Medium Voltage Load Soft Sensing And Identification Algorithms Of Auxiliary Power Supply Systems On Emus

This paper focuses on the medium voltage load state monitoring of Electrical Multiple Unit(EMU)Auxiliary Power Supply Systems(APSS),and carries out load identification based on soft sensing method.The purpose is to use the aggregate load data to achieve operating state identification and energy consumption disaggregation of multiple appliances.The main contents of this paper are described as follows:(1)The mathematical description of load identification of EMU APSS is constructed based on its topological structure.The characteristics of two load identification application scenarios,car-level load identification and train-level load identification,are analyzed.A load identification strategy based on source signal modeling,mixed process modeling and separation matrix inference is proposed in the framework of blind source separation.The measured data collection method of EMU APSS load is introduced.The measured data for train-level load identification experiment are faced with the problem of packet loss and poor synchronization.A simulation method based on operation mechanism modeling and measured data correction is proposed to realize high-precision simulation of the EMU APSS loads.The average percentage errors between the simulated data and the measured data are less than 7%,and the Pearson correlation coefficients are more than 0.90.(2)To evaluate the differences of the APSS loads,load identification experiments based on transient features are carried out under the assumption that the loads operate asynchronously.The differences of the APSS load are evaluated by load identification performance.Firstly,in the shallow machine learning framework,an integrated model based on data down sampling is proposed to avoid the influence of sample imbalance on the identification results.Under the proposed shallow integration framework,the load identification performance of different forms of transient features is studied,and the transient time series are finally selected as the transient feature.Then,a hybrid deep learning model based on multivariate fusion,time series data augmentation and deep neural network calculation is proposed.Compared with the Dynamic Time Warping(DTW)template matching model which is famous for transient time series processing capability,the proposed hybrid deep learning model has the same accuracy level and has advantage in test efficiency.The results show that the overall recognition accuracies of the proposed model for loads over different cars are greater than 0.98,and the Macro-F1 scores are greater than 0.94.The results show that the fingerprints of medium voltage load on the EMU auxiliary power supply system are different greatly.It is feasible to carry out load identification under the condition of simultaneous operation of multiple loads.(3)To achieve car-level APSS load identification,a hybrid MDLDTW-SDL-PKR model based on piecewise Sparse Dictionary Learning(SDL)and prior knowledge Revise(PKR)is proposed.The piecewise SDL is applied to overcome the difficulty of extracting overall load features caused by frequent APSS loads start and stop.The change point detection algorithm based on the Minimum Description Length(MDL)is used to segmentate the steady process and the transient process of the APSS load.The SDL models of the steady process and the transient process are established respectively.The atoms of each load are obtained by clustering method for steady process.For the transient process,DTW matching method is used to determine the best matched aggregate load,and the single load transient component at the corresponding position is defined as the transient process atom.In the unsupervised sparse coding matrix optimization stage,the steady state process sparse coding and transient process phase coding are realized by minimizing the decomposition error and maximizing the sparsity degree of the coding matrix.The steady state process coding and transient process coding are carried out alternately supported by each other.In order to overcome the problem that the load magnitude varies greatly and the small power loads are submerged by the large ones,a prior knowledge constraint mechanism is proposed.The PKR module is proposed to make full use of the dependency relationship between different loads in the APSS,and the load disaggregation accuracy is improved through synchronous and mutually exclusive constraints.The effectiveness of the proposed model is verified by case study based on measured data.The performance of the model is significantly better than that of the benchmark model,and the overall M-F1 score and decomposition accuracy over different cars are greater than 0.86.(4)To achieve train-level APSS load identification,a load identification model consisting of association rule model and sparse Hidden Markov Model(HMM)is proposed.The possible load state combinations of the whole APSS are huge due to the variety and quantity of loads,and the traditional load disaggregation methods fail due to the large computational complexity.To overcome this problem,the strategy of hierarchical disaggregation is proposed.In the first level of disaggregation,the operation principles of several APSS loads are used to mining association rules.The relationship between speed and power as well as the relationship between voltage and power are established.In this way,the several appliances including traction transformer fan,traction converter fan,traction motor fan,traction transformer oil pump,and traction converter water pump are identified and disaggregated.Then the identified loads are removed from the aggregate load,and the HMM is used for second level disaggregation.In order to describe the operation state of the APSS,a state quantification method considering both the peak value and the valley value of the power distribution is proposed.On this basis,the sparsity of state transition matrix and transmission matrix of EMU APSS is fully utilized,and the sparse Viterbi algorithm is used to realize efficient decoding of the residual load.The simulation case study based on 9 actual trains shows that the difference between the predicted energy consumption proportion and the actual energy consumption proportion for all involved loads is within ±1%.The proposed method can achieve the efficient disaggregation of APSS load on the premise of ensuring accuracy.