AquaAware: A Two-Phase Learning System for Water Conservation

AquaAware is a project designed to reduce water consumption in homes by providing real-time recommendations for various tasks like brushing teeth, washing hands, showering, and washing dishes. The system employs a two-phase machine learning algorithm to first train users on best practices and then further optimize their personal habits.

Phase 1: Supervised Learning for Best Practices

Initially, the system uses a dataset containing optimal water consumption levels and time durations for common household tasks.

• Algorithm: A Decision Tree classifier is trained on this “best practice” data.
• Functionality: As the user consumes water, the device records the usage and time. The Decision Tree predicts the ongoing task and provides real-time recommendations to help the user align their habits with water-saving standards.
• Goal: The primary objective of this phase is to guide the user towards adopting more sustainable habits and reducing initial water wastage.

Phase 2: Unsupervised Learning for Personalized Optimization

Once the user has consistently adopted the recommended best practices, the system transitions to a personalized optimization phase.

• Data Collection: The system begins to store the user’s new, more efficient water usage patterns in a new dataset, but without task labels.
• Unsupervised Learning: Since the device cannot automatically classify these new patterns, the K-Means clustering algorithm is used. Knowing the number of tasks beforehand, K-Means groups the unlabeled data into distinct clusters, effectively re-identifying the tasks based on the user’s unique behavior.
• Retraining: Once the new data is labeled through clustering, a new Decision Tree model is trained on this personalized dataset.
• Goal: This second model provides even more refined predictions and recommendations, pushing the user to consume only what is strictly necessary, thereby minimizing waste beyond the standard best practices.

Design Choices

• Decision Tree was selected as the supervised learning algorithm due to its high accuracy and significantly faster response times compared to alternatives like K-Nearest Neighbors (KNN).
• K-Means was chosen for the unsupervised learning task because the number of clusters (i.e., the number of distinct household tasks) was known in advance, making it a highly efficient choice.