A multi-reservoir machine learning model for forecasting reservoir storage at monthly and seasonal timescales

Reservoirs are a vital aspect of water resource management in many regions worldwide, used to meet domestic and irrigation demand, for hydropower, flood control, and maintaining river flow for ecological and navigational purposes.  It is important to have a reliable forecast for reservoir status to ensure efficient operation of individual reservoirs and the wider water resource system, particularly during unusually wet or dry periods. This forecasting will become increasingly important under a changing climate and growing demand for water and hydropower. In this work, we focus on sub-seasonal to seasonal forecasts, providing vital information for water users and enabling them to make informed decisions on management strategies with seasonal lead-times.

There has been research on forecasting reservoir status (i.e. reservoirs storage, inflow, outflow, level, or storage anomaly) at different lead-times with machine learning (ML) methods, with most studies producing reservoir-specific models. In this work, we produce an Extra Trees (ET) regressor model for multiple reservoirs over Europe, trained on historical monthly storage, rainfall and temperature, along with static catchment characteristics from the relevant CAMELS (Catchment Attributes and Meteorology for Large-sample Studies) datasets. This model is used to forecast storage at one- and three- month lead times for reservoirs in the region, including reservoirs unseen by the model in the training period. A multi-reservoir model makes use of all the available data, allowing forecasts for reservoirs with limited historical data, and improves prediction performance under extremes compared to a single-reservoir model.

The model has already shown promising results for predicting reservoir storages using observed climatic inputs, and this work will investigate the model skill in forecast mode, using ensemble climate hindcasts, for reservoirs across Europe. It is anticipated that this model will provide a computationally- efficient forecasting tool with relatively low input data requirements that can be used to forecast reservoir storage in the modelled area, with potential to expand to a global extent.