Hierarchical attention-based context-aware network for red tide forecasting

Abstract

Chlorophyll forecasting is helpful for understanding characteristics of red tides, thus enabling early warning. In practice, it is formulated as a time series forecasting problem, which aims to predict the future chlorophyll concentration with the observation of various exogenous factors (e.g., PH, turbidity, etc.) and corresponding historical chlorophyll. However, existing methods hardly satisfy the chlorophyll forecasting because the interaction in observed data are complicated and changeable. In this work, we propose a fine-grained hierarchical attention-based context-aware network. The model consists of two hierarchical attention networks, with one named EF-net focusing on exogenous factors and the other named TF-net executing on chlorophyll. EF-net introduces factor-level and sequence-level attention to learn exogenous factors and time steps that are beneficial to prediction. TF-net firstly conducts context-aware attention to capture the interaction in historical time series from EF-net. Then a specially designed contextual long short-term memory network employs the interactive information to benefit the accuracy for chlorophyll. To suppress noisy information, we employ a gated fusion method to fuse the outputs of EF-net and TF-net. Experiments with two real world data sets show that our proposed model achieves 16.02%, 10.65%, and 22.45% improvements in average MAE, RMSE, and MAPE in seven-step-ahead predictions compared with baseline methods. Visualization of attention weights shows that sea surface temperature, air temperature, standard atmospheric pressure, and PH are significant for chlorophyll prediction.