基于自注意力机制和改进的K-BiLSTM的水产养殖水体溶解氧含量预测模型

Prediction model of dissolved oxygen content in aquaculture water based on self-attention mechanism and improved K-BiLSTM

为精确预测水产养殖水体溶解氧含量,本研究提出一种基于自注意力机制(ATTN)和改进的K-means聚类-基于残差和批标准化(BN)的双向长短期记忆网络(BiLSTM)的水产养殖水体溶解氧含量预测模型。首先,根据环境数据的相似性,使用改进的K-means算法将数据划分成若干个类别;然后,在BiLSTM基础上构建残差连接和加入BN完成高层次特征提取,利用BiLSTM的长期记忆能力保存特征信息;最后,引入自注意力机制突出不同时间节点数据特征的重要性,进一步提升模型的性能。试验结果表明,本研究提出的基于自注意力机制和改进的K-BiLSTM模型的平均绝对误差为0.238、均方根误差为0.322、平均绝对百分比误差为0.035,与单一的BP模型、CNN-LSTM模型、传统的K-means-基于残差和BN的BiLSTM-ATTN等模型相比具有更优的预测性能和泛化能力。

In order to accurately predict the content of dissolved oxygen(DO)in aquaculture water,a prediction model of dissolved oxygen content in aquaculture water based on self-attention mechanism(ATTN)and improved K-means clustering-bidirectional long-term and short-term memory network(BiLSTM)was proposed.Firstly,according to the similarity of environmental data,the improved K-means algorithm was used to divide environmental data into several categories.Then,based on BiLSTM,residual connection was constructed and batch normalization(BN)was added to complete high-level feature extraction,and the feature information was saved by the long-term memory ability of BiLSTM.Finally,the self-attention mechanism was introduced to highlight the importance of data characteristics at different time nodes,which further improved the performance of the model.The experimental results showed that the mean absolute error(MAE),root mean square error(RMSE)and average absolute percentage error(MAPE)of the hybrid model based on self-attention mechanism and improved K-BiLSTM were 0.238,0.322 and 0.035,respectively.Compared with single BP model,CNN-LSTM model and traditional K-means-BiLSTM-ATTN model based on residual and BN,the model constructed in this study had better prediction performance and generalization ability.