Prediction of Shenzhen PM2.5 Hourly Concentration Based on LSTM

Abstract

PM2.5 pollution remains a key public health issue in coastal megacities such as Shenzhen, where human activity patterns and meteorological conditions drive rapid short-term changes in air quality. Holiday periods can alter travel, industrial, and household emission patterns, creating unique challenges for accurate air quality forecasting. This paper investigates how holiday periods affect short-term PM2.5 fluctuations and their predictability, utilizing hourly monitoring data from Shenzhen spanning October 2025 to January 2026. A Long Short-Term Memory (LSTM) network is constructed for prediction, complemented by independent samples t-tests for statistical validation. The findings reveal an average PM2.5 concentration of 21.45 μg/m 3 during the period, with marginally elevated levels observed on holidays (p=0.573). The LSTM model yields an overall Mean Absolute Error (MAE) of 7.00 μg/m 3 , registering higher errors on holidays (10.40 μg/m 3 ) than on regular days (6.86 μg/m 3 ). Comparative experiments confirm its superiority over ARIMA and linear regression models, demonstrating LSTM's efficacy in capturing holiday-induced volatility and its value for air quality governance.