Power Systems Optimisation and Management by Distributed Methods

Abstract

The research in this thesis mainly focuses on the power system management and optimisation by distributed methods. The traditional power system management and optimisation approaches are based on centralised fashions. However, it is difficult or impossible to establish a single centralised control centre in modern distributed power systems. As the amount of data in the power system continues to increase, centralised algorithms are limited by insufficient computing power. To solve the problems mentioned above, we will look into the distributed optimisation problem in power system. Firstly, to deal with the initialisation error and local physical constraints, a distributed demand side management strategy is studied by maximising the total welfare from demand side to supply side. To achieve the objectives of demand side management, controllable power units generate their optimal power reference based on the proposed distributed algorithm by coordinating information with the neighbours. Thus, it is a completely distributed algorithm, and the analysis denotes that the proposed algorithm can solve the optimal economic dispatch problem in an initialisation-free approach and adapt to the plug-and-play operation. Secondly, a uncertain power market environment is considered in this thesis. To maximise the benefit of a battery energy storage system in power markets, a novel reinforcement learning based optimal bidding strategy is investigated. The objective is to extend the battery life and maximise the benefit for battery energy storage system owners. Using this strategy, the battery energy storage system can make different decisions based on various environmental states. Meanwhile, the proposed bidding strategy overcomes the discrete limits by the function approximation approach. Finally, from the above researches, it is noticed that the data volume and customer privacy may influence the reliability of the power systems. To avoid the computing burden and consider the data privacy. A novel Markovian switching based distributed deep belief networks for short term load forecasting is introduced. The proposed algorithm trains the short-term load forecasting model with local datasets and update the model parameters by communicating with connected neighbours, which does not need to transfer any information about the load data. In addition, it can simultaneously achieve fast training speed and superior robustness to cyberattacks due to the Markovian switching structure and the unsupervised training process of the deep belief networks.

Date of Award	1 Aug 2021
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Zhirun Hu (Supervisor) & Zhengtao Ding (Supervisor)