Multihorizons transfer strategy for continuous online prediction of time‐series data in complex systems

Abstract

The sustainability online prediction is of great significance for higher horizon time-series prediction in the future, and it embodies higher application value in equipment fault prediction and health management. However, compared with one-step time-series prediction, continuous online prediction faces many uncertainties, including error accumulation and lack of information. To realize continuous online prediction of time-series data in complex systems, this paper proposes a continuous online prediction strategy based on multihorizons transfer (OnMultiHorTS), which is used for continuous online prediction tasks of timeseries data. The algorithm aims to use source domain data to provide more effective information for target prediction tasks. However, the time-varying characteristics of time-series data often lead to large differences in data distribution over a long time span, which is difficult to guarantee the assumption that the data are the same distribution. How to construct more effective source domain information based on historical data and existing data, and apply it to the target domain prediction tasks, is one of the focuses of our OnMultiHorTS algorithm. In addition, different from the typical iterative and multistep advance prediction methods, the proposed algorithm regards different prediction tasks as different horizons, which are