Abstract:Real-world chemical processes often exhibit stochastic dynamics with non-trivial correlations and state-dependent fluctuations. However, most process models simply add stationary noise terms to a deterministic prediction, which can lead to inaccurate predictions. This work proposes using conditional normalizing flows as discrete-time models (DTMs) to learn the stochastic dynamics of chemical processes. Normalizing flows learn an explicit expression of the system states' probability density function (PDF) given prior states and control inputs. The resulting model naturally allows for formulating stochastic and probabilistic setpoint-tracking objectives and chance constraints. In applications to a continuous reactor and a reactor cascade, the normalizing flow yields stable simulations over long time horizons and high-quality results in stochastic and probabilistic MPC formulation for open-loop control. Furthermore, a chance-constrained optimization finds reliable startup controls for the reactor cascade with stochastic reactions. In conclusion, the conditional normalizing flow presents an excellent choice for modeling nonlinear stochastic dynamics.
What problem does this paper attempt to address?
### What problems does this paper attempt to solve?
This paper aims to solve the problems of complex stochastic dynamics modeling and optimization in chemical processes. Specifically, traditional discrete - time models (DTMs) usually describe system behavior by simply adding a static noise term to deterministic predictions, and this method may lead to inaccurate predictions. To solve this problem, the authors propose using conditional normalizing flows as discrete - time models (DTMs) to learn and simulate stochastic dynamics in chemical processes.
#### Main problems include:
1. **Limitations of traditional models**:
- Most existing DTMs can only predict the most likely outcome of the system state, ignoring the randomness and uncertainty of the system.
- These models usually assume that the noise is additive, zero - mean, and independent of the system state, which fails to capture complex non - Gaussian distributions and state - dependent randomness.
2. **Need for more accurate stochastic dynamics modeling tools**:
- Chemical processes in the real world often exhibit non - trivial correlations and state - dependent fluctuations, so more flexible modeling tools are needed to describe these complex stochastic behaviors.
- Conditional normalizing flows can learn the probability density function (PDF) of the system state, thus providing a natural way to handle randomness and probability constraints.
3. **Application requirements**:
- In practical application scenarios such as continuous reactors and reactor cascades, it is necessary to be able to perform stable simulations over a long time range and provide high - quality stochastic and probabilistic MPC (model - predictive control) results in open - loop control.
- It is necessary to find reliable start - up control strategies in the presence of stochastic reactions.
#### Solutions:
- **Using conditional normalizing flows**: By using conditional normalizing flows to learn the stochastic dynamics of the system, non - linear, non - Gaussian, and state - dependent stochastic behaviors can be better captured.
- **Probabilistic set - point tracking objectives**: Using the conditionally PDF explicitly expressed by the normalizing flow, probabilistic set - point tracking objectives and chance constraints can be formulated, thus achieving more precise control.
- **Long - term stability**: In the applications of continuous reactors and reactor cascades, the normalizing flow model has demonstrated stable simulation performance over a long time range.
In conclusion, by introducing conditional normalizing flows as DTMs, this paper solves the deficiencies of existing models in describing complex stochastic dynamics and provides a new method for the modeling and optimization of chemical processes.