Ozone air concentration trend attributes assist hours-ahead forecasts from univariate recorded data avoiding exogenous data inputs

Abstract

Multiple trend attributes extracted from univariate hourly ozone recorded data can be effective in forecasting ozone concentrations at near surface sites for t0 to t + 12 h ahead without recourse to exogeneous variables. The method is evaluated with datasets from three cities less than 100 km apart in central/eastern England, each with more than forty thousand data records for the period 2016 to 2020. The 2020 recorded ozone distribution values are higher in all three cities than for 2016 to 2019 due, at least in part, to COVID-19 lockdowns limiting vehicle emissions. Fifteen attributes extracted from the recorded ozone trend for the past twelve hours are added to each hourly data record. The attributes include seasonal components, some prior-hour values, averages, differences and rates of change. Two multi-linear regression and eight machine-learning (ML) models are used to predict 2019 and 2020 hourly ozone values with the attribute-endowed datasets. The forecasting accuracy of all but one of these models outperforms that of an autoregression model applied to the univariate recorded ozone trends. The support vector machine model achieves the highest ozone forecasting accuracy for hours ahead t0 to t + 12. However, nine of the other models also providing credible and consistent forecasts for the datasets for all three cities. Coefficient analysis of the multi-linear regression models reveals the flexibility with which each of the trend attributes is used in predicting different hours ahead in the t0 to t + 12 range. The attribute-endowed datasets also enable the ML models to assign different relative weights to each attribute for the different hours-ahead being forecasts. This capability introduces additional dimensions that are not available to autoregressive or moving-average models applied to univariate ozone trends.