This is the first project of my Ph.D. career. It aims to sell a closed-loop predict-then-optimize (CPO) idea.

To achieve the lowest operating cost, power system operations generally use an open-loop predict-then-optimize (OPO) process:

  1. Predict uncertainties, e.g., wind power, as accurately as possible.
  2. Optimize the operation plan, e.g., unit commitment, using the predictions.

Alt text

Open-loop predict-then-optimize.

Alt text

Closed-loop predict-then-optimize.

The OPO process does make sense. Yet, due to the nonlinearity and complexity of power systems, the relationship between prediction accuracy and operating cost is non-monotonic and asymmetric. This implies that a more accurate prediction may NOT lead to a lower-cost operating plan.

Alt text

Testing results on the IEEE 118-bus system. Each dot represents a specific testing sample with a prediction error and the corresponding increase in operating cost, while the red line indicates the overall trend. As observed, Point A has a smaller mean absolute percentage error (MAPE), but it leads to a higher operating cost compared to Point B.

To this end, we are selling you the CPO idea, which features that:

  1. It feeds the operating cost back to the prediction phase.
  2. The predictor training evaluates prediction quality via operating costs rather than statistical accuracy criterion (e.g., MAPE).

Related Papers

  • [1] Xianbang Chen, Yafei Yang, Yikui Liu, Lei Wu. "Feature-driven economic improvement for network-constrained unit commitment: A closed-loop predict-and-optimize framework," IEEE Transactions on Power Systems, 2021. [PDF »] [Code »]
  • [2] Xianbang Chen, Yikui Liu, Lei Wu. "Towards improving unit commitment economics: An add-on tailor for renewable energy and reserve predictions," IEEE Transactions on Sustainable Energy, 2024. [PDF »]