基于季节分解和长短期记忆的北京市鸡蛋价格预测

Egg price forecasting in Beijing market using seasonal-trend decomposition procedures based on seasonal decomposition and long-short term memory

准确把握北京市场鸡蛋价格波动特征和规律,及时预测鸡蛋价格波动趋势,不仅是农业进入新发展阶段的首都"菜篮子"工程建设的需要,而且有利于社会经济的稳定发展。该文选取北京市月度鸡蛋价格作为试验数据,在对北京市鸡蛋价格历史数据分析的基础上,根据鸡蛋价格序列的非线性、季节性和周期性特征,提出一种基于时间序列季节性分解(Seasonal-trend Decomposition Procedure Based on Loess,STL)和长短期记忆网络(Long-short Term Memory,LSTM)组合的鸡蛋价格预测模型。通过采用LSTM模型实现对由STL方法分解的鸡蛋价格波动成分的趋势成分及剩余成分和用季节朴素方法(Seasonal-na?ve, Sna?ve)对鸡蛋价格波动的季节成分分别进行预测,可以获取未来鸡蛋价格的综合预测值。研究结果表明:2000-2018年北京市鸡蛋价格在整体呈现上升趋势,且存在"春低秋高"的季节性和随机波动特征;该研究构建的STL-LSTM模型在预测步长分别为1、3、6时的均方根误差分别为0.19、0.33、0.43;平均绝对百分比误差分别为1.91、3.53、4.58,均优于长短期记忆网络、支持向量回归(Support Vector Regression,SVR)和差分整合移动平均自回归(Autoregressive Integrated Moving Average Model,ARIMA)模型,可以为预测预警北京市场鸡蛋价格异常波动情况、为行业和政府主管部门保障北京市场鸡蛋供应决策提供参考依据。

Egg price has been attracting public attentions from every community in Beijing market. It is necessary to obtain timely information for the fluctuation of the future egg price, particularly on the demand and supply of table eggs for human consumption. A lot of efforts have been made to accurately forecast future egg price in short, medium or long terms. However, there are many factors affecting egg prices to make the prediction challenging. In this paper, a hybrid model was proposed to forecast egg price by combining seasonal-trend decomposition procedures based on loess(STL) and long short-term memory(LSTM), denoted as STL-LSTM. In decomposition, a time series can be splitted into three components: seasonality, trends and remainder fluctuation. A more stable variance can be obtained from the non-linear, seasonal and periodic each part of egg price. Then, LSTM can be used to capture appropriate behaviors and predict precisely the trends and remainder parts of egg price, respectively, while the seasonal-naive method can be used to predict seasonal trends in a 12-month cycle. The results from three parts were summarized into a total price forecast. The egg price data that used in this study were collected from the China animal husbandry, covering from January 2000 to December 2018 in Beijing egg markets. The monthly data from January 2000 to December 2017 were used as training set, whereas the 12 monthly data in 2018 were used as testing set in the proposed model. The method was evaluated by using the relative error(RE), root mean square error(RMSE) and the mean absolute error percentage(MAPE). The results show that there was an overall upward trend for the egg price in the Beijing market from January 2000 to December 2018, with the seasonal fluctuation of "low spring and high autumn", and random fluctuations. The decomposition indicated that the trend component was the main contributor to egg price fluctuations, where the contribution rate decreased from 71.18% to 56.84% during the test period. The influence of seasonal and remaining components on egg prices increased in recent years, with the contribution rates of 34.24% and 8.92%, respectively. In STL-LSTM model, when the step size was given as 1, 3 and 6, the evaluating indexes were optimum: the relative error of 3.67%, 6.49% and 7.22%, the root mean square errors of 0.19, 0.33, and 0.43, and the average absolute percentage errors of 1.91, 3.53, and 4.58. In terms of the evaluating indexes, the proposed STL-LSTM model demonstrated most efficiency to predict egg prices, compared with the previous models, such as separate LSTM, support vector regression(SVR) and the autoregressive integrated moving average(ARIMA). The proposed model can be expected to extend on price predictions of other similar agricultural product in the future. The findings can provide a great potential to accurately forecast the future egg price for market strategies in animal husbandry.