Abstract
Harmful algal blooms (HABs) often cause great harm to fishery production and the safety of human lives. Therefore, the detection and prediction of HABs has become an important issue. Machine learning has been increasingly used to predict HABs at home and abroad. However, few of them can capture the sudden change of Chl-a in advance and handle the long-term dependencies appropriately. In order to address these challenges, the Long Short-Term Memory (LSTM) based spatial-temporal attentions model for Chlorophyll-a (Chl-a) concentration prediction is proposed, a model which can capture the correlation between various factors and Chl-a adaptively and catch dynamic temporal information from previous time intervals for making predictions. The model can also capture the stage of Chl-a when values soar as red tide breaks out in advance. Due to the instability of the current Chl-a concentration prediction model, the model is also applied to make a prediction about the forecast reliability, to have a basic understanding of the range and fluctuation of model errors and provide a reference to describe the range of marine disasters. The data used in the experiment is retrieved from Fujian Marine Forecasts Station from 2009 to 2011 and is combined into 8-dimension data. Results show that the proposed approach performs better than other Chl-a prediction algorithms (such as Attention LSTM and Seq2seq and back propagation). The result of error prediction also reveals that the error forecast method possesses established advantages for red tides prevention and control.
Highlights
Natural disasters occur more and more frequently with global warming
Our main contributions are as follows: (1) First, we developed the DA-Recurrent Neural Network (RNN) to predict the Chl-a values, we found that dual attention mechanism performed better than other models both on root mean squared error (RMSE) and mean absolute error (MAE), and dual-stage attention based RNN (DA-RNN) could predict the mutation of Chl-a value better than other models, which is significant for red tide prevention
Because the cost and loss caused by red tide disaster are huge, the instability of red tide prediction needs to be forecasted in advance, so DA-RNN is applied to predict the fluctuation of model errors, which can provide a reference for the forecast reliability
Summary
Natural disasters occur more and more frequently with global warming. Abnormal natural disasters usually occur suddenly, such as rainstorms, heavy fog, and earthquakes. These natural disasters have the characteristics of suddenness, short duration, and serious impact, which bring difficulties to forecast and control. Most of the existing environment forecast models use the total average error to evaluate prediction results, and those models will have great error fluctuations when natural disasters occur suddenly. When predicting the occurrence of natural disasters, we should know the prediction results of the model, and grasp the fluctuation of model errors, which can provide a reference for prediction reliability
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
