Analyzing the influence of spatial rainfall variability on reservoir outflow through interpretable data-driven modelling: an application in Tropical mountain basin

Spatial rainfall variability directly impacts hydrological basin responses, especially in regions with complex interactions between hydrometeorological phenomena and physical factors. Understanding this influence supports the development of more accurate flow forecasting models, enhancing practical applications in water resources management. However, current studies often overlook the effect of rainfall spatial distribution on forecasting outcomes, despite its critical role in shaping hydrological responses. Such forecasts are essential for planning water distribution across sectors like human consumption, agriculture, and energy generation. In Latin America and the Caribbean, where hydropower supplies approximately half of the region's electricity, accurate forecasts are crucial. Ecuador, for instance, relies on hydropower for over 85% of its energy needs, underscoring the necessity for reliable hydrological forecasts, particularly in the Andean tropical mountain basins where major hydropower plants are located. One of the most important hydropower systems in the country, supplied by its largest reservoir, currently lacks an operational hydrological forecast to support its management, which is urgently needed due to the ongoing drought conditions impacting hydropower production. To address this, we developed data-driven models, known for their ability to handle data complexity and outperform conceptual or physical models in scenarios with high complexity and limited precise data, which is characteristic of our study area. These models analyze the influence of spatial rainfall variability on reservoir outflow forecasting using an interpretable approach to identify the most impactful rainfall data configurations. We employed satellite-based rainfall data from IMERG and GSMaP, which have shown promising results in nearby basins, across five top-down configurations: mean rainfall, climatological rainfall regions, seasonal clusters, travel time regions, and spatially distributed data. The models, based on neural networks including (RNN) Recurrent Neural Networks and Long-Short Term Memory (LSTM) architectures, are configured to provide forecasts from hourly to daily scales across these scenarios, supporting practical operational applications. Initial experiments indicate strong model performance, with NSE values ranging from 0.9 to 0.45 for hourly forecasts at lead times from 3 to 24 hours. To enhance interpretability, methods such as SHAP (SHapley Additive exPlanations) are applied to understand how rainfall data conditions model performance under different hydrological scenarios. With this approach, we aim to identify the optimal rainfall data setups for improved forecasting models in these basin settings.