A load switch control method for a distribution box

Abstract

The present invention discloses a load switch control method for a distribution box, which relates to the technical field of automatic control systems and comprises the following steps: obtaining real-time load data of a to-be-tested area, and constructing a dynamic switch threshold control model based on LSTM to generate an initial threshold interval; constructing a load behavior map according to an optimal clustering model, mapping the current load data to the load behavior map, and regulating the initial threshold interval in combination with the fluctuation interval of historical load data to obtain a corrected threshold interval; extracting a distribution feature vector and a classification accuracy parameter of the current load data as first data, and training an optimal classification model based on the first data; inputting the corrected threshold interval into the optimal classification model to identify the current load state label; the present application constructs a load behavior map, evaluates the load state in combination with the fluctuation interval of historical load data, and then adjusts the control strategy to achieve dynamic regulation of the switch threshold interval, and controls the load switch according to the correction result.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic control systems, and more particularly to a load switch control method for a distribution box. Background Art

[0002] With the rapid development of modern smart grids, the safety, stability, and economic efficiency of distribution systems, as the key link connecting power generation and users, are receiving increasing attention. Traditional load switch control methods, particularly in industrial parks, residential communities, and commercial complexes, where power loads exhibit high-frequency fluctuations and complex dynamic characteristics, have struggled to meet the high standards required for energy efficiency optimization and safety management. Currently, most distribution box load switch control strategies still rely on fixed thresholds or simple rules, lacking the ability to perceive and respond to real-time load trends. These control strategies are prone to false triggering, delayed response, and energy waste, severely hindering improvements in distribution system control efficiency and automation. Therefore, to meet the high-precision control requirements of future smart distribution systems, a load switch control method with adaptive learning and dynamic adjustment capabilities is needed to achieve more refined and intelligent load management.

[0003] For example, the invention patent with the announcement number: CN119105293B discloses an adaptive control method and system for intelligent switching equipment, which relates to the field of intelligent switching equipment control technology. The adaptive control method for intelligent switching equipment includes the following steps: load data collection and processing, fluctuation compliance evaluation and judgment, balanced optimization evaluation and judgment, and fault detection evaluation and feedback. The present invention obtains load-related data by pre-processing the initial load data, and then performs load evaluation on the preset industrial park area based on the load-related data to determine whether to perform load balancing optimization, and then evaluates the effect of the load balancing optimization and determines whether to perform equipment fault detection. Finally, the switch equipment fault evaluation index after the equipment fault detection is performed is obtained and a comprehensive evaluation is performed to obtain a comprehensive evaluation result, thereby achieving the effect of improving the adaptive control efficiency of the intelligent switching equipment and solving the problem of low adaptive control efficiency of the intelligent switching equipment in the prior art.

[0004] For example, the invention patent with announcement number: CN118466219B discloses an optimization method, system, device and medium for a flexible power load control device. The method obtains historical power load data of the target power grid to construct a historical power load database; obtains a first local optimization result of the flexible power load control device based on the historical power load database; obtains real-time operation data of the target power grid to update the historical power load database and obtain an optimized power load database; corrects the first local optimization result based on the optimized power load database to obtain a second local optimization result; performs global optimization on the flexible power load control device of the target power grid based on the second local optimization result; and dynamically optimizes the flexible power load control device of the target power grid based on the global optimization result. The method provided in this application realizes local optimization, global optimization and dynamic optimization of the flexible power load control device, improves the control effect and adapts to the needs of intelligent power grid.

[0005] The above disclosed technical solutions have at least the following technical problems:

[0006] Traditional distribution box load control methods are often based on static threshold settings and cannot automatically adjust control strategies based on real-time load fluctuations. This results in a slow response to sudden load anomalies, impacting system stability. Furthermore, traditional load state identification methods fail to adaptively optimize classification features based on different data distribution characteristics, and the lack of a fusion mechanism limits identification accuracy. To address these issues, the present invention proposes a solution. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a load switch control method for a distribution box. By constructing a load behavior map and combining it with the fluctuation range of historical load data, the load state can be evaluated more scientifically, and then the control strategy can be adjusted to achieve dynamic regulation of the switch threshold range, and the load switch can be controlled according to the correction result.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A load switch control method for a distribution box includes the following steps: obtaining real-time load data of a test area, building a dynamic switch threshold control model based on LSTM, and generating an initial threshold interval; building a load behavior map based on an optimal clustering model, mapping current load data to the load behavior map, and regulating the initial threshold interval in combination with historical load data fluctuation intervals to obtain a revised threshold interval; extracting a distribution feature vector and a classification accuracy parameter of the current load data as first data, and training an optimal classification model based on the first data; and inputting the revised threshold interval into the optimal classification model to identify a current load state label.

[0010] In a preferred embodiment, the real-time load data of the area to be tested is obtained, and a dynamic switch threshold control model is constructed based on LSTM to generate an initial threshold range, specifically as follows: each phase line and circuit of the distribution box in the area to be tested is monitored in real time, and the electrical parameter data of each circuit is synchronously collected at a fixed period to form multi-channel parallel load data; the real-time load data is preprocessed by the edge computing node, and each data stream is time-synchronized based on the NTP protocol; the preprocessed load data is organized in chronological order, and the time series is reorganized using the sliding window technology, and the load data of every N consecutive time points constitute a single time series sample, and each sample constitutes a dimensional tensor; the time series sample is input into the LSTM model for training to obtain a trained dynamic switch threshold model, and the load fluctuation trend is predicted based on the LSTM model; the dynamic threshold range is constructed according to the predicted value output by the LSTM model combined with the dynamic switch threshold model.

[0011] In a preferred embodiment, the predicted value output by the LSTM model is combined with a dynamic switching threshold model to construct a dynamic threshold range, specifically as follows: based on the load data in the window, the standard deviation of the load data in the window is output; based on the predicted load trend and the standard deviation of the load data in the sliding window, a dynamic threshold model is constructed to obtain a dynamic threshold range.

[0012] In a preferred embodiment, the optimal clustering model is specifically as follows: preprocessing the load data, and reducing the dimension of the load characteristic data based on the principal component analysis method to obtain sample data; mapping similar sample data to the same bucket through hash coding based on the local sensitive hashing algorithm to generate hash values; clustering the hash values, and obtaining the local neighborhood of each data point based on the cosine similarity method; extracting local features based on the data in the local neighborhood; evaluating the clustering effect of the clustering model based on the local features, and screening out the optimal clustering model.

[0013] In a preferred embodiment, the load behavior map is constructed based on the optimal clustering model, specifically as follows: historical load data of the area to be measured is obtained, and sliding window segmentation is performed, each window contains time series samples of a preset length; the time series samples are constructed into a tensor, and the tensor is input into the clustering model for training, and the load behavior category to which each time window belongs is output; the sample tensor is clustered by the trained clustering model, and according to the clustering results, the time window data belonging to the same category are classified into the same group to obtain several load behavior categories, and each category constitutes a load behavior map.

[0014] In a preferred embodiment, the current load data is mapped to the load behavior map, and the initial threshold range is regulated in combination with the historical load data fluctuation range, specifically as follows: feature extraction is performed on the real-time load data in the sliding window at the current moment to form a tensor with the same dimension as the load behavior map; the similarity between the current feature vector and the central vector of all load behavior maps is obtained based on the cosine similarity method, and the load behavior category map with the highest similarity is selected as the mapping map of the current window; the maximum deviation value within the historical load fluctuation range of the category is extracted from the mapping map, and is used to represent the upper offset correction amount and the lower offset correction amount as correction factors respectively; the initial threshold range is controlled by the correction factor and combined with the standard deviation of the sliding window load data to obtain the corrected threshold range.

[0015] In a preferred embodiment, the distribution feature vector and classification accuracy parameter of the current load data are extracted as the first data, and the optimal classification model is trained based on the first data, specifically as follows: first data of the load data is obtained, the first data includes distribution characteristics and classification accuracy; load data of the area to be measured is obtained, the distribution characteristics of the load data are obtained, and the distribution form of the load data is analyzed by a histogram method; load data characteristics are screened based on the distribution characteristics of the load data, and a load data feature subset is constructed; the load data feature subset is input as a time series sample into the classification model for classification, and the classification accuracy of the classification model is evaluated based on cross-validation, and the best classification model is screened according to the classification accuracy.

[0016] In a preferred embodiment, the load data feature screening based on the distribution characteristics of the load data is specifically as follows: obtaining the distribution form of the load data and extracting characteristic indicators that are correlated with the data distribution form; evaluating the contribution of each characteristic indicator to the load data classification performance based on a feature importance ranking method based on correlation analysis; based on the evaluation results, screening out the characteristic indicators that have the best effect on improving the classification accuracy under the current load data distribution characteristics, and constructing a load data feature subset.

[0017] In a preferred embodiment, the corrected threshold interval is input into the optimal classification model to identify the current load state label, specifically as follows: each data point is clustered by the optimal clustering model to output the initial load state label; the load feature data of the current window is input into the trained classification model to obtain the predicted load state label; the output results of the clustering model and the classification model are subjected to label fusion; the load state label obtained after fusion is used as the final identification result, and the load switch of the target distribution box is controlled.

[0018] The technical effects and advantages of the load switch control method of a distribution box of the present invention are as follows:

[0019] 1. This invention utilizes a long short-term memory (LSTM) network to perform time-series analysis on load data from distribution boxes, thereby constructing a dynamic threshold model. The LSTM model has a strong ability to capture time-series dependencies and effectively learns historical fluctuation patterns in load data, thereby predicting load fluctuation trends. By combining the predicted results with the load standard deviation within a sliding window, the load control threshold is dynamically adjusted, enabling the load switch to automatically adjust based on real-time load conditions. The key advantage of this method is its adaptive ability to address scenarios with high load fluctuations, avoiding the drawbacks of traditional fixed threshold methods and enabling more accurate identification of real-time load status.

[0020] 2. The present invention performs sliding window segmentation on historical load data and uses a clustering algorithm to generate a load behavior map. The present invention can quickly identify the current load status under different load fluctuation patterns. The construction of the load behavior map analyzes typical load fluctuation patterns, such as peak load, valley load, and periodic load, so that the distribution system can map the load status based on the similarity between the real-time data window and the load behavior map. Through this method, the behavior category corresponding to the current load data can be accurately identified, and the control strategy can be adjusted in a targeted manner, effectively improving the prediction accuracy of the load fluctuation range and the control response speed.

[0021] 3. This invention employs multiple clustering models (such as clustering methods based on locality-sensitive hashing) to identify patterns in historical load data, constructing load behavior maps and providing data support for different load states. Combined with the LSTM prediction model, this approach enables customized threshold adjustments for different load behavior categories, demonstrating the system's adaptability and scalability to complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention is a flow chart of a load switch control method for a distribution box. DETAILED DESCRIPTION

[0023] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0024] Example 1, Figure 1 The present invention provides a load switch control method for a distribution box, comprising the following steps:

[0025] S1, obtain historical load data, and adjust the pre-acquired dynamic switch threshold according to the fluctuation range of the historical load data.

[0026] In this embodiment, historical load data is obtained, and the pre-acquired dynamic switch threshold is adjusted according to the fluctuation range of the historical load data, as follows:

[0027] The distribution box system is equipped with electrical parameter acquisition devices such as voltage, current, and frequency to monitor each phase line and circuit in the test area in real time. The active power, reactive power, voltage RMS, current RMS, and frequency parameters of each circuit are synchronously collected at a fixed period to form multi-channel parallel load data.

[0028] The real-time load data is pre-processed at the millisecond level through edge computing nodes, and the sensor data streams are synchronized based on the NTP protocol to generate structured data packets with global timestamps.

[0029] The preprocessed load data is organized in chronological order and reorganized into time series using sliding window technology. The load data of every N consecutive time points (for example, N=30) constitute a time series sample. The dimension of each sample is where F is the feature dimension, and the data at each time point contains several load features (such as active power, reactive power, voltage, current, and frequency);

[0030] Time series samples are input into a long short-term memory neural network (LSTM) model for training to obtain a trained dynamic switch threshold model. The LSTM model learns the temporal dependencies in historical load data to predict the load fluctuation trend at the next moment.

[0031] The dynamic threshold range is constructed based on the predicted value output by the LSTM model and the dynamic switch threshold model.

[0032] In this implementation, the dynamic threshold range is constructed based on the predicted value output by the LSTM model and the dynamic switch threshold model, as follows:

[0033] Based on the load data in the window, the standard deviation of the load data in the window is output to measure the fluctuation intensity of the load data;

[0034] According to the predicted load trend and the standard deviation of the load data in the sliding window, a dynamic threshold model is constructed, and the dynamic threshold range is obtained according to the dynamic threshold model.

[0035] The dynamic threshold model is as follows:

[0036]

[0037]

[0038] Where: is the upper load limit, is the predicted load value at the next moment, is the upper load limit, is a predetermined constant coefficient that is adjusted according to the tolerance of load fluctuations. is the standard deviation of the load data within the sliding window.

[0039] It's important to note that the adaptive adjustment function of the dynamic threshold optimizes itself based on real-time fluctuations in load data. When the standard deviation of the load data is large, the threshold range is expanded to accommodate larger fluctuations; when the load data fluctuations are smaller, the threshold range is narrowed to reduce the risk of false alarms. Furthermore, the dynamic threshold adjustment can incorporate a smoothing factor to prevent excessive threshold changes caused by abnormal fluctuations in a short period of time. This smoothing factor can be implemented using exponential smoothing to ensure threshold stability.

[0040] S2, obtain the clustering effects of different clustering models and obtain the optimal clustering model.

[0041] In this embodiment, the clustering effects of different clustering models are obtained to obtain the optimal clustering model, as follows:

[0042] Preprocess the load characteristic data and reduce its dimension based on the principal component analysis method to obtain sample data;

[0043] Based on the locality-sensitive hashing algorithm, similar sample data is mapped to the same bucket through hash coding to generate a hash value;

[0044] Cluster the hash values ​​and obtain the local neighborhood of each data point based on the cosine similarity method;

[0045] Extracting local features based on data in a local neighborhood, wherein the local features include cluster spacing, cluster boundary clarity, silhouette coefficient, cohesion, and separation;

[0046] The clustering effect of the clustering model is evaluated according to the local features, and the best clustering model is screened out according to the evaluation result. The clustering model includes K-means and hierarchical clustering.

[0047] The formula for calculating cohesion is as follows:

[0048]

[0049] The calculation formula of separation degree is as follows:

[0050]

[0051] The clustering effect evaluation formula is as follows:

[0052]

[0053] Where: is the cohesion, is the kth cluster, is the i-th data point in the cluster, is the jth data point in the cluster, is the separation, is the cluster center of the kth cluster, is the cluster center of the mth cluster, is the number of samples in the dataset, that is, the total number of all data points, is the clustering effect evaluation result, is the total number of clusters.

[0054] It's important to note that high cohesion indicates that data points within the same cluster are very close or compact. This means that the data points within the cluster are highly similar, leading to more accurate and effective clustering results. High cohesion generally improves clustering effectiveness. Low cohesion indicates that the data points within the cluster are highly dissimilar, resulting in poor clustering effectiveness. Large distances between data points may indicate inaccurate clustering.

[0055] Furthermore, high separation (higher Separation values) indicate greater differences between clusters. Larger distances between cluster centers indicate more distinct clustering results and better performance. This means that the boundaries between clusters are clear and the clustering model is effectively distinguishing between samples. Low separation (lower Separation values) indicate less differentiation between clusters and poor clustering performance. Close distances between cluster centers can lead to ambiguous clustering results and poor classification results.

[0056] Finally, a high clustering effect evaluation result indicates that the clustering results have both strong cohesion (tight clusters) and strong separation (clear clusters), indicating a good clustering effect. A low clustering effect evaluation result indicates that the clustering results have poor cohesion (large differences in data points within clusters) and poor separation (indistinct clusters), indicating a poor clustering effect.

[0057] S3, based on the optimal clustering model, construct a load behavior map, map the current load data to the load behavior map, and adjust the initial threshold range in combination with the historical load data fluctuation range to obtain the revised threshold range.

[0058] In this embodiment, a load behavior map is constructed based on the optimal clustering model, as follows:

[0059] Evaluate the clustering effect of each clustering model and obtain the best clustering model;

[0060] Obtain historical load data of the area to be tested and segment the data into sliding windows, where each window contains time series samples of a preset length;

[0061] Input the dimensional tensor into the clustering algorithm for training, and output the load behavior category to which each time window belongs, wherein the load behavior category includes but is not limited to peak load mode, valley load mode, periodic load mode, and burst load mode;

[0062] All sample tensors are clustered by clustering algorithms to obtain several load behavior categories. Each category constitutes a typical load spectrum. Each spectrum contains its cluster center eigenvector and the corresponding load fluctuation range.

[0063] In this embodiment, the current load data is mapped to the load behavior map, and the initial threshold range is adjusted in combination with the historical load data fluctuation range, as follows:

[0064] Perform feature extraction on the real-time load data within the sliding window at the current moment to form a tensor with the same dimension as the load behavior map for similarity matching;

[0065] Based on the cosine similarity method, the similarity between the current feature vector and the central vector of all load behavior maps is obtained, and the load behavior category map with the highest similarity is selected as the mapping map of the current window;

[0066] The maximum deviation value within the historical load fluctuation range of the category is extracted from the mapping spectrum and used to represent the upper deviation correction amount and the lower deviation correction amount as correction factors respectively;

[0067] The initial threshold interval is controlled by the correction factor and combined with the standard deviation of the sliding window load data to obtain the corrected threshold interval.

[0068] The revised dynamic threshold model is as follows:

[0069]

[0070]

[0071] Where: is the upper load limit, is the predicted load value at the next moment, is the upper load limit, is a predetermined constant coefficient that is adjusted according to the tolerance of load fluctuations. is the standard deviation of the load data within the sliding window, is the empirical positive offset correction recorded in the mapped atlas, is the empirical negative offset correction recorded in the mapped atlas.

[0072] S4, extracting the distribution feature vector and classification accuracy parameter of the current load data as the first data, training the optimal classification model based on the first data, and inputting the corrected threshold interval into the optimal classification model to identify the current load state label, wherein the first data includes the distribution feature and the classification accuracy.

[0073] In this embodiment, the distribution feature vector and classification accuracy parameter of the current load data are extracted as the first data, and the optimal classification model is trained based on the first data, specifically as follows:

[0074] Acquire first data of load data, wherein the first data includes distribution characteristics and classification accuracy;

[0075] Obtain load data of the area to be measured, obtain the distribution characteristics of the load data, and analyze the distribution form of the load data using the histogram method;

[0076] Screening load data features based on the distribution characteristics of load data and constructing a subset of load data features;

[0077] The load data feature subset is input into the classification model as a time series sample for classification, and the classification accuracy of the classification model is evaluated based on cross-validation, and the best classification model is selected according to the classification accuracy.

[0078] In this embodiment, load data of the area to be measured is obtained, and first data of the load data is obtained. The first data includes distribution characteristics and expected classification accuracy, as follows:

[0079] Obtaining load data of the area to be measured, obtaining distribution characteristics of the load data, and analyzing the distribution form of the load data using a histogram method, wherein the distribution characteristics include skewness and kurtosis, and the distribution form includes normal distribution and skewed distribution;

[0080] Screening load data features based on distribution features of the load data and constructing a load data feature subset, wherein the load data features include time features and load characteristics;

[0081] The load data feature subset is input into the classification model as a time series sample for classification, and the classification accuracy of the classification model is evaluated based on cross-validation. The best classification model is selected according to the classification accuracy. The classification model includes decision tree, support vector machine, K nearest neighbor, and neural network.

[0082] In this embodiment, the load data characteristics are screened based on the distribution characteristics of the load data, specifically as follows:

[0083] Obtaining a distribution form of load data, wherein the distribution form includes normal distribution, skewed distribution, and multimodal distribution;

[0084] According to the distribution type, characteristic indicators related to the data distribution morphology are extracted, and the characteristic indicators include mean, variance, skewness, kurtosis, maximum value, and minimum value;

[0085] A feature importance ranking method based on correlation analysis is used to evaluate the contribution of each feature index to the load data classification performance;

[0086] According to the evaluation results, the characteristic indicators that have the best effect on improving classification accuracy under the current load data distribution characteristics are screened out, and a load data feature subset is constructed.

[0087] It should be noted that the higher the score of the system's adaptability and subsequent tuning efficiency, the more adaptable the current feature selection and clustering strategy is, and the faster the system can enter a stable state. If the score is low, the system needs to repeatedly adjust the clustering parameters, feature sets, or reselect the model, which increases the operation and maintenance costs and model iteration cycle.

[0088] Therefore, in this embodiment, by introducing a performance score evaluation mechanism to optimize the candidate clustering models, the system's recognition accuracy and control response speed for load status can be effectively improved, thereby enhancing the intelligence, reliability and energy-saving effect of the distribution system operation, which has significant practical value and engineering promotion significance.

[0089] In this embodiment, the modified threshold interval is input into the optimal classification model to identify the current load state label, as follows:

[0090] Cluster each data point using the optimal clustering model and output the initial load state label;

[0091] Input the load characteristic data of the current window into a trained classification model to obtain a predicted load state label, wherein the classification model includes a decision tree model, a support vector machine model, a K-nearest neighbor model, or a neural network model;

[0092] Perform label fusion on the output results of the clustering model and the classification model. The fusion methods include weighted voting method, confidence weighted average method or label consistency test method;

[0093] The load state label obtained after fusion is used as the final recognition result;

[0094] According to the load status tag finally identified, the load switch of the target distribution box is controlled.

[0095] The load status label generally includes:

[0096] Normal load: Load fluctuation is within the preset normal range and no intervention measures are required;

[0097] Overload state: Load fluctuation exceeds the preset safety threshold, and the system needs to automatically take measures, such as switching the load switch to avoid system damage;

[0098] Light load state: Load fluctuation is small, and the system can perform energy-saving regulation and control as needed to optimize load utilization efficiency;

[0099] The control operation strategy for the load switch of the target distribution box is as follows:

[0100] Normal load: Maintain the current load status and keep the power distribution system running normally

[0101] Overload status: When the load status is overload, the system immediately controls the target load switch, cuts off or adjusts the power supply, and protects the electrical equipment from damage;

[0102] Light load state: When the load state is light load, the system can adjust the switch of the load circuit according to demand to improve the system operation efficiency.

[0103] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0104] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0105] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0106] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0107] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0108] Finally: The above description is only a preferred embodiment of the present invention and is 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 in the scope of protection of the present invention.

Claims

1. A load switch control method for a distribution box, characterized in that: The steps include: Obtain real-time load data for the area to be tested, build a dynamic switch threshold control model based on LSTM, and generate the initial threshold range; According to the optimal clustering model, a load behavior map is constructed, the current load data is mapped to the load behavior map, and the initial threshold interval is adjusted in combination with the fluctuation interval of historical load data to obtain the revised threshold interval; Extracting a distribution feature vector of current load data and an expected classification accuracy parameter as first data, and screening an optimal classification model based on the first data; The modified threshold interval is input into the optimal classification model to identify the current load status label.

2. The load switch control method of the distribution box according to claim 1, characterized in that: The real-time load data of the area to be tested is obtained, and a dynamic switch threshold control model is constructed based on LSTM to generate the initial threshold range, as follows: Real-time monitoring of each phase line and circuit of the distribution box in the test area is carried out, and the electrical parameter data of each circuit is synchronously collected at a fixed period to form multi-channel parallel load data; Real-time load data is pre-processed through edge computing nodes, and time synchronization of each data stream is performed based on the NTP protocol; The preprocessed load data is organized in chronological order and the sliding window technique is used to reorganize the time series. The load data of every N consecutive time points constitute a single time series sample, and each sample constitutes a dimensional tensor. Input time series samples into the LSTM model for training to obtain a trained dynamic switch threshold model, and then predict the load fluctuation trend based on the LSTM model; The dynamic threshold range is constructed based on the predicted value output by the LSTM model and the dynamic switch threshold model.

3. The load switch control method of the distribution box according to claim 2, characterized in that: The predicted value output by the LSTM model is combined with the dynamic switch threshold model to construct a dynamic threshold range, as follows: Based on the load data in the window, output the standard deviation of the load data in the window; According to the predicted load trend and the standard deviation of the load data in the sliding window, a dynamic threshold model is constructed to obtain the dynamic threshold range.

4. The load switch control method of the distribution box according to claim 3, characterized in that: The optimal clustering model is as follows: Preprocess the load data and reduce the dimension of the load characteristic data based on the principal component analysis method to obtain sample data; Based on the locality-sensitive hashing algorithm, similar sample data is mapped to the same bucket through hash coding to generate a hash value; Cluster the hash values ​​and obtain the local neighborhood of each data point based on the cosine similarity method; Extract local features based on the data in the local neighborhood; The clustering effect of the clustering model is evaluated based on local features, and the best clustering model is selected.

5. The load switch control method of the distribution box according to claim 4, characterized in that: According to the optimal clustering model, a load behavior map is constructed, as follows: Obtain historical load data of the area to be tested and perform sliding window segmentation, where each window contains time series samples of a preset length; The time series samples are constructed into tensors, and the tensors are input into the clustering model for training, and the load behavior category to which each time window belongs is output; The sample tensors are clustered using the trained clustering model. Based on the clustering results, the time window data belonging to the same category are grouped together to obtain several load behavior categories, each of which constitutes a load behavior map.

6. The load switch control method of the distribution box according to claim 5, characterized in that: The current load data is mapped to the load behavior map, and the initial threshold range is adjusted in combination with the historical load data fluctuation range, as follows: Perform feature extraction on the real-time load data within the sliding window at the current moment to form a tensor with the same dimension as the load behavior map; Based on the cosine similarity method, the similarity between the current feature vector and the central vector of all load behavior maps is obtained, and the load behavior category map with the highest similarity is selected as the mapping map of the current window; The maximum deviation value within the historical load fluctuation range of the category is extracted from the mapping spectrum and used to represent the upper deviation correction amount and the lower deviation correction amount as correction factors respectively; The initial threshold interval is controlled by the correction factor and combined with the standard deviation of the sliding window load data to obtain the corrected threshold interval.

7. The load switch control method of the distribution box according to claim 6, characterized in that: The distribution feature vector of the current load data and the expected classification accuracy parameter are extracted as the first data, and the optimal classification model is screened based on the first data, specifically as follows: Acquire first data of load data, wherein the first data includes distribution characteristics and classification accuracy; Obtain load data of the area to be measured, obtain the distribution characteristics of the load data, and analyze the distribution form of the load data using the histogram method; Screening load data features based on the distribution characteristics of load data and constructing a subset of load data features; The load data feature subset is input into the classification model as a time series sample for classification, and the classification accuracy of the classification model is evaluated based on cross-validation, and the best classification model is selected according to the classification accuracy.

8. The load switch control method of the distribution box according to claim 7, characterized in that: The load data feature screening based on the distribution feature of the load data is as follows: Obtain the distribution form of load data and extract characteristic indicators related to the data distribution form; A feature importance ranking method based on correlation analysis is used to evaluate the contribution of each feature index to the load data classification performance; According to the evaluation results, the characteristic indicators that have the best effect on improving classification accuracy under the current load data distribution characteristics are screened out, and a load data feature subset is constructed.

9. The load switch control method of the distribution box according to claim 8, characterized in that: The modified threshold interval is input into the optimal classification model to identify the current load state label, as follows: Cluster each data point using the optimal clustering model and output the initial load state label; Input the load characteristic data of the current window into the trained classification model to obtain the predicted load state label; Perform label fusion on the output results of clustering model and classification model; The fused load status label is used as the final recognition result, and the load switch of the target distribution box is controlled.

Images