The escalating global issue of plastic waste, with over 380 million tons produced annually and recycling rates at a mere 9 %, necessitates more effective identification and sorting methods. Conventional approaches struggle with visually similar polymers, reducing the efficiency of mechanical recycling. In this study, we aimed to develop a methodological framework for accurate plastic classification by combining machine learning with handheld Raman spectroscopy. A dataset of 3000 spectra was collected from 10 common plastic types (PET, HDPE, PVC, LDPE, PP, PS, ABS, PC, PLA, PTFE) under varied measurement conditions. Raman spectroscopy provides unique molecular fingerprints for different plastics, while machine learning enhances classification accuracy, particularly for complex mixtures or contaminated samples encountered in recycling streams. To address the classification challenge, we proposed a branched neural network architecture (Branched PCA-Net), inspired by Deep&Wide design and operating on Principal Component Analysis (PCA) reduced spectral data. This network employs separate paths for high-, medium-, and low-variance principal components prior to final classification. The model achieved over 99 % accuracy on the test dataset, with perfect classification for 7 out of 10 plastics and high accuracy for the remaining three. Robustness was further validated on new samples measured under different conditions, confirming strong generalization capabilities. This study provides a methodological contribution through the introduction of the Branched PCA-Net architecture, which shows high potential for spectroscopic data analysis. While not directly scalable to high-throughput sorting, the approach offers significant promise for quality control and targeted plastic identification in recycling workflows.