stem from extensive experience in operational hydrological forecasting and focus on some of the most pressing needs in the field. These objectives are manifold and can be described on two levels. On a general level, PREDICT will contribute to saving lives, promoting the economy, and increasing the know-how and competitiveness of the private sector. It will do this by: a) Enhancing the state-of-the-art capability of forecasting hydrological events reliably and efficiently, b) Exploring decision-support tools that can cope with the rich information provided by probabilistic forecasts, and c) Implementing findings in an operational forecasting portal that will cover some of the main Portuguese catchments, directly contributing to society and showcasing the techniques and technology involved. On a deeper level, the project will address the following scientific questions related to hydrological forecasting: 1) To what extent is the performance of the latest generation of meteorological forecasts limiting for hydrological forecasting tasks from short to seasonal prediction horizons? 2) Can Temporal Fusion Transformers (TFTs) overcome the predictive performance of classical hydrological models? 3) What are the predictive limits of TFTs during extreme events? 4) Can the statistical reliability of Temporal Fusion Transformers be enhanced probabilistic post-processors more powerful than quantile regression in the output layer? 5) Can TFTs be successfully used in ungauged catchments, effectively encoding and using implicit knowledge about hydrological processes? And 6) Can hydrologic predictive uncertainty affect previous conclusions about the impacts of climate change? Additionally, it will attend to the following practical needs: 7) Analyze deep learning models to extract the key information involved in predictions, 8) Devise ways of generating statistically robust synthetic historical forecasts based on surrogate observations, 9) Propose statistically robust methods to generate streamflow timeseries coherent with one-shot probabilistic discharge and accumulated volume forecasts, and 10) Demonstrate that the proposed forecasting framework is readily applicable to newly monitored catchments. PREDICT will also show that, by exploring advances high-performance computing, multi-objective optimization coupled with probabilistic forecasts is achievable in real-time and can lead to better outcomes in terms of flood and drought preparedness and mitigation, as well as more reliable water supply and more efficient hydropower production.