Efficient and robust time series prediction model based on REMD-MMLP with temporal-window

Abstract

Time-series prediction plays a critical role in realizing industrial automation. However, due to the non-linear, non-stationary and skew distribution of time series, many existing time series forecasting models have poor generalization performance and forecast accuracy. Although the comprehensive performance of deep learning is better, it takes too much time. To this end, we proposes a hybrid model based on REMD-MMLP (Recursive Empirical Mode Decomposition-Storage Multilayer Perceptron) and Temporal-Window. In REMD-MMLP, we first define a new time window mechanism based on DTW (Dynamic time warping) to search for the best training set of the model. Then, we propose a new REMD algorithm, which decomposes the input data into multiple IMF sub-sequences, and uses these IMF sequence time series as the hidden features of the data. Finally, we design a new storage mechanism in MLP to fully mine the historical information in the time series data to realize the storage and inheritance of the historical information to improve the prediction accuracy of the model. Our comprehensive experiments are conducted using seven different real-world datasets and contrasting models. The final results show that the REMD-MMLP model can be widely used in different practical application scenarios, with strong generality and higher prediction accuracy than state-of-the-art.