NEURAL NETWORK MODEL FOR DETECTING LEAKS IN THE WATER SUPPLY AND DISTRIBUTION SYSTEM OF THE SOUTHERN COAST OF CRIMEA

The water supply and distribution system often face the problem of leaks, which pose risks to the uninterrupted water supply of the population. At the same time, traditional diagnostic methods often do not allow leaks to be detected before obvious consequences appear. Within the framework of this study, a machine learning model has been developed designed for early detection of leaks in SPRW. To form a training sample, hydraulic modeling was performed using the ZuluGIS geoinformation system. During the modeling process, the following factors were taken into account: the dynamics of water consumption by users; possible distortions and interference in the data; different leakage rates and other significant factors. The subject. The reliability of urban water supply systems, taking into account topographical features. Materials and research methods: The analysis of the technical condition of the pipelines of the Yalta gas pipeline system was carried out. A machine learning (MMO) model for detecting leaks in the Yalta open air defense system has been investigated. Data on water pressure under leak conditions and the absence of leaks were obtained using the ZuluGIS geoinformation system, taking into account factors such as changing user requirements, data interference, the degree of leaks, etc. An artificial neural network (INS-OU) model has been developed to detect leaks. For this purpose, data on water pressure in a group of monitoring nodes was used. Unlike existing approaches using time series analysis, water pressure data is used by the INS-OU model to determine the spatial relationship between data in monitoring nodes at a given time. Results. The conducted research has shown that an artificial neural network (ANN) is able to accurately distinguish between two states of the system — the presence or absence of leakage. However, a balanced data set that includes examples of both scenarios is crucial for the effective operation of the model. In real-world conditions, this is difficult, since the SPRV mainly operates normally, and leakage cases are relatively rare. The study also revealed that the accuracy of the models depends on the location of the leak: when leaks occur in the sensor coverage area, the models demonstrate high accuracy; outside the sensor monitoring area, the prediction accuracy decreases significantly. The results obtained allow us to formulate recommendations on the optimal placement of monitoring sensors, ensuring the necessary coverage of the controlled area. Conclusions. A data-driven approach using ANN models is promising for fast and reliable leak detection. The rationale is that the spatial structure of water pressure and its changes during leakage depend on the structure of the water distribution network, which allows us to obtain information about the conditions in the urban water supply network.