【摘要】 为精准捕捉不同因素与粮食产量变化趋势之间的关系，为农业生产等提供更为可靠和有效的支持，文中提出基于长短期记忆网络的粮食产量趋势预测方法。在采用WT-EEMD方法处理原始粮食产量样本数据的过程中，利用小波变换（WT）分解获取其高低频分量，使用集合经验模态分解（EEMD）获取其模态分量，对分解结果进行拼接，完成粮食产量变化趋势特征的捕捉；通过基于动态相关性的特征选择方法对影响粮食产量的关键特征作筛选；将提取的特征集与历史粮食产量时间序列一起输入到基于LSTM网络的产量预测模型中，利用粒子群优化算法改进模型参数；通过学习特征向量对粮食产量的长期影响后，输出粮食产量预测结果。实验结果表明：该方法可实现粮食产量趋势预测，预测误差不超过1.63%；经过特征提取与选择，粮食产量预测的R~2指标值可达到0.93，平均绝对误差为0.41，验证了所提方法的有效性。更多还原

【Abstract】 A grain yield trend prediction method based on long short-term memory networks is studied to accurately capture the relationship between different factors and grain yield trends, and provide more reliable and effective support for agricultural production. In the course of processing the raw grain yield sample data with the WT-EEMD method, the high and low frequency components are decomposed and obtained by wavelet transform(WT). The modal components are obtained by ensemble empirical mode decomposition(EEMD). The trend characteristics of grain yield changes are captured by concatenating the decomposition results. The key features that affect grain yield are screened with a feature selection method based on dynamic correlation. The extracted feature set and historical grain yield time series are input into an LSTM networks based yield prediction model. The model parameters are optimized with particle swarm optimization(PSO) algorithm. After learning the long-term impact of feature vectors on grain yield, the grain yield prediction results are output. The experimental results show that the method can achieve grain yield trend prediction with a prediction error of no more than 1.63%. After feature extraction and selection, the value of R~2 for predicting grain yield can reach 0.93, with a mean absolute error(MAE) of 0.41, which verifies the validity of the proposed method.更多还原