Research On Identification And Correction Methods Of Bad Data In Synchrophasor Measurement Systems

Fast and accurate electrical measurements in dynamic processes are critical to the safety of increasingly complex power systems.Synchronous phasor measurement units(PMUs),as one of the most effective dynamic measurement tools,can provide data support for power system monitoring and control.However,at present,PMUs on the site have different levels of data quality issues.Reports suggest that around 10%to 17%of PMU data in North America has a certain degree of problems,while the figures increase to as high as 20%to 30%in China.The poor quality of the data severely restricts the performance and prospect of PMU in the dynamic safety monitoring of power systems.Based on the characteristics of different types of PMU bad data,the dissertation proposes several methods to detect and correct bad data efficiently and accurately.All those approaches provide strong support to enhance the quality of PMU measurement data.The dissertation analyzes the characteristics of various types of PMU bad data and divides them into occasional bad data and continuous bad data.The PMU bad data detection and correction framework is proposed:for the occasional bad data with high occurrence probability and relatively easy to handle,it is detected and corrected based on a single PMU to improve the calculation efficiency;for the continuous bad data with low occurrence probability and difficult to handle,it is detected and corrected based on multiple PMUs to improve the calculation accuracy.Furthermore,an initial screening method for occasional bad data based on slope characteristics and a steady and dynamic data identification method based on decision trees are proposed.They can preprocess and classify bad PMU data,and provide a basis for the subsequent detection and correction.The initial screening accuracy of the proposed method is about 99.4%,and the identification accuracy is about 99.1%.Occasional bad data under steady conditions is usually hard to distinguish from the normal ones as the deviation is relatively smaller.To solve this issue,a spectral clustering-based method is proposed which can achieve high accuracy of bad data detection with deviation greater than 0.5%.This is realized by firstly mapping PMU data from low-dimensional space to high-dimensional space,amplifying the deviation characteristics of bad data,and then clustering.Moreover,the correction method based on the improved cubic spline interpolation function is proposed.It can automatically assign priorities due to the relative positions of the bad data,and then apply the non-linear interpolation.So the average comprehensive vector error of the corrected data is reduced to 0.3%.On the other hand,occasional bad data under dynamic conditions of power systems are also hard to detect due to the normal fluctuations in the data.Thus,the thesis proposes a double-layer network model based on long and short-term memory(LSTM).The LSTM network manages to learn the dynamic data patterns and detect the bad data from their damaging effect towards the future trending of other data.The bad data with a deviation greater than 0.5%under dynamic conditions will be detected.Furthermore,a component decomposition method for data under dynamic conditons based on singular value decomposition is proposed for the correction.The method will reconstruct the original data precisely so that the error rate of occasional bad data will below 0.05%after the correction.Finally,for PMU continuous bad data,which is usually in large quantities and with strong continuity,a similarity calculation method of multiple PMUs based on dynamic time warping theory on the spatial scale is proposed.With the improved spectral clustering algorithm,the PMU data spatial grouping can be achieved.Furthermore,on the time scale,a two-way LSTM is established,which maps the relationship between the measurements of multiple PMUs to the measurement of a single PMU,so that the former can be used to predict the latter.The predicted data will be compared to the measurement with certain thresholds to identify the quality of the data and implement a reliable correction at the same time.The proposed method can effectively detect continuous bad data with a deviation greater than 0.5%,and the average error of the corrected data is about 0.02%.