Optimising Deep Learning at the Edge for Accurate Hourly Air Quality Prediction

Area of Science:

• Environmental Science
• Computer Science
• Machine Learning

Background:

• Centralized machine learning models struggle with resource-constrained edge devices for air quality monitoring.
• Processing multi-dimensional, multi-location sensor data for accurate air quality prediction is challenging.

Purpose of the Study:

• To design a novel hybrid deep learning model for hourly PM2.5 prediction.
• To optimize this model for edge device deployment, focusing on accuracy and latency.
• To evaluate the performance of the optimized model on Raspberry Pi devices.

Main Methods:

• Developed a hybrid deep learning model combining 1D Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM).
• Optimized the model for edge devices (Raspberry Pi 3 Model B+ and 4 Model B) using post-training quantization.
• Evaluated model performance using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), and measured execution latency.

Main Results:

• The proposed hybrid CNN-LSTM model outperformed existing deep learning models in PM2.5 prediction accuracy.
• Model optimization reduced file size significantly, with quantization further decreasing it.
• The optimized model demonstrated efficient performance on edge devices, with Raspberry Pi 4 executing twice as fast as Raspberry Pi 3.

Conclusions:

• The novel hybrid deep learning model is effective for accurate, low-latency PM2.5 prediction on resource-constrained edge devices.
• Model optimization and quantization are crucial for deploying advanced AI on edge hardware.
• This approach enables practical, real-time air quality monitoring in diverse environments.