Modular risk prediction method and device for power distribution terminal, and storable medium
Through the modular risk prediction method, multi-level stacked integrated learning model and isolated forest model are used to collect and process multi-source data of power distribution terminals, realizing risk prediction of power distribution terminals, and solving the universality and adaptability in the existing technology The problem of insufficient sexuality improves the accuracy of prediction and the flexibility of the system.
Patent Information
- Application Number
- CN202410958055.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-12
AI Technical Summary
The risk prediction methods of existing power distribution terminals are insufficient in terms of versatility and adaptability, and the risks of specific components or subsystems cannot be accurately identified, resulting in low prediction accuracy.
Modular risk prediction method is adopted to collect multi-source historical data, form a terminal historical database, extract multi-source historical data of the working module, perform preprocessing and model training, and use multi-level stacked integrated learning model and isolated forest model for real-time detection and Warning, calculate the overall risk probability.
It improves the accuracy and adaptability of risk prediction of power distribution terminals, and can manage and monitor various working modules more refinedly, reduce errors, improve overall prediction accuracy and system flexibility and reliability.
Smart Images

Figure CN118941075A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of operation, maintenance and management of power distribution terminals, and specifically to a modular risk prediction method, device and storable medium for power distribution terminals. Background Art
[0002] The distribution terminal is a key device in the power system, used for the distribution and management of electric energy in the distribution network. It is usually located after the distribution transformer and is responsible for transmitting electric energy from the distribution network to the final power user. In the power system, the reliability and safety of the distribution terminal equipment are directly related to the operational safety and reliability of the entire power grid. Since the failure rate of the distribution terminal equipment itself is extremely high, once a failure occurs, it will bring great safety hazards to the power grid. Therefore, it is particularly important to predict the risks of the distribution terminal equipment.
[0003] In existing research, the Chinese patent application number CN201811347988.3 proposes a modular risk prediction method, device and storable medium for distribution terminals. First, the historical operation and maintenance data of the distribution terminal are counted and a fault database is generated; the operation and maintenance indicator data of each working module are obtained from the database according to the preset detection indicators; the entropy weight coefficient of the operation and maintenance indicator data is calculated by the entropy weight method, and the operation and maintenance indicator data is screened and classified according to the correlation between the detection indicators and the working modules, and the risk probability value of each module is obtained by weighted summation, and finally the reliability prediction value of the distribution terminal is calculated by the upper and lower limit method. However, this method is designed for a specific distribution terminal and cannot be directly applied to other types of equipment or systems, and its versatility is poor. In addition, the preset detection indicators and correlation relationships cannot be adjusted in time in actual applications, resulting in poor adaptability of the method to new fault types and operation and maintenance indicators. The Chinese patent application number CN202010379002.1 proposes a distribution box early warning system and early warning method, and the system includes a collection unit, a network unit, a processing unit and a terminal device. The acquisition unit obtains the monitoring data in the distribution box and transmits the data to the processing unit through the network unit. The processing unit inputs the monitoring data into the preset trend prediction model for prediction and generates warning information. The terminal device receives and displays these warning information. The acquisition unit also obtains the age data of each electrical equipment in the distribution box and sends it to the processing unit through the network unit to determine whether to generate age warning information. The system uses part of the monitoring data as sample data for learning and training, and the rest of the data as verification data. The specific steps include calculating the node outputs of the hidden nodes and the output layer, determining the number of hidden nodes, and calculating the output error between the actual and the target until the error reaches the expected value to terminate the learning. However, the system uses the distribution box as a whole for risk warning, and may not be able to accurately identify and predict the risks of specific components or subsystems, resulting in low overall prediction accuracy. The Chinese patent application number CN202211039556.2 proposes a distribution terminal status assessment method based on a modified weighted fault tree model, including obtaining multi-dimensional influencing factors of the distribution terminal, constructing a fault tree model including fault types and functional modules, calculating the correlation between fault types and functional modules through an association mining algorithm, and modifying the fault tree model based on the correlation to form a weighted fault tree model, and then using the model to assess the status of the distribution terminal. Due to the complexity of the model, errors or data deviations in any link may be propagated and amplified throughout the entire assessment process, affecting the accuracy and reliability of the final assessment results.
[0004] At present, the accuracy of risk prediction for distribution terminals still needs to be further improved.
[0005] To this end, a modular risk prediction method, device and storable medium for distribution terminals are proposed. Summary of the invention
[0006] The object of the present invention is to provide a modular risk prediction method, device and storable medium for distribution terminals, collect multi-source historical data of a specified distribution terminal, including environmental data, equipment status data, equipment operation data, load data and historical event data, to form a terminal historical database; extract multi-source historical data of N working modules from the terminal historical database to form N module historical data sets, these working modules include terminal equipment module, monitoring and metering module, relay protection module, communication module, control module, power management module, power quality module, safety management module, environmental monitoring module and auxiliary module; pre-process the historical data sets of N working modules respectively to obtain N pre-processing sets; perform model training according to the pre-processing sets respectively to obtain N multi-level stacked integrated learning models; collect real-time data of N working modules of the distribution terminal to obtain N first real-time data; use an isolation forest model to detect each real-time data to obtain N first abnormal data and N first normal data, and issue an early warning based on the first abnormal data; input the first normal data into the multi-level stacked integrated learning model to obtain N module risk probabilities, and calculate the overall risk probability based on the N module risk probabilities. Finally, the module risk probability is compared with the set threshold, and an early warning is issued when the threshold is exceeded.
[0007] To achieve the above object, the present invention provides the following technical solutions:
[0008] A modular risk prediction method for a power distribution terminal, comprising:
[0009] Collect multi-source historical data of the designated power distribution terminal, including environmental data, equipment status data, equipment operation data, load data and historical event data, to obtain a terminal historical database;
[0010] Extracting the multi-source historical data of N working modules from the terminal history database to form N module history data sets; the N working modules include a terminal device module, a monitoring and metering module, a relay protection module, a communication module, a control module, a power management module, a power quality module, a safety management module, an environmental monitoring module and an auxiliary module; preprocessing the N module history data sets respectively to obtain N preprocessing sets;
[0011] According to the N preprocessing sets, model training is performed respectively to obtain N multi-level stacked integrated learning models; each of the working modules corresponds to one multi-level stacked integrated learning model;
[0012] Collecting real-time data of the N working modules of the power distribution terminal to obtain N first real-time data;
[0013] Using an isolation forest model to detect each of the first real-time data respectively, to obtain N first abnormal data and N first normal data;
[0014] Obtain an abnormal module according to the N first abnormal data, and issue an early warning for the abnormal module;
[0015] Inputting N of the first normal data into N of the multi-level stacked ensemble learning models respectively to obtain N module risk probabilities;
[0016] Calculating the overall risk probability of the power distribution terminal according to the risk probabilities of the N modules;
[0017] An early warning is issued based on the risk probabilities of the N modules and the risk thresholds of the N modules.
[0018] Furthermore, the historical event data includes historical power distribution terminal fault information, and the power distribution terminal fault information includes faulty equipment and fault type.
[0019] Furthermore, the working module is specifically:
[0020] The terminal equipment modules are used to control and distribute electrical energy, and the terminal equipment modules include various types of circuit breakers, switches and distribution transformers;
[0021] The monitoring and metering module is used to monitor the electric energy parameters and measure the electric energy consumption in real time, and the monitoring and metering module includes various sensors, electric energy meters and data acquisition systems;
[0022] The relay protection module is used to protect the equipment and line safety of the power distribution terminal; the relay protection module includes a relay protection device and a protection relay;
[0023] The communication module is used to realize data communication between the power distribution terminal and the upper control center and / or designated equipment; the communication module includes a communication terminal, a router and a network switch;
[0024] The control module is used to implement local and / or remote control functions to enable the power distribution terminal to operate normally; the control module includes a controller and a programmable logic controller;
[0025] The power management module is used to manage the supply and distribution of power; the power management module includes a power management system and an uninterruptible power supply;
[0026] The power quality module is used to monitor and improve power quality, and the power quality module includes a filter and a capacitor bank;
[0027] The security management module is used to ensure the physical security and information security of the power distribution terminal; the security management module includes a video monitoring system, a firewall and a security gateway;
[0028] The environmental monitoring module is used to monitor the environmental parameters of the environment in which the power distribution terminal is located; the environmental monitoring module includes a temperature and humidity sensor and an air quality monitor;
[0029] The auxiliary module is used to provide supplementary and extended functions of the power distribution terminal.
[0030] Furthermore, according to the N preprocessing sets, model training is performed respectively to obtain the N multi-level stacked ensemble learning models, which are specifically:
[0031] Dividing the preprocessed set of each module into a training set, a validation set and a test set;
[0032] According to the training set of each module, M basic machine learning models are trained respectively, including LightGBM model, XGBoost model, random forest model, SVM model and DNN model; M represents the total number of basic machine learning models; each basic machine learning model uses a different algorithm, and each basic machine learning model is the optimal model obtained by training with the corresponding algorithm;
[0033] For each of the modules, use the M basic machine learning models to predict the test set to generate a first-level prediction result data set; the first-level prediction result data set includes the prediction results of each of the basic machine learning models;
[0034] Using the first-level prediction result data set to train the first-layer stacking model, and output the prediction result;
[0035] Taking the prediction results of the first layer stacking model as input, generating a secondary prediction result data set;
[0036] Using the secondary prediction result dataset to train a second-layer stacking model;
[0037] The first layer stacking model and the second layer stacking model are integrated to obtain the multi-level stacking integrated learning model.
[0038] Furthermore, the multi-level stacked integrated learning model also includes an adaptive weighted fusion layer;
[0039] The adaptive weighting layer comprises:
[0040] Using the validation set to evaluate the performance of each of the basic machine learning models, and calculating the weight of each of the basic machine learning models to obtain M basic weights;
[0041] Dynamically adjust each of the basic weights using an optimization algorithm to obtain M adjusted weights;
[0042] Using the M adjustment weights, weighted combination is performed on the prediction results of each of the basic machine learning models to generate a weighted prediction result;
[0043] The weighted prediction results are used as the first-level prediction result data set.
[0044] Furthermore, the first layer stacking model and the second layer stacking model are specifically:
[0045] The first layer of stacked models includes linear regression, random forest and DNN models;
[0046] The second layer of stacked models includes an XGBoost model, an LSTM model, and an MLP model.
[0047] Further, according to the risk probabilities of the N modules, calculating the overall risk probability of the power distribution terminal includes:
[0048] A node is defined for each working module to obtain N nodes, each node represents the working state of the corresponding working module; a node is added to represent the overall state of the power distribution terminal; the working state includes normal and fault;
[0049] Defining directed edges between nodes according to the terminal history database and / or expert knowledge to obtain a Bayesian network includes the following steps:
[0050] Acquire N module historical data sets from the terminal historical database; preprocess the module historical data sets, including missing value processing and outlier processing, to obtain module preprocessed historical data sets; perform correlation analysis based on the module preprocessed historical data sets, calculate the correlation coefficients between the modules, and obtain a module correlation coefficient matrix;
[0051] According to the module correlation coefficient matrix, setting an initial structure and / or constraint conditions;
[0052] Using the Hill Climb Search algorithm to perform a structural search on the module preprocessing historical data set, during the search process, performing an optimization search based on the initial structure and / or the constraint conditions to obtain multiple candidate structures; calculating the BIC score of each candidate structure to obtain multiple candidate BIC scores; selecting the candidate structure corresponding to the lowest candidate BIC score as the final output to obtain the Bayesian network;
[0053] Establishing a conditional probability table; the conditional probability table represents the probability distribution of a node under a given condition of its parent node;
[0054] The risk probabilities of the N modules are input into the Bayesian network, and the overall risk probability of the power distribution terminal is calculated by a variable elimination algorithm.
[0055] Furthermore, according to the risk probabilities of the N modules and the risk thresholds of the N modules, the warning is specifically performed as follows:
[0056] The module risk threshold is set for each module; when the module risk probability of the module is greater than the module risk threshold, the module is identified as a risk module; and an early warning is issued for the risk module.
[0057] A modular risk prediction device for a power distribution terminal, comprising:
[0058] A multi-source historical data collection unit is used to collect multi-source historical data of a specified power distribution terminal, including environmental data, equipment status data, equipment operation data, load data and historical event data, to obtain a terminal historical database;
[0059] A module data extraction unit, used to extract the multi-source historical data of N working modules from the terminal history database to form N module history data sets; the N working modules include a terminal device module, a monitoring and metering module, a relay protection module, a communication module, a control module, a power management module, a power quality module, a safety management module, an environmental monitoring module and an auxiliary module;
[0060] A data preprocessing unit, used for preprocessing the N module historical data sets respectively to obtain N preprocessed sets;
[0061] The model building unit is used to perform model training according to the N preprocessing sets to obtain N multi-level stacked integrated learning models; each working module corresponds to one multi-level stacked integrated learning model.
[0062] A real-time data collection unit, used to collect real-time data of the N working modules of the power distribution terminal to obtain N first real-time data;
[0063] an anomaly detection unit, used to detect each of the first real-time data using an isolation forest model to obtain N first anomaly data and N first normal data;
[0064] A module early warning unit, configured to obtain an abnormal module according to the N first abnormal data, and issue an early warning for the abnormal module;
[0065] A risk probability calculation unit, used for inputting the N first normal data into the N multi-level stacked integrated learning models respectively to obtain N module risk probabilities;
[0066] An overall risk assessment unit, used to calculate the overall risk probability of the power distribution terminal according to the risk probabilities of the N modules;
[0067] The comprehensive early warning unit is used to issue an early warning based on the risk probabilities of the N modules and the risk thresholds of the N modules.
[0068] A computer storable medium stores a computer program, which implements the method proposed in the present disclosure when executed by a processor.
[0069] Compared with the prior art, the present invention has the following beneficial effects:
[0070] 1. The present invention sorts out the working modules of the distribution terminal, and divides the working modules into terminal equipment module, monitoring and metering module, relay protection module, communication module, control module, power management module, power quality module, safety management module, environmental monitoring module and auxiliary module. The setting of auxiliary modules provides flexible expansion space for the system. According to specific needs, it is possible to easily increase or decrease modules and expand functions to meet the needs of different application scenarios and improve the adaptability and flexibility of the system. Independently dividing the various functional modules of the distribution terminal can achieve more refined management and monitoring, and can accurately identify faulty equipment and modules in historical data. Through modular design and detailed functional division, each working module runs independently but works together to provide high-quality data support for the risk prediction model. It helps to improve the accuracy of risk prediction.
[0071] 2. The present invention introduces a multi-level stacked integrated learning model. Specifically, by training multiple basic machine learning models (such as LightGBM, XGBoost, random forest, SVM and DNN), and using the stacked model to further integrate and optimize the prediction results of these models, it is possible to comprehensively utilize the advantages of each model and reduce the errors and deviations of a single model, thereby improving the accuracy of the overall prediction. The multi-level structural design ensures the comprehensive utilization of data and the robustness of the model, enabling the system to more effectively respond to complex risk scenarios. In addition, by independently processing and training the preprocessing sets of each module, the module risk prediction probability error of the N working modules is reduced, further improving the overall risk prediction accuracy of the distribution terminal.
[0072] 3. Utilize the terminal historical database and correlation analysis, combined with the Hill Climb Search algorithm to perform data-driven structural learning and optimization search to ensure the scientificity and rationality of the Bayesian network structure. The Bayesian network can dynamically adjust to adapt to the complexity and changes in practical applications. Through the conditional probability table and variable elimination algorithm, it can effectively handle the uncertainty and noise in the data to ensure the accuracy and robustness of the prediction results. The modular design ensures the independence and integrity of each module, making the system have good scalability and maintainability. At the same time, this method realizes real-time input and rapid calculation of the overall risk probability, improves the efficiency and flexibility of risk management, provides reliable risk prediction support, and helps to improve the accuracy of risk prediction for the entire distribution terminal. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1 A flow chart of a power system provided by an embodiment of the present invention;
[0074] Figure 2 A flow chart of a modular risk prediction method for a power distribution terminal provided by an embodiment of the present invention;
[0075] Figure 3 A structural diagram of a multi-level stacked ensemble learning model provided by an embodiment of the present invention;
[0076] Figure 4 A structural diagram of a modular risk prediction device for a power distribution terminal provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0077] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0078] Embodiment 1:
[0079] like Figure 1 As shown in the figure, the power system process covers the complete process from power generation to end-user use. Power distribution is an important link between power transmission and end-users. The power distribution system needs to ensure the reliability and stability of power supply, and timely detect and handle faults through various protection devices and intelligent control systems to prevent power outages and ensure continuous and stable power supply for users.
[0080] Industrial Park A has multiple production workshops and office areas, which have high requirements for the reliability and safety of power supply. The power distribution system of the industrial park needs real-time monitoring and prediction to ensure that the production process is not affected by power problems and to ensure the normal operation of the office environment. Industrial Park A has applied a modular risk prediction method for power distribution terminals, including:
[0081] See also Figure 2 S10 in which multi-source historical data of a designated power distribution terminal is collected, including environmental data, equipment status data, equipment operation data, load data, and historical event data, to obtain a terminal historical database;
[0082] Table 1. Sample table of environmental data
[0083]
[0084] Table 2. Sample table of equipment status data
[0085]
[0086] Table 3. Sample table of equipment operation data
[0087]
[0088] Table 4. Sample load data table
[0089]
[0090] Table 5. Sample table of historical event data
[0091]
[0092] As an embodiment of the present invention, samples of environmental data are shown in Table 1; samples of equipment status data are shown in Table 2; samples of equipment operation data are shown in Table 3; samples of load data are shown in Table 4; and samples of historical event data are shown in Table 5.
[0093] Furthermore, the historical event data includes historical distribution terminal fault information, and the distribution terminal fault information includes faulty equipment and fault type. In this embodiment, the predicted risk probability is specifically the probability of a fault occurring.
[0094] By incorporating historical fault information into the data set, a data foundation is laid for the establishment of subsequent N module historical data sets; comprehensive historical event data supports data-driven decision-making, helps optimize maintenance plans and equipment configuration, and includes historical distribution terminal fault information, especially fault modules and fault types, so that the system can use this data to identify and learn fault patterns. This historical data provides a rich information foundation for the model, helps improve the accuracy of predictions, and identifies potential faults in a timely manner.
[0095] See also Figure 2 S20 in which the multi-source historical data of N working modules are extracted from the terminal historical database to form N module historical data sets; the N working modules include a terminal device module, a monitoring and metering module, a relay protection module, a communication module, a control module, a power management module, a power quality module, a safety management module, an environmental monitoring module and an auxiliary module; the N module historical data sets are preprocessed respectively to obtain N preprocessed sets;
[0096] Furthermore, the working module is specifically:
[0097] The terminal equipment modules are used to control and distribute electrical energy, and the terminal equipment modules include various types of circuit breakers, switches and distribution transformers;
[0098] The monitoring and metering module is used to monitor the electric energy parameters and measure the electric energy consumption in real time, and the monitoring and metering module includes various sensors, electric energy meters and data acquisition systems;
[0099] The relay protection module is used to protect the equipment and line safety of the power distribution terminal; the relay protection module includes a relay protection device and a protection relay;
[0100] The communication module is used to realize data communication between the power distribution terminal and the upper control center and / or designated equipment; the communication module includes a communication terminal, a router and a network switch;
[0101] The control module is used to implement local and / or remote control functions to enable the power distribution terminal to operate normally; the control module includes a controller and a programmable logic controller;
[0102] The power management module is used to manage the supply and distribution of power; the power management module includes a power management system and an uninterruptible power supply;
[0103] The power quality module is used to monitor and improve power quality, and the power quality module includes a filter and a capacitor bank;
[0104] The security management module is used to ensure the physical security and information security of the power distribution terminal; the security management module includes a video monitoring system, a firewall and a security gateway;
[0105] The environmental monitoring module is used to monitor the environmental parameters of the environment in which the power distribution terminal is located; the environmental monitoring module includes a temperature and humidity sensor and an air quality monitor;
[0106] The auxiliary module is used to provide supplementary and extended functions of the power distribution terminal.
[0107] In this embodiment, first, N working modules are identified, such as terminal equipment modules, monitoring and metering modules, and relay protection modules. Each working module involves different types of data. For example, the terminal equipment module requires data such as equipment status and operating time, and the monitoring and metering module requires data such as voltage and current. The data type and parameters corresponding to each working module are identified, and the parameters include equipment type and fault type; the data required by each working module are extracted from the historical database to form an independent module historical data set. For example, all data related to the terminal equipment are extracted from the historical database to form a historical data set of the terminal equipment module. The module historical data set corresponding to each working module consists of records from multiple data sources, covering all key performance indicators of the working module.
[0108] The division of the above working modules can refine and clarify the functions and responsibilities of each working module of the distribution terminal, and realize efficient collaborative work between modules. Each module is responsible for a specific function, which makes the responsibilities of each module clearer and easier to manage and maintain. Modular design can not only improve the overall reliability of the system, but also quickly locate and replace the module if a module fails without affecting the normal operation of other modules, thereby reducing the impact of the failure on the entire system; it also facilitates the maintenance and upgrade of the system. Maintenance personnel can inspect, repair or upgrade a single module as needed without having to shut down the entire system for maintenance, reducing maintenance costs and time. In addition, modular design makes the system more flexible and scalable, and modules can be added, removed or adjusted according to actual needs, which facilitates the expansion and function adjustment of the system and meets the application requirements under different scenarios and conditions. By splitting the distribution terminal into multiple working modules, it is easier to perform fault diagnosis and risk warning of the auxiliary distribution terminal system. The data of each working module is collected and analyzed independently, which reduces confusion and errors in the data processing process and improves the accuracy of data processing. The abnormal module can be quickly located and corresponding measures can be taken to improve the efficiency of fault handling. Modular design also helps to optimize resource allocation, reasonably allocate resources according to the needs of different modules, improve resource utilization and avoid resource waste.
[0109] As an implementation mode of the present invention, N module historical data sets are preprocessed respectively, including data cleaning, data standardization, feature extraction and feature selection, to obtain N preprocessed sets;
[0110] See also Figure 2 In S30, model training is performed respectively according to the N preprocessing sets to obtain N multi-level stacked ensemble learning models; each of the working modules corresponds to one multi-level stacked ensemble learning model;
[0111] Furthermore, according to the N preprocessing sets, model training is performed respectively to obtain the N multi-level stacked ensemble learning models, which are specifically:
[0112] Dividing the preprocessed set of each module into a training set, a validation set and a test set;
[0113] According to the training set of each module, M basic machine learning models are trained respectively, including LightGBM model, XGBoost model, random forest model, SVM model and DNN model; M represents the total number of basic machine learning models; each basic machine learning model uses a different algorithm, and each basic machine learning model is the optimal model obtained by training with the corresponding algorithm;
[0114] For each of the modules, use the M basic machine learning models to predict the test set to generate a first-level prediction result data set; the first-level prediction result data set includes the prediction results of each of the basic machine learning models;
[0115] Using the first-level prediction result data set to train the first-layer stacking model, and output the prediction result;
[0116] Taking the prediction results of the first layer stacking model as input, generating a secondary prediction result data set;
[0117] Using the secondary prediction result dataset to train a second-layer stacking model;
[0118] The first layer stacking model and the second layer stacking model are integrated to obtain the multi-level stacking integrated learning model, such as Figure 3 shown.
[0119] By introducing a multi-level stacked ensemble learning model, the present invention significantly improves the accuracy, robustness and real-time performance of risk prediction for distribution terminals. The advantages of multiple basic machine learning models (such as LightGBM, XGBoost, random forest, SVM and DNN) are combined to reduce the errors and deviations of a single model, thereby improving the accuracy of the overall prediction. The prediction results of each model are further optimized and integrated through a layered stacking model, maximizing the use of information from different models, and significantly improving the accuracy of risk prediction. The multi-level structure makes the model more robust in the face of different types of data and complex application scenarios, reduces overfitting problems, and improves generalization capabilities. The modular design ensures the independence and synergy of each module, facilitates system maintenance and expansion, and provides high-quality data support by independently processing and training the data sets of each module. It realizes rapid processing and risk prediction of real-time data, improves the response speed and flexibility of the system, and ultimately provides strong support for the safe and reliable operation of distribution terminals.
[0120] Furthermore, the multi-level stacked integrated learning model also includes an adaptive weighted fusion layer;
[0121] The adaptive weighting layer comprises:
[0122] Using the validation set to evaluate the performance of each of the basic machine learning models, and calculating the weight of each of the basic machine learning models to obtain M basic weights;
[0123] Dynamically adjust each of the basic weights using an optimization algorithm to obtain M adjusted weights;
[0124] Using the M adjustment weights, weighted combination is performed on the prediction results of each of the basic machine learning models to generate a weighted prediction result;
[0125] The weighted prediction results are used as the first-level prediction result data set.
[0126] By introducing the adaptive weighted fusion layer, the prediction accuracy and robustness of the multi-level stacked ensemble learning model have been further improved. The adaptive weighted fusion layer uses the validation set to evaluate the performance of each basic machine learning model, calculates and dynamically adjusts the weight of each model, and realizes the weighted combination of the prediction results of each model. The application of the optimization algorithm ensures the rationality and efficiency of the weight adjustment, so that the final generated weighted prediction results can maximize the advantages of different models and significantly improve the prediction accuracy and reliability. Using the weighted prediction results as the first-level prediction result data set enables the subsequent stacked models to learn and predict based on better data, further improving the performance of the overall risk prediction model. Ultimately, the system can make risk predictions more accurately and robustly, providing strong support for the safe and stable operation of distribution terminals.
[0127] Furthermore, the first layer stacking model and the second layer stacking model are specifically:
[0128] The first layer of stacked models includes linear regression, random forest and DNN models;
[0129] The second layer of stacked models includes an XGBoost model, an LSTM model, and an MLP model.
[0130] By concretizing the design of the first and second layer stacking models, the present invention significantly improves the accuracy and robustness of risk prediction of distribution terminals. The first layer stacking model includes linear regression, random forest and DNN models, which preliminarily integrate the prediction results of each model by utilizing the simplicity of linear models, the decision tree advantages of random forests and the deep learning capabilities of DNNs. The second layer stacking model includes XGBoost, LSTM and MLP models, which further optimize and improve the prediction effect by combining the efficient gradient boosting of XGBoost, the time series processing capabilities of LSTM and the multi-layer perceptron characteristics of MLP. Through this diversified model combination and multi-level stacking structure, the system can make full use of the characteristics of the input data, improve the accuracy and reliability of the prediction, and enhance the robustness and stability of the model.
[0131] See also Figure 2 In S40, real-time data of N working modules of the power distribution terminal are collected to obtain N first real-time data;
[0132] See also Figure 2 In S50, each of the first real-time data is detected using an isolation forest model to obtain N first abnormal data and N first normal data;
[0133] For each working module, the isolation forest model outputs two parts of results:
[0134] First abnormal data: the set of data points detected as abnormal in this module.
[0135] First normal data: a set of data points detected as normal in this module.
[0136] If all the real-time data of a certain working module is detected as normal data, the first abnormal data set of the module will be empty. This means that no abnormality is found in the current data of the working module; if all the real-time data of a certain working module is detected as abnormal data, the first normal data set of the module will be empty. This means that all data points in the current data of the working module are considered abnormal.
[0137] See also Figure 2 In S60, an abnormal module is obtained according to the N first abnormal data, and an early warning is issued for the abnormal module;
[0138] As an implementation manner of the present invention, if any of the first abnormal data is not empty, an abnormal module is obtained according to N of the first abnormal data, and an early warning is issued for the abnormal module;
[0139] See also Figure 2In S70, the N first normal data are respectively input into the N multi-level stacked ensemble learning models to obtain N module risk probabilities;
[0140] See also Figure 2 S80, calculating the overall risk probability of the power distribution terminal according to the risk probabilities of the N modules;
[0141] Further, according to the risk probabilities of the N modules, calculating the overall risk probability of the power distribution terminal includes:
[0142] A node is defined for each working module to obtain N nodes, each node represents the working state of the corresponding working module; a node is added to represent the overall state of the power distribution terminal, as shown in Table 6; the working state includes normal and fault;
[0143] Table 6. Bayesian network node definition table
[0144] Node Name Node Number Terminal equipment module Node1 Monitoring and metering modules Node2 Relay protection module Node3 Communication Module Node4 Control Module Node5 Power Management Module Node6 Power Quality Module Node7 Security Management Module Node8 Environmental monitoring module Node9 Auxiliary modules Node10 Overall status of distribution terminal Node11
[0145] Defining directed edges between nodes according to the terminal history database and / or expert knowledge to obtain a Bayesian network includes the following steps:
[0146] Acquire N module historical data sets from the terminal historical database; preprocess the module historical data sets, including missing value processing and outlier processing, to obtain module preprocessed historical data sets; perform correlation analysis based on the module preprocessed historical data sets, calculate the correlation coefficients between the modules, and obtain a module correlation coefficient matrix;
[0147] According to the module correlation coefficient matrix, setting an initial structure and / or constraint conditions;
[0148] Using the Hill Climb Search algorithm to perform a structural search on the module preprocessing historical data set, during the search process, performing an optimization search based on the initial structure and / or the constraint conditions to obtain multiple candidate structures; calculating the BIC score of each candidate structure to obtain multiple candidate BIC scores; selecting the candidate structure corresponding to the lowest candidate BIC score as the final output to obtain the Bayesian network;
[0149] Establishing a conditional probability table; the conditional probability table represents the probability distribution of a node under a given condition of its parent node;
[0150] The risk probabilities of the N modules are input into the Bayesian network, and the overall risk probability of the power distribution terminal is calculated by a variable elimination algorithm.
[0151] By defining each working module as a node and constructing a Bayesian network based on historical data and expert knowledge, the dependencies and mutual influences between modules can be fully captured. The Hill Climb Search algorithm is used for structural search and optimization to ensure the rationality and efficiency of the network structure, and the uncertainty in the data is processed through the conditional probability table to provide more accurate prediction results. Finally, by inputting the risk probability of each module into the Bayesian network and performing variable elimination algorithm calculations, the overall risk of the distribution terminal can be accurately evaluated, providing reliable decision support for the safe operation of the system.
[0152] See also Figure 2 In S90, an early warning is issued according to the risk probabilities of the N modules and the risk thresholds of the N modules.
[0153] Furthermore, according to the risk probabilities of the N modules and the risk thresholds of the N modules, the warning is specifically performed as follows:
[0154] The module risk threshold is set for each module; when the module risk probability of the module is greater than the module risk threshold, the module is identified as a risk module; and an early warning is issued for the risk module.
[0155] Table 7. Risk warning monitoring table
[0156]
[0157] Table 7 is used to record and manage the risk probability, risk threshold and corresponding warning information of each working module, helping to timely discover and deal with potential risks and ensure the safe and stable operation of the distribution system. The warning level is set based on the difference between the risk probability and the risk threshold, and the risk is divided into multiple levels (such as high, medium and low).
[0158] By setting independent risk thresholds for each module, refined risk management can be performed on each module. This method can flexibly adjust the thresholds according to the characteristics and importance of different modules to ensure more accurate risk identification for each module. By real-time monitoring of the risk probability of the module, when the risk probability exceeds the set threshold, the system can immediately identify and issue an early warning. This early warning mode can achieve early warning, timely detect potential risks, prevent problems in a single module from spreading to the entire system, and enhance the overall stability and reliability of the distribution terminal.
[0159] Embodiment 2:
[0160] Industrial Park B includes multiple manufacturing workshops and research centers, and has extremely high requirements for the continuity and safety of power supply. In order to ensure the smooth progress of the manufacturing process and the normal development of research activities, the power distribution system of the industrial park needs to have real-time monitoring and risk prediction capabilities. Industrial Park B adopts an advanced modular risk prediction method and device for distribution terminals to ensure that there are no interruptions or failures in power supply, thereby ensuring the efficient and safe operation of the entire park.
[0161] A modular risk prediction device for a power distribution terminal, such as Figure 4 As shown, including:
[0162] A multi-source historical data collection unit is used to collect multi-source historical data of a specified power distribution terminal, including environmental data, equipment status data, equipment operation data, load data and historical event data, to obtain a terminal historical database;
[0163] A module data extraction unit, used to extract the multi-source historical data of N working modules from the terminal history database to form N module history data sets; the N working modules include a terminal device module, a monitoring and metering module, a relay protection module, a communication module, a control module, a power management module, a power quality module, a safety management module, an environmental monitoring module and an auxiliary module;
[0164] A data preprocessing unit, used for preprocessing the N module historical data sets respectively to obtain N preprocessed sets;
[0165] The model building unit is used to perform model training according to the N preprocessing sets to obtain N multi-level stacked integrated learning models; each working module corresponds to one multi-level stacked integrated learning model.
[0166] A real-time data collection unit, used to collect real-time data of the N working modules of the power distribution terminal to obtain N first real-time data;
[0167] an anomaly detection unit, used to detect each of the first real-time data using an isolation forest model to obtain N first anomaly data and N first normal data;
[0168] A module early warning unit, configured to obtain an abnormal module according to the N first abnormal data, and issue an early warning for the abnormal module;
[0169] A risk probability calculation unit, used for inputting the N first normal data into the N multi-level stacked integrated learning models respectively to obtain N module risk probabilities;
[0170] An overall risk assessment unit, used to calculate the overall risk probability of the power distribution terminal according to the risk probabilities of the N modules;
[0171] The comprehensive early warning unit is used to issue an early warning based on the risk probabilities of the N modules and the risk thresholds of the N modules.
[0172] In order to verify the effectiveness of the method in the present invention, several groups of comparative experiments are designed, and the following methods need to be compared:
[0173] Traditional approach: Use a single model to predict risk for the entire distribution terminal.
[0174] Holistic prediction (multi-level stacked ensemble): Use a single multi-level stacked ensemble learning model to make predictions for the entire distribution terminal.
[0175] Module-by-module prediction (multi-level stacking integration): Use a multi-level stacking integration learning model to make predictions for each working module and calculate the overall risk probability.
[0176] Module-by-module prediction (random forest): Use the random forest model to make predictions for each working module and calculate the overall risk probability.
[0177] Module-by-module prediction (support vector machine): Use the support vector machine model to make predictions for each working module and calculate the overall risk probability.
[0178] To control variables, the same historical data and real-time data of distribution terminals are used; the same missing value processing, outlier processing and standardization processing are performed; the same evaluation indicators (accuracy, recall rate and F1 value) are used to measure model performance; and the model is run under the same computing environment and configuration.
[0179] Table 8. Comparison of experimental results
[0180]
[0181] Through the comparative experimental data in Table 8, the superiority of the modular risk prediction method can be scientifically evaluated. The experimental results show that the modular prediction (multi-level stacking integration) method is significantly superior to the traditional method and the overall prediction method in terms of accuracy, recall rate and F1 value. And under the same modular prediction, the modular prediction effect of the multi-level stacking integration learning model is the best. The modular risk prediction method proposed in the present invention can more accurately predict the risks of distribution terminals and improve the accuracy of risk prediction of distribution terminals.
[0182] Embodiment three:
[0183] Industrial Park C includes multiple manufacturing workshops, data centers and office buildings, and has extremely high requirements for the continuity and safety of power supply. In order to ensure the smooth progress of the manufacturing process, the stable operation of the data center and the normal development of office activities, the power distribution system of the industrial park needs to have real-time monitoring and risk prediction capabilities. To this end, Industrial Park C uses a computer storable medium that stores a modular risk prediction program for distribution terminals to ensure that there are no interruptions or failures in the power supply, thereby ensuring the efficient and safe operation of the entire park.
[0184] A computer storable medium stores a computer program, which implements the method proposed in the present disclosure when executed by a processor.
[0185] The specific implementation of the computer-readable storage medium of the present invention is basically the same as the various embodiments of the modular risk prediction device and method for the above-mentioned distribution terminal, and will not be repeated here.
[0186] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A modular risk prediction method for distribution terminals, characterized in that: include: Collect multi-source historical data of the designated power distribution terminal, including environmental data, equipment status data, equipment operation data, load data and historical event data, to obtain a terminal historical database; Extracting the multi-source historical data of N working modules from the terminal history database to form N module history data sets; the N working modules include a terminal device module, a monitoring and metering module, a relay protection module, a communication module, a control module, a power management module, a power quality module, a safety management module, an environmental monitoring module and an auxiliary module; Preprocessing the N module historical data sets respectively to obtain N preprocessed sets; According to the N preprocessing sets, model training is performed respectively to obtain N multi-level stacked integrated learning models; each of the working modules corresponds to one multi-level stacked integrated learning model; Collecting real-time data of the N working modules of the power distribution terminal to obtain N first real-time data; Using an isolation forest model to detect each of the first real-time data respectively, to obtain N first abnormal data and N first normal data; Obtain an abnormal module according to the N first abnormal data, and issue an early warning for the abnormal module; Inputting N of the first normal data into N of the multi-level stacked ensemble learning models respectively to obtain N module risk probabilities; Calculate the overall risk probability based on the risk probabilities of the N modules; An early warning is issued based on the risk probabilities of the N modules and the risk thresholds of the N modules.
2. A modular risk prediction method for a power distribution terminal according to claim 1, characterized in that: The historical event data includes historical power distribution terminal fault information, and the power distribution terminal fault information includes faulty equipment and fault type.
3. A modular risk prediction method for a power distribution terminal according to claim 1, characterized in that: The working modules are specifically: The terminal equipment modules are used to control and distribute electrical energy, and the terminal equipment modules include various types of circuit breakers, switches and distribution transformers; The monitoring and metering module is used to monitor the electric energy parameters and measure the electric energy consumption in real time, and the monitoring and metering module includes various sensors, electric energy meters and data acquisition systems; The relay protection module is used to protect the equipment and line safety of the power distribution terminal; the relay protection module includes a relay protection device and a protection relay; The communication module is used to realize data communication between the power distribution terminal and the upper control center and / or designated equipment; the communication module includes a communication terminal, a router and a network switch; The control module is used to implement local and / or remote control functions to enable the power distribution terminal to operate normally; the control module includes a controller and a programmable logic controller; The power management module is used to manage the supply and distribution of power; the power management module includes a power management system and an uninterruptible power supply; The power quality module is used to monitor and improve power quality, and the power quality module includes a filter and a capacitor bank; The security management module is used to ensure the physical security and information security of the power distribution terminal; the security management module includes a video monitoring system, a firewall and a security gateway; The environmental monitoring module is used to monitor the environmental parameters of the environment in which the power distribution terminal is located; the environmental monitoring module includes a temperature and humidity sensor and an air quality monitor; The auxiliary module is used to provide supplementary and extended functions of the power distribution terminal.
4. A modular risk prediction method for a power distribution terminal according to claim 1, characterized in that: According to the N preprocessing sets, model training is performed respectively to obtain the N multi-level stacked ensemble learning models, which are specifically: Dividing the preprocessed set of each module into a training set, a validation set and a test set; According to the training set of each module, M basic machine learning models are trained respectively, including LightGBM model, XGBoost model, random forest model, SVM model and DNN model; M represents the total number of basic machine learning models; each basic machine learning model uses a different algorithm, and each basic machine learning model is the optimal model obtained by training with the corresponding algorithm; For each of the modules, use the M basic machine learning models to predict the test set to generate a first-level prediction result data set; the first-level prediction result data set includes the prediction results of each of the basic machine learning models; Using the first-level prediction result data set to train the first-layer stacking model, and output the prediction result; Taking the prediction results of the first layer stacking model as input, generating a secondary prediction result data set; Using the secondary prediction result dataset to train a second layer stacking model; The first layer stacking model and the second layer stacking model are integrated to obtain the multi-level stacking integrated learning model.
5. A modular risk prediction method for distribution terminals according to claim 4, characterized in that: The multi-level stacked integrated learning model also includes an adaptive weighted fusion layer; The adaptive weighting layer comprises: Using the validation set to evaluate the performance of each of the basic machine learning models, and calculating the weight of each of the basic machine learning models to obtain M basic weights; Dynamically adjust each of the basic weights using an optimization algorithm to obtain M adjusted weights; Using the M adjustment weights, weighted combination is performed on the prediction results of each of the basic machine learning models to generate a weighted prediction result; The weighted prediction results are used as the first-level prediction result data set.
6. A modular risk prediction method for a power distribution terminal according to claim 4, characterized in that: The first layer stacking model and the second layer stacking model are specifically: The first layer of stacked models includes linear regression, random forest and DNN models; The second layer of stacked models includes an XGBoost model, an LSTM model, and an MLP model.
7. A modular risk prediction method for a power distribution terminal according to claim 1, characterized in that: Calculating the overall risk probability according to the N module risk probabilities includes: A node is defined for each working module to obtain N nodes, each node represents the working state of the corresponding working module; a node is added to represent the overall state of the power distribution terminal; the working state includes normal and fault; Defining directed edges between nodes according to the terminal history database and / or expert knowledge to obtain a Bayesian network includes the following steps: Acquire N module historical data sets from the terminal historical database; preprocess the module historical data sets, including missing value processing and outlier processing, to obtain module preprocessed historical data sets; perform correlation analysis based on the module preprocessed historical data sets, calculate the correlation coefficients between the modules, and obtain a module correlation coefficient matrix; According to the module correlation coefficient matrix, setting an initial structure and / or constraint conditions; Using the Hill Climb Search algorithm to perform a structural search on the module preprocessing historical data set, during the search process, performing an optimization search based on the initial structure and / or the constraint conditions to obtain multiple candidate structures; calculating the BIC score of each candidate structure to obtain multiple candidate BIC scores; selecting the candidate structure corresponding to the lowest candidate BIC score as the final output to obtain the Bayesian network; Establishing a conditional probability table; the conditional probability table represents the probability distribution of a node under a given condition of its parent node; The risk probabilities of the N modules are input into the Bayesian network, and the overall risk probability of the power distribution terminal is calculated by a variable elimination algorithm.
8. A modular risk prediction method for a power distribution terminal according to claim 1, characterized in that: According to the risk probabilities of the N modules and the risk thresholds of the N modules, the warning is specifically performed as follows: The module risk threshold is set for each module; when the module risk probability of the module is greater than the module risk threshold, the module is identified as a risk module; An early warning is issued for the risk module.
9. A modular risk prediction device for a power distribution terminal, characterized in that: include: A multi-source historical data collection unit is used to collect multi-source historical data of a specified power distribution terminal, including environmental data, equipment status data, equipment operation data, load data and historical event data, to obtain a terminal historical database; A module data extraction unit, used to extract the multi-source historical data of N working modules from the terminal history database to form N module history data sets; the N working modules include a terminal device module, a monitoring and metering module, a relay protection module, a communication module, a control module, a power management module, a power quality module, a safety management module, an environmental monitoring module and an auxiliary module; A data preprocessing unit, used for preprocessing the N module historical data sets respectively to obtain N preprocessed sets; A model building unit, used to perform model training according to the N preprocessing sets, respectively, to obtain N multi-level stacked ensemble learning models; each of the working modules corresponds to one of the multi-level stacked ensemble learning models; A real-time data collection unit, used to collect real-time data of the N working modules of the power distribution terminal to obtain N first real-time data; an anomaly detection unit, used to detect each of the first real-time data using an isolation forest model to obtain N first anomaly data and N first normal data; A module early warning unit, configured to obtain an abnormal module according to the N first abnormal data, and issue an early warning for the abnormal module; A risk probability calculation unit, used for inputting the N first normal data into the N multi-level stacked integrated learning models respectively to obtain N module risk probabilities; An overall risk assessment unit, used to calculate the overall risk probability according to the risk probabilities of the N modules; The comprehensive early warning unit is used to issue an early warning based on the risk probabilities of the N modules and the risk thresholds of the N modules.
10. A computer storable medium, characterized in that: The computer storable medium stores a computer program, and when the computer program is executed by a processor, the steps of the modular risk prediction method for distribution terminals according to any one of claims 1 to 8 are implemented.
Citation Information
Cited By
Electric power engineering intelligent management and control system and method based on big data
CN119886567A
