The accessibility to freshwater suitable for human use is a modern problem in many countries of the world. One of the well-known methods to overcome this problem is the reverse osmosis (RO). The performance of a reverse osmosis unit integrated to a recovery energy system was experimentally investigated under various operating system pressures (10, 15, 20, 25, 30, 35, 40, 45, 50, 55, and 60 bar) and recovery ratios (10%, 20%, 30%, 40%, and 50%). Moreover, a hybrid machine learning model composed of Long Short-term Memory (LSTM) neural network optimized by artificial hummingbirds’ algorithm (AHA) was developed to predict permeate flow and power saving of the investigated RO unit. The inputs of the models, in case of power saving, were recovery ratio and system pressure; while system pressure was the input of the models in case of permeate flow. AHA was employed to optimize the performance of pure LSTM via determining the optimal values of the model parameters. A considerable enhancement in prediction accuracy of the optimized model was observed compared with pure model. The coefficient of determination during testing phase of power saving prediction was 0.997 and 0.981 for LSTM-AHA and LSTM, respectively. While it was 0.992 and 0.97 for LSTM-AHA and LSTM, respectively, in case of permeate flow prediction. Furthermore, the saving in consumed power of the RO unit was declined with increasing the recovery ratio. Therefore, the best saving in consumed power was obtained for the recovery ratio of 10%, where it reported more than 85%.