Design, development and testing a hybrid control model for RCCI engine using double Wiebe function and random forest machine learning. (August 2021)

• Record Type: Journal Article
• Authors: Mishra, Chinmaya; Subbarao, P.M.V.
• Is Part Of: Control engineering practice. Volume 113(2021)
• Publication Date: 2021-08
• Subjects: Combustion metrics -- Cyclic variations -- Double Wiebe function -- Hybrid control model -- RCCI engine -- Random forest machine learning; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2021.104857
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: Reactivity controlled compression ignition (RCCI) engine technology significantly reduces emissions of NO x and soot while improving thermal efficiency. RCCI engine development requires optimum in-cylinder combustion metrics as desired objective across different operating conditions. Physics-inspired control models are a cost-competitive option to achieve these objectives in engine control systems. Combining these control models with machine learning can offer computational efficiency, higher accuracy and lower number of experiments needed for control system calibration. In this context, the present paper proposed a novel hybrid control model by combining physics-inspired parametrized double Wiebe function (D-W) and Random Forest Machine Learning (RFML) to predict both the average and cyclic variation trends of combustion metrics in an RCCI engine. In this work, firstly, an optimized D-W parameters matrix was formulated that can stochastically reconstruct pressure cycles in agreement with 99.7% of the experimental confidence interval. RFML was then employed to develop correlative learning models between experimental control variables such as premix ratio, engine load, injection timing etc. with corresponding D-W parameters matrix. The hybrid control model was used to predict peak pressure (PP), indicated mean effective pressure (IMEP), crank angle for 50% of fuel mass fraction burnt ( θ 50 ) and maximum pressure rise rate (MPRR) across 60 test cases and 36000 cycles. Results showed an error mean and standard deviation ( μ err ± σ err ) of 0.068 ± 1.840 bar, 0.007 ± 0.264 bar, 0.15 ± 0.98° CA and 0.018 ± 0.221 bar/° CA between predictions and experiment for PP, IMEP, θ 50 and MPRR, respectively in the tenable RCCI operation regime. Proposed hybrid control model exhibited good accuracy and was able to capture the local cyclic variation trends. Graphical abstract: Highlights: Stochastic computational reconstruction of RCCI combustion pressure profile using double Wiebe function. Development of correlative machine learning models between control variables and optimized double Wiebe parameters. Hybrid control model synthesis by integrating machine learning models with double Wiebe function. Cycle to cycle prediction and statistical study.