The global energy demand is still primarily reliant on fossil-fueled thermal power plants. With the growing share of renewables, these plants must frequently adjust their loads. Maintaining, or ideally increasing operational efficiency under these conditions is crucial. Increasing the efficiency of such systems directly reduces associated greenhouse gas emissions, but it requires sophisticated models and monitoring systems. Data-driven models have proven their value here, as they can be used for monitoring, operational state estimation, and prediction. However, they are also sensitive to (1) the training approach, (2) the selected feature set, (3) and the algorithm used. Using operational data, we comprehensively investigate these model parameters for performance prediction in a thermal plant for process steam generation. Specifically, four regression algorithms are evaluated for the prediction of the highly fluctuating live steam flow with two training approaches and three feature subsets of the raw dataset. Furthermore, manual and automatic clustering methods are used to identify different states of operation regarding the fuel amounts used in the combustion chamber. Our results show that the live steam flow is predicted with excellent accuracy for a testing period of one month (R2=0.994 and NMAE=0.55%) when using a dynamic training approach and a comprehensive feature set comprised of 48 features representing the combustion process. It is also seen that the statically trained model predicts various load changes with strong accuracy and that the accuracy of the dynamically trained model can be approached by incorporating the cluster information into the static model. These models reflect the plant’s physical intricacies under varying loads, where deviations from the predicted live steam flow indicate unwanted long-term drifts. They can be directly implemented to help operators detect inefficiencies and optimize plant performance.