【摘要】 针对PM_2.5浓度对于出行和交通规划的影响，对PM_2.5浓度预测进行了研究。主要解决了传统预测模型输入样本不稳定、运算时间长、预测效果不稳定、缺乏灵活性、泛化能力不足、精确度低等难题。引入了深度学习长短期记忆网络（LSTM）框架，设计基于LSTM基本的预测模型做初步预测，主要通过利用已知的空气质量数据训练模型，然后随机选取2个监测站（顺义和怀柔），用训练好的模型预测这2个监测站未来连续60 h的PM_2.5浓度。为了进一步优化预测模型的参数，引入贝叶斯优化（BO）算法，重新训练模型并进行预测。最后通过对LSTM基本预测模型和BO之后的LSTM模型的预测效果进行对比、评估和分析，发现BO算法对于预测模型时序记忆能力的提高有优化作用，预测效果更好，模型泛化能力更强。更多还原

【Abstract】 In view of the influence of PM_2.5 concentration on travel and traffic planning, the prediction of PM_2.5concentration was studied. It mainly solves the problems of unstable input sample, long operation time, unstable prediction effect, lack of flexibility, insufficient generalization ability and low accuracy of the traditional prediction model. In this paper, the deep learning long short-term memory（LSTM） network framework was introduced to design a basic prediction model based on long-and short-term memory network for preliminary prediction. The model was trained mainly by using known air quality data, and then two monitoring stations（Shunyi and Huairou selected in this paper） were randomly selected to predict the PM_2.5 concentration of the two monitoring stations for 60 consecutive hours in the future with the trained model. In order to further optimize the parameters of the prediction model, the Bayesian optimization algorithm is introduced to retrain the model and make the prediction. Finally, by comparing, evaluating and analyzing the prediction effect of the LSTM basic prediction model and the LSTM model after Bayesian optimization, it is found that the Bayesian optimization（BO） algorithm has an optimization effect on the improvement of the sequential memory ability of the prediction model, and the prediction effect is better and the model generalization ability is stronger.更多还原