Multi-step PM_2.5 Concentrations Forecasting Based On Multi-source Data Fusion And City Collaborative Strategy

The vigorous development of global economic activity and urbanization and industrialization has led to the gradual deterioration of air quality.The deterioration of air quality has caused climate change,loss of biodiversity,and even respiratory diseases,with particulate matter with a diameter of less than 2.5μm(PM_2.5)being the main pollutant.In order to protect the health and safety of residents and provide reliable air quality warning information,it is necessary to propose a high-precision PM_2.5 concentrations（PCs）multi-step ahead forecasting system.Due to factors such as industrial,transportation,regulations and human interference,PCs have nonlinearity,non-stationarity,and high complexity,making it a challenging research topic to forecast PCs.After analyzing existing PCs forecasting studies,this article believes that considering both the fusion of multiple data sources and the collaborative impact mechanism of surrounding cities or stations can improve the accuracy of PCs forecasting.Based on this,two PCs mixed forecasting models are designed,specifically:（1）A multi-scale integrated learning multi-step ahead point forecasting model for PCs considering the city collaboration forecasting strategy.Considering that previous studies often focused on the data of the target city itself and ignored the collaborative effects among cities in the same region,this model proposes a multi-scale integrated learning method to forecast the daily PCs of the target city by modeling the air and climate indicators of the target city and the PCs data of nearby cities.First,the model smoothes multiple data sources through singular spectrum analysis and selects features based on data forecasting ability,spatial distance,and other factors.Next,multivariate empirical modal decomposition is used to explore the potential correlation between multiple features.Then,the average Hurst index matches each time scale with the corresponding predictor for multi-step ahead forecasting.Finally,the forecasting values of all time scales are added to obtain the PCs forecasting result of the target city.（2）A multi-step ahead point-interval forecasting model for PCs based on multi-dimensional decomposition and kernel density estimation.Existing research mainly focuses on point forecasting and ignores the importance of uncertainty research.Therefore,this study proposes a new hourly PCs point-interval forecasting model that quantifies the uncertainty of the forecast while improving the accuracy of the target monitoring site’s forecasting.First,the model considers the impact of surrounding stations on the target station’s PCs changes,adjusts the lag of the PCs of the surrounding stations using time-lag correlation,and extracts features from the selected multiple data sources using the proposed mutual information and decision tree regression combination method.Next,successive multivariate variational mode decomposition is used to explore multi-modal components（MMC）with the same frequency modulation and central frequency in the chosen features,and bidirectional gated recurrent unit are used to match each element in MMC with time factors for multi-step ahead point forecasting.Finally,non-parametric kernel density estimation is used to extend the upper and lower bounds of the prediction interval.The prediction interval is fine-tuned using a velocity-constrained multi-objective particle swarm optimization algorithm to obtain multi-step ahead interval forecasting results with confidence levels of 80%,90%,and 99%.This article uses different datasets to conduct empirical analysis for the two proposed models to verify their effectiveness and robustness.Through the analysis of the evaluation indicators,statistical tests,and further discussions of the two models and their benchmark models,the conclusion shows that the proposed models perform better in PCs forecasting performance after fully considering the fusion of multiple data sources and the city collaboration strategy.It can provide reliable decision-making basis for air warning departments and meet their needs in different application scenarios.