基于特征提取和LSTM的PM _2.5浓度预测模型

PM _2.5 Concentration Prediction Model Based on Feature Extraction and LSTM

摘要: PM_2.5是空气污染中对人体危害最大的一类污染物之一，对PM_2.5的准确预测可以为人们社会活动的决策制定提供可靠依据。使用自编码器(Auto-encoder)能够达到较高的预测精度，但是需要训练的参数量大，所需的计算资源也会增加，于是本文提出了一种待训练参数比Auto-encoder减少了24%的PM_2.5预测模型FE-LSTM (Feature Extraction-LSTM)。FE-LSTM模型基于SE注意力机制、卷积神经网络(CNN)和长短期记忆网络(LSTM)等模块，先使用CNN提取出输入张量的初始特征，再通过SE注意力对特征张量按通道加权，通过全连接和重构后使用LSTM得出污染浓度的预测值。在北京PM_2.5数据集上，对FE-LSTM模型及对比模型进行了训练和测试，结果表明，本文所提出的FE-LSTM模型的预测精度优于其他对比模型。

Abstract: PM_2.5 is one of the most harmful pollutants in air pollution, and accurate prediction of PM_2.5 can pro-vide a reliable basis for decision making in people’s social activities. This paper proposes a PM_2.5 prediction model, FE-LSTM (Feature Extraction-LSTM), with 24% fewer parameters to be trained than Auto-encoder, which can achieve high prediction accuracy but requires a large number of training parameters and increased computational resources. The FE-LSTM model is based on the SE attention mechanism, convolutional neural network and long short-term memory network mod-ules. The initial features of the input tensor are first extracted using CNN, then the feature tensor is weighted by channel through SE attention, and the predicted pollution concentration is derived us-ing LSTM after full concatenation and reconstruction. The FE-LSTM model and the comparison mod-els were trained and tested on the Beijing PM_2.5 dataset, and the results showed that the prediction accuracy of the FE-LSTM model proposed in this paper was better than that of the other comparison models.