<p> Thermal errors significantly affect the precision of machine tools, often contributing to 40–70% of total machining inaccuracies. Accurate compensation of these errors requires the strategic selection of temperature measurement points (TMPs) to ensure model robustness and efficiency. This study proposes a novel methodology for TMP selection that integrates long short-term memory (LSTM) neural networks with Shapley Additive Explanations (SHAP), offering a data-driven, interpretable framework for identifying critical sensor locations. The proposed LSTM-SHAP framework captures complex, nonlinear, and time-dependent thermal behaviors, and ranks TMPs based on their dynamic contribution to thermally induced structural displacements. The methodology was validated through both simulation using an ANSYS-based spindle model and an experiment on a 5-axis vertical machining center. In simulation studies, TMPs selected by LSTM-SHAP strongly aligned with those identified by thermal modal analysis (TMA) and grey relational analysis (GRA). In real-world testing, the LSTM-SHAP method demonstrated consistent performance and interpretability under varying operating conditions, successfully identifying key sensors correlated with thermal displacement. The convergence of TMP rankings across the three methodologies highlights the reliability of the proposed approach. This research establishes LSTM-SHAP as a powerful and scalable tool for optimizing sensor configurations in thermal error compensation systems, particularly in ultra-precision manufacturing environments.</p>