A computationally efficient neural network model for short-term PM2.5 forecasting in urban areas

Liudmila Gorodetskaya, Kirill Garinskikh, Nursultan Kuldeyev, Viktors I. Gopejenko, Alexey Beloussov

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.30837/2522-9818.2026.3.005
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Urban air pollution caused by fine particulate matter with an aerodynamic diameter of less than 2.5 micrometers (PM2.5) remains a major environmental and public health challenge, especially in metropolitan areas with spatially heterogeneous pollutant distribution. Purpose. The study aims to develop and evaluate a computationally efficient short-term forecasting model for predicting PM2.5 concentrations in an urban agglomeration while accounting for both temporal dynamics and spatial relationships between monitoring stations. Objectives. The objectives include preprocessing and standardization of sensor data, construction of local spatio-temporal input subsets for each target station using data from the three nearest neighboring sensors, integration of inter-sensor distances as spatial features, and assessment of predictive performance. Methods. The dataset comprises hourly PM2.5 measurements from 32 stationary sensors collected over an 11-month period. Preprocessing involved anomaly removal, temporal alignment to a uniform hourly grid, z-normalization, and formation of sliding windows of 10-time steps. Spatial context was incorporated using a distance matrix based on geographic coordinates. A multilayer perceptron neural network with fully connected layers, batch normalization, and dropout was trained to predict concentrations one hour ahead. Performance was evaluated on temporally separated training, validation, and test sets using the coefficient of determination, mean absolute error, root mean square error, mean absolute percentage error, and symmetric mean absolute percentage error. Results. The model demonstrated stable convergence and consistent generalization. The coefficient of determination reached 0.80 on validation data and 0.79 on the test set, with 0.81 on the full dataset. The correlation between observed and predicted values was 0.91. Conclusions. The findings indicate that a moderate-complexity multilayer perceptron with explicit spatial feature integration can provide reasonably accurate and computationally efficient short-term PM2.5 forecasts, representing a practical solution for urban air quality monitoring.