基于CNN-BiLSTM-Attetion的银杏液流预测模型及环境因子影响研究

Research of ginkgo sap flow prediction model based on CNN-BiLSTM-Attetion and the impact of environmental factors

树木液流受生理活动和多重环境因子的共同作用,表现为非线性和随机性特征,为预测模型的精确度带来挑战。对此,结合CNN卷积层、BiLSTM双向网络结构和注意力机制的优势分别对树干液流序列的局部特征、长期依赖和关键信息进行提取,并根据自测银杏液流数据集构建基于CNN-BiLSTM-Attetion的树干液流预测模型。该模型的R^(2)、MSE和MAE分别为0.9773、0.0029和0.0134,相较于CNN、BiLSTM、XGBoost、RNN和TCN建立的模型均有不同程度的提高。另外,还利用特征工程对环境因子的重要性进行排名,分析银杏树干液流在生长季初期对环境因子的响应规律,对银杏生长季初期的灌溉和养护提供理论依据。

Sap flow is subject to the combined effects of physiological activities and multiple environmental factors,and exhibits nonlinear and stochastic characteristics,which poses a challenge to the accuracy of prediction models.In this regard,the advantages of CNN convolutional layer,BiLSTM bidirectional network structure and attention mechanism are combined to extract the local features,long-term dependence and key information of sap flow sequences,respectively,and the CNN-BiLSTM-Attetion sap flow prediction model is constructed according to the self-test ginkgo sap flow data set.The model has the R^(2),MSE,and MAE of 0.9773,0.0029,and 0.0134,respectively,which are all improved in varying degrees compared with the CNN,BiLSTM,XGBoost,RNN and TCN.In addition,feature engineering is also used to rank the importance of environmental factors and analyze the response regularity of ginkgo sap flow to environmental factors at the beginning of the growing season,which provides a theoretical basis for irrigation and maintenance of ginkgo at the beginning of the growing season.