基于注意力机制的seq2seq模型在PM_(2.5)浓度预测中的研究

Prediction of PM_(2.5)Concentration Based on seq2seq Model with Attention Mechanism

目前PM_(2.5)浓度预测研究主要是对未来1 h的污染物浓度进行预测,不能满足污染物浓度较长时间细粒度预测的应用需求。构建了基于注意力机制的序列到序列(sequence to sequence,seq2seq)模型。模型主要由编码器、解码器和注意力模块3部分构成,其中,编码器用于提取时间特征,解码器使用注意力模块动态计算每个时刻的背景变量,从而预测未来时刻的PM_(2.5)浓度。使用2015—2018年北京市12个空气监测站点的小时级别的PM_(2.5)观测数据进行实验,并将结果与基准模型进行比较。结果表明,该模型预测结果较好。

The current research on the prediction of PM_(2.5)concentration mainly focus on predicting the concentration of pollutants in the future 1 h,which cannot meet the application requirements of fine-grained prediction of pollutant concentration for a long time.Based on the attention mechanism,this article builds a Sequence to Sequence(seq2seq)model,which is mainly composed of a three-part encoder,decoder,and attention module where the encoder is used to extract temporal features,and the decoder uses the attention module to dynamically calculate the background variable at each moment to predict PM_(2.5)concentration data in unknown time.The encoder and decoder in the model use a single-layer LSTM(long short-term memory)model structure to achieve long-term prediction goals.Finally,the hourly PM_(2.5)observation data of 12 air monitoring stations in Beijing from 2015 to 2018 is used in the experiment to compare with the benchmark model.The results show that the model used in this paper can achieve better prediction results.