Abstract:

Batch processes are inherently more difficult to model than continuous processes due to their complex dynamics, non-steady-state operation and batch-to-batch variation, especially under new operation condition and with small samples.To improve the generalization performance of the global learning based modeling methods, such as least squares support vector regression (LSSVR), a local LSSVR (LLSSVR) was proposed for modeling batch processes.By combining the optimized parameters of LSSVR from previous batches off-line, an online LLSSVR built a model for each set of new data to be predicted in a new batch by considering only the neighbors of this query.The proposed LLSSVR algorithm was applied to the online prediction of biomass concentration in the penicillin fed-batch process.The simulations showed that LLSSVR could predict the biomass concentration online just from the second batch, which was more accurate and robust to batch-to-batch variation than LSSVR.

Key words: 局部最小二乘支持向量机回归, 在线建模, 间歇过程, 发酵

摘要：

当间歇生产切换于不同的工艺条件时，由于新工况下的样本一般很少，且批次间存在着不确定性（由于原材料波动或过程动态特性波动等），基于全局学习的建模方法（如最小二乘支持向量机回归，LSSVR）建立的模型泛化性能不强。将局部学习融入LSSVR中，提出一种局部LSSVR（local LSSVR, LLSSVR）的间歇过程在线建模方法。结合前一批次离线优化后的LSSVR参数，针对待预测新样本在线选择与之相关的近邻样本集并基于此进行建模。以建立青霉素发酵过程的菌体浓度为例，验证了LLSSVR算法能够从过程的第2个生产批次开始在线建立较准确的预报模型，较LSSVR有着更好的推广能力、适应性和鲁棒性。

关键词: 局部最小二乘支持向量机回归, 在线建模, 间歇过程, 发酵