Machine learning is nipping at the heels of conventional physical modeling of air quality predictions in more and more places. The latest is Johannesburg, South Africa, where computer engineer Tapiwa M. Chiwewe at the newly opened IBM Research lab is adapting IBM’s air quality prediction software to local needs and adding new capabilities. The work is an expansion of the so-called Green Horizons initiative, in which IBM researchers partnered with Chinese government researchers and officials, starting two years ago.
Last month, Chiwewe presented some of the Johannesburg lab’s first results, involving ground-level ozone level predictions, at the 14th International Conference on Industrial Informatics in Poitiers, France. “You can do a lot of physics to understand how ozone is found in different places,” he says, “but what we did is we just collected a lot of data and trained these machines on it and they were able to predict [local ozone levels] without any knowledge of how ozone works in the atmosphere.”
Like China, South Africa relies on coal power for a large part of its energy and suffers the consequences in the form of toxic, particulate-laden air. While Chiwewe says that he and his South African colleagues were able to re-use some of the air quality forecasting tool developed by their colleagues in China, they must also adapt it to local particularities. Johannesburg, for example, has a long mining history. The mining industry has left many exposed tailings piles and strong winds regularly pick up the finer particles, reducing air quality in neighborhoods downwind. Chiwewe says he hopes to develop a tool that might pick up on signs of rising winds and alert nearby residents.
Johannesburg also lacks Beijing’s dense network of air quality monitoring stations: It has just eight stations compared to Beijing, which has has 35, according to one report. The IBM system is designed to incorporate data also from other lower-cost sensors that might include just one or two types of measurements (such as particulate matter) rather than the full set of gases and particles measured at the main stations—Beijing may have around 1,000 of the smaller ones, Chiwewe says. So his team must adapt the “teaching” stage of their machine learning system to work with much sparser data and make up for it in creative ways. Until they have more sources of ground data they are working on an intermediate fix: so-called “virtual stations” that might use data from such remote-sensing platforms as satellites.
All this should help guide authorities, who for now are providing IBM Research with data from public monitoring stations in return for free access to the resulting forecasts. As the forecasts mature, officials there or elsewhere might use them to order heavily polluting power plants to reduce production during inversions or other smog-inducing weather patterns. They might also use the tools in reverse to identify the sources of pollution, helping to enforce existing laws. Longer-term forecasts could help officials plan road placement and zoning to try to reduce emissions, or at least their health consequences.