Multi-Factor Water Level Prediction Based on IndRNN-Attention

Abstract

Accurate prediction of hydrologic time series, such as water level and water volume, is essential for effective water resource management and plays a crucial role in detecting water transfers. Existing water level prediction methods usually only consider a single factor (such as the historical water level data), and do not fully consider other factors such as flows from the upstream stations that affect the water level. To address this problem, in this paper, we propose a multi-factor water level prediction model that combines an independently recurrent neural network (IndRNN) with an attention mechanism. Our model overcomes the gradient disappearance problem of traditional RNNs and improves prediction accuracy by encoding the historical input data that influence the water level. Additionally, by combining with the attention mechanism, the model is capable of capturing the contributions of specific historical moments for water level prediction. Experimental results demonstrate the effectiveness of our proposed method. Besides, the findings of this paper have practical implications for industry supervisors and cargo-carrying ships, providing scientific guidance for precise ship pre-scheduling.