Simulation and prediction of streamflow based on phase space reconstruction and deep learning algorithm

Developing low data-dependent, efficient, practical and accurate modeling techniques can provide effective solutions for hydrological simulation and prediction in areas with limited data availability.From a data-driven perspective, a composite streamflow prediction model, PSR-LSTM, which integrates Phase Space Reconstruction (PSR) and Long Short-Term Memory (LSTM) networks, was proposed in this study and validated globally over ten river basins (stations) in different climate zones.The results indicate that the PSR-LSTM can effectively extract multi-dimensional sub-space hydrological features and accurately predict streamflow changes at different time scales.Compared to LSTM, the Nash efficiency coefficient of PSR-LSTM in predictions of future 1 to 9 timesteps is increased by an average of 1.49% to 9.77% over the ten river basins; the root mean square error is reduced by an average of 17.01% to 19.72%.The dependency on the amount of training data is reduced by 25% to 33% for PSR-LSTM compared to LSTM.The research findings obtained in this study provide insights into hydrological simulation and prediction in data-scarce river basins.