Predictive Performance Evaluation of ARIMA and Hybrid ARIMA-LSTM Models for Particulate Matter Concentration

Abstract

This study objectively evaluates prediction models for particulate matter policy for industrial stakeholders by comparing the ARIMA (Auto Regressive Integrated Moving Average) and hybrid ARIMA-LSTM (Long Short-Term Memory) models for predicting air quality data in industrial environments. For PM 1.0 concentration, the ARIMA model has an RMSE of 8.29 and an error ratio of 0.45, while the hybrid ARIMA-LSTM model achieves an RMSE of 3.54 and an error ratio of 0.22. For PM 2.5 concentration, the ARIMA model shows an RMSE of 6.61 and an error ratio of 0.66, compared to the hybrid ARIMA-LSTM model's RMSE of 2.68 and an error ratio of 0.19. The best ARIMA models identified are (2,0,1) for PM 1.0 and (1,0,1) for PM 2.5. The hybrid ARIMA-LSTM model outperforms ARIMA, with improved RMSE and error ratio values by approximately 57.30% and 51.11% for PM 1.0, and 59.46% and 71.21% for PM 2.5, respectively. This superior performance is due to the hybrid model's ability to handle variable-length sequences and capture long-term relationships, making it more resistant to noise and enhancing prediction accuracy.