Micro-grid Energy Management Methodologies Based On Adaptive Dynamic Programming

As an important part of smart grid,Micro-grid Energy Management System(MEMS)is of great significance in achieving supply-demand balance,efficient utilization of renewable energy,and economical operation of power grids.Due to the diversification of loads and the high proportion of renewable energy generation,MEMS are characterized by complexity,time variability and uncertainty,which results in difficulty in establishing an accurate mechanism model and solving it in real time.To solve these problems,this thesis focuses on studying several issues about modelling and optimizations in MEMS.The main contributions of this thesis are as follows.(1)Aiming at the difficulty in accurately establishing the model and solving the optimal scheduling problem caused by the complex and diverse load data of MEMS in the park,the research on modeling and algorithm is carried out.Firstly,a multi-neural network fusion algorithm is proposed to predict the electricity consumption of different types of users combined with different influencing factors.Then,in order to solve the dynamic power scheduling problem in the MEMS,an adaptive dynamic programming(ADP)micro-grid energy management method based on the multi-neural network fusion prediction algorithm is employed to optimize the charging and discharging power of the energy storage system.The optimization goal of the MEMS is to reduce the life loss and power of the micro-grid energy storage system.Finally,the convergence of the proposed algorithm is proved mathematically when the initial value is arbitrary semi-positive definite function,and the effectiveness is verified via the hardware-in-the-loop(HIL)platform.(2)Aiming to solve the real-time problem of building MEMS in the park,the real-time optimization algorithm is further studied,and a kernel based ADP energy management method is proposed.Firstly,to solve the uncertainty of renewable energy power generation and load power consumption in the building micro-grid,the long-short-term-memory neural network is used to predict their generation and consumption power respectively.Secondly,in view of the problem that the actual approximation ability of the neural network in the traditional ADP algorithm is affected by many factors such as the learning step size,the number of hidden layers,and the initial weight,the kernel method is employed to construct the performance index function of ADP to avoid these problems.Then,a nuclear ADP micro-grid energy management method based on prediction technology is proposed,and an approximate optimal control strategy is obtained to realize orderly and real-time charging and discharging scheduling of electric vehicles in buildings,thereby ensuring economical and low-carbon operation of building micro-grids.Finally,the functional verification of building MEMS is conducted on the HIL platform.(3)Aiming at the problem of complex and diverse load modeling in residential MEMS in buildings,the refined modeling of load is further studied,and a prediction algorithm of Gate Recurrent Unit-Bidirectional Encoder Representations from Transformers(GRU-BERT)is developed.First,in view of the fact that the traditional forecasting methods are only related to the historical power data,the residential load is divided into meteorological sensitive load,human activity sensitive load and non-sensitive load.Second,the electricity consumption and renewable energy generation of different types of loads are predicted by using the GRU-BERT prediction algorithm combined with meteorological temperature,humidity,radiation intensity and human activities.Third,the charging and discharging power of the energy storage system in the residential micro-grid is optimized by using the energy management method based on ADP.The optimization goal of residential MEMS is to improve the utilization rate of renewable energy and reduce the electricity cost of residential users.Finally,the functional verification of residential MEMS is implemented on a HIL platform.Finally,a conclusion is drawn and the future research is prospected.