门限重复单元的PM2.5浓度预报方法

PM2.5 concentration real-time forecasting method based on GRU model

针对当前我国重污染天气实时的空气质量预报问题,该文提出了一种基于长短期记忆神经网络的PM2.5浓度实时预报方法。此方法结合了北京市地面空气质量监测数据、天气预报模式的气象预报数据及东亚地区污染物排放清单进行分析,在将高层大气状态及排放状况融入了预报模型的同时,利用LSTM模型模拟区域PM2.5浓度的时序连续变化特征,建立了0~72h的区域PM2.5浓度实时预报模型。实验证明,该方法可以有效表征大气污染物变化的时序特征,从而进行更为精准的长时PM2.5浓度预报。同时,使用门限重复单元作为LSTM神经网络的核心,在保障模型精度的同时,进一步减少了模型训练时间,提高了模型的计算效率。

In order to study the real-time air quality forecasting system suitable for the heavy-polluted weather in China,a PM2.5 concentration real-time forecasting method based on GRU was put forward.This method combined the ground air quality monitoring data of Beijing,the WRF meteorological data and the INDEX-B pollutant emission list to put the high-level atmospheric conditions and pollutant emission conditions into the forecasting model,and established the real-time forecasting model of PM2.5 concentration from 0 to 72 hours based on GRU.Experiments showed that the method could better simulate the temporal state of air pollutant to carry out more accurate long-term PM2.5 concentration forecasting,otherwise,using GRU as the core of the neural network reduced the training time of the model and improved the computational efficiency of the model while ensuring the accuracy of the model.