Evaluation of nine machine learning regression algorithms for calibration of low-cost PM2.5 sensor

Low-cost sensors (LCS) can construct a high spatial and temporal resolution PM2.5 network but are affected by environmental parameters such as relative humidity and temperature. The data generated by LCS are inaccurate and require calibration against a reference instrument. This study has applied nine machine learning (ML) regression algorithms for Plantower PMS 5003 LCS calibration and compared their performance. The nine ML algorithms applied in this study are: (a) Multiple Linear Regression (MLR); (b) Lasso regression (L1); (c) Ridge regression (L2); (d) Support Vector Regression (SVR); (e) k-Nearest Neighbour (kNN); (f) Multilayer Perceptron (MLP); (g) Regression Tree (RT); (h) Random Forest (RF); (i) Gradient Boosting (GB). The comparison exhibits that kNN, RF and GB have the best performance out of all the algorithms with train scores of 0.99 and test scores of 0.97, 0.96 and 0.95 respectively. This study validates the capability of ML algorithms for the calibration of LCS.

Automated summary

This study compared nine machine learning algorithms for the calibration of low-cost sensors, with kNN, RF, and GB showing the best performance.