The aerospace sector is in the constant need of large-scale, highly precise components manufactured in challenging materials with tight tolerances by chip removal processes [1]. When machining, ambient temperature changes, the heat generated by machine moving components and motors, and the chip removal process itself, lead to dimensional changes in structural components of machine tools, and a subsequent positioning error. Thermally induced positioning errors of large-scale machines can represent up to a 75% of the total error [2]. To this day, avoiding temperature-induced errors in the machined workpiece remains a challenge. This work presents a methodology for measuring the impact of ambient and machine usage temperature changes, and a simple yet useful relation between temperature or deformation measurements and Tool Centre Point (TCP) positioning errors. The research was experimentally conducted in a large-scale, five-axis machining centre, in which, apart from numerous internal sensor values and CNC variables, a total of 49 temperature and 14 Integral Deformation Sensor measurements were recorded and correlated to tool positioning errors. Errors for different temperature conditions were measured using a calibrated metrological artifact. The obtained model, capable of estimating thermal errors in TCP can be easily applied to compensate them in real time with low computing requirements, as proposed at the end of the article. Therefore, implementing it to compensate thermal deformations in industrial environment is straightforward. In this article, two similar models are presented, one based on the temperature measurements, while the other considers deformation measurements. The main objective of this research was not only to predict thermal errors, but to do it with a volumetric approach. Thus, the proposed models, which both show accurate error predictions, consist of two staged non-linear models. Firstly, volumetric error is modelled by a non-linear multivariable fit depending on a low number of parameters. In a second stage, each parameter is fitted by other non-linear multivariable functions depending on the most representative sensors, thus considering thermal effects. The results showed a 77% error reduction when using the model based on temperature measurements.