A reliable forecast for oil production represents one of the biggest challenges in the oil and gas industry and contributes to the planning and decision making of oil companies. Because of that, this work uses intelligent well valves settings and data driven methodology to explore the advantages and the performance of machine learning algorithms in the forecasting of oil, gas and water production. In order to do so, two database containing historical data series of oil, gas and water production were used. The first was generated synthetically (through reservoir simulation) and consisted of the average monthly production of an injection well over a period of 10 years, as well as the configuration of 3 of its valves. The second database used the production data provided by the state of South Dakota, located in the United States, and consisted of the daily production average and the overall well status (active or not producing) from several oil producing wells in a period ranging from 1950 to 2018. In order to test the methodology, several experiments were performed combining proxy with Artificial Neural Network Algorithms (Multilayer Perceptron) and deep learning recurrent neural networks (Simple Recurrent Neural Networks, long short-term memory, Gated Recurrent Units), which were named smart proxy and deep smart proxy, respectively. The results showed that the deep smart proxy model was very promising. Using the Gated Recurrent Units network with bi-directional layers (GRUB), a reduction of 66 percent in the RMSE error and 79 percent in the MAE error was obtained when compared to smart proxy models with Artificial Neural Networks. The deep smart proxy models with bidirectional layers generated a significant improvement in prediction and error reduction in both databases tests ( i.e. tests with simulated production data (synthetic case) and with the observed production data (real case), resulting in a variation of up to 75 percent in RMSE and 85 percent in MAE). The normalized RMSE error in the GRUB network was of 0.53 percent in the observed database and 0.65 percent in the synthetic database. It is important to notice that the Deep smart proxy models achieved very similar performances when comparing the LSTMB and GRUB network in both databases (synthetic and real production), surpassing in all cases the results obtained with the MLP smart proxy model.