Knowledge graph platform and method for low-voltage prediction in combination with deep learning

Abstract

The invention relates to the technical field of low-voltage prediction, in particular to a knowledge graph platform and method for low-voltage prediction in combination with deep learning, and the method comprises the steps: collecting and integrating the power grid topological data, equipment parameters, historical operation data and low-voltage response measure knowledge from a power grid business system and a customer service platform, constructing a structured low-voltage knowledge graph; based on the data set extracted from the low-voltage knowledge graph, training is carried out through a deep learning model, and a low-voltage prediction model is generated; an incremental learning process is periodically adopted, and the low-voltage prediction model is iteratively updated according to newly-collected power grid change data; and based on the iteratively updated low-voltage prediction model, predicting a target transformer area, and providing an alarm and a response measure according to a prediction result. Knowledge distillation incremental learning is adopted, it is ensured that the model has high prediction precision for historical and newly-added power grid states, and continuous and stable evolution of the model is achieved.

Description

Technical Field

[0001] The present invention relates to the field of low voltage prediction technology, and in particular to a knowledge graph platform and method for low voltage prediction combined with deep learning. Background Art

[0002] With the increasing penetration of renewable energy, particularly distributed photovoltaic power generation systems, in distribution networks, the operational characteristics of power grids are undergoing profound and complex changes. The integration of renewable energy sources is characterized by significant intermittent and volatile characteristics, posing unprecedented challenges to grid voltage stability. Low-voltage events, a key factor impacting power quality and reliability, are also increasing in frequency and uncertainty. Therefore, to ensure the safe and stable operation of modern power grids, the development of intelligent technologies that can accurately and real-timely predict low-voltage events has become an urgent need within the industry. Currently, existing model updating technologies have significant limitations when dealing with dynamic grid data. On the one hand, when new grid operational data is generated, completely retraining the model by merging historical data with new data can ensure model performance, but this approach incurs significant computational overhead, requires a lengthy training cycle, and consumes significant hardware resources, making it difficult to meet the grid's requirements for rapid model iteration and near-real-time response. On the other hand, fine-tuning existing models using only new data, while efficient, can lead to significant loss of previously learned patterns in historical data as the model learns new knowledge, a phenomenon known as catastrophic forgetting. Not only does it cause the model's predictive ability in historical scenarios to deteriorate sharply, but it also damages the model's overall generalization ability and long-term reliability, making it unable to become a stable and reliable forecasting tool. Summary of the Invention

[0003] The purpose of the present invention is to provide a knowledge graph platform and method for low voltage prediction combined with deep learning. It adopts knowledge distillation incremental learning. On the one hand, it efficiently utilizes new data to update the model, reducing training costs; on the other hand, it effectively solves the catastrophic forgetting problem, ensuring that the model has high prediction accuracy for both historical and newly added power grid states, and realizing the continuous and stable evolution of the model.

[0004] The present invention is achieved through the following technical solutions: The knowledge graph platform that combines deep learning for low voltage prediction includes: The knowledge graph module aggregates and integrates grid topology data, equipment parameters, historical operating data, and low-voltage response measures from the grid business system and customer service platform to build a structured low-voltage knowledge graph. A prediction module, which generates a low voltage prediction model based on a data set extracted from the low voltage knowledge graph and trained by a deep learning model; The iterative update module periodically uses an incremental learning process to iteratively update the low voltage prediction model based on newly collected grid change data; The application module predicts the target substation based on the iteratively updated low voltage prediction model and provides warnings and response measures according to the prediction results.

[0005] Optionally, the constructing of a structured low-voltage knowledge graph further includes: Identify and extract entities, relationships, and attributes from the grid topology data, equipment parameters, historical operating data, and low voltage response knowledge; The extracted entities, relationships and attributes are uniformly converted into a triple data structure to construct the structured low-voltage knowledge graph.

[0006] Optionally, the step of constructing a structured low voltage knowledge graph further includes: The triples are stored in a graph database, and the entities and relationships in the low voltage knowledge graph are graphically displayed using visualization technology.

[0007] Optionally, the specific training steps of the low voltage prediction model are: The deep learning model adopts an LSTM model, and the LSTM model is configured to receive a feature vector integrating internal factors of the power grid and external meteorological factors, wherein the internal factors of the power grid include line length, line diameter, and transformer capacity, and the external meteorological factors include temperature and rainfall; Dividing the data set extracted from the low voltage knowledge graph to obtain a training data set and a test data set for model learning; performing normalization processing on the training data set; The training dataset is input into the LSTM model for training, and the loss value between the prediction results and the test dataset is calculated. The parameters of the LSTM model are updated in reverse based on the loss value until the maximum number of iterations is reached, and the trained low voltage prediction model is output.

[0008] Optionally, the training step further includes: using cross entropy as a calculation function of the loss value to evaluate the difference between the prediction result of the low voltage prediction model and the true value.

[0009] Optionally, the iterative update module includes: an incremental learning triggering process driven by key power grid events: Monitor and identify key grid events, including: newly occurring low voltage events, grid connection events for newly commissioned sites, or completion of distribution network transformation events; When the critical power grid event is identified, an incremental learning process is triggered, and the latest data related to the event is directionally extracted as the newly collected power grid change data.

[0010] Optionally, the incremental learning process adopts the LwF algorithm, and the specific steps are: The old low voltage prediction model is used as a teacher model to predict the power grid change data, and a preset temperature coefficient is applied to smooth the output probability of the teacher model to generate a soft label; Obtaining a distillation loss by comparing the difference between the predicted probability output by the current model for the power grid change data and the soft label; The task loss is obtained by comparing the difference between the predicted probability output by the current model for the power grid change data and the true label corresponding to the data; The distillation loss and the task loss are weightedly summed by balancing hyperparameters to form a combined loss, and the parameters of the low voltage prediction model are optimized and updated based on the combined loss.

[0011] Optionally, providing warnings and response measures based on the prediction results may include: Make predictions for the target substation area and generate dynamic alarms including risk level and impact range when the predicted low voltage probability exceeds the preset threshold; Based on the substation information in the dynamic alarm, query the low voltage knowledge graph to retrieve and match response measures; The dynamic alarms and the matching countermeasures are integrated into a decision support report and output.

[0012] Optionally, providing an alarm and response measures based on the prediction results further includes: The information of the alarm substations where low voltage has been confirmed is reversely injected into the low voltage knowledge graph and updated on the visual interface with a highlighted mark.

[0013] The knowledge graph method for low voltage prediction using deep learning includes the following steps: Gather and integrate grid topology data, equipment parameters, historical operating data, and low-voltage response knowledge from the grid business system and customer service platform to build a structured low-voltage knowledge graph; Generate a low voltage prediction model based on a data set extracted from the low voltage knowledge graph and trained by a deep learning model; Periodically adopt an incremental learning process to iteratively update the low voltage prediction model based on newly collected grid change data; Based on the iteratively updated low voltage prediction model, predictions are made for the target substation, and warnings and response measures are provided according to the prediction results.

[0014] The technical solution of the present invention has at least the following advantages and beneficial effects: By iteratively updating the model using only newly added data, this invention greatly improves training efficiency and significantly reduces computing resource consumption and time costs, enabling rapid and agile model deployment and timely response to the latest changes in grid status. Furthermore, this invention creatively introduces a learning-without-forgetting (LwF) mechanism based on knowledge distillation. During training, the new model not only learns the standard answers to new data but also imitates the decision-making process of the old model when processing this data. This dual-objective optimization strategy not only ensures that the model can accurately grasp emerging grid operating modes, but also effectively consolidates and retains the memory of historical knowledge, perfectly overcoming the problem of catastrophic forgetting. Ultimately, a high-precision low-voltage prediction model is obtained whose performance continues to improve as it evolves. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the process of the knowledge graph method for low voltage prediction combined with deep learning provided by the present invention; Figure 2 Schematic diagram of the knowledge graph platform provided by the present invention. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0017] The present invention provides a knowledge graph platform for low voltage prediction combined with deep learning. Figure 1 and Figure 2 The platform logically and functionally consists of a knowledge graph module, a prediction module, an iterative update module, and an application module. The following details the construction, functions, and collaborative relationships of each module.

[0018] Construction and implementation of the knowledge graph module: The core task of this module is to aggregate and integrate multi-source heterogeneous data to build a structured and visual low-voltage domain knowledge graph, providing a high-quality, easily accessible data foundation for subsequent prediction model training and application decision-making. The specific implementation steps are as follows: 1.1 Aggregation and Integration of Data Sources This platform first uses the data collection module to aggregate various types of data related to low voltage prediction from multiple business systems. These data sources mainly include: power grid business systems: such as marketing systems, electricity consumption collection systems, PMS (production management systems), etc. From them, grid topology data (such as hierarchical topology information of substations-lines-substations), equipment parameters (such as distribution transformer capacity, wire diameter, distribution line length), historical operation data (such as substation load history data, historical low voltage event records, number of substation users, load information of important users) are obtained. Customer service platform: such as the 95598 platform. Unstructured low voltage complaint information is obtained, which can be used as supplementary verification of low voltage events. External data interface: Real-time access to external meteorological data through API and other means, including temperature, wind speed, rainfall, and historical meteorological disaster records.

[0019] 1.2 Knowledge Extraction and Triple Conversion After gathering the raw data, this module performs knowledge extraction on these structured, semi-structured and unstructured data. The core step is to identify and extract the core entities, relations and attributes.

[0020] Entity Definition: Define various entity types based on business needs. For example, equipment entities include substations, lines, substations, and transformers. Environmental entities include meteorological events (such as high temperatures and heavy rain) and geographic locations (such as a county in Dazhou City). Event entities include low voltage events and low voltage complaints.

[0021] Relationship Definition: Defines the association between entities. For example, topological relationships include <line, connection, substation> and <substation, belonging, line>. Action relationships include <low voltage event, action measure, capacitor switching>. Attribute relationships include <substation A, distribution transformer capacity, 500kVA>.

[0022] Triple conversion: Extracted entities, relationships, and attributes are uniformly converted into a triple (Subject, Predicate, Object) data structure. This step is key to building a structured knowledge graph. For example, a piece of information stating that substation A belongs to substation B is converted into the triple <substation A, substation to which it belongs, substation B>. Voltage information obtained from the electricity consumption collection system can be organized into <substation, time period, minimum voltage value>. Unstructured data (such as complaint texts) is parsed manually or through natural language processing techniques, and converted manually or semi-automatically into a standard triple format. Before conversion, data is preprocessed, including deduplication and missing value correction, to ensure data quality.

[0023] 1.3 Graph Database Storage and Visualization Graph database storage: The generated massive triple data is stored in a professional graph database. In this embodiment, the Neo4j graph database is preferably used. Compared with traditional relational databases, graph databases can more efficiently store and query complex relationships between entities, and are very suitable for the application scenarios of knowledge graphs. Visual display: In order to facilitate the understanding and use of operation and maintenance personnel, this module uses front-end visualization technology (such as the D3.js library) to graphically display the knowledge graph stored in the graph database. Users can intuitively see the connection relationships between various substations, lines, and substations, as well as related equipment parameters and historical events on the visual interface, realizing clear visualization of entity-relationships, greatly facilitating data query, analysis, and update.

[0024] Construction and implementation of the prediction module: The core task of this module is to use the high-quality dataset provided by the knowledge graph to train a high-precision low voltage prediction model through deep learning technology. The specific training steps are as follows: 2.1 Model selection and feature vector construction Considering the close correlation between low voltage occurrence and time series characteristics (such as weather changes and peak electricity demand), this embodiment uses a long short-term memory (LSTM) network as the core deep learning model. LSTM is a special type of recurrent neural network (RNN). Through its internal input gate, forget gate, and output gate design, it can effectively capture long-term dependencies in time series data and alleviate the vanishing and exploding gradient problems that are common in traditional RNNs.

[0025] In order to comprehensively predict low voltage, the feature vector of the model input integrates the internal factors of the power grid and the external meteorological factors. Internal factors of the power grid: data that directly reflects the physical characteristics of the power grid, such as wire diameter, distribution transformer capacity, distribution line length, number of users in the substation, etc. These data are directly extracted from the constructed knowledge graph for specific substations. External meteorological factors: external environmental data that affects power load and line status, such as temperature, wind speed, rainfall, extreme weather records, etc. Time series data: time labels such as months, weekdays, holidays, etc., used to capture power consumption patterns. Label data: low voltage labels. In the early stage when the amount of data is limited, it can be a classification task label marked with the presence or absence of low voltage phenomenon (0 or 1).

[0026] 2.2 Dataset Partitioning and Preprocessing Data from the past three years (which can be adjusted according to actual conditions) is extracted from the knowledge graph module to construct an initial dataset. The dataset construction tends to select substations or their neighboring substations where low voltage has occurred. The dataset is then divided into a training dataset for model learning and a test dataset for evaluating model performance. In this embodiment, a 5-fold cross-validation method is used to more robustly evaluate model performance. The original data is divided into five subsets, four of which are used as training sets and one as a validation set in turn, repeated five times, and the average result is taken as the performance indicator of the model. Because the dimensions and numerical ranges of different features in the dataset (such as distribution line length and number of users in the substation) vary greatly, the training dataset needs to be normalized to eliminate the impact of such differences on model training. This embodiment uses the maximum and minimum normalization method to scale the data to the interval [0, 1].

[0027] 2.3 Model Training and Parameter Update Input the normalized training dataset into the configured LSTM model for training.

[0028] After each forward propagation, the model outputs a prediction (e.g., the probability of a low voltage event). To quantify the difference between the prediction and the true label, this example uses cross-entropy as the loss function. The cross-entropy loss function is widely used in classification tasks and effectively measures the difference between two probability distributions.

[0029] Based on the calculated loss value, the gradient of the loss function with respect to each model parameter is calculated through the back-propagation algorithm, and an optimizer (such as Adam) is used to update the network parameters (weights and biases) of the LSTM model along the direction of gradient descent.

[0030] Repeat the above process of forward propagation-loss calculation-backward update until the preset maximum number of iterations is reached or the loss value converges to the target range, and finally output a trained low voltage prediction model with prediction capability.

[0031] Construction and Implementation of the Iterative Update Module: Due to the dynamic nature of grid structure and operating conditions (e.g., equipment upgrades, new users), initially trained models can deviate from actual scenarios over time, leading to decreased prediction accuracy. This module aims to continuously iteratively update the model through incremental learning at a low computational cost, ensuring its long-term effectiveness.

[0032] 3.1 Triggering Process of Incremental Learning This module does not learn continuously but is instead triggered by key grid events. Key Grid Event Monitoring: The system continuously monitors and identifies several key grid events that could significantly change the data distribution. These include: New Low Voltage Events: The first or frequent occurrence of low voltage in a substation; the associated data becomes an important learning sample. New Site Connection Events: New substations, lines, or substations introduce new topologies and load characteristics. Distribution Network Renovation Completion Events: The renovation and upgrade of existing lines, transformers, etc., alters their physical parameters and operating characteristics. Triggering and Data Extraction: When the system identifies any of these key grid events, it automatically triggers the incremental learning process. The system then specifically extracts the latest data directly related to the event from the knowledge graph (such as substation information for the newly commissioned site, equipment parameters in the post-renovation area, and recent operational and meteorological data for the newly low-voltage substation). This data constitutes the newly collected grid change data for model updates. This event-driven, targeted learning approach is more efficient and targeted than periodic full updates.

[0033] 3.2 Specific Algorithm Implementation of Incremental Learning To avoid catastrophic forgetting while learning new knowledge, this embodiment uses the LwF (Learning without Forgetting) algorithm to perform incremental learning. The specific steps are as follows: An existing, trained low-voltage prediction model is used as a teacher model. Newly collected grid change data is fed into this teacher model for prediction, resulting in an output probability. To convey richer information, a preset temperature coefficient T (T > 1) is applied to smooth the teacher model's output probability. The resulting smoothing is called a soft label. Soft labels contain more information about similarities between categories than the original hard labels (0 or 1).

[0034] New data is also input into the model that needs to be updated. It is also smoothed using the temperature coefficient T to obtain its predicted probability. Distillation loss is calculated by comparing the smoothed probabilities output by the current model with the soft labels generated by the teacher model, typically using cross-entropy or KL divergence. The goal of this loss is to enable the new model to learn and emulate the thinking of the old model when processing new data. Simultaneously, task loss is calculated by comparing the (unsmoothed) predicted probabilities output by the current model for the new data with the true labels (hard labels) corresponding to the data. The goal of this loss is to enable the new model to learn the patterns inherent in the new data. The two losses are weighted and summed by balancing the hyperparameter λ to form the final combined loss function. The parameters of the low-voltage prediction model are optimized and updated based on this combined loss. While the model learns the new data distribution, knowledge distillation preserves the generalization capabilities learned from the old data, achieving efficient and stable iterative updates.

[0035] Construction and implementation of the application module: This module is the ultimate value export of the entire platform. It uses the latest prediction model after iterative updates to provide real-time prediction alarms and intelligent decision-making support for power grid operation and maintenance personnel.

[0036] 4.1 Dynamic Alerts and Risk Assessment The application module performs periodic polling predictions on all target areas.

[0037] When the model predicts that the probability of low voltage occurring in a substation exceeds the preset alarm threshold, the system automatically generates a dynamic alarm. This alarm goes beyond simply stating that low voltage will occur and includes more detailed information, such as the predicted risk level (which can be divided by probability), the time period in which the occurrence is likely, and the range of users that may be affected.

[0038] 4.2 Response Measure Matching and Decision Support Once an alarm is generated, the system automatically queries the knowledge graph based on the substation information (such as the substation ID) in the alarm. The knowledge graph then searches for "action relationships" associated with that substation or similar low-voltage events, matching historically effective response measures such as "switching capacitors," "adjusting transformer taps," or "issuing a notice for orderly power consumption."

[0039] The system integrates dynamic alarm information with matched response measures to create a clear decision-making support report, which is then sent to relevant management personnel. This report enables maintenance personnel to quickly understand risk situations and obtain feasible solutions, significantly improving the efficiency and accuracy of emergency response.

[0040] 4.3 Visual Feedback and Closed-Loop Update For substations that have issued alarms, the application module will inject the alarm information back into the knowledge graph module. On the knowledge graph's visual interface, substation nodes in alarm status will be highlighted in a special color (such as red), making them clearly visible in the complex power grid topology.

[0041] Once the alarm is confirmed as a true low-voltage event, all relevant data (including the event itself, the handling process, and the results) will be solidified into the knowledge graph as new knowledge, forming a complete data closed loop. This not only provides valuable samples for the next round of incremental learning, but also continuously enriches and improves the knowledge graph itself.

[0042] In summary, the embodiment of the present invention constructs a knowledge graph module, a prediction module, an iterative update module and an application module. In actual application, it can also include a data acquisition module as a specific reference for data acquisition collaboration. Figure 2 and make them work together, successfully integrating multi-source heterogeneous data, achieving accurate prediction of low voltage events, low-cost iterative updates of models, and intelligent alarm and decision support, providing a complete, efficient and intelligent platform solution for solving low voltage problems in distribution networks.

[0043] The above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A knowledge graph platform combining deep learning for low voltage prediction, characterized by: include: The knowledge graph module aggregates and integrates grid topology data, equipment parameters, historical operating data, and low-voltage response measures from the grid business system and customer service platform to build a structured low-voltage knowledge graph. A prediction module, which generates a low voltage prediction model based on a data set extracted from the low voltage knowledge graph and trained by a deep learning model; The iterative update module periodically uses an incremental learning process to iteratively update the low voltage prediction model based on newly collected grid change data; The application module predicts the target substation based on the iteratively updated low voltage prediction model and provides warnings and response measures according to the prediction results.

2. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 1 is characterized in that: The constructing of a structured low voltage knowledge graph further includes: Identify and extract entities, relationships, and attributes from the grid topology data, equipment parameters, historical operating data, and low voltage response knowledge; The extracted entities, relationships and attributes are uniformly converted into a triple data structure to construct the structured low-voltage knowledge graph.

3. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 2 is characterized in that: The constructing of a structured low voltage knowledge graph further includes: The triples are stored in a graph database, and the entities and relationships in the low voltage knowledge graph are graphically displayed using visualization technology.

4. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 3 is characterized in that: The specific training steps of the low voltage prediction model are as follows: The deep learning model adopts an LSTM model, and the LSTM model is configured to receive a feature vector integrating internal factors of the power grid and external meteorological factors, wherein the internal factors of the power grid include line length, line diameter, and transformer capacity, and the external meteorological factors include temperature and rainfall; Dividing the data set extracted from the low voltage knowledge graph to obtain a training data set and a test data set for model learning; performing normalization processing on the training data set; The training dataset is input into the LSTM model for training, and the loss value between the prediction results and the test dataset is calculated. The parameters of the LSTM model are updated in reverse based on the loss value until the maximum number of iterations is reached, and the trained low voltage prediction model is output.

5. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 4 is characterized in that: The training step further includes: using cross entropy as a calculation function of the loss value to evaluate the difference between the prediction result of the low voltage prediction model and the true value.

6. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 5, characterized in that: The iterative update module includes: an incremental learning triggering process driven by key power grid events: Monitor and identify key grid events, including: newly occurring low voltage events, grid connection events for newly commissioned sites, or completion of distribution network transformation events; When the critical power grid event is identified, an incremental learning process is triggered, and the latest data related to the event is directionally extracted as the newly collected power grid change data.

7. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 6, characterized in that: The incremental learning process adopts the LwF algorithm, and its specific steps are as follows: The old low voltage prediction model is used as a teacher model to predict the power grid change data, and a preset temperature coefficient is applied to smooth the output probability of the teacher model to generate a soft label; Obtaining a distillation loss by comparing the difference between the predicted probability output by the current model for the power grid change data and the soft label; The task loss is obtained by comparing the difference between the predicted probability output by the current model for the power grid change data and the true label corresponding to the data; The distillation loss and the task loss are weightedly summed by balancing hyperparameters to form a combined loss, and the parameters of the low voltage prediction model are optimized and updated based on the combined loss.

8. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 7, characterized in that: Providing warnings and response measures based on prediction results specifically includes: Make predictions for the target substation area and generate dynamic alarms including risk level and impact range when the predicted low voltage probability exceeds the preset threshold; Based on the substation information in the dynamic alarm, query the low voltage knowledge graph to retrieve and match response measures; The dynamic alarms and the matching countermeasures are integrated into a decision support report and output.

9. The knowledge graph platform for low voltage prediction combined with deep learning according to claim 8, characterized in that: Providing warnings and response measures based on the prediction results further includes: The information of the alarm substations where low voltage has been confirmed is reversely injected into the low voltage knowledge graph and updated on the visual interface with a highlighted mark.

10. A knowledge graph method for low voltage prediction combined with deep learning, characterized in that: The steps of the method include: Gather and integrate grid topology data, equipment parameters, historical operating data, and low-voltage response knowledge from the grid business system and customer service platform to build a structured low-voltage knowledge graph; Generate a low voltage prediction model based on a data set extracted from the low voltage knowledge graph and trained by a deep learning model; Periodically adopt an incremental learning process to iteratively update the low voltage prediction model based on newly collected grid change data; Based on the iteratively updated low voltage prediction model, predictions are made for the target substation, and warnings and response measures are provided according to the prediction results.

Images