Combining spatial pyramid pooling and long short-term memory network to predict PM2.5 concentration

Abstract

Deep learning algorithms have been effective in predicting PM2.5. A deep learning algorithm integrating the convolutional neural networks (CNNs) and LSTM networks is proposed in this study to predict PM2.5. Missing data problems commonly exist in the datasets used as input features by the deep learning algorithms. To solve the problem of missing data, traditionally, the dataset is constructed into the input features of the deep learning algorithm after interpolation or elimination processes. The constructed features may have different sizes, rendering the deep learning algorithm invalid. Alternatively, to make the algorithm valid, features with different sizes must be deleted before entering the deep learning algorithm, which leads to information loss. A spatial pyramid pooling (SPP) net is added between the CNN and LSTM to construct a CNN-SPP-LSTM network that can adapt to input features of different sizes. To determine the correspondence between the pollutants, meteorological changes within 24 h, and PM2.5, a newly proposed time-shift Pearson algorithm was proposed to analyze this correlation. The analysis confirmed that using PM2.5 and meteorological characteristics data of the past 24 h to construct the input features of the proposed algorithm to predict the future PM2.5 data is reasonable. Then, two different types of feature vectors with fixed and variable sizes were organized as the input of the proposed CNN-SPP-LSTM to predict PM2.5 data. The LSTM-dense and CNN-LSTM-dense algorithms were utilized as the benchmark method. Comparing the proposed method to traditional LSTM-dense, we observed that the prediction ability of CNN-SPP-LSTM is better than that of traditional LSTM. The proposed method is an effective method for handling missing data problems.