Dual-stream transformer-attention fusion network for short-term carbon price prediction

Accurate prediction of carbon price provides important references for relevant companies, investors, and policymakers. However, the nonlinear, non-stationary, and random of carbon price time series pose the current prediction models a challenging task for carbon price prediction. Considering deep learning, especially for Transformer, has got a promising space in time series prediction. Therefore, this study develops a dual-stream Transformer-attention fusion network (DTF-Net), which contains three modules: multi-scale extraction, dual-stream Transformer, and attention fusion module. Firstly, the external variables of atmospheric pollution and the target variable of carbon price are taken as inputs, and the multi-scale extraction module is constructed via multiple one-dimensional convolutions with kernels to mine features on different time scales, enhancing feature engineering. Then, inspired by the idea of “divide and conquer”, the dual-stream Transformer module is applied to independently capture multivariate internal relationships and univariate temporal dependencies, improving feature learning. Finally, the attention fusion module designs the attention mechanism to generate real-time weights and dynamically integrate the above features, highlighting core features. In summary, the proposed DTF-Net network has not only relatively high prediction accuracy but also refined designs and clear multi-layer functions. Five experiments under two carbon price datasets from Hubei and Beijing carbon markets in China show the average improvements of mean absolute percentage error (MAPE) are 63.70 % and 64.55 %, 54.51 % and 53.03 %, and 57.04 % and 52.24 % for recursive, parallel and hybrid methods, respectively. The proposed DTF-Net outperforms eighteen benchmark models and is an addition to predicting carbon prices.