基于Informer的PM2.5浓度预测

针对现有PM_2.5浓度时序预测模型预测精度不高的问题,基于Informer建立了1个Seq2Seq的单站点PM_2.5浓度多步时序预测模型,以历史污染物数据和气象数据为输入,实现对未来一段时间PM_2.5浓度的预测。所构建模型基于ProbSparse (概率稀疏)自注意力机制提取所输入的序列信息,能够广泛地捕获输入序列的长期依赖信息,并对影响因子之间复杂的非线性关系进行建模,从而提高预测准确度。采用北京市2015-2019年逐小时空气污染物数据与气象数据进行模型训练、验证和测试,建立与循环神经网络(RNN)、长短期记忆网络(LSTM)的对比实验并与其他现有研究方法进行比较,结果表明:对未来1~6 h的PM_2.5浓度时序预测,Informer的平均绝对误差(MAE)、均方根误差(RMSE)和可决系数(R2)指标均为最好,实现了较为准确的预测。

PREDICTION OF PM2.5 CONCENTRATION BASED ON INFORMER

For improving the low accuracy of the existing models for time series prediction of PM_2.5 concentration,a Seq2Seq multi-step PM_2.5 concentration prediction model for single-site based on Informer was proposed.With a series of air pollutant data and meteorological data in the past,Informer could make a forecast for PM_2.5 concentration in the future.The constructed model extracted the information of the input sequence based on the probsparse self-attention mechanism,which could widely capture the long-range dependency of the input sequence and model the complex nonlinearity between features,to improve the prediction accuracy eventually.The hourly air pollutant data and meteorological data of Beijing from 2015 to 2019 were used for training,validation and testing.Compared with RNN,LSTM and other existing models,the MAE,RMSE and R2 metrics of Informer were the best for the time series prediction of PM_2.5 concentration in the next 1 to 6 hours,and then a more accurate prediction was realized.