【摘要】 为了充分挖掘多因素数据间的时空特征信息，解决在多种因素相互影响下不能准确预测PM2.5值的问题，提出了一种融合了局部加权回归的周期趋势分解(STL,seasonal-trend decomposition procedure based on loess)算法、卷积长短期记忆网络(ConvLSTM,convolutional long short-term memory network)和门控循环单元(GRU,gated recurrent unit)的PM2.5预测方法；首先利用STL算法将PM2.5数据进行分解，将分解得到的序列分别与其他因素相融合；搭建ConvLSTM-GRU模型，并利用贝叶斯寻优算法进行超参数寻优；将融合数据传入ConvLSTM网络中进行时空特征提取，再将提取后的特征序列传入GRU网络中进行预测；通过与ConvLSTM-GRU模型、CNN-GRU模型以及GRU模型的预测结果进行比较实验，证明所提模型具有误差小、预测效果好等特点。更多还原

【Abstract】 In order to fully mine the spatiotemporal feature information between multi-factor data, and solve the problem that PM2.5 value cannot be accurately predicted under the influence of multiple factors, a PM2.5 prediction method is proposed to combine with a seasonal-trend decomposition procedure based on Loess(STL) algorithm, convolutional long short-term memory network(ConvLSTM) and gated recurrent unit(GRU). Firstly, the STL algorithm is used to decompose the PM2.5 data and fuse the decomposed sequence with other factors; The ConvLSTM-GRU model is built, and the Bayesian optimization algorithm is used to search for super parameters; The fused data is transferred to extract the ConvLSTM network for time-space feature, and then the extracted feature sequence is transferred to the GRU network for the prediction. Compared with the prediction results of the ConvLSTM-GRU model, CNN-GRU model and GRU model, the results show that the proposed model has the characteristics of small error and good prediction effect.更多还原