Resumo:
Climate change and the indiscriminate use of water have led to the scarcity of water resources, intensifying the water crisis. These factors put at risk the efficient supply of water to the population. Consequently, the development of techniques and methodologies to ensure the optimal operation of a water distribution network becomes essential. Therefore, this work developed a method for forecasting and optimizing losses in a benchmark water distribution network. The proposed methodology consisted of using three tools jointly, with the objective of predicting the worst loss scenarios of the study network for the next 24 hours of operation. These tools were: an adaptive digital twin prototype for simultaneous data exchange between server and client; a recurrent LSTM neural network for forecasting the worst loss scenarios; and a genetic algorithm for optimizing these scenarios based on the ideal configuration of operational parameters: pumps, valves, and reservoirs. For this purpose, the Python programming language and its libraries were used, such as WebSocket, Scikit-Learn, and Keras (TensorFlow). The most adverse loss scenarios found ranged between 51% and 53%, representing an average increase of 15.44 percentage points compared to the system's baseline condition. By applying the genetic algorithm, it was possible to achieve an average reduction of 11.42 percentage points in these scenarios, as well as identifying the operational configurations responsible for this improvement, allowing their incorporation as a mitigation strategy. Thus, the main innovation of this work lies in the integration of these three approaches into a single framework, enabling not only prediction but also optimization of critical operating conditions. The use of the digital twin allowed for analyses to be performed in a virtual environment, without risks to the real system, while the combination with artificial intelligence and optimization techniques contributed to increased operational efficiency and the anticipation and mitigation of adverse loss scenarios.
