Chemiresistive gas sensors are a crucial tool for monitoring gases on a large scale. In order to properly estimate certain gas concentrations and to differentiate between different gases, pattern recognition algorithms, such as neural networks, are used to analyze the complex signals describing the resistivity or conductivity on the sensor material, e.g. MOX or graphene. These algorithms are usually trained on experimental data based on sample sensors measured under either controlled laboratory conditions or in open scenarios involving a reference device. However, in the production process of such low cost sensor technologies, small variations in the physical properties of the sensors can occur. This means that the reaction of a single sensor to a certain concentration can slightly vary from the original sensor used for algorithm development. An example for such a variation would be the operating and heating temperature of the device. In order to study the influence of such variations on the overall performance of pattern recognition algorithms, we used a stochastic simulation model of a graphene-based gas sensor to generate data of different concentration profiles exposed to sensors with different heating settings. Subsequently, we trained machine learning models on different subsets of the synthetic data to estimate the influence of temperature variations on the prediction outcome and to study which training data configurations might increase the performance of the sensors under varying input conditions. Our results show that different temperatures can steadily lower the performance of the algorithms. Moreover, a well-balanced training set featuring several measuring temperatures can increase the robustness of the prediction algorithms.