Recent years have seen a variety of applications in machine learning that consider optimization as a tool for inference learning. Embedding optimization problems as optimization layers in deep neural networks can capture useful inductive bias, such as domain specific knowledge and priors. Our target is to connect the date-driven learning model and physics-driven optimization models in an end-to-end differentiable learning framework. Under this framework, the learning model will be more explainable and robust while the optimization model will be more time-efficient.

 

With the advent of the era of big data and artificial intelligence, data security has become an issue of increasing concern. The concept of federated learning has also emerged from data security: the essence of federated machine learning is a distributed machine learning framework, which makes full use of the data and computing power of multiple institutions to meet the requirements of user privacy protection and data security. Enable multiple parties to collaboratively build a general machine learning model, effectively solving the problem among multiple agents. This project aims to build a set of trustworthy federated learning algorithm framework and extend it to real-world applications, such as smart grids. Compared with the traditional federated learning algorithm in protecting data privacy, this federated learning framework will also take into account the robustness and reliability of the model. Many issues related to data security in the smart grid can be completed within this framework, such as distributed power flow optimization problems, large-scale electric vehicle charging and discharging coordination problems, and distributed demand response, etc.

Innovative breakthroughs in the new generation of smart grid and battery technology have promoted the vigorous development of electric vehicles (EVs). However, due to the large-scale popularity of EVs, the large-scale charging and discharging of EVs in smart grid may cause many potential effects, such as voltage fluctuations, load changes and power loss. There is relatively few work that can meet the needs of joint coordination between large-scale electric vehicle charging/discharging and optimal power flow (EV-OPF). Therefore, there is an urgent need to develop key technologies that can solve this problem. This project first establishes a dynamic optimization model for the large-scale EV-OPF problem and then designs a mixed-integer non-convex optimization algorithm to solve this problem. In order to tackle the uncertainty of EVs’ charging demand and time of arrival, a model predictive control model is developed. Taking into account the shortcomings of centralized algorithms in system robustness, computing and communication costs, and data privacy, this project aims to investigate non-convex distributed optimization algorithms suitable for this problem, and to study their convergence mechanisms.