Subseasonal to seasonal prediction and forecasting of reservoir storage using a machine learning model in the UK, Spain and France

A multi-reservoir machine learning model has been built for seasonal to subseasonal reservoir storage predictions on a monthly timestep, including a range of reservoirs from the UK, France and Spain; and applied in forecast mode in the UK. The model demonstrates strong predictive skill across all the reservoirs, with a median NSE of 0.93 (1 month ahead) and 0.79 (3 months ahead). It outperforms a benchmark SARIMAX model in almost all cases, with a median NSE skill increase of 0.41 (1 month ahead) and 0.48 (3 months ahead). The model performance drops for those reservoirs without a clear seasonal signal. Monthly and seasonal ensemble forecasts were produced for a subset of reservoirs in the UK. The model shows good forecast skill for those reservoirs, with a median Continuous Ranked Probability Skill Score of 0.45 and 0.42 for 1 month ahead and seasonal forecasts, respectively. The model could be used to support operational monthly reservoir storage forecasts over the UK, and for reservoirs in Spain and France where a clear seasonal signal exists. There is potential for expanding this model to new regions, including areas with low volumes of accessible historical data.