The interaction between intrinsically disordered proteins and lipids plays a pivotal role in various cellular functions and human diseases. Accurately identifying disordered lipid-binding residues (DLBRs) is essential for elucidating protein functions and advancing drug development. However, only a limited number of native annotations for DLBRs are currently identified. Existing computational methods for DLBR prediction often exhibit insufficient predictive accuracy and require multiple feature extraction processes, underscoring the demand for an efficient and precise predictive approach. We propose DDLB, a predictor that utilizes a protein language model (PLM) and a hierarchical architecture for DLBR prediction. Protein sequences are encoded by the PLM to reduce reliance on time-consuming database searches, then passed through the hierarchical architecture to predict DLBRs. Firstly, inputs are divided into several regions to calculate their binding propensities. Next, potential disordered lipid-binding regions are processed by residue level layers to identify specific DLBRs, while residues in non-binding regions are all labeled as non-binding. Evaluation results reveal that DDLB surpasses existing methods, particularly in identifying DLBRs within disordered regions compared to disordered lipid-binding residue, lipid-binding residue, and disordered binding residue predictors. Furthermore, ablation analysis demonstrates that the hierarchical architecture offers more accurate predictions compared to a single-layer architecture.