面向PM2.5预测的递归随机森林与多层神经网络集成模型

Recursive Stochastic Ensemble Model of Multi-layer Neural Network and Random Forest for PM Prediction

不同地区的PM2.5浓度由于其产生来源和地理气象条件不同,通常表现出不同的变化趋势。以往研究中一般主观选择特征集,经常出现考虑不充分导致拟合度低或过拟合现象。本文创新性地提出了递归随机森林特征消除多层神经网络(RRFMLP)的空气质量预测模型,充分考虑各个影响因素,又自行适应特定地区的实际情况,避免人为主观因素的影响,具有较高的拟合度和较低的过拟合度。通过对比实验,验证了该集成模型可以优先考虑最相关因素,并在挑选17个特征时表现出了最好的预测准确度。

The concentration of PM2.5 in different regions usually shows different trends due to different sources and geographical and meteorological conditions.In the existing studies, subjective selection of feature sets generally leads to low fitting or excessive features due to insufficient consideration.This paper innovatively proposes a recursive random forest feature based multi-layer perceptron(RRFMLP) as an air quality prediction model,fully considers each influencing factor,adapts itself to the actual situation in a specific region,and avoids the impact of human subjective factors.This model shows fit and lower degree of overfitting.By comparing experiments,it is verified that the integrated model can give priority to the most relevant factors, and it shows the best prediction accuracy when selecting 17 features.