Intelligent prediction method for line loss of regional power transmission line

Abstract

The invention discloses an intelligent prediction method for line loss of a regional power transmission line. The method comprises the following steps: sequentially carrying out data preprocessing of cleaning, preprocessing, clustering analysis and coding on acquired data; an EMD method is adopted to decompose a non-stationary line loss time sequence in the data set, IMF components are reserved, an automatic singular spectrum analysis Sliding SSA method is adopted to decompose a stationary line loss time sequence in the data set, and main trend components are reserved; the discrete IMF component and the discrete main trend component in the line loss data are subjected to CountEncoder feature coding; analyzing the training set through the improved LSTM model to obtain a final model trained by the improved LSTM model; evaluating the performance of the model in the power transmission line loss prediction according to the prediction value and the true value error; according to the method, signal decomposition is carried out through EMD and automatic singular spectrum analysis Sliding SSA, the accuracy of preprocessing the line loss data of the regional power transmission line is improved, and the accuracy of model prediction is improved through stacking improvement of the LSTM model and introduction of an attention mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence power transmission line loss prediction, and more particularly, to an intelligent prediction method for regional power transmission line loss. Background Art

[0002] During the transmission of electric energy in the power grid, a part of active power loss will be generated due to factors such as impedance through various power components. This part of the loss is the line loss. The ratio of the line loss to the power supply is the line loss rate. The line loss is one of the important indicators of the power grid. Accurate calculation of the theoretical line loss can provide data support for the refined management of power grid enterprises and the formulation of energy-saving and loss-reduction measures, helping to achieve the "dual carbon" goal and respond to the call for energy conservation, emission reduction and green development in China. With the development of the power system, the structure of the transmission line is becoming increasingly complex and the amount of power grid data is getting larger and larger. Therefore, it is of great significance to propose a more accurate line loss prediction method.

[0003] The research on the prediction of power transmission line loss can be traced back to the 1930s. Researchers in various countries have proposed a variety of methods for predicting the line loss of distribution and transmission lines. So far, the methods for calculating the line loss of transmission lines can be divided into two categories. One is the traditional theoretical calculation methods and related improved algorithms that have been widely used based on physical models, including the root mean square current method, equivalent resistance method, average current method, power flow method, etc. These methods are based on the physical model of the power grid, and the required data includes current, voltage, line resistance, distribution network topology, etc. The other is the prediction methods proposed by using the learning ability of machine learning for large-scale data, such as prediction methods based on support vector machines, neural networks, heuristic algorithms, etc. The line loss prediction model is obtained through learning and training of a large amount of data.

[0004] The result of line loss prediction is as follows. First, from the perspective of traditional line loss prediction methods based on physical models, loss formula calculation methods such as the equivalent resistance method have simple formulas and have been used by power supply enterprises for many years. However, there are many simplified and estimated parts in them, resulting in the problem that the accuracy of line loss prediction is not high and there is a certain gap with the actual situation, which affects the management level and loss reduction efficiency of power supply enterprises. Among traditional algorithms, the power flow prediction method has relatively higher accuracy, but it has the problem of high requirements for the quality of original data. In addition to load power, power flow calculation also requires knowledge of the network topology and line resistance. However, it is often difficult to collect this information for low-voltage transmission lines due to their complex structures, making the applicable voltage level of this method limited. In recent years, research on line loss prediction methods based on machine learning has emerged, but there is not much research, and more in-depth research is urgently needed in this regard. Among the current machine learning methods for line loss prediction of transmission lines, the standard BP neural network model is a relatively mature main calculation method, but it has the problem of lack of effective adjustment of hidden layer weights and thresholds, leaving room for further improvement in the prediction accuracy of transmission line losses. In addition, through machine learning methods, the quality of data preprocessing is uneven, and the decomposition of data signals is not comprehensively and accurately processed, resulting in low accuracy of the line loss prediction model. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent prediction method for line losses of regional transmission lines to solve the above problems existing in the prior art.

[0006] Specifically, this application is as follows:

[0007] An intelligent prediction method for line losses of regional transmission lines includes the following steps:

[0008] S1. Obtain data related to load factor, meteorological data, and power data;

[0009] S2. Perform data preprocessing on the obtained data in sequence, including cleaning, preprocessing, clustering analysis, and encoding;

[0010] S3. Use the EMD method to decompose the non-stationary line loss time series in the dataset, and retain the IMF components, where the IMF components include discrete IMF components and continuous IMF components; use the automatic singular spectrum analysis Sliding SSA method to decompose the stationary line loss time series in the dataset, and retain the main trend components, where the main trend components include discrete main trend components and continuous main trend components;

[0011] S4. Perform CountEncoder feature encoding on the discrete IMF components and discrete main trend components, so that the component data is converted into numbers between 1 and n. Perform mean normalization encoding on the continuous IMF components and continuous main trend components, and use the window sliding method to perform time series window processing on the encoded data, construct the dataset required for training the neural network model LSTM, and divide the dataset into a training set and a test set. The dataset includes a load rate sample set, a meteorological data sample set, and an electricity quantity data sample set.

[0012] S5. Construct an improved LSTM model. The improved LSTM model uses a long short-term memory network and is based on an attention mechanism. Analyze the training set through the improved LSTM model to obtain the final model after training the improved LSTM model.

[0013] S6. Input the test set into the improved LSTM model to obtain the predicted values of the final model. Evaluate the performance of the model in predicting the transmission line loss according to the predicted values, true values, combined with the predicted fluctuation values, predicted error values, and predicted error ratios.

[0014] Further, the meteorological data in S1 includes temperature, humidity, and wind speed; the electricity quantity data includes active output electricity quantity, active input electricity quantity, reactive output electricity quantity, and reactive input electricity quantity; the load rate represents the ratio of the actual active output electricity quantity to the maximum active output electricity quantity; the load rate, meteorological data, and electricity quantity data all include historical data and real-time data.

[0015] Further, the specific implementation process of the data preprocessing work of performing clustering analysis on the obtained data in S2 is as follows:

[0016] S21. Perform initial clustering on the data related to the load rate, meteorological data, and electricity quantity data through the Canopy clustering algorithm to generate K clustering centers.

[0017] S22. Use the K clustering centers as the initial centers of the k-means clustering algorithm to perform fine clustering on the data related to the load rate, meteorological data, and electricity quantity data, and divide the data into two categories: stationary line loss time series data and non-stationary line loss time series data. After classifying the data through the k-means clustering algorithm, take the classification with the densest distribution and the highest score after classification as the stationary line loss time series data, and take all the remaining classifications as non-stationary line loss time series data.

[0018] Further, the specific implementation process of S21 is as follows:

[0019] S211. Set the data set List, set the thresholds T1 and T2, and T1>T2.

[0020] S212. Arbitrarily select a data point A, remove data point A from the List, and use data point A as a Canopy set;

[0021] S213. Calculate the distance d from all points in the set List to data point A;

[0022] S214. If T2 < d < T1, then use the corresponding data point as a weakly associated point of data point A;

[0023] S215. If d < T2, then use the data point as a strongly associated point of data point A, and remove this data point from the set List and no longer select it as the center point;

[0024] S216. Finally, obtain K clustering center points. Traverse the clusters where the K clustering center points are located. If the number of points in a certain cluster is less than a certain number, it is considered that there are noise points here, and directly delete this center point and the cluster where it is located.

[0025] Furthermore, the specific implementation process of S22 is as follows:

[0026] S221. Set the K clustering center points obtained by the Canopy clustering algorithm as the number of clustering clusters K;

[0027] S222. Calculate the Euclidean distance from all points to each centroid;

[0028] S223. Clustering division. For each sample, assign it to the cluster closest to the clustering centroid. The formula is:

[0029]

[0030] where, X i represents the i-th sample point, represents the clustering center of the k-th cluster at the t-th clustering, represents the cluster to which the i-th sample point belongs at the t-th clustering;

[0031] S224. Define a cost function and calculate the average clustering error:

[0032]

[0033] where, X i represents the i-th sample point, represents the clustering center of the cluster to which the i-th sample point belongs, J(c,a) represents the clustering mean square error, and N cluster represents the number of clustering samples;

[0034] S225. Update the clustering center, calculate the average value of each clustering, and use it as the new clustering center;

[0035] S226. Recalculate the Euclidean distances from all points to each new cluster center and perform clustering division;

[0036] S227. Calculate the clustering average error again;

[0037] S228. Verify whether the two clustering average errors are the same or whether the two conditions of the iteration times are met. If one of the two conditions is satisfied, end and output the result; otherwise, return to S225.

[0038] Furthermore, in S3, the EMD method is used to decompose the non-stationary line loss time series in the dataset, and the specific implementation process of retaining the IMF components is as follows:

[0039] S31. Find all local maximum points and local minimum points of the non-stationary original line loss time series X(t);

[0040] S32. On the basis of finding all local maximum points and local minimum points, use the cubic spline interpolation method to interpolate the local maximum points to obtain the upper envelope line U(t) of the original time series, and interpolate the local minimum points to obtain the lower envelope line L(t) of the original time series;

[0041] S33. Calculate the mean of the upper and lower envelope lines:

[0042] S34. Subtract the mean m(t) of the upper and lower envelope lines from the original time series to obtain the residual series h(t) = X(t) - m(t);

[0043] S35. Preset the IMF threshold. If the residual series h(t) meets the IMF threshold, then this residual series is an IMF of the original time series, denoted as C(t); otherwise, repeat S31 - S34 for this residual series until the preset IMF threshold is met, and use the remaining part R(t) = X(t) - C(t) to replace the original time series X(t);

[0044] S36. Repeat steps S31 - S35 for R(t) until the finally obtained residual series is a monotonic function, and then end the decomposition process.

[0045] Furthermore, in S3, the specific implementation process of using the Automatic Singular Spectrum Analysis Sliding SSA method to decompose the stationary line loss time series in the dataset is as follows:

[0046] S37. Construct a trajectory matrix X for the stationary original time series and perform singular value decomposition on it;

[0047] S38. Perform diagonal averaging on each component obtained after singular value decomposition to obtain a Hankel matrix;

[0048] S39. Perform hierarchical clustering on each of the Hankel matrices obtained in step S38 to obtain different groupings;

[0049] S40. Add the components in each of the groupings obtained in step S39 to obtain physically interpretable components.

[0050] Further, the specific implementation process of S37 includes:

[0051] S371. Establish a trajectory matrix. Given a time series S(t) = {s1, s1,..., s n}, use a sliding window of length L to take values from this time series. Starting from s1, every time L data are taken, they are used as a column of the matrix, and then the window is shifted one position to the right, repeating the value-taking operation until the right end of the window reaches the end of the time series. The finally obtained matrix is:

[0052]

[0053] Among them, the matrix X is a Hankel matrix with L rows and K = N - L + 1 columns;

[0054] S372. Singular value decomposition. Decompose the matrix X into the product of a singular matrix and a diagonal matrix composed of singular values. The calculation formula for singular value decomposition is:

[0055]

[0056] Among them, R = min[L, K] is the rank of the matrix X, Σ is a diagonal matrix composed of singular values σ i (i = 1, 2,..., R), X i are the matrices obtained after singular value decomposition, and there is: Among them, u i and v i respectively represent the vectors of the i-th column in the left singular matrix U and the right singular matrix V.

[0057] Further, the specific structure and steps of improving the LSTM model in S4 include:

[0058] S41. The improved LSTM model is a three-layer stacked LSTM-attention network. Its improvement lies in adopting a three-layer LSTM network architecture. The output of the first layer of LSTM is connected to the input of the second layer of LSTM, and the output of the second layer of LSTM is connected to the input of the third layer of LSTM. And an attention mechanism is used to weight and merge the hidden state time series of the first layer of LSTM, the second layer of LSTM, and the third layer of LSTM to obtain a weighted hidden state time series. Finally, a fully connected layer is added to capture the key information of the weighted hidden state time series and map it to the space of the final transmission line power loss prediction;

[0059] S42. In the first - layer LSTM, use the load - rate sample set obtained in S4 as the training set and input it into the first - layer LSTM to obtain the hidden - state time series T1 of the first - layer LSTM, T1={T1, T2,... T n}, T n represents the hidden state of the nth time series of the first - layer LSTM, and n is the length of the hidden - state time series;

[0060] S43. In the second - layer LSTM, use the meteorological - data sample set training set obtained in S4 and input it into the second - layer LSTM to obtain the hidden - state time series T2 of the second - layer LSTM, T2={T'1, T'2,... T' n}, T2 represents the hidden state of the nth time series of the second - layer LSTM;

[0061] S44. In the third - layer LSTM, use the power - consumption data sample set training set obtained in S4 and input it into the third - layer LSTM to obtain the hidden - state time series T3 of the third - layer LSTM, and T3 represents the hidden state of the nth time series of the third - layer LSTM;

[0062] S45. In the attention - mechanism layer, use the attention mechanism to perform weighted combination on the hidden - state time series Q1 of the first - layer LSTM, the hidden - state time series Q2 of the second - layer LSTM, and the hidden - state time series Q3 of the third - layer LSTM;

[0063] First, calculate the attention scores of the hidden - state time series of the three - layer LSTM, including two steps: calculating the weighted values and calculating the attention scores;

[0064] The formula for calculating the weighted values is as follows:

[0065] k1 = w1 * q1 + c1;

[0066] k2 = w2 * q2 + c2;

[0067] k3 = w3 * q3 + c3;

[0068] where k1, k2, and k3 are the weighted values corresponding to the first - layer LSTM, the second - layer LSTM, and the third - layer LSTM respectively, q1, q2, and q3 are the hidden states of the time series corresponding to the first - layer LSTM, the second - layer LSTM, and the third - layer LSTM respectively, c1, c2, and c3 represent the biases of the first - layer LSTM, the second - layer LSTM, and the third - layer LSTM respectively, and w1, w2, and w3 are the weighted - value coefficients of the first - layer LSTM, the second - layer LSTM, and the third - layer LSTM respectively;

[0069] The formula for calculating the attention scores is as follows:

[0070] f1 = tanh(k1);

[0071] f2 = tanh(k2);

[0072] f3 = tanh(k3);

[0073] Among them, f1, f2, and f3 are the attention scores corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively;

[0074] Secondly, it is to use the attention scores to weight and fuse the hidden state time series T' of the three-layer LSTM and obtain the final weighted hidden state time series. The calculation formula is as follows:

[0075] T' = f1 * T1 + f2 * T2 + f3 * T3.

[0076] Furthermore, the evaluation of the performance of the model in the transmission line loss prediction according to the predicted value and the true value combined with the predicted fluctuation value, the predicted error value, and the predicted error ratio in S6 includes:

[0077] Evaluating the difference between the actual line loss value and the predicted line loss value of the transmission line loss, obtaining the predicted line loss value predicted by the improved LSTM model, calculating the actual line loss value of the transmission line and the predicted line loss value of the improved LSTM model, and using the formula: Get the predicted fluctuation value QS; where P t represents the actual line loss value at the t-th time, represents the predicted line loss value predicted at the t-th time, t represents the time index subscript predicted by the improved LSTM model, and N represents the total number of time index subscripts predicted by the improved LSTM model in the record; calculate the average of the absolute value of the difference between the predicted line loss value and the actual line loss value, using the formula: Get the predicted error value QA; calculate the ratio of the predicted line loss value to the actual line loss value, using the formula:

[0078] Get the predicted error ratio QP.

[0079] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0080] The embodiments of the present invention provide a method for obtaining data related to load rate, meteorological data, and power consumption data; performing data preprocessing on the obtained data in sequence, including cleaning, preprocessing, clustering analysis, and encoding; decomposing the non-stationary line loss time series in the dataset using the EMD method, retaining the IMF components, and decomposing the stationary line loss time series in the dataset using the Automatic Singular Spectrum Analysis Sliding SSA method, retaining the main trend components; performing CountEncoder feature encoding on the discrete IMF components and discrete main trend components in the line loss data, converting the component data into numbers between 1 and n, and performing mean normalization encoding on the remaining continuous IMF component data and continuous main trend components. Then, using the window sliding method to perform time series window processing on the encoded data, constructing the dataset required for training the neural network model LSTM, and dividing the dataset into a training set and a test set; constructing an improved LSTM model based on the attention mechanism; analyzing the training set through the improved LSTM model to obtain the final model after training the improved LSTM model; inputting the test set into the improved LSTM model to obtain the predicted values of the final model; evaluating the performance of the model in predicting the line loss of transmission lines according to the predicted values, true values, combined with the predicted fluctuation values, predicted error values, and predicted error ratios; the present invention decomposes signals through EMD and Automatic Singular Spectrum Analysis Sliding SSA, improves the accuracy of data preprocessing for regional transmission line line losses, and improves the accuracy of model prediction through the stacking improvement of the LSTM model and the introduction of the attention mechanism. Description of the Drawings

[0081] Figure 1 is a schematic flow chart of an intelligent prediction method for regional transmission line line losses provided by an embodiment of the present invention;

[0082] Figure 2 is a schematic flow chart of the Canopy algorithm of an intelligent prediction method for regional transmission line line losses provided by an embodiment of the present invention;

[0083] Figure 3 is a schematic flow chart of the k-means algorithm of an intelligent prediction method for regional transmission line line losses provided by an embodiment of the present invention;

[0084] Figure 4 is a diagram of the IMF components decomposed by the EMD method of an intelligent prediction method for regional transmission line line losses provided by an embodiment of the present invention;

[0085] Figure 5 is a diagram of the components decomposed by the Sliding SSA method of an intelligent prediction method for regional transmission line line losses provided by an embodiment of the present invention;

[0086] Figure 6It is a comparison graph of the predicted values and the true values of the improved LSTM model for an intelligent prediction method of regional transmission line loss provided by an embodiment of the present invention;

[0087] Figure 7 It is a comparison graph before and after using EMD and Sliding SSA of the improved LSTM model for an intelligent prediction method of regional transmission line loss provided by an embodiment of the present invention. Specific implementation manners

[0088] The present invention will be described in detail below with reference to the accompanying drawings.

[0089] Embodiment 1

[0090] An embodiment of the present invention provides an intelligent prediction method for regional transmission line loss, such as Figure 1 , including the following steps:

[0091] An intelligent prediction method for regional transmission line loss, such as Figure 1 shown, including the following steps:

[0092] S1. Obtain data related to load factor, meteorological data, and power consumption data;

[0093] S2. Perform data preprocessing on the obtained data in sequence, including cleaning, preprocessing, clustering analysis, and encoding;

[0094] S3. Use the EMD method to decompose the non-stationary line loss time series in the dataset, and retain the IMF components, where the IMF components include discrete IMF components and continuous IMF components; use the automatic singular spectrum analysis Sliding SSA method to decompose the stationary line loss time series in the dataset, and retain the main trend components, where the main trend components include discrete main trend components and continuous main trend components;

[0095] S4. Perform CountEncoder feature encoding on the discrete IMF components and discrete main trend components, so that the component data is converted into numbers between 1 and n. The continuous IMF components and continuous main trend components are subjected to mean normalization encoding, and the encoded data is processed by a window sliding method for time series window processing to construct a dataset required for training the neural network model LSTM, and the dataset is divided into a training set and a test set. The dataset includes a load factor sample set, a meteorological data sample set, and a power consumption data sample set;

[0096] S5. Construct an improved LSTM model, where the improved LSTM model uses a long short-term memory network and is based on an attention mechanism; analyze the training set through the improved LSTM model to obtain the final model after training the improved LSTM model;

[0097] S6. Input the test set into the improved LSTM model to obtain the predicted values of the final model; evaluate the performance of the model in predicting the transmission line loss based on the predicted values, true values, combined with the predicted fluctuation value, predicted error value, and predicted error ratio.

[0098] Specifically, obtain data related to load factor, meteorological data, and power consumption data; perform data preprocessing on the obtained data in sequence, including cleaning, preprocessing, clustering analysis, and encoding; use the EMD method to decompose the non-stationary line loss time series in the dataset, retain the IMF components, and use the automatic singular spectrum analysis Sliding SSA method to decompose the stationary line loss time series in the dataset, retain the main trend components; perform CountEncoder feature encoding on the discrete IMF components and discrete main trend components in the line loss data to convert the component data into numbers between 1 and n, and perform mean normalization encoding on the remaining continuous IMF component data and continuous main trend components, and use the window sliding method to perform time series window processing on the encoded data to construct the dataset required for training the neural network model LSTM, and divide the dataset into a training set and a test set; construct an improved LSTM model based on the attention mechanism; analyze the training set through the improved LSTM model to obtain the final model after training the improved LSTM model; input the test set into the improved LSTM model to obtain the predicted values of the final model; evaluate the performance of the model in predicting the transmission line loss based on the predicted values, true values, combined with the predicted fluctuation value, predicted error value, and predicted error ratio; through signal decomposition by EMD and automatic singular spectrum analysis Sliding SSA, improve the accuracy of data preprocessing for regional transmission line losses, and through the stacking improvement of the LSTM model and the introduction of the attention mechanism, improve the accuracy of model prediction.

[0099] In the above embodiment, specifically, the meteorological data in S1 includes temperature, humidity, and wind speed; the power consumption data includes active output power, active input power, reactive output power, and reactive input power; the load factor represents the ratio of the actual active output power to the maximum active output power; the load factor, meteorological data, and power consumption data all include historical data and real-time data.

[0100] In the above embodiment, specifically, the specific implementation process of the data preprocessing work of performing clustering analysis on the obtained data in S2 is as follows:

[0101] S21. Perform initial clustering on the data related to load factor, meteorological data, and power consumption data through the Canopy clustering algorithm to generate K clustering centers.

[0102] S22. Use the K cluster centers as the initial centers of the k-means clustering algorithm to perform fine clustering on the data related to the load rate, meteorological data, and power consumption data, and divide the data into two categories: stable line loss time series data and unstable line loss time series data. After classifying the data through the k-means clustering algorithm, select the category with the densest distribution and the highest score after classification as the stable line loss time series data, and regard all the remaining classifications as the unstable line loss time series data.

[0103] In the above embodiment, specifically, the specific implementation process of S21 is as follows:

[0104] S211. Set the data set List, set the thresholds T1 and T2, and T1 > T2;

[0105] S212. Arbitrarily select a data point A, remove the data point A from List, and use the data point A as a Canopy set;

[0106] S213. Calculate the distance d from all points in the set List to the data point A;

[0107] S214. If T2 < d < T1, then regard the corresponding data point as a weak associated point of the data point A;

[0108] S215. If d < T2, then regard the data point as a strong associated point of the data point A, and remove this data point from the set List and no longer select it as a center point;

[0109] S216. Finally, obtain K cluster centers, traverse the clusters where the K cluster centers are located. If the number of points in a certain cluster is less than a certain number, it is considered that there are noise points here, and directly delete this center point and its corresponding cluster.

[0110] Furthermore, the specific implementation process of S22 is as follows, as Figure 3 shown:

[0111] S221. Set the K cluster centers obtained by the Canopy clustering algorithm as the number of clustering clusters K;

[0112] S222. Calculate the Euclidean distance from all points to each centroid;

[0113] S223. Clustering division. For each sample, assign it to the cluster closest to the clustering centroid. The formula is:

[0114]

[0115] where, X i represents the i-th sample point, represents the clustering center of the k-th cluster at the t-th clustering, Denote the cluster to which the \(i\)-th sample point belongs in the \(t\)-th clustering.

[0116] S224. Define a cost function and calculate the average clustering error:

[0117]

[0118] Among them, \(X\) i denotes the \(i\)-th sample point, denotes the clustering center of the cluster to which the \(i\)-th sample point belongs, \(J(c,a)\) represents the clustering mean square error, and \(N\) cluster denotes the number of clustering samples;

[0119] S225. Update the clustering center, calculate the average value of each cluster, and use it as the new clustering center;

[0120] S226. Recalculate the Euclidean distances from all points to the new clustering centers and perform clustering division;

[0121] S227. Calculate the average clustering error again;

[0122] S228. Verify whether the two conditions that the two average clustering errors are the same or have reached the number of iterations are met. If one of the two conditions is satisfied, end and output the result; otherwise, return to S225.

[0123] Specifically, in this embodiment, the Canopy clustering algorithm is a fast clustering algorithm. Compared with traditional clustering algorithms, the Canopy clustering algorithm does not need to specify the number of clustering clusters in advance. It can obtain the number of clustering clusters with only one pass through the data. Although the Canopy clustering accuracy is relatively low, it can obtain the optimal number of clustering clusters and has a very fast running speed. Therefore, the Canopy clustering algorithm has high practical value, especially suitable for being used in combination with the k-means clustering algorithm. In clustering algorithms, the most cumbersome step is to set the similarity between data, while the Canopy algorithm chooses a simple method to calculate the similarity between data and puts similar data into a set, which is the Canopy. There will be several clustering circles in the Canopy algorithm and they can overlap. Data points will not exist alone outside the Canopy. Canopy clustering refers to dividing the center points of data. Draw small circles with the Canopy as the center and a radius of \(2T\), and draw large circles with a radius of \(1T\). The data points inside the small circles belong to the Canopy, and the data points outside the small circles and inside the large circles can be used as new Canopies to divide the data. Canopy clustering does not need to specify the number of clusters in advance, but depends on the similarity between data points to obtain multiple different clustering clusters. The Canopy algorithm process in this embodiment is as Figure 2As shown in the figure, using the Canopy algorithm to perform rough clustering before the k-means fine clustering can take the number of Canopies as the preset k value for k-means clustering, and the Canopy clustering reduces the number of similarity data that needs to be calculated subsequently, improving the calculation efficiency and enhancing the clustering effect at the same time.

[0124] In the above embodiment, specifically, in S3, the EMD method is used to decompose the non-stationary line loss time series in the dataset, and the specific implementation process of retaining the IMF components is as follows:

[0125] S31. Find all local maximum points and local minimum points of the original non-stationary line loss time series X(t);

[0126] S32. On the basis of finding all local maximum points and local minimum points, use the cubic spline interpolation method to interpolate the local maximum points to obtain the upper envelope U(t) of the original time series, and interpolate the local minimum points to obtain the lower envelope L(t) of the original time series;

[0127] S33. Calculate the mean of the upper and lower envelopes:

[0128] S34. Subtract the mean m(t) of the upper and lower envelopes from the original time series to obtain the residual series h(t) = X(t) - m(t);

[0129] S35. Preset the IMF threshold. If the residual series h(t) meets the IMF threshold, then this residual series is an IMF of the original time series, denoted as C(t). Otherwise, repeat S31 - S34 for this residual series until the preset IMF threshold is met, and use the remaining part R(t) = X(t) - C(t) to replace the original time series X(t);

[0130] S36. Repeat steps S31 - S35 for R(t) until the finally obtained residual series is a monotonic function, and then end the decomposition process.

[0131] Specifically, the original time series is finally decomposed into multiple IMFs and a monotonic residual series, expressed as

[0132]

[0133] where n represents the number of decomposed IMFs. As shown in Figure 4 the figure, C i (t) represents the i-th IMF, and C1(t), C2(t),..., C n(t) varies from high to low in frequency, and r(t) represents the monotonic residual sequence. The EMD method can retain as much information of the original time series as possible and reduce data loss. In this embodiment, after the original time series is decomposed, a series of components with different frequencies are obtained. Therefore, the LSTM neural network model can learn the components of each frequency after decomposition in a targeted manner, improve the fineness of feature extraction, and the information loss rate of this decomposition method is also relatively low, thereby improving the prediction accuracy of the LSTM neural network model.

[0134] In the above embodiment, specifically, the specific implementation process of using the Automatic Singular Spectrum Analysis Sliding SSA method to decompose the stationary line loss time series in the dataset in S3 is as follows:

[0135] S37. Construct a trajectory matrix X for the original stationary line loss time series and perform singular value decomposition on it;

[0136] S38. Perform diagonal averaging on each component obtained after singular value decomposition to obtain a Hankel matrix;

[0137] S39. Perform hierarchical clustering on each Hankel matrix obtained in step S38 to obtain different groups;

[0138] S40. Add the components in each group obtained in step S39 to obtain physically interpretable components, as Figure 5 shown.

[0139] In the above embodiment, specifically, the specific implementation process of S37 includes:

[0140] S371. Establish a trajectory matrix. Given a time series S(t) = {s1, s1,..., s n}, use a sliding window of length L to take values from this time series. Starting from s1, every time L data are taken, they are used as a column of the matrix, and then the window is shifted one position to the right, repeating the value-taking operation until the right end of the window reaches the end of the time series. The finally obtained matrix is:

[0141]

[0142] where the matrix X is a Hankel matrix with L rows and K = N - L + 1 columns;

[0143] S372. Singular value decomposition. Decompose the matrix X into the product of a singular matrix and a diagonal matrix composed of singular values. The calculation formula for singular value decomposition is:

[0144]

[0145] where R = min[L, K] is the rank of the matrix X, and Σ is composed of singular values σi a diagonal matrix composed of (i = 1, 2,..., R), X i are the matrices obtained after singular value decomposition, and there are: where u i and v i respectively represent the vectors of the i-th column in the left singular matrix U and the right singular matrix V.

[0146] Specifically, Sliding SSA decomposes the original time series into a series of components with different frequencies and physical interpretability. Therefore, it enables the LSTM neural network model to learn each decomposed component specifically, improving the fineness of feature extraction, and the information loss rate of this decomposition method is also relatively low, thus improving the prediction accuracy of the neural network.

[0147] In the above embodiment, specifically, the specific structure and steps of improving the LSTM model in S4 include:

[0148] S41. The improved LSTM model is a three-layer stacked LSTM-attention network. Its improvement lies in adopting a three-layer LSTM network architecture. The output of the first-layer LSTM is connected to the input of the second-layer LSTM, and the output of the second-layer LSTM is connected to the input of the third-layer LSTM. And an attention mechanism is used to perform weighted merging on the hidden state time series of the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM to obtain a weighted hidden state time series. Finally, a fully connected layer is added to capture the key information of the weighted hidden state time series and map it to the space of the final transmission line loss prediction.

[0149] S42. In the first-layer LSTM, the load rate sample set obtained in S4 is used as the training set and input into the first-layer LSTM to obtain the hidden state time series T1 of the first-layer LSTM, T1 = {T1, T2,... T n}, T n represents the hidden state of the n-th time series of the first-layer LSTM, and n is the length of the hidden state time series.

[0150] S43. In the second-layer LSTM, the meteorological data sample set training set obtained in S4 is input into the second-layer LSTM to obtain the hidden state time series column T2 of the second-layer LSTM, T2 = {T'1, T'2,... T' n}, T2 represents the hidden state of the n-th time series of the second-layer LSTM.

[0151] S44. In the third-layer LSTM, the electricity quantity data sample set training set obtained in S4 is input into the third-layer LSTM to obtain the hidden state time series column T3 of the third-layer LSTM, and T3 represents the hidden state of the n-th time series of the third-layer LSTM.

[0152] S45. In the attention mechanism layer, the attention mechanism is used to perform weighted combination on the hidden state time series Q1 of the first-layer LSTM, the hidden state time series Q2 of the second-layer LSTM, and the hidden state time series Q3 of the third-layer LSTM;

[0153] First, calculate the attention scores of the hidden state time series of the three-layer LSTM, including two steps of calculating the weighted values and calculating the attention scores;

[0154] The formula for calculating the weighted value is as follows:

[0155] k1 = w1 * q1 + c1;

[0156] k2 = w2 * q2 + c2;

[0157] k3 = w3 * w3 + c3;

[0158] Among them, k1, k2, and k3 are the weighted values corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively, q1, q2, and q3 are the hidden states of the time series corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively, c1, c2, and c3 represent the biases of the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively, and w1, w2, and w3 are the weighted value coefficients of the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively;

[0159] The formula for calculating the attention score is as follows:

[0160] f1 = tanh(k1);

[0161] f2 = tanh(k2);

[0162] f3 = tanh(k3);

[0163] Among them, f1, f2, and f3 are the attention scores corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively;

[0164] Secondly, use the attention scores to perform weighted fusion on the hidden state time series T' of the three-layer LSTM and obtain the final weighted hidden state time series. The calculation formula is as follows:

[0165] t' = f1 * t1 + f2 * t2 + f3 * T3.

[0166] In the above embodiment, specifically, the evaluation of the performance of the model in the transmission line loss prediction according to the predicted value, the true value, the predicted fluctuation value, the predicted error value, and the predicted error ratio in S6 includes:

[0167] Evaluate the difference between the actual line loss value and the predicted line loss value of the transmission line, obtain the predicted line loss value predicted by the improved LSTM model, calculate the actual line loss value of the transmission line and the predicted line loss value of the improved LSTM model, and use the formula: to obtain the predicted fluctuation value QS; where P t represents the actual line loss value at the t-th time, represents the predicted line loss value predicted at the t-th time, t represents the time index subscript predicted by the improved LSTM model, and N represents the total number of time index subscripts predicted by the improved LSTM model in the record; calculate the average of the absolute value of the difference between the predicted line loss value and the actual line loss value, and use the formula: to obtain the predicted error value QA; calculate the ratio of the predicted line loss value to the actual line loss value, and use the formula:

[0168] to obtain the predicted error ratio QP.

[0169] Specifically, as Figure 7 shown, the left figure shows that the data used by the improved LSTM model has not been processed by the Sliding SSA method and the EMD method, and the right figure shows that the data used by the improved LSTM model has been processed by the Sliding SSA method and the EMD method. After comparison, it is found that using the data preprocessed by the Sliding SSA method and the EMD method can effectively solve the problems of uneven distribution and large fluctuation of the predicted value of the line loss rate, and improve the prediction accuracy.

[0170] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The structures required to construct such systems will be apparent from the above description. In addition, the present invention is not directed to any particular programming language. It should be understood that the content of the present invention described herein can be implemented using various programming languages, and the description of the specific language above is for the purpose of disclosing the best mode of the present invention.

[0171] In the specification provided herein, a large number of specific details are set forth. However, it can be understood that the embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0172] Similarly, it should be understood that, for the purpose of streamlining the present disclosure and facilitating the understanding of one or more of the various inventive aspects, in the foregoing description of the exemplary embodiments of the present invention, the various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, the disclosed method should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as reflected in the following claims, the inventive aspects lie in less than all of the features of the single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate embodiment of the present invention.

[0173] Those skilled in the art can understand that the modules in the devices in the embodiments can be adaptively changed and disposed in one or more devices different from the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and in addition, they can be divided into multiple sub-modules or sub-units or sub-components. Except that at least some of such features and / or processes or units are mutually exclusive, any combination can be adopted to combine all the features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all the processes or units of any method or device so disclosed. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) can be replaced by an alternative feature that provides the same, equivalent, or similar purpose.

[0174] In addition, those skilled in the art can understand that although some of the embodiments herein include certain features included in other embodiments rather than other features, the combination of the features of different embodiments means that it is within the scope of the present invention and forms different embodiments. For example, in the following claims, any one of the claimed embodiments can be used in any combination.

[0175] The various component embodiments of the present invention can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) can be used in practice to implement some or all of the functions of some or all of the components in the device according to the embodiments of the present invention. The present invention can also be implemented as a device or device program (for example, a computer program and a computer program product) for executing part or all of the methods described herein. Such a program for implementing the present invention can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

Claims

1. An intelligent prediction method for line losses of regional transmission lines, characterized in that, It includes the following steps: S1. Obtain data related to load rate, meteorological data, and power consumption data; S2. Perform data preprocessing on the obtained data in sequence, including data cleaning, preprocessing, clustering analysis, and encoding; S3. Use the EMD method to decompose the non-stationary line loss time series in the dataset, and retain the IMF components, where the IMF components include discrete IMF components and continuous IMF components; use the automatic singular spectrum analysis Sliding SSA method to decompose the stationary line loss time series in the dataset, and retain the main trend components, where the main trend components include discrete main trend components and continuous main trend components; S4. Perform CountEncoder feature encoding on the discrete IMF components and discrete main trend components, so that the component data is converted into numbers between 1 and n. Perform mean normalization encoding on the continuous IMF components and continuous main trend components, and use the window sliding method to perform time series window processing on the encoded data to construct the dataset required for training the neural network model LSTM, and divide the dataset into a training set and a test set. The dataset includes a load rate sample set, a meteorological data sample set, and a power consumption data sample set; S5. Construct an improved LSTM model, which uses a long short-term memory network and is based on an attention mechanism; Analyze the training set through the improved LSTM model to obtain the final model after training the improved LSTM model; S6. Input the test set into the improved LSTM model to obtain the predicted values of the final model; Evaluate the performance of the model in predicting the line loss of transmission lines according to the predicted values, true values, combined with the predicted fluctuation values, predicted error values, and predicted error ratios.

2. The intelligent prediction method for line loss of regional transmission lines according to claim 1, characterized in that The meteorological data in S1 includes temperature, humidity, and wind speed; the power consumption data includes active output power, active input power, reactive output power, and reactive input power; the load rate represents the ratio of the actual active output power to the maximum active output power; the load rate, meteorological data, and power consumption data all include historical data and real-time data.

3. The intelligent prediction method for line loss of regional transmission lines according to claim 1, characterized in that The specific implementation process of the data preprocessing work of clustering analysis on the obtained data in S2 is as follows: S21. Perform initial clustering on the data related to load rate, meteorological data, and power consumption data through the Canopy clustering algorithm to generate K clustering centers; S22. Use the K clustering centers as the initial centers of the k-means clustering algorithm to perform fine clustering on the data related to load rate, meteorological data, and power consumption data, and divide the data into two categories: stationary line loss time series data and non-stationary line loss time series data; After classifying the data through the k-means clustering algorithm, classify the most densely distributed and highest-scoring classification as stationary line loss time series data, and classify all the remaining others as non-stationary line loss time series data.

4. The intelligent prediction method for line loss of regional transmission lines according to claim 3, wherein The specific implementation process of S21 is as follows: S211. Set the data set List, set the thresholds T1 and T2, and T1>T2; S212. Arbitrarily select a data point A, remove data point A from the List, and use data point A as a Canopy set; S213. Calculate the distance d from all points in the set List to data point A; S214. If T2 < d < T1, then use the corresponding data point as a weak associated point of data point A; S215. If d < T2, then use the data point as a strong associated point of data point A, and remove this data point from the set List and no longer select it as the center point; S216. Finally, obtain K clustering center points. Traverse the clusters where the K clustering center points are located. If the number of points in a certain cluster is less than a certain number, it is considered that there are noise points here, and directly delete this center point and its corresponding cluster.

5. The intelligent prediction method for line loss of regional transmission lines according to claim 3, characterized in that, The specific implementation process of S22 is as follows: S221. Set the K clustering center points obtained by the Canopy clustering algorithm as the number of clustering clusters K; S222. Calculate the Euclidean distance from all points to each centroid; S223. Clustering division. For each sample, allocate it to the cluster closest to the clustering centroid. The formula is: Among them, X i represents the i-th sample point, represents the clustering center of the k-th cluster at the t-th clustering, represents the cluster to which the i-th sample point belongs at the t-th clustering; S224. Define a cost function and calculate the clustering average error: Among them, X i represents the i-th sample point, represents the cluster center of the cluster to which the i-th sample point belongs, J(c,a) represents the clustering mean square error, and N cluster represents the number of clustering samples; S225. Update the clustering center, calculate the average value of each cluster, and use it as the new clustering center; S226. Recalculate the Euclidean distance from all points to each new clustering center and perform clustering division; S227. Calculate the clustering average error again; S228. Verify whether the two conditions that the two clustering average errors are the same or have reached the number of iterations are met. If one of the two conditions is satisfied, end and output the result, otherwise return to S225.

6. The intelligent prediction method for line loss of regional transmission lines according to claim 1, characterized in that In S3, the EMD method is used to decompose the non-stationary line loss time series in the dataset, and the specific implementation process of retaining the IMF component is as follows: S31. Find all local maximum points and local minimum points of the non-stationary line loss original time series X(t); S32. On the basis of finding all local maximum points and local minimum points, use the cubic spline interpolation method to interpolate the local maximum points to obtain the upper envelope line U(t) of the original time series, and interpolate the local minimum points to obtain the lower envelope line L(t) of the original time series; S33. Calculate the mean of the upper and lower envelopes: S34. Subtract the mean value m(t) of the upper and lower envelope lines from the original time series to obtain the residual series h(t) = X(t) - m(t); S35. Preset the IMF threshold. If the residual series h(t) meets the IMF threshold, then this residual series is an IMF of the original time series, denoted as C(t). Otherwise, repeat S31 - S34 for this residual series until the preset IMF threshold is met, and use the remaining part R(t) = X(t) - C(t) to replace the original time series X(t); S36. Repeat steps S31 - S35 for R(t) until the final obtained residual series is a monotonic function, and then end the decomposition process.

7. The intelligent prediction method for line loss of regional transmission lines according to claim 1, wherein The specific implementation process of using the automatic singular spectrum analysis Sliding SSA method in S3 to decompose the stationary line loss time series in the dataset is as follows: S37. Construct a trajectory matrix X for the stationary line loss original time series and perform singular value decomposition on it; S38. Diagonal average each component obtained from the singular value decomposition to obtain a Hankel matrix; S39. Perform hierarchical clustering on each Hankel matrix obtained in step S38 to obtain different groups; S40. Add the components in each group obtained in step S39 to obtain physically interpretable components.

8. The intelligent prediction method for line loss of regional transmission lines according to claim 7, characterized in that The specific implementation process of S37 includes: S371. Establish a trajectory matrix. Assume there is a time series S(t) = {s1, s1,..., s n}. Use a sliding window of length L to extract values from this time series. Starting from s1, every time L data points are taken, they are used as a column of the matrix. Then the window is shifted one position to the right, and the value extraction operation is repeated until the right end of the window reaches the end of the time series. The finally obtained matrix is: Among them, matrix X is a Hankel matrix with L rows and K = N - L + 1 columns; S372. Singular value decomposition, decompose matrix X into the product of a singular matrix and a diagonal matrix composed of singular values. The calculation formula for singular value decomposition is: where \(R = \min[L, K]\) is the rank of matrix \(X\), and \(\Sigma\) is a diagonal matrix composed of singular values \(\sigma\) i (\(i = 1, 2, \cdots, R\)), and \(X\) i are the matrices obtained after singular value decomposition, and we have: where \(u\) i and \(v\) i respectively represent the vectors of the \(i\)-th column in the left singular matrix \(U\) and the right singular matrix \(V\).

9. The intelligent prediction method for line loss of regional transmission lines according to claim 1, characterized in that The specific structure and steps of the improved LSTM model in S4 include: S41. The improved LSTM model is a three-layer stacked LSTM-attention network. Its improvement lies in adopting a three-layer LSTM network architecture. The output of the first-layer LSTM is connected to the input of the second-layer LSTM, and the output of the second-layer LSTM is connected to the input of the third-layer LSTM. An attention mechanism is used to weight and merge the hidden state time series of the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM to obtain a weighted hidden state time series. Finally, a fully connected layer is added to capture the key information of the weighted hidden state time series and map it to the space of the final transmission line loss prediction; S42. In the first-layer LSTM, use the load rate sample set obtained in S4 as the training set and input it into the first-layer LSTM to obtain the hidden state time series T1 of the first-layer LSTM, where T1 = {T1, T2,... T n}, and T n represents the hidden state of the nth time series of the first-layer LSTM, and n is the length of the hidden state time series; S43. In the second-layer LSTM, input the training set of the meteorological data sample set obtained in S4 into the second-layer LSTM to obtain the hidden state time series column T2 of the second-layer LSTM, where T2 = {T'1, T'2,... T' n}, and T2 represents the hidden state of the nth time series of the second-layer LSTM; S44. In the third-layer LSTM, input the power data sample set training set obtained in S4 into the third-layer LSTM to obtain the hidden state time series column T3 of the third-layer LSTM, where T3 represents the hidden state of the nth time series of the third-layer LSTM; S45. In the attention mechanism layer, use the attention mechanism to weight and merge the hidden state time series Q1 of the first-layer LSTM, the hidden state time series Q2 of the second-layer LSTM, and the hidden state time series Q3 of the third-layer LSTM; First, calculate the attention scores of the hidden state time series of the three-layer LSTM, including two steps: calculating the weighted value and calculating the attention score; The formula for calculating the weighted value is as follows: k1 = w1 * q1 + c1; k2 = w2 * q2 + c2; k3 = w3 * w3 + c3; Among them, k1, k2, and k3 are the weighted values corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively, q1, q2, and q3 are the hidden states of the time series corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively, c1, c2, and c3 represent the biases of the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively, and w1, w2, and w3 are the weighted value coefficients of the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively; The formula for calculating the attention score is as follows: f1 = tanh(k1); f2 = tanh(k2); f3 = tanh(k3); Among them, f1, f2, and f3 are the attention scores corresponding to the first-layer LSTM, the second-layer LSTM, and the third-layer LSTM respectively; Secondly, the time series T' of the hidden states of the three-layer LSTM is weighted and fused using the attention scores, and the final weighted time series of the hidden states is obtained. The calculation formula is as follows: T' = f1 * T1 + f2 * T2 + f3 * T3.

10. The intelligent prediction method for line loss of regional transmission lines according to claim 1, characterized in that The evaluation of the performance of the model in the transmission line loss prediction according to the predicted value and the true value by combining the predicted fluctuation value, the predicted error value, and the predicted error ratio described in S6 includes: Evaluate the difference between the actual line loss value and the predicted line loss value of the transmission line, obtain the predicted line loss value predicted by the improved LSTM model, calculate the actual line loss value of the transmission line and the predicted line loss value of the improved LSTM model, and use the formula: to obtain the predicted fluctuation value QS; where P t represents the actual line loss value at the t-th time, represents the predicted line loss value predicted at the t-th time, t represents the time index subscript predicted by the improved LSTM model, and N represents the total number of time index subscripts predicted by the improved LSTM model in the record; calculate the average of the absolute value of the difference between the predicted line loss value and the actual line loss value, and use the formula: to obtain the predicted error value QA; calculate the ratio of the predicted line loss value to the actual line loss value, and use the formula: Obtain the prediction error ratio QP.

Images