Intelligent dynamic voltage sag compensation control method and system

Abstract

The invention provides an intelligent dynamic voltage sag compensation control method and system. The method comprises the following steps: acquiring monitoring data from key nodes and equipment of a power grid, and uploading the monitoring data to a cloud for processing to obtain training data; establishing a power grid state prediction model by using deep learning, predicting a voltage trend and a sag occurrence probability through real-time monitoring data, obtaining voltage sag prediction data, positioning a potential source and analyzing a result; training a multi-device cooperative control strategy by adopting deep reinforcement learning, formulating an optimization scheme according to an analysis result, and generating a multi-device cooperative control scheme; and the distributed execution module is used for coordinating and adjusting equipment such as a transformer and a generator set to realize dynamic compensation of voltage sag. By adopting artificial intelligence and Internet of Things technologies, the intelligent and collaborative level is realized, and dynamic analysis, judgment, optimization and correction can be efficiently and accurately carried out on voltage sags of different types and time scales, so that the power grid compensation level is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grids, and specifically relates to an intelligent dynamic voltage sag compensation control method and system. Background Art

[0002] With the progress of technology, digital devices have been widely used in various fields of industry, commerce and residential life. These devices usually rely on the voltage provided by the power grid to work properly. Voltage sag refers to the sudden drop in the effective value of the power frequency voltage at a certain point in the power system, and the duration usually does not exceed 10 seconds, and then returns to normal. Voltage sag may cause serious problems such as digital device crashes, data loss, and unstable motor operation. Summary of the Invention

[0003] Based on the above problems, the present invention proposes an intelligent dynamic voltage sag compensation control method and system. By adopting advanced artificial intelligence and Internet of Things technologies, the intelligent and collaborative level is achieved, and it can dynamically analyze, judge and optimize and correct different types and time scales of voltage sags efficiently and accurately, thus significantly improving the power grid compensation level.

[0004] In view of this, one aspect of the present invention proposes an intelligent dynamic voltage sag compensation control method, including: Identifying each key node and key device of the power grid; Obtaining the time series data of voltage, current, load and environmental parameters from the key nodes and key devices to obtain the first monitoring data; Uploading the first monitoring data to the cloud processing platform; The cloud processing platform preprocesses the first monitoring data and extracts feature quantities to obtain the first training data; Combining the first training data, establishing a power grid state prediction model through deep learning; Real-time monitoring the time series data of voltage, current, load and environmental parameters to obtain the first real-time monitoring data; According to the first real-time monitoring data and the power grid state prediction model, real-time predicting the voltage level trend and judging the occurrence probability of voltage sag to obtain voltage sag prediction data; Locating the potential source of voltage sag according to the voltage sag prediction data to obtain the voltage sag analysis result; Training a multi-device collaborative control strategy by using deep reinforcement learning; Determining an optimization plan according to the voltage sag analysis result; Generating a multi-device collaborative control plan according to the optimization plan and the multi-device collaborative control strategy; Combined with the multi-device collaborative control solution, a distributed execution module is used to coordinately adjust multiple devices including transformers and generator sets to dynamically correct the voltage level and achieve the compensation of voltage sags.

[0005] Preferably, the step of identifying each key node and key device of the power grid includes: Obtain the historical operation data of each node and device in the power grid; Apply a deep learning algorithm to extract features and classify the historical operation data, and find out the node devices with large parameter fluctuation amplitude or whose change mode conforms to the preset special conditions; Combined with the power grid topology structure data, identify the intersection points of the main lines and the load concentration areas; Analyze the historical fault data and identify the first nodes and first devices with the probability of failure greater than the first preset probability; Confirm and improve the first nodes and the first devices to obtain the determination results of each key node and key device of the power grid.

[0006] Preferably, the step of establishing a power grid state prediction model through deep learning in combination with the first training data includes: Divide the first training data into a training data set and a test data set; Select an appropriate deep learning model; Set the neurons in the output layer according to the number of targets and set the number of neurons in the hidden layer; Use the training data set to train the model parameters using an optimizer; Adjust the weights through network backpropagation to minimize the prediction error; Test the model on the test data set, optimize the hyperparameters to reduce the error, obtain the power grid state prediction model, and form the power grid state prediction model program code to be deployed on the cloud processing platform.

[0007] Preferably, the step of obtaining the voltage sag prediction data by predicting the voltage level trend in real time according to the first real-time monitoring data and the power grid state prediction model and judging the occurrence probability of voltage sags includes: Input the first real-time monitoring data into the pre-trained power grid state prediction model; Use the deep learning model in the power grid state prediction model to predict the voltage level at a given time node to obtain the predicted voltage level value; According to the predicted voltage level value, judge whether the voltage drops at a given time point through the classification model in the power grid state prediction model, and obtain the prediction result of the voltage sag occurrence time; Based on the prediction result of the voltage sag occurrence time, combined with the power grid topology structure data, predict the key nodes or devices where the voltage sag occurs to obtain the prediction result of the voltage sag occurrence point; Use the regression model in the power grid state prediction model to predict the duration of the voltage drop and the magnitude of the voltage drop; Integrate the above voltage level prediction values, the prediction result of the voltage sag occurrence time, the prediction result of the voltage sag occurrence point, the duration of the voltage drop, and the magnitude of the voltage drop to obtain complete voltage sag prediction data.

[0008] Preferably, the step of locating the potential source of the voltage sag based on the voltage sag prediction data to obtain the voltage sag analysis result includes: Identify the possible occurrence points and trend mutation regions of the voltage sag according to the voltage sag prediction data; Use a graph neural network to analyze the power grid topology structure data and find the up and down influence relationships; Apply rule-driven or constrained learning methods to infer potential influence sources; Confirm the potential source by integrating historical operation and maintenance data; Output the result analysis report of the possible causes and occurrence locations of the voltage sag.

[0009] Preferably, the step of training a multi-device collaborative control strategy using deep reinforcement learning includes: Establish an environmental scenario simulation including a multi-control device set; Design a reward function to reward the improvement amplitude of the voltage level; Use a deep neural network as the policy network; Obtain the environmental state update and reward by using the actions taken under the policy network; Optimize the network parameters through the policy gradient algorithm to find an efficient policy that meets the preset value criteria; Repeatedly test different collaborative schemes and select the one with the best effect as the multi-device collaborative control strategy.

[0010] Preferably, the step of determining the optimization plan according to the voltage sag analysis result includes: Judge the potential cause of the voltage sag according to the voltage sag analysis result; Generate several alternative restoration plans according to the potential cause; Simulate each alternative restoration plan to obtain the simulation effect index and the simulation cost index; The historical effect indicators and historical cost data extracted from the historical restoration plan are sent into multi-objective decision tree training to obtain a first decision tree model, and the simulated effect indicators and the simulated cost indicators are input into the first decision tree model to obtain a first alternative restoration plan; Select reference samples from historical successful cases; Combined with the reference samples, use the first decision tree model to give a reference plan, and adjust the first alternative restoration plan according to the reference plan to obtain an optimized plan.

[0011] Preferably, the step of generating a multi-device collaborative control plan according to the optimized plan and the multi-device collaborative control strategy includes: Determine the set of target devices to be controlled according to the optimized plan; According to the multi-device collaborative control strategy, match the corresponding device set in the policy network obtained by reinforcement learning; Adjust the policy network structure to match the actual control device attributes; Specify a corresponding control action space for each device; Search for the optimal joint action in the policy network according to the optimization goal; Output a joint control instruction sequence specific to each device to obtain a device collaborative control plan.

[0012] Preferably, the step of combining the multi-device collaborative control plan to perform collaborative adjustment on multiple devices including transformers and generator sets through a distributed execution module to dynamically correct the voltage level and achieve voltage sag compensation includes: Send the multi-device collaborative control plan to the execution module of each execution device; The execution module adjusts its electrical device parameters in real time according to the corresponding control sequence extracted from the device collaborative control plan; The devices cooperate with each other through communication to converge to the target voltage value; Real-time monitor the voltage level condition and perform closed-loop regulation to achieve dynamic compensation of voltage sag within the predicted section.

[0013] Another aspect of the present invention provides an intelligent dynamic voltage sag compensation control system, which is used to execute the intelligent dynamic voltage sag compensation control method as described in any one of the foregoing, including: an Internet of Things platform and a cloud processing platform; wherein, The cloud processing platform is configured to: identify each key node and key device of the power grid; Obtain the time-series data of voltage, current, load, and environmental parameters from key nodes and key devices to obtain the first monitoring data, and upload the first monitoring data to the cloud processing platform; monitor the time-series data of voltage, current, load, and environmental parameters in real time to obtain the first real-time monitoring data; The cloud processing platform is configured to: preprocess the first monitoring data and extract feature quantities to obtain the first training data; Combine the first training data and establish a power grid state prediction model through deep learning; According to the first real-time monitoring data and the power grid state prediction model, predict the voltage level trend in real time and judge the occurrence probability of voltage sags to obtain voltage sag prediction data; Locate the potential source of voltage sags according to the voltage sag prediction data to obtain the voltage sag analysis result; Train a multi-device collaborative control strategy using deep reinforcement learning; Determine an optimization plan according to the voltage sag analysis result; Generate a multi-device collaborative control plan according to the optimization plan and the multi-device collaborative control strategy; Combine the multi-device collaborative control plan and dynamically correct the voltage level through the distributed execution module to perform collaborative adjustment on multiple devices including transformers and generator sets to achieve the compensation of voltage sags.

[0014] Adopting the technical solution of the present invention, the intelligent dynamic voltage sag compensation control method includes: identifying each key node and key device of the power grid; obtaining the time-series data of voltage, current, load and environmental parameters from the key nodes and key devices to obtain the first monitoring data; uploading the first monitoring data to the cloud processing platform; the cloud processing platform preprocesses the first monitoring data and extracts feature quantities to obtain the first training data; combining the first training data, establishing a power grid state prediction model through deep learning; real-time monitoring the time-series data of voltage, current, load and environmental parameters to obtain the first real-time monitoring data; according to the first real-time monitoring data and the power grid state prediction model, real-time predicting the voltage level trend and judging the occurrence probability of voltage sag to obtain voltage sag prediction data; positioning the potential source of voltage sag according to the voltage sag prediction data to obtain the voltage sag analysis result; training a multi-device collaborative control strategy by using deep reinforcement learning; determining an optimization plan according to the voltage sag analysis result; generating a multi-device collaborative control plan according to the optimization plan and the multi-device collaborative control strategy; combining the multi-device collaborative control plan, and performing collaborative adjustment on multiple devices including transformers and generator sets through a distributed execution module to dynamically correct the voltage level, so as to realize the compensation of voltage sag. By adopting advanced artificial intelligence and Internet of Things technologies, the intelligent and collaborative levels are realized, and dynamic analysis, judgment and optimization correction can be efficiently and accurately carried out for different types and time scales of voltage sags, thereby significantly improving the power grid compensation level. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flowchart of an intelligent dynamic voltage sag compensation control method provided by an embodiment of the present invention; Figure 2 is a schematic block diagram of an intelligent dynamic voltage sag compensation control system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] In order to be able to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0017] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0018] In the description, claims and the above-mentioned drawings of this application, terms such as "first", "second", etc. are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0019] Reference to "embodiment" herein means that a specific feature, structure or characteristic described in connection with the embodiment can be included in at least one embodiment of this application. The phrase appears in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0020] Next, with reference to Figures 1 to 2 A smart dynamic voltage sag compensation control method and system provided according to some embodiments of the present invention will be described.

[0021] As Figure 1 shown, an embodiment of the present invention provides a smart dynamic voltage sag compensation control method, including: Identifying each key node and key device of the power grid; Obtaining the time series data of voltage, current, load and environmental parameters from the key nodes and key devices to obtain the first monitoring data; In this step, high-precision time series operation characteristics of the power grid nodes can be obtained, reflecting the dynamic change law; through big data volume support for load pattern learning and analysis, the prediction accuracy can be improved; the transmission is efficient and safe, reducing the communication link pressure; providing basic data to support subsequent in-depth analysis and decision-making; realizing dynamic real-time monitoring of power grid parameters, providing reference for intelligent analysis and decision-making; accurately and efficiently obtaining the first monitoring data set of the power grid, laying a foundation for intelligent analysis.

[0022] Uploading the first monitoring data to the cloud processing platform; In this step, the data obtained from the power grid is centrally managed and monitored on the cloud platform, efficiently and reliably transmitting large-scale distributed real-time data; forming the basis of the power grid operation data set, providing samples for deep learning; standardizing the storage and management of data, facilitating in-depth mining applications; obtaining source data in real time, helping to make accurate decision control; realizing cloud data synchronization and calculation; this solution realizes the safe and efficient upload of monitoring data to the cloud, laying the foundation for intelligent analysis applications.

[0023] The cloud processing platform preprocesses the first monitoring data and extracts feature quantities to obtain first training data; In this step, preprocessing such as denoising and interpolation is performed on the first monitoring data to clean the valid data; statistical methods are applied to extract time-domain features (such as mean, deviation, etc.); signal processing methods are applied to extract frequency-domain features (such as power spectrum, etc.); correlation features between data (such as autocorrelation spectrum) are extracted; the feature dimension is reduced through dimensionality reduction (such as PCA); the corresponding features are identified and integrated using label information; and it is converted into a standard machine learning training sample format for storage. Through this solution, the data quality can be improved, laying a foundation for subsequent analysis; a variety of methods are applied to extract distinct distributed features; high-quality samples with low redundancy are formed, which helps the model to learn; the automation of data preprocessing and feature engineering is realized; high-quality input is provided to support the effective training of the modeling algorithm; and a knowledge base foundation for the operation rules of the power grid is automatically constructed. This step realizes automatic and efficient feature extraction, providing high-quality data support for deep learning.

[0024] Combined with the first training data, a power grid state prediction model is established through deep learning; The time-series data of voltage, current, load, and environmental parameters are monitored in real time to obtain first real-time monitoring data; In this step, data can be read from the management system / platform of the power grid, and the read data is uploaded to the cloud platform in real time through the communication module; the platform performs preliminary verification to identify complete and legal real-time monitoring data segments; the time stamp is extracted, and the data segments are subjected to time series restoration processing; and the monitoring data is stored in the distributed data storage system according to the time series continuity. Through this solution, high-frequency operation dynamic data of power grid nodes can be obtained in real time, a time series data set that truly reflects the real-time state of the power grid is formed, and all-day monitoring coverage of important node parameters is achieved; the source data can be obtained in a timely manner, providing real-time reference for analysis and decision-making; the time dimension is introduced to enrich the ability to describe the operation mechanism of the power grid; and real-time online learning is supported to verify the power grid prediction and control model. This method can efficiently obtain the real-time time series operation data support of important power grid nodes.

[0025] According to the first real-time monitoring data and the power grid state prediction model, the voltage level trend is predicted in real time and the occurrence probability of voltage sags is judged to obtain voltage sag prediction data; Based on the voltage sag prediction data, the potential sources of voltage sags are located to obtain voltage sag analysis results; A multi-device collaborative control strategy is trained using deep reinforcement learning; An optimization plan is determined according to the voltage sag analysis results; According to the optimization plan and the multi-device collaborative control strategy, a multi-device collaborative control plan is generated; In combination with the multi-device collaborative control solution, a distributed execution module is used to collaboratively adjust multiple devices including transformers and generator sets to dynamically correct the voltage level and achieve the compensation of voltage sags.

[0026] Adopting the technical solution of this embodiment, advanced artificial intelligence and Internet of Things technologies are used to achieve intelligent and collaborative levels, and can efficiently and accurately perform dynamic analysis, judgment, and optimization and correction for different types and time scales of voltage sags, thereby significantly improving the grid compensation level.

[0027] In some possible implementation manners of the present invention, the step of identifying each key node and key device of the power grid includes: Obtain the historical operation data (such as physical parameters such as voltage and current) of each node and device in the power grid; In this step, the historical operation data refers to the operation monitoring data generated and recorded by each power grid node and device in the past period of time (such as one year). These data are sourced from various monitoring devices set in the power grid (for example, voltage and current sensors set at key nodes). The historical operation data includes various physical parameter values collected by these monitoring devices, such as voltage values, current values, and other electrical parameters; the historical operation data may also record other operation-related information, such as context information such as the time stamp of the recorded data; obtaining these historical operation data can be achieved by extracting and saving these acquisition records from the monitoring devices, or by extracting these long-term stored data from the historical database of the power grid management system; obtaining the historical operation monitoring data sets of these nodes and devices over a long period provides rich sample references for subsequent data analysis and modeling. In short, obtaining the historical operation data of each node and device in the power grid is to obtain the long-term recorded data of various monitored physical parameters during their past operations.

[0028] Apply a deep learning algorithm to perform feature extraction and classification on the historical operation data to find node devices with large parameter fluctuation amplitudes or whose change modes meet preset special conditions; In this step, preprocess the historical operation monitoring data of each node device collected (such as denoising, standardization, etc.), and apply deep learning methods such as convolutional neural networks or recurrent neural networks to automatically learn implicit features from the time series data, such as extracting abstract feature expressions of operation behavior patterns like periodic features, frequency features, correlation features, etc.; according to the preset classification criteria (such as whether the parameter fluctuation amplitude exceeds a certain threshold, etc.), give the category labels of the monitoring points; use this deep learning classification model to train and predict the historical monitoring data, and find out the node devices that meet specific conditions (such as the fluctuation value exceeds the preset fluctuation value); the special conditions can be set manually or automatically learned based on existing abnormal data; based on this, identify the key regional points of the power grid operation that may have problems or are worthy of key optimization; the purpose of this step is to intelligently identify the key points and abnormalities from a large amount of historical data through deep learning means, providing a reference for subsequent analysis and early warning.

[0029] Combine the power grid topology structure data to identify the intersection points of the main lines and the load concentration areas. In this step, the power grid topology structure data includes the connection relationship information between each node and device. Through the topology structure, the main lines can be analyzed, that is, the lines with the largest load or strong transmission capacity; there are often multiple intersection points on the main lines. Due to the influence of the power grid load distribution, etc.; use network analysis algorithms to identify the important intersection nodes on the main lines. These intersection nodes often have concentrated loads and are prone to power distribution problems; at the same time, by analyzing the aggregation degree of the load point positions, identify the load concentration areas. The load aggregation situation in these areas is complex and is also prone to voltage quality problems. Therefore, in this step, by comprehensively considering information such as the power grid structure and load distribution, intelligently identify the regional points with high operation and maintenance difficulties as the key points.

[0030] Analyze the historical fault data to identify the first nodes and first devices whose probability of occurrence of faults is greater than the first preset probability. In this step, collect various fault data of the power grid within a certain period of time, such as records of line faults, equipment faults, etc.; count the number of faults of each node or device in the historical data; calculate the historical probability of each node or device having a fault according to the total historical data scale; preset a probability threshold as the first preset probability; compare the historical fault probabilities of each node or device with this threshold, and find out those nodes or devices whose historical fault probabilities are higher than the threshold. These nodes or devices with probabilities greater than the threshold are the first nodes and first devices with higher fault risks; the purpose of this step is to identify the key node devices with higher historical fault probabilities through operation and maintenance records as the key prevention and control objects; analyze their potential fault rules to prevent possible problems.

[0031] Confirm and improve the first node and the first device to obtain the determination results of the key nodes and key devices of the power grid.

[0032] Through the solution of this embodiment, potential bottlenecks and key areas of the power grid can be accurately located, improving the monitoring accuracy; refined monitoring is adopted for the identified key nodes and devices to improve the monitoring quality; key objects of concern are provided for subsequent prediction and decision-making to improve efficiency and quality; electronic supporting facilities can be installed in a targeted manner to optimize resource allocation; it plays an important role in aspects such as production safety and power grid operation stability management; the above methods can efficiently and specifically identify the key nodes and devices of the power grid, enhancing the construction level of the smart grid.

[0033] In some possible implementation manners of the present invention, the step of establishing a power grid state prediction model by deep learning in combination with the first training data includes: Divide the first training data into a training data set and a test data set; Select an appropriate deep learning model (such as CNN, RNN, etc.); In this step, according to the task characteristics, different types of deep learning models are selected (such as CNN (Convolutional Neural Network), suitable for processing data with topological structures such as images and sounds; RNN (Recurrent Neural Network), suitable for processing sequential data such as natural language and time series data, with memory capabilities; transformer, used for processing sequence data with relatively loose order, such as machine translation; ResNet, which has good effects in image classification tasks, with a principle similar to CNN but can build deeper networks; GAN, Generative Adversarial Network, used for sample generation tasks); according to whether the task contains features such as time series and spatial structure, select a suitable network type (comprehensive factors such as model complexity, training difficulty, and performance also need to be considered); continuously accumulate experience and rich selection practices to select the deep learning model architecture most suitable for the task requirements; the purpose of this step is to find the best-performing deep learning technical solution based on the task characteristics.

[0034] Set the number of neurons in the output layer according to the number of targets, and set the number of neurons in the hidden layer; In this step, the number of neurons in the output layer depends on the number of target variables of the prediction task. If predicting the category of a discrete variable, the softmax function is used in the output layer, and the number of nodes is equal to the number of categories. If predicting a continuous variable, the linear function is used in the output layer, and the number of nodes is 1. For simultaneous prediction of multiple targets, the number of output layer nodes is the total number of target variables. The number of neurons in the hidden layer is generally initially determined within a range with reference to empirical formulas, such as twice the size of the input and output layers. Then, through actual measurement and iteration adjustment, the size of the hidden layer is reasonably expanded or contracted to improve the effect. The number of hidden layers also needs to be appropriately selected. Generally, it is relatively deep but not too deep to avoid overfitting. The purpose is to set the network output format according to the prediction requirements, and reasonably set the hidden layer space to search for a better model structure.

[0035] Use the training data set to train the model parameters using an optimizer. In this step, the model is trained using the training data set. The role of the training data set is to provide sample data and labels. The model training process requires iterative optimization of the model parameters to make it fit the pattern of the training data. Optimizers (such as SGD, Adam, etc.) are optimization algorithms used to update the values of the model parameters. In each round of iteration, the loss is calculated based on the samples and labels, and the optimizer continuously optimizes the parameters according to the loss gradient. For example, SGD updates a small number of parameters each time in the direction of gradient descent, and Adam has an adaptive learning rate, etc. After multiple rounds of iteration, it helps the model continuously learn the pattern of the training data set.

[0036] Adjust the weights through network backpropagation to minimize the prediction error. Test the model on the test data set, optimize the hyperparameters to reduce the error, obtain the power grid state prediction model, and form the program code of the power grid state prediction model to be deployed on the cloud processing platform.

[0037] In this embodiment, the power grid state prediction model is a general power grid operation rule learning model. From this general model, specialized submodels for predicting specific tasks can be extracted, such as a deep learning model for predicting voltage values, a classification model for judging whether it is decreasing, a regression model for predicting the duration and amplitude, etc. These submodels are all based on the feature extraction ability trained by the original power grid state prediction model. They are fine-tuned using the parameters extracted by the general model and output specific prediction targets.

[0038] The solution of this embodiment realizes automatic learning of power grid rules based on a large amount of real data; extracts implicit feature relationships to achieve efficient prediction; generates a power grid behavior prediction model with high accuracy; realizes intelligent prediction ability to provide support for power grid decision-making; can learn and improve the model online according to new data at any time; forms an effective power grid operation knowledge expression; this method uses deep learning to automatically construct a high-quality power grid prediction model.

[0039] In some possible embodiments of the present invention, the step of predicting the voltage level trend in real time according to the first real-time monitoring data and the power grid state prediction model and judging the occurrence probability of voltage sag to obtain voltage sag prediction data includes: Input the first real-time monitoring data into the pre-trained power grid state prediction model; Use the deep learning model in the power grid state prediction model to predict the voltage level at a given time node to obtain a voltage level prediction value; In this step, the power grid state prediction model is the deep learning model obtained by training with a large amount of historical operation data before; this model implicitly extracts the internal correlation of various operation parameters of the power grid. Given the time point to be predicted as the model input, the deep learning structure (such as CNN, RNN, etc.) inside the power grid state prediction model will perform prediction calculations for the input time point; using the pattern recognition function, infer the voltage level value of each node corresponding to this time point according to historical rules; the predicted voltage level here refers to the corresponding voltage value, which is a specific numerical output; this realizes the numerical prediction of the voltage level at a future time point based on the deep learning model; so this step uses the overall model trained before to perform single-point prediction to obtain the specific voltage prediction value.

[0040] According to the voltage level prediction value, judge whether the voltage drops at the given time point through the classification model in the power grid state prediction model to obtain the prediction result of the voltage sag occurrence time; In this step, when the power grid state prediction model was trained before, different sub-models were generated / extracted for different prediction tasks. Among them, the classification model is used for the binary classification task of judging whether the voltage drops. The voltage level prediction value at the given time point has been obtained. Send this voltage prediction value into the classification model to realize the classification of whether it drops; if the classification result is a drop, it means that the voltage may sag at this predicted time point, otherwise the voltage remains stable; in this way, the prediction result of the possible time range of voltage sag is obtained through classification; using the combination of numerical prediction and classification prediction to make a more refined judgment on the voltage change trend; so this step uses the sub-model to further predict the time when the voltage drops to perform a more refined early warning analysis of voltage quality.

[0041] According to the prediction result of the voltage sag occurrence time, combined with the power grid topology structure data, predict the key nodes or devices where the voltage sag occurs to obtain the prediction result of the voltage sag occurrence point; In this step, the time range when voltage sags may occur has been predicted. Then, by combining the topological structure data of the power grid, the transmission relationship between current and power is analyzed. Through algorithms such as graph neural networks, the key nodes or devices affecting the voltage at this time point are identified, such as the intersection points of the main lines within the affected range and the areas with concentrated loads. These nodes and devices may be the potential cause points for voltage sags. Therefore, based on the topological analysis, the key areas or devices where voltage sags are likely to occur are predicted, and the prediction result of the voltage sag occurrence point refers to these identified nodes or devices. The purpose of this step is to combine time prediction and space prediction to further accurately locate the specific areas or devices where voltage problems may occur.

[0042] Use the regression model in the power grid state prediction model to predict the duration of the voltage drop and the magnitude of the voltage drop. In this step, during the training process of the power grid state prediction model, prediction models for different subtasks are generated / extracted. The regression model is used to predict continuous-value variables. Given the voltage sag occurrence point and time, but not knowing the duration and the magnitude of the drop, these two target variables are input into the regression model for prediction. The model will learn the voltage change pattern based on a large number of historical samples, and then predict the possible duration of the voltage drop under the given conditions. At the same time, it predicts the magnitude of the voltage drop, that is, the numerical depth of the drop. These two prediction results can describe the severity of the voltage problem prediction. By supplementing the voltage problem details through regression prediction, the comprehensiveness of problem analysis and early warning is achieved.

[0043] Integrate the above voltage level prediction values, the prediction results of the voltage sag occurrence time, the prediction results of the voltage sag occurrence point, the duration of the voltage drop, and the magnitude of the voltage drop to obtain complete voltage sag prediction data.

[0044] In this embodiment, the voltage sag prediction data includes the time when a sag may occur, the nodes / devices where the sag occurs, the possible duration, the number of voltage drops, etc.

[0045] The solution of this embodiment can achieve real-time prediction of key indicators of voltage sags, provide the time range when voltage anomalies may occur, identify the key regional nodes or devices where voltage sags may occur, predict the duration and the degree of the voltage drop, provide a reference for compensation control, and the prediction results have stronger guidance and operability, providing a comprehensive reference basis for in-depth analysis of voltage quality problems. This method can systematically integrate and real-time predict multiple indicators of voltage sags.

[0046] In some possible implementation manners of the present invention, the step of locating the potential source of the voltage sag according to the voltage sag prediction data to obtain the voltage sag analysis result includes: Identify the possible occurrence points and trend mutation regions of voltage sags based on the voltage sag prediction data; In this step, through the combined prediction of the previous models, various prediction data related to voltage sags are obtained, such as time, location, duration, etc. These data constitute the prediction description data set of voltage sags; use an algorithm (such as a graph neural network) to discover the internal relationships in the data: find the nodes with obvious drops in voltage prediction values as the potential occurrence points of voltage sags; analyze the node data in the surrounding areas to identify the range of regions where the voltage drop trend may undergo mutations. These two identification results are: the possible occurrence points of voltage sags and the regions where the voltage drop trend may undergo mutations. The purpose of this step is to mine the key clues of voltage problems from the data set obtained by calculating the power grid state prediction and provide a targeted basis for subsequent responses.

[0047] Use a graph neural network to analyze the power grid topology data and find the up-down influence relationships; In this step, the topology of the power grid can be regarded as a graph structure, where the nodes represent various devices / regions and the edges represent the connection relationships. As a deep learning algorithm, the graph neural network can process and learn data of graph structure types; take the power grid topology data as input and pass it to the graph neural network model, and the network will automatically learn various influence relationships between nodes, such as the up-down relationships like the flow direction and intensity of electric energy in the network. Eventually, the upstream and downstream regions of a certain node can be identified; determine which regions or devices have influence relationships and the directionality of the influence; the purpose of this step is to discover more implicit connection rules on the basis of the original topology with the help of the powerful graph neural network algorithm.

[0048] Apply rule-driven or constraint learning methods to infer potential influence sources; Through the previous steps, the up-down relationships and influence paths in the power grid have been found, but it is difficult to directly obtain the exact cause of a certain voltage change. In this step, rule-driven or constraint learning methods can be used: extract expert knowledge, set customized rules or constraint conditions for influence propagation; add these rules to the model learning process; while the model learns actual cases, it also takes into account meeting the set rule constraints; by this way of constraining the model learning, potential influence sources that meet the rules can be better found. The purpose of this step is to combine expert knowledge, guide the model learning process, and specifically infer the potential internal causes of power grid problems; limit the possible space with rules or constraints to improve the accuracy and interpretability of problem positioning.

[0049] Confirm the potential sources (such as load clusters, line faults, etc.) by integrating historical operation and maintenance data; In this step, through multiple previous technical means, potential impact sources are initially inferred. However, these are only based on model learning and may not be 100% accurate. It is necessary to further verify by comprehensively comparing with historical actual operation and maintenance data: check whether relevant problems have occurred in the historical operation and maintenance data of this potential source point, such as whether the load cluster has experienced frequent overload situations, whether there are fault records of different levels such as short circuits in the line, and other historical accident records related to the potential source point. Through direct data matching, the initial inference results are confirmed from multiple dimensions; if multiple pieces of evidence indicate that it is the real cause, the potential source is finally determined; the purpose is to combine historical data and finally identify the true internal cause source of the power grid problem in the way of multi-source corroboration.

[0050] Output an analysis report on the possible causes and occurrence locations of voltage sags.

[0051] In this step, through a series of previous analysis and calculations, potential prediction results of voltage sags are obtained, including a series of prediction indicators such as possible causes, occurrence times, locations, etc. Generate a detailed report based on these modeling analysis results. The content of the report should include: description of the analysis process, various technical means used, each prediction result obtained, the multi-evidence analysis process of possible causes, the finally determined possible causes and locations, and the interpretation of other reference information. The purpose of this step is to display and describe the modeling results in an organized manner by generating a detailed report, providing a scientific basis for subsequent work.

[0052] In this embodiment, the hidden cause area of voltage drop is targeted for positioning, the hidden rules of power grid operation are intelligently inferred, providing an important reference basis for power grid fault handling, improving the efficiency of power grid problem diagnosis and repair, and supporting the decision-making of power grid optimization and improvement work; this method realizes intelligent tracing of the source of voltage problems through in-depth analysis, enhancing the power grid quality monitoring ability.

[0053] In some possible implementation manners of the present invention, the step of training a multi-device collaborative control strategy by using deep reinforcement learning includes: Establish an environmental scenario simulation including a set of multiple control devices; In this step, in order to test and verify the control strategy, it is necessary to formulate a simulation scenario; this scenario should include multiple control devices in the power grid, such as generator sets, substation equipment, etc.; these control devices are in the same virtual environment and maintain the working attributes of the actual devices, such as technical parameters such as the types and capacities of different generator sets; they have logical relationships, influence each other, and can be dynamically regulated; through the environmental scenario, an aggregate of multiple main control points in the power grid is simulated, enabling the strategy to be verified in such a near-real multi-control environment, providing authenticity and a basis for subsequent strategy inspection and optimization; the purpose of this step is to establish a functionally complete and controllable virtual test platform, which is an important basis for subsequent steps.

[0054] Design a reward function to reward the improvement in voltage level; In this step, when using the reinforcement learning method to train the control strategy, a reward function needs to be defined. The reward function determines the target direction of the learning strategy. Here, the control objective is to improve the voltage level. Therefore, a reward function can be defined to measure the magnitude of the voltage change after control. Specifically, it can be set as the absolute value or percentage of the voltage increase after control. The more the voltage increases, the higher the reward value. In this way, the strategy will continuously optimize the operation during the learning process to maximize the voltage increase, thus achieving the ultimate control goal, which is to improve the voltage level. Designing a reasonable reward function is the key to reinforcement learning. Here, rewarding the improvement in voltage level is in line with the control objective.

[0055] Adopt a deep neural network as the policy network; In this step, when using the reinforcement learning method to train the control strategy, a policy network is required to fit and solve the strategy. The role of the policy network is to learn how to select the optimal behavior according to the environmental state. As a powerful function approximation tool, the deep neural network can fit any function. It has very good representation ability and can learn the mapping relationship between states and actions in a complex environment. Here, a deep neural network is selected to establish the policy network model. For example, a convolutional neural network combined with a graph neural network is used to process the input of power grid structure data, and the optimal control behavior in different states is found through learning, so as to realize the solution of the power grid control strategy using deep learning technology. Deep learning has obvious advantages in representation learning ability, so it is very suitable as the policy network model.

[0056] Obtain the environmental state update and reward based on the actions taken under the policy network; In this step, the policy network constructed by the deep neural network will select specific control actions according to the state. After this action is implemented to control the environment through simulation, the environmental state is updated, such as adjusting the power generation of the generator set or connecting a certain line, etc. According to the voltage level change before and after control, it is evaluated whether this action successfully causes a level increase. If it increases, a positive reward is given according to the preset reward function, otherwise a negative reward or no reward is given. The new environmental state and the obtained reward can be used as the input for the next link and continuously iteratively learn. The purpose of this step is to obtain the feedback of the action consequences through trial and error, promote the continuous progress of the policy network, and summarize better control behaviors. The key lies in the closed-loop mechanism of "action - state change - reward" realized by environmental simulation.

[0057] Optimize the network parameters through the policy gradient algorithm to find an efficient strategy that meets the preset value standard; In this step, after the policy network continuously explores the environment, it records the experiences of each state-action-reward. At this time, the policy gradient algorithm can be used to optimize the network parameters and improve the policy level. The policy gradient algorithm calculates how the policy network parameters should be fine-tuned based on the collected samples to maximize the reward. For example, it feeds the gradient information back into the network to increase the likelihood of beneficial actions and reduce harmful actions. As the policy network parameters are continuously optimized, the network can always give higher-value action selections. Through repeated iterations, the policy network can approximate the most efficient control strategy in the environment. By combining environment interaction and optimization algorithms, the policy network can continuously grow and find the best solution for power grid control.

[0058] Repeatedly test different cooperation schemes and select the one with the best effect as the multi-device cooperation control strategy.

[0059] In the previous step, a control policy network was obtained through deep reinforcement learning, but this policy only targets single-device control. In actual power grid control, multi-device cooperation is required. In this step, different multi-device cooperation control schemes (such as centralized command schemes, distributed negotiation schemes, etc.) need to be tested and compared here. Each scheme corresponds to a multi-device version of the policy network. Repeatedly use the reinforcement learning algorithm for optimization and testing, and select the policy scheme with the best performance in the simulation environment. The multi-device cooperation policy network corresponding to this scheme is the required result. The purpose of this step is to select the best-performing cooperation control mode scheme through comparison for the actual multi-device environment.

[0060] In this embodiment, the best multi-device optimization control scheme is automatically learned; considering the influence relationship between devices, true cooperation is achieved; aiming to maximize the voltage quality index, a directly implementable multi-device linkage scheme is generated; the difficulty of manual design is eliminated, and the control optimization efficiency is improved; the control policy has robustness and generality; this method realizes the generation of an automatic optimization control policy through reinforcement learning.

[0061] In some possible implementation manners of the present invention, the step of determining the optimization scheme according to the voltage sag analysis result includes: Judge the potential cause of the voltage sag according to the voltage sag analysis result; In the previous section, a variety of models and algorithms were used for comprehensive analysis to obtain the prediction results of voltage sags, which included relevant prediction indicators such as time, location, and duration. At the same time, multiple rounds of verification were used to obtain the most likely internal cause of the voltage problem. At this time, based on these analysis results, several potential causes of voltage sags can be identified, such as load cluster overload, line failure, equipment abnormality, etc.; if a certain result is judged to be consistent with the trend of multiple prediction indicators and consistent with the advantages of historical data, it can be preliminarily identified as a potential main cause. For example, if the location coincides with the historical overload cluster, the possible cause is judged to be increased load. By comprehensively judging various results, the most inferential internal potential cause of the voltage problem is locked.

[0062] generating several alternative recovery plans based on the potential causes; Through the previous analysis, the potential causes of voltage sag have been confirmed. In this step, a corresponding restoration plan needs to be formulated to solve the voltage problem. According to different potential causes, several types of solutions are proposed, such as: if the cause is excessive load, a load reduction or backup power supply plan is proposed; if the cause is a line fault, a plan to repair or reclose a line is proposed; if the cause is equipment abnormality, a plan to repair or replace the equipment is proposed; and other types of alternative plans. Each plan describes the operation steps and expected results. The purpose of this step is to provide multiple feasible restoration countermeasures for the different potential causes determined by the analysis, for reference in subsequent decision-making.

[0063] Simulate each alternative restoration plan to obtain simulation effect index and simulation cost index; Several types of alternative voltage restoration schemes have been proposed before. In order to select the best scheme, it is necessary to use environmental simulation to evaluate the effect of each scheme: import the operation steps in the scheme into the simulation environment for execution; observe whether the voltage level in the simulation environment is restored after the scheme is implemented; estimate the time and cost investment for the implementation of the scheme; compare the control effect and cost of each scheme; find out which schemes are effective and low-cost in the simulation test; provide reference for selecting the best alternative scheme. The purpose of this step is to further evaluate the feasibility and advantages and disadvantages of the alternative schemes by simulating the actual environment, and provide reference for subsequent decision-making.

[0064] The historical effect index and the historical cost data extracted from the historical restoration scheme are fed into a multi-objective decision tree training to obtain a first decision tree model, and the simulation effect index and the simulation cost index are fed into the first decision tree model to obtain a first alternative restoration scheme; Historical effect indicators and historical cost data extracted from historical restoration plans. These data form multi-objective evaluation factors. Taking these data as the training set, they are fed into a multi-objective decision tree model for training (decision tree is a popular multi-objective training algorithm that can learn complex relationships and find the balance point for optimizing each objective). The trained decision tree model can then judge which alternative plan to prefer based on the newly input effect and cost indicators to achieve the best comprehensive effect. Previously, each alternative voltage restoration plan was simulated and evaluated, and the simulated effect indicators (such as the degree of voltage level recovery) and simulated cost indicators of each plan were collected. The simulated effect indicators and simulated cost indicators obtained from the simulation of each alternative restoration plan are input into the first decision tree model to judge which alternative plan to prefer to achieve the best comprehensive effect. The purpose of this step is to find a rule model for optimal plan selection under the premise of meeting the balance between effect and cost through machine learning.

[0065] Select reference samples from historical successful cases; Currently, some analysis results of voltage problems and alternative restoration plans have been initially obtained through previous modeling work. However, to ensure accuracy, these modeling works still need further verification. In this step, it is proposed to select reference samples from actual successful cases: collect successful cases of similar types of voltage problems in the company's historical operation and maintenance. Such cases should include the full process of problem description, exploration process, cause determination, implementation plan, etc. Select some representative successful cases as reference samples. These samples will have real operability and can be used to verify and optimize the analysis plan or model established in the current work. The purpose of this step is to find samples with strong practicality, enrich the current research, and improve its application value.

[0066] Combined with the reference samples, use the first decision tree model to give a reference plan, and adjust the first alternative restoration plan according to the reference plan to obtain an optimized plan.

[0067] In this step, a certain number of successful case samples collected in the previous steps, which record the whole process of problems, analysis, and solution practices, are now input into the decision tree model. Each sample is regarded as a data point with multiple feature values. The features include problem types, analysis results, solutions adopted, etc. The first decision tree model learns the rules through these modelings. It can give an optimized solution reference for a newly input problem (for example, select a practical solution that was used for a similar problem in the past and had good results). The first alternative restoration plan is adjusted according to the reference plan to obtain an optimized plan (such as adjusting the execution time, execution object, etc.). The purpose of this step is to train a reusable plan selection model using real case samples, give reference plan suggestions for new problems, and then revise / adjust the selected alternative plan according to the reference suggestions to obtain an optimized plan for the voltage problem.

[0068] The solution of this embodiment has clear goals and scientific plan selection. It takes into account both the restoration effect and cost to obtain the optimal solution, considers historical experience to avoid repeating mistakes, realizes automatic restoration plan selection, recommends optimized plans that can be directly implemented, improves the problem handling efficiency and success rate, and ensures the quality of plan selection through data-driven decision support.

[0069] In some possible implementation manners of the present invention, the step of generating a multi-device collaborative control plan according to the optimized plan and the multi-device collaborative control strategy includes: Determine the set of target devices to be controlled according to the optimized plan; In this step, the specific operation objects that may be involved in the optimized plan need to be clarified. For example, for the opening and closing operation of a certain line, it is necessary to determine the specific switch stations for opening and closing, or if it is to adjust the power generation, it is necessary to determine which generating units to adjust. Therefore, according to the requirements of the optimized plan, the range of the possible set of target devices is determined in the power grid equipment topological structure, and the spatial range is narrowed down according to the needs of the optimized plan to obtain a clear list of devices to be controlled. The purpose of this step is to lay a foundation for subsequent specific control operations and clarify the specific device objects that need to be controlled and adjusted.

[0070] Match the corresponding device set in the policy network obtained by reinforcement learning according to the multi-device collaborative control strategy; Previously, a multi-device collaborative control strategy has been obtained using reinforcement learning. This strategy is implemented by a deep learning policy network and can select actions for different states. In this step, according to the previously determined set of target devices, they are corresponded one by one in the policy network structure. For example, several generators or substations are corresponded to several nodes in the policy network. In this way, the policy network can execute control actions for this combination of devices, enabling the previously obtained multi-device collaborative control strategy to be specifically applied to this set of target devices that need to be controlled currently. The purpose of this step is to specifically direct the control strategy to this group of devices that actually need to be controlled in order to perform subsequent actual control operations.

[0071] Adjust the policy network structure to match the actual control device attributes; The previously trained policy network was designed for general multi-device control and was not fine-tuned according to the device situation in a specific power grid. In this step, it is necessary to make adjustments according to the technical attributes of the actual control target device set: understand the main parameters and interaction relationships of these devices, such as the performance curves of different generator sets, the capacity of lines, etc.; accordingly, modify and enrich the attribute expression capabilities of several nodes corresponding to the devices in the policy network, such as adding input parameters or hidden layer structures to improve the accuracy of simulating these devices; at the same time, keep the original functions of the policy network unchanged. The purpose of this step is to deeply optimize the policy network structure according to the actual situation of the actual control object to make it better cooperate with the control requirements.

[0072] Specify the corresponding control action space for each device; The previously output of the policy network is a general control action but not specific to a particular device. In this step, according to the attributes of each control target device, it is necessary to specify a dedicated range of selectable control actions for it, set its controllable parameters according to the device type, etc., such as the power generation range or the maximum change degree of a generator set, the combination selection of the opening and closing states of a certain line, etc.; clarify the specific control action space of each device and establish a "device-control action" mapping table so that when the policy network selects an action, it only considers the allowed range of each device. The purpose of this step is to standardize the output of the policy network to ensure that the control actions given by it are within the feasible range of each specific device.

[0073] Search for the optimal joint action in the policy network according to the optimization goal; After the above fine-tuning, the policy network can output control actions for specific device groups. However, these actions are not combined, and joint adjustment is still needed to find the optimal strategy. By using optimization algorithms (such as the deep Q-learning algorithm), the improvement of power grid operation efficiency is used as the reward function. Based on the decisions of the policy network, the action selections of each device are combined. Rewards are obtained according to the combined effects, and the network parameters are continuously optimized. During the training process, the joint action combination with the maximum reward is searched for. The output result is the optimal joint control strategy for scheduling these specific devices. The purpose of this step is to combine the actions of multiple devices and find the optimal multi-device joint operation plan to achieve the control goal through an optimization algorithm.

[0074] Output the joint control instruction sequence specific to each device to obtain the device collaborative control plan.

[0075] Previously, through deep reinforcement learning, a policy network for multi-device joint control was obtained. This network can output the optimal joint action plan. However, this joint action has not been refined to the level of each individual device. In this step, according to the optimization results, for each specific device, its specific control action in the joint plan is specified (for example: device A changes its power generation to X at time T1; device B switches its switch to state Y at time T2, etc.), forming a complete time-series instruction set, and finally outputting the clear corresponding relationship of "device-action-time". The purpose of this step is to obtain a detailed device collaborative control plan description that can be directly issued for execution to solve the actual control task.

[0076] The solution of this embodiment generates an operation plan by combining the optimization plan and the collaborative control requirements, specifies the specific control tasks and responsible scopes of each device, and coordinates each device in an optimal joint manner to achieve the control goal; eliminates the difficulty of manual allocation, automatically generates an executable plan; improves the control efficiency and quality, and promotes the construction of a smart grid; this method realizes the deep integration of the optimization plan and the collaborative control.

[0077] In some possible implementation manners of the present invention, the step of combining the multi-device collaborative control plan and performing collaborative adjustment on multiple devices including transformers and generator sets through a distributed execution module to dynamically correct the voltage level and achieve voltage sag compensation includes: Send the multi-device collaborative control plan to the execution module of each execution device; The execution module adjusts its electrical equipment parameters in real time according to the corresponding control sequence extracted from the device collaborative control plan (such as adjusting the tap ratio of the transformer, the power output of the generator set, etc.); The devices communicate with each other to collaboratively converge to the target voltage value; In this step, a combined control instruction sequence scheme for each device is obtained. However, at this time, the control objective is the voltage level of the overall power grid, and each device needs to communicate and jointly control to gradually approach the voltage target value: the device implements control to adjust the initial value according to its own instruction, the device transmits new voltage data through the communication network, the device receives the new voltages of other devices, recalculates the combined control error, adjusts itself again according to the error according to the instruction, enters the next control cycle, and continuously repeats the above process. The combined control error gradually converges, and finally the collaborative control of the voltage value and the convergence control target value are achieved. The purpose of this step is to describe how multiple devices complete the overall control task through information interaction.

[0078] Real-time monitor the voltage level condition and perform closed-loop regulation to achieve dynamic compensation for voltage sags within the predicted section.

[0079] The solution of this embodiment verifies the control effect in a real environment and specifically implements it, coordinates multiple devices to conduct electric energy and achieve fine regulation, responds to grid changes in real time to quickly correct voltage anomalies, improves power supply quality and grid operation reliability, realizes the collaborative scheduling ability of smart grid devices, and this method realizes automatic and refined dynamic adjustment of voltage level.

[0080] Please refer to Figure 2 , another embodiment of the present invention provides an intelligent dynamic voltage sag compensation control system, and the system is used to execute the intelligent dynamic voltage sag compensation control method as described above, including: an Internet of Things platform, a cloud processing platform; wherein, The cloud processing platform is configured to: identify each key node and key device of the power grid; Obtain the time-series data of voltage, current, load and environmental parameters from the key nodes and key devices, obtain the first monitoring data, and upload the first monitoring data to the cloud processing platform; real-time monitor the time-series data of voltage, current, load and environmental parameters to obtain the first real-time monitoring data; The cloud processing platform is configured to: preprocess the first monitoring data and extract feature quantities to obtain the first training data; Combine the first training data and establish a power grid state prediction model through deep learning; According to the first real-time monitoring data and the power grid state prediction model, real-time predict the voltage level trend and judge the occurrence probability of voltage sags to obtain voltage sag prediction data; Locate the potential source of voltage sags according to the voltage sag prediction data to obtain the voltage sag analysis result; Adopt deep reinforcement learning to train a multi-device collaborative control strategy; Determine an optimization plan according to the voltage sag analysis result; Generate a multi-device collaborative control solution according to the optimization solution and the multi-device collaborative control strategy; In combination with the multi-device collaborative control solution, a distributed execution module is used to perform collaborative adjustment on multiple devices including transformers and generator sets to dynamically correct the voltage level and achieve compensation for voltage sags.

[0081] It should be known that Figure 2 The block diagram of the intelligent dynamic voltage sag compensation control system shown is only for illustration, and the number of each module shown does not limit the protection scope of the present invention. The intelligent dynamic voltage sag compensation control system provided in this embodiment can be used to execute the solutions of each embodiment of the corresponding intelligent dynamic voltage sag compensation control method. For the specific implementation process, please refer to the descriptions of each method embodiment and will not be elaborated here.

[0082] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0083] In the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0084] In several embodiments provided by the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the above division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be in an electrical or other form.

[0085] The units described as separate components above may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0086] In addition, in each embodiment of the present application, the functional units can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0087] If the above integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the above methods in each embodiment of the present application. The aforementioned memory includes: USB flash drives, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), mobile hard disks, magnetic disks, or optical discs and other media that can store program codes.

[0088] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program. This program can be stored in a computer-readable memory. The memory can include: flash drives, read-only memories (abbreviation: ROM), random access memories (abbreviation: RAM), magnetic disks, or optical discs, etc.

[0089] The above has introduced the embodiments of the present application in detail. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

[0090] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions without departing from the spirit and scope of the present invention, and can make various modifications and alterations, including combinations of the above different functions and implementation steps, including software and hardware implementation manners, all within the protection scope of the present invention.

Claims

1. An intelligent dynamic voltage sag compensation control method [admin1], characterized in that, Including: Identifying each key node and key device of the power grid; Obtaining time-series data of voltage, current, load, and environmental parameters from the key nodes and key devices to obtain first monitoring data; Uploading the first monitoring data to the cloud processing platform; The cloud processing platform preprocesses the first monitoring data and extracts feature quantities to obtain first training data; Combining the first training data, establishing a power grid state prediction model through deep learning; Real-time monitoring of time-series data of voltage, current, load, and environmental parameters to obtain first real-time monitoring data; According to the first real-time monitoring data and the power grid state prediction model, real-time predicting the voltage level trend and judging the occurrence probability of voltage sags to obtain voltage sag prediction data; Locating the potential source of voltage sags according to the voltage sag prediction data to obtain voltage sag analysis results; Training a multi-device collaborative control strategy using deep reinforcement learning; Determining an optimization plan according to the voltage sag analysis results; Generating a multi-device collaborative control plan according to the optimization plan and the multi-device collaborative control strategy; Combining the multi-device collaborative control plan, and dynamically correcting the voltage level through the distributed execution module to perform collaborative adjustment on multiple devices including transformers and generator sets to achieve compensation for voltage sags.

2. The intelligent dynamic voltage sag compensation control method according to claim 1, wherein The step of identifying each key node and key device of the power grid includes: Obtaining historical operation data of each node and device in the power grid; Applying a deep learning algorithm to extract features and classify the historical operation data, and finding node devices with large parameter fluctuation ranges or whose change patterns meet preset special conditions; Combining power grid topology structure data to identify the intersection points of main lines and load concentration areas; Analyzing historical fault data to identify the first nodes and first devices with a failure probability greater than a first preset probability; Confirming and improving the first nodes and the first devices to obtain the determination results of each key node and key device of the power grid.

3. The intelligent dynamic voltage sag compensation control method according to claim 2, characterized in that, The step of combining the first training data and establishing a power grid state prediction model through deep learning includes: Dividing the first training data into a training data set and a test data set; Selecting an appropriate deep learning model; Setting the number of output layer neurons according to the number of targets and setting the number of hidden layer neurons; Using the training data set to train model parameters using an optimizer; Adjusting the weights through network backpropagation to minimize the prediction error; Testing the model on the test data set, optimizing hyperparameters to reduce errors, obtaining a power grid state prediction model, and forming a power grid state prediction model program code for deployment on the cloud processing platform.

4. The intelligent dynamic voltage sag compensation control method according to claim 3, wherein, The step of real-time predicting the voltage level trend and judging the occurrence probability of voltage sags according to the first real-time monitoring data and the power grid state prediction model to obtain voltage sag prediction data includes: Inputting the first real-time monitoring data into the pre-trained power grid state prediction model; Using the deep learning model in the power grid state prediction model to predict the voltage level at a given time node to obtain a voltage level prediction value; According to the predicted value of the voltage level, determine whether the voltage drops at a given time point through the classification model in the power grid state prediction model, and obtain the prediction result of the voltage sag occurrence time; According to the prediction result of the voltage sag occurrence time, combined with the power grid topology structure data, predict the key nodes or devices where the voltage sag occurs, and obtain the prediction result of the voltage sag occurrence point; Apply the regression model in the power grid state prediction model to predict the duration of the voltage drop and the magnitude of the voltage drop; Integrate the above predicted value of the voltage level, the prediction result of the voltage sag occurrence time, the prediction result of the voltage sag occurrence point, the duration of the voltage drop, and the magnitude of the voltage drop to obtain complete voltage sag prediction data.

5. The intelligent dynamic voltage sag compensation control method according to claim 4, wherein The step of locating the potential source of the voltage sag according to the voltage sag prediction data to obtain the voltage sag analysis result includes: Identify the possible occurrence points and trend mutation regions of the voltage sag according to the voltage sag prediction data; Use the graph neural network to analyze the power grid topology structure data and find the upstream and downstream influence relationships; Apply rule-driven or constrained learning methods to infer potential influence sources; Confirm potential sources by integrating historical operation and maintenance data; Output the result analysis report of the possible causes and occurrence locations of the voltage sag.

6. The intelligent dynamic voltage sag compensation control method according to claim 5, wherein The step of training a multi-device collaborative control strategy using deep reinforcement learning includes: Establish an environmental scenario simulation including a set of multi-control devices; Design a reward function to reward the improvement amplitude of the voltage level; Use a deep neural network as the policy network; Obtain environmental state updates and rewards by taking actions under the policy network; Optimize network parameters through the policy gradient algorithm to find an efficient policy that meets the preset value criteria; Repeatedly test different collaborative schemes and select the one with the best effect as the multi-device collaborative control strategy.

7. The intelligent dynamic voltage sag compensation control method according to claim 6, wherein The step of determining an optimization plan according to the voltage sag analysis result includes: Judge the potential causes of the voltage sag according to the voltage sag analysis result; Generate several alternative restoration plans according to the potential causes; Simulate each alternative restoration plan to obtain simulation effect indicators and simulation cost indicators; Send the historical effect indicators and historical cost data extracted from the historical restoration plans into the multi-objective decision tree for training to obtain the first decision tree model, and input the simulation effect indicators and the simulation cost indicators into the first decision tree model to obtain the first alternative restoration plan; Select reference samples from historical successful cases; Combine the reference samples, use the first decision tree model to give a reference plan, and adjust the first alternative restoration plan according to the reference plan to obtain an optimization plan.

8. The intelligent dynamic voltage sag compensation control method according to claim 7, characterized in that The step of generating a multi-device collaborative control plan according to the optimization plan and the multi-device collaborative control strategy includes: Determine the set of target devices to be controlled according to the optimization plan; According to the multi-device collaborative control strategy, match the corresponding device set in the policy network obtained by reinforcement learning; Adjust the policy network structure to match the actual control device attributes; Specify the corresponding control action space for each device; Search for the optimal joint action in the policy network according to the optimization goal; Output the combined control instruction sequence for each device to obtain a device collaborative control solution.

9. The intelligent dynamic voltage sag compensation control method according to claim 8, wherein The step of combining the multi-device collaborative control solution and dynamically correcting the voltage level through collaborative adjustment of multiple devices including transformers and generator sets by a distributed execution module to achieve voltage sag compensation includes: Send the multi-device collaborative control solution to the execution module of each execution device; The execution module adjusts its electrical device parameters in real time according to the corresponding control sequence extracted from the device collaborative control solution; The devices converge on the target voltage value through communication and collaborative control; Monitor the voltage level status in real time and perform closed-loop regulation to achieve dynamic compensation for voltage sags within the predicted section.

10. An intelligent dynamic voltage sag compensation control system, characterized in that, The system is used to execute the intelligent dynamic voltage sag compensation control method according to any one of claims 1 to 9, including: an Internet of Things platform and a cloud processing platform; wherein, The cloud processing platform is configured to: identify each key node and key device of the power grid; Obtain the time-series data of voltage, current, load, and environmental parameters from the key nodes and key devices to obtain first monitoring data, and upload the first monitoring data to the cloud processing platform; monitor the time-series data of voltage, current, load, and environmental parameters in real time to obtain first real-time monitoring data; The cloud processing platform is configured to: preprocess the first monitoring data and extract feature quantities to obtain first training data; Combine the first training data and establish a power grid state prediction model through deep learning; According to the first real-time monitoring data and the power grid state prediction model, predict the voltage level trend in real time and judge the occurrence probability of voltage sags to obtain voltage sag prediction data; Locate the potential source of voltage sags according to the voltage sag prediction data to obtain a voltage sag analysis result; Train a multi-device collaborative control strategy using deep reinforcement learning; Determine an optimization plan according to the voltage sag analysis result; Generate a multi-device collaborative control plan according to the optimization plan and the multi-device collaborative control strategy; Combine the multi-device collaborative control plan and dynamically correct the voltage level through collaborative adjustment of multiple devices including transformers and generator sets by a distributed execution module to achieve voltage sag compensation.

Images