Abstract: Composite solid propellants play a critical role in the aerospace field, with their mechanical properties being jointly influenced by formulation and manufacturing processes. However, the underlying non-linear correlation mechanism between propellant casting parameters and mechanical properties remains ambiguous, and conventional deep learning models suffer from high computational costs and low efficiency in data-constrained industrial scenarios. To address these challenges, this study proposes a prediction method based on the stacking positive approximate kernel-based broad learning system (Stacking PAK-BLS). Rooted in the broad learning system framework, this method innovatively introduces a positive random feature technique to project feature nodes into an approximate kernel space, which not only overcomes the uncertainties associated with random mapping in traditional BLS and enhances non-linear fitting capabilities, but also significantly reduces computational costs. Concurrently, the integration of a Stacking ensemble strategy effectively mitigates batch fluctuations during production, thereby improving the generalization and robustness of the model. Comparative experiments conducted on the industry-recognized Pensim 2.0 benchmark platform demonstrate that the Stacking PAK-BLS model achieves an average coefficient of determination ( R^2 ) of 0.9552 and a root-mean-square error ( E_\textRMS ) as low as 0.074, significantly outperforming a variety of existing mainstream models. Furthermore, based on engineering empirical validation using 400 sets of real-world production data for composite solid propellants, the model yields average relative errors of merely 2.71% and 2.92% for predicting the maximum tensile strength and maximum elongation under room temperature conditions, respectively, thereby fully validating the high-precision advantages and practical value of the proposed method in real engineering applications.