An agile and intelligent dynamic economic emission dispatcher based on multi-objective proximal policy optimization

The recent development in power systems claims the improvement of the efficiency to solve dynamic economic emission dispatch (DEED) problem. However, the popular pure optimization framework based on evolutionary algorithms only starts the compute-intensive optimization process after receiving the power demand, which causes significant delays To address this limitation, this research defines a novel dynamic economic emission dispatcher (DEEDer) problem and the dispatcher learning framework. It is different from previous research and the most current algorithms in the advantage of transferring the on-line compute-intensive optimization task to off-line. To solve the dispatcher, we model the dispatching process as a conditional deterministic Markov Decision Process (MDP) and propose the multi-objective proximal policy optimization (MOPPO). The Benchmark Test Set with 10,000 different dispatching tasks for 5-unit and 10-unit system is released to evaluate the generalization of the dispatcher. The experiment results indicate that the neural network dispatcher trained with MOPPO is hundreds to thousands of times faster than the state-of-the-art pure optimization algorithms. Meanwhile, the dispatcher not only has comparable performance as the state-of-the-art multi-objective optimization algorithms in standard dispatching task but also shows generalization on generating Pareto optimal solutions given any possible dispatching task. Following the DEEDer framework, the proposed method makes it possible to dispatch power in a more agile way as the timely response to the changing of power demand while still controlling economic and emission. (C) 2020 Elsevier B.V. All rights reserved.

Automated summary

The research defines a novel dynamic economic emission dispatcher problem and learning framework, transferring the optimization tasks offline and using multi-objective proximal policy optimization to significantly improve the speed and performance of the neural network dispatcher, showcasing its generalization capabilities.