A Multi-task Learning Model for Daily Activity Forecast in Smart Home.

Area of Science:

• Artificial Intelligence
• Smart Home Technology
• Machine Learning

Background:

• Daily activity forecasting is crucial for smart home residents.
• Existing methods often focus on either activity category or occurrence time prediction.
• Sequential single-task models yield suboptimal performance.

Purpose of the Study:

• To develop an integrated model for predicting both daily activity category and occurrence time.
• To leverage multi-task learning for improved accuracy and efficiency in activity forecasting.
• To address the limitations of existing single-task learning approaches.

Main Methods:

• A multi-task learning framework was designed to mutually and iteratively forecast activity category and occurrence time.
• Raw sensor data was pre-processed into a feature space representing daily activities.
• A parallel model combining Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) was employed.

Main Results:

• The proposed model demonstrated superior performance across five datasets compared to state-of-the-art single-task models.
• Accuracy was improved by at least 2.22%.
• Normalized Mean Absolute Error (NMAE), Normalized Root Mean Square Error (NRMSE), and R-squared (R²) metrics showed significant improvements of at least 1.542%, 7.79%, and 1.69%, respectively.

Conclusions:

• The multi-task learning approach effectively integrates category and occurrence time forecasting for daily activities.
• The CNN-Bi-LSTM architecture provides a robust solution for complex activity pattern recognition in smart homes.
• This model offers a significant advancement in the accuracy and reliability of smart home activity prediction systems.