PMU Data Reconstruction Method Based on Spatiotemporal Information Feature Fusion

Abstract

The present invention relates to a PMU data reconstruction method based on spatio-temporal information feature fusion. This method analyzes the synchronize phasor measurement data measured on-site, summarizes four cases of low-quality data, and uses mathematical analysis and correlation analysis methods to identify the operating state of the system; then combines the multi-view learning method with the power grid operation mechanism to obtain four view methods with spatio-temporal information features, and uses the four view methods to preliminarily predict the reconstruction values of low-quality target data; finally, trains the non-low-quality data of the target system, adaptively updates the weights of each prediction value through the overall error of each of the four algorithms in the transient and steady-state stages, and integrates the prediction results of the four view methods to obtain the final data reconstruction value. The method of the present invention can effectively perform real-time data reconstruction on dynamic systems, providing an effective guarantee for the application of synchronize phasor data in power systems.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system dispatching automation (including renewable energy), and in particular relates to a PMU data reconstruction method based on spatiotemporal information feature fusion. Background Art

[0002] In recent years, synchronized phasor measurement units (PMUs) have been widely deployed and applied in actual power grids. Compared with supervisory control and data acquisition (SCADA), they have the advantages of high accuracy, strict synchronization of the entire network, and short update cycle. They can directly measure the node status, providing an opportunity to achieve real-time global state estimation of the power system at the ten-millisecond level. Taking my country's power grid as an example, various power grid companies have deployed more than 4,000 sets of PMU devices, covering 500kV and above trunk transmission networks and important 220kV plants and lines, and are also gradually being used in distribution networks. Although the voltage and current phasors of the power system can be measured by PMUs, due to measurement quality issues, PMU measurement values ​​cannot be directly used as the real-time state of the power grid. The synchronized phasor measurement values ​​are affected by the voltage / current transformer (PT / CT) error and the PMU's own measurement, and there are certain measurement errors. In actual power grids, the field measurement environment of PMU data is often more complex and harsher than indoor simulation / test conditions. At the same time, the operation and maintenance of on-site PMUs are often unsatisfactory, resulting in large measurement errors and bad data in PMU measurements. Relevant test results show that some PMUs cannot accurately perform phasor calculations for complex signals containing dynamic information, which may produce large errors or bad data. The steady-state measurement accuracy test of on-site PMUs also found that the amplitude and phase angle errors may be as high as 2% and 1.5°, respectively, far exceeding the standard range specified by the national standard. Due to the lack of system-level functional applications based on PMUs, PMUs are not taken seriously in actual power grids and lack effective maintenance, resulting in insufficient accuracy, reliability and completeness of existing PMU measurement data. If the data quality problem cannot be improved, PMUs may be marginalized as advanced power grid "recorders" or even inaccurate recorders. Therefore, how to reconstruct low-quality data in real time online has become an important issue that needs to be solved in the power system.

[0003] After low-quality data is detected and identified, directly removing abnormal data is a relatively rare approach. Usually, data reconstruction is performed on it. For low-quality data in a steady state, interpolation methods are mainly used for data reconstruction. For example, some literature performs PMU data reconstruction based on Lagrange interpolation; some literature improves the cubic spline interpolation function and uses a priority allocation strategy with dynamically changing interpolation intervals to reconstruct different types of missing data. The above two methods only require single-channel device data. However, due to limited contained information, they are not applicable to transient situations and long-term continuous low-quality data reconstruction. For this reason, research on low-quality data reconstruction at the regional level has been proposed, and existing research methods mostly focus on the low-rank characteristics of matrices and machine learning directions. Some literature performs data reconstruction based on the low-rank characteristics of the PMU measurement matrix through the singular value threshold method in the low-rank matrix; some literature improves it and can perform real-time online data reconstruction on long-term continuous low-quality PMU measurements. However, the above methods cannot perform PMU data reconstruction when all-channel data is lost. In response to this, some literature uses a recovery method based on the alternating direction multiplier method to reconstruct all-channel PMU data, but the disadvantage is that the calculation time is relatively long and it is difficult to be used for real-time situation awareness. In addition, neural networks and ensemble learning in machine learning methods have also been widely applied to PMU measurement data reconstruction. Some literature uses gated recurrent unit neural networks for PMU measurement data reconstruction; some literature is based on the graph convolutional recurrent neural network method and mines the spatio-temporal correlation of data through training to recover and predict PMU measurements; some literature is based on the improved generative adversarial network method for PMU data reconstruction and realizes it by mining spatio-temporal information through training on regional measurement data. However, machine learning methods require a large amount of offline training and are prone to problems such as overfitting and insufficient generalization ability. In addition, some literature uses the robust principal component analysis method to achieve data reconstruction, but the interference data generated by abnormal events will be misjudged as low-quality data and thus be wrongly reconstructed; in response to this, some literature is based on the kernel principal component analysis method and avoids misjudgment by detecting perturbation events. Summary of the Invention

[0004] The object of the present invention is to provide a PMU data reconstruction method based on spatio-temporal information feature fusion in view of the increasingly prominent roles of the timeliness, correlation, and spatiality of power system measurement data in the research on low-quality data reconstruction. First, a multi-dimensional view learning method based on spatio-temporal information features is proposed to obtain preliminary predicted values from the global and local perspectives of space and time; further, four methods with spatio-temporal information feature views are fused in combination with the operating state of the system to achieve accurate and reliable data reconstruction under dynamic system conditions.

[0005] To achieve the above object, the technical solution of the present invention is: a PMU data reconstruction method based on spatio-temporal information feature fusion, including the following steps:

[0006] Step S1: Establish a system operation state identification method to determine the operation stage of the system at the latest moment when the measurement data x ij is located; if it is in a steady state, go to Step S2; if it is in a transient state, go to Step S4;

[0007] Step S2: When in a steady state, use the GTV method to detect and identify whether x ij is low-quality data; if so, go to the next step; if not, stop the calculation;

[0008] Step S3: Use GTV, GSV, LTV, and LSV to perform a preliminary prediction on x ij ; then use the overall error adaptive weight allocation obtained by training non-low-quality data under steady state to perform data reconstruction on x ij ;

[0009] Step S4: When in a transient state, use the GSV method to detect and identify whether x ij is low-quality data; if so, go to the next step; if not, stop the calculation;

[0010] Step S5: Use GSV and LSV to perform a preliminary prediction on x ij ; then use the overall error adaptive weight allocation obtained by training non-low-quality data under transient state to perform data reconstruction on x ij ;

[0011] In an embodiment of the present invention, Step S1 is specifically implemented as follows:

[0012] Step S11: System mutation identification (the system enters a transient state from a steady state):

[0013] PMUs are installed on n nodes of the regional power grid. Let the data y t of a certain node at the latest moment t be different from the data y t-1 at the previous moment, and the absolute value of the difference is ε:

[0014] ε = |y t - y t-1 | (1)

[0015] When the difference ε exceeds the normal range, it may be caused by an abnormal event in the system or the presence of low-quality data. When it is caused by an abnormal event in the system, both the average and median of the differences of these n nodes will change significantly; when it is caused by low-quality data, only the average will change significantly. Therefore, the following system mutation detection formula is obtained:

[0016]

[0017] Among them, when θ tWhen it is equal to 1, it indicates a mutation occurring at time t, θ t When it is equal to 0, it indicates no mutation; ε mean and ε median represent the average and median of the differences of n node values; E mean and E median represent the thresholds of the average and median of the differences, which are determined through system testing and experience.

[0018] Step S12, System transient - steady - state identification method (the system returns from the transient state to the steady state):

[0019] During the transient stage, the measured values of adjacent nodes will show co - directional fluctuations. At this time, the correlation between the measured values of the two nodes increases significantly. As the system transitions to the steady state, the correlation slowly decreases; when the transient fluctuations end and no co - directional fluctuations occur anymore, only uncertain errors remain, and at this time, the correlation between the measured values of the two nodes drops to the lowest. The Spearman rank - correlation coefficient method has strong robustness for data with outliers. Therefore, the following formula is used to calculate the correlation coefficient ρ between adjacent nodes:

[0020]

[0021] where, T represents the length of the selected data; r t represents the rank difference of the measured values at both ends of adjacent nodes at time t. The rank of a number is the position of this number in a column sorted from smallest to largest.

[0022] To reduce randomness, the correlation coefficients ρ between both ends of each branch in the region are accumulated and then averaged:

[0023]

[0024] where, ρ mean is the average correlation coefficient of h branches; ρ H is the correlation coefficient between both ends of the H - th branch.

[0025] Generally speaking, 0 - 0.09 indicates no correlation, 0.1 - 0.3 indicates weak correlation, 0.3 - 0.5 indicates medium correlation, and 0.5 - 1.0 indicates strong correlation. Therefore, the following discriminant is set:

[0026]

[0027] where, S t represents the state of the system operation at time t. When S t = 1, it is the transient transition stage of the system; when S t = 0, it is the stable stage of the system.

[0028] In an embodiment of the present invention, in steps S2 and S4, the method for identifying low-quality data is as follows:

[0029] The spatio-temporal MVL algorithm (i.e., the spatio-temporal MVL data reconstruction method from the global and local multi-dimensional perspectives of spatio-temporal information fusion constructed in step S3) can be used for identifying low-quality data. Due to the different data volatility in the transient and steady-state stages of the system, different low-quality data identification methods are adopted for different system operation stages.

[0030] PMU data has a high frequency, resulting in a very short sampling time. Therefore, it is necessary to quickly and accurately detect and identify low-quality data. Among the four proposed spatio-temporal MVL algorithms, the GTV method has a good prediction effect for the steady state; the GSV method has a good prediction effect for the next moment in the transient stage when the measurement data at the previous moment is accurate. Therefore, when performing real-time online data cleaning and reconstruction on PMU measurement data, the GTV method can be used to identify low-quality data in the steady state, and the GSV method can be used to identify low-quality data in the transient state.

[0031]

[0032] Where φ ij = 1 indicates the presence of low-quality data at node i at time j, and φ ij = 0 indicates normal data; And respectively represent the thresholds of the residuals between the prediction results of the GTV method and the GSV method and the measured values, which are obtained by testing non-missing data.

[0033] In an embodiment of the present invention, the specific implementation manner of using GTV, GSV, LTV, and LSV to perform a preliminary prediction on x ij is as follows:

[0034] Since the power system is a dynamic system with strong inertia, many components in the power system, represented by generators, are dynamic components. These components respond according to inputs and historical states, making the state of the power system at the current moment closely related to the state at the previous moment, and there is a strong time correlation in the system state. Under the same system operation state and the output data of the same device, within a certain time window T win , the continuous measurement data obtained on a single-channel PMU device has time correlation. Given the sampling time interval ΔT, using x to replace specific types of measurement data such as the node voltage amplitude V and the node voltage phase angle θ, its measurement matrix X within the time window T win is expressed as:

[0035] X = [x1, x2,..., xi ,..., x n T , 1 ≤ i ≤ n (7)

[0036]

[0037] where x i = [x i1 , x i2 ,..., x im represents the continuous measurement data related to node i, and m = T win / ΔT represents the total number of sampling points in the sliding time window.

[0038] Therefore, the invention proposes to reconstruct the low-quality PMU data based on the basic idea of multiview-based learning (MVL). Based on the spatio-temporal correlation exhibited by the measurement data of each PMU device within the regional power grid within a certain time range, low-quality data detection and identification are performed on the measurements, and through the extraction of spatio-temporal information characteristics of the measurement data, a spatio-temporal MVL data reconstruction method with global and local multi-dimensional perspectives of spatio-temporal information fusion is constructed.

[0039] Global Temporal View (GTV):

[0040] The GTV data recovery method is constructed based on the exponential smoothing method. For the low-quality target data x ij at the j-th moment of node i in the measurement matrix X, select the previous m gt consecutive time node data to form a long time series

[0041]

[0042] Based on this long time series, predict x ij to obtain the GTV predicted value

[0043]

[0044] where x it represents the data at time t of node i; α ∈ (0, 1) represents the smoothing parameter in the exponential smoothing method, and in the GTV method, according to the distance between x it and the target data x ij , weights are assigned through α(1 - α) j-t-1 .

[0045] ​In order to enable the GTV method to change according to the phased characteristics of the time series, it is necessary to set α to be adaptively selected. By using the comprehensive prediction error E of a series of predicted values and measured values before the j-th moment j to determine α:

[0046] E j =βe j-1 +(1 - β)E j-1 (11)

[0047] where β can be set to 0.1 as the initial value of α; e j-1 is the prediction error at the (j - 1)-th moment. At the same time, since α ∈ (0, 1), it is necessary to normalize E j :

[0048] M j =β|e j-1 |+(1 - β)M j-1 (12)

[0049] α j =|E j | / M j (13)

[0050] where M j is the absolute smoothing error.

[0051] Global spatial view (GSV):

[0052] The GSV data recovery method is inspired by some spatial interpolation methods in geoscience and is implemented based on the inverse distance weight (IDW). Since the electrical distance between grid nodes is quantified by the equivalent impedance, the impedance matrix Z all is obtained by the given admittance matrix Y of the entire regional system all =Y all -1 . Then, select the rows and columns of n relevant nodes equipped with PMUs from it to form a small impedance matrix Use to calculate the electrical distance between each node:

[0053] d ik =|(z ii -z ik )-(z ki -z kk ) 2 ,(1 ≤ i, k ≤ n) (14)

[0054] where d ikrepresents the electrical distance between node i and node k; z ik Indicated in The impedance between node i and node k is then normalized:

[0055]

[0056] Among them, D ik is the normalized electrical distance between node i and node k. The D value can be calculated between n nodes to form a spatial distance matrix D = [D ik ] n×n .

[0057] Utilize low-quality target data x ij The data of other nodes at the same time j form a long spatial sequence

[0058] x gs =[x 1j ,x 2j ,...,x (i-1)j ,x (i+1)j ,...,x nj ] (16)

[0059] Based on the long sequence in this space, ij Predict the GSV predicted value

[0060]

[0061] Among them, first use the j-1 time x i(j-1) / x k(j-1) The obtained ratio is x kj Adjust to get x ij The estimated value is then assigned weights through D based on the IDW method. Since low-quality data may appear simultaneously on multiple nodes at time t, all low-quality data at time t can be replaced with the data of the previous moment, and then replaced one by one using the results of the GSV method.

[0062] Local temporal view (LTV)

[0063] The LTV data recovery method is based on the correlation similarity method. That is, the node k in a certain area is in t∈(jm lt ,j-1) The measured value x kt Compared with the measured value x obtained at time j kj The similarity between The measurement value x of node i in the same area at time t itThe similarity between it and the measurement value x obtained at time j ij is correlated. Based on this, the local spatio-temporal matrix X is selected: L :

[0064]

[0065] To reconstruct the low-quality target data x ij , using the similarity of the measurement values of other nodes in X L at time t for time j to generate an average similarity on node i

[0066]

[0067] At the m ij consecutive time node data before x lt each generate Combined to form a time similarity sequence, used to assign a weight value to x it , thereby predicting the LTV prediction value for x ij

[0068]

[0069] Local spatial view (LSV)

[0070] The calculation method of the LSV data recovery method is similar to that of LTV. Using the similarity of the measurement values of node k on node i at t∈(j - m lt , j - 1) to generate an average similarity at time j

[0071]

[0072] At the data of other nodes at the same time j as x ij each generate Combined to form a spatial similarity sequence, used to assign a weight value to x kj , thereby predicting the LSV prediction value for x ij

[0073]

[0074] ​​​In an embodiment of the present invention, in steps S3 and S5, the overall error obtained by training non-low-quality data using a spatial and temporal information feature fusion model (STIFFM) is adaptively weighted to x ij for data reconstruction, specifically as follows:

[0075] Due to the constraints of electrical connections and the characteristics of components in the power system, there are significant temporal and spatial separations in the system state. Using only the measurement information of a single section and linear line parameters cannot fully characterize the global spatio-temporal information features of the power system. Therefore, it is necessary to integrate four algorithms in the spatio-temporal MVL to construct a STIFFM data recovery method.

[0076] Since the two temporal methods, GTV and LTV, cannot predict values well when dealing with disturbance events, only the spatial methods, GSV and LSV, are used for combined prediction during the transient process. Four algorithms in the spatio-temporal MVL are used to train the non-low-quality data of the target system, and the weights of each prediction value are adaptively updated according to the overall error of the four algorithms in the transient and steady-state stages:

[0077]

[0078] where, is the final result obtained by STIFFM; L gsv1 and L lsv1 are the overall errors of the GSV and LSV methods in the transient transition stage of the S t =1 system, and their reciprocals are used as weights; L gtv0 and L gsv0 and L ltv0 and L lsv0 are the overall errors of the GTV, GSV, LTV, and LSV methods in the stable stage of the S t =0 system, and their reciprocals are used as weights; the overall error situation can be calculated online through non-low-quality data and their predicted values.

[0079] Compared with the prior art, the present invention has the following beneficial effects: A PMU data reconstruction method based on spatio-temporal information feature fusion provided by the present invention first proposes a multi-dimensional view learning method based on spatio-temporal information features to obtain preliminary predicted values from the global and local perspectives of time and space; further, it combines the operating state of the system to fuse four methods with spatio-temporal information feature views to achieve accurate and reliable data reconstruction under dynamic system conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 is the system flow chart of the present invention;

[0081] Figure 2 This is the topology diagram of the IEEE-39 bus system in the embodiment of the present invention;

[0082] Figure 3 This is the identification result diagram of the system operation state in the embodiment of the present invention;

[0083] Figure 4 This is the comparison of single-point low-quality data reconstruction under steady state in the embodiment of the present invention;

[0084] Figure 5 This is the comparison of continuous low-quality data reconstruction under steady state in the embodiment of the present invention;

[0085] Figure 6 This is the comparison of single-point low-quality data reconstruction under transient state in the embodiment of the present invention;

[0086] Figure 7 This is the comparison of continuous low-quality data reconstruction under transient state in the embodiment of the present invention. Detailed implementation manners

[0087] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings.

[0088] As Figure 1 shown, a PMU data reconstruction method based on spatio-temporal information feature fusion proposed in the embodiment of the present invention includes the following steps:

[0089] Step 1: Establish a system operation state identification method to determine the operation stage of the system at the latest moment when x ij is located. If it is in a steady state, go to Step 2; if it is in a transient state, go to Step 4;

[0090] Step 2: When in a steady state, use the GTV method to detect and identify whether x ij is low-quality data. If so, go to the next step; if not, stop the calculation;

[0091] Step 3: Use GTV, GV, LTV, and LV to make a preliminary prediction for x ij ; then use the overall error adaptively allocated weights obtained by training non-low-quality data under steady state to reconstruct the data of x ij ;

[0092] Step 4: When in a transient state, use the GV method to detect and identify whether x ij is low-quality data. If so, go to the next step; if not, stop the calculation;

[0093] Step 5: Use GV and LV to make a preliminary prediction for x ij ; then use the overall error adaptively allocated weights obtained by training non-low-quality data under transient state to reconstruct the data of xij Perform data reconstruction.

[0094] For the above PMU data reconstruction method based on spatio-temporal information feature fusion, the implementation of step 1 includes the following steps:

[0095] (1) System mutation identification (the system enters the transient state from the steady state):

[0096] PMUs are installed on n nodes of the regional power grid. Let the data y of a certain node at the latest moment t t , and the absolute value of the difference between the data y t-1 at the previous moment is ε:

[0097] ε = |y t - y t-1 | (1)

[0098] When the difference ε exceeds the normal range, it may be due to an abnormal event in the system or the presence of low-quality data. When it is caused by an abnormal event in the system, both the average and median of the differences of these n nodes will change significantly; when it is caused by low-quality data, only the average will change significantly. Therefore, the following system mutation detection formula is obtained:

[0099]

[0100] Among them, when θ t = 1, it means that a mutation occurs at time t, and when θ t = 0, it means that there is no mutation; ε mean and ε median represent the average and median of the differences of n nodes; E mean and E median represent the thresholds of the average and median of the differences, which are determined through system testing and experience.

[0101] (2) System transient and steady-state identification method (the system returns to the steady state from the transient state):

[0102] During the transient stage, the measured values of adjacent nodes will show co-directional fluctuations. At this time, the correlation between the measured values of the two nodes increases significantly. As the system transitions to the steady state, the correlation slowly decreases; when the transient fluctuations end and no co-directional fluctuations occur, only uncertain errors remain, and at this time, the correlation between the measured values of the two nodes drops to the lowest. The Spearman rank correlation coefficient method has strong robustness for data with outliers. Therefore, the following formula is used to calculate the correlation coefficient ρ between adjacent nodes:

[0103]

[0104] Among them, T represents the length of the selected data; r tIt represents the rank difference of the measured values at both ends of adjacent nodes at time t. The rank of a number is the position of this number after sorting the column it belongs to from small to large.

[0105] To reduce randomness, the correlation coefficients ρ between both ends of each branch in the area are accumulated and then the average value is calculated:

[0106]

[0107] Among them, ρ mean is the average correlation coefficient of h branches; ρ H is the correlation coefficient between both ends of the H-th branch.

[0108] Generally speaking, 0 - 0.09 means no correlation, 0.1 - 0.3 means weak correlation, 0.3 - 0.5 means medium correlation, and 0.5 - 1.0 means strong correlation. For this reason, the following discriminant is set:

[0109]

[0110] Among them, S t represents the state in which the system operates at time t. When S t = 1, it is the transient transition stage of the system; when S t = 0, it is the stable stage of the system.

[0111] For the above PMU data reconstruction method based on spatio-temporal information feature fusion, in steps 2 and 4, the spatio-temporal MVL algorithm is used to identify low-quality data. Due to the different data volatility in the transient and steady-state stages of the system, different low-quality data identification methods are adopted for different system operation stages.

[0112] PMU data has high frequency, making its sampling time very short. For this, it is necessary to quickly and accurately detect and identify low-quality data. Among the four spatio-temporal MVL algorithms proposed, the GTV method has a good prediction effect for the steady state; the GSV method has a good prediction effect for the next moment in the transient stage when the measured data at the previous moment is accurate. For this reason, when performing real-time online data cleaning and reconstruction on PMU measurement data, the GTV method can be used to identify low-quality data in the steady state, and the GSV method can be used to identify low-quality data in the transient state.

[0113]

[0114] Among them, φ ij = 1 indicates that low-quality data appears at node i at time j, and φ ij = 0 indicates that the data is normal; and They respectively represent the thresholds of the residuals between the prediction results of the GTV method and the GSV method and the measured values, which are obtained by testing the non-missing data.

[0115] For the above PMU data reconstruction method based on spatio-temporal information feature fusion, in step 3, using GTV, GSV, LTV, and LSV to perform a preliminary prediction on x ij The implementation of the preliminary prediction includes the following steps:

[0116] Since the power system is a dynamic system with strong inertia, many components in the power system, represented by generators, are dynamic components. These components respond according to the input and historical states, making the state of the power system at the current moment closely related to the state at the previous moment, and there is a strong time correlation in the system state. Under the same system operating state and the output data of the same device, within a certain time window T win The continuous measurement data obtained on a single-channel PMU device has temporal correlation. Given the sampling time interval ΔT, using x to replace specific types of measurement data such as the node voltage magnitude V and the node voltage phase angle θ, its measurement matrix X within the time window T win is expressed as:

[0117] X = [x1, x2,..., x i ,..., x n T , 1 ≤ i ≤ n (7)

[0118]

[0119] Among them, x i = [x i1 , x i2 ,..., x im represents the continuous measurement data related to node i, and m = T win / ΔT represents the total number of sampling points in the sliding time window.

[0120] Therefore, the invention proposes the basic idea of multiview-based learning (MVL) to reconstruct PMU low-quality data. Based on the spatio-temporal correlation exhibited by the measurement data of each PMU device within a certain time range in the regional power grid, low-quality data detection and identification are performed on the measurements, and through the extraction of spatio-temporal information features of the measurement data, a spatio-temporal MVL data reconstruction method in the global and local multi-dimensional perspectives of spatio-temporal information fusion is constructed.

[0121] (1) Global temporal view (GTV):

[0122] The GTV data recovery method is constructed based on the exponential smoothing method. For the low-quality target data x at the j-th moment of node i in the measurement matrix X ij , select the data of its previous m gt consecutive time node data to form a long time series

[0123]

[0124] Based on this long time series, predict x ij to obtain the GTV predicted value

[0125]

[0126] Among them, x it represents the data at time t of node i; α∈(0,1) represents the smoothing parameter in the exponential smoothing method. In the GTV method, according to the distance between x it and the target data x ij , the weight is assigned through α(1-α) j-t-1 .

[0127] In order to enable the GTV method to change according to the stage characteristics of the time series, α needs to be set for adaptive selection. By using the comprehensive prediction error E j of a series of predicted values and measured values before the j-th moment to determine α:

[0128] E j = βe j-1 +(1-β)E j-1 (11)

[0129] Among them, β can be set to 0.1 as the initial value of α; e j-1 is the prediction error at the (j - 1)-th moment. At the same time, since α∈(0,1), it is necessary to normalize E j :

[0130] M j = β|e j-1 |+(1-β)M j-1 (12)

[0131] α j = |E j | / M j (13)

[0132] Among them, M j is the absolute smoothing error.

[0133] (2) Global Spatial View (GSV):

[0134] The GSV data recovery method is inspired by some spatial interpolation methods in geoscience and is implemented based on the inverse distance weight (IDW). Since the electrical distance between grid nodes is quantified by the equivalent impedance, the impedance matrix Z is obtained through the admittance matrix Y of the given entire regional system all to obtain the impedance matrix Z all = Y all -1 . Then, the rows and columns of n relevant nodes equipped with PMUs are selected from it to form a small impedance matrix Using calculate the electrical distance between each node:

[0135] d ik = |(z ii - z ik ) - (z ki - z kk ) 2 , (1 ≤ i, k ≤ n) (14)

[0136] where d ik represents the electrical distance between node i and node k; z ik represents the impedance between node i and node k in . Then, normalization is performed:

[0137]

[0138] where D ik is the electrical distance between node i and node k after normalization. The D values can be calculated pairwise between n nodes to form a spatial distance matrix D = [D ik n×n .

[0139] Using the low-quality target data x ij and the data of other nodes at the same moment j, a spatial long sequence is formed

[0140] x gs = [x 1j , x 2j ,..., x (i-1)j , x (i+1)j ,..., x nj (16)

[0141] Based on this spatial long sequence, the GSV predicted value is obtained by predicting x ij

[0142] ​​

[0143] Among them, first use the ratio obtained from x i(j-1) / x k(j-1) at time j - 1 to adjust x kj and obtain the estimated value of x ij . Then, based on the IDW method, allocate weights through D. Since at time t, low-quality data may appear simultaneously on multiple nodes, for this, all low-quality data at time t can be first replaced with the data from the previous moment, and then the results of the GSV method are used for one-by-one replacement.

[0144] (3) Local Temporal View (LTV)

[0145] The LTV data recovery method is established based on the correlation similarity method. That is, the similarity between the measurement value x lt obtained by node k in a certain area at time t ∈ (j - m kt , j - 1) and the measurement value x kj obtained by it at time j is correlated and similar to the similarity between the measurement value x it obtained by node i in the same area at time t and the measurement value x ij obtained by it at time j . Based on this, select the local spatio-temporal matrix X L :

[0146]

[0147] to reconstruct the low-quality target data x ij . Use the similarity of the measurement values of other nodes in X L at time t for time j to generate the average similarity at node i

[0148]

[0149] at the m ij consecutive time node data before x lt to generate a combined time similarity sequence respectively, which is used to assign weight values to x it and thereby predict the LTV predicted value ij for x

[0150]

[0151] (4) Local Spatial View (LSV)

[0152] The calculation method of the LSV data recovery method is similar to that of LTV. Using at t∈(j - m lt , j - 1), the similarity of the measurement value of node k to node i Generate the average similarity at time j

[0153]

[0154] At x ij At the data of other nodes at the same time j, generate respectively Combine to form a spatial similarity sequence for x kj Assign weight values, thereby for x ij Predict to obtain the LSV predicted value

[0155]

[0156] For the above PMU data reconstruction method based on spatio - temporal information feature fusion, in steps S3 and S5, use the spatial and temporal information feature fusion model (STIFFM) to adaptively allocate weights to the overall error obtained from training non - low - quality data, and perform data reconstruction on x ij as follows:

[0157] Due to the constraints of electrical connections and the characteristics of components in the power system, there are significant temporal and spatial separations in the system state. Only using the measurement information of a single section and linear line parameters cannot fully characterize the global spatio - temporal information features of the power system. Therefore, it is necessary to integrate the four algorithms in spatio - temporal MVL to construct the STIFFM data recovery method.

[0158] Since the two time - based methods GTV and LTV cannot predict values well when dealing with disturbance events, only the spatial methods GSV and LSV are used for combined prediction during the transient process. Use the four algorithms in spatio - temporal MVL to train the non - low - quality data of the target system, and adaptively update the weights of each predicted value through the overall error of the four algorithms in the transient - steady state stage:

[0159]

[0160] Among them, is the final result obtained by STIFFM; L gsv1 、L lsv1 are the overall errors of the GSV and LSV methods in the transient transition stage of the S t =1 system, and their reciprocals are used as weights; Lgtv0 , L gsv0 , L ltv0 , L lsv0 are the overall errors of the GTV, GSV, LTV, and LSV methods in the stable stage of the S t = 0 system, and their respective reciprocals are used as weights; the overall error situation can be calculated online through non-low-quality data and its predicted values.

[0161] Based on the above model and process design, the present invention uses the IEEE 39-node system for simulation verification. The network structure of the IEEE 39-node system is as Figure 2 shown. PMU devices are deployed on all 39 nodes of the system for online monitoring. The sampling time interval ΔT = 0.01 s is set, Gaussian distributed noise with σ = 0.005 pu is added to the steady-state simulation data, and Gaussian distributed noise with σ = 0.007 pu is added to the transient simulation data. To simulate the actual errors of the network model and parameters, Gaussian distributed noise with σ = 0.02 pu is added to the spatial distance matrix D. For the spatio-temporal MVL algorithm, its view size is set to m gt = 30, m lt = 5.

[0162] To simulate the high proportion of new energy access and various possible abnormal situations in the actual power grid, a wind turbine model and new energy disturbance data are added in the simulation, and various transient scenarios such as short circuits, open circuits, and large and small disturbances are set in the system. At the same time, bad data and missing values are randomly set at each node.

[0163] (1) System operating state judgment

[0164] Through the system mutation detection method in step S11, the system disturbance data and low-quality data are distinguished; through the system steady-state and transient identification method in step S12, the stable stage and transient transition stage of the system are distinguished. The simulation results are as Figure 3 shown.

[0165] In Figure 3 , a short circuit fault occurred at 2.0 s and reclosing was performed at 2.1 s. The median of the difference at these two places changed significantly, so it is system interference data. Starting from 4.6 s, the average correlation coefficient is lower than 0.1, so the system transitions from the transient stage to the stable stage. The time between the system mutation and the transient-steady state critical point is the transient transition stage.

[0166] (2) System stable stage simulation

[0167] The disturbances of the system in the steady state mainly come from new energy fluctuations and system noise. For the simulation cases in the stable stage, both single-point low-quality and continuous low-quality conditions are set. Under the stable stage of the system, the results obtained from experiments on single-point low-quality data, non-continuous low-quality data, and continuous low-quality data are as Figure 4 、 5 and shown in Table 1.

[0168] Table 1 Data reconstruction results of 6 methods in the steady state

[0169]

[0170]

[0171] As can be seen from Table 1, in the case where the measured values contain noise and only the measurement data at the current moment and before of the low-quality target values are used, except for HI among them, the other 5 data recovery methods can effectively reconstruct single-point low-quality data in the stable stage. In the case of non-continuous multi-point low-quality data, the errors all increase to some extent, and HI and AI increase more. And since HI, LI, and AI are all linear interpolation methods, when there is continuous multi-point low-quality data, their errors increase significantly. And from Figure 5 it can be seen that the errors of STIFFM, LSTM, and ANN basically remain unchanged in the case of multi-point continuous low-quality data. The STIFFM method and LSTM proposed in the invention can effectively perform data reconstruction in the steady state.

[0172] (3) Simulation of the transient transition stage of the system

[0173] When the system is in the transient state, 6 methods are respectively used to reconstruct the low-quality PMU data. The results obtained from experiments on single-point low-quality data, non-continuous low-quality data, and continuous low-quality data are as Figure 6 、 7 and shown in Table 2.

[0174] Table 2 Data reconstruction results of 6 methods in the transient state

[0175]

[0176] As can be seen from Table 2, in the transient transition stage, the errors in various low-quality cases all increase to some extent. Among the three interpolation methods of HI, LI, and AI, when facing continuous low-quality data, the deviations are too large. And LSTM and ANN, which performed well in the steady state, when facing continuous multi-point low-quality data in the transient stage, from Figure 7It can also be seen that data reconstruction cannot be effectively carried out. Moreover, LSTM and ANN are the results after targeted training. In the face of various abnormal events in the actual power grid, it will be even more difficult to adopt their data reconstruction results. As shown in Table 2, the STIFFM method proposed in the invention has the best reconstruction effect in the face of various transient low-quality situations. And this method does not need to be specially trained. In the face of the transient process generated by various abnormal events, it can effectively and accurately reconstruct discontinuous and continuous PMU low-quality measurement data.

[0177] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention, when the functional effects produced do not exceed the scope of the technical solution of the present invention, shall fall within the protection scope of the present invention.

Claims

1. A PMU data reconstruction method based on spatio-temporal information feature fusion, characterized in that It includes the following steps: Step S1: Establish a system operation state identification method to determine the operation stage of the system at the latest moment when the measurement data x ij is located; if it is in a steady state, go to step S2; if it is in a transient state, go to step S4; Step S2: When in a steady state, use the GTV method to detect and identify x ij whether it is low-quality data; if so, proceed to the next step; if not, stop the calculation; Step S3: Use GTV, GSV, LTV, and LSV to preliminarily predict x ij respectively; then use the overall error adaptively allocated weights trained on non-low-quality data under steady state to reconstruct the data of x ij ​ Step S4. When in the transient state, use the GSV method to detect and identify x ij whether it is low-quality data; if so, proceed to the next step; if not, stop the calculation; Step S5: Use GSV and LSV to make a preliminary prediction for x ij ; then use the overall error adaptively allocated weights trained on non-low-quality data in the transient state to reconstruct the data for x ij ​ In step S3 and step S5, the overall error obtained from training non-low-quality data using the spatio-temporal information feature fusion model STIFFM is adaptively weighted to x ij for data reconstruction, specifically as follows: Due to the constraints of electrical connections and the characteristics of components in the power system, there are significant temporal and spatial separations in the system state. Using only the measurement information of a single section and linear line parameters cannot fully characterize the global spatio-temporal information characteristics of the power system. Therefore, four algorithms, namely GTV, GSV, LTV, and LSV, in the spatio-temporal MVL data reconstruction method from the global and local multi-dimensional perspectives of spatio-temporal information fusion are integrated to construct the STIFFM data recovery method; Since GTV and LTV cannot well predict values when dealing with disturbance events, only the spatial methods GSV and LSV are used for combined prediction during the transient process; four algorithms in the spatio-temporal MVL data reconstruction method from the global and local multi-dimensional perspectives of spatio-temporal information fusion are used to train the non-low-quality data of the target system, and the weights of each prediction value are adaptively updated through the overall errors of the four algorithms in the transient and steady-state stages: Among them, is the final result obtained by STIFFM; L gsv1 , L lsv1 are the overall errors of the GSV and LSV methods in the transient transition stage of the S t =1 system, and their respective reciprocals are used as weights; L gtv0 , L gsv0 , L ltv0 , L lsv0 are the overall errors of the GTV, GSV, LTV, and LSV methods in the stable stage of the S t =0 system, and their respective reciprocals are used as weights; respectively represent the GSV prediction value, GTV prediction value, LTV prediction value, and LSV prediction value predicted for x ij .

2. The PMU data reconstruction method based on spatio-temporal information feature fusion according to claim 1, wherein The specific implementation of step S1 is as follows: Step S11, system mutation identification: PMUs are installed at n nodes of the regional power grid. Let y be the data of a certain node at the latest moment t t , and the absolute value of the difference between the data y t-1 at the previous moment is ε: ε = |y t - y t-1 | (1) When the difference ε exceeds the normal range, and if it is caused by an abnormal event in the system, both the average and median of these n node differences will change; if it is caused by low-quality data, only the average will change. Therefore, the following system mutation detection formula is obtained: Among them, when θ t = 1, it indicates a mutation occurring at time t, and when θ t = 0, it indicates no mutation; ε mean and ε median represent the average and median of the differences of n nodes; E mean and E median represent the thresholds of the average and median of the differences. Step S12, system transient and steady-state identification method: During the transient stage, the measured values of adjacent nodes will show the same-direction fluctuation. At this time, the correlation between the measured values of the two nodes increases. As the transition to the steady state occurs, the correlation slowly decreases; when the transient fluctuation ends and there is no longer the same-direction fluctuation, only the uncertain error remains, and at this time, the correlation between the measured values of the two nodes drops to the lowest; the following formula is used to calculate the correlation coefficient ρ between adjacent nodes: where T represents the selected data length; r t represents the rank difference of the measured values at both ends of adjacent nodes at time t; the rank of a number is its position in a column after sorting the column in ascending order; To reduce randomness, the correlation coefficients ρ between both ends of each branch in the region are accumulated and then averaged: Among them, ρ mean is the average correlation coefficient of h branches; ρ H is the correlation coefficient between both ends of the H-th branch; The following discriminant is set: Among them, S t represents the state of the system operation at time t. When S t = 1, it is the transient transition stage of the system; when S t = 0, it is the stable stage of the system.

3. The PMU data reconstruction method based on spatio-temporal information feature fusion according to claim 2, wherein In steps S2 and S4, detect and identify x ij The calculation formula for whether it is low-quality data is as follows: Among them, φ ij = 1 indicates that low-quality data appears at node i at time j, and φ ij = 0 indicates that the data is normal; and respectively represent the thresholds of the residuals between the predicted results of the GTV method and the GSV method and the measured values.

4. The PMU data reconstruction method based on spatio-temporal information feature fusion according to claim 2, wherein In step S3, the specific implementation method for making a preliminary prediction of x using GTV, GSV, LTV, and LSV is as follows: ij The specific implementation method for making a preliminary prediction of x using GTV, GSV, LTV, and LSV is as follows: Under the same system operating state and for the output data of the same device, within the time window T win the continuous measurement data obtained on a single-channel PMU device are temporally correlated; Given a sampling time interval ΔT, x is used to replace the measurement data including the node voltage magnitude V and the node voltage phase angle θ, and the measurement matrix X within the time window T win is expressed as: X = [x1, x2,..., x i ,..., x n T , 1 ≤ i ≤ n (7)​ And it is expressed in the following form: where x i = [x i1 , x i2 ,..., x im represents the continuous measurement data related to node i, and m = T win / ΔT represents the total number of sampling points in the sliding time window; The basic idea of multi-view learning MVL is proposed to reconstruct the low-quality PMU data, that is, based on the spatio-temporal correlation exhibited by the measurement data of each PMU device in the regional power grid within a certain time range, the low-quality data detection and identification of the measurement are carried out, and through the extraction of the spatio-temporal information characteristics of the measurement data, a spatio-temporal MVL data reconstruction method from the global and local multi-dimensional perspectives of spatio-temporal information fusion is constructed, and the specific implementation is as follows: Global time view GTV: The GTV is constructed based on the exponential smoothing method; for the low-quality target data x of node i at the j-th moment in the measurement matrix X ij , select the data of its previous m gt consecutive time node data to form a long time series Based on a long time series for x ij the predicted value of GTV is obtained through prediction where x it represents the data at time t of node i; α ∈ (0, 1) represents the smoothing parameter in the exponential smoothing method, and in GTV, according to x it and the distance to the target data x ij weights are assigned through α(1 - α) j-t-1 . By utilizing the comprehensive prediction error E of a series of predicted values and measured values before the j-th moment j to determine α: E j = βe j-1 + (1 - β)E j-1 (11) Among them, β is set to 0.1 and used as the initial value of α; e j-1 is the prediction error at the (j - 1)th moment; meanwhile, since α ∈ (0, 1), it is necessary to perform normalization on E j as follows: M j = β|e j-1 | + (1 - β)M j-1 (12) α j = |E j | / M j (13) Among them, M j is the absolute smoothing error; Global space view GSV: The GSV is implemented based on the inverse distance weighted method (IDW). Since the electrical distance between grid nodes is quantified by the equivalent impedance, the admittance matrix Y of the entire regional system is given to obtain the impedance matrix Z all and then the rows and columns of n relevant nodes equipped with PMUs are selected from it to form a small impedance matrix all = Y all -1 and the electrical distances between nodes are calculated using by : d ik = |(z ii - z ik ) - (z ki - z kk )| 2 , (1 ≤ i, k ≤ n) (14) Among them, d ik represents the electrical distance between node i and node k; z ik represents the impedance between node i and node k; then normalization is performed: Among them, D ik is the electrical distance between the normalized node i and node k. The D values are calculated pairwise among the n nodes to form a spatial distance matrix D = [D ik n×n ;​ Utilize low-quality target data x ij Other node data at the same moment j to form a long spatial sequence x gs = [x 1j , x 2j ,..., x (i-1)j , x (i+1)j ,..., x nj (16) Based on the spatial long sequence for x ij the GSV predicted value is obtained by prediction Among them, first use the ratio obtained from x i(j-1) / x k(j-1) at time j - 1 to adjust x kj to obtain an estimated value of x ij Then, based on the IDW method, weights are assigned through D. Since at time t, low-quality data may appear simultaneously on multiple nodes, for this, first replace all the low-quality data at time t with the data from the previous moment, and then use the results of the GSV method for one-by-one replacement; Local time view LTV: The LTV is established based on the relevant similarity method, that is, the measurement value x obtained by node k in a certain area at time t ∈ (j - m lt , j - 1) kt and the measurement value x obtained by it at time j kj have a similarity with the measurement value x of node i in the same area at time t it and the measurement value x obtained by it at time j ij have a similarity which is relevantly similar; based on this, the local spatio-temporal matrix X L is selected as follows: To reconstruct the low-quality target data x ij , the similarity of the measured values of other nodes in X L at time t for time j is used to generate the average similarity on node i At x ij At the m lt consecutive time node data before, generate respectively Combine to form a time similarity sequence, used to assign a weight value to x it Thereby, for x ij Predict to obtain the LTV prediction value Local space view LSV: The calculation method of LSV is similar to that of LTV; by using the similarity of the measurement value of node k to node i at time t ∈ (j - m lt , j - 1), generate the average similarity at time j At x ij At the same moment j, other node data are each generated Combined to form a spatial similarity sequence for x kj Assign a weight value to x, thereby ij Predict to obtain the LSV prediction value

Images