基于混合CNN-LSTM结构的PM_(2.5)浓度预测深度学习模型

A Hybrid Deep Learning Model for PM_(2.5) Concentration Prediction Based on Hybrid CNN-LSTM Structure

随着工业化和城市化进程的加快,空气污染成为亟待解决的问题。PM_(2.5)是大气中的主要污染物。现有的模型,如循环神经网络(Recurrent Neural Network,RNN)和长短期记忆网络(Long Short Term Memory,LSTM)可以进行即时预测,而进行长期预测则较为困难。为了解决这个问题,笔者提出一个基于混合CNN-LSTM结构的PM_(2.5)预测模型,将CNN结构与LSTM模型结合。本文分别开展了对单元和多元数据集的实验,对于多元数据集,考虑了多种气象因素。实验结果表明,该模型无论在单元还是多元数据集上的表现都优于LSTM模型。基于多元数据集,该模型的NRMSE值为3.66%,说明通过添加CNN结构可以提取更多有效信息并提升预测的精确性。

With the acceleration of industrialization and urbanization,air pollution has become an urgent problem.PM_(2.5)is the main pollutant in the atmosphere harmful to human health.Existing models such as RNN and LSTM can make immediate prediction,while long-term prediction is more difficult.To solve this problem,this paper proposes a PM_(2.5)prediction model based on a hybrid CNN-LSTM structure,which combines the CNN structure with the LSTM model.In this paper,experiments on unit and multivariate data sets are carried out respectively.For multivariate data sets,many meteorological factors are considered.The results show that our model performs better than LSTM model on both units and multivariate data sets.Based on the multivariate dataset,our model achieves a great result with the NRMSE value is 3.66%,indicating that by adding CNN structure,our model can extract more effective information and improve the accuracy of prediction.