Artificial intelligence for streamflow prediction in river basins: a use case in Mar Menor
Alejandro Cisterna‐García, Aurora González-Vidal, Antonio Martínez Ibarra, Yu Ye, Antonio Guillén-Teruel, Luis Bernal-Escobedo, Antonio Skármeta
Abstract
Streamflow prediction is crucial for efficient water resource management, flood forecasting and environmental protection. This is even more important in areas particularly vulnerable to environmental changes such our study area-the Mar Menor basin in the Region of Murcia, Spain-with a specific emphasis on the Albujón watercourse, a significant contributor to the Mar Menor. Utilizing data from stream gauge stations, nearby rain gauge stations, and piezometers, our research forecasts streamflow at two critical points: "La Puebla" and "Desembocadura" along the watercourse. Targeting short-term forecasts of 1, 12, and 24 hours, our study employs Machine and Deep Learning techniques after data preprocessing, which includes station selection, data granularity adjustment, and feature selection. A state-of-the-art data augmentation technique was used to balance periods of low and high streamflow. Results show that Random Forest slightly outperforms LSTM for 1-hour forecasts (NSE > 0.89, MAE < 0.01), while Long Short Term Memory with data augmentation excels for 12 and 24-hour forecasts (NSE > 0.12, MAE < 0.05). This is noteworthy in areas with torrential rains causing rapid streamflow increases, a more challenging yet less studied scenario in forecasting. The findings contribute to addressing the challenges associated with streamflow prediction in vulnerable regions.