MODELING AND OPTIMIZATION OF ACTIVE DISTRIBUTION NETWORK OPERATION BASED ON DEEP LEARNING

With the increasing penetration of renewable energy sources (RES), energy storages, and flexible loads, active distribution networks are becoming more sophisticated and are facing more uncertainties. Traditional model-based methods cannot fully satisfy the analysis and control requirements of active distribution networks due to a variety of reasons. For example, Traditional model-based methods for modeling power profiles of RES are difficult to accurately capture the probability distribution characteristics and volatility of power curves, because they need to artificially assume the probability density function of wind power curves. In addition, these methods are not universal, because the volatility and the probability distribution of power curves vary from region to region. As one of the data-driven approaches, deep learning technology shows state-of-the-art performance in many fields compared with other data-driven methods. It can directly learn from historical data without requiring any simplifications and assumptions of the system’s physical model. In recent years, many famous frameworks have been proposed in the field of image vision, such as recurrent neural networks, convolution neural networks, and generative adversarial networks. Although the advanced deep learning technology is beneficial for the integration of RES and flexible loads into active distribution networks, it cannot directly process the data from the active distribution network, which is quite different from the image data. Therefore, it is necessary to redesign the structure and readjust parameters of these deep neural networks according to the data characteristics of RES and flexible loads before using them.

The goal of this doctoral program is to develop and apply deep neural networks to improve the economy and stability of the active distribution network with RES, energy storages, and flexible loads. Specifically, intelligent technologies from the image restoration field will be used to fill the missing data of the power curve for RES and flexible loads, such as electrical vehicles, electric boilers, and heat pumps. Regression techniques are well suited for learning patterns from the data, and therefore are utilized to predict energy flexibility from deferrable loads and renewable energy generation such as wind and solar energy. In addition, advanced generative networks can be employed to capture the uncertainties of RES and flexible loads by generating a set of possible scenarios. Reinforcement learning can be used to provide optimal control actions to operators amid uncertainties of active distribution networks.

Funding: China Scholarship Council (CSC)