A data-driven approach for the assessment of the thermal stratification of reservoirs based on readily available data

A data-driven approach to assess the occurrence of thermal stratification and the depth of the thermocline in water masses has been developed and tested in the Spanish reservoir of El Val. The novelty of this approach is that it relies only on readily available data that can be collected for almost any reservoir, providing water managers with a transferable tool that can be easily adapted for any other reservoirs or lakes. The input variables were meteorological data, the water level and the output flow. A non-supervised clustering technique, k-means, was used to identify the stratification period with unlabelled data, that is to say inferring patterns when data from the target variable is not available. As a supervised method, Artificial Neural Networks were used to classify a given day as having or not having stratification and, in the positive case, to infer the depth of the thermocline. The classification showed a very high accuracy (96%) and the estimation of the thermocline depth showed a mean absolute error (MAE) of 1.94 m and 1.99 m in the training and test fractions, respectively. Shapley Additive Explanations (SHAP) values were used to improve the explainability and they revealed that the most important features to infer the depth of the thermocline were the water level, the daily average solar irradiance and the air temperature.

The authors acknowledge the support from the Project CPP2021-008510 “Un Servicio Web Para la evaluación Sectorial del Cambio Climático a Medida”, funded by the Ministerio de Ciencia e Innovación. The authors also would like to acknowledge the data availability from the Confederación Hidrográfica del Ebro (CHE) through the SAICA Ebro and the SAIH Ebro.

Peer reviewed