Among the parameters that allow evaluating the degree of pollution of a river, one of the most efficient is dissolved oxygen concentration (DO), as it reflects the balance between the production and consumption of oxygen in aquatic ecosystems. Due to the dynamic characteristic of DO concentration, especially in rivers and wetlands, it is highly recommended to generate DO models periodically for aquatic ecosystems so that quality control measures can be optimized over a time horizon. To this end, implementation of different artificial intelligence (AI) techniques have been suggested in the relevant literature and among these techniques, neural networks (ANNs) have been successfully applied to water quality estimates. In this context, the present work describes the development of a model in Convolutional Neural Networks – CNN (Convolutional Neural Networks) with the objective of simulating the self-cleaning potential of the Rio Alegria, located in the municipality of Medianeira in the State of Paraná . The data used was the same as that collected by Schütz in 2014, so that it was possible to compare the performance of the CNN network with the performance of the Feed Forward Network (FFN), developed with the same set of data collected in 2014. The models were developed based on data on the quality of the river's water and the effluent that is incorporated into the watercourse throughout the studied interval. The model was named, CNN5. Tests and validations were carried out by varying the network architecture with cross-validation to estimate dissolved oxygen. Considering the results referring to the simulations carried out with the CNN5 model, where the simulated OD values are compared (from the combination of the weights that the network assigns to each input), with the results collected in the field and it can be concluded that a CNN network can be used to predict dissolved oxygen in the waters of a stretch of river, with an overall accuracy of 0.90, while FFN5 managed to achieve an accuracy of 77% for the same group of data. processing, a reduction from 12 hours to two minutes can be observed.