Time Series Generative Adversarial Network Controller for Long-Term Smart Generation Control of Microgrids

Linfei Yin,Bin Zhang
DOI: https://doi.org/10.1016/j.apenergy.2020.116069
IF: 11.2
2020-01-01
Applied Energy
Abstract:The conventional combined generation control framework of microgrids, which contains two time-scales, i.e., the time slot of economic dispatch is set to 15 min; and the total time slot of smart generation control and generation command dispatch is set to 4 s, could lead to uncoordinated problems. To avoid uncoordinated problems, this paper proposes a long-term smart generation control framework with a single time-scale to replace the conventional combined generation control framework with two time-scales, and then proposes time series generative adversarial network controller for long-term smart generation control of microgrids. The proposed time series generative adversarial network controller contains reinforcement learning, generator deep neural networks, and discriminator deep neural networks. The generator deep neural networks generate predicted states from multiple historical states, multiple historical actions, and multiple long-term actions. The discriminator deep neural networks judge whether the data from the generator deep neural networks or real-life data. This paper compares the proposed controller with conventional optimization algorithms and control algorithms, which are applied for economic dispatch, smart generation control, and generation commands dispatch in microgrids. The numerical simulation results under Hainan Power Grid, IEEE 300-bus power system, and IEEE 1951-bus power system verify that the proposed time series generative adversarial network controller can simultaneously obtain higher control performance and smaller economic cost than conventional combined control algorithm and optimization algorithms in the long-term. Consequently, the uncoordinated problem of economic dispatch, smart generation control, and generation commands dispatch can be solved by the proposed approach with one single long-term time-scale.
What problem does this paper attempt to address?