Method and system for repairing incomplete time series data

Abstract

The invention relates to an incomplete time series data restoration method and system, and belongs to the technical field of photovoltaic output data preprocessing. The method specifically comprises the following steps: S1, detecting abnormal values and missing values of multi-dimensional photovoltaic power generation data and meteorological related factor data, and processing the abnormal data as missing data; s2, performing low-rank approximation and feature extraction on the processed data; s3, building a cost matrix in the DTW to obtain similar day matching data of the missing data; and S4, filling missing data to complete data recovery. According to the scheme, the accuracy and the reliability of photovoltaic output prediction are remarkably improved by comprehensively considering the historical integrity of the data and the multi-dimensional meteorological factors. In addition, due to the flexibility and adaptability of the scheme, the method is not only suitable for the photovoltaic field, but also can be popularized to other occasions needing time sequence data recovery.

Description

Technical Field

[0001] The invention belongs to the technical field of preprocessing photovoltaic output data and relates to a method and system for repairing incomplete time series data. Background Art

[0002] As energy issues become increasingly prominent, countries are paying more and more attention to the photovoltaic industry, and the global photovoltaic installed capacity is increasing year by year. Affected by environmental factors such as sunshine conditions, temperature and seasonal changes, photovoltaic output objectively has significant fluctuations, resulting in drastic changes in power generation in a short period of time. In addition, photovoltaic output also has intermittent characteristics, which are mainly affected by the sunshine cycle and seasonal changes, and are manifested in the interruption of photovoltaic power generation at night and on cloudy days. These characteristics make photovoltaic power generation difficult to predict, and when it is connected to the public power grid, it will have a certain impact on the stability of the power grid. Considering the uncertainty of photovoltaic output, in order to improve the overall economy and reliability of the power system, the current global energy policy tends to encourage renewable energy to participate in the regulation and trading of medium- and long-term power markets. Under this framework, the accuracy of photovoltaic output forecasting becomes particularly critical, which is directly related to the stability of power supply and affects energy trading decisions and future power grid planning.

[0003] Since medium- and long-term forecasts need to more accurately capture the long-term dependencies and seasonal patterns in time series, the data quality requirements are higher than those of short-term and ultra-short-term forecasts. Medium- and long-term forecast models must process and analyze data with a longer time span to ensure that these patterns can be identified and simulated. However, during the data collection process, equipment failures may cause data collection interruptions or inaccurate records, and omissions in records may cause missing data fragments, which will have a negative impact on subsequent forecast results. In order to address this problem, most current data preprocessing research focuses on statistical techniques, such as mean filling, front and back value filling, etc. Although the operation is simple, it usually ignores the dynamic characteristics of the data and cannot restore the original trend and complex patterns of the data. Some scholars use machine learning, deep learning and other methods to deal with the complex relationship between photovoltaic data and meteorological factors, but a lot of feature engineering is required to improve accuracy. Although some scholars use historical data similarity to fill missing values, they usually fail to fully consider the importance of time alignment, which is crucial to ensure that the filled data can accurately reflect the dynamic characteristics of photovoltaic output. If there is a lack of precise time alignment, even if similar historical data is used for filling, it may not be possible to truly restore the changes in photovoltaic output, thus affecting the accuracy and reliability of the prediction model. Summary of the invention

[0004] In view of this, in order to solve the above-mentioned deficiencies in the prior art, the present invention proposes a method and system for repairing incomplete time series data, which aims to enhance the historical integrity of photovoltaic output data and improve the generalization performance of the prediction model through similarity analysis and time series alignment. The method and system provided by the present invention first identify and mark the outliers and missing values ​​in the data set, and then use historical similar day data for intelligent filling to ensure that the filled data set can truly reflect the long-term dynamics of photovoltaic output.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A method for repairing an incomplete time series, the method specifically comprising:

[0007] S1: For multi-dimensional photovoltaic power generation data and meteorological related factor data, first detect outliers and missing values, and treat the outliers as missing data;

[0008] S2: Perform low-rank approximation and feature extraction on the processed data;

[0009] S3: Obtain similar day matching data for missing data by constructing a cost matrix in DTW;

[0010] S4: Fill in the missing data and complete data recovery.

[0011] Further, in step S1, the abnormal value and missing value detection specifically includes:

[0012] S11: For abnormal values, there are usually two situations: glitches and extreme values. Glitches refer to data values ​​at a certain time point that are significantly different from the ringing time point, resulting in a thorny shape in the data curve. Extreme values ​​refer to abnormal data points when the load data value at a non-peak or non-trough time point on a certain day is significantly higher than the peak time point value or lower than the trough time point value. If these values ​​are used as inputs for subsequent model training, the error value of the machine learning process will be too large. In order to detect abnormal values, the standard deviation method is first used to measure the degree of deviation of a single data point from the average value. Therefore, for the standard deviation It is necessary to ensure that z|<3, at which point the preliminary marking of overall outliers is completed;

[0013] S12: Since IQR has a stronger adaptability to the distribution of data, the time series is further marked on the basis of S11, and the smaller boundary points Q1 and the larger boundary points Q3 in the time series are selected. At this time, there is a lower boundary limit L l =Q1-1.5×IQR and upper boundary L u =Q3+1.5×IQR, when the value of the data point x t Satisfy xt <L l or x t >L u When the data point satisfies both S11 and S12, the data point is regarded as an outlier and processed as a missing value.

[0014] S13: For all missing values ​​x t Use statistical methods to count missing values ​​and perform preliminary missing value filling. First, find the backward non-missing value x of the missing value. t-1 and the forward non-missing values ​​x t+1 , use its mean to preliminarily fill in the missing values, that is,

[0015] Further, in step S2, the low-rank approximation and feature extraction specifically include:

[0016] S21: Normalize the input data. In the photovoltaic output data, the value of 0 every day means that the equipment has no output or the equipment is turned off. In order to ensure that these key information are not lost in subsequent analysis, the input time series needs to be further normalized. The data point x corresponding to each location t , and its normalized result is

[0017] S22: Create an n×p dimensional data matrix Z and construct a covariance matrix to express the linear relationship between meteorological factors and photovoltaic output data in the data set The covariance matrix is ​​a p×p matrix, each element c ij Represents the covariance between features;

[0018] S23: Perform eigenvalue decomposition on the covariance matrix C, i.e., |λI-C|, and obtain the eigenvalues ​​λ1,λ2,...,λ p , after the eigenvalues ​​are arranged in descending order, solve the corresponding eigenvectors c1, c2, ..., c p ;

[0019] S24: Select the eigenvectors corresponding to the first k largest eigenvalues ​​according to the required amount of information, and calculate the contribution rate of each component And its cumulative contribution rate The value of each component is calculated by:

[0020]

[0021] For each sample, calculate its comprehensive meteorological coefficient F = η1z1 + η2z2 + ... + η i z i +...+ηm z m .

[0022] Further, in step S3, the similar day matching specifically includes:

[0023] S31: The outliers have been marked through S1, and these marked data x sign The date is taken as the center, and the period T is selected as a time window. If the missing dates are close to the beginning or end of the data set, the time window needs to be adjusted to fit the data boundaries;

[0024] S32: Enumerate all time windows with a time interval of T in the data set, and ensure that the data in each time window is complete and without missing data; for each time window, use the DTW algorithm to calculate the similarity score with the reference time window; the DTW algorithm finds the best alignment by minimizing the overall distance between time series, and arranges similar time series in ascending order according to the DTW score. The lower the score, the smaller the difference between the sequences, so as to obtain several sequences with higher similarity; for the target time series X and the candidate time series T containing missing values, construct a distance matrix D[i,j]=(X[i]-T[j]) 2 , initialize the DTW matrix, that is, let DTW[0,0] = D[0,0], and the remaining DTW[0,j] and DTW[i,0] are initialized to infinity, and then gradually fill the entire matrix:

[0025] DTW[i,j]=D[i,j]+min(DTW[i-1,j],DTW[i,j-1],DTW[i-1,j-1]);

[0026] S33: When the time series requires a high enough granularity of the sequence, the sequence filtered out by S32 is further matched, and the similarity between the day and the day of missing data is compared. The day with the lowest DTW score is selected as the most similar date. If the missing data involves a whole day or multiple consecutive data points, the corresponding data of the similar day of the selected date is directly used to completely replace the missing value.

[0027] Further, in step S4, the data recovery specifically includes: in step S24, the corresponding principal component coefficients are obtained, and then the inverse transformation of PCA is used to complete the final data recovery for the obtained new sequence, z i =c i1 x1+c i2 x2+…+c ij x j +…+c ip x p

[0029] The beneficial effects of the present invention are:

[0030] The technical solution provided by the present invention significantly improves the accuracy and reliability of photovoltaic output forecasting by comprehensively considering the historical integrity of data and multi-dimensional meteorological factors. In addition, the flexibility and adaptability of this solution make it not only suitable for the photovoltaic field, but also can be extended to other occasions that require time series data repair.

[0031] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0033] Figure 1 This is a system architecture diagram of the present invention;

[0034] Figure 2 A diagram of the hash tree structure designed for the present invention. DETAILED DESCRIPTION

[0035] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0036] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0037] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0038] See also Figure 1-2 , a method for repairing incomplete time series data, in order to reduce the volatility and randomness of the prediction results, first detect the outliers and treat them as missing data, record the location of the missing points after simply filling the missing data, and standardize the data, then use PCA to perform low-rank approximation on the data, and obtain the main variation direction of the data set, for the data after dimensionality reduction, use DTW to match similar days, and complete the filling of the data set, and finally restore the filled data to multi-dimensional data to achieve data reconstruction to enhance the accuracy of subsequent predictions. To achieve the above purpose.

[0039] In this embodiment, the technical solution provided by the present invention is as follows:

[0040] The present invention proposes a strategy for repairing incomplete time series data. Considering that the current data preprocessing can be divided into two links: detection and correction of abnormal data. The missing values ​​are directly counted by statistical methods, and the abnormal values ​​are identified by threshold detection method; after the position of the abnormal data is identified and counted, it is necessary to correct it. In the method of this article, the abnormal values ​​and missing values ​​are detected, and their positions are marked. If it is an abnormal value, it is made a missing value. When processing abnormal data, the standard score method (Z-Score) or the box plot (Interquartile Range, IQR) method is often used. Z-Score is a method for measuring the degree of deviation of a single data point from the average value. When the data set is approximately normally distributed, it can be used to identify values ​​in the data that are significantly different from the average value, and when the number of samples of the photovoltaic output data set is large enough, the data can be regarded as obeying the normal distribution. The specific proof is given in the simulation analysis part. The box plot rule is a graphical method for displaying data distribution characteristics. Considering its robustness to abnormal values, it is applicable to data of any distribution type.

[0041] In PV output datasets, extreme Z-Score values ​​usually indicate equipment failure, sudden environmental changes, or data recording errors. However, for biased datasets containing a large number of extreme values, the applicability of Z-Score is limited. In this case, Z-Score cannot distinguish between true outliers and extreme but normal values ​​in the dataset. IQR, on the other hand, can help identify data points that do not conform to seasonal patterns or deviate from the normal range due to equipment failure when analyzing PV data. A single outlier detection method, whether based on Z-Score or IQR, may be overly sensitive to specific patterns or noise in the data. For example, Z-Score may perform well under the assumption of Gaussian distribution, but may fail when faced with data with heavy-tailed distribution. Similarly, IQR may not accurately capture outliers when there are multiple patterns or seasonal changes in the dataset. Therefore, combining multiple detection methods can provide more comprehensive outlier identification capabilities. This paper adopts a combined detection strategy of Z-Score and IQR, which is based on the coordinated application of Z-Score and IQR. The theoretical basis of the strategy is that Z-Score and IQR have their own advantages and applicable scenarios in outlier detection. Z-Score performs well when the data is assumed to be Gaussian distributed, while IQR is more adaptable to the distribution of the data. These two methods are combined to achieve comprehensive identification of outliers while reducing reliance on a single detection algorithm. First, use Z-Score for preliminary outlier screening to identify points that deviate greatly from the mean, and further refine the identification of outliers through the IQR method to ensure that points that do not conform to the distribution characteristics of the data set are accurately marked:

[0042]

[0043] where x t represents the value at time t, μ is the mean of the sequence, and z is the standard deviation. Q1 and Q3 represent the first quartile and the third quartile respectively. Q1 is the boundary point of the smaller part of the data set, indicating that 25% of the data points in the data are less than this value. Q3 represents the boundary point of the larger part of the data set, indicating that 75% of the data points in the data are less than this value. L l and L u are the lower and upper limits of the box plot method.

[0044] For all data points x t Check: When x t <L l or x t >L u, it is marked as an outlier. If it has been marked as an outlier by the Z-Score method, its status is confirmed again, otherwise its marking status is cancelled, and finally all marked data points are missing, making them missing values, and the missing data are preliminarily filled, that is, Indicates the missing values ​​that need to be filled, x t-1 and x t+1 represent the backward non-missing values ​​and forward non-missing values ​​of missing values ​​respectively.

[0045] When analyzing photovoltaic output data sets, the data has inherent high dimensionality and complex correlations between features. This data structure makes the direct application of prediction models time-consuming and may cause dimensionality disasters. To solve this problem, considering that the principal component analysis (PCA) method can project the data into a low-dimensional space consisting of principal components with maximized variance through linear transformation, this process not only reduces the dimensionality of the data, but also retains the most important change information in the data set, that is, those features that contribute the most to the data variance. These principal components can not only represent the key features of the original data, but also in large-scale time series analysis, PCA can ensure the undistortion of the data and ensure that the reduced-dimensional data can be accurately restored to the original state. Compared with other dimensionality reduction techniques, the mathematical principles and algorithm implementation of PCA are relatively simple and computationally efficient, which makes PCA particularly effective when processing large-scale data sets. In addition, PCA does not rely on the distribution assumptions of the data, which enables it to maintain good performance when facing non-normally distributed data. When considering multi-factor photovoltaic output prediction, the dimensions of different characteristic factors are different. The photovoltaic output data range is often larger than other characteristic factors, which leads to the fact that this factor occupies a dominant position in the model algorithm when making model predictions, while other features with smaller values ​​will be submerged, making it impossible for the model to effectively learn the laws of these features. Therefore, it is necessary to normalize the input data, and the commonly used normalization methods are Min-Max normalization, decimal calibration normalization, logarithmic transformation, square root transformation, etc. In the photovoltaic output data, the daily value of 0 means that the equipment has no output or the equipment is shut down. In order to ensure that these key information is not lost in subsequent analysis, and considering that the Min-Max normalization method can not only retain the original distribution shape and relative relationship of the data, but also has a simple calculation process, compared with other methods, it has the characteristics of simplicity, speed and retention of the original data distribution, so the minimum-maximum normalization method is used. This method can scale the original data to a specified range, making the model more suitable for processing data of different dimensions and magnitudes, and improving the model's generalization ability for different input features, that is: where x max and x min are the maximum and minimum values ​​in the data sequence respectively, and x is the normalized value.

[0046] For the normalized results, an n×p-dimensional data matrix Z is created, and a covariance matrix C is constructed to express the linear relationship between meteorological factors and photovoltaic output data in the data set. Where Z T represents the transpose of the matrix Z. The covariance matrix is ​​a p×p matrix, where each element c ij Represents the covariance between feature i and feature j. Since the eigenvalue size of the constructed covariance matrix reflects the importance of the corresponding eigenvector in the data and the ability to explain the data variance, the factors with larger eigenvalues ​​can be used as the main variation direction for feature construction in the subsequent mid- and long-term photovoltaic output forecast. Therefore, the covariance matrix C is decomposed by eigenvalue, that is, |λI-C|, and the eigenvalues ​​λ1,λ2,...,λ p , after the eigenvalues ​​are arranged in descending order, solve the corresponding eigenvectors c1, c2, ..., c p .

[0047] According to the required amount of information, select the eigenvectors corresponding to the first k largest eigenvalues. The formula for the principal component contribution rate is: The cumulative contribution rate is Where m is the number of principal components. The corresponding eigenvalues ​​with cumulative contribution rates of 80% to 95% are usually selected as principal components. The values ​​of each principal component are calculated according to the following formula.

[0048]

[0049] Where c ij is the element of the eigenvector matrix, z i For each sample, calculate its comprehensive meteorological coefficient F = η1z1+η2z2+...+η i z i +...+η m z m , where η i is the contribution rate, which indicates the weight of each output component.

[0050] Considering that the Dynamic Time Warping (DTW) algorithm is very robust to noise and outliers in time series, and can handle the scaling and offset problems of time series, it can provide reliable similar day matching even when the data is incomplete. Photovoltaic output data is affected by a variety of meteorological factors, resulting in strong volatility. The DTW algorithm is not only very robust to noise and outliers in time series, but also can handle the scaling and offset problems of time series caused by changes in meteorological conditions. This means that even in the case of data fluctuations caused by meteorological factors such as sunshine intensity and cloud cover, the DTW algorithm can still accurately identify the dates that are most similar to the target time series in history. The specific steps and processes are as follows:

[0051] Select the time window: During the abnormal data preprocessing, all abnormal values ​​and missing values ​​have been marked. sign The date is taken as the center and the period T is selected as a time window. If the missing dates are near the beginning or end of the dataset, you may need to adjust the time window to accommodate the data boundaries.

[0052] Preliminary DTW matching: Enumerate all time windows with a time interval of T in the dataset. Ensure that the data in each time window is complete and not missing. For each time window, use the DTW algorithm to calculate the similarity score with the reference time window. The DTW algorithm finds the best alignment by minimizing the overall distance between time series. Similar time series are arranged in ascending order according to the DTW score. The lower the score, the smaller the difference between the sequences, thus obtaining several sequences with higher similarity.

[0053] For the target time series X and the candidate time series T containing missing values, construct a distance matrix

[0054] D[i,j]=(X[i]-T[j]) 2 Where and j represent the indexes of time series X and T respectively. Initialize the DTW matrix, that is, let DTW[0,0] = D[0,0], and the rest of DTW[0,j] and DTW[i,0] are initialized to infinity, and then gradually fill the entire matrix: DTW[i,j] = D[i,j] + min(DTW[i-1,j], DTW[i,j-1], DTW[i-1,j-1])

[0055] Day-level DTW matching: In the months selected in the previous step, a separate DTW analysis is performed on each day. The same method as the preliminary DTW matching is used to calculate the cumulative distance matrix. The similarity between the day and the day of missing data is compared, and the day with the lowest DTW score is selected as the most similar day. If the missing data involves a whole day or multiple consecutive data points, the corresponding data of the similar day of the selected day is directly used to completely replace the missing value, that is, x sign =x t .

[0056] For the missing moment data in the day, the present invention not only considers the overall pattern of similar days, but also pays attention to the fluctuations on a smaller time scale, and interpolates the missing points more smoothly by using the data of similar days and the existing data points around. This interpolation strategy takes into account the granularity of the data, that is, the distance and frequency of the data points in time, and ensures that the filled data is consistent with the adjacent data points in terms of statistical characteristics and time series characteristics. In this way, the time series of photovoltaic output can be reconstructed more accurately, and the accuracy and smoothness of the filling can be ensured even when the data granularity is fine.

[0057] Data restoration: In order to reduce the impact of subsequent predictions, the data needs to be restored. Since each data contains a principal component coefficient when performing PCA low-rank approximation, there is z i =c i1 x1+c i2 x2+…+c ij x j +…+c ip x p .

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for repairing an incomplete time series, characterized in that: The method specifically includes: S1: For multi-dimensional photovoltaic power generation data and meteorological related factor data, first detect outliers and missing values, and treat the outliers as missing data; S2: Perform low-rank approximation and feature extraction on the processed data; S3: Obtain similar day matching data for missing data by constructing a cost matrix in DTW; S4: Fill in the missing data and complete data recovery.

2. The incomplete time series repair method according to claim 1, characterized in that: In step S1, the abnormal value and missing value detection specifically includes: S11: For abnormal values, there are usually two situations: glitches and extreme values. Glitches refer to the data value at a certain time point that is significantly different from the ringing time point, resulting in a thorny shape in the data curve. Extreme values ​​refer to abnormal data points when the load data value at a non-peak or non-trough time point on a certain day is significantly higher than the peak time point value or lower than the trough time point value. In order to detect abnormal values, the standard deviation method is first used to measure the degree of deviation of a single data point from the average value. Therefore, for the standard deviation It is necessary to ensure that |z| < 3, at which point the preliminary marking of overall outliers is completed; S12: Since IQR has a stronger adaptability to the distribution of data, the time series is further marked on the basis of S11, and the smaller boundary points Q1 and the larger boundary points Q3 in the time series are selected. At this time, there is a lower boundary limit L l =Q1-1.5×IQR and upper boundary L u =Q3+1.5×IQR, when the value of the data point x t Satisfy x t <L l or x t >L u When the data point satisfies both S11 and S12, the data point is regarded as an outlier and processed as a missing value. S13: For all missing values ​​x t Use statistical methods to count missing values ​​and perform preliminary missing value filling. First, find the backward non-missing value x of the missing value. t-1 and the forward non-missing values ​​x t+1 , use its mean to preliminarily fill in the missing values, that is, 3. The incomplete time series repair method according to claim 2, characterized in that: In step S2, the low-rank approximation and feature extraction specifically include: S21: Normalize the input data. In the photovoltaic output data, the value of 0 every day means that the equipment has no output or the equipment is turned off. The corresponding data point x at each location t , and its normalized result is S22: Create an n×p dimensional data matrix Z and construct a covariance matrix C to express the linear relationship between meteorological factors and photovoltaic output data in the data set The covariance matrix C is a p×p matrix, each element c ij represents the covariance between feature i and feature j; S23: Perform eigenvalue decomposition on the covariance matrix C, i.e. |λI-C|, and obtain the eigenvalues ​​λ1,λ2,...,λ p , after the eigenvalues ​​are arranged in descending order, solve the corresponding eigenvectors c1, c2, ..., c p ; S24: Select the eigenvectors corresponding to the first k largest eigenvalues ​​according to the required amount of information, and calculate the contribution rate of each component And its cumulative contribution rate The value of each component is calculated by: For each sample, calculate its comprehensive meteorological coefficient F = η1z1+η2z2+...+η i z i +...+η m z m .

4. The incomplete time series repair method according to claim 3, characterized in that: In step S3, the similar day matching specifically includes: S31: The outliers have been marked through S1, and these marked data x sign The date is taken as the center, and the period T is selected as a time window. If the missing dates are close to the beginning or end of the data set, the time window needs to be adjusted to fit the data boundaries; S32: Enumerate all time windows with a time interval of T in the data set, and ensure that the data in each time window is complete and without missing data; for each time window, use the DTW algorithm to calculate the similarity score with the reference time window; the DTW algorithm finds the best alignment by minimizing the overall distance between time series, and arranges similar time series in ascending order according to the DTW score. The lower the score, the smaller the difference between the sequences, so as to obtain several sequences with higher similarity; for the target time series X and the candidate time series T containing missing values, construct a distance matrix D[i,j]=(X[i]-T[j]) 2 , initialize the DTW matrix, that is, let DTW[0,0] = D[0,0], and the remaining DTW[0,j] and DTW[i,0] are initialized to infinity, and then gradually fill the entire matrix: DTW[i,j]=D[i,j]+min(DTW[i-1,j],DTW[i,j-1],DTW[i-1,j-1]); S33: When the time series requires a high enough granularity of the sequence, the sequence filtered out by S32 is further matched, and the similarity between the day and the day of missing data is compared. The day with the lowest DTW score is selected as the most similar date. If the missing data involves a whole day or multiple consecutive data points, the corresponding data of the similar day of the selected date is directly used to completely replace the missing value.

5. The incomplete time series repair method according to claim 4, characterized in that: In step S4, the data recovery specifically includes: in step S24, the corresponding principal component coefficients are obtained, and then the inverse transformation of PCA is used to complete the final data recovery for the obtained new sequence, z i =c i1 x1+c i2 x2+...+c ij x j +…+c ip x p .

6. An incomplete time series repair system, characterized by: The system adopts the method as claimed in any one of claims 1 to 5.

Images