为了对空气污染物浓度进行准确预测, 提出了基于改进萤火虫优化方法(IGSO) 的支持向量机回归(SVR)空气污染物浓度预测模 型. 首先, 利用佳点集理论、拥挤度以及变步长策略对萤火虫优 化算法进行改进; 其次, 根据空气污染物浓度时间序列数据构造 训练集, 运用IGSO算法寻找SVR的最优参数; 最后, 利用基于最 优参数的SVR 实现对空气污染物浓度的预测. 通过两部分的实 验说明文章所提方法的性能. 1) 在8个标准测试函数上进行多次对比实验, 结果显示IGSO算法相比于基于其他改进策略的萤火虫优化方法能够寻找到更优的目标函数值且方差较小, 实验表明改进萤火虫优化算法在稳定性及求解精度方面性能较优. 2) 对京津冀地区空气污染物浓度进行实验, 结果显示如下, 首先, 相比于萤火虫优化算法、粒子群优化算法以及遗传算法, 文章基于IGSO对SVR参数的多次寻优结果波动较小, 并且所得SVR模型的交叉验证误差及其方差较小; 其次, 与基于上述其他优化算法的SVR、 基于网格搜索的SVR以及BP神经网络相比, 文章方法对测试集的预测精度较高. 因此, 基于IGSO的SVR空气污染物浓度预测模型具有较高稳定性及预测精度.

In order to predict the air pollutant concentration accurately, air pollutant concentration forecast model of support vector regression (SVR) based on improved glowworm swarm optimization (IGSO) algorithm is proposed. Firstly, the glowworm swarm optimization algorithm is improved by using good point set, a swarm degree and varied step. Secondly, the training set is constructed according to the time series data of air pollutant concentration, and the optimal parameters of SVR are searched by IGSO algorithm. Finally, the air pollutant concentration is forecast by the SVR with the optimal parameters. Experiments in two parts show the performance of the proposed method. 1) Repeated experiments on 8 benchmark global optimization problems show that IGSO algorithm can find better objective function values with smaller variance than GSO algorithms based on other improved strategies.The experiments demonstrate that the IGSO algorithm has good performance on stability and computational accuracy. 2) The results of experiments on air pollutant concentration in three cities of Jing-Jin-Ji region are as follows. First of all, compared with the glowworm swarm optimization algorithm, particle swarm optimization algorithm and genetic algorithm, the multiple optimization results of SVR parameters based on IGSO are less volatile, and the cross validation error of SVR model and its variance are smaller. In addition, the proposed method has higher prediction accuracy for the test sets than SVR models based on other optimization algorithms mentioned above, SVR model based on grid search and BP neural network. Therefore, the air pollutant concentration forecast model of SVR based on IGSO algorithm (IGSOSVR) has higher stability and prediction accuracy.