Automation of drinking water reservoirs using artificial intelligence

In this thesis study, it is aimed to enable existing drinking water control systems to be remotely monitored and suitable for machine learning-based data analysis without any hardware modification, by integrating a LoRa-based communication module. In this context, the study demonstrates that existing infrastructure can be digitized with a low-cost, energy-efficient, and recyclable module. Approximately 4 million rows of data were collected at 10-second intervals from a drinking water network located in Antalya between October 2022 and September 2023. This raw data was converted into 10-minute time series and subjected to analysis. During data preparation, erroneous and missing values were cleaned; time series analyses were performed on pH, redox, temperature, and chlorine levels. Seasonal distributions and day-night differences were statistically examined, and the findings were used as contextual features in anomaly detection models. For anomaly detection, unsupervised labeling was first performed using the Isolation Forest algorithm, followed by supervised modeling using Random Forest, k-NN, Logistic Regression, and Gradient Boosting. Contextual variables based on time were added to the dataset through feature engineering, which significantly improved model accuracy. The Logistic Regression algorithm achieved up to 84% F1 score, while Random Forest consistently demonstrated the highest performance across all scenarios. As a result, it has been proven that the system can be made "smart" through the integration of a communication module and software-based data analysis. The proposed approach offers a field-applicable, scalable, and sustainable solution through both hardware integration and advanced data analysis processes.