Modelling and reforecasting real-time reservoir operation and outflow with neural networks: case study of the multi-purpose Sirikit reservoir in Thailand

Abstract

Real-time reservoir operations are highly dependent on decisions made by reservoir operators, which are difficult to simulate accurately with process-based hydrological models. Data-driven models, particularly those based on machine learning (ML), have been shown to be able to overcome the limitations typically encountered in process-based models. Despite a large number of ML studies in reservoir operation modelling, only few studies have focused on ML model performance in real-time reservoir operation and outflow forecasting. This study aims to investigate the capabilities of the Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) and Gated Recurring Unit (GRU) in simulating and reforecasting real-time (daily) reservoir operation and outflow, considering uncertainties in input data, training-testing periods and different algorithms. A major, multi-purpose reservoir, namely the Sirikit reservoir, in the upper Chao Phraya River basin Thailand was used as the case study. The main inputs for the ML operation models included the daily reservoir outflow, inflow, storage and the month of the year. We applied the distributed wflow_sbm model for inflow simulation (using MSWEP precipitation data) and inflow reforecasting (using ECMWF precipitation data). Daily reservoir storage was obtained from observations and real-time recalculation based on the reservoir water balance. The ML operation models were trained and tested with 10-fold cross-validation. Results show that RNN, LSTM and GRU can reconstruct real-time reservoir operation and provide accurate outflow when training data cover both regular and extreme conditions. For multi-day reforecasting, the model performances are appropriate for the current day up to 2-day lead times for low outflows and up to 6-7 days for high outflows. GRU is potentially the most accurate, robust and convenient model to be used in practice. We conclude that with some further improvements, the ML operation models can be effective and applicable tools to support decision-making for real-time operational water management.