适应时间序列预测不确定性的在线组合概率模型及应用研究

In recent years, most of researches about combination probabilistic forecasts for time series forecasting are static-offline, which are not able to exploit the non-stationary in time series for the multistep prediction; moreover, there is a still lack of studies on how to select sub-models and determine the weight of the combined forecasting model, all of these make the current models are not sustainable for forecasting complex time series with noise. To fill the gaps of existing studies, this project aims to establish an online combined probabilistic forecasting model using information theory and artificial intelligence algorithms, then the established model is applied in wind power grid management. The research contents can be divided into four parts: (1) Detect and analyze the noise in time series; (2) The sub-model selection algorithm of combined point forecasting model is generalized. Combined with neighborhood mutual information theory, the sub-model selection algorithm of combined probabilistic forecasting model based on the minimum redundancy and maximum relevance is proposed; (3) Propose an artificial intelligence double-online weight searching method without non-negative limit to confirm the weight; (4) Apply the online combined probabilistic forecasting model based on the upper theories in the wind power time series forecasting in the wind power plants.

近年来，针对时间序列预测的组合概率预测大多为静态离线模型，在预测步长较长时无法有效地挖掘时间序列的非平稳特征。并且针对组合概率预测子模型的选择以及权重确定方式的研究仍有欠缺，无法适应含噪声的复杂时间序列的预测。针对上述问题，本项目以信息理论和人工智能算法为研究手段，构建在线组合概率预测模型，并将其应用到风电并网管理中。研究内容分为四部分：(1)对时间序列的噪声进行探测和分析；(2)将组合点预测模型的子模型选择算法进行推广，结合邻域互信息理论，提出基于最小冗余度和最大相关度的组合概率预测模型的子模型选择算法；(3)提出基于无非负限制的人工智能搜索双在线权重确定方式；(4)将基于上述研究理论的在线组合概率预测模型应用于风电场中风能时间序列的预测中。

现有的预测模型大多为静态离线模型，在预测步长较长时无法有效地挖掘时间序列的非平稳特征。本项目针对非平稳时间序列进行如下几个方面的研究：第一、结合深度学习模型提取非平稳时间序列的特征，对不同频段的特征进行训练从而得到更具鲁棒性的预测模型；第二、考虑点预测误差的异方差性问题，对不同点位的预测误差进行划分，从而提取误差中的残留信息，采用Bootstrap方法估计动态误差信息的置信区间从而得到时间序列中的不确定信息，从而量化时间序列中的不确定性信息；第三、针对非平稳的时间序列构建了一种基于改进后的引导注意力机制模型，结合多时间序列的时空特征，避免差分造成的信息遗失，设计一种在线动态组合模式，实现对时间序列的复杂特征提取；第四、构建基于多目标和贝叶斯优化的深度学习区间预测模型，依据分位数回归理论计算出电力负荷在不同分位点处的预测区间，再通过有效性检验，筛选出合理的预测模型。本项目的完成既有理论依据又有实践依据，不仅可以完善非平稳时间序列预测的理论基础，同时能够为不同场景下的时间序列预测问题提供参考依据。