Power distribution network dynamic protection setting method facing demand resource side

Abstract

The invention discloses a demand resource side-oriented power distribution network dynamic protection setting method, which comprises the following steps of: 1, acquiring multi-source data which comprises demand side resource data, power distribution network operation data and environment and meteorological data, and preprocessing the multi-source data; step 2, constructing a dynamic model of demand side resources based on the preprocessed demand side resource data; 3, key features are extracted from the dynamic model, the power distribution network operation data and the environment and meteorological data, and a multi-dimensional feature vector library adaptive to a deep learning setting model is constructed; 4, constructing a deep learning setting model by taking the multi-dimensional feature vector library as input; 5, outputting a setting scheme by the deep learning setting model, and issuing the setting scheme to a corresponding protection device; 6, after the protection device receives the setting scheme, local setting configuration is automatically updated, and the problem that existing power distribution network protection setting is difficult to adapt to demand side fluctuation and complex faults is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and in particular to a dynamic protection setting method for distribution networks oriented towards the demand and resource side. Background Technology

[0002] With the accelerated construction of new power systems and the process of new urbanization, the power distribution network is characterized by its expanded scale, diversified resources, and more complex operating conditions. On the one hand, distributed energy and energy storage devices are widely connected to the user side, and the demand-side energy supply system has become an important part of the power distribution network. On the other hand, extreme weather disasters occur frequently, leading to a surge in short circuits, grounding faults, and other issues in the power distribution network. Vulnerability has become a core bottleneck restricting the safe and reliable operation of the power system.

[0003] Currently, vulnerability mitigation and protection settings in distribution networks mainly rely on two technical approaches: Traditional supply-side optimization approaches involve increasing grid redundancy through hardware investments such as adding substations, upgrading transmission lines, and expanding power supply equipment to enhance grid resilience. However, this approach has significant limitations: high construction costs, long construction periods, and an inability to dynamically adapt to demand-side load fluctuations and the uncertainty of distributed energy output, making it difficult to cope with rapidly changing grid conditions.

[0004] Data-driven intelligent optimization path: With the development of big data, deep learning, and edge computing technologies, the industry has begun to explore protection setting decision systems based on multi-source data. For example, features such as fault current and voltage are extracted through wavelet transform and Fourier transform, and historical fault data is used to train models to adjust protection parameters. However, this path still has key technical shortcomings: insufficient spatiotemporal correlation and fusion of multi-source data, and limited accuracy in abnormal data removal and missing value repair; protection setting models mostly rely on fixed algorithms, making it difficult to cope with topology changes and various types of faults after the integration of distributed power sources, resulting in frequent false protection and missed protection phenomena; at the same time, the operating characteristics of demand-side resources are not fully incorporated into the protection setting framework, resulting in limited load recovery capabilities of the distribution network under extreme disturbances.

[0005] Furthermore, existing technologies generally neglect the value of demand-side resources as backup power sources during grid failures, and fail to form a complete technical chain for data acquisition, model building, and implementation of setting schemes. The synergy between protection setting and the characteristics of demand-side resources and the governance of grid vulnerability is lacking, which cannot meet the actual needs of safe and efficient operation of distribution networks. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies in distribution network vulnerability management and protection settings, which are difficult to adapt to demand-side fluctuations and complex faults in distribution networks, and lack adaptability in practical applications. This invention provides a dynamic protection setting method for distribution networks oriented towards demand-side resources. By collecting and preprocessing multi-source data, constructing a dynamic model of demand-side resources and a multi-dimensional feature vector library adapted to a deep learning setting model, a deep learning setting model is built to output setting schemes and distribute them to protection devices so that they can automatically update their configurations, thus achieving the effect of dynamic protection setting for distribution networks oriented towards demand-side resources.

[0007] The objective of this invention is achieved through the following technical solution: A dynamic protection setting method for distribution networks oriented towards the demand and resource side includes the following steps: Step 1: Collect multi-source data, including demand-side resource data, power distribution network operation data, and environmental and meteorological data, and preprocess the multi-source data. Step 2: Based on the preprocessed demand-side resource data, construct a dynamic model of demand-side resources; Step 3: Extract key features from dynamic models, power distribution network operation data, and environmental and meteorological data to construct a multi-dimensional feature vector library adapted to the deep learning tuning model; Step 4: Using the multidimensional feature vector library as input, construct a deep learning tuning model; Step 5: The deep learning tuning model outputs a tuning scheme and sends it to the corresponding protection device. Step 6: After receiving the setting scheme, the protection device automatically updates the local setting configuration.

[0008] Preferably, the demand-side resource data includes distributed power sources and energy storage devices. The dynamic model of the demand-side resources includes a dynamic output model of distributed power sources and a dynamic charge / discharge constraint model of energy storage devices. The dynamic output model of distributed power sources is a model that dynamically correlates influencing factors with output, and the influencing factors are periodically updated. The dynamic charge / discharge constraint model of energy storage devices is a charge / discharge power constraint model based on the energy storage state of charge, establishing a mapping relationship between the real-time energy storage state of charge and the allowable charge / discharge power range, which serves as a constraint condition for the tuning parameters to ensure that the physical limits of the energy storage devices are not exceeded during the tuning process.

[0009] Preferably, in step 3, the key features of the dynamic model include the distributed power generation output fluctuation coefficient, the energy storage state of charge change rate, and the total demand-side output; the key features of the distribution network include the effective value of line current, the node voltage deviation rate, and the power flow direction; and the environmental and meteorological data include the light intensity change rate, wind speed change rate, temperature change rate, and extreme weather level.

[0010] Preferably, the construction of the multi-dimensional feature vector library adapted to the deep learning tuning model specifically includes: Based on timestamp-geographic location tags, key features of dynamic models, distribution network operation data, and environmental and meteorological data from the same power grid node at the same time are fused into a single feature vector. The data within the feature vector is then standardized, and feature vectors from multiple time points constitute a multi-dimensional feature vector library.

[0011] Preferably, in step 4, the deep learning tuning model includes: The overcurrent protection threshold is dynamically predicted and outputs a dynamic overcurrent threshold to provide trigger boundaries for subsequent protection actions. The zero-sequence protection time limit is dynamically corrected, and the zero-sequence protection time limit is output to adapt to the fault scenario, which is used to limit the time of subsequent protection actions; The adaptive protection coordination strategy is generated by taking the dynamic overcurrent threshold and zero-sequence protection time limit as inputs and using the protection action timing output by the supervised learning model as the tuning scheme.

[0012] Preferably, the dynamic prediction of the overcurrent protection threshold specifically refers to: Dynamic model and environmental and meteorological data features are extracted from a multidimensional feature vector library. The mapping relationship between features and overcurrent thresholds in historical data is learned through an LSTM network. The probability density function of normal load current is established by combining kernel density estimation to ensure that the threshold maintains a balance between avoiding normal fluctuations and responding quickly to faults. At the same time, an attention mechanism is introduced to increase the weight value of fault features and realize the dynamic adjustment of the threshold.

[0013] Preferably, the dynamic correction of the zero-sequence protection time limit is specifically as follows: The distribution network topology is represented as an adjacency matrix. Zero-sequence current, fault distance, and load capacity features are extracted. The basic time limit is calculated using a fuzzy logic rule base, and then a correction formula is used for dynamic correction. This is to avoid time limit coordination failure caused by changes in zero-sequence current distribution due to the access of distributed power sources.

[0014] Preferably, the adaptive protection coordination strategy is generated as follows: The supervised learning model is trained using historical fault data to automatically adjust the timing of overcurrent protection, zero-sequence protection, or differential protection.

[0015] As a preferred approach, the dynamic protection setting method for distribution networks oriented towards the demand and resource side further optimizes the deep learning setting model, specifically as follows: Collect correlation data on tuning parameters, demand-side resource status, and execution results to form feedback samples; Analyze feedback samples to correct key characteristic parameters on the demand side and reduce the error between actual output or power and forecast; Add fuzzy rules to optimize the correction accuracy of overcurrent threshold and zero-sequence protection time limit.

[0016] Preferably, the process of optimizing the deep learning tuning model is carried out continuously at a set period.

[0017] The beneficial effects of this invention are: This invention effectively solves the problem that existing distribution network protection settings are difficult to adapt to demand-side fluctuations and complex faults. By collecting and preprocessing multi-source data, constructing a demand-side dynamic model and a multi-dimensional feature vector library, building a deep learning setting model, and automatically issuing updated setting schemes, it realizes dynamic protection settings for distribution networks oriented towards demand-side resources, improves the adaptability of protection actions to demand-side changes and complex power grid operating conditions, and meets the real-time requirements of fault response.

[0018] It improves the accuracy of demand-side resource modeling and feature extraction, builds a dynamic model based on preprocessed high-quality multi-source data, accurately depicts the output law of distributed power sources and the charging and discharging constraints of energy storage, and provides reliable input for deep learning tuning models by combining a spatiotemporal fusion multi-dimensional feature library, reducing false protection and missed protection caused by modeling deviation and invalid features.

[0019] It simplifies the protection setting implementation process, completing the entire process from model output setting scheme to automatic update configuration of protection devices without manual intervention. This meets the millisecond-level response requirements of distribution network faults, reduces manual operation and maintenance costs, and provides complete technical support for subsequent protection setting model optimization and distribution network vulnerability reduction. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0024] Example: Dynamic protection setting methods for distribution networks oriented towards the demand and resource sides, such as Figure 1 As shown, it includes the following steps: Step 1: Collect multi-source data, including demand-side resource data, power distribution network operation data, and environmental and meteorological data, and preprocess the multi-source data. Step 2: Based on the preprocessed demand-side resource data, construct a dynamic model of demand-side resources; Step 3: Extract key features from dynamic models, power distribution network operation data, and environmental and meteorological data to construct a multi-dimensional feature vector library adapted to the deep learning tuning model; Step 4: Using the multidimensional feature vector library as input, construct a deep learning tuning model; Step 5: The deep learning tuning model outputs a tuning scheme and sends it to the corresponding protection device. Step 6: After receiving the setting scheme, the protection device automatically updates the local setting configuration.

[0025] The demand-side resource data includes distributed power sources and energy storage devices. The dynamic model of the demand-side resources includes a dynamic output model of distributed power sources and a dynamic charge / discharge constraint model of energy storage devices. The dynamic output model of distributed power sources is a model that dynamically correlates influencing factors with output, and the influencing factors are updated periodically. The dynamic charge / discharge constraint model of energy storage devices is a charge / discharge power constraint model based on the energy storage state of charge, establishing a mapping relationship between the real-time energy storage state of charge and the allowable charge / discharge power range, which serves as a constraint condition for the tuning parameters to ensure that the physical limits of the energy storage devices are not exceeded during the tuning process.

[0026] Distributed energy sources include photovoltaic (PV), wind power, and clean fossil fuel power generation. Taking PV power generation as an example, a three-dimensional model of "solar intensity - temperature - output" is established, with the following formula: in, To contribute practically, The output is under standard test conditions, and G is the actual light intensity. Where α is the standard illumination intensity, α is the power temperature coefficient, and T is the actual temperature. The standard temperature is used; model parameters are dynamically updated using real-time meteorological data to improve the accuracy of power output prediction.

[0027] The dynamic charge / discharge constraint model for energy storage devices considers charge / discharge efficiency and SOC limitations, and establishes charge / discharge power constraints: During charging: During discharge: Among them, P ch / P dis For charging / discharging power, P ch,max / P dis,max The maximum charge / discharge power is given by SOC, which is the current state of charge. max / SOC min The upper and lower limits of SOC are used; the model can dynamically reflect the available power of energy storage under different SOCs, providing equipment constraint basis for protection settings.

[0028] The SOC data of the energy storage is collected every second by the edge nodes, and the SOC change rate is calculated by combining the charging and discharging power. , or ch / or dis For charge / discharge efficiency, E bat (To determine the rated energy storage capacity (kWh)), adjust the SOC parameters in the model.

[0029] In step 3, the key features of the dynamic model include the distributed power generation output fluctuation coefficient, the energy storage state of charge change rate, and the total demand-side output. The key features of the distribution network include the effective value of line current, the node voltage deviation rate, and the power flow direction. The environmental and meteorological data include the rate of change of light intensity, the rate of change of wind force, the rate of change of temperature, and the extreme weather level. A fault feature library is also constructed, and based on historical fault data, fault current mutation rate, zero-sequence current amplitude, and fault duration are extracted.

[0030] The construction of the multi-dimensional feature vector library adapted to the deep learning tuning model specifically involves: Based on timestamp-geographic location tags, key features of dynamic models, distribution network operation data, and environmental and meteorological data from the same power grid node at the same time are fused into a single feature vector. The data within this feature vector is then standardized, and feature vectors from multiple time points constitute a multi-dimensional feature vector library. For example, a feature vector can be represented as: photovoltaic fluctuation coefficient 0.2 + current 120A + voltage deviation -2% + irradiance variation rate 0.1W / (㎡・min); the standardization process uses Z-score standardization. , x For the original values ​​of a certain feature, m The characteristic mean, sThe standard deviation of the features is used. The Pearson correlation coefficient method is used to calculate the correlation between each feature and the accuracy of the protection action, and redundant features with correlation coefficients <0.3 are eliminated to reduce the computational load of the model.

[0031] In step 4, the deep learning tuning model includes: The overcurrent protection threshold is dynamically predicted and outputs a dynamic overcurrent threshold to provide trigger boundaries for subsequent protection actions. The zero-sequence protection time limit is dynamically corrected, and the zero-sequence protection time limit is output to adapt to the fault scenario, which is used to limit the time of subsequent protection actions; The adaptive protection coordination strategy is generated by taking the dynamic overcurrent threshold and zero-sequence protection time limit as inputs and using the protection action timing output by the supervised learning model as the tuning scheme.

[0032] The aforementioned dynamic prediction of overcurrent protection threshold specifically refers to: Dynamic model and environmental and meteorological data features are extracted from a multidimensional feature vector library. The mapping relationship between features and overcurrent thresholds in historical data is learned through an LSTM network. The probability density function of normal load current is established by combining kernel density estimation to ensure that the threshold maintains a balance between avoiding normal fluctuations and responding quickly to faults. At the same time, an attention mechanism is introduced to increase the weight value of fault features and realize the dynamic adjustment of the threshold.

[0033] The LSTM network specifically includes: Input layer: receiving a standardized 12-dimensional feature vector, including demand-side output, grid current, and environmental features; Hidden layer: consisting of 3 layers of LSTM units, each with 64 neurons, which learn long-term features through a gating mechanism, such as the impact of intraday photovoltaic output peak-valley variations on current; Attention layer: assigning higher weights to key features such as fault current mutation rate and total demand-side output, strengthening features that have a significant impact on setting; Output layer: outputting the predicted overcurrent protection threshold value.

[0034] The aforementioned dynamic correction of the zero-sequence protection time limit specifically refers to: The distribution network topology is represented as an adjacency matrix. Zero-sequence current, fault distance, and load capacity features are extracted. The basic time limit is calculated using a fuzzy logic rule base, and then a correction formula is used for dynamic correction. This is to avoid time limit coordination failure caused by changes in zero-sequence current distribution due to the access of distributed power sources.

[0035] Specifically, based on the distribution network topology, the distance d between the fault point and the protection device is calculated, combined with the load capacity P. load (Based on demand-side modeling results), the basic time limit is obtained through fuzzy inference. t base Using the corrected formula Dynamic correction is performed, among which... k 1 is the distance correction factor. k2 represents the load correction factor, and the final output is the zero-sequence protection time limit. t 0.

[0036] The aforementioned adaptive protection coordination strategy generation is specifically as follows: The supervised learning model is trained using historical fault data to automatically adjust the timing of overcurrent protection, zero-sequence protection, or differential protection. In this embodiment, the supervised learning model can be trained using a random forest algorithm to learn the mapping relationship between demand-side resource status and protection action timing.

[0037] The real-time strategy output is as follows: extract the current demand-side resource status and grid fault type features; input the supervised learning model and output the protection action timing, such as overcurrent protection action delay of 0.2s and zero-sequence protection action delay of 0.5s; if the demand-side resource output drops sharply, such as photovoltaic output dropping by 50% within 10 minutes, then shorten the overcurrent protection action delay, such as from 0.2s to 0.1s, and prioritize the clearing of faults.

[0038] The false protection rate after the statistical strategy is implemented is calculated. If the false protection rate is greater than 0.1%, erroneous samples are added to retrain the model until the false protection rate is less than or equal to 0.1%.

[0039] The demand-resource-side-oriented dynamic protection setting method for distribution networks also optimizes the deep learning setting model, specifically: Collect correlation data on tuning parameters, demand-side resource status, and execution results to form feedback samples; Analyze feedback samples to correct key characteristic parameters on the demand side and reduce the error between actual output or power and forecast; Add fuzzy rules to optimize the correction accuracy of overcurrent threshold and zero-sequence protection time limit.

[0040] Specifically, key indicators are extracted from the execution status monitoring module every hour, such as fault response time = 0.6s > target 0.5s, false protection rate = 0.2% > target 0.1%; and related data of setting parameters, demand-side resource status, and execution effect are collected to form feedback samples.

[0041] If the photovoltaic power output prediction error is 10% > the target of 8%, then the power temperature coefficient will be corrected based on the feedback samples. α For example, if the energy storage power calculation error is 4% > the target of 3%, then the charge and discharge efficiency should be corrected. or ch / or dis If the overcurrent protection response time exceeds the limit, add feedback samples to retrain the LSTM model and adjust the attention layer weights, such as increasing the weight of the fault current mutation rate; if the zero-sequence time limit is mismatched, supplement the fuzzy rule base, such as adding a rule to shorten the time limit by 10% under extreme weather conditions.

[0042] The process of optimizing the deep learning tuning model is carried out continuously at a set period.

[0043] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0044] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A dynamic protection setting method for distribution networks oriented towards the demand and resource side, characterized by: Includes the following steps: Step 1: Collect multi-source data, including demand-side resource data, power distribution network operation data, and environmental and meteorological data, and preprocess the multi-source data. Step 2: Based on the preprocessed demand-side resource data, construct a dynamic model of demand-side resources; Step 3: Extract key features from dynamic models, power distribution network operation data, and environmental and meteorological data to construct a multi-dimensional feature vector library adapted to the deep learning tuning model; Step 4: Using the multidimensional feature vector library as input, construct a deep learning tuning model; Step 5: The deep learning tuning model outputs a tuning scheme and sends it to the corresponding protection device. Step 6: After receiving the setting scheme, the protection device automatically updates the local setting configuration.

2. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 1, characterized in that, The demand-side resource data includes distributed power sources and energy storage devices. The dynamic model of the demand-side resources includes a dynamic output model of distributed power sources and a dynamic charge / discharge constraint model of energy storage devices. The dynamic output model of distributed power sources is a model that dynamically correlates influencing factors with output, and the influencing factors are updated periodically. The dynamic charge / discharge constraint model of energy storage devices is a charge / discharge power constraint model based on the energy storage state of charge, establishing a mapping relationship between the real-time energy storage state of charge and the allowable charge / discharge power range, which serves as a constraint condition for the tuning parameters to ensure that the physical limits of the energy storage devices are not exceeded during the tuning process.

3. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 2, characterized in that, In step 3, the key features of the dynamic model include the output fluctuation coefficient of distributed power sources, the rate of change of energy storage state of charge, and the total output of the demand side. The key features of the distribution network include the effective value of line current, the node voltage deviation rate, and the power flow direction. The environmental and meteorological data include the rate of change of light intensity, the rate of change of wind force, the rate of change of temperature, and the extreme weather level.

4. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 3, characterized in that, The construction of the multi-dimensional feature vector library adapted to the deep learning tuning model specifically involves: Based on timestamp-geographic location tags, key features of dynamic models, distribution network operation data, and environmental and meteorological data from the same power grid node at the same time are fused into a single feature vector. The data within the feature vector is then standardized, and feature vectors from multiple time points constitute a multi-dimensional feature vector library.

5. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 1, characterized in that, In step 4, the deep learning tuning model includes: The overcurrent protection threshold is dynamically predicted and outputs a dynamic overcurrent threshold to provide trigger boundaries for subsequent protection actions. The zero-sequence protection time limit is dynamically corrected, and the zero-sequence protection time limit is output to adapt to the fault scenario, which is used to limit the time of subsequent protection actions; The adaptive protection coordination strategy is generated by taking the dynamic overcurrent threshold and zero-sequence protection time limit as inputs and using the protection action timing output by the supervised learning model as the tuning scheme.

6. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 5, characterized in that, The aforementioned dynamic prediction of overcurrent protection threshold specifically refers to: Dynamic model and environmental and meteorological data features are extracted from a multidimensional feature vector library. The mapping relationship between features and overcurrent thresholds in historical data is learned through an LSTM network. The probability density function of normal load current is established by combining kernel density estimation to ensure that the threshold maintains a balance between avoiding normal fluctuations and responding quickly to faults. At the same time, an attention mechanism is introduced to increase the weight value of fault features and realize the dynamic adjustment of the threshold.

7. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 6, characterized in that, The aforementioned dynamic correction of the zero-sequence protection time limit specifically refers to: The distribution network topology is represented as an adjacency matrix. Zero-sequence current, fault distance, and load capacity features are extracted. The basic time limit is calculated using a fuzzy logic rule base, and then a correction formula is used for dynamic correction. This is to avoid time limit coordination failure caused by changes in zero-sequence current distribution due to the access of distributed power sources.

8. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 7, characterized in that, The aforementioned adaptive protection coordination strategy generation is specifically as follows: The supervised learning model is trained using historical fault data to automatically adjust the timing of overcurrent protection, zero-sequence protection, or differential protection.

9. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 1, characterized in that, Furthermore, the deep learning tuning model was optimized, specifically as follows: Collect correlation data on tuning parameters, demand-side resource status, and execution results to form feedback samples; Analyze feedback samples to correct key characteristic parameters on the demand side and reduce the error between actual output or power and forecast; Add fuzzy rules to optimize the correction accuracy of overcurrent threshold and zero-sequence protection time limit.

10. The dynamic protection setting method for distribution networks oriented towards the demand and resource side as described in claim 9, characterized in that, The process of optimizing the deep learning tuning model is carried out continuously at a set period.

Images