基于改进多孔算法的时间序列预测

针对小波分析技术存在的边界问题,提出一种改进的多孔算法。使用该算法得到的系数序列,在具备时移不变性的同时,消除了右侧边界存在数据畸变的现象,使小波分析技术结合神经网络等传统预测模型的方法应用于单变量时间序列预测任务具备可行性。为进一步提高预测精度,引入了神经网络集成技术以改善网络泛化能力。实验表明,这种组合预测模型预测效果与稳定性优于传统预测模型。

Time Series Forecasting Based on Improved (A) Trous Algorithm

Aiming at the boundary problem of wavelets transforms, an improved A trous algorithm was proposed. The coefficient sequences decomposed by this novel method possessed time-invariant capability and eliminated the data distortion phenomenon around right boundary, which made it feasible that the traditional forecasting models such as Neural Networks combining with wavelet transforms could apply to the uni-variant time series forecasting task. In order to improve the network generalization ability, Neural Network Ensembles was introduced into this hybrid model. Experiments result stipulate that the forecasting performance and stability of the hybrid forecasting model is superior to the traditional forecasting model.