摘要： 大眼金枪鱼（Thunnus obesus）是太平洋延绳钓的主捕鱼种之一，针对多数传统预报模型存在的问题，提出了基于经验模态分解和双向长短时记忆神经网络（EMD-BiLSTM）的渔场预报新模型，以实现一种新的面向渔业应用的产量预报方法。首先，通过经验模态分解机制（EMD）对单位捕捞努力量渔获量（CPUE）序列进行分解，得到不同尺度的分解分量（IMF）；然后结合各影响因子对IMF分量分别建立双向长短时记忆神经网络渔场预报模型（Bi-LSTM），使神经网络的数据处理优势得以充分发挥；最后整合各项结果作为最终预报值。结果显示：与Bi-LSTM模型相比，均方根误差和绝对误差分别降低0.053和0.018；与BP模型相比，均方根误差和绝对误差分别降低0.208和0.048。研究表明，EMD-BiLSTM模型具有较高的预报准确率，可为渔场预报相关研究提供一种新思路。

关键词: 渔场预报, EMD-BiLSTM, 神经网络, 分解分量, 大眼金枪鱼

Abstract: Thunnus obesus is one of the main fish species for longline fishing in the Pacific. Aiming at the problems of most traditional forecast models, a new fishing ground forecast model based on empirical mode decomposition and two-way long and short-term memory neural network (EMD-BiLSTM) is proposed to realize a new production forecast method for fishery applications. First, the empirical mode decomposition mechanism (EMD) is used to decompose the catch per unit effort (CPUE) sequence to obtain decomposition components (IMF) of different scales. Then the influencing factors are combined to establish the two-way long and short-term memory neural network fishing ground forecast model (Bi-LSTM) respectively for the IMF components, so that the data processing advantages of the neural network can be fully utilized. Finally, the results are integrated as the final forecast value. The results show that compared with the Bi-LSTM model, the root mean square error and absolute error are reduced by 0.053 and 0.018, respectively; compared with the BP model, the root mean square error and absolute error are reduced by 0.208 and 0.048, respectively. Studies have shown that the EMD-BiLSTM model has a high forecast accuracy rate, which provides a new idea for related research on fishing ground forecast.

Key words: fishing ground forecast, EMD-BiLSTM, neural network, decomposition components, Thunnus obesus