A Hydrological Data Prediction Model Based on LSTM with Attention Mechanism

With the rapid development of IoT, big data and artificial intelligence, research application data-driven hydrological models are increasing. However, when conducting time series analysis, many prediction often directly based on following assumptions: hydrologic normal, homogeneous, smooth non-trending, which not always all true. To address related issues, a solution for short-term forecasting is proposed. Firstly, feature test conducted to verify whether non-trending; secondly, sequence-to-sequence (seq2seq)-based water level model (LSTM-seq2seq) proposed improve accuracy prediction. The uses long memory neural network (LSTM) as an encoding layer encode historical flow sequence into context vector, another LSTM decoding decode vector in order predict target runoff, by superimposing attention mechanism, aiming at improving accuracy. Using experimental regarding Chu River, compared other analysis normality, smoothness, homogeneity trending different data. results show that greater than set without these characteristics with trend. Flow Runcheng, Wuzhi, Baima Temple, Longmen Town, Dongwan, Lu’s Tongguan used input sets train evaluate model. Metrics RMSE NSE convergence speed LSTM-seq2seq LSTM-BP higher models. Furthermore, process fastest among