Carbon Monitor AutoForecast-Asia: a real-time emission estimates of the residential sector for Asian major emitters with an automatic machine learning framework

Carbon monitoring is crucial for mitigating urban climate change and expediting urban responses to emission changes. This requires monitoring and reporting emissions on a timely basis. However, existing inventories mainly consider scope-1 (in boundary emissions) with a time lag of more than 2 years. The Carbon Monitor (carbonmonitor.org) and the following Carbon Monitor Cities (cities.carbonmonitor.org) project thus developed the near-real-time (NRT) inventories for the global high emitters worldwide at the country level and city level, pushing forwards the efforts of global NRT monitoring in an around time lag of 3 months to ensure the unprecedented timeliness of carbon data. However, the immediate monitoring of emissions in more rapid governmental responses and short-term future policy design is still urgently needed in the future, especially for getting real-time estimates of major emitters by developing the “nowcasting” and even forecasting framework. We choose the residential sector as the initial exploration of this kind given the sectoral characteristics of Asian big emitters, including China, Japan, and India.

By developing an automatic machine learning ensemble framework (Carbon Monitor AutoForecast-Asia), we essentially achieve the nowcasting from the perspective of city-level emission dynamics. Specifically, this framework utilizes data from multiple sources, including the Carbon Monitor datasets, Carbon Monitor Cities datasets, and the ERA5 reanalysis datasets. Time-series and tree-based models are incorporated into the entire framework with automatic finetuning pipelines, as well as designed algorithms for capturing the seasonalities of daily emissions, and the holiday impacts as demonstrated in our previous studies. We use the Carbon Monitor and the Carbon Monitor Cities datasets up to the end of 2021 to train and test the entire framework by parallelizing the framework of more than 400 cities for acceleration. After running correction models to reduce the induced uncertainties from the Carbon Monitor to the Carbon Monitor Cities, we get the R squared metrics of 0.95, 0.94, 0.97 for China, India, and Japan respectively at the country level. Note that, we use the Carbon Monitor data from 2022.1.1 to the latest for comparisons in this procedure, and the framework could nowcast the residential emissions in the time lag of 1 week.

Subsequently, deep learning-based models (i.e., LSTM) are used as the baselines with the same configurations (i.e., train and test splits) and we find that the ensemble results could outperform the baseline models in terms of common metrics, including MAE, MSE, RMSE, MAPE. This suggests that the real-time estimates of emissions may depend on more complicated ensemble methods rather than the deep learning models due to the trade-off between the small volume of near-real-time data and the complex patterns of daily emissions.

In the future, we may include more high emitters globally by extending the developed framework to at least satisfy the needs of real-time and even future estimates of the residential sector.