| Time series analysis has a wide range of applications in daily life and industrial production.Time series forecasting provides support for decision-making in various fields,which has attracted much attention from researchers.With the rapid development of deep learning,using deep neural networks for time series forecasting has become prevalent in academia and industry.However,existing research focused on deterministic models that lack explicit modeling of probability distributions,which is challenging to capture potential factors and susceptible to uncertainty.Besides,existing probabilistic models usually construct independent Gaussian distributions directly,which ignore the intra-and interdependencies and relations between temporal components.These drawbacks result in low accuracy in forecasting and further make it difficult to achieve high-quality decisionmaking.To break the above limitations,this paper introduces a flexible distribution for probabilistic modeling.Moreover,we establish probabilistic sequential dependencies and analyze seasonal and trend components in time series for better forecasting.There are three main parts:(1)building non-Gaussian distributions through autoregressive latent variables construction and normalizing flow distribution transformation for probabilistic time series modeling;(2)introducing a nonlinear state space model and Copula mechanism to capture evolutional patterns of observations and hidden variables in dynamic systems.Sequential dependencies and influence from external factors are also taken into consideration in probabilistic analysis;(3)designing a framework based on variational inference and mutual information theory to disentangle seasonal and trend components in temporal signals.The proposed methods are evaluated on various datasets,including the power,finance,and meteorology industries.Compared with the most advanced models,these methods have achieved superior performance.Moreover,the proposed forecasting methods have been deployed and operated in hydropower stations for over a year,creating economic and social value. |
