马树森 题为 “THCVformer: Modeling Temporal-dependencies, Heterogeneity, and Correlations for Enhanced Time-series Forecasting” 的论文已被《Knowledge-Based Systems》接受发表。该论文摘要如下：

Multivariate time-series forecasting (MTSF) poses a significant challenge due to the intricate and nonlinear interdependencies among diverse timestamps and variables. Extracting temporal features from the variable sequence is crucial to improving model performance in such tasks. Learning the heterogeneity of variables and analysing their correlations are equally important. To construct a more robust predictive model, this paper incorporates the temporal dependencies, heterogeneity, and correlation of variables into the predictive framework, proposing a new model named THCVformer. THCVformer first deploys a Heterogeneity Learner to capture the statistical features of the input sequence for different variables, i.e. variance and mean. Subsequently, through an embedding layer, the THCVformer effectively captures the positional information of input sequences and dependencies between different time steps, thereby learning the temporal characteristics of variables. Finally, the THCVformer employs Multivariate Sparse-attention to clip weak attention, reducing the interference of irrelevant variables on target variable prediction and enhancing the model’s perception and learning of useful information among variables. Experimental results show that compared with the previous state-of-the-art models, THCVformer can reduce mean squared errors by 9.2% and 40.5% for multistep and single-step forecasting tasks, respectively.