Adaptive local weighted kernel-based regression for online modeling of batch processes

Fed-batch processes are inherently more difficult to characterize than continuous processes due to the variations under different operation stages, drifting and small-sample condition. The classical kernel-based regression (KR) methods, e.g., least squares support vector regression (LSSVR), aim to achieve a universal generalization performance, which may fail in some local regions when applied to batch process modeling. Local LSSVR model which only uses the neighbors of the query instance helps improve the accuracy, but it generally leads to a heavy computation load. Inspired by the idea of universal and local learning simultaneously, an adaptive local weighted kernel-based regression (ALW-KR) method is proposed. That is, adaptive weights are assigned to corresponding samples based on the similarity measurement, followed by a recursive updating to obtain local models. This ALW-KR framework is applied to the prediction of biomass concentration in the penicillin fed-batch process. The experimental results show that the proposed ALW-KR model could predict the biomass concentration more accurate and robust to batch-to-batch variation than traditional KR methods.