Investigation of the problem of medium-term forecasting of reservoir inflows using stochastic techniques and machine learning tools

The management of water resources is an issue that has been of concern to humanity for many centuries. In modern times, this implies the need to manage water in reservoirs, artificial or natural, in order to ensure full and uninterrupted coverage of the needs of water supply and irrigation, but more generally of all human activities that require water. In particular, in Attica, whose needs are covered by a system of three reservoirs, the Evinos, Mornos and Yliki, it would be particularly useful to be able to forecast the inflows into the reservoirs for short periods in the future, in order to be able to better plan the allocation of water from them. In this paper, we investigate whether artificial feedback neural networks, namely LSTMs (Long Short-term Memory Networks), can be used to predict reservoir inflows for the March – September period of the hydrological year, using rainfall and basin inflows for the October – February period as forecast data. The data used for the model training process (training and validation in terms of LSTMs) are synthetically generated from historical basin data, while the original historical data are used for model evaluation. The results of the predictions are compared with the simple linear model, using metrics suitable for hydrological problems, such as NSE (Nash – Sutcliffe Efficiency) and KGE (Kling – Gupta Efficiency). At the same time, the aim is to optimize the hyperparameters of the models to achieve the best results. Ultimately, it is concluded that the use of LSTM neural networks for predicting reservoir inflows is promising, and can be used, either alone or as part of hybrid models, as supporting tools in decision making systems for water resource management by the responsible agencies.