基于时域卷积网络与Transformer的茶园蒸散量预测模型

Evapotranspiration Prediction Model of Tea Garden Based on Temporal Convolutional Network and Transformer

摘要:

在茶园水资源管理中，蒸散量（Evapotranspiration，ET）是评估作物水分需求的关键指标，由于茶园蒸散量预测具有时序性、不稳定性以及非线性耦合等特点，目前的茶园蒸散量预测模型存在预测精度较低的问题，针对此问题本文提出了一种新型的茶园蒸散量预测模型。首先使用互信息算法（Mutual information，MI）与主成分分析算法（Principal component analysis，PCA）相融合的数据处理算法（MIPCA），筛选强相关的特征并提取主成分；其次将时域卷积网络（Temporal convolutional network，TCN）与Transformer融合，利用灰狼算法（Grey wolf optimization，GWO）优化超参数，捕捉茶园数据的全局依赖关系；最后整合2个网络构建了MIPCA-TCN-GWO-Transformer模型，通过消融试验和对比试验验证了模型性能，并对模型在不同时间步长下的性能进行测试。结果表明，该模型平均绝对百分比误差（Mean absolute percentage error，MAPE）、均方根误差（Root mean square error，RMSE）和决定系数（Coefficient of determination, R2）3个评价指标分别为0.015 mm/d、0.312 mm/d和0.962，优于长短期记忆模型 (Long short term memory，LSTM)等传统预测模型。在小时尺度、日尺度和月尺度下的R2分别为0.986、0.978和0.946，在不同时间步长下展现了良好的适应性和准确性。本文构建的MIPCA-TCN-GWO-Transformer模型具有较高的预测精度和稳定性，可为茶园水资源优化管理和灌溉制度制定提供科学参考。

Abstract:

In tea garden water resource management, accurately assessing crop water requirements is crucial, with evapotranspiration (ET) serving as a key indicator. The challenges posed by the time series nature, instability, and non-linear coupling in tea garden data were addressed by introducing a novel evapotranspiration prediction model. Firstly, a data processing algorithm, mutual information-principal component analysis (MIPCA), was employed to integrate mutual information (MI) and principal component analysis (PCA), facilitating the selection of features strongly correlated with tea garden transpiration and the extraction of principal components. Subsequently, the temporal convolutional networks (TCN) was integrated with Transformer to construct a new model. Specifically, the grey wolf optimization (GWO) algorithm was employed to optimize the hyperparameters of the TCN, followed by the utilization of the Transformer to capture global dependencies. Ultimately, the two networks were integrated to propose the hybrid model MIPCA-TCN-GWO-Transformer. The model performance was validated through ablation experiments and comparative analyses, while also examining the model’s performance across different time scales. The results showed that the model’s three evaluation indicators such as mean absolute percentage error (MAPE), root mean square error (RMSE) and coefficient of determination (R2) were 0.015 mm/d, 0.312 mm/d and 0.962, respectively, which as better than that of traditional prediction models such as long short term memory (LSTM). R2 at hourly scale, daily scale and monthly scale were 0.986, 0.978 and 0.946, respectively, showing good adaptability and accuracy at different time scales. The MIPCA-TCN-GWO-Transformer model constructed had high prediction accuracy and can provide scientific reference for the optimal management of tea garden water resources and the formulation of irrigation systems.