A Multi-Head Attention-Enhanced Fusion Model for Cross-Domain Short-Term Time Series Forecasting

Abstract

With the rapid advancement of artificial intelligence technologies in the era of big data, time series forecasting has become indispensable in critical fields such as environmental monitoring and financial market analysis. However, the existing forecasting models often encounter performance limitations when extracting high-dimensional features and generally cannot dynamically focus on critical information within long-term sequences. To address these challenges, this study proposes a multi-head attention fusion model (MAFM) designed to enhance the predictive accuracy and modelling capability for high-dimensional and nonlinear data across diverse application scenarios. Experiments were conducted on two heterogeneous datasets from the environmental and financial domains. After the key hyperparameters of the MAFM were optimized through an orthogonal experimental design, the model achieved coefficients of determination exceeding 0.90 on both datasets. Furthermore, the results of four comparative experiments demonstrate that the MAFM consistently outperforms traditional machine learning models, including support vector regression and extreme gradient boosting, as well as state-of-the-art deep learning models such as long short-term memory, temporal convolutional networks, and transformers. Compared with the best-performing baseline model on each sub-dataset, the MAFM reduced the mean squared error by 44.4%, 8.3%, 29.4%, and 65.5%, respectively, highlighting its superior predictive performance and strong generalization capability. In summary, the proposed MAFM provides an efficient, robust, and interpretable solution for time series forecasting tasks across multiple domains. Its outstanding performance demonstrates significant potential for practical applications in environmental monitoring, financial forecasting, and other real-world scenarios.