Assessing the value of meteorological reforecast data to predict inflow volumes over a Canadian snow-dominated catchment using a deep learning model

Hydrological forecasting is essential across multiple sectors, including hydroelectric power generation, flood prediction and mitigation, and water resource management. In this field of research, Machine Learning (ML) models have shown promising results and are increasingly used to replace traditional hydrological models.

This work presents a novel framework for forecasting 14-day inflow volumes to a hydropower reservoir using deep-learning models and atmospheric reforecasts in a Canadian catchment. The forecasting framework investigates whether Long Short-Term Memory (LSTM) models can directly forecast inflow volumes without relying on intermediate daily streamflow predictions, and whether integrating meteorological reforecast data during training can enhance model performance and forecast quality.

Three LSTM models were trained using various combinations of meteorological data from the European Centre for Medium-Range Weather Forecasts (ECMWF), including ERA5 reanalysis data and probabilistic reforecast datasets. The target hydrological forecast is the 14-day cumulative inflow volume to the reservoir. The first model is trained exclusively with ERA5 data, the second using a combination of ERA5 data and reforecasts, and the third combining the training datasets of the first two models. The models are then used to generate hydrological forecasts using ECMWF ensemble meteorological forecasts and assessed with quantitative metrics such as the Kling-Gupta Efficiency (KGE), Continuous Ranked Probability Score (CRPS), and Average Bin Distance to Uniformity (ABDU).

Results indicate that the three LSTM models can directly predict 14-day cumulative inflow volumes with reasonable accuracy and reliability, yielding strong performance metrics. However, no single model consistently outperforms the others. The model trained solely on reanalysis data exhibits greater variability in its predictions, resulting in lower accuracy but higher reliability. Results also vary seasonally. These findings suggest that incorporating meteorological reforecast data during training offers valuable potential for improving inflow volume forecasts within specific seasons and depends on the desired trade-off between accuracy and reliability.

Overall, it can be stated that LSTM models are a promising alternative to current operational models for inflow volume forecasting, although further research is necessary to understand how to fully exploit their potential and ensure their applicability and transferability into an operational context.