Enhancing low-cost PM2.5 air quality monitoring sensors through sensor calibration using regression and echo-state network models.

Low-cost PM2.5 sensors have gained popularity due to their simplicity and low maintenance. However, calibration challenges often limit their accuracy and reliability. In South Asian countries, strong seasonal meteorological variation can require dynamic sensor calibration approaches. This study investigated sensor performance of TSI BlueSky low-cost PM2.5 sensors within a harmonized sensor network using linear regression (LR), multiple linear regression (MLR), and echo-state network (ESN) models across two distinct climatic zones in Sri Lanka. Results indicated that the harmonized sensors were generally in close agreement with one another, with 10% variability. The results of LR, MLR, and ESN indicate that all three models significantly enhance performance with increasing temporal averaging windows, especially from the simple LR model at 24-h averaging in the wet season (R2 = 0.60, MAE = 0.93 µgm-3, MAPE = 7.81%) and dry season (R2 = 0.85, MAE = 1.98 µgm-3, MAPE = 9.17%). Applying wet season calibration values to dry season data results in a significant mean absolute percentage error (26.57%). The effectiveness of calibration models developed in Colombo decreased when they were applied to PM2.5 data from Kandy, suggesting that they have limited transferability across different climatic zones. Conversely, models that were specific to Kandy and incorporated temperature and relative humidity demonstrated a substantial increase in precision during both wet (R2 = 0.84, MAE = 1.78 µgm-3, MAPE = 11.71%) and dry (R2 = 0.92, MAE = 1.59 µgm-3, MAPE = 7.73%) seasons. These findings emphasize the critical need for both temporal and spatial calibration strategies to enhance PM2.5 estimation reliability for low-cost sensors.