Detecting household toilet flushing events using machine learning

Flow chart showing the process of water end use disaggregation. From left to right: household water patterns, flow rate time series/physical features extracted from water patterns and comparison of two training data types for random forest. Household water consumption plays a big role in water demand and supply management. Water utilities have growing interest in insights into how water is used in different types of households at a microcomponent level. In this study, a random forest model was trained to detect household toilet flushing events using different types of input features. The water consumption data were collected in the southwest of the UK, and an online survey was conducted to obtain essential information on household water end-use characteristics. Two types of random forest-based models are developed with different input features: one is trained with physical features and the other is trained with flow rate time series features. The test F1-score of the model using physical features reaches 88.46%, which is higher than the F1-score of 85.62% achieved using time series input. As comparison, a gated recurrent unit model was also trained with time series data, which achieved an F1-score of 86.18%. Based on the detection results, toilet use profiles for the two household groups are generated, revealing different behaviors between two groups. Overall, the proposed model demonstrates promising detection accuracies and holds potential for application in supporting household water demand disaggregation and management.