统计与数据科学学院知识库

在面对全球粮食市场的不确定性和挑战，粮食产量预测变得至关重要。粮食产量预测不仅可以帮助我们及时应对自然灾害、气候变化等因素对农业生产的影响，还能为国家粮食安全战略的制定提供关键信息。通过粮食产量预测，我们能够更好地规划农业生产、合理配置资源，进而保障粮食供应，维护国家经济的稳定和人民的福祉。因此，加强粮食产量预测工作，提高预测准确性和可靠性，对于确保粮食安全、促进农业可持续发展具有重要意义。

本文通过收集1949年至2022年的我国粮食产量数据和影响因素等数据，结合机器学习算法的强大分析能力，建立了粮食产量预测模型。接着为了比较不同的数据集对模型进行预测的效果，继续收集了1949年至2022年江西省粮食产量数据和影响因素等数据用来训练和评估模型在面对不同数据特征和分布时的稳健性和可靠性。结果显示，两种不同的数据集在模型训练与测试中均表现良好。在这两个数据集上，BP神经网络和随机森林模型表现出较好的训练和测试效果，其次是支持向量回归模型和长短期记忆神经网络模型。

本文提出采用Stacking集成算法，将RF、BP、SVR以及LSTM模型进行融合，构建Stacking集成模型。将构建后的Stacking算法集成模型与以往的平均法组合模型进行预测效果对比分析，并采用交叉验证的方法来进一步评估模型的稳定性。实验结果表明，相较于传统的平均法集成模型与单一机器学习模型，Stacking算法构建的集成模型均展现出优异的预测效果。在5折交叉验证中，RF-BPNN-SVR-Stacking模型预测效果显著强于其它模型；在10折交叉验证中，RF-BPNN-SVR-LSTM-Stacking模型的预测实验误差最小，预测效果最好。RF-BPNN-SVR-Stacking、RF-BPNN-SVR-LSTM-Stacking模型相较于传统的平均法集成模型在不同的交叉验证方式下均展现出优异的预测效果，为粮食产量预测提供了新的优化思路，有望提升预测精度。

In the face of uncertainty and challenges in the global grain market, predicting grain production has become crucial. Grain production forecasting not only helps us timely respond to factors such as natural disasters and climate change affecting agricultural production, but also provides key information for formulating national food security strategies. Through grain production forecasting, we can better plan agricultural production, allocate resources reasonably, thereby ensuring food supply, maintaining economic stability, and safeguarding the well-being of the people. Therefore, strengthening grain production forecasting work and improving forecasting accuracy and reliability are of great significance for ensuring food security and promoting sustainable agricultural development.

This paper establishes grain production prediction models by collecting data on grain production from 1949 to 2022 in China, along with other influencing factors, and leveraging the powerful analytical capabilities of machine learning algorithms. To compare the effects of different datasets on model predictions, grain production data and influencing factors from Jiangxi Province from 1949 to 2022 were collected to train and evaluate the robustness and reliability of the models when facing different data features and distributions. The results show that both datasets performed well in model training and testing. On both datasets, the BP neural network and random forest models exhibited good training and testing performance, followed by the SVR model and long short-term memory neural network model.

This paper proposes the use of Stacking ensemble learning algorithm to integrate RF, BP, SVR, and LSTM models, constructing a Stacking ensemble model. The predictive performance of the constructed Stacking ensemble model is compared with that of traditional averaging ensemble models, and the stability of the models is further evaluated using cross-validation. Experimental results demonstrate that, compared to traditional averaging ensemble models and single machine learning models, the Stacking ensemble models exhibit excellent predictive performance. In 5-fold cross-validation, the RF-BPNN-SVR-Stacking model significantly outperformed other models; in 10-fold cross-validation, the RF-BPNN-SVR-LSTM-Stacking model showed the smallest experimental prediction error and the best predictive performance.

The RF-BPNN-SVR-Stacking and RF-BPNN-SVR-LSTM-Stacking models demonstrate excellent predictive performance under different cross-validation methods compared to traditional averaging ensemble models, providing new optimization ideas for grain production prediction and potentially improving prediction accuracy.