CTDFormer: controllable trend decomposition with dual-seasonal attention for multivariate time series forecasting

Abstract Time series forecasting refers to constructing models based on historical observations to predict future values or trends. This method is widely applied in fields such as marine environmental monitoring, meteorological prediction, and power load forecasting. To address the issue of unstable trend prediction in multivariate time series and further enhance the capability to capture dependencies across variables, a multivariate time series forecasting model with controllable trend decomposition (CTDFormer) is proposed. This model comprises three modules: 1) the controllable trend decomposition module introduces control parameters along with similarity constraints to decompose the series optimally, yielding trend and seasonal information at an appropriate granularity; 2) the independent trend modeling module eliminates incidental interference between variables, independently predicts the decomposed trend information, and progressively constructs a global trend representation through fine-grained subsequence analysis for each variable; and 3) the dual-seasonal attention module integrates correlation analysis of seasonal information within and between variables, enhancing the model’s ability to capture complex seasonal dependencies. Experimental results demonstrate that CTDFormer achieves significant improvements in forecasting accuracy with only 6.73 million parameters, attaining an optimal balance between computational efficiency and forecasting performance.