In this report, our objective is to develop a center-of-gravity-based method in solving the problems of the various sources of uncertainties for data-driven prediction models. The uncertainty sources include input uncertainty, model uncertainty, parameters uncertainty, and so forth. Involving prediction uncertainty in data-driven prediction models can enhance the reliability and credibility of the model outputs. In general, uncertainty is estimated by prediction interval, which can provide more information about forecasting results. However, for non-linear data-driven prediction models, such as neural networks, fuzzy systems, and so forth, conventional methods for building prediction interval suffer from the assumption of data distribution (such as the assumption of constant variance), large computational complexity or difficult computation, and the characteristics of the model errors that result in generating inappropriate prediction interval. Besides, because of increased uncertainty from longer forecasting steps, constructing effectively the prediction interval of multistep-ahead forecasting is an important issue. Therefore, in this report, we develop a center-of-gravity-based neural network method for building the prediction interval of non-linear data-driven prediction models in order to avoid the restrictive condition of data distribution, and the problem of difficult computation. Moreover, analyzing the center of gravity for all model output errors and integrating the training process of neural networks generate more appropriate prediction interval.
|Effective start/end date||2018/08/01 → 2019/07/31|
- prediction models
- prediction interval
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.