A hybrid annual runoff prediction model using echo state network and gated recurrent unit based on sand cat swarm optimization with Markov chain error correction method

Abstract

Reliable annual runoff prediction is crucial for efficient water resource planning. Therefore, this study proposes a hybrid model based on the combination of sand cat swarm optimization (SCSO), echo state network (ESN), gated recurrent unit (GRU), least squares method (LSM), and Markov chain (MC) models to improve the accuracy of annual runoff prediction. First, correlation analysis is conducted on multifactor data related to runoff to determine the input of the model. Second, the SCSO algorithm is used to optimize the parameters of the ESN and GRU models, and the SCSO-ESN and SCSO-GRU models are established. Next, the LSM is used to couple the prediction results of the SCSO-ESN and SCSO-GRU models to obtain the initial prediction results of the SCSO-ESN-GRU model. Finally, the initial prediction results are corrected for errors using MC to get the final prediction results. Two stations are selected as experimental stations, and five evaluation indicators are chosen to reflect the model's predictive performance at the experimental stations. The results show that the combined prediction model corrected by the MC achieved the optimal prediction performance at both experimental stations. This study emphasizes that using a combination prediction model based on MC correction can significantly improve the accuracy of prediction.