丁鹏, 邹晓荣, 丁淑仪, 白思琦. 基于CNN-BiLSTM模型的黄鳍金枪鱼渔获量与气候因子关系研究[J]. 南方水产科学, 2024, 20(2): 19-26. DOI: 10.12131/20230190
引用本文: 丁鹏, 邹晓荣, 丁淑仪, 白思琦. 基于CNN-BiLSTM模型的黄鳍金枪鱼渔获量与气候因子关系研究[J]. 南方水产科学, 2024, 20(2): 19-26. DOI: 10.12131/20230190
DING Peng, ZOU Xiaorong, DING Shuyi, BAI Siqi. Study on relationship between Thunnus albacares catches and climatic factors based on CNN-BiLSTM model[J]. South China Fisheries Science, 2024, 20(2): 19-26. DOI: 10.12131/20230190
Citation: DING Peng, ZOU Xiaorong, DING Shuyi, BAI Siqi. Study on relationship between Thunnus albacares catches and climatic factors based on CNN-BiLSTM model[J]. South China Fisheries Science, 2024, 20(2): 19-26. DOI: 10.12131/20230190

基于CNN-BiLSTM模型的黄鳍金枪鱼渔获量与气候因子关系研究

Study on relationship between Thunnus albacares catches and climatic factors based on CNN-BiLSTM model

  • 摘要: 为探究气候因子对黄鳍金枪鱼渔获量的影响,根据1960—2021年的南方涛动指数 (SOI)、太平洋年代际涛动 (PDO)、北大西洋涛动 (NAO)、北太平洋指数 (NPI)、全球海气温度异常指标 (dT) 以及厄尔尼诺相关指标 (Niño1+2、Niño3、Niño4以及Niño3.4) 等9种气候因子数据和全球黄鳍金枪鱼渔获量数据,采用相关性分析、BP神经网络、长短期记忆网络 (LSTM) 模型、双向长短期记忆网络 (BiLSTM) 模型和卷积神经网络结合双向长短期记忆网络 (CNN-BiLSTM) 模型对低频气候因子与黄鳍金枪鱼渔获量的关系进行了研究。结果表明,气候变化表征因子对黄鳍金枪鱼渔获量的重要性依次为dT>SOI>Niño1+2>PDO>NPI>NAO,其对应的最佳滞后年限分别为0、11、6、5、15、0年。CNN-BiLSTM模型的预测效果最优,其后依次为BiLSTM模型、LSTM模型、BP神经网络模型。最优预测模型显示预测值与实际值的拟合优度为0.887,平均绝对误差为0.125,均方根误差为0.154,预测值与实际值变化趋势基本一致,模型拟合效果良好。

     

    Abstract: To explore the impact of climatic factors on Thunnus albacares catches, we studied its relationship with low-frequency climatic factors by using correlation analysis, BP neural network, LSTM model, BiLSTM model and CNN-BiLSTM model based on the data of nine climate factors, including Southern Oscillation Index (SOI), Pacific Decadal Oscillation (PDO), North Atlantic Oscillation (NAO), North Pacific Index (NPI), global sea-air temperature anomaly index (dT), El Niño-related indexes (Niño1+2, Niño3, Niño4, Niño3.4) from 1960 to 2021, as well as global T. albacares catches data. The results show that the importance of climate change characterization factors on T. albacares catches followed a descending order of dT>SOI>Niño1+2>PDO>NPI>NAO, whose corresponding optimal lag periods were 0, 11, 6, 5, 15 and 0 years, respectively. CNN-BiLSTM model had the highest prediction accuracy, followed by BiLSTM, LSTM and BP. The goodness of fit between the predicted and actual values of CNN-BiLSTM model was 0.887, with a mean absolute error of 0.125 and a root mean square error of 0.154. The trend of predicted values and actual values was basically consistent, indicating a good model fitting effect.

     

/

返回文章
返回