基于小波变换和LSSVM-DE的天然气日负荷组合预测模型

为了提高天然气短期负荷的预测精度,提出了基于小波变换和LSSVM-DE(Least Squares Support Vector MachineDifferential Evolution)的天然气日负荷组合预测模型,首先,采用Mallat快速算法对天然气日负荷实际采集数据样本时间序列进行小波分解;其次,对分解出来的高频分量和低频分量分别建立LSSVM预测模型,各分量的模型参数分别采用DE进行优化,以期得到更准确的预测结果;最后,分别对各分量的预测结果进行小波重构。以某市实际采集的样本数据为例,并将重构结果与单独应用LSSVM预测模型及ANN(Artificial Neural Networks)预测模型进行对比分析。结果表明:小波变换和LSSVM-DE预测模型的预测精度分别比单独应用LS-SVM和ANN预测模型高出1.662%、1.14%、3.96%、2.99%、15.53%和1.942%、1.01%、3.07%、1.86%、12.26%。该结论预示着将小波变换和LSSVM-DE理论相结合对天然气日负荷时间序列进行预测是一种行之有效的方法。

A forecasting model of natural gas daily load based on wavelet transform and LSSVM-DE

〗In order to improve the accuracy of predicting the short term natural gas load, a forecasting model of natural gas daily load was built based on wavelet transform and LSSVM – DE （Least Squares Support Vector Machine – Differential Evolution）. First, the Mallat algorithm was applied to conduct the wavelet decomposition of the time series in a sample of the actual gas daily load data. Then, the LSSVM forecasting model was established for the decomposed high and low frequency components respectively, the parameters of which were optimized by using the DE to achieve more accurate forecasting results. Finally, we re constructed the wavelet of each component′s forecasting result. In a case study from a certain city, a comparative analysis was made of the forecasting results between the combined use of wavelet transform and LSSVM – DE and the independent use of LSSVM or ANN （artificial neural networks）. The forecasting model based on wavelet transform and LSSVM – DE was validated with a high prediction accuracy and the resulted relative mean square error, normalization mean square error, normalization absolute square error, normalization root mean square error, maximum absolute error resulted from the combined use of wavelet transform and LSSVM – DE were lower than those from the independent use of LSSVM or ANN by 1.662%, 1.14%, 3.96%, 2.99%, 15.53%, 1.942%, 1.01%, 3.07%, 1.86% and 12.26% respectively. In conclusion, this study provides a practical and feasible method for predicting the time series of natural gas daily load.