Multivariate Time Series Forecasting through Automated Feature Extraction and Transformer-Based Modeling

Abstract

This paper addresses the challenges of feature redundancy and complex high-dimensional dependencies in multivariate time series forecasting. A forecasting method is proposed by combining TSFresh-based feature engineering with the Temporal Fusion Transformer. The method first applies TSFresh to perform automated feature extraction and selection on raw time series data. This process reduces input dimensionality and enhances feature representation. Then, the Temporal Fusion Transformer is used to model temporal dependencies and inter-variable relationships. It integrates dynamic variable selection, gated residual networks, and multi-head attention to achieve accurate future sequence prediction. Experimental results on the Electricity multivariate load dataset show that the proposed model outperforms existing mainstream methods in terms of MAE, MSE, and R². It also shows stable performance in hyperparameter sensitivity analysis and robustness testing. These results confirm the effectiveness and reliability of the method in complex multivariate time series forecasting scenarios.