深度置信网络(DBN)是一种常用的深度学习模型, 在雾霾预测 领域得到了广泛的应用. 然而, 利用传统的DBN进行雾霾预测时, 无监 督学习阶段各节点的连接权值和节点阈值的随机初始化会导致学习速度 慢、容易陷入局部最优等问题. 为此, 文章提出了基于参数优化深度 置信网络的雾霾预测模型. 首先, 构建融合多种变异策略的改进人工蜂群算法(IABC), 从理论上证明了算法的有效性, 并利用6个标准测试函数验证了其收敛速度和寻优精度优于其他人工蜂群算法; 其次, 将IABC算法用于DBN的连接权值和节点阈值的参数寻优, 通过DBN的无监督特征学习和有监督微调, 建立基于参数优化深度置信网络(PODBN)的预测模型; 最后, 利用雾霾数据集和UCI标准数据集验证了该预测模型的有效性, 实验结果表明该模型的预测精度和稳定性优于传统的DBN、FA-DBN及PSO-DBN等模型.

Deep belief network (DBN) is one of the popular deep learning models and has been widely used in haze prediction. However, when traditional DBN is used in haze prediction, the random initialization of connecting weight and threshold of each node in the unsupervised learning stage of DBN may lead to various problems including slow learning speed and local optimal solution. For this purpose, this paper proposes a novel haze prediction method based on parameter-optimized deep belief network. First, an improved artificial bee colony (IABC) algorithm integrating multiple mutation strategies is established. Then the effectiveness of such algorithm is proved from theory and six standard test functions are used to demonstrate that its convergence speed and accuracy are superior to other artificial bee colony algorithms. Second, the IABC algorithm is used to optimize the initial weight and node threshold. Through the unsupervised feature learning and supervised fine tuning of DBN, a novel prediction model based on parameter-optimized DBN (PODBN) is established. Finally, haze datasets and UCI datasets are used to demonstrate the effectiveness of the prediction model. Experimental results show that the prediction accuracy and stability of the model are better than DBN, FA-DBN, PSO-DBN and other DBN models.