The DeNOx System: Machine Learning for Process Control

At Hogdalenverket, waste from Stockholm households is burned to produce heat and power for the Stockholm area. Hogdalenverket has been instructed by the Swedish National Environment Protection Board to reduce it's emissions of nitrogen oxides (NO x ). One way to reduce NO x is by selective non catalytic reduction, which can be obtained by injecting ammonia (NH 3 ) into the combustion chamber. The achieved reduction is not always enough to meet the given regulations, but can be improved by optimizing the amount of ammonia. The optimal amount is affected by factors such as the load of the boiler, variations in the waste etc. In order to find optimal amounts of ammonia, Hogdalenverket has made a series of tests where a number of parameters were measured. The amount of data has proven to be too large for manual analysis and machine learning has therefore been applied in order to automatically produce a set of control rules. Experiments show that accurate control rules can be learned in thi..