基于误差修正的多模型融合空气质量预测

Multi Model Fusion Air Quality Prediction Based on Error Correction

摘要: 由于空气质量指数序列的复杂性和非线性，创造性提出将误差修正与多模型融合相结合的预测方法。首先，利用变分模态分解(VMD)将原始不平稳的空气质量指数序列分解为多个不同时间尺度的平稳固有模态分量；其次，使用改进的SA-BiGRU用于空气质量预测，叠加各个子序列得到空气质量指数初始预测值，实现了长距离时间模式的特征提取；最后，在初始预测模型的基础上建立误差修正模型，通过SVR预测模型得到训练集的预测误差，与初步预测结果用加法器合并，增强模型的表达能力。与单一模型BP、LSTM以及混合模型VMD-LSTM、VMD-GRU、VMD-BiGRU、VMD-SA-BiGRU模型对比，其预测的平均绝对误差分别降低了32.037%、24.581%、18.134%、11.448%、9.320%、5.802%。实验结果表明，VMD-SA-BiGRU-SVR模型在对空气质量指数进行预测时具有更高的精度，预测性能更优异。

Abstract: Due to the complexity and nonlinearity of the air quality index, a creative prediction method com-bining error correction with multi model fusion is proposed. Firstly, variational mode decomposi-tion (VMD) is used to decompose the original unstable air quality index sequence into multiple sta-tionary natural mode components at different time scales. Secondly, the improved SA-BiGRU was used for air quality prediction, and the initial predicted values of the air quality index were ob-tained by overlaying various sub-sequences, achieving feature extraction for long-distance time patterns. Finally, an error correction model is established on the basis of the initial prediction mod-el, and the prediction error of the training set is obtained through the SVR prediction model. The initial prediction results are combined with an adder to enhance the model’s expression ability. Compared with the single model BP, LSTM and the mixed model VMD-LSTM, VMD-GRU, VMD-BiGRU and VMD-SA-BiGRU, the average absolute error of its prediction decreased by 32.037%, 24.581%, 18.134%, 11.448%, 9.320% and 5.802% respectively. The experimental results show that the VMD-SA-BiGRU-SVR model has higher accuracy and better prediction performance in predicting air quality index.