长短期记忆神经网络在叶绿素a浓度预测中的应用

针对传统人工神经网络对叶绿素a浓度预测存在训练速度慢、收敛精度低、易陷入局部最优,尤其是无法灵活的利用任意长度的历史信息对叶绿素a浓度进行预测等问题,本文根据海洋各要素与叶绿素a浓度之间的长短期依赖程度,对叶绿素a浓度与各要素间的关系进行界定,分别将各要素与叶绿素a浓度之间的长期依赖关系与短期依赖关系分割开来,并且在长短期记忆(Long Short-Term Memory, LSTM)神经网络模型的基础上构建融合的LSTM预测模型,模型中的长期依赖关系与短期依赖关系分别使用不同的神经元,最终在模型的最上层进行长短期融合。本文选取三都澳站位的连续监测资料作为实验数据,实验结果表明本文构建的模型不仅具有训练误差下降快的优点,与其他3种经典的神经网络模型相比,预测精度也有显著提高。

Application of long term and short term memory neural network in prediction of chlorophyll a concentration

Prediction of chlorophyll a concentration in traditional artificial network methods has some disadvantages, such as slower training speed, lower convergence precision, and easy to fall into local optimum situation. In particular, it is not possible to flexibly use historical information of any length to predict chlorophyll a concentration. To solve these problems, this paper defines the relationship between chlorophyll a concentration and various elements, depending on the long-term and short-term dependence between elements and the concentration of chlorophyll a. In this way, the long-term dependence between each element and the chlorophyll a concentration is separated from the short-term dependence. Then, based on the Long Short-Term Memory (LSTM), a merged LSTM prediction model was proposed. In this model, short and long term dependencies were presented respectively by different neurons and finally merged at the top of the model. The experimental data involves the continuous monitoring data of the station of Sandu Ao. The main result includes that the model has the advantage of fast reduction of training error, but also has significantly higher prediction accuracy compared with other three classical neural network models.