LSTM (Long Short Term Memory)-based method for filling station transformer data of intelligent fusion terminal according to working conditions

Abstract

The invention provides an LSTM (Long Short Term Memory)-based intelligent fusion terminal station transformer data working condition dividing filling method. The method comprises the steps of data acquisition and working condition division: carrying out working condition refining and pile dividing on station area data uploaded by an intelligent fusion terminal according to a photovoltaic equipment operation state OpSt; data preprocessing: carrying out abnormal value detection, missing value preliminary filling, filtering denoising and normalization processing on the data after working condition division; sub-working-condition model training: aiming at each type of working conditions, respectively constructing and adopting the preprocessed data to train a corresponding LSTM data filling model; and missing value classification and filling: calculating the working condition category of to-be-filled data according to the similarity, calling the trained LSTM data filling model under the corresponding working condition, and performing prediction filling on the missing value. The method can solve the problems of high dimension, non-stationarity and complex missing modes in the monitoring data of the photovoltaic transformer area.

Description

Technical Field

[0001] This invention relates to power system data processing, specifically to a method for filling in the working condition of transformer data in an intelligent fusion terminal based on LSTM. Background Technology

[0002] With the large-scale integration of distributed photovoltaic (PV) power generation into the distribution network, power quality issues such as voltage exceeding limits, harmonic pollution, and three-phase imbalance are becoming increasingly prominent. Voltage, current, and other power data are prerequisites for power quality analysis and management in the distribution network; therefore, real-time and accurate monitoring is crucial. As a key link connecting the distribution network and end-users, transformer substations have seen widespread adoption of smart distribution transformer combine terminal units (SCUs) in recent years to achieve comprehensive perception and real-time tracking of their operational status. These SCUs can acquire and analyze substation power data in real time, possessing high flexibility and responsiveness. However, in real-world applications, changes in the data acquisition environment, network fluctuations during data upload, equipment malfunctions, and human error can all affect the accuracy of SCU data acquisition. Therefore, data loss inevitably occurs during the actual data upload process of the SCU, severely impacting data quality and hindering subsequent analysis and evaluation of power quality on the user side. Thus, accurately filling in missing SCU data is one of the important prerequisites for ensuring the power supply quality of the substations.

[0003] Although methods for imputing missing data in solar substation data units (SCUs) have been studied to some extent, several key issues and challenges remain. First, existing research does not consider the high dimensionality of solar substation data, making it difficult to maintain physical consistency in the imputation results, thus affecting imputation accuracy and generalization ability. Second, most existing studies do not consider the non-stationarity of solar substation data, making imputation methods based on static statistical assumptions or fixed patterns difficult to accurately adapt. Finally, the complex missing data patterns of substation data mean that current methods perform well in handling random missing data, but when faced with large-scale continuous missing data, interpolation methods are prone to error accumulation, while machine learning-based imputation methods may significantly reduce prediction accuracy due to the long span of missing data. Therefore, there is an urgent need for an efficient data imputation method that adapts to the complex characteristics of solar substation data. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a method for filling data from LSTM-based intelligent converged terminal transformers under different operating conditions. This method considers dynamic influencing factors such as photovoltaic output and load power variations, constructs an LSTM-based time-series data filling model for different operating conditions, and employs a data stacking method for transformer areas based on photovoltaic operating conditions. The stacked data is then used to construct the filling model for the corresponding operating condition. The method is simple, easy to implement, has strong generalization ability, high accuracy, and can adapt to changes in grid conditions and topology. Compared to existing methods, it has significant advantages in data filling for photovoltaic transformer areas in distribution networks.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for filling in the functional data of an intelligent converged terminal using LSTM, comprising the following steps:

[0007] Data acquisition and operating condition classification: The operating condition data of the transformer substations uploaded by the smart fusion terminal is refined and classified according to the operating status OpSt of the photovoltaic equipment;

[0008] Data preprocessing: outlier detection, preliminary missing value imputation, filtering and denoising, and normalization are performed on the data after the work conditions are divided.

[0009] Training of models for different working conditions: For each type of working condition, a corresponding LSTM data-filling model is constructed and trained using preprocessed data;

[0010] Missing value classification and imputation: For the data to be imputed, the working condition category to which the data to be imputed belongs is calculated based on similarity, and the LSTM data imputation model trained under the corresponding working condition is called to predict and impute the missing values.

[0011] Furthermore, the distribution area data includes transformer data and photovoltaic data, and the photovoltaic data includes the voltage, current, active power and reactive power of the photovoltaic equipment.

[0012] Furthermore, the data acquisition and operating condition classification include:

[0013] Based on the photovoltaic equipment operating status OpSt, the data of the distribution area is initially divided into photovoltaic operation period data and shutdown period data. The photovoltaic operation period data includes photovoltaic data and transformer data, while the shutdown period data only includes transformer data.

[0014] K-means clustering was applied to the photovoltaic operation period data and the shutdown period data respectively to further refine them into different operating condition categories.

[0015] Furthermore, the K-means clustering method is applied to the photovoltaic operation period data and the shutdown period data respectively to further refine them into different operating condition categories, including:

[0016] After determining the number of clusters K1 and K2 for the photovoltaic operation period data and the shutdown period data, K-means clustering was used to divide the dataset M under different operating conditions into:

[0017] (1)

[0018] (2)

[0019] (3)

[0020] Where M 1,j M represents the dataset corresponding to operating condition j during the photovoltaic operation period. 2,m Let K1 represent the time set corresponding to the operating condition m during the downtime, where K1≥j≥1 and K2≥m≥1.

[0021] Furthermore, the number of clusters K1 and K2 are determined by the elbow rule. With the goal of minimizing the sum of squared errors (SSE), the optimal number of clusters for K-means is found by plotting the relationship between SSE and K, where the rate of decrease of SSE slows down significantly.

[0022] Furthermore, the filtering and denoising in the data preprocessing uses the Savitzky-Golay filtering algorithm.

[0023] Furthermore, in the training of the sub-working condition model, for each type of working condition, a corresponding LSTM data filling model is constructed and trained using preprocessed data, specifically including:

[0024] The input matrix for the photovoltaic operation period is as follows:

[0025] (4)

[0026] The input matrix for the downtime period is:

[0027] (5)

[0028] Where U, I, P, and Q represent the load voltage, current, active power, and reactive power of phase A of the transformer, respectively, and U', I', P', and Q' represent the voltage, current, active power, and reactive power output by the inverter on the photovoltaic equipment side, respectively.

[0029] For each working condition M i,j If M i,j Given N time series, each of length T, the nth time series is represented as:

[0030] (6)

[0031] Where i = 1, 2, j = 1, 2, ..., Ki , Let be the input vector at time t. For each time step t, the LSTM state update formula is:

[0032] (7)

[0033] Among them, h t It is the hidden state of the LSTM at time t, x t h is the current input vector. t-1 It is the hidden state of the previous time step, c t-1 Let f represent the cell state at the previous time step, and let f denote the nonlinear mapping function of the LSTM. For each time step t, the output of the LSTM is:

[0034] (8)

[0035] Wo is the weight matrix of the output layer, b o The bias vector for the output layer, for operating condition M i,j The filled results for all time series are as follows:

[0036] (9)

[0037] Among them, M̂ i,j Let x̂ represent the dataset under the j-th working condition in the i-th time period after padding. N (i,j) Indicates the filled M̂ i,j The Nth filler value in the dataset.

[0038] Furthermore, the classification and imputation of missing values ​​specifically includes:

[0039] Define working condition M i,j The characteristic mean is:

[0040] (10)

[0041] in, It is working condition M i,j The mean of the k-th feature, N i,j It is working condition M i,j The number of samples in It is the k-th feature value of the n-th sample;

[0042] Calculate the feature mean of data within the same timestamp of the missing values, and calculate the other data for the missing points and the working condition M. i,j Euclidean distance:

[0043] (11)

[0044] x 缺失 The missing part is replaced by the mean of the first non-missing value before and after it, and the result is obtained by comparing d for all working conditions. i,j Value, select the smallest d i,j The missing values ​​are assigned to the corresponding working conditions, and the missing values ​​are filled in using the LSTM data filling model trained for the corresponding working conditions.

[0045] The beneficial effects of this invention are:

[0046] 1. A data clustering method for photovoltaic (PV) distribution areas based on PV operating conditions is proposed. First, the data is initially classified according to the operating status of the PV equipment. Then, the data clustering is further refined based on different operating conditions to improve data matching and provide a more accurate classification basis for model filling.

[0047] 2. A separate LSTM filling method is adopted, which trains LSTM models for different working conditions, reducing interference between working conditions and improving the adaptability and accuracy of filling.

[0048] 3. This method can be applied to smart grid distribution automation systems, which helps improve the integrity of SCU data and provides solid data support for distribution network status monitoring, power quality analysis and intelligent decision-making for large-scale photovoltaic access. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the working condition filling method for intelligent converged terminal station data based on LSTM, as implemented in this invention.

[0050] Figure 2 This is a diagram illustrating the effect of filling missing SCU data in a photovoltaic substation, as provided in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for filling in the working conditions of intelligent converged terminal station data based on LSTM, including the following steps:

[0053] Step 1: Data stacking based on photovoltaic operating conditions

[0054] 1) Preliminary data classification based on photovoltaic operating conditions

[0055] This invention first divides the data collected by the SCU into Class I data and Class II data according to the photovoltaic operating status (OpSt) recorded by the SCU. The specific division rules are as follows:

[0056] ① When OpSt = 1, the photovoltaic equipment is in operation and enters the Class I time period (photovoltaic operating period). At this time, Class I data includes photovoltaic data and transformer data.

[0057] ② When OpSt = 0, the photovoltaic equipment stops working and enters the Class II time period (photovoltaic shutdown period). At this time, Class II data only includes transformer data.

[0058] 2) Data partitioning based on K-means clustering

[0059] This invention, based on the initial division of data into Category I (PV operating period) and Category II (PV shutdown period), introduces the K-means clustering algorithm to refine and group each category of data. Before K-means clustering, it is necessary to determine the optimal number of clusters K, according to the Sum of Squares due to Error (SSE) formula:

[0060] (12)

[0061] Among them, C i It is the i-th cluster, and p is C i The sample points in the middle, m i It is C i The center of mass.

[0062] The optimal K can be selected by plotting the relationship between SSE and the number of clusters K using the elbow theorem. The optimal K can be determined by observing the inflection point where the rate of decrease in the SSE curve slows down significantly.

[0063] After determining the number of clusters K1K2 for Class I and Class II time periods, K-means clustering is used to divide the dataset M under different working conditions for Class I and Class II time periods into:

[0064] (1)

[0065] (2)

[0066] (3)

[0067] Where M 1,j M represents the dataset corresponding to working condition j in time period I. 2,mLet K1 represent the time set corresponding to work condition m in time period II (K1≥j≥1, K2≥m≥1). Finally, each split dataset will be used to train the LSTM padding model for the corresponding work condition.

[0068] Step 2: Data Preprocessing

[0069] To improve the quality of LSTM model training data, outlier handling, missing value imputation, and noise reduction preprocessing are performed on the data after data piling. Outlier handling is defined as follows: if a value is higher than the sum of its preceding and following non-missing values, or higher than twice the preceding or following non-missing values; if a missing value is a single point, it is imputed using the average of the preceding and following time points; if there are consecutive missing values ​​greater than one, it is imputed using the previous normal value. The Savitzky-Golay filtering algorithm effectively removes high-frequency noise while preserving the signal structure. Finally, to accelerate model computation and eliminate the influence of dimensions, the processed data is normalized.

[0070] (13)

[0071] x represents the preprocessed data, 𝑥 min The minimum value in the preprocessed data, 𝑥 𝑚𝑎𝑥 is the maximum value of the preprocessed data, and 𝑥' is the data after standardization.

[0072] Step 3: Train the LSTM data filling model for the corresponding working conditions

[0073] The input matrix for type I time period is:

[0074] (4)

[0075] The input matrix for type II time period is

[0076] (5)

[0077] Wherein, U, I, P, and Q represent the load voltage, current, active power, and reactive power of phase A of the transformer, respectively, and U', I', P', and Q' represent the voltage, current, active power, and reactive power output by the inverter from the photovoltaic equipment side, respectively.

[0078] For each working condition M i,j (i=1, 2, j=1, 2,...,Ki): If M i,j Given N time series, each of length T, the nth time series can be represented as:

[0079] (6)

[0080] in Let be the input vector at time t. For each time step t, the LSTM state update formula is:

[0081] (7)

[0082] Among them, h t It is the hidden state of the LSTM at time t, x t h is the current input vector. t-1 It is the hidden state of the previous time step, c t-1 Let f represent the cell state at the previous time step, and let f denote the nonlinear mapping function of the LSTM. For each time step t, the output of the LSTM is:

[0083] (8)

[0084] Wo is the weight matrix of the output layer, b o This is the bias vector for the output layer. For operating condition M... i,j The filled results for all time series are as follows:

[0085] (9)

[0086] Among them, M̂ i,j Let x̂ represent the dataset under the j-th working condition in the i-th time period after padding. N (i,j) Indicates the filled M̂ i,j The Nth filler value in the dataset.

[0087] Step 4: Fill in missing values ​​according to working conditions

[0088] Define working condition M i,j The characteristic mean is:

[0089] (10)

[0090] in, It is working condition M i,j The mean of the k-th feature. N i,j It is working condition M i,j The number of samples in the sample. It is the k-th feature value of the n-th sample.

[0091] Then, the feature mean is calculated using data from the same timestamp as the missing values, and the other data for the missing points are compared with the working condition M. i,j Euclidean distance:

[0092] (11)

[0093] x 缺失The missing part is replaced with the mean of the first non-missing value before and after it, resulting in the final value. This is achieved by comparing the d values ​​for all operating conditions. i,j Value, select the smallest d i,j The missing values ​​are then assigned to the corresponding working conditions. Finally, the corresponding working condition model trained in step three is used to fill in the missing values.

[0094] Comparative analysis was conducted to verify the accuracy of the infill model:

[0095] This embodiment utilizes data from a photovoltaic power generation area in Southwest China on September 30th, with 70% as the training set and 30% as the test set. The missing data imputation performance of this invention is compared with that of condition-specific LSTM, non-condition-specific LSTM, RNN, and GRU models to test its accuracy. The imputation results of different models are shown below. Figure 2 As shown.

[0096] from Figure 2 As can be seen, in the region of large current fluctuations (18:00-19:00), the condition-specific LSTM model can more accurately capture the changing trend of the actual data (black line), while the non-condition-specific LSTM (red line), RNN (green line), and GRU (purple line) exhibit significant deviations and lags. In the region of small current fluctuations (19:30-23:30), the fluctuation amplitude of the condition-specific LSTM model is closer to the actual value, and it can better avoid excessive filling errors, while other models have relatively large errors and find it difficult to achieve accurate filling when the data has small fluctuations. Compared with the non-condition-specific LSTM, RNN, and GRU models, the condition-specific LSTM model can better capture the data trend and has the best filling effect.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for filling in the working condition data of an intelligent fusion terminal based on LSTM, characterized in that, Includes the following steps: Data acquisition and operating condition classification: The operating condition data of the transformer substations uploaded by the smart fusion terminal is refined and classified according to the operating status OpSt of the photovoltaic equipment; Data preprocessing: outlier detection, preliminary missing value imputation, filtering and denoising, and normalization are performed on the data after the work conditions are divided. Training of models for different working conditions: For each type of working condition, a corresponding LSTM data-filling model is constructed and trained using preprocessed data; Missing value classification and imputation: For the data to be imputed, the working condition category to which the data to be imputed belongs is calculated based on similarity, and the LSTM data imputation model trained under the corresponding working condition is called to predict and impute the missing values.

2. The method according to claim 1, characterized in that, The data for the distribution area includes transformer data and photovoltaic data. The photovoltaic data includes the voltage, current, active power, and reactive power of the photovoltaic equipment.

3. The method according to claim 1, characterized in that, The data acquisition and operating condition classification include: Based on the photovoltaic equipment operating status OpSt, the data of the distribution area is initially divided into photovoltaic operation period data and shutdown period data. The photovoltaic operation period data includes photovoltaic data and transformer data, while the shutdown period data only includes transformer data. K-means clustering was applied to the photovoltaic operation period data and the shutdown period data respectively to further refine them into different operating condition categories.

4. The method according to claim 3, characterized in that, The K-means clustering method was applied to the photovoltaic operation period data and the shutdown period data respectively, further refining them into different operating condition categories, including: After determining the number of clusters K1 and K2 for the photovoltaic operation period data and the shutdown period data, K-means clustering was used to divide the dataset M under different operating conditions into: （1）; （2）; （3）; Where M 1,j M represents the dataset corresponding to operating condition j during the photovoltaic operation period. 2,m Let K1 represent the time set corresponding to the operating condition m during the downtime, where K1≥j≥1 and K2≥m≥1.

5. The method according to claim 4, characterized in that, The number of clusters K1 and K2 is determined by the elbow rule. With the goal of minimizing the sum of squared errors (SSE), the optimal number of clusters for K-means is found at the point where the rate of decrease of SSE slows down significantly by plotting the relationship between SSE and K.

6. The method according to claim 1, characterized in that, The filtering and denoising in the data preprocessing uses the Savitzky-Golay filtering algorithm.

7. The method according to claim 5, characterized in that, In the training of the work condition-specific model, for each type of work condition, a corresponding LSTM data-filling model is constructed and trained using preprocessed data, specifically including: The input matrix for the photovoltaic operation period is as follows: （4）; The input matrix for the downtime period is: （5）; Where U, I, P, and Q represent the load voltage, current, active power, and reactive power of phase A of the transformer, respectively, and U', I', P', and Q' represent the voltage, current, active power, and reactive power output by the inverter on the photovoltaic equipment side, respectively. For each working condition M i,j If M i,j Given N time series, each of length T, the nth time series is represented as: （6）; Where i = 1, 2, j = 1, 2, ..., K i , Let be the input vector at time t. For each time step t, the LSTM state update formula is: （7）; Among them, h t It is the hidden state of the LSTM at time t, x t h is the current input vector. t-1 It is the hidden state of the previous time step, c t-1 Let f represent the cell state at the previous time step, and let f denote the nonlinear mapping function of the LSTM. For each time step t, the output of the LSTM is: （8）; Wo is the weight matrix of the output layer, b o The bias vector for the output layer, for operating condition M i,j The filled results for all time series are as follows: （9）; Among them, M̂ i,j Let x̂ represent the dataset under the j-th working condition in the i-th time period after padding. N (i,j) Indicates the filled M̂ i,j The Nth filler value in the dataset.

8. The method according to claim 7, characterized in that, The classification and imputation of missing values ​​specifically includes: Define working condition M i,j The characteristic mean is: （10）; in, It is working condition M i,j The mean of the k-th feature, N i,j It is working condition M i,j The number of samples in It is the k-th feature value of the n-th sample; Calculate the feature mean of data within the same timestamp of the missing values, and calculate the other data for the missing points and the working condition M. i,j Euclidean distance: （11）; x 缺失 The missing part is replaced by the mean of the first non-missing value before and after it, and the result is obtained by comparing d for all working conditions. i,j Value, select the smallest d i,j The missing values ​​are assigned to the corresponding working conditions, and the missing values ​​are filled in using the LSTM data filling model trained for the corresponding working conditions.

Images