上海市PM_(2.5)浓度延伸期预测模型的构建及评估

Establishment and evaluation of prediction model for PM_(2.5)concentration extension period in Shanghai

基于PM_(2.5)浓度、次季节到季节预测项目(S2S)三家模式(CMA、UK和ECMWF)历史回报气象要素(2000~2014年)作为预测因子,利用长短时记忆神经网络(LSTM)深度学习方法和轻量级梯度提升机(LGBM)算法开发了上海PM_(2.5)延伸期预测模型,并应用LSTM模型建立了上海秋冬季(11月~次年2月)PM_(2.5)浓度延伸期预测融合模型.结果显示:融合模型11~40d逐候预测与实况的相关系数为0.47~0.76,比单一模型相关系数上升23.5%~31.1%;融合模型RMSE介于19~25.1μg/m^(3)之间,较单一模型下降19%~19.3%.融合模型能够较好地预报出未来11~40d上海秋冬季PM_(2.5)浓度的总体趋势、浓度峰值谷值的变化及发生时间等关键特征,逐候HSS技巧评分在0.18~0.5之间,显示出较好的预测技巧.对于典型污染过程个例的预测而言,融合模型在不同的预测时效预测准确率存在差异,提前11−40d的总体预测准确率为75.5%.对于持续3d及以上的3次污染过程,11~40d预测准确率达到100%.融合模型的预报时效可达40d,是目前污染数值预报模型(一般96~240h)预报时效的近4倍,且运算速度快,能够节省大量计算资源和时间成本.

In this study,two newly developed long short-term memory(LSTM)model and Light Gradient Boosting Machine(LGBM)algorithms were introduced for application in extended-range forecasting of PM_(2.5)in Shanghai by incorporating three members of the sub-seasonal-to-seasonal prediction project(S2S)forecasting,six prediction models were obtained.Therefore,based on six models forecast results,an accurate~40d PM_(2.5)prediction fusion model over Shanghai was developed by LSTM algorithm,providing new insights for air pollution extended-range forecasting.The evaluation results indicated the fusion model exhibited not only much better accuracy but also captured the pollution process more closely compared to than any of the six single model.The correlation coefficients for the fusion model forecasts on lead times of 11~40 days ranged from 0.47 to 0.76,23.5%~31.1%higher than other six single model.The root mean square errors(RMSE)for the fusion model forecasts on lead times of 11~40 days ranged from 19 to 25.1μg/m^(3),19%~19.3%lower than other six single model.The fusion model could not only better predict the overall trend of PM_(2.5)concentration,but also the occurrence time of peak and valley,and its Heidke Skill Scores(HSS)was between 0.18 and 0.5,showing a good prediction skill.The fusion model could predict pollution episodes at a lead of 11~40 days,and the overall prediction accuracy was 75.5%at a lead of 11~40 days.For three pollution episodes lasting for 3 days or more,the prediction accuracy reached 100%at a lead of 11~40 days.The prediction efficiency of the fusion model was up to 40days,nearly four times that of the current pollution numerical prediction model(generally 96~240h),and the calculation speed was fast,which could save a lot of computing resources and time costs.