深度集成框架的玉米产量预测

Corn Yield Prediction with a Deep Ensemble Framework

摘要: 玉米产量预测对农业生产决策与粮食安全保障具有重要意义。针对传统机器学习模型在农业时间序列预测中对非线性特征和长期依赖关系建模能力不足，以及农业统计数据样本规模有限等问题，本文以河北省玉米产量数据为预测对象，构建了一种融合深度学习与集成学习的玉米产量预测方法。首先，采用灰色关联分析、多元相关系数分析及机器学习特征筛选方法，对玉米产量影响因素进行综合评估；在此基础上，提出一种融合连续小波变换(CWT)、卷积神经网络(CNN)、双向长短期记忆网络(BiLSTM)和Transformer注意力机制的深度预测模型，下文简称为CLT-Net。引入残差感知动态权重机制的Stacking组合预测模型，并结合LSTM优化的条件生成对抗网络(LSTM-CGAN)对训练样本进行数据增强。实验结果表明，与BP神经网络、支持向量机等传统模型相比，CLT-Net在预测精度和稳定性方面均具有明显优势；融合LSTM-CGAN数据增强的组合模型预测准确率最高可达99.58%。研究结果表明，所提出的方法在玉米产量预测中具有较好的预测性能和稳定性。

Abstract: Maize yield prediction plays an important role in agricultural production decision-making and food security assurance. To address the limitations of traditional machine learning models in modeling nonlinear characteristics and long-term dependencies in agricultural time series, as well as the problem of limited agricultural statistical samples, this study takes maize yield data from Hebei Province as the research object and develops a maize yield prediction method integrating deep learning and ensemble learning. First, the influencing factors of maize yield are comprehensively evaluated using grey relational analysis, multivariate correlation analysis, and machine learning–based feature selection methods. On this basis, a deep prediction model integrating continuous wavelet transform (CWT), convolutional neural networks (CNN), bidirectional long short-term memory networks (BiLSTM), and a Transformer attention mechanism is proposed, hereinafter referred to as CLT-Net. Furthermore, a Stacking ensemble prediction model with a residual-aware dynamic weighting mechanism is introduced, and an LSTM-optimized conditional generative adversarial network (LSTM-CGAN) is employed to augment the training samples. Experimental results show that, compared with traditional models such as backpropagation neural networks and support vector machines, CLT-Net exhibits significantly superior prediction accuracy and stability. The ensemble model combined with LSTM-CGAN-based data augmentation achieves a maximum prediction accuracy of 99.58%. The results indicate that the proposed method demonstrates strong predictive performance and robustness in maize yield prediction.