BACKGROUND Pain is common and clinically consequential in children with cerebral palsy (CP), but retrospective pain severity assessment from electronic health records (EHRs) is difficult. Pain evidence is often dispersed across long narrative notes, many children have limited ability to self-report, and structured pain scores may be influenced by proxy reporting and encounter context. OBJECTIVE This study introduces the first systematic benchmark for three-class pain severity classification from clinical notes of children with CP and evaluates whether modern Natural Language Processing (NLP) and large language model approaches can recover clinically meaningful severity signals from real-world EHR documentation. METHODS We conducted a retrospective study using 15,969 de-identified clinical notes from 1,467 children with CP treated across the Shriners Children’s Hospital Network over 19 years. Notes were linked to encounter-level numeric pain ratings and grouped into No/Mild (0–3), Moderate (4–6), and Severe (7–10) classes. Using patient-stratified splits, we compared supervised encoder-fusion models, explanation-augmented local language models, frozen instruction-following LLMs, and a late-fusion ensemble. RESULTS Performance improved as models incorporated broader context and complementary decision patterns. The best supervised encoder-fusion model achieved a macro F1 of 0.600, while explanation-augmented local LLMs improved to 0.620 when clinical notes were combined with generated explanations. Few-shot GPT-5 was the strongest standalone model in this held-out test set, achieving a macro F1 of 0.644. The best overall performance was achieved by a late-fusion ensemble of Bio-Mistral-7B, zero-shot GPT-5, and few-shot GPT-5, reaching a macro F1 of 0.66. SHapley Additive exPlanations (SHAP) analysis showed that No/Mild predictions were driven mainly by negated pain language, whereas Moderate and Severe predictions were associated with explicit pain terms and severity cues. CONCLUSIONS Modern NLP and LLM-based approaches can extract clinically meaningful pain severity signals from real-world EHR notes of children with CP. Performance was strongest when models used long-context modeling, explanation-guided supervision, and complementary prediction patterns. Remaining errors reflected the subjective, proxy-mediated, and temporally variable nature of pain documentation. Therefore, these models should support retrospective pain research, cohort identification, and documentation review. To our knowledge, this is the first study to benchmark multi-level pain severity classification from unstructured EHR notes of children with CP.