Demand forecasting is a key component of planning in the aeronautical sector, where intermittent and irregular demand patterns pose significant challenges for decision-making processes. Forecast combinations have emerged as an effective strategy to improve predictive performance by leveraging the complementary strengths of diverse models. However, the quality of such combinations depends heavily on the generation schemes for combination weights. Existing learning approaches, particularly those based on meta-learning, have shown strong performance but can require substantial computational resources, limiting their practicality in large-scale or real-time environments. This study investigates LightGBM, a learning algorithm based on decision trees, as an efficient meta-learning tool for generating forecast combination weights in intermittent demand, using real data from the aeronautical sector. The current study proposes a forecasting combination framework in which LightGBM learns to assign weights based on time series features, allowing for flexible adaptation to diverse demand patterns. The method is compared to XGBoost, particle swarm optimisation and traditional combination approaches across multiple forecasting horizons. The experiments include a diverse pool of standard forecasting models with parameter variations grounded in the literature, ensuring a wider range of behaviours and method diversity. Results show that the LightGBM combination achieves forecasting accuracy comparable to XGBoost (slightly better in some cases), while consistently reducing training times by $96.11{\rm{\% }}$ on average compared to state-of-the-art approaches. These findings demonstrate that the proposed LightGBM provides a scalable, high-performance solution for forecast combinations, particularly suited to environments with intermittent demand patterns.