改进的Prophet融合误差预测模型应用于大气二氧化硫时序预测  

Improved prophet fusion error prediction model applied to atmospheric sulfur dioxide time series prediction

作　　者：虞益军 曾国辉[1] 黄勃[1] 刘瑾[1] 张亦栩 尹玲 周科亮[1] Yu Yijun;Zeng Guohui;Huang Bo;Liu Jin;Zhang YiXu;Yin Ling;Zhou Keliang(School of Elctrical and Electronic Engineering,Shanghai University of Engineering Science,Shanghai,201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《南京大学学报（自然科学版）》2022年第3期440-447,共8页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金青年基金(61802251);上海市科委科技创新行动计划(22S31903700,21S31904200)。

摘　　要：近年来,天气预报中的空气质量预报成为大众尤为关心的热点,由于二氧化硫对空气质量水平变化的影响较大,因此准确预测二氧化硫的浓度变化尤为重要.采用XGBoost模型对Prophet模型的预测误差进行修正,建立改进的Prophet融合误差预测模型,对于空气质量中的关键指标二氧化硫进行时序预测.将时序数据输入Prophet模型,对Prophet生成的预测结果与源输入比较求出残差,构建关于残差的时序序列,利用XGBoost进行残差时序建模,获取残差的修正值,将修正值返回输入到Prophet模型.通过上述步骤,构建特定时序数据融合模型.实验数据表明,融合模型在预测结果中的平均绝对误差和均方根误差分别为1.08和1.38,与Prophet相比,误差指标分别降低2.47,2.45;与差分整合移动平均自回归模型相比,误差指标分别降低0.49,0.47;与XGBoost模型相比,误差指标分别降低0.54,0.52.证明融合模型的预测精度优于上述模型.In the past decades,air quality has become a hot topic of special interest to the public in weather forecasting. Sulfur dioxide contributes significantly to changes in air quality levels. It is especially important to accurately predict changes in sulfur dioxide concentration. In this paper,extreme Gradient Boosting(XGBoost) is used to correct the prediction error of Prophet model. An improved Prophet fusion error prediction model is established based on this technique. Sulfur dioxide time series are used as data,and the time series data are input into the Prophet model. The prediction results generated by Prophet are compared with the source input to find the residuals,and a time series about the residuals is constructed. The residual time series modeling is performed with XGBoost to obtain the correction values of the residuals. Finally,the correction values are returned to the input to the Prophet model. The experimental data show that the mean absolute error and root mean square error of the fusion model in the prediction results are 1.08 and 1.38. Compared with Prophet, the error metrics of our algorithm reduces by 2.47 and 2.45. Compared with Autoregressive Integrated Moving Average model, the error metrics of our algorithm reduces by 0.49 and 0.47. Compared with XGBoost model, the error metrics of our algorithm reduces by 0.54and 0.52. The results proved that the prediction accuracy of the fusion model is better than the above models.

关 键 词：XGBoost PROPHET 时序序列预测 融合预测模型 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]