Determining oil and water production in petroleum reservoirs involves conducting several simulations, which can be very costly due to the complexity of flow physics in real-world conditions. The high computational expense of these simulations has led to the development of proxy models that utilize historical production data, offering a more efficient alternative to traditional reservoir simulators. Various machine learning models, including recurrent neural networks (RNN), convolutional neural networks (CNN) and support vector regression (SVR), have been used in creating these proxies. However, these individual models have limitations, particularly in situations characterized by high complexity and uncertainty. In this work, we propose a proxy model based on ensemble methods to enhance accuracy and address uncertainties related to predictive modeling. The approach involves constructing a heterogeneous ensemble (different types of machine learning models), specifically integrating RNN and CNN. Additionally, using diverse activation functions (AFs). The goal is to leverage each model type’s strengths with different AFs for better performance. Our methodology was applied to a reservoir model that is well-documented in the literature. The results indicated that the proposal outperformed both homogeneous ensembles and individual models, reducing the mean and variance of the root mean square error (RMSE). Over ten runs, it demonstrated a maximum improvement of 23.3% in the mean and 67.36% in the standard deviation of the RMSE, compared to the RNN model on the test set.