Evaluation of data-driven and process-based real-time flow forecasting techniques for informing operation of surface water abstraction

This paper presents an approach to managing surface water abstraction utilizing real-time flow forecasting and control techniques. To evaluate the effectiveness of alternative data-driven and process-based methods, flow forecasts at a case study site (River Dove, UK) using (1) a probability-distributed rainfall-runoff model (PDM), (2) PDM coupled with an autoregressive integrated moving average (ARIMA) error predictor, and (3) a long short-term memory (LSTM) neural network are integrated into a water resources management model coupled with genetic algorithm optimization to simulate and compare water abstractions, reservoir storage, downstream river flows, and pumping energy costs. When compared to historical data, results show that both PDM plus ARIMA and LSTM forecasts led to improved water abstraction operations, i.e., increased water abstraction volumes during dry periods while maintaining river environmental flows, as well as reduced pumping costs. Cost savings were found to be sensitive to the accuracy of the forecasting technique only within specific flow ranges. This study demonstrates the water resource benefits of real-time flow forecasting in supporting flexible water pumping schedules and further discusses the benefits of alternative modeling approaches in the specific context of controlling water abstraction.