基于长短期记忆网络偏差校正的空气质量二次预测模型研究

Air Quality Secondary Prediction via LSTM-Based Bias Correction

摘要: 大气污染对人类生产生活和身体健康具有重要影响，研究空气污染物浓度的精准预测对提前预警和污染防控具有重要意义。本文立足于气象条件对污染物浓度的影响，并考虑邻近监测点实测数据对一次预报模型的校准作用，从60个预测变量中筛选出最优特征组合，构建8种典型深度学习模型，重点对长短期记忆网络(LSTM)的结构、超参数设置与性能评估进行系统分析。实验结果表明，LSTM模型在R²、MAE和RMSE三项指标上均优于其他对比模型，能够实现快速、准确的污染物浓度二次预测，并有效识别首要污染物。

Abstract: Air pollution significantly impacts human production, daily life, and public health. Research on accurate prediction of air pollutant concentrations is of great importance for early warning and pollution prevention and control. This study examines the influence of meteorological conditions on pollutant concentrations while considering the calibration effect of measured data from adjacent monitoring stations on primary prediction models. By selecting an optimal combination of features from 60 predictor variables, eight typical deep learning models are constructed, with a systematic focus on the architecture, hyperparameter configuration, and performance evaluation of Long Short-Term Memory (LSTM) networks. Experimental results demonstrate that the LSTM model outperforms other comparative models in terms of R², MAE, and RMSE, enabling rapid and accurate secondary prediction of pollutant concentrations and effectively identifying primary pollutants.