融合空间尺度特征的时空序列预测建模方法

假定原始的时空数据由较大尺度下的全局趋势项与较小尺度下的局部偏差项构成,提出了融合空间尺度特性的时空序列建模方法。首先,将原始数据转换为较大尺度下的数据,此部分反映原始数据的趋势特征。然后,将趋势部分剔除,剩余部分反映原始数据的偏差特征。最后,用灰色系统模型和BP神经网络模型分别对趋势项和偏差项建模,它们的组合预测结果即为原始时空序列预测值。采用该方法对实际的年降水数据以及日平均PM2.5浓度数据进行预测建模分析,实验结果表明:融合空间尺度特性的时空序列预测模型可以用于多空间尺度预测,并且预测精度优于不考虑空间尺度特性的建模方法。

A New Method of Modeling Spatio-temporal Sequence by Considering Spatial Scale Characteristics

The paper presents a new method of modeling spatio-temporal sequence data by considering spatial scale characteristics, where original data is regarded as the combination of a general trend term representing at a larger scale and a local bias term representing at a smaller scale. The original data is first converted into a larger-scale data that represents the general trend of original data. Then, this general trend term is removed and the left part is the bias term of original data. Finally, prediction models of the trend term and the bias term are respectively built, and the combination of the two models is used to predict the original spatio-temporal sequence. Application in annual precipitation and daily average of PM2.5 concentration shows that this proposed method can be used to obtain multi-scale spatial predictions, and is better in accuracy than those models without consideration of spatial scale characteristics.