Transformer-based models excel in capturing long-term dependencies, making them promising for Long Sequence Time-Series Forecasting (LSTF). However, despite the use of positional encoding, these models struggle to effectively integrate temporal order information. While time series decomposition techniques can enhance model performance, variations in decomposed components may hinder fitting and training speed. Additionally, the volatility of time series data may reveal similarities across different numerical levels, indicating potential underlying patterns. To cope with these issues, we propose the Global Decomposition Long Sequence Forecasting Model (GDLFM), which features: (1) global decomposition of time series and multi-error backpropagation to minimize interference among components and improve prediction accuracy; (2) dual-dimensional encoding (DDE) that extracts features from both temporal and numerical dimensions using multiple multilayer perceptrons; and (3) a baseline strategy that focuses on value changes relative to the baseline for sequence prediction. Experimental results show that GDLFM enhances long sequence forecasting accuracy compared to baseline methods, achieving at least a 21 \(\%\) reduction in mean square error and a 15 \(\%\) reduction in mean absolute error across six datasets when compared to six baseline models.