KIE-STCformer: Key Information Enhanced Spatio-Temporal Correction Transformer for Time Series Forecasting

Many existing time series forecasting models primarily rely on past time series points to directly forecast future multi-step sequences. However, two key problems are confronted with the existing approaches. First, within the fixed input space of size N, the initial N sequence data points may include redundant information. Second, the direct multi-step forecasting strategy often overlooks the temporal dependencies between consecutive forecasting steps. To address these issues, we propose an enhanced Transformer-based model named KIE-STCformer, designed to tackle input redundancy and weak temporal-spatial correlations in multi-step forecasting. First, we introduce a novel input data selection strategy to filter out redundant information, ensuring the input samples contain more relevant and key predictive signals. Second, an external attention mechanism that captures inter-sample correlations across different time points is integrated into the model, enabling it to leverage external relationships that significantly impact forecasting outcomes. To mitigate the weak spatial-temporal correlations typically observed in forecasted sequences, a spatial correction network with a mask matrix is designed, and a temporal correction network embedded with an agent attention mechanism is constructed. Extensive experiments are conducted on six widely-used benchmark datasets and the experimental results demonstrate that KIE-STCformer achieves superior accuracy compared to other transformer-based models.