Accurate spatial prediction of forage nutritional quality is critical for optimizing livestock production and supporting regional food security planning. This study developed an interpretable machine learning framework based on Random Forest (RF) and Support Vector Machine (SVM) algorithms to generate high-resolution maps of key forage nutritional parameters-crude protein (CP), ether extract (EE), acid detergent fiber (ADF), and neutral detergent fiber (NDF)-across natural grasslands in the Ili River Valley, Xinjiang, China. Our approach integrated 40 environmental covariates and explicitly accounted for both immediate and legacy climate effects through time-lag and accumulation analyses. Predictor selection was optimized using forward feature selection combined with k-fold nearest-neighbor distance matching cross-validation to reduce overfitting. Results indicated that climate variables significantly influenced forage quality via time-lag and accumulation effects, with distinct drivers identified for nutrient content versus pools: CP and EE content were primarily influenced by climate, whereas ADF and NDF content as well as all nutrient pools were predominantly driven by vegetation traits. Model performance varied considerably, with nutrient pools (R2 = 0.63-0.73) being predicted more accurately than nutrient content (R2 = 0.42-0.58). Spatially, high-quality forage was concentrated in mountain meadows and temperate meadow steppes, while low-quality forage was more prevalent in temperate desert steppes. Uncertainty analysis revealed higher prediction errors in desert steppes, likely due to limited sampling density and extreme environmental conditions. These findings offer a scientific basis for targeted grassland management and underscore the importance of appropriate model selection and covariate inclusion for reliable spatial predictions of forage quality.