Power lightweight storage method based on time series data compression

Abstract

The application discloses a power lightweight storage method based on time series data compression, comprising the following steps: S1, acquiring time series data stream from power collection terminals in real time, and outputting voltage and current fluctuation trend identification; S2, storing summary records as indexes for partition storage; S3, performing power grid anomaly analysis based on compressed data packets and original data packets of the partition storage; S4, constructing a deep learning load prediction model cascaded by a time series convolution network and a long short-term memory network, and outputting power load prediction values; and S5, calculating load prediction deviation based on the power load prediction values, and combining the fluctuation trend identification to identify and grade early warning of power grid abnormal changes. The application outputs fluctuation trend identification through S1, and constructs a three-level storage system of a first-level high-speed storage area, a second-level high-density storage area and a third-level independent storage area, so that efficient unified integration of power data resources is realized, and the problems of storage space expansion and low data retrieval efficiency are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power data storage technology, and more specifically to a lightweight power storage method based on time-series data compression. Background Technology

[0002] With the comprehensive construction of smart grids, the power system generates massive amounts of time-series data every moment. For example, smart meters deployed in distribution network areas collect and transmit operational data, including voltage, current, and power, every 15 minutes; while synchronous phasor measurement units (PMUs) used for monitoring power grid safety and stability transmit data at a frequency of tens or even hundreds of times per second. This data is a crucial foundation for accurate load forecasting and real-time identification of power grid anomalies.

[0003] However, when storing such massive amounts of power time-series data, existing technologies typically store the data directly on servers or cloud databases in a full, lossless manner to meet the granularity requirements of short-term load forecasting and the high-frequency detail needs of power grid anomaly analysis. While this storage method ensures data integrity, the massive amount of time-series data occupies most of the storage space. When load forecasting models need to retrieve data from a longer historical time window (e.g., the past year) for model training, or when anomaly diagnosis requires retrospectively querying high-frequency voltage and current waveforms from several days ago, the large data volume leads to enormous I / O throughput pressure, resulting in storage space expansion and low data retrieval efficiency. Furthermore, while conventional full-file compression techniques (such as ZIP and GZIP) can solve the storage space problem, any data query or analysis must first decompress the entire file, failing to simultaneously meet the needs of dimensionality reduction and analysis. Summary of the Invention

[0004] To address this, the present invention provides a lightweight power storage method based on time-series data compression to solve the problems of storage space expansion and low data retrieval efficiency in the prior art; at the same time, simple compression strategies cannot meet the needs of dimensionality reduction and analysis.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A lightweight power storage method based on time-series data compression includes the following steps:

[0007] S1: Real-time acquisition of time-series data streams from power acquisition terminals, calculation of the variance of voltage and current and the slope of linear fitting of voltage and current, comparison with preset thresholds, and output of voltage and current fluctuation trend indicators.

[0008] S2: Determine the fluctuation level based on the voltage fluctuation trend indicator and the current fluctuation trend indicator, generate the corresponding compressed data packet and raw data packet, and then store them in partitions using the summary record as an index;

[0009] S3: Power grid anomaly analysis based on compressed data packets and raw data packets stored in partitioned storage;

[0010] S4: Construct a deep learning load prediction model consisting of a cascaded temporal convolutional network and a long short-term memory network to form macro-load trend feature vectors and micro-load morphology feature vectors. Then perform temporal alignment and splicing to output the predicted power load value.

[0011] S5: Calculate the load prediction deviation based on the predicted load value, and then combine it with the fluctuation trend indicator to identify and classify the power grid anomalies.

[0012] Furthermore, the specific steps of S1 are as follows:

[0013] S1.1: Using the first duration as a rolling time window, the voltage transformer and current transformer are used to collect and upload a time-series data stream including voltage and current in real time.

[0014] S1.2: Extract the summary features of the time series data stream for the current time period, and combine them with the start timestamp of the current time period to form a summary record;

[0015] S1.3: Using the second time period as the observation window, and combining the voltage mean sequence in the summary record, calculate the voltage variance and voltage linear fitting slope respectively;

[0016] S1.4: Using the second time period as the observation window, and combining the current mean sequence in the summary record, calculate the current variance and the current linear fitting slope respectively;

[0017] S1.5: Compare the voltage variance, voltage linear fitting slope, current variance, and current linear fitting slope with preset thresholds respectively, and output voltage and current fluctuation trend indicators to represent the overall operating status for the second duration.

[0018] Furthermore, the specific steps of S2 are as follows:

[0019] S2.1: Determine the fluctuation level of the current second duration based on the fluctuation trend indicators of voltage and current output by S1;

[0020] S2.2: Based on the fluctuation level of the current second time period, perform differentiated compression and storage operations on the original high-frequency time-series data stream within the current second time period:

[0021] If the fluctuation level is steady state, the original high-frequency time series data stream is lossily compressed using the rotating door compression algorithm to generate a steady-state compressed data packet.

[0022] If the fluctuation level is slowly varying, the original high-frequency time series data stream is losslessly compressed using Huffman coding to generate slowly varying compressed data packets.

[0023] If the fluctuation level is severe, no compression is performed on the original high-frequency time-series data stream; it is retained as the original data packet.

[0024] S2.3: Use the summary record generated by S1.2 as an index to associate it with the compressed or raw data packet of the corresponding second duration, and store them in different storage areas of the database respectively.

[0025] Furthermore, the determination steps of S2.1 are as follows: if the fluctuation trends of voltage and current are both stable, then it is determined to be a steady state;

[0026] If both voltage and current fluctuation trends are indicated by a gradual increase or decrease, or one is stable while the other is a gradual increase or decrease, then it is judged as a gradual change.

[0027] If the fluctuation trend of voltage or current is identified as violent oscillation, it is determined to be a violent state.

[0028] Furthermore, the specific steps of S3 are as follows:

[0029] S3.1: When performing power grid anomaly diagnosis, first scan the summary records of voltage peak value, voltage valley value, current peak value and current valley value in the first-level high-speed storage area, indexed by timestamp;

[0030] S3.2: If the voltage peak, voltage valley, current peak, and current valley are all within the preset normal range, it is determined that there is no abnormality in the current time period, and there is no need to access the lower-level second or third-level storage area, thus realizing the rapid elimination of abnormalities.

[0031] S3.3: If any of the voltage peak value, voltage valley value, current peak value, and current valley value exceeds the normal range, the corresponding data packet is retrieved from the second-level high-density storage area or the third-level independent storage area according to the summary record to confirm and locate the specific abnormal event and the time of occurrence.

[0032] Furthermore, during the training phase of the deep learning load prediction model, based on the span requirements of the training data time window, the summary records covering all first-time periods within the training period are quickly read from the first-level high-speed storage area using timestamps as indexes, and the average current sequence is extracted to form a macro load trend feature vector.

[0033] Simultaneously, steady-state compressed data packets and slowly varying compressed data packets corresponding to the time period are retrieved from the second-level high-density storage area, and after decompression, they are restored to current time-series data with the first time-length granularity, forming a micro-load morphology feature vector.

[0034] Furthermore, the specific steps of S5 are as follows:

[0035] S5.1: At the end of the current second time window, obtain the average current value of the current period from the summary record generated in real time in S1.1, and at the same time read the load prediction value for the current period from the prediction result buffer in S4.3, and calculate the load prediction deviation rate.

[0036] S5.2: When the load prediction deviation rate exceeds the preset first-level deviation threshold, read the current fluctuation trend indicator and voltage fluctuation trend indicator, and combine them with the extreme value scan results of the summary record to jointly determine the anomaly type;

[0037] S5.3: After any warning is triggered, the steady-state compressed data packet, slowly varying compressed data packet, or original data packet for the current period is retrieved from the second-level high-density storage area or the third-level independent storage area according to the summary record in S2.3. After decompression, the original high-frequency voltage timing data stream and current timing data stream are restored.

[0038] S5.4: Perform Fast Fourier Transform on the restored voltage and current time-series data streams to analyze the fundamental frequency and harmonic components, and identify frequency and harmonic anomalies.

[0039] Furthermore, the load prediction deviation rate The calculation formula is as follows:

[0040]

[0041] in, This is the load forecast value for the current second time window. This represents the average current value for the current second time window.

[0042] The present invention has the following advantages: The present invention calculates and outputs the fluctuation trend of voltage variance, voltage linear fitting slope, current variance and current linear fitting slope in real time through S1, combined with the differentiated hierarchical compression strategy of time series data stream based on fluctuation trend in S2, and constructs a three-level storage system with summary records as index, namely the first-level high-speed storage area, the second-level high-density storage area and the third-level independent storage area, thereby realizing the efficient and unified integration of power data resources and effectively solving the problems of storage space expansion and low data retrieval efficiency caused by the existing full storage method.

[0043] Meanwhile, S4 uses the average current sequence in the first-level high-speed storage area to form a macro-load trend feature vector, and uses the decompressed and restored current time series data in the second-level high-density storage area to form a micro-load morphology feature vector. The two are fused and input into a deep learning load prediction model composed of a cascaded temporal convolutional network and a long short-term memory network for training and online prediction. This enables the model to learn both the long-term evolution trend of the load and the intraday short-term fluctuation pattern, thereby improving the accuracy of short-term, medium-term and long-term power load prediction for the power grid and providing reliable decision support for the optimal allocation of power resources.

[0044] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0045] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0046] Figure 1 This is a flowchart illustrating the implementation of the power lightweight storage method based on time-series data compression according to the present invention. Detailed Implementation

[0047] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figure 1 A lightweight power storage method based on time-series data compression includes the following steps:

[0049] S1: Real-time acquisition of time-series data streams from power acquisition terminals, calculation of the variance of voltage and current and the slope of linear fitting of voltage and current, comparison with preset thresholds, and output of voltage and current fluctuation trend indicators to represent the overall operating status of the power grid.

[0050] The specific steps of S1 are as follows:

[0051] S1.1: Using a first duration (e.g., 15 minutes) as a rolling time window, the voltage transformer and current transformer are used to collect and upload a time-series data stream including voltage and current in real time;

[0052] The primary side of the voltage transformer is connected in parallel to the monitored power line. It converts the high voltage on the primary side into a low voltage signal on the secondary side proportionally. After analog-to-digital conversion, it forms a voltage timing data stream. Its function is to achieve high voltage isolation and precise scaling, ensure measurement safety, and provide real voltage waveform data.

[0053] The primary side of the current transformer is connected in series or clamped to the conductor of the monitored line. It converts the large primary current into a small secondary current signal proportionally. The signal is then converted into a voltage by a sampling resistor and then converted into a current timing data stream by analog-to-digital conversion. Its function is to non-intrusively sense the current carrying capacity of the line and provide an accurate current data source that reflects changes in electrical load.

[0054] S1.2: Extract summary features of the time-series data stream for the current time period, including average voltage, average current, peak voltage, valley voltage, peak current, and valley current. The extracted summary features, together with the start timestamp of the current time period, constitute a summary record. This record serves as a condensed expression of the steady-state operation of the current time period, is passed to subsequent trend analysis, and ultimately serves as the basic storage unit of the first-level storage area in S2.

[0055] S1.3: Using the second time period as the observation window, and combining it with the voltage mean sequence in the summary record, calculate the voltage variance respectively. Slope of linear fitting with voltage Voltage variance Used to indicate the severity of data fluctuations within the current time period. Voltage linear fit slope This indicates the overall trend and rate of change of the average voltage over the second time period, used to determine whether the data is stable, gradually increasing, or gradually decreasing. The calculation formula is as follows:

[0056]

[0057] in, This is the average of the voltage values ​​over the second time period; For the second duration The average voltage value of the first time window. This represents the total number of windows in the first duration.

[0058] voltage linear fitting slope The calculation formula is as follows:

[0059]

[0060] in, For the first The first duration window.

[0061] S1.4: Similarly, using the second time period as the observation window, and combining the current mean sequence in the summary record, calculate the current variance respectively. and the slope of the linear fitting of the current The calculation formula is as follows:

[0062]

[0063] in, This is the average value of the current over the second time period; For the second duration The average current of the first time window.

[0064] Current linear fitting slope The calculation formula is as follows:

[0065]

[0066] in, For the first The first duration window.

[0067] S1.5: The voltage variance will be... Voltage linear fitting slope Current variance and the slope of the linear fitting of the current The output voltage and current fluctuation trends are compared with preset thresholds to indicate the overall operating status during the second time period. An example of the rule is as follows:

[0068] like If the voltage is stable, then the voltage is marked as stable; if and Then the voltage is marked as a gradual increase; if and Then the voltage is marked as a gradual decrease; if or If so, the marked voltage will oscillate violently.

[0069] in, , , , This is a threshold constant determined based on voltage level and historical operating data.

[0070] The current fluctuation trend indicator is generated using the same method to obtain the corresponding current fluctuation trend indicator.

[0071] like If the current is stable, then the current is marked as stable; if Then the marked current is a gradual increase; if If the current is marked as gradually decreasing; if or If so, the current will be marked as violently oscillating.

[0072] in, , , , This is a current discrimination threshold constant determined based on the line's rated current and historical operating data.

[0073] S2: Determine the fluctuation level as steady state, slow change state or violent state based on the voltage fluctuation trend indicator and the current fluctuation trend indicator, and generate corresponding compressed data packets and raw data packets, and then store them in partitions using summary records as indexes.

[0074] The specific steps of S2 are as follows:

[0075] S2.1: Determine the fluctuation level of the current second duration based on the fluctuation trend indicators of voltage and current output by S1: if the fluctuation trend indicators of voltage and current are both stable, it is determined to be a steady state.

[0076] If both voltage and current fluctuation trends are indicated by a gradual increase or decrease, or one is stable while the other is a gradual increase or decrease, then it is judged as a gradual change.

[0077] If the fluctuation trend of either voltage or current is characterized as violent oscillation, then it is determined to be a violent state.

[0078] S2.2: Based on the fluctuation level of the current second time period, perform differentiated compression and storage operations on the original high-frequency time-series data stream within the current second time period:

[0079] If the fluctuation level is steady state, the original high-frequency time series data stream is lossily compressed using the rotating door compression algorithm to generate a steady-state compressed data packet.

[0080] If the fluctuation level is slowly varying, the original high-frequency time series data stream is losslessly compressed using Huffman coding to generate slowly varying compressed data packets.

[0081] If the fluctuation level is severe, no compression is performed on the original high-frequency time-series data stream; it is retained as the original data packet.

[0082] S2.3: Use the summary record generated by S1.2 as an index to associate it with the compressed or raw data packet of the corresponding second duration, and store them in different storage areas of the database respectively; wherein, the summary record is stored in the first-level high-speed storage area, the steady-state compressed data packet and the slowly varying compressed data packet are stored in the second-level high-density storage area, and the raw data packet is stored in the third-level independent storage area.

[0083] S3: Power grid anomaly analysis based on compressed data packets and raw data packets stored in partitioned storage: First, scan the voltage peak, voltage valley, current peak and current valley recorded in the first-level high-speed storage area. If all are within the normal range, it is determined that there is no anomaly. If any one exceeds the limit, the corresponding data packet is retrieved from the second-level high-density storage area or the third-level independent storage area, decompressed and then subjected to waveform backtracking analysis.

[0084] The specific steps for S3 are as follows:

[0085] S3.1: When performing power grid anomaly diagnosis, first scan the summary records of voltage peak value, voltage valley value, current peak value and current valley value in the first-level high-speed storage area, indexed by timestamp;

[0086] S3.2: If the voltage peak, voltage valley, current peak, and current valley are all within the preset normal range, it is determined that there is no abnormality in the current time period, and there is no need to access the lower-level second or third-level storage area, thus realizing the rapid elimination of abnormalities.

[0087] S3.3: If any of the voltage peak, voltage valley, current peak, or current valley values ​​exceeds the normal range, the corresponding data packet shall be retrieved from the second-level high-density storage area or the third-level independent storage area according to the summary record to confirm and locate the specific abnormal event and its occurrence time.

[0088] S4: Construct a deep learning load prediction model consisting of a cascaded temporal convolutional network and a long short-term memory network to generate macro-load trend feature vectors and micro-load morphology feature vectors. Then, perform temporal alignment and splicing to output the predicted electricity load value.

[0089] The specific steps for S4 are as follows:

[0090] S4.1: Construct a load prediction model based on deep learning. This model consists of a cascaded structure of a temporal convolutional network and a long short-term memory network. During the model training phase, based on the span requirements of the training data time window, the summary records covering all first-time periods within the training period are quickly read from the first-level high-speed storage area using timestamps as indices. The average current sequence is extracted to form a macro-load trend feature vector. Simultaneously, steady-state compressed data packets and slowly varying compressed data packets for the corresponding time periods are retrieved from the second-level high-density storage area. After decompression, they are restored to current time-series data at the first-time-length granularity to form a micro-load morphology feature vector.

[0091] S4.2: The macro load trend feature vector and micro load shape feature vector of the same time window are time-series aligned and spliced, and input into the load prediction model. The average actual current at future time is used as a label for supervised training, so that the model learns the inherent mapping relationship between the long-term evolution trend of load and the short-term fluctuation pattern.

[0092] S4.3: During online prediction, for a given prediction start time stamp, the summary record of the historical time window before that moment is read from the first-level high-speed storage area, the average current sequence is extracted, and the corresponding compressed data package is retrieved from the second-level high-density storage area, decompressed to obtain microscopic morphological features, and input into the trained load prediction model. The output is a current average prediction sequence covering short-term, medium-term or long-term, i.e., the predicted electricity load value. The prediction result, together with the prediction timestamp, is stored in the prediction result cache area.

[0093] S5: Calculates load prediction deviation based on predicted electrical load values, and then uses fluctuation trend indicators to identify and provide graded early warnings for power grid anomalies. This facilitates deviation analysis between real-time load prediction values ​​and actual average current values, and links with the fluctuation trend indicators in S1 to achieve accurate anomaly identification based on prediction deviation.

[0094] The specific steps for S5 are as follows:

[0095] S5.1: At the end of the current second time window, obtain the average current for the current period from the summary record generated in real time in S1.1. Simultaneously, the load forecast value for the current period is read from the forecast result buffer of S4.3, and the load forecast deviation rate is calculated. The calculation formula is as follows:

[0096]

[0097] in, This is the load forecast value for the current second time window.

[0098] S5.2: When the load prediction deviation rate exceeds the preset first-level deviation threshold, read the current fluctuation trend indicator and voltage fluctuation trend indicator of the current period output by S1.4, and combine them with the extreme value scan results of the summary record in S3.1 to perform joint judgment of the anomaly type:

[0099] If the current fluctuation trend is marked as violent oscillation, and the load prediction deviation rate is positive and exceeds the first-level deviation threshold, and the current peak exceeds the preset multiple of the line's rated current carrying capacity, then it is determined to be a sudden impact load anomaly, triggering a first-level current impact warning.

[0100] If the current fluctuation trend is marked as violent oscillation, and the load prediction deviation rate is negative and exceeds the first-level deviation threshold, while the current valley value is lower than the preset minimum current carrying capacity threshold, then it is determined to be a sudden load shedding anomaly, triggering a first-level warning for a sudden load drop.

[0101] If the voltage fluctuation trend is identified as violent oscillation, and the voltage peak or voltage trough exceeds the normal range set in S3.2, it is determined to be an abnormal voltage over-limit, triggering a first-level voltage over-limit warning.

[0102] If the current fluctuation trend is marked as a gradual increase or decrease, and the load prediction deviation rate exceeds the level 2 deviation threshold for multiple consecutive second durations, it is determined to be a trend-based abnormal load transfer, triggering a level 2 warning for load trend anomalies.

[0103] S5.3: After any warning is triggered, the steady-state compressed data packet, slowly varying compressed data packet, or original data packet for the current time period is retrieved from the second-level high-density storage area or the third-level independent storage area according to the summary record in S2.3. After decompression, the original high-frequency voltage timing data stream and current timing data stream are restored.

[0104] S5.4: Perform Fast Fourier Transform on the restored voltage and current time-series data streams to analyze the fundamental frequency and harmonic components, and complete the supplementary identification of frequency and harmonic anomalies.

[0105] If the fundamental frequency deviates from the rated frequency by more than the preset frequency threshold, it is determined to be a frequency anomaly, and a first-level warning for frequency exceeding the limit is triggered.

[0106] If the total harmonic distortion rate exceeds the preset harmonic threshold, it is determined to be a harmonic distortion fluctuation, triggering a level two warning for harmonic exceedance.

[0107] The aforementioned Fast Fourier Transform algorithm refers to a general term for numerical calculation methods that convert discrete time-domain sampled signals into complex frequency-domain spectra. It is mainly used to perform spectral analysis on the original high-frequency voltage and current time-series data streams after decompression and reconstruction, thereby extracting the fundamental frequency components included in the signal and calculating the proportion of each harmonic amplitude relative to the fundamental amplitude to determine whether there is a frequency anomaly.

[0108] S5.5: Summarizes all early warning information from S5.2 and S5.4, generates summary records including early warning type, early warning level, trigger timestamp, load prediction deviation rate, and related information, pushes them to the operation and maintenance monitoring system in real time, and stores them in the early warning event database for subsequent statistical analysis and model feedback optimization.

[0109] This invention utilizes S1 to perform real-time calculation and output of voltage variance, voltage linear fitting slope, current variance, and current linear fitting slope, along with fluctuation trend identification. Combined with S2's differentiated hierarchical compression strategy for time-series data streams based on fluctuation trend identification, and using summary records as indexes to construct a three-tiered storage system—a first-tier high-speed storage area, a second-tier high-density storage area, and a third-tier independent storage area—this achieves efficient and unified integration of power data resources, promoting the digital and intelligent upgrading of power systems. It effectively solves the problems of storage space expansion and low data retrieval efficiency caused by existing full-volume storage methods.

[0110] Meanwhile, S3 first scans the voltage peak, voltage valley, current peak and current valley recorded in the first-level high-speed storage area to quickly eliminate anomalies. Only for suspicious periods with extreme values ​​exceeding the limit, data packets in the deep storage area are retrieved as needed for decompression and backtracking. S5 further analyzes the deviation between the actual average current and the load prediction value output by S4, and identifies the anomaly type in conjunction with the current fluctuation trend indicator and voltage fluctuation trend indicator. This transforms the diagnosis of power grid anomalies from the traditional passive post-event repair to proactive prediction based on data, improving power supply reliability and reducing the failure rate of power equipment, as well as reducing the manpower and maintenance costs of power grid operation and maintenance.

[0111] In addition, S4 uses the average current sequence in the first-level high-speed storage area to form a macro-load trend feature vector, and uses the decompressed and restored current time series data in the second-level high-density storage area to form a micro-load morphology feature vector. The two are fused and input into a deep learning load prediction model composed of a cascaded temporal convolutional network and a long short-term memory network for training and online prediction. This allows the model to learn both the long-term evolution trend of the load and the intraday short-term fluctuation pattern, thereby improving the accuracy of short-term, medium-term and long-term power load prediction for the power grid and providing reliable decision support for the optimal allocation of power resources.

[0112] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lightweight power storage method based on time-series data compression, characterized in that, Includes the following steps: S1: Real-time acquisition of time-series data streams from power acquisition terminals, calculation of the variance of voltage and current and the slope of linear fitting of voltage and current, comparison with preset thresholds, and output of voltage and current fluctuation trend indicators. S2: Determine the fluctuation level based on the voltage fluctuation trend indicator and the current fluctuation trend indicator, generate the corresponding compressed data packet and raw data packet, and then store them in partitions using the summary record as an index; S3: Power grid anomaly analysis based on compressed data packets and raw data packets stored in partitioned storage; S4: Construct a deep learning load prediction model consisting of a cascaded temporal convolutional network and a long short-term memory network to form macro-load trend feature vectors and micro-load morphology feature vectors. Then perform temporal alignment and splicing to output the predicted power load value. S5: Calculate the load prediction deviation based on the predicted load value, and then combine it with the fluctuation trend indicator to identify and classify the power grid anomalies.

2. The power lightweight storage method based on time-series data compression according to claim 1, characterized in that, The specific steps of S1 are as follows: S1.1: Using the first duration as a rolling time window, the voltage transformer and current transformer are used to collect and upload a time-series data stream including voltage and current in real time. S1.2: Extract the summary features of the time series data stream for the current time period, and combine them with the start timestamp of the current time period to form a summary record; S1.3: Using the second time period as the observation window, and combining the voltage mean sequence in the summary record, calculate the voltage variance and voltage linear fitting slope respectively; S1.4: Using the second time period as the observation window, and combining the current mean sequence in the summary record, calculate the current variance and the current linear fitting slope respectively; S1.5: Compare the voltage variance, voltage linear fitting slope, current variance, and current linear fitting slope with preset thresholds respectively, and output voltage and current fluctuation trend indicators to represent the overall operating status for the second duration.

3. The lightweight power storage method based on time-series data compression according to claim 1, characterized in that, The specific steps of S2 are as follows: S2.1: Determine the fluctuation level of the current second duration based on the fluctuation trend indicators of voltage and current output by S1; S2.2: Based on the fluctuation level of the current second time period, perform differentiated compression and storage operations on the original high-frequency time-series data stream within the current second time period: If the fluctuation level is steady state, the original high-frequency time series data stream is lossily compressed using the rotating door compression algorithm to generate a steady-state compressed data packet. If the fluctuation level is slowly varying, the original high-frequency time series data stream is losslessly compressed using Huffman coding to generate slowly varying compressed data packets. If the fluctuation level is severe, no compression is performed on the original high-frequency time-series data stream; it is retained as the original data packet. S2.3: Use the summary record generated by S1.2 as an index to associate it with the compressed or raw data packet of the corresponding second duration, and store them in different storage areas of the database respectively.

4. The power lightweight storage method based on time-series data compression according to claim 1, characterized in that, The determination steps for S2.1 are as follows: if the fluctuation trends of voltage and current are both stable, then it is determined to be a steady state; If both voltage and current fluctuation trends are indicated by a gradual increase or decrease, or one is stable while the other is a gradual increase or decrease, then it is judged as a gradual change. If the fluctuation trend of voltage or current is identified as violent oscillation, it is determined to be a violent state.

5. The power lightweight storage method based on time-series data compression according to claim 1, characterized in that, The specific steps of S3 are as follows: S3.1: When performing power grid anomaly diagnosis, first scan the summary records of voltage peak value, voltage valley value, current peak value and current valley value in the first-level high-speed storage area, indexed by timestamp; S3.2: If the voltage peak, voltage valley, current peak, and current valley are all within the preset normal range, it is determined that there is no abnormality in the current time period, and there is no need to access the lower-level second or third-level storage area, thus realizing the rapid elimination of abnormalities. S3.3: If any of the voltage peak value, voltage valley value, current peak value, and current valley value exceeds the normal range, the corresponding data packet is retrieved from the second-level high-density storage area or the third-level independent storage area according to the summary record to confirm and locate the specific abnormal event and the time of occurrence.

6. The power lightweight storage method based on time-series data compression according to claim 1, characterized in that, During the training phase of the deep learning load prediction model, based on the span requirements of the training data time window, the summary records covering all first-time periods within the training period are quickly read from the first-level high-speed storage area using timestamps as indexes, and the average current sequence is extracted to form a macro load trend feature vector. Simultaneously, steady-state compressed data packets and slowly varying compressed data packets corresponding to the time period are retrieved from the second-level high-density storage area, and after decompression, they are restored to current time-series data with the first time-length granularity, forming a micro-load morphology feature vector.

7. The power lightweight storage method based on time-series data compression according to claim 1, characterized in that, The specific steps of S5 are as follows: S5.1: At the end of the current second time window, obtain the average current value of the current period from the summary record generated in real time in S1.1, and at the same time read the load prediction value for the current period from the prediction result buffer in S4.3, and calculate the load prediction deviation rate. S5.2: When the load prediction deviation rate exceeds the preset first-level deviation threshold, read the current fluctuation trend indicator and voltage fluctuation trend indicator, and combine them with the extreme value scan results of the summary record to jointly determine the anomaly type; S5.3: After any warning is triggered, the steady-state compressed data packet, slowly varying compressed data packet, or original data packet for the current period is retrieved from the second-level high-density storage area or the third-level independent storage area according to the summary record in S2.

3. After decompression, the original high-frequency voltage timing data stream and current timing data stream are restored. S5.4: Perform Fast Fourier Transform on the restored voltage and current time-series data streams to analyze the fundamental frequency and harmonic components, and identify frequency and harmonic anomalies.

8. The power lightweight storage method based on time-series data compression according to claim 1, characterized in that, The load prediction deviation rate The calculation formula is as follows: in, This is the load forecast value for the current second time window. This represents the average current value for the current second time window.

Images