Power grid frequency prediction method and system based on time sequence prediction algorithm

Abstract

The invention discloses a power grid frequency prediction method and system based on a time sequence prediction algorithm, and the method mainly comprises the steps: constructing an adjacent matrix of a prediction power grid based on the topological relation between power grid sites; historical data of the power grid are selected and preprocessed, wherein the preprocessing comprises data cleaning, standardization processing and time alignment; constructing a power grid frequency prediction model which comprises an improved dynamic graph learning module, a linear space convolution layer and a time convolution layer; and inputting the adjacency matrix and the preprocessed data into a trained power grid frequency prediction model, so that the power grid frequency change in the future 10 seconds can be predicted with high precision. According to the method, the accuracy of power grid frequency prediction is improved by enhancing the learning ability of the model for dynamic characteristics.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and specifically relates to a power grid frequency prediction method based on a time series prediction algorithm, and also relates to a power grid frequency prediction system based on a time series prediction algorithm. Background Art

[0002] Frequency prediction plays a crucial role in the operation of power systems. It is not only one of the important indicators to ensure the safe, stable and efficient operation of the power grid, but also a key factor in maintaining the stability of the power system. Therefore, quickly and accurately analyzing the dynamic frequency of the power grid is of great significance for formulating corresponding frequency stability control measures, improving system frequency stability, and preventing system frequency collapse. With the continuous expansion of the scale of power systems and the widespread access of new energy, the stability of power grid operation faces unprecedented challenges. Traditional frequency prediction methods mainly rely on physical models and statistical methods. However, these methods have certain limitations in dealing with the dynamic characteristics of large-scale and complex power systems and are difficult to meet the requirements of modern power grids.

[0003] CN118630791A proposes a physical-information driven power grid frequency prediction method considering the participation of new energy in frequency regulation. This method first constructs a frequency prediction physical model and a frequency prediction information model, then trains the frequency prediction information model, fuses the frequency prediction physical model and the frequency prediction information model, constructs a physical-information driven power grid frequency prediction model and trains this model. For the trained model, based on the active power disturbance information of the power grid, a dynamic prediction result of the power grid frequency is generated. The analysis method based on the physical information model is applicable to the transient analysis of traditional power grids. However, compared with traditional power grids, the frequency stability problem of new power systems is more complex, and the traditional physical information model analysis method has shown deficiencies in the accuracy and real-time performance of transient frequency prediction.

[0004] CN115864431A proposes a power grid frequency prediction method considering the influence of the primary frequency regulation start threshold of wind power. This method obtains a power grid system frequency response model and key parameters of wind power participating in the primary frequency regulation of the power grid, determines the wind power frequency regulation start threshold value and regulation rate, solves the dynamic frequency deviation based on the power grid system frequency response model, obtains the maximum system frequency deviation by analytically expressing the maximum value of the dynamic frequency deviation, and gets the lowest frequency prediction value; obtains the steady-state frequency deviation of the system and gets the steady-state frequency prediction value. However, the simulation calculation method is usually applicable to accurately simulate the dynamic response of the system in the time dimension, has a high computational complexity, low simulation efficiency, is difficult to handle large-scale power systems, is very sensitive to system parameters, and parameter errors may lead to deviations in simulation results. Summary of the Invention

[0005] In order to improve the accuracy of power grid frequency prediction and enhance the model's learning ability for dynamic features, a power grid frequency prediction model method based on a time series prediction algorithm is proposed.

[0006] The present invention adopts the following technical solutions.

[0007] In the first aspect of the present invention, a power grid frequency prediction method based on a time series prediction algorithm is proposed, specifically as follows: Step 1: Collect the historical operation data of the power grid related to the power grid frequency and perform preprocessing to obtain the topological relationship between power grid nodes at the corresponding moments of the historical data, and construct a power grid adjacency matrix; Step 2: Construct a power grid frequency prediction model, which includes an improved dynamic graph learning module, a linear space convolutional layer, and a linear time convolutional layer connected in sequence; Step 3: Use the historical data collected and preprocessed in Step 1 and the corresponding power grid adjacency matrix as input samples to train the power grid frequency prediction model; Step 4: Collect the current power grid operation data related to the power grid frequency in real time, construct the adjacency matrix of the power grid to be predicted, and input it into the trained power grid frequency prediction model to obtain the prediction result of the power grid frequency.

[0008] Preferably, in Step 1, collecting the historical operation data of the power grid related to the power grid frequency includes: Collecting frequency and active power data through a Phasor Measurement Unit (PMU), and collecting load data through a Supervisory Control and Data Acquisition (SCADA) system.

[0009] Preferably, in Step 1, the construction of the power grid adjacency matrix is specifically as follows: Construct an adjacency matrix based on the relationship between each site to represent the topological relationship between power grid nodes; in the adjacency matrix, nodes are divided into two categories: the first category is the units that can provide active power data; the second category is the units that can provide load data.

[0010] Preferably, in Step 1, the data preprocessing includes data cleaning, normalization processing, and time alignment.

[0011] Preferably, in Step 2, the improved dynamic graph learning module includes: constructing an adaptive weight block feature, introducing an improved dynamic and static embedding vector to construct a dynamic graph matrix, and optimizing the graph structure based on the attention mechanism.

[0012] Preferably, the adaptive weight block feature is specifically as follows: For the historical frequency sequence of the target node, calculate the importance score weight of each time node, and use the concatenation operation to concatenate the adaptive weight features of each node as a block,

[0013] Wherein, is the importance score for each historical time weight.

[0014] Preferably, the importance score is specifically:

[0015] Wherein, is a non-linear activation function, is a learnable parameter for measuring at importance, is a bias term.

[0016] Preferably, the improved static embedding vector:

[0017] Wherein, is the th data of the vector composed of the active power and load data of the site , is the number of sites.

[0018] Preferably, the improved dynamic embedding vector:

[0019] Wherein, is used to capture temporal information, enabling to remember historical states and predict future states.

[0020] The second aspect of the present invention proposes a power grid frequency prediction system based on a time series prediction algorithm, based on the power grid frequency prediction method described in the first aspect, specifically: Power grid data acquisition and preprocessing module: Collect historical power grid operation data related to the power grid frequency and perform preprocessing to obtain the topological relationship between power grid nodes at the corresponding moments of the historical data, and construct a power grid adjacency matrix; Power grid frequency prediction model construction module: Construct a power grid frequency prediction model composed of an improved dynamic graph learning module, a linear space convolutional layer, and a time convolutional layer; Power grid frequency prediction model training module: Use the collected and preprocessed historical data and the corresponding power grid adjacency matrix as input samples to train the power grid frequency prediction model; Frequency prediction module: Real-time collect current power grid operation data related to the power grid frequency, construct an adjacency matrix of the power grid to be predicted, and input it into the trained power grid frequency prediction model to obtain the prediction result of the power grid frequency.

[0021] The present invention has the following beneficial technical effects compared with the prior art: (1) By improving the adaptive weights of the block-level dynamic graph linear module and assigning different weights to historical time data, the time data that has a great impact on the power grid frequency prediction can be enhanced, thereby further improving the prediction accuracy. (2) Through the spatio-temporal joint modeling strategy that integrates the power grid topology relationship information, on the basis of modeling in the time dimension, the topology relationship information between sites is further introduced, which can capture the dynamic characteristics of the power system more comprehensively, make up for the limitations of single-time feature modeling, and more accurately predict the dynamic changes of the frequency, thereby significantly improving the accuracy of frequency prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic flowchart of the power grid frequency prediction method based on the time series prediction algorithm of the present invention; Figure 2 is the training process of the power grid frequency prediction model based on the time series prediction algorithm of the present invention; Figure 3 is a relationship graph between Ganzi power grid stations in an embodiment of the present invention; Figure 4 is an effect diagram of the power grid frequency prediction model in an embodiment of the present invention; Figure 5 is an effect diagram of the predicted power grid frequency under the influence of an upstream disturbance in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the spirit of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0024] As shown in the attached Figure 1 figures, the present invention specifically adopts the following technical solutions: Step 1: Collect the historical operation data of the power grid related to the power grid frequency and perform preprocessing to obtain the topological relationship between power grid stations at the corresponding moments of the historical data, and construct a power grid adjacency matrix; Step 2: Construct a power grid frequency prediction model, which includes an improved dynamic graph learning module, a linear spatial convolution layer, and a linear temporal convolution layer connected in sequence; Step 3: Use the historical data collected and preprocessed in Step 1 and the corresponding power grid adjacency matrix as input samples to train the power grid frequency prediction model; Step 4: Collect the current grid operation data related to the grid frequency in real time, construct the adjacency matrix of the grid to be predicted, and input it into the trained grid frequency prediction model to obtain the prediction result of the grid frequency.

[0025] In Step 1, the data preprocessing is specifically as follows: Collect historical data through the Phasor Measurement Unit (PMU) and the Supervisory Control and Data Acquisition (SCADA) system. The historical data includes: Set multiple sampling moments. For each sampling moment, obtain the frequency data for the past 10 seconds, the active power data for the past 10 seconds, the load data for the past 10 seconds, the active power data for the next 10 seconds, and the load data for the next 10 seconds relative to that moment.

[0026] Specifically, constructing the adjacency matrix of the grid to be predicted: In the actual grid operation scenario, there is a specific topological relationship between stations. An adjacency matrix is constructed based on the mutual relationships between stations and between stations and units to represent this topological relationship. In this topological relationship, nodes are divided into two categories: The first category is the units that can provide active power data; the second category is the units that can provide load data.

[0027] As Figure 2 shown, in Step 2, the Adaptive Weight Block-level Dynamic Graph Learning Module and the construction of the grid frequency prediction model are specifically as follows: (1) Adaptive time weight calculation: For the input short historical grid frequency time series, for node , given the historical frequency values for the previous steps, since in the historical 10-second data, the influence degrees of different times on the grid frequency prediction are different, an adaptive weight mechanism is introduced to assign different weights to historical times, so that the information of key times can be enhanced, thereby further improving the prediction accuracy. Calculate the importance score for each time:

[0028] where is the non-linear activation function, is a learnable parameter used to measure the importance of at , and is the bias term. The Adaptive Weight Block-level Dynamic Graph Learning Module uses the concatenation operation to splice the features of the nodes together as a block. The formula is as follows:

[0029] (2)Construct the dynamic graph adjacency matrix: In the dynamic graph learning module, an improved static embedding vector is introduced and the dynamic embedding , Improved static embedding vector:

[0030] In the formula, is the th data of the vector composed of the active power and load data of the site , and is the number of sites. The static embedding is encoded by learnable parameters and does not change with time, representing the inherent node attributes; The dynamic embedding changes with time and corresponds to the temporal characteristics of the nodes.

[0031]

[0032] In the formula, is used to capture temporal information, enabling to remember the historical state and predict the future state; Construct the dynamic graph matrix using the embedding vector:

[0033] (3)Optimize the dynamic graph matrix with the attention mechanism: Use the attention weight to represent the weight of the edge between node and node . The constructed graph structure is used to capture the time-varying correlation between any node pair. The formula is as follows:

[0034]

[0035] In the formula, is the non-linear activation function, is the trainable attention weight vector, and is the trainable linear transformation matrix.

[0036] The attention weight acts on the dynamic graph matrix:

[0037] Then, a hybrid multi-hop feature propagation operator is introduced for information propagation between nodes. The formula is as follows:

[0038]

[0039] In the formula the row is , representing the node representation at the th hop.

[0040] After the feature propagation of the th hop, a series of node representations at different hop counts can be obtained , where

[0041] is the propagation depth. Then, the node representations at all hop counts are aggregated to obtain the final representation of the node, and the formula is as follows:

[0042] Step 3: Use the historical data collected and preprocessed in Step 1 and the corresponding power grid adjacency matrix as input samples to train the power grid frequency prediction model; Step 4: Collect the current power grid operation data related to the power grid frequency in real time, construct the adjacency matrix of the power grid to be predicted, and input it into the trained power grid frequency prediction model to obtain the prediction result of the power grid frequency.

[0043] Taking the Ganzi Power Grid as an example, the topological relationship between stations is as Figure 3 . Based on the actual operation historical data of the Ganzi Power Grid, the PMU data and SCADA data from October 22, 2024 to October 28, 2024 are selected as research samples. Among them, the PMU data provides accurate generator active power information, and the SCADA data records the detailed load change situation. The data collection time span is 7 days, and continuous recording is carried out at a time resolution of 1 second, obtaining a total of 604,800 groups of effective data samples, ensuring the integrity and timeliness of the data. In the data preprocessing stage, the original data is first cleaned to identify and remove outliers. Subsequently, the processed data is divided into a training set, a validation set, and a test set according to a ratio of 6:2:2, containing 362,880, 120,960, and 120,960 groups of data respectively.

[0044] The model of the present invention is implemented based on the Pytorch framework, and the test environment is configured as: Ubuntu20.04 LTS operating system, Intel CPU 2.50GHz, GeForce RTX 4080 GPU. In the frequency prediction model, the loss function is used to measure the difference between the model prediction result and the true result, so as to achieve the backpropagation of errors and update the model parameters. The present invention adopts the root mean square error as the loss function and uses the Adam optimizer to optimize the model parameters. By adjusting the hyperparameters of this model through multiple experiments, the finally determined parameter values are shown in Table 1.

[0045] Table 1. Main parameter settings table of the model

[0046] When evaluating the model performance, this article adopts the root mean square error and the mean absolute percentage error as the measurement indicators. Denote the values of the root mean square error and the mean absolute percentage error as 、 . The calculation formulas of each index are shown in the following formula:

[0047]

[0048] In the formula is the total number of test samples; is the predicted value of the frequency; is the true value of the frequency. 、 The smaller the values of

[0049] , the more accurate the model prediction. For the frequency prediction results under the scenario of power grid downstream disturbance, the fitting analysis of the true frequency curve and the predicted frequency curve shows that the root mean square error ranges from 0.003 to 0.009, and the mean absolute percentage error ranges from 0.005% to 0.010%. This result further verifies the robustness of the model when the power grid frequency fluctuates due to downstream disturbance. The predicted frequency curve can better capture the change trend of the true frequency curve, and at the same time control the prediction error within an acceptable range. To sum up, whether under the condition of upstream disturbance or downstream disturbance, the model shows high prediction accuracy and stability.

[0050] When evaluating the frequency prediction effect on the test set, the obtained root mean square error (RMSE) is 0.00268, and the mean absolute percentage error (MAPE) is 0.004%. To further visually display the prediction results, a fitting curve is drawn, as shown in Figure 4 . In the figure, the solid curve represents the true value, and the dashed curve represents the model predicted value. From Figure 4It can be clearly observed that the model has an ideal fitting effect on the true frequency and can achieve relatively accurate prediction of the frequency change within the next 10 seconds. This result indicates that the proposed model has high accuracy and reliability in the frequency prediction task.

[0051] The present invention further evaluates the frequency prediction performance of the model under the condition of power grid disturbance. As Figure 5 shown, when the power grid is subjected to an upward disturbance and causes frequency fluctuations, the fitting results of the predicted frequency curve and the true frequency curve show that the root mean square error is between 0.003 and 0.005, and the mean absolute percentage error is between 0.005% and 0.009%. This result indicates that even under the influence of an upward disturbance on the power grid frequency, the predicted frequency curve of the model can still closely track the dynamic changes of the true frequency curve, and the prediction error remains at a low level.

[0052] To further illustrate the prediction effect of the method proposed in this paper, it is compared with the models trained by time series prediction algorithms such as TCN, TiDE, and TSMixerx, and the results are shown in Table 2.

[0053] Table 2 Comparison of different modeling methods

[0054] It can be seen from the experimental results in Table 2 that the present model shows significant advantages in the power grid frequency prediction task. Compared with the baseline model that does not consider spatial information, the present model realizes a significant improvement in prediction accuracy by fusing spatio-temporal features. Specifically, the root mean square error of the present model is reduced by 2.44557, 0.00782, and 0.00744 respectively compared with the baseline model, and the mean absolute percentage error is reduced by 0.128%, 0.009%, and 0.008% respectively. Analyzing from the mechanism, traditional time series models only rely on historical time series data and are difficult to fully reflect the spatial characteristics of disturbance propagation in the power system. However, the present model, through the spatio-temporal joint modeling strategy that integrates the power grid topology relationship information, further introduces the topology relationship information between stations on the basis of time dimension modeling, can capture the dynamic characteristics of the power system more comprehensively, makes up for the limitations of single time feature modeling, can better describe the electrical coupling relationship between stations, and can more accurately predict the dynamic changes of frequency, thus significantly improving the accuracy of frequency prediction.

[0055] Embodiment 2 of the present invention proposes a power grid frequency prediction system based on time series prediction using the method described in Embodiment 1 of the present invention, specifically as follows: Adjacency matrix construction module: Construct an adjacency matrix according to the relationship between each station of the power grid; Power grid data acquisition and preprocessing module: Collect historical data of the power grid and perform cleaning, normalization, and time alignment; Power grid frequency prediction model module: Construct a power grid frequency prediction model consisting of an adaptive weight dynamic graph learning module, a linear space convolutional layer, and a temporal convolutional layer; Model training module: Input the adjacency matrix and preprocessed data in the training set into the constructed power grid frequency prediction model, train the model, and determine the model parameter values; Frequency prediction module: Input the data in the test set into the trained power grid frequency prediction model to generate a power grid frequency prediction sequence.

[0056] The present disclosure may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0057] The computer-readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device. The computer-readable storage medium may be, for example—but not limited to—an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device such as a punch card or raised structures in grooves having instructions stored thereon, and any suitable combination of the foregoing. The computer-readable storage medium as used herein is not construed as being a transitory signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0058] The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to respective computing / processing devices, or may be downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include a copper transmission cable, an optical fiber transmission, a wireless transmission, a router, a firewall, a switch, a gateway computer, and / or an edge server. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0059] Computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider). In some embodiments, by using the state information of the computer-readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer-readable program instructions to implement various aspects of the present disclosure.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or make equivalent replacements, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A power grid frequency prediction method based on a time series prediction algorithm, characterized in that, It includes the following steps: Step 1: Collect the historical operation data of the power grid related to the power grid frequency and perform preprocessing to obtain the topological relationship between power grid sites corresponding to the historical data, and construct a power grid adjacency matrix; Step 2: Construct a power grid frequency prediction model, which includes an improved dynamic graph learning module, a linear space convolutional layer, and a linear time convolutional layer connected in sequence; Step 3: Use the historical data collected and preprocessed in Step 1 and the corresponding power grid adjacency matrix as input samples to train the power grid frequency prediction model; Step 4: Collect the current power grid operation data related to the power grid frequency in real time, construct the adjacency matrix of the power grid to be predicted, and input it into the trained power grid frequency prediction model to obtain the prediction result of the power grid frequency.

2. A power grid frequency prediction method based on a time series prediction algorithm according to claim 1, characterized in that: In Step 1, the historical operation data of the power grid related to the power grid frequency is collected, including: Collect frequency and active power data through a Phasor Measurement Unit (PMU), and collect load data through a Supervisory Control and Data Acquisition (SCADA) system.

3. A power grid frequency prediction method based on a time series prediction algorithm according to claim 1, characterized in that: In Step 1, the construction of the power grid adjacency matrix is specifically: Construct an adjacency matrix based on the relationship between each site to represent the topological relationship between power grid sites; in the adjacency matrix, nodes are divided into two categories: the first category is the units that can provide active power data; the second category is the units that can provide load data.

4. A power grid frequency prediction method based on a time series prediction algorithm according to claim 1, characterized in that: In Step 1, data preprocessing includes data cleaning, normalization processing, and time alignment.

5. A power grid frequency prediction method based on a time series prediction algorithm according to claim 1, characterized in that: In Step 2, the improved dynamic graph learning module includes: constructing an adaptive weight block feature, introducing an improved dynamic and static embedding vector to construct a dynamic graph matrix, and optimizing the graph structure based on the attention mechanism.

6. A power grid frequency prediction method based on a time series prediction algorithm according to claim 5, characterized in that: The adaptive weight block feature is specifically: For the historical frequency sequence of the target node, calculate the importance score weight of each time node, and use the concatenation operation to concatenate the adaptive weight features of each node as a block. In the formula, is the weight of each historical time importance score .

7. A power grid frequency prediction method based on a time series prediction algorithm according to claim 6, characterized in that: The importance score is specifically: In the formula, is a non-linear activation function, is a learnable parameter used to measure at the importance at, is the bias term.

8. A power grid frequency prediction method based on a time series prediction algorithm according to claim 5, characterized in that: The improved static embedding vector: In the formula, is the -th data of the vector composed of the active power and load data of the site, is the number of sites.

9. A power grid frequency prediction method based on a time series prediction algorithm according to claim 5, characterized in that: The improved dynamic embedding vector: In the formula, is used to capture timing information, so that can remember historical states and predict future states.

10. A power grid frequency prediction system using the power grid frequency prediction method according to any one of claims 1-9, characterized in that: Power grid data acquisition and preprocessing module: Collect the historical operation data of the power grid related to the power grid frequency and perform preprocessing to obtain the topological relationship between power grid nodes at the corresponding moments of the historical data, and construct a power grid adjacency matrix; Power grid frequency prediction model construction module: Construct a power grid frequency prediction model composed of an improved dynamic graph learning module, a linear space convolution layer, and a temporal convolution layer; Power grid frequency prediction model training module: Use the collected and preprocessed historical data and the corresponding power grid adjacency matrix as input samples to train the power grid frequency prediction model; Frequency prediction module: Collect the current power grid operation data related to the power grid frequency in real time, construct the adjacency matrix of the power grid to be predicted, and input it into the trained power grid frequency prediction model to obtain the prediction result of the power grid frequency.

Images