基于时间序列GA-SVR的水产品价格预测模型及验证

Forecasting model and validation for aquatic product price based on time series GA-SVR

水产品价格的准确预测有助于合理规划水产养殖,正确引导水产行业的发展。根据水产品价格序列的非线性、非平稳和周期性特点,提出了一种基于时间序列遗传优化(genetic algorithm,GA)支持向量回归(support vector regression,SVR)的水产品价格预测模型。该模型首先通过时间序列分析方法对价格序列进行平稳性检验和确定相关阶数,得到训练数据集;再利用遗传算法对支持向量回归模型的参数组合进行寻优,使用优化后的参数建立支持向量回归模型,然后使用模型进行预测。分别选取桂鱼、基围虾、梭子蟹的价格数据对模型进行验证,选取2011-2014年的数据作为训练集,对2015年价格进行预测,结果表明:桂鱼、基围虾、梭子蟹的平均绝对误差分别为6.70%、7.82%、14.76%,均方根误差分别为5.853 1、23.701 1、13.858 0,且优于基于时间序列的SVR模型及BPANN模型的预测结果,可以为水产品价格的预测提供依据。

Fluctuations in aquatic product prices have an important impact on the development of the aquaculture industry. Accurate forecasting results can enable farmers to keep abreast of changes in the market and rationally plan aquaculture. Based on the non-linear, non-stationary and periodicity of the aquatic product price series, a genetic algorithm（GA） support vector regression（SVR） model based on time series for forecasting aquatic product price was presented in this paper. Firstly, the time series method was applied to the price series, the autocorrelation function was used to judge the stability, and the partial correlation coefficient was used to judge the data items, then the training data set was obtained. After that, the genetic algorithm was used to optimize the parameters of support vector regression. The parameters of SVR based on radial basis kernel function were kernel function coefficient, penalty factor, and loss parameter. We designed these three parameters by using real number coding individual representation. We used the selection operation to select the mean square error as the fitness function, to calculate the fitness value of each individual, and to select the individuals with better fitness value. By use of the crossover operation, we selected the point intersecting as crossover operator with different individuals, respectively in a corresponding position to a certain probability. The nature of mutation operation was used to enhance the local search algorithm, and avoid falling into the local minimum. We mutated individual to a certain probability and change its current value, then generated new population. We introduced the mechanism of 5-fold cross validation to the process of each iteration to obtain the optimized parameter combination. Finally, the support vector regression model was established by using the optimized parameters to forecast the price of aquatic products in the next period. In this paper, we selected mandarin fish, metapenaeus ensis and portunus trituberculatus as the experimental objects. The experimental data we used were the value of aquatic product price from January 2011 to December 2015 of Beijing Xinfadi market website（http：//www.xinfadi.com.cn）. After craw ling the web data-including 1,541 records of mandarin fish, 1,525 records of metapenaeus ensis and 1,430 records of portunus trituberculatus, we calculated the monthly average price to represent the price of a period. We trained the proposed model by using data from 2011 to 2014, and forecasted the price of the next year. Through comparing with the real value, the mean absolute percent error of mandarin fish, metapenaeus ensis and portunus trituberculatus was 6.70%, 7.82% and 14.76%, with corresponding root mean square error of 5.8531, 23.7011 and 13.8580, respectively. After surveying the market, we found that the results of forecasting were more in line with the actual situation. In this paper, the SVR model and the BP neural network model based on time series were all used in contrast experiment of our model. The experiment results showed that our model was superior. According to the characteristics of aquatic product price in this paper, we proposed a combined model for the determination of the relevant items of the aquatic product price series, the selection of the kernel function and the parameter optimization. The results showed that the proposed model can provide the basis for the forecasting of aquatic product price.