The interplay between model diversity and predictive accuracy in ensemble forecasting represents a pivotal area of current research focus. Addressing the prevalent challenges of overfitting and suboptimal predictive accuracy in ensemble forecasting, this study proposes a two-stage dynamic selection ensemble strategy predicated on diversity regularization. In the first stage, a novel ensemble forecasting diversity regularization strategy is devised by instituting a loss function that judiciously balances diversity against predictive accuracy. The second stage introduces a dynamic selection ensemble methodology capable of proficiently identifying candidate predictive models that adeptly reconcile diversity with predictive accuracy. Experimental outcomes derived from publicly available datasets attest to the proposed two-stage selection ensemble strategy’s efficacy in notably enhancing predictive accuracy while concurrently bolstering model diversity.This dual advancement substantially elevates the ensemble forecasting model’s generalization capacity and reduces predictive errors.