改进白鲸优化卷积网络在银川市空气质量预测中的应用

Application of Improved Beluga Whale Optimization convolutional network for air quality prediction in Yinchuan City

空气质量指数(Air Quality Index,AQI)的精准预测对环境治理具有重要意义。研究针对影响银川市空气质量的PM_(2.5)、PM_(10)等6项污染物指标,提出基于因子分析法-改进白鲸优化算法-卷积神经网络(Factor Analysis-Improved Beluga Whale Optimization-Convolutional Neural Network,FA-IBWO-CNN)的复合AQI预测模型。该模型利用FA法对影响空气质量的6项污染物指标进行相关性分析,并通过计算因子载荷矩阵将新的因子映射到旧的污染物指标上,从而提出一种新的空气污染指标因子表示方式。在此基础上,采用IBWO算法与动态阈值策略和白鲸患病策略,计算训练深度神经网络所需的超参数,改善超参数寻优能力并提高模型收敛速率。研究以CNN作为基线模型,通过IBWO算法优化CNN的全连接层神经数和学习率,实现对银川市AQI预测。利用银川市历史空气质量数据进行试验,结果显示:FA-IBWO-CNN模型与未经优化的CNN模型相比,平均绝对误差(N_(MAE))、均方根误差(N_(RMSE))和平均百分比绝对误差(N_(MAPE))分别提升了56.15%、50.28%和13.943百分点,在预测方面表现出良好的性能。

The Air Quality Index(AQI) is a vital indicator for assessing and monitoring air pollution levels,offering scientific evidence for environmental protection.Accurate prediction of the AQI is essential for government efforts in implementing targeted pollution control measures and regulations,thereby enhancing urban air quality.Therefore,this paper proposes an air quality index prediction model that combines Factor Analysis(FA) with an Improved Beluga Whale Optimization(IBWO) algorithm to optimize a Convolutional Neural Network(CNN).The model utilizes historical air quality monitoring data from Yinchuan City,spanning from October 28,2013,to May 31,2023,to validate its effectiveness.First,Factor Analysis is employed to examine six pollution indicators in the dataset:PM_(2.5),PM_(10),CO,NO_2,SO_2,and O_3.The factor loading matrix is then used to calculate the score of each pollution indicator for each new factor,ultimately resulting in a novel factor representation scheme for air pollution indicators.Subsequently,a CNN-based AQI prediction model is developed for forecasting the AQI in Yinchuan City.Additionally,this paper introduces a dynamic threshold strategy and a Beluga Whale sickness strategy to enhance the original Beluga Whale Optimization(BWO) algorithm.The improved IBWO is employed to optimize two hyperparameters of the CNN model:the number of neurons in the fully connected layer and the learning rate.This optimization aims to identify the optimal hyperparameters that achieve the best prediction performance.Finally,the proposed model is compared with other existing models.The research results indicate that,compared to the ELM model,SVR model,and CNN models optimized by other intelligent optimization algorithms,the proposed model demonstrates superior performance across all metrics,including average absolute error(N_(MAE)),normalized root mean square error(N_(RMSE)),goodness of fit(R~2),and normalized mean absolute percentage error(N_(MAPE)).Therefore,the model proposed in this paper demonstrates strong predictive performance and wide applicability.It offers accurate predictions in practical applications,enabling individuals to make informed decisions and effective plans.