DAMixer: A dual-stage attention-based mixer model for multivariate time series forecasting

To improve prediction accuracy, computational efficiency, and interpretability, and to address feature redundancy and high nonlinearity from multiple data sources, we propose DAMixer—a dual-stage attention-based mixer model for multivariate time series forecasting. This model leverages feature attention to extract variable features and temporal attention to capture time patterns. In the first stage, a dynamic variable selection module (VSM) and an optimized temporal convolutional network (TCN) with parallel convolutions enhance multivariate feature extraction and computational efficiency. The second stage integrates an efficient temporal attention mechanism based on long short-term memory (LSTM) to capture time patterns across variables and improve interpretability through a quantile loss function and attention monitoring. Experimental results demonstrate that DAMixer outperforms Transformers and graph neural networks (GNNs), reducing mean squared error (MSE) by up to 40.31% and 8.05%, and mean absolute error (MAE) by 36.45% and 12.63%, respectively. Additionally, DAMixer demonstrates significant advantages across multiple performance metrics, including the coefficient of determination ($R^2$), the number of model parameters, convergence speed, and inference time.