<p>In model selection for combination forecasting, it is essential not only to consider the relevance and redundancy of models but also to emphasize their interactions. However, current methods often overlook this aspect, resulting in the selection of model subsets that fail to fully account for the synergistic effects between models, thereby limiting improvements in predictive accuracy. To address this issue, this paper proposes a novel model selection method based on mutual information theory, which defines relevance, redundancy, and interaction within combination forecasting. The method first selects the model outputs most correlated with the actual values through mutual information maximization and then assesses the potential interactions between the selected outputs and other candidate models to form an interaction set. From this set, the superior subset is selected according to the criteria of maximizing mutual information and minimizing error. This approach efficiently selects superior subsets without using an exhaustive search over all combinations. Empirical analysis using carbon price datasets confirms that the selected subsets outperform their derived subsets, the best individual predictions, and those chosen by the benchmark selection algorithms. The results further demonstrate that the incorporation of model interactions in the selection process effectively enhances forecasting performance.</p>