Abstract:
The multi-step prediction of time series uses historical time series data to forecast their future development trends at multiple time points, in order to formulate corresponding production control strategies in advance. In this paper, a new multi-step prediction algorithm based on time series decomposition is proposed. To address the significant impact of time series decomposition parameters on prediction results, an adaptive variational mode decomposition strategy optimized by genetic algorithm is proposed. For sub-series with different modalities, a random forest algorithm is proposed to establish various base learners to fully mine the information of each sub-series. A real-time decomposition framework is proposed to address the problem that the one-time decomposition strategy is only applicable for simulation research and cannot be applied in practice. The whole decomposition and prediction process is embedded in this framework. Finally, experiments on several public time series datasets have shown that the proposed prediction algorithm has higher accuracy compared to the comparison algorithms. Although the final prediction result seems to has lower accuracy than the ones in some literatures, it is more feasible and practical.