| 基于信号分解和深度学习的农产品价格预测 |
| |
| 引用本文： | 王润周,张新生,王明虎.基于信号分解和深度学习的农产品价格预测[J].农业工程学报,2022,38(24):256-267. |
| |
| 作者姓名： | 王润周 张新生 王明虎 |
| |
| 作者单位： | 西安建筑科技大学管理学院，西安 710055 |
| |
| 基金项目： | 国家自然科学基金(41877527)；陕西省教育厅重点科学研究计划项目(20JT033) |
| |
| 摘 要： | 农产品价格的稳定对社会经济与农业发展有重要意义，但农产品价格的波动具有非平稳、非线性、波动性大的特性，较难精确预测。该研究基于信号分解和深度学习，提出一种分解-重构-提取-关联-输出的农产品价格预测模型（CT-BiSeq2seq），并且加入平均气温、养殖成本（大猪配合饲料与尿素价格）、群众关注度等多维度数据来提高模型的预测精度。首先，采用互补集合经验模态分解（Complementary Ensemble Empirical Mode Decomposition，CEEMD）方法把复杂的原始价格序列分解为简单序列。其次，分析皮尔逊相关系数及分解后的子序列，把原始价格序列重构为进高频项、低频项、残差项。再经过时间卷积网络（Temporal Convolutional Network，TCN）提取重构序列的数据特征。随后，构建Biseq2seq模型，解码器引入双向长短期记忆网络（Bi-directional Long Short-Term Memory，Bi-LSTM）网络加强序列数据间的全局关联。最后，通过解码器的LSTM网络输出预测值。以北京丰台区批发市场的白条猪肉价格进行实证分析，该研究提出的CT-BiSeq2seq模型的预测性能显著优于其他基准模型，在滞后天数为11 d达到最优效果。另一方面，在其他数据集也有精确和稳定的预测效果，菠菜、苹果，鸡蛋的均方误差分别为0.627 7(元/kg)、0.463 2(元/kg)、0.552 6(元/kg)，平均绝对误差分别为0.543 1(元/kg)、0.442 5(元/kg)、0.533 9(元/kg)，平均绝对百分比误差分别为3.204 7%、2.236 1%、2.231 4%。同时根据不同数据集的结果发现，价格波动大的农产品适合采用较大的滞后天数，价格波动小的农产品适合采用较小的滞后天数。该模型可以为预测农产品的价格波动提供参考。
|
| 关 键 词： | 农产品价格预测 互补集合经验模态分解 时间卷积网络 双向序列到序列模型 长短期记忆网络 |
| 收稿时间： | 2022/8/26 0:00:00 |
| 修稿时间： | 2022/10/26 0:00:00 |
Agricultural product price prediction based on signal decomposition and deep learning |
| |
| Wang Runzhou,Zhang Xinsheng,Wang Minghu.Agricultural product price prediction based on signal decomposition and deep learning[J].Transactions of the Chinese Society of Agricultural Engineering,2022,38(24):256-267. |
| |
| Authors: | Wang Runzhou Zhang Xinsheng Wang Minghu |
| |
| Institution: | College of Management, Xi''an University of Architecture and Technology, Xi''an 710055, China |
| |
| Abstract: | Abstract: The stability of agricultural product prices is of great significance to social economy and agricultural development, but the fluctuation of agricultural product prices is characterized by non-stationary, non-linear, and high volatility, and it is difficult to predict accurately. Based on signal decomposition and deep learning, this paper proposes a decomposition-reconstruction-extraction-associated-output agricultural product price prediction model (CT-BiSeq2seq) and adds multi-dimensional data such as the average temperature, fertilizer cost (price of pig formula feed and urea), and public attention to improve the model prediction accuracy. Firstly, the complex original price series are divided into simple series by using the complementary ensemble empirical mode decomposition method (CEEMD). Secondly, the original price series is reconstructed into high-frequency items, low-frequency items, and residual items by analyzing the Pearson correlation coefficient and the decomposed subsequence. Next, the data features of the reconstructed sequence are extracted through a temporal convolutional network (TCN). At this time, the input dimension is 7-dimensional data, which can extract the factors that affect the price of agricultural products, and the output steps are equal to the input steps. Subsequently, a Biseq2seq model is constructed which is composed of an encoder and a decoder, and its encoder introduces a bi-directional Long Short-Term Memory network (Bi-LSTM) to strengthen the global correlation between sequence data. Finally, the decoder introduces the LSTM network to output prediction value which can output the predictive value of the number of any steps. Taking the pork price of the Fengtai District wholesale market in Beijing for empirical analysis, the prediction performance of the CT-BiSeq2seq model proposed in this paper is remarkably better than other benchmark models, and the number of lags reached the optimal effect in 11 days. The mean square error, the mean absolute error, and the mean absolute error are 0.661 1 (rmb/kg), 0.501 4 (rmb/kg), and 2.113 8% respectively. This shows that in agricultural product price forecasting, too few days lag is easy to fall into local optimum and cannot fully reflect the overall characteristics. If the lag days are too long, overfitting is easy to occur, and the prediction accuracy will also be reduced. On the other hand, this model also has an accurate and stable prediction effect in other data sets. The mean square error of spinach, apple, and egg is 0.627 7 (rmb/kg), 0.463 2(rmb/kg), and 0.5526(rmb/kg) respectively, the mean absolute error is 0.543 1(rmb/kg), 0.442 5(rmb/kg), and 0.533 9(rmb/kg) respectively, and the mean absolute percentage error is 3.204 7%, 2.236 1% and 2.231 4% respectively. At the same time, according to the results of different data sets, it is found that agricultural products with large price fluctuations are suitable for large lag steps, while agricultural products with small price fluctuations are suitable for small lag steps. For agricultural products with large price changes, the large number of lag days can completely learn the trend of price change. For agricultural products with smaller price changes, due to the relatively stable trend of price change, the short lag days can fit the time sequence relationship. Specifically, the prices of spinach and eggs fluctuate greatly in the data range, and the loss error reaches the minimum when the lag days are 11 days. Apple''s price fluctuates less in the data range, and the loss error reaches the minimum when the lag days are 7 days. This model can provide a reference for forecasting the price fluctuation of agricultural products. |
| |
| Keywords: | agricultural price forecast complementary ensemble empirical mode decomposition temporal convolutional network bi-directional sequence to sequence model long-short term memory |
|
| 点击此处可从《农业工程学报》浏览原始摘要信息 |
| 点击此处可从《农业工程学报》下载免费的PDF全文 |
|
