基于贝叶斯优化VMD-LSTM的成都市近地面臭氧浓度预测

Research on Prediction of Near-Surface Ozone Concentration in Chengdu Based on Bayesian Optimization VMD-LSTM

摘要: 近地面臭氧已成为影响成都市夏季空气质量达标率和公众健康的首要污染物。臭氧的生成受气象条件、前体物排放及光化学反应等多重因素的非线性控制，传统单一预测模型难以实现高精度预报。针对臭氧浓度的波动性与非线性特征，本研究提出了一种基于贝叶斯优化(Bayesian Optimization, BO)与变分模态分解(Variational Mode Decomposition, VMD)相融合的长短期记忆网络(LSTM)臭氧预测模型(BO-VMD-LSTM)。利用成都市历史逐日空气质量与气象观测数据，通过VMD技术将复杂的臭氧序列分解为多个相对平稳的子序列；再利用LSTM网络捕捉时间的依赖性；最后创新性地引入严格模式的贝叶斯优化，对模型的四个参数(分解层数$K$ 、惩罚因子$\alpha$ 、隐藏层节点数及学习率)进行全局联合寻优。实验结果表明，在最优参数配置下，模型在测试集上取得了优异的预测性能($R^2=0.9044\pm 0.0021$ , $\text{RMSE}=10.0600\pm 0.1096\mu \text{g}/{\text{m}}^3$ , $\text{MAE}=7.7218\pm 0.0925\mu \text{g}/{\text{m}}^3$ )。该模型为成都市重污染天气应急管控和公众健康防护提供了科学的高精度预警工具。

Abstract: Near-surface ozone has become the primary pollutant affecting the summer air quality compliance rate and public health in Chengdu. The formation of ozone is non-linearly controlled by multiple factors such as meteorological conditions, precursor emissions, and photochemical reactions, making it difficult for traditional single-prediction models to achieve high-precision forecasting. To address the volatility and non-linear characteristics of ozone concentration, this study proposes an ozone prediction model (BO-VMD-LSTM) integrating Long Short-Term Memory (LSTM) network with Bayesian Optimization (BO) and Variational Mode Decomposition (VMD). Utilizing historical daily air quality and meteorological observation data in Chengdu, complex ozone sequences were decomposed into multiple relatively stable sub-sequences via VMD technology; the LSTM network was then employed to capture temporal dependencies; finally, a strict-mode Bayesian optimization was innovatively introduced to perform global joint optimization on four model parameters (decomposition level $K$ , penalty factor $\alpha$ , number of hidden layer nodes, and learning rate). Experimental results show that under the optimal parameter configuration, the model achieves excellent prediction performance on the test set ($R^2=0.9044\pm 0.0021$ , $\text{RMSE}=10.0600\pm 0.1096\mu \text{g}/{\text{m}}^3$ , $\text{MAE}=7.7218\pm 0.0925\mu \text{g}/{\text{m}}^3$ ). This model provides a scientific and high-precision early warning tool for emergency control of heavy air pollution and public health protection in Chengdu.