A Daily Air Pollutant Concentration Prediction Framework Combining Successive Variational Mode Decomposition and Bidirectional Long Short-Term Memory Network

Abstract

Precise and efficient air quality prediction plays a vital role in safeguarding public health and informing policy-making. Fine particulate matter, specifically PM2.5 and PM10, serves as a crucial indicator for assessing and managing air pollution levels. In this paper, a daily pollution concentration prediction model combining successive variational mode decomposition (SVMD) and a bidirectional long short-term memory (BiLSTM) neural network is proposed. Firstly, SVMD is used as an unsupervised feature-learning method to divide data into intrinsic mode functions (IMFs) and to extract frequency features and improve short-term trend prediction. Secondly, the BiLSTM network is introduced for supervised learning to capture small changes in the air pollutant sequence and perform prediction of the decomposed sequence. Furthermore, the Bayesian optimization (BO) algorithm is employed to identify the optimal key parameters of the BiLSTM model. Lastly, the predicted values are reconstructed to generate the final prediction results for the daily PM2.5 and PM10 datasets. The prediction performance of the proposed model is validated using the daily PM2.5 and PM10 datasets collected from the China Environmental Monitoring Center in Tianshui, Gansu, and Wuhan, Hubei. The results show that SVMD can smooth the original series more effectively than other decomposition methods, and that the BO-BiLSTM method is better than other LSTM-based models, thereby proving that the proposed model has excellent feasibility and accuracy.