Titolo della tesi: Dynamics of Reservoir Management: AI-Driven Prediction and the Role of Rainfall Erosivity
Water reservoirs play a fundamental role in water resource management, ensuring the availability of water for irrigation, potable consumption, hydropower production, and flood control. However, increasing climate variability, anthropogenic pressures, and sedimentation processes are threatening their efficiency and long-term sustainability. Among the key challenges, sedimentation reduces water storage capacity, affects water quality, and compromises reservoir operation, leading to significant economic and environmental consequences. Understanding the dynamics of sediment transport and deposition is therefore essential to optimize reservoir management and mitigate the impacts of sedimentation. This dissertation explores two interconnected aspects of water resource management: the use of advanced data-driven models for sub-seasonal inflow forecasting and the estimation of rainfall erosivity to improve soil conservation strategies.
Hydrological forecasting is essential for water resource management, particularly for reservoir operations, drought mitigation, and flood control. Sub-seasonal forecasts (up to six weeks ahead) help bridge the gap between short-term weather predictions and long-term climate projections, offering insights into hydrological extremes. However, climate change has increased uncertainties in reservoir inflow estimation, making accurate forecasts crucial for effective management.
Challenges remain in extending the accuracy of predictions to longer lead times and extreme events. In small basins like the Órbigo system (Spain), sub-seasonal forecasts are not yet considered reliable for reservoir operations, highlighting the need for improved predictive approaches. Data-driven models, particularly Long Short-Term Memory (LSTM) networks, have shown promise in capturing non-linear hydrological processes and often outperform traditional models in streamflow forecasting. Given their ability to model time dependencies effectively, LSTMs represent a potential solution for improving inflow predictions in localized river systems. Here, we investigate the potential of LSTM networks to predict inflows into the Barrios de Luna reservoir (Spain) at lead times ranging from one week to one month. Results demonstrate that integrating observed data with ECMWF Extended Range forecasts enhances predictive accuracy, with LSTM outperforming traditional hydrological models and capturing complex hydrological dynamics. However, challenges remain in accurately forecasting peak flows, particularly in winter and autumn when seasonal patterns are less predictable.
Beyond hydrological variability, another critical factor influencing reservoir management is sediment transport, which reduces storage capacity and compromises operational efficiency. Therefore, the second part of this work focuses on rainfall-induced soil erosion, a primary driver of sedimentation in reservoirs. Using an extensive dataset from the Lazio region (Italy), we assess rainfall erosivity and develop regression models to estimate the R-factor based on coarser precipitation data. Our findings highlight the importance of site-specific calibration in erosivity estimation, as relationships derived from other regions may lead to over- or underestimations. The study provides valuable insights for improving soil conservation practices and reducing sediment inflow into reservoirs.
By integrating hydrological forecasting and erosion assessment, this research contributes to more effective and sustainable reservoir management strategies. The findings emphasize the need for advanced predictive tools to anticipate hydrological variability and soil erosion, ultimately supporting decision-making processes aimed at extending the functional lifespan of reservoirs and ensuring water security in the face of climate change.