Seasonal forecasting of dam water resources using optimized hybrid models under unprecedented drought conditions

Abstract

In the Oum Er Rbia watershed, Morocco, dam water resources play a crucial role in prolonged drought conditions, particularly in the case of the Al Massira Dam, which has been a strategic reservoir for drought resilience since its inauguration. Optimized pipelines of explainable artificial intelligence (XAI) models were developed for monthly forecasts of water resource variations at the Al Massira dam, which has been affected by unprecedented drought since 2019. The architectures of the models developed incorporate Bayesian optimization via Optuna for identifying the best hyperparameters, advanced feature selection methods, and lagged regressors of teleconnection indices, drought indices, and hydroclimatic variables. The performance of the models was first evaluated in terms of their ability to forecast dam water volume up to 6 months ahead under near-normal hydroclimate conditions. Next, model performance was assessed under a scenario of unusual changes in time series. The light gradient boosting machine (LightGBM) showed high uncertainty when forecasting water volumes under unusual drought conditions, with Skill= 75.1 % and NMAE= 11.2 %. The Bayesian probabilistic LSTM (ProbLSTM) reached the maximum predictive skill score (Skill=86.2 %, NMAE=3.6 %), followed by the generalized additive model (GAM) (Skill=85.3 % and NMAE=3.4 %). Overall, from an operational perspective, ProbLSTM and the GAM are preferable for seasonal forecasting because of their low performance variability under a scenario of unusual changes in time series and their high predictive performance. • Gradient boosting showed high forecast uncertainty under unprecedented changes. • Optimized neural network models and GAM are preferable for seasonal forecasting. • Bayesian optimization is very useful for achieving high predictive performance. • Optimized models developed achieved high predictive performance, even with limited historical data. • Dam water volume forecasts up to 6 months in advance were achieved with an MAE of 17.7 Mm3.