Abstract: Accurate multi-component concentration measurement of coke oven gas plays a critical role in optimizing coking processes, improving energy utilization efficiency, reducing pollutant emissions, and ensuring production safety. To enhance measurement accuracy in multi-component coke oven gas analysis using Raman spectroscopy, this study proposed a successive projections algorithm-support vector machine (SPA-SVM) model based on feature extraction and small-sample modeling principles. First, the prediction accuracy of full-spectrum machine learning models was compared. Leave-one-out cross-validation (LOOCV) identified SVM with minimum deviation. To further enhance model speed and precision, the SPA algorithm screened feature points in coke oven gas spectral data. Feature-based modeling retrieved multi-component concentrations. Results demonstrate that H₂, CO₂, CO, N₂, and CH₄ concentration predictions achieve determination coefficients (R²) of 0.9918, 0.9975, 0.9985, 0.9872, and 0.9958, respectively. The mean absolute percentage errors (MAPEs) measure 1.2%, 0.5%, 0.6%, 2.6%, and 0.8%, respectively. This approach delivers superior accuracy over full-spectrum modeling methods. This research provides methodological guidance for precise multi-component coke oven gas measurement.