Fuse wire configuration auxiliary decision-making system and method based on big data analysis
The fuse configuration auxiliary decision-making system, which combines big data analysis with LSTM time series model and random forest classification model, solves the problem of fuse configuration relying on human experience, realizes accurate and dynamic fuse configuration decisions, and improves configuration accuracy and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-31
AI Technical Summary
The current configuration of fuses mainly relies on manual experience and lacks scientific data support, which may lead to improper configuration, unplanned power outages, or an expansion of the scope of accidents. Existing data analysis methods are complex and have limited effectiveness.
A fuse configuration auxiliary decision-making system based on big data analysis is adopted. Combining LSTM time series model and random forest classification model, the model is dynamically updated by analyzing the operation data and fault data of fuses to provide accurate fuse configuration suggestions.
It enables simplified and reliable fuse configuration decisions, improves configuration accuracy, lowers the threshold for data acquisition and model training, and can continuously optimize decisions based on changes in power grid conditions.
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power distribution network operation and maintenance technology, and relates to a fuse configuration auxiliary decision-making system and method based on big data analysis. Background Technology
[0002] Drop-out fuses are crucial protective devices in power distribution networks. Currently, their selection and configuration rely heavily on manual experience, lacking scientific data support. This approach easily leads to two problems: first, under-configuration may cause false tripping under normal load, resulting in unplanned power outages; second, over-configuration may cause failure to trip during faults, expanding the scope of the accident. While some data analysis-based methods have emerged in existing technologies, they often involve complex models, high implementation costs, and fail to effectively combine time-series load characteristics with fault characteristics, resulting in limited effectiveness in practical applications. Summary of the Invention
[0003] This invention proposes a fuse configuration auxiliary decision-making system and method based on big data analysis. The system is simple to implement and reliable.
[0004] The technical solution of this invention is implemented as follows:
[0005] Technical Topic 1
[0006] A fuse configuration auxiliary decision-making system based on big data analysis includes:
[0007] The multi-source data acquisition module is used to collect the operating data and fault data of the fuse;
[0008] The distributed data storage module adopts a combined architecture of HDFS distributed file system and InfluxDB time-series database; InfluxDB is used to store recent real-time running data; HDFS is used for partitioned storage of historical running data and fault data.
[0009] The hybrid algorithm data processing module includes a data preprocessing unit, an LSTM time series model, and a random forest classification model. The data preprocessing unit is used to perform outlier removal and standardization on the running data and fault data. The LSTM time series model is used to output a suggested value for the rated current of the fuse based on the preprocessed load current time series data. The random forest classification model is used to output a suggested fuse material type based on the preprocessed fault data.
[0010] The dynamic update module is used to periodically or trigger a full update of the model based on an event that exceeds the limit of the number of failures; the update process includes calling new historical data to retrain the LSTM time series model and the random forest classification model, and replacing the parameters of the old model;
[0011] The decision output module is used to receive and directly output the rated current recommendation value generated by the LSTM timing model and the material type recommendation generated by the random forest classification model, which together serve as the final fuse configuration scheme.
[0012] Preferably, the operating data includes load current collected at fixed time intervals.
[0013] Preferably, the fault data includes peak short-circuit current, annual number of fuse failures, and fault type.
[0014] Preferably, the data preprocessing unit performs outlier removal using the 3σ principle and performs Z-score standardization on numerical data.
[0015] Preferably, the LSTM time series model is trained using the Adam optimizer with mean squared error (MSE) as the loss function; the criterion for determining whether the model training is complete is that the average deviation between the predicted current value and the actual adapted current value on the test set does not exceed 8%.
[0016] Preferably, the random forest classification model is trained using Gini coefficients as the node splitting criterion; the criterion for determining whether the model training is complete is that the prediction accuracy of the filament material type on the test set is not less than 90%.
[0017] Preferably, the preset cycle of the dynamic update module is once a week, and the event that triggers the update is when the number of failures of the same fuse reaches or exceeds 2 times in a single week; the update process includes calling historical data from the most recent month to retrain the model and replacing the old model parameters.
[0018] Technical Theme Two
[0019] The present invention also provides a method for assisting decision-making on fuse configuration in the system described in Technical Subject 1, comprising the following steps:
[0020] S1: System initialization: Retrieve complete historical data from the past six months from the distributed data storage module, perform the first model training, and generate the initial LSTM time series model and random forest classification model;
[0021] S2: Data Acquisition and Storage: The multi-source data acquisition module continuously acquires the operating data and fault data of the fuse and stores them in the distributed data storage module;
[0022] S3: Data Preprocessing: Call the data preprocessing unit to remove outliers and standardize the stored data, and divide it into training and test sets;
[0023] S4: Hybrid model training: Using the training set data, train the LSTM time series model and the random forest classification model with fixed parameters respectively, and validate them on the test set. Save the model after it meets the target.
[0024] S5: Configuration scheme generation: For the target circuit breaker, its preprocessed data is input into the trained model, and the decision output module directly outputs the model calculation result as the configuration scheme.
[0025] S6: Dynamic Model Update: Repeat steps S3-S4 according to a preset period or trigger event to complete the model update iteration.
[0026] The beneficial effects of the present invention using the above technical solution are as follows:
[0027] 1. The fuse configuration auxiliary decision-making system and method based on big data analysis provided by this invention provides accurate fuse configuration suggestions to operation and maintenance personnel by constructing a hybrid algorithm module composed of LSTM time series model and random forest classification model and analyzing historical operation data and fault data.
[0028] 2. In this invention, the LSTM model effectively captures the temporal patterns of the load, and the random forest accurately associates fault characteristics with fuse materials. The combination of the two improves the overall configuration accuracy. The solution of this invention is simplified, using a combination of mature basic models, reducing the required data dimensions and lowering the threshold for data acquisition and model training. Through the dynamic update module, this invention enables the system to continuously optimize the model according to changes in the power grid operating status, maintaining the advancement of decision-making. Detailed Implementation
[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0030] This invention proposes a fuse configuration auxiliary decision-making system based on big data analysis, comprising:
[0031] The multi-source data acquisition module is used to collect the operating data and fault data of the fuse;
[0032] The distributed data storage module adopts a combined architecture of HDFS distributed file system and InfluxDB time-series database; InfluxDB is used to store recent real-time running data; HDFS is used for partitioned storage of historical running data and fault data.
[0033] The hybrid algorithm data processing module includes a data preprocessing unit, an LSTM time series model, and a random forest classification model. The data preprocessing unit is used to perform outlier removal and standardization on the running data and fault data. The LSTM time series model is used to output a suggested value for the rated current of the fuse based on the preprocessed load current time series data. The random forest classification model is used to output a suggested fuse material type based on the preprocessed fault data.
[0034] The dynamic update module is used to periodically or trigger a full update of the model based on an event that exceeds the limit of the number of failures; the update process includes calling new historical data to retrain the LSTM time series model and the random forest classification model, and replacing the parameters of the old model;
[0035] The decision output module is used to receive and directly output the rated current recommendation value generated by the LSTM timing model and the material type recommendation generated by the random forest classification model, which together serve as the final fuse configuration scheme.
[0036] The network structure of the LSTM time series model is as follows: the input layer receives a load current sequence of fixed length 2016 data points, which is 7 days of data with a sampling interval of 5 minutes; two LSTM hidden layers, each containing 32 units; the output layer is a fully connected neuron; its training parameters are fixed as follows: batch size 32, number of iterations 50, learning rate 0.001, using the Adam optimizer, and using mean squared error (MSE) as the loss function; the criterion for judging the completion of model training is that the average absolute percentage error between the predicted current value and the actual adapted current value on the test set does not exceed 8%.
[0037] The input features of the random forest classification model are peak short-circuit current, annual number of fuse failures, and numerically encoded fault types. The model parameters are fixed as follows: 50 decision trees, a maximum depth of 8 layers per tree, and Gini coefficients as the node splitting criteria. The criterion for judging the completion of model training is that the prediction accuracy of fuse material type on the test set is not less than 90%.
[0038] The fuse configuration auxiliary decision-making method based on this system includes the following steps:
[0039] S1: System initialization: Retrieve complete historical data from the past six months from the distributed data storage module, perform the first model training, and generate the initial LSTM time series model and random forest classification model;
[0040] S2: Data Acquisition and Storage: The multi-source data acquisition module continuously acquires the operating data and fault data of the fuse and stores them in the distributed data storage module;
[0041] S3: Data Preprocessing: Call the data preprocessing unit to remove outliers and standardize the stored data, and divide it into training and test sets;
[0042] S4: Hybrid model training: Using the training set data, train the LSTM time series model and the random forest classification model with fixed parameters respectively, and validate them on the test set. Save the model after it meets the target.
[0043] S5: Configuration scheme generation: For the target circuit breaker, its preprocessed data is input into the trained model, and the decision output module directly outputs the model calculation result as the configuration scheme.
[0044] S6: Dynamic Model Update: Repeat steps S3-S4 according to a preset period or trigger event to complete the model update iteration.
[0045] The details are as follows:
[0046] I. System Architecture
[0047] The implementation of this system relies on the collaborative work of intelligent monitoring terminals (such as current transformers and fault indicators) deployed at the distribution network site and the back-end data processing center. The specific data flow and module connection relationships are as follows:
[0048] Data upload: The intelligent monitoring terminal collects the load current (unit: A) of the line where the fuse is located every 5 minutes, and records the peak short-circuit current (unit: kA) when a fault occurs. This data is transmitted to the back-end data access layer via a wireless network.
[0049] Data storage and distribution: The data access layer writes the parsed data to the InfluxDB time-series database (retaining data from the most recent 3 months) and archives the data daily to the HDFS distributed file system (stored in a "year / month" directory structure). The hybrid algorithm data processing module queries real-time data from InfluxDB and loads historical data from HDFS via Java / Python clients.
[0050] Core processing: After receiving data processing or training instructions, the hybrid algorithm data processing module first calls the data preprocessing unit, and then feeds the processed data in parallel into the LSTM time series model (implemented based on the TensorFlow / Keras framework) and the random forest classification model (implemented based on the Scikit-learn framework) for training or prediction.
[0051] Decision and Feedback: The outputs of the two models are directly encapsulated into a JSON-formatted configuration scheme by the decision output module and provided to the front-end operation and maintenance management system via a RESTful API. Meanwhile, the dynamic update module, as an independent scheduled task (e.g., using Linux Cron or Apache Airflow scheduling), periodically checks update conditions and triggers model retraining.
[0052] II. Construction of Hybrid Algorithm Data Processing Module
[0053] 1. Data Preparation
[0054] Data sources: Loading at least one year of historical data for a specific fuse from HDFS. The LSTM model uses its load current time series. The Random Forest model uses all its fault event records.
[0055] Objective definition of the target variable:
[0056] LSTM target value (objective calculation): Target rated current value = MAX (historical load current sequence) × safety factor K. Wherein, K is an empirical value set according to industry standards, usually taken as 1.3 to 1.5.
[0057] Random Forest Target Value: Establishes a mapping relationship between fuse material type and maximum short-circuit current peak value. For example, define the rule: if the short-circuit current peak value is ≤8kA, then the target material is pure copper (code 0); if the short-circuit current peak value is >8kA, then the target material is copper-silver alloy (code 1). This mapping relationship is based on the breaking capacity specifications in the national standards for fuse materials.
[0058] 2. Training of LSTM Temporal Model
[0059] Input data construction: Taking the load current sequence of a fuse as an example, a sliding window of length 2016 (7 days × 288 points / day) is extracted sequentially as a training sample. The target value corresponding to this sample is the target rated current value calculated according to the above formula.
[0060] Training and Validation: After training, calculate the mean absolute percentage error (MAPE) of the predicted values on the test set. If the MAPE is ≤ 8%, the model is considered to have passed training, and the model file (.h5 format) is saved for later use.
[0061] 3. Training the Random Forest Classification Model
[0062] Input feature construction: Each fault event is treated as a sample, and its feature vector is [peak short-circuit current, number of fuse blowouts per year, fault type (0 / 1)].
[0063] Training and Validation: After training, calculate the classification accuracy on the test set. If the accuracy is ≥90%, the model is considered successfully trained, and the model file is saved.
[0064] III. Decision Output and Dynamic Update Logic
[0065] Decision output: Configuration scheme = LSTM output value (unit: A) + Random Forest output value (material type). For example, the output is "68A, copper-silver alloy".
[0066] Dynamically updated:
[0067] Initialization: When the system is deployed for the first time, it will automatically perform a full training, with the data range being at least 6 months of historical data available in HDFS.
[0068] Periodic Update: Thereafter, every Monday at 2:00 AM, the dynamic update module will be automatically triggered, calling the new data from the past month, repeating the training process in Part 2, generating a new model and hot-replacing the old model.
[0069] Event-triggered update: If any fuse records ≥2 failures within a week, the system will immediately trigger a model update for that fuse.
[0070] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A big data analytics based fuse configuration aided decision system for fuses, characterized in that, The application relates to a fuse configuration system based on hybrid algorithm. The application comprises: a multi-source data acquisition module for acquiring operation data and fault data of a fuse; a distributed data storage module adopting a combined architecture of an HDFS distributed file system and an InfluxDB time series database; the InfluxDB is used for storing recent real-time operation data; and the HDFS is used for partitioned storage of historical operation data and fault data; a hybrid algorithm data processing module comprising a data preprocessing unit, an LSTM time series model and a random forest classification model; the data preprocessing unit is used for performing outlier elimination and standardization processing on the operation data and the fault data; the LSTM time series model is used for outputting a fuse rated current suggestion value according to preprocessed load current time series data; the random forest classification model is used for outputting a fuse material type suggestion according to preprocessed fault data; a dynamic updating module is used for triggering full updating of the model on a regular basis or based on a fault frequency exceeding event; the updating process comprises calling new historical data to retrain the LSTM time series model and the random forest classification model, and replacing old model parameters; 2. A big data analytics based auxiliary decision system for fuse configuration of a fuse link as claimed in claim 1, wherein, a decision output module is used for receiving and directly outputting the rated current suggestion value generated by the LSTM time series model and the material type suggestion generated by the random forest classification model, which are used as a final fuse configuration scheme.
3. A big data analytics based auxiliary decision system for fuse configuration of a fuse link as claimed in claim 1, wherein, The operation data comprises load current acquired at fixed time intervals.
4. A big data analytics based fuse configuration aided decision system as claimed in claim 1, wherein, The fault data comprises short-circuit current peak value, annual fuse blowing frequency and fault type.
5. A big data analytics based auxiliary decision system for fuse configuration of a fuse link as claimed in claim 1, wherein, The data preprocessing unit performs outlier elimination by adopting a 3sigma principle, and performs Z-score standardization processing on numerical data.
6. A big data analytics based auxiliary decision system for fuse configuration of a fuse link as claimed in claim 1, wherein, The LSTM time series model is trained by using an Adam optimizer, and mean square error (MSE) is used as a loss function; the determination criterion for completion of model training is that the average deviation of predicted current values and actual adaptive current values on a test set is not more than 8%.
7. A big data analytics based fuse configuration aided decision system as claimed in claim 1, wherein, The random forest classification model is trained by using a Gini coefficient as a node splitting standard; and the determination criterion for completion of model training is that the fuse material type prediction accuracy on a test set is not less than 90%.
8. A method for assisting the decision of a fuse configuration of a fuse link based on the system according to any one of claims 1 to 7, characterized in that, The preset period of the dynamic updating module is once a week, and the event triggering updating is that the fault frequency of the same fuse in a single week reaches or exceeds 2 times; the updating process comprises calling historical data of the last month to retrain the model and replacing old model parameters. The application comprises the following steps: S1: system initialization: calling complete historical data of the past six months from the distributed data storage module, performing first model training, and generating an initial LSTM time series model and a random forest classification model; S2: data acquisition and storage: continuously acquiring operation data and fault data of a fuse through a multi-source data acquisition module, and storing the data in the distributed data storage module; S3: data preprocessing: calling a data preprocessing unit, performing outlier elimination and standardization processing on the stored data, and dividing the data into a training set and a test set; S4: hybrid model training: training an LSTM time series model and a random forest classification model by using training set data and fixed parameters, and verifying on a test set, and saving the model after reaching the standard; and S5: decision output: receiving and directly outputting a rated current suggestion value generated by the LSTM time series model and a material type suggestion generated by the random forest classification model, which are used as a final fuse configuration scheme. S5: Configuration scheme generation: for the target fuse, input the preprocessed data into the trained model, and the decision output module directly outputs the model operation result as the configuration scheme; S6: Model dynamic update: according to the preset period or trigger event, repeat steps S3-S4 to complete the update iteration of the model.