Thermal errors in machine tools significantly impact the precision of manufactured workpieces, particularly in high-accuracy processes such as grinding, where even minor deviations can lead to defects. Traditional mitigation strategies, such as active cooling and warm-up cycles, increase energy consumption and reduce productivity. Data-driven compensation models offer a more sustainable alternative by predicting and compensating for thermal deformations based on sensor data. This study explores different time-dependent models such as autoregressive models with exogenous inputs (ARX), long short-term memory (LSTM) networks, and temporal convolutional networks (TCN). These effectively capture thermal hysteresis effects but require initialization strategies to function immediately after machine start-up. To address this challenge, we evaluate various initialization, or “padding,” techniques, including zero initialization, leveraging prior temperature measurements, virtual predictions, and a dual-model approach that integrates a non-time-dependent (NTD) model for robust start-up compensation. Furthermore, to improve data efficiency in non-uniformly sampled scenarios, we assess different upsampling strategies such as linear interpolation, piecewise cubic Hermite interpolating polynomials (PCHIP), cubic splines, and Gaussian process regression (GPR) model-based interpolation. The proposed methods are validated on a five-axis grinding machine equipped with 13 temperature sensors and a 3D touch trigger probe. Results demonstrate that model-based interpolation significantly enhances compensation accuracy by incorporating temperature fluctuations between measurement cycles, while hybrid initialization strategies ensure accurate error compensation during critical start-up phases. The combination of optimized initialization and resampling strategies achieves a \(\sim 75\%\) reduction in average thermal error to below \(3.6\ \upmu \) m, providing a robust and energy-efficient solution for precision manufacturing.