Inflow volume forecasting using regional deep learning models trained on operational meteorological ensemble forecasts in Canada

Hydropower reservoirs typically require inflow forecasts to allow water resources managers to optimize drawdown rates and improve infrastructure efficiency. Usually, operators use physically-based or conceptual hydrological models to forecast streamflow for a desired lead-time, and then evaluate the total inflows for the period of interest. Recently, deep-learning models have been shown to provide better streamflow forecasts than classical hydrological models in certain cases. They have also shown better performance when trained on a multitude of donor catchments to increase the number of available data for training.

This work presents a novel method to provide inflow forecasts volumes directly, i.e. without first generating day-to-day streamflow, by using a deep-learning model trained on ensemble meteorological forecasts and observed inflow volumes for given lead-times. Furthermore, the model makes use of large-scale datasets during its training, by including data from 200 catchments in Canada. The model is then applied to a hydropower system reservoir to estimate 14-day inflow volume forecasts. The model shows promising results in terms of accuracy and reliability, and it is demonstrated that the addition of extra donor catchments during training helps increase the forecast performance. Furthermore, training the model using forecasted meteorological data as the inputs helps further increase model performance. This work demonstrates the potential residing in training regional models using meteorological forecasts for ensemble inflow volumes forecasts.