Room Occupancy Estimation

This project explores room occupancy estimation using ML techniques, analyzing environmental sensor data to optimize energy consumption and space utilization efficiency. Employing Support Vector Machines(SVM), Logistic Regression, and Neural Networks(NN), alongside Principal Component Analysis(PCA) for dimensionality reduction, achieved a precision score of 98.7% in occupancy prediction, highlighting the role of ML in smarter infrastructure planning and resource management.

Data Description

The dataset used in this study originates from the UCI Machine Learning repository and is credited to Adarsh Pal Singh, Vivek Jain, Sachin Chaudhari, Frank Alexander Kraemer, Stefan Werner, and Vishal Garg. The dataset is prominently featured in the research paper titled "Machine Learning-Based Occupancy Estimation Using Multivariate Sensor Nodes," presented at the 2018 IEEE Globecom Workshops (GC Wkshps).

Link: UCI Machine Learning Repository - Room Occupancy Estimation

The dataset consists of 10,129 instances with 19 features, including time series data on temperature, light, CO2 levels, PIR, and sound, gathered from sensor nodes in a 6m x 4.6m room.

EDA

• Feature Engineering: Combined the 'Date' and 'Time' columns to create a new 'Date_time' column that helped to categorize room occupancy into different time segments ('Morning,' 'Afternoon,' 'Evening,' 'Night') by creating Time_of_Day column and converted them to numerical form using LabelEncoder. The 'Date_time' column was dropped to adhere to the independence assumption, facilitating the application of ML algorithms, which perform optimally under this condition.
• Data Preprocessing: High correlation among features S1_temp, S3_temp, and S4_temp, suggesting their similar impact on the target variable, thus warranting the removal of two features (S1_temp and S3_temp) to streamline the dataset's dimensionality.

Modeling

Summary

Model	Precision	Accuracy	Recall	F1 Score
Logistic Regression without PCA	95.42%	98.47%	95.02%	95.22%
Logistic Regression with PCA	83.69%	92.60%	76.22%	79.78%
SVM without PCA	94.71%	94.66%	94.66%	94.57%
SVM with PCA	83.94%	83.98%	83.98%	83.85%
Neural Networks	98.7%	98.62%	98.62%	98.61%