基于深度信念网络的PM2.5预测

提出一种基于深度信念网络(deep belief networks, DBNs)的区域PM_2.5日均值预测方法,讨论了训练数据选择方式,并优化了DBNs参数设置。通过相关实验并与基于径向基神经网络(radial basis function, RBF)和反向传播神经网络(back propagation, BP)方法比较,验证了基于DBNs方法的可行性和预测精度。实验结果表明:基于DBNs的方法,区域(西安市)预测PM_2.5日均值与观测日均值之间均方差(mean square error, MSE)为8.47×10^-4mg²/m⁶;而采用相同数据集,基于RBF和BP的方法均方差为1.30×10^-3mg²/m⁶和1.96×10^-3mg²/m⁶。比较分析表明:基于DBNs的方法能较好预测区域整体PM_2.5的日均值变化趋势,显著优于基于神经网络和径向基网络方法的预测结果。

A prediction method of atmospheric PM2.5 based on DBNs

A DBNs-based (deep belief networks) method for forecasting the daily average concentrations of PM_2.5 in Xian was proposed. Besides, the way to select training data set as well as the DBNs parameters was optimized. Then relative experiments and comparison with methods based on BP (back propagation) and RBF (radial basis function) artificial neural network confirmed the feasibility and precision of DBNs. The results showed that the MSE (mean square error) between DBNs simulated PM_2.5 daily average concentrations and observed ones was 8.47×10^-4 mg²/m⁶, while the MSE of RBF and BP was 1.30×10^-3 mg²/m⁶ and 1.96×10^-3 mg²/m⁶ respectively. Therefore the DBNs-based method was fit for prediction of PM_2.5 concentrations and it predicted more accurately than those methods based on RBF and BP artificial neural network.