Performance of artificial neural networks and linear models to quantify 4 trace gas species in an oil and gas production region with low-cost sensors

[Display omitted] •An increasing number of people live near oil and gas production activities.•Industry related air quality impacts have let to growing public health concerns.•Low-cost sensors may support distributed air quality measurements in these areas.•Sensor field calibration is useful in quantification of some relevant gas species.•Artificial neural networks performed generally well and better than linear models. We tested the performance of regression via inverse linear models (LMs), direct LMs, and artificial neural networks (ANNs) towards field calibration of low-cost gas sensors in an area influenced by oil and gas production activities to quantify O3, CO, CO2, and CH4 in ambient air. Sensors were co-located with reference measurements in Greeley, Colorado. We selected a three-month period of data in the spring of 2017 to conduct our analysis. Approximately two months of measurements bookending the middle test month were used for model training. We found that ANNs generally outperformed LMs and that direct LMs generally outperformed inverted LMs. An analysis of model residuals during the test period revealed that ANNs were better able to mitigate curvature and linear trends relative to direct LMs with the same set of inputs.