Topics in Catalysis Vols. 16/17, Nos. 1–4, 2001 337 Inferential NO x emission prediction in feedforward reductant control for SCR H.C. Krijnsen * , J.C.M. van Leeuwen, R. Bakker, H.P.A. Calis and C.M. van den Bleek Faculty of Applied Sciences, DelftChemTech, Delft University of Technology, Julianalaan 136, 2628 BL Delft, The Netherlands E-mail: h.c.krijnsen@tnw.tudelft.nl To adequately control the reductant flow for the selective catalytic reduction of NO x in diesel exhaust gas a tool is required that is capable of accurately and quickly predicting NO x emissions from the engine’s operating variables. Two algorithms for non-linear modelling are evaluated: neural networks (Solla et al., Adv. in Neural Information Processing Systems 12 (MIT Press, Five Cambridge Center, Cambridge, MA, 2000)) and the split & fit algorithm (Bakker et al., submitted for publication to NIPS). Measurements were carried out on a transient automotive diesel engine and a semi-stationary diesel engine. Both algorithms gave excellent predictions with a short computation time (0.03–0.13 ms). This makes them very promising tools in automotive catalytic NO x emission control. KEY WORDS: neural network; split & fit algorithm; diesel exhaust gas; nitrogen oxides; deNOxing; SCR 1. Introduction One of the potential solutions for NO x emission abate- ment is selective catalytic reduction (SCR). SCR involves adding a reductant to the exhaust gas for catalytic NO x re- moval. The required reductant flow depends on the exhaust flow rate, the NO x concentration, the required NO x reduc- tion and the catalyst conditions. Using on-line NO x analysis equipment in a reductant dosage control system is not an op- tion both from economical and practical point of view, as it is expensive and susceptible to soot plugging. Further- more, NO x analysers have to be frequently calibrated and serviced. On top of this the dead-time between exhaust gas sampling and NO x analysis is in the order of seconds, which will cause an intolerable time lag for the reductant dosage control action. Therefore, an alternative for NO x emission measurement is desired [1]. The aim of this paper is to show that it is possible to predict the NO x emission from the diesel engine’s operating variables. For this so-called “soft- sensoring” or inferential measurement, application of a non- linear black-box modelling algorithm is needed that requires nothing but a large set of measurements to learn how to pre- dict the NO x emissions from the operating variables. Two black-box algorithms are evaluated in this paper: (1) An artificial neural network (ANN); (2) The split & fit algorithm of Bakker et al. [2]. The split & fit algorithm (s&f) splits the input data into a number of disjoint regions, each of which has its own local linear model. The results of these algorithms will be compared to the com- monly used linear fit and engine map. This approach will be followed for both a transient and semi-stationary diesel en- gine. * To whom correspondence should be addressed. It is not the aim of this paper to explain in detail the principles and mathematical details of the applied black box models, because (1) this is not relevant for the purpose of demonstrating that black box models can be applied to SCR, and (2) various adequate test books are available to provide the reader with principles and details on black box mod- elling. 2. Experimental set-up Nitrogen oxide measurements were performed at TNO Automotive, using a standard 6 cylinder, 12 l heavy-duty diesel engine and a transient engine dynamometer. The en- gine used was a direct injection turbocharged intercooled engine with a maximum engine power of about 300 kW at 2000 rpm. Measurements (NO x emission, intake air temper- ature, pressure, engine load and engine speed) were done at 1 Hz using the European transient test cycle (ETC). During the ETC cycle, city, rural and highway conditions are mim- icked. The data of this transiently operating engine are used to estimate whether or not the black-box algorithms are ca- pable of NO x emission prediction from such a typical tran- siently operating engine. Furthermore, engine measurements were performed on a semi-stationary LPW3 diesel engine at our laboratory. The LPW3 is a three-cylinder water-cooled naturally aspirated diesel engine fitted in a Wilson LD 12.5/W4 generator with a rated power of 8.0 kW and a constant engine speed of 1500 rpm. The fuel used during all the experiments was a summer quality diesel fuel containing 0.04 wt% sulphur. The LPW3 engine was used both to validate the applicabil- ity of NO x emission prediction models and to validate the inferential NO x estimation in time. In order to keep the inlet conditions of the combustion air constant during the experi- ments, dried air was used for deNOx experiments; to assess the effect of intake air temperature and humidity, these vari- 1022-5528/01/0900-0337$19.50/0  2001 Plenum Publishing Corporation