Internal combustion (IC) engines produce during the first minutes after cold start higher emissions of hazardous gases. Experimental data from a state-of-the-art turbo-charged 3-cylinder, 999 cc gasoline engine are used to predict cold start emissions using two Machine Learning (ML) models: fully-connected feedforward Artificial Neural Networks (ANN) and encoder-decoder (ED) recurrent neural networks. Engine parameters and various temperatures are used as input for the models. Including ambient temperature as a training parameter enables both models to predict NOx, CO, and unburned hydrocarbon (UHC) emissions during cold starts across various environmental conditions. The dataset includes time series recordings from a Worldwide harmonized Light-duty vehicles Test Cycle (WLTC) and Real Driving Emissions (RDE) cycles at ambient and initial engine temperatures ranging from -20°C to +23°C. In total, 21 cases are considered, comprising 8 different ambient temperatures and 5 distinct driving cycles. Each case contains a sequence of 2500 samples collected at a rate of 5 Hz. The training process utilized 7 input variables and 3 output variables. A first scenario assessed the ability of the models to predict emissions at ambient temperatures not included in the training process. A second, more challenging scenario, tested the capability of the models to predict emissions for completely unknown driving cycles at temperature levels included in the training. While both models predicted the validation cases with good accuracy in the first scenario, the ANN model failed to predict the data in the second scenario. In contrast, the ED model proved to be more robust, yielding satisfactory results also for the second scenario. The low computational demands of the encoder-decoder model during the prediction phase render it suitable for real-time applications and effective for emission control purposes.