Precision machine tools are critical components in the manufacture of high value components used in industries such as energy, automotive and aviation industries. The performance of machine tools can be severely affected by thermal errors arising from heat generated by internal and external sources as well as by the machining process. Empirical models are widely used to predict thermal errors which are then compensated by the numerical controller of the machine tools. Most of these models use temperature measurements from key structural points as inputs to predict thermal error at the tool centre point. However, literature shows that the accuracy of these predictions is not robust to changes in the machining conditions such as changes in spindle speeds and feed rates. This problem is aggravated by the fact that obtaining training data for empirical modelling is expensive in time and cost due to the associated machine downtime. Hence models are required to learn from relatively small datasets, which limits the sizes of models that can be trained and the robustness of their performance. The research aim of this thesis is to provide a framework for improving thermal error modelling under these settings. The first challenge faced when using temperature-based empirical thermal error models is determining an appropriate set of inputs for the model as well as the model’s structure. Sensor selection often involves tuning multiple parameters which limits the ease of implementation of the methods. This thesis proposes a novel model selection approach which automates selecting both the inputs and the structure of the empirical model making it easy to implement. The approach was implemented using both Adaptive Neuro-Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN) models and tested on a diverse set of machining cycles and operation parameters on two different machine tools. This thesis also proposes a novel model switching framework that uses clustering to identify and compare similarities in temperature data which are used to infer the prevailing machine operational conditions. With this, a model can compare new data to the data it has experienced during training. This forms a basis for deciding which model to use in predicting thermal errors. Proper Orthogonal Decomposition (POD) and Hidden Markov Models with Gaussian Mixture Model emissions (GMM-HMM) models are used in performing similarity measurement and classification. This forms the basis for a model-switching strategy in which multiple targeted models are used to overcome the limitations faced by using a single model. The results from validation experiments show that model selection robustly reduces the thermal errors from an overall maximum of 41 µm to residual errors with range values below 10 µm and Root Mean Square Error (RMSE) values below 5 µm. The Y-axis thermal errors in one machine tool can be reduced by 60 to 76% under different machining conditions. Results from classification of temperature data show that disparate machining cycles can be classified with an accuracy of over 97%. Additionally, data from the same machining cycle whose spindle speeds differ by over 2,000 rpm can be classified with an accuracy of over 73%. The results from model switching were positive and showed that through model switching, the predicted output could have better properties than the outputs from the individual models being switched between. However, further work is required to reduce the discontinuities introduced because of false class assignments in the switching process.