基于深度学习与随机森林的PM_(2.5)浓度预测模型

A PM_(2.5)prediction model based on deep learning and random forest

针对PM_(2.5)浓度预测中传统机器学习算法无法对数据内部隐藏特征进行深层次挖掘,而深度学习算法在数据较少情况下效果不佳的问题,综合考虑深度学习与随机森林的特点,提出一种基于深度学习与随机森林的PM_(2.5)浓度预测组合模型。模型以气溶胶光学厚度(AOD)遥感数据、气象再分析数据和PM_(2.5)地面观测数据构建训练数据集,通过深度学习方法对训练数据内部深层次隐含特征进行提取,将提取得到的隐含特征用于随机森林模型训练,并使用随机森林回归算法得到PM_(2.5)浓度的预测值。为验证方法的有效性,以河南省区域2018年—2019年的PM_(2.5)浓度估算为例,将原始特征与利用CNN、LSTM和CNN_LSTM所提取特征共同构建的新特征分别通过随机森林回归、支持向量回归以及K近邻回归等3种传统机器学习方法进行训练和预测。实验结果表明,在较少数据情况下PMCOM模型无论是在整体预测还是在分季节预测场景下均具有较好的预测精度,其中以LSTM为特征选择器,RF为回归器的组合模型是本实验的最优模型,在即使只有35%的数据作为训练样本时,整体预测实验中R^(2)仍可达0.89,各季节预测实验中R^(2)均在0.75以上。

At present,the situation of environmental pollution in China is grim,among which regional compound air pollution dominated by PM_(2.5)is the most prominent.Aerosol Optical Depth(AOD)is a key physical quantity used to characterize the degree of atmospheric turbidity,which represents the intensity of aerosol light reduction.Many studies have shown that there is a strong correlation between AOD and PM_(2.5).Using the AOD data obtained by satellite remote sensing combined with other influencing factors to analyze the change mechanism of PM_(2.5)is of great significance to air pollution prevention and the protection of human health.The diffusion of PM_(2.5)is an extremely complicated process,and the PM_(2.5)prediction model based on the statistical regression method can only describe a relatively simple nonlinear relationship.However,the estimation of PM_(2.5)is considered to be a more complex multivariable nonlinear problem.Compared with statistical regression models,the PM_(2.5)prediction model based on traditional machine learning algorithms can deal with more complex nonlinear problems.However,its ability to process historical data is still limited,so it is difficult to mine the variation law of pollutant concentrations from the perspective of big data.Compared with the traditional machine learning method,the models based on deep learning can dig deep features hidden in historical data.However,the AOD remote sensing data are affected by image time resolution and pixel cloud pollution,which will greatly reduce the effective data.Because the construction of a deep learning method depends on a large amount of training data,less training data will seriously affect the model accuracy.Aiming at the problem that the traditional machine learning algorithm cannot deeply mine the hidden association features in data and the deep learning algorithm has a poor effect under the condition of less data,a combined model of PM_(2.5)prediction based on deep learning and random forest is proposed.The model builds a training dataset with AOD remote sensing data,meteorological reanalysis data and PM_(2.5)ground observation data.The deep hidden features in the training data are extracted by the powerful feature extraction ability of the deep learning model first.Then,the extracted hidden features are used in the training of the random forest model,and the predicted value of PM_(2.5)concentration is obtained by the random forest regression algorithm.To verify the effectiveness of this method,a series of experiments were carried out.The results demonstrate that PMCOM has better prediction accuracy in both overall prediction and seasonal prediction scenarios.The combination of random forest and long-and short-term memory neural networks is the best for this experiment.Even when only 35%of the data are used for training,R^(2) in the overall prediction experiment can reach 0.89,and R^(2) in each season prediction experiment is also above 0.75.The combination of deep learning and random forest can reduce the dependence of deep learning models on the amount of data by random forest and make full use of the high-level hidden features of existing historical data.In this way,it makes up for the deficiency of mining the internal associated features of data by a random forest model and improves the prediction accuracy of PM_(2.5)concentration.