Despite the intensive development of FE simulations for cold forging applications over the last decades, they are still prone to errors due to, among other things, inaccurate material and friction modelling. The use of advanced friction models can reduce the error caused by friction modelling. [1] However, existing models for cold forging are often limited to a specific application and require extensive tribometer testing for parameter determination. This work presents a new method for efficient data collection through time series analysis, which significantly reduces the number of tribometer tests required. The new method also allows the use of deep learning algorithms for friction modelling. Using the new method, five different friction models, including one deep learning model, are trained and implemented in the FE simulation. Using two typical forming processes for validation, it is shown that the use of a feed-forward neural network friction model reduces the relative error of the FE simulation by ~59% compared to simple friction models. Compared to the state of the art method, the time series based data collection approach reduces the necessary experimental testing by 62 %. Furthermore, the advanced friction models presented are not limited to a specific process, but can be used for any type of cold forging simulation.