Real-Time Risk Analysis with Optimization Proxies

Wenbo Chen,Mathieu Tanneau,Pascal Van Hentenryck

2023-10-05

Abstract:The increasing penetration of renewable generation and distributed energy resources requires new operating practices for power systems, wherein risk is explicitly quantified and managed. However, traditional risk-assessment frameworks are not fast enough for real-time operations, because they require numerous simulations, each of which requires solving multiple economic dispatch problems sequentially. The paper addresses this computational challenge by proposing proxy-based risk assessment, wherein optimization proxies are trained to learn the input-to-output mapping of an economic dispatch optimization solver. Once trained, the proxies make predictions in milliseconds, thereby enabling real-time risk assessment. The paper leverages self-supervised learning and end-to-end-feasible architecture to achieve high-quality sequential predictions. Numerical experiments on large systems demonstrate the scalability and accuracy of the proposed approach.

Optimization and Control

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the computational efficiency of traditional risk assessment frameworks in real - time operations of power systems. Specifically, with the increasing penetration of renewable energy sources (such as wind and solar energy) and distributed energy resources (DERs), the operation of power systems requires new methods to explicitly quantify and manage risks. However, traditional risk assessment methods rely on a large number of Monte Carlo (MC) simulations, and each simulation needs to solve multiple economic dispatch problems in turn, which makes these methods difficult to implement in real - time operations.

To address this computational challenge, the paper proposes a risk assessment method based on surrogate models, namely **Risk Assessment with End - to - End Learning and Repair (RA - E2ELR)**. This method uses a fast optimization surrogate model instead of the traditional optimization solver, so it can complete the prediction within milliseconds and achieve real - time risk assessment. Through self - supervised learning and an end - to - end feasible architecture, RA - E2ELR can efficiently handle large - scale power systems and provide high - precision risk assessment results.

### Summary of Main Problems

1. **Computational Bottleneck of Traditional Risk Assessment Methods**:
- Traditional methods rely on a large number of MC simulations, and each simulation needs to solve multiple economic dispatch problems.
- The computational cost of this method is high and cannot meet the requirements of real - time operations.

2. **Requirement for Real - Time Risk Assessment**:
- With the increase of renewable energy and distributed energy resources, power systems need more efficient real - time risk assessment methods.
- New methods must be able to complete a large number of simulations in a short time to adapt to the rapidly changing power demand and supply situations.

3. **Proposed Solution**:
- Use optimization surrogates instead of traditional optimization solvers.
- Through self - supervised learning and an end - to - end feasible architecture, ensure the prediction speed and accuracy of the surrogate model.
- Achieve a prediction time at the millisecond level to support real - time risk assessment.

### Formula Examples

The formulas involved in the paper include:

- **Objective Function of the Economic Dispatch Model**:
\[
\min_{p_t} c^\top p_t+M_{\text{th}}\|\xi_{\text{th}}\|_1
\]
where \(c\) is the generation cost vector, \(p_t\) is the generation dispatch, \(M_{\text{th}}\) is the thermal violation penalty coefficient, and \(\xi_{\text{th}}\) is the thermal violation variable.

- **Conditional Value at Risk (CVaR)**:
\[
\text{CE}(Q_t)=\mathbb{E}[Q_t\mid Q_t\geq Q_t(\alpha)]\approx\frac{\sum_{s\in S}\hat{Q}_s^t\cdot1_{\hat{Q}_s^t\geq\hat{Q}_t(\alpha)}}{\sum_{s\in S}1_{\hat{Q}_s^t\geq\hat{Q}_t(\alpha)}}
\]
where \(\alpha\) is the given significance level, and \(\hat{Q}_s^t\) is the predicted value of scenario \(s\) at time \(t\).

Through these improvements, the paper demonstrates the scalability and accuracy of the RA - E2ELR method in large - scale power systems, providing an effective tool for real - time risk management in power systems.