Modeling of Particulate Pollutants Using a Memory-Based Recurrent Neural Network Implemented on an FPGA.

Area of Science:

• Environmental Science
• Computer Engineering
• Artificial Intelligence

Background:

• Accurate air quality assessment is vital for protecting ecosystems and human health.
• Criteria air pollutants include nitrogen dioxide (NO2), carbon monoxide (CO), and particulate matter (PM10, PM2.5).
• Overexposure to these pollutants poses significant risks.

Purpose of the Study:

• To train and implement a Long Short-Term Memory (LSTM) neural network on an FPGA board for air pollutant prediction.
• To evaluate the performance of a modified LSTM architecture on resource-constrained systems.
• To assess the feasibility of integrating predictive models into embedded systems for real-time air quality monitoring.

Main Methods:

• Development and training of an LSTM neural network model.
• Implementation of the trained model on a Field-Programmable Gate Array (FPGA) board.
• Comparative analysis of the FPGA-implemented model against the original LSTM network.

Main Results:

• The FPGA-implemented LSTM network demonstrated an 11% performance improvement over the original model.
• The modified architecture maintained prediction accuracy despite a reduction in neurons.
• Accurate predictions were achieved for both 24-hour and 72-hour time frames.
• Feasibility of integrating prediction networks into limited systems like FPGAs was confirmed.

Conclusions:

• The proposed LSTM network on an FPGA board offers an efficient solution for air pollutant modeling and prediction.
• This approach enables the deployment of advanced AI models in embedded systems without compromising accuracy.
• The findings highlight potential for enhanced air quality monitoring systems and further model optimization.