Development of predictive NOx model for on-road heavy-duty diesel engines.

摘要

Emissions models currently employed by EPA and CARB do not account for the variations in engine operation and their effect on emissions. Alternatively, this study, demonstrates the feasibility of using Engine Control Module (ECM) broadcast parameters such as Engine Speed, Engine Torque, Injection Timing, Fueling Rate, Manifold Air Temperature, Manifold Air Pressure, Coolant Temperature and Oil Temperature as inputs to in order to predict engine-out exhaust NO x emissions. These parameters were obtained when the engine operates in the Not-to-Exceed (NTE) zone, (which is defined by 40 CFR §86.1370-2007) for a continuous time period of at least 30s in length.;This study taps into the in-use emissions measurement capabilities and the vast databases that reside at the National Research Center for Alternative Fuels Engine and Emissions (CAFEE), and combines them with an advanced statistical modeling technique called Multivariate Adaptive Regression Splines (MARS) to predict NOx emissions. The MARS technique is an adaptive piece-wise regression approach that can be configured to fit models with terms that represent nonlinear effects and interactions among input variables.;In this study, an on-board portable emissions measurement system called the Mobile Emissions Measurement System (MEMS), developed at West Virginia University (WVU) was used to record in-use, continuous NOx emissions along with ECM broadcast parameters from 60 heavy-duty diesel-powered vehicles from model years 2001, 2002 and 2003. The vehicles were classified according to their engine model and model year and four vehicles were tested for each category. The vehicles were tested over different routes which included a mix of urban and highway driving conditions.;Data collected from the on-road tests of a vehicle(s) were combined to form the calibration and validation datasets. 'Calibration' dataset was used to create a predictive model using MARS. Validation datasets which were independent of the 'calibration' datasets were used to check the accuracy of the model predictions. Results indicate that the predictive models developed proved highly successful with the range of uncertainty in predictions within +/- 20% of the actual value.