基于小波变换和优化预测器的瓦斯浓度预测

Gas Concentration Forecasting Approach Based on Wavelet Transform and Optimized Predictor

为了准确预测煤矿瓦斯浓度,提出一种基于小波变换和优化预测器的预测方法.用瓦斯浓度序列分解到小波函数空间(或尺度函数空间)上的能量作为尺度能量,依据尺度能量与满足预测精度的最大误差能量的比值关系,确定小波最佳分解级数.通过小波分解降低瓦斯浓度序列的复杂度,引入延时映射,将小波分解后各个分量转化为具有历史特征的新样本分别进行预测,所得到结果进行叠加为最终预测结果.提出基于预测残差方差比检验的最佳延时单元数确定方法,以预测残差的F检验值作为适应值,采用粒子群算法优化预测器的参数.结果表明:单一的BP人工神经网络(BP_ANN)或支持向量机(SVM)所建立的预测方法对某些瓦斯浓度突变数据的预测有过拟合现象,其预测结果的平均误差大于23%,小波变换后的组合预测方法对于瓦斯浓度突变数据具有较好的跟踪能力和反应速度,采用预测模型的最佳参数后,预测器性能显著提高,平均误差小于10%,表明所建议的方法是可行和有效的.

In order to accurately predict the gas concentration,an approach based on wavelet transform and optimized predictor was proposed.The energy of gas concentration series decomposed into wavelet or scaling functional space was used as the scaling energy,based on which,the best wavelet decomposition level was determined according to the ratio of the scaling energy and the largest forecasting biases energy under the allowed precision.Wavelet decomposition was used to reduce the complexity of gas concentration series.A delay mapping was introduced to transform the wavelet components into new samples with historical characteristics,which were used in the prediction,respectively.The final forecasting result was obtained by combining all the predicted results.The method based on variance ratio test of the prediction variances was suggested to determine the best number of the delay units.Taking F test of the prediction residual as the fitness value,the parameters of the forecaster were optimized by a PSO based program.The results showed that simply implanted ANN or SVM based forecasters were with over-fitting problems in case of gas concentration suddenly changed and the averaged prediction biases were larger than 23%.The proposed approach based on wavelet transform and SVM had improved tracking ability and dynamic behavior.The performance of the forecaster was remarkably improved to obtain the averaged biases within 10% using the optimized parameters,which indicated that the suggested approach was feasible and effective.