完美拟合与虚幻最优：数据驱动模型如何在实时优化中失效
文章背景与核心概要
实时优化（RTO）高度依赖准确的工艺模型来确定最佳操作条件。尽管现代机器学习模型和数据-driven方法可以轻松拟合历史工厂数据并通过标准验证测试，但本文证明了良好的预测拟合并不能保证可靠的经济性能。
通过醋酸乙烯酯单体（VAM）基准工艺，作者研究了结构化混合模型（将质量平衡/热力学与神经网络动力学相结合）以及完全数据-driven的神经常微分方程（Neural ODE）模型。研究结果揭示了两种主要的失效模式：第一，虚幻最优（Phantom Optima），即训练出的模型会产生大量虚假的最优解，而不是识别出真实工厂的唯一经济最优解；第二，优化器漂移（Optimizer Drift），即便是面对无噪声数据和理想初始化，随机梯度训练也可能会天然地向产生次优 RTO 解的权重漂移——这意味着所识别出的模型实际上是训练优化器自身的产物。
因此，作者提出建议：任何用于实时优化的数据-driven模型，在部署到实际工厂之前，都必须在面向决策的基准测试中进行严格测试，以确保成功恢复工厂最优解。
Summary
Real-time optimization (RTO) relies heavily on accurate process models to determine optimal operating conditions. While modern machine learning models and data-driven methods can easily fit historical plant data and pass standard validation tests, this paper demonstrates that a good predictive fit does not guarantee reliable economic performance.
Using a vinyl acetate monomer benchmark process, the authors investigate both structured hybrid models (combining mass balances/thermodynamics with neural network kinetics) and fully data-driven neural ODE models. Their findings reveal two major failure modes: 1. Phantom Optima: Instead of identifying the unique economic optimum of the actual plant, trained models generate numerous false ("phantom") optima. 2. Optimizer Drift: Even with noise-free data and ideal initializations, stochastic gradient training can inherently drift toward weights that produce subpar RTO solutions—meaning the identified model acts as an artifact of the training optimizer itself.
Consequently, the authors propose that any data-driven model intended for RTO should be rigorously tested on a decision-oriented benchmark to successfully recover plant optima before being deployed in real-world plants.