A TCN-Transformer Hybrid Deep Learning Model for Multivariate Time Series Forecasting with Multi-source Feature Fusion

Abstract Accurate multivariate time series forecasting is a fundamental task for intelligent operation and decision-making across industrial scenarios. Existing hybrid time series models fail to simultaneously capture fine-grained local fluctuations and long-range periodic dependencies, and often neglect the coupled effects of heterogeneous external features, resulting in limited accuracy under frequent external shocks. To address these issues, this paper proposes a dual-branch parallel TCN-Transformer hybrid deep learning model for daily multivariate time series forecasting. First, a multi-source feature dataset is constructed by integrating historical values, calendar attributes, policy variables and meteorological features. A dual-branch encoding architecture is designed: the TCN branch extracts local short-term fluctuations via dilated causal convolutions, while the Transformer branch captures long-term periodic dependencies through multi-head self-attention. Adaptive average pooling and feature concatenation achieve deep fusion of local and global representations. Taking daily inbound passenger flow of the Hong Kong-Macau Ferry Terminal as an empirical case, experiments on a three-year dataset show that the proposed model achieves an R² of 0.953, an RMSE of 357.93 and an MAE of 274.27 on the test set, outperforming single TCN, Transformer, LSTM and CNN baselines. The model provides an effective solution for multivariate time series forecasting with multi-source heterogeneous features.