基于季节趋势分解的PM_2.5浓度混合预测模型

A Hybrid Prediction Model for PM_2.5 Concentration Based on Seasonal Trend Decomposition

摘要

为提高PM_2.5浓度预测精度，提出了基于季节趋势分解的时间序列混合预测模型（Hybrid-X12）。首先，使用季节趋势分解算法将PM_2.5时序分解为趋势-循环、季节和不规则子序列；然后，分别使用自回归移动平均模型（Autoregressive Integrated Moving Average Model,ARIMA）、长短期记忆网络（Long Short Term Memory,LSTM）和支持向量机（Support Vector Machine,SVM）对以上子序列进行预测；最后，集成子序列结果得到最终的预测结果。仿真实验选用华北地区主要六个城市PM_2.5月浓度数据，平均绝对误差（Mean Absolute Error,MAE）、均方根误差（Root Mean Square Error,RMSE）和一致性指数（Index of Agreement,IA）为模型评价指标。实验结果证明混合预测模型能明显提高预测精度，与传统单一模型ARIMA、LSTM和SVM相比，以北京为例，MAE分别降低了18.72%、60.14%和43.15%，验证了季节趋势分解算法有助于时序季节趋势信息挖掘，针对不同特征子序列选择合适的算法充分发挥不同模型优势，为PM_2.5浓度预测提供了新思路。

Abstract

To improve the accuracy of PM_2.5 concentration prediction, this paper introduces a hybrid time series forecasting model with seasonal-trend decomposition (Hybrid-X12). Firstly, the seasonal-trend decomposition algorithm performs the task of decomposing PM_2.5 time series into trend-cycle, seasonal and irregular sub-series; Then, ARIMA (Autoregressive Integrated Moving Average Model), LSTM (Long Short Term Memory), and SVM (Support Vector Machine) are applied to the above sub-series prediction tasks respectively; Finally, the final prediction result comes from the integration of the predicted results of sub-series. The simulation experiment selected PM_2.5 monthly concentration from six major cities in North China and used MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and IA (Index of Agreement) as model evaluation indicators. The experimental results demonstrated that the hybrid prediction system can significantly enhance prediction accuracy. Compared with the traditional single model ARIMA, LSTM and SVM, the MAE of the proposed model in Beijing is reduced by 18.72%, 60.14% and 43.15%, respectively. This verifies that the seasonal-trend decomposition algorithm is helpful for mining seasonal-trend information in time series. It can be concluded that selecting appropriate algorithms for sub-series with different characteristics ensures the full utilization of the advantages of different models, providing new ideas for PM_2.5 concentration prediction.

Graphical abstract

关键词

PM_2.5浓度预测 / 季节趋势分解 / 自回归移动平均模型 / 长短期记忆网络 / 支持向量机

Key words

PM_2.5 concentration prediction / seasonal trend decomposition / autoregressive integrated moving average model / long short term memory / support vector machine

基金资助

国家社会科学基金(20BTJ045)