Exploring Copula-based Bayesian Model Averaging with multiple ANNs for PM2.5 ensemble forecasts

Abstract

Quantifying predictive uncertainty of ensemble air quality forecast is very crucial and challenging. This study integrated a Copula-based Bayesian Model Averaging (CBMA) and multiple deterministic artificial neural networks (ANNs) to make accurate ensemble probabilistic PM2.5 forecasts. The new approach (CBMA), has a flexible structure that grants the posterior distribution to have any shape owing to the Copula function. The CBMA approach could remove the data transformation and bias correction procedures as it is done in the original BMA, which was taken as the benchmark. The air quality in Taipei City of Taiwan was selected as a study case to evaluate the applicability and reliability of the proposed approach. Three kinds of air quality monitoring stations denoted heavy traffic loads, intensive commercial trading and human intervention, and a natural circumstance with fewer human activities respectively. The forecasts of PM2.5 concentrations were regarded as a math function involving meteorological and air quality variables, using long-term (2010–2018) hourly observational datasets. Firstly, four deterministic ANN models were established and evaluated to provide inputs for ensemble forecasting. Then, the two post-processing techniques (i.e. CBMA and BMA) were employed to produce ensemble probabilistic forecasts based on the forecasts obtained from multiple ANN models. The results demonstrated that the CBMA not only could outperform the BMA but also could provide a practical and reliable approach as a complement to multiple deterministic ANN models to create ensemble probabilistic forecasts. From horizons t+1 up to t+4, the CBMA approach could drive up the Containing Ratio (CR) values by 3.12% ˗ 9.58% as well as reduce the average Relative Band-width (RB) values by 8.63% ˗ 34.48% and the Continuous Ranked Probability Score (CRPS) values by 7.62% ˗ 32.89%, in comparison with the BMA one. Consequently, the predictive uncertainty could be alleviated while model reliability and PM2.5 forecast accuracy could be considerably increased.