融合数值模式预报数据的深度学习PM2.5浓度预测模型

A DEEP LEARNING MODEL FOR FORECASTING PM2.5 COMBINED WITH NUMERICAL MODEL DATA

PM2.5污染问题是中国近年来引起广泛关注的环境问题,对PM2.5浓度进行预报有重要意义.传统的预报方法是基于空气动力学理论的数值模式预报方法.最近几年深度学习方法被广泛应用于PM2.5浓度预报问题.之前的深度学习预报方法主要是使用观测站的观测数据建立单点式的预报模型.本文使用ConvLSTM深度神经网络建立模型,在中国及周边区域的PM2.5数据集上实现了网格化的序列到序列预报.模型通过卷积模块提取空间特征,通过LSTM模块提取时间特征,适合解决PM2.5网格化预报问题.同时,模型中使用了再分析数据和模式数据两种不同来源的数据结合起来进行预报,融合了深度学习方法和传统数值模式方法.实验表明,模型的均方根误差比数值模式预报下降30.2%,具有良好的预报效果.

Particulate Matter(PM2.5)pollution caused widespread concern recent years in China.It is very significant to predict PM2.5 concentration.Traditional prediction method is numerical model method,which is based on aerodynamics.Deep learning model is applied to forecast PM2.5 concentration these years.The data used by past research is mainly from monitoring station.In this paper,a sequential grid forecast model is proposed in PM2.5 dataset among China and adjacent regions,and the model is based on Convolutional Long Short-Term Memory(ConvLSTM)deep learning neural network.This model is suitable for PM2.5 prediction,as the convolutional module can extract spatial feature and the LSTM module can extract time feature.The model uses both re-analysis data and numerical model data,and combine deep learning method and numerical model method.The experimental results showed that the Root Mean Square Error(RMSE)decrease 30.23,compared with numerical model method.