Time series forecasting using singular spectrum analysis, fuzzy systems and neural networks.

Area of Science:

• Data Science
• Computational Statistics
• Time Series Analysis

Background:

• Hybrid methodologies are increasingly adopted in research to leverage the strengths of diverse methods.
• Combining methods can overcome the limitations inherent in individual techniques.
• Time series forecasting remains a critical area requiring robust and accurate predictive models.

Purpose of the Study:

• To present novel hybrid methodologies for time series forecasting.
• To introduce two specific hybrid models: SSA-LRF combined with NN, and SSA-LRF combined with WFTS.
• To demonstrate the efficacy of these hybrid models in handling complex time series data.

Main Methods:

• Development of a hybrid model integrating Singular Spectrum Analysis with Linear Recurrent Formula (SSA-LRF) and Neural Networks (NN).
• Development of a second hybrid model combining SSA-LRF with Weighted Fuzzy Time Series (WFTS).
• Application and evaluation of both hybrid models on load data series and other time series datasets.

Main Results:

• The proposed hybrid methods are effective for load data series and other time series.
• Both hybrid methods successfully address deterministic patterns and nonlinear stochastic behavior within the data.
• Significant improvements in forecasting accuracy were observed compared to individual methods and existing hybrid approaches.

Conclusions:

• The presented hybrid methodologies offer enhanced performance for time series forecasting.
• The integration of SSA-LRF with NN or WFTS provides a powerful framework for capturing complex data dynamics.
• These hybrid models represent a valuable advancement in the field of time series prediction.