Research On Data-driven-based State Prediction Technology Of Electric Energy System

Lithium-ion batteries(LIBs),as a representative of the latest generation of green electricity energy system,have revolutionized energy storage technology and promoted the progress of science and technology greatly.It is not only deeply involved in all aspects of our daily life and work,but also an indispensable energy source for national defense economy and military equipment.The specific demand of electric equipment determines specific requirements for LIBs,such as high safety,high reliability,high environmental adaptability and so on.State of Charge(So C)and Remaining Useful Lifetime(RUL),as the two major state characterization indexes of lithium-ion battery system,can provide quantitative basis for the performance of LIBs.However,it is difficult to predict So C and RUL accurately and reliably,due to the complex electrochemical reactions and degradation mechanism during charging and discharging.The data-driven methods,based on a large amount of training data,can mine hidden feature information and state variables by machine learning method.A model is directly modeled by using the data-driven method and the state variation trend is learned without considering the internal working mechanism on LIBs.Based on the above considerations,this thesis focuses on the So C estimation and RUL prediction method research based on data-driven for LIBs,and mainly completes the following work:1.A capacity prediction algorithm is proposed based on long short-term memory recurrent neural network(LSTM RNN).The battery data exists in the form of time-series,and So C and RUL can be characterized by current capacity and the maximum discharge capacity,respectively.Therefore,both of them can be seen as a regression problem,thus a kind of capacity prediction algorithm,which selects LSTM RNN as the basic network,is proposed to predict lithium-ion battery system state.2.So C estimation and RUL prediction for LIBs are conducted based on the capacity model.Features used to drive the prediction model are selected based on the analysis of the relationship between charging/discharging parameters and state degradation.A sliding window technology is used to cut out the input time series length of the model based on the characteristics of long time series and local information of the charging-discharging data.With a variable output sequence,it is possible to evaluate multistep ahead prediction.A cycling prediction technology is proposed to implement an indirect RUL prediction.Extensive experimental results prove the accuracy of the proposed So C estimation method and RUL prediction method.3.Development of a lithium-ion battery status estimation platform.According to the working characteristics and scenarios of LIBs,the operating condition of LIBs in electric vehicle scenarios is simulated and designed.Based on the research results of So C and RUL,an embedded state prediction module is designed and an online monitoring platform is developed,which realize real-time So C estimation and RUL prediction effectively.