A novel forecasting strategy for improving the performance of deep learning models

In this research, a new strategy is introduced for the development of robust, efficient and reliable deep learning time-series models, which is based on a sophisticated algorithmic framework. The novelty and the rationale of the proposed framework is based on efficiently handling multivariate time-series data in order to obtain "high-quality"and "suitable"training data for fitting a deep learning time-series model. The suitability of the time-series data was achieved through the application of a series of transformations for addressing non-stationarity, which is usually presented in many real-world time-series and it is mainly responsible for the performance degradation of a time-series model as well as its forecasting unreliability. The conducted numerical experiments performed on four real-world multivariate datasets, utilizing two state-of-the-art deep learning models, provided convincing experimental evidence about the efficacy and efficiency of our approach.