Active power distribution network line loss assessment method based on BP clustering hybrid model
An improved algorithm based on a BP clustering hybrid model was developed to address the shortcomings of the BP neural network line loss model in terms of accuracy and adaptability. This resulted in high-precision, fast-response line loss assessment, adapting to the dynamic changes in the distribution network and reducing computational and resource consumption.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-04-07
AI Technical Summary
Existing BP neural network line loss models have shortcomings in assessment accuracy and cannot be updated adaptively, resulting in inaccurate line loss assessment and a large workload, especially when the distribution network structure changes and it is difficult to adapt quickly.
A hybrid BP clustering model is adopted, which performs joint clustering by improving the K-Means++ algorithm to generate running scenario labels. The hybrid BP clustering model is constructed by selecting features based on the maximum information coefficient, and online incremental learning is carried out by combining the sliding window mechanism and the exponential decay strategy to achieve adaptive updating of the model.
It significantly improves the accuracy of line loss assessment, reduces the mean absolute error, supports rapid migration and refresh, provides real-time decision-making support, meets plug-and-play requirements, and reduces manual intervention and computational resource consumption.
Smart Images
Figure CN121808428A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for assessing power grid line losses. Background Technology
[0002] With the rapid development of my country's economy, electricity load is constantly increasing. Distribution network line losses account for approximately 40% of the total power grid losses, making the line loss problem extremely serious. Line loss, as an important comprehensive indicator, is closely related to the profitability of power companies. It comprehensively reflects the management, planning, design, and operational levels of power companies. To reduce losses in the transmission and distribution of electrical energy, improve the corporate and social benefits of the power industry, and promote its better development, it is essential to attach great importance to line loss calculation. However, distribution networks are characterized by numerous devices, complex lines, low voltage levels, and direct connections to users. These characteristics lead to difficulties in data collection, complex calculations, and a large workload in actual line loss calculations. Currently, line loss calculation methods such as the equivalent resistance method, the maximum load loss hour method, the average current method, and the root mean square current method are widely used in the actual production of power companies. However, in the actual operation of the power grid system, due to the complexity of the distribution network structure and the connections between lines, any changes to the network structure require updates to the relevant power grid line diagrams used for calculations, which greatly increases the workload and difficulty for relevant power departments.
[0003] The rise of artificial intelligence technology in recent years has provided new ideas for research on line loss calculation. Some scholars have introduced machine learning algorithms, such as BP neural networks and RBF neural networks, into theoretical line loss calculation. These algorithms leverage the powerful learning capabilities of machine learning to abstract the line loss calculation process into a multi-dimensional feature extraction process related to line loss. By using a large amount of historical data generated during power grid operation and employing multi-level machine learning models, the essential features of this massive dataset are extracted to discover the potential relationship between features and line loss, thereby simplifying the calculation process and better meeting the power industry's needs for line loss calculation. However, existing BP neural network line loss models generally use static input features and do not dynamically cluster and classify the operating scenarios of distribution transformers. This leads to aliasing of different loads and photovoltaic modes, resulting in slow network convergence; abnormal measurement data is not pre-cleaned, causing prediction drift; and the model has poor generalization ability, requiring retraining when changing distribution transformers. Summary of the Invention
[0004] The purpose of this invention is to solve the problems of poor line loss assessment accuracy and inability to adaptively update existing BP neural network line loss models, and to propose an active distribution network line loss assessment method based on a BP clustering hybrid model.
[0005] The active distribution network line loss assessment method based on a BP clustering hybrid model described in this invention includes the following steps:
[0006] Obtain the electrical characteristic parameters of the transformer area and construct the original measurement dataset;
[0007] The daily load curve and distributed power output curve in the original measurement dataset are normalized and filtered smoothed to form a standardized two-dimensional curve.
[0008] The standardized two-dimensional curves are jointly clustered using an improved K-Means++ algorithm to generate... The improved K-Means++ algorithm introduces a contour coefficient maximization strategy in the initial center selection stage, so that the joint cluster contour coefficient is ≥0.72.
[0009] Based on the maximum information coefficient, 11 features with the highest correlation to line loss rate were selected from the electrical characteristic parameters of the transformer area to form the input vector of the BP clustering hybrid model;
[0010] Regarding the aforementioned Each of the three operating scenarios is independently trained into a BP clustering hybrid model.
[0011] During online operation, the system identifies the operational scenario label of the current transformer area in real time and calls the corresponding BP clustering hybrid model. Using the operational data of the line to be evaluated as input, the system obtains the predicted line loss value for this line.
[0012] Furthermore, the BP clustering hybrid model is a BP sub-network with 11 input layer nodes, a 64×32 double hidden layer structure in the hidden layer, and 1 output layer node.
[0013] Furthermore, the normalization adopts a min-max linear transformation, the filter window length is 15 points, and the polynomial order is 3.
[0014] Furthermore, the specific formula for normalization is:
[0015]
[0016] in, The data is after normalization; This is the original data; The maximum value in the original data; This is the minimum value in the original data.
[0017] Furthermore, the weighting ratio of the daily load curve to the distributed power generation output curve is 1:1.
[0018] Furthermore, the 11-dimensional electrical characteristics include: average active power, average reactive power, three-phase voltage imbalance, average power factor, proportion of reverse power generation of distributed power sources, load factor, maximum load factor, load fluctuation coefficient, photovoltaic output fluctuation coefficient, daily power supply, and equivalent impedance modulus.
[0019] Furthermore, when the cumulative number of new samples in any running scenario exceeds the preset threshold N=200, online incremental learning of the BP clustering hybrid model for that scenario is triggered. A sliding window mechanism is used, and the weights of old samples are forgotten with an exponential decay coefficient of 0.95.
[0020] Furthermore, the aforementioned In each running scenario label The value is determined jointly by the elbow rule and the profile coefficient, and is taken as follows: =5.
[0021] Compared with the prior art, the present invention has the following advantages:
[0022] The accuracy of line loss assessment is significantly improved. Validated in 326 photovoltaic substations in a certain area, the mean absolute error (MAE) is 0.18%, a 42% reduction compared to a single BP network. When the model is migrated to a new substation, the prediction error on the first day is less than 1%, meeting the "plug and play" requirement. The system is deployed on a single unit. It features minute-level online updates and ultra-low latency, with inference time of less than 50 ms per substation, supporting 15-minute line loss rate updates, providing real-time decision-making support for line loss anomaly warnings, network reconfiguration, and voltage and reactive power optimization. It also features adaptive online evolution. When the operating scenario changes (addition of residential photovoltaic systems, seasonal load migration), incremental learning is triggered when the cumulative number of samples exceeds 200. The old sample index forgetting coefficient is 0.95, and weight hot updates can be completed within 2 minutes without manual intervention. The model lifecycle has been upgraded from a fixed quarterly period to "continuous lifelong learning." Attached Figure Description
[0023] Figure 1 This is a flowchart of an active distribution network line loss assessment method based on a BP clustering hybrid model, as described in Specific Implementation Method 1. Detailed Implementation
[0024] Specific Implementation Method 1: Combination Figure 1 This embodiment describes an active distribution network line loss assessment method based on a BP clustering hybrid model, characterized by the following steps:
[0025] Obtain the electrical characteristic parameters of the transformer area and construct the original measurement dataset;
[0026] The daily load curve and distributed power output curve in the original measurement dataset are normalized and filtered smoothed to form a standardized two-dimensional curve.
[0027] The standardized two-dimensional curves are jointly clustered using an improved K-Means++ algorithm to generate... The improved K-Means++ algorithm introduces a contour coefficient maximization strategy in the initial center selection stage, so that the joint cluster contour coefficient is ≥0.72.
[0028] Based on the maximum information coefficient, 11 features with the highest correlation to line loss rate were selected from the electrical characteristic parameters of the transformer area to form the input vector of the BP clustering hybrid model;
[0029] Regarding the aforementioned Each of the three operating scenarios is independently trained into a BP clustering hybrid model.
[0030] During online operation, the system identifies the operational scenario label of the current transformer area in real time and calls the corresponding BP clustering hybrid model. Using the operational data of the line to be evaluated as input, the system obtains the predicted line loss value for this line.
[0031] In this implementation, the accuracy of line loss assessment is significantly improved. Validated across 326 photovoltaic distribution areas in a specific region, the average absolute error (MAE) is 0.18%, a 42% reduction compared to a single BP network. When the model is migrated to a new distribution area, the prediction error on the first day is less than 1%, meeting the "plug-and-play" requirement. The system is deployed on a single unit; it features minute-level online updates and ultra-low latency, with inference time per distribution area less than 50 ms, supporting 15-minute-level line loss rate updates, providing real-time decision-making support for line loss anomaly warnings, network reconfiguration, and voltage and reactive power optimization. It also features adaptive online evolution, adapting to changes in the operating scenario (addition of residential photovoltaic installations, seasonal load migration).
[0032] Specific Implementation Method Two: This implementation method further defines the active distribution network line loss assessment method based on the BP clustering hybrid model described in Specific Implementation Method One. In this implementation method, the BP clustering hybrid model is a BP sub-network with 11 input layer nodes, a 64×32 double hidden layer structure in the hidden layer, and 1 output layer node.
[0033] In this embodiment, the dual-hidden-layer structure reduces the MAE by 0.05 percentage points compared to the single-hidden-layer (11-128-1) structure on the validation set.
[0034] Specific Implementation Method 3: This implementation method further defines the active distribution network line loss assessment method based on the BP clustering hybrid model described in Specific Implementation Method 1. In this implementation method, the normalization adopts the min-max linear transformation, the filter window length is 15 points, and the polynomial order is 3.
[0035] In this embodiment, compared with no filtering, MAE is reduced by 0.04 percentage points; the 15-minute spike caused by HPLC communication is effectively suppressed, and the curve smoothness is improved by 32% (quantified by the second derivative norm).
[0036] Specific Implementation Method Four: This implementation method further defines the active distribution network line loss assessment method based on a BP clustering hybrid model described in Specific Implementation Method Three. In this implementation method, the specific normalization formula is as follows:
[0037]
[0038] in, The data is after normalization; This is the original data; The maximum value in the original data; This is the minimum value in the original data.
[0039] In this embodiment, normalizing historical data can unify the magnitude of different types of data and improve the prediction accuracy of the model output.
[0040] Specific Implementation Method Five: This implementation method further defines the active distribution network line loss assessment method based on the BP clustering hybrid model described in Specific Implementation Method One. In this implementation method, the weight ratio of the daily load curve to the distributed power generation output curve is 1:1.
[0041] In this embodiment, compared with the "load only" scenario, the clustering profile coefficient after introducing photovoltaics increased from 0.61 to 0.72; the "high photovoltaic power generation + light load" high line loss scenario was accurately identified, and the prediction error of this type of transformer area was reduced from 0.35% to 0.19%.
[0042] Specific Implementation Method Six: This implementation method further defines the active distribution network line loss assessment method based on the BP clustering hybrid model described in Specific Implementation Method One. In this implementation method, the 11-dimensional electrical characteristics include: average active power, average reactive power, three-phase voltage imbalance, average power factor, proportion of reverse power generation of distributed sources, load factor, maximum load factor, load fluctuation coefficient, photovoltaic output fluctuation coefficient, daily power supply, and equivalent impedance modulus.
[0043] In this implementation, the original 57-dimensional features are filtered down to 11 dimensions using the maximum information coefficient, retaining 98.7% mutual information, reducing training time by 55%, and memory usage by 50%.
[0044] Specific Implementation Method Seven: This implementation method further defines the active distribution network line loss assessment method based on the BP clustering hybrid model described in Specific Implementation Method One. In this implementation method, when the cumulative number of new samples in any operating scenario exceeds the preset threshold N = 200, online incremental learning of the BP clustering hybrid model for that scenario is triggered. A sliding window mechanism is adopted, and the weights of old samples are forgotten with an exponential decay coefficient of 0.95.
[0045] In this implementation, the triggering conditions are: more than 200 new samples are added in the same scenario; the sliding window size is 1000; the old sample weight index decay number is equal to 0.95; the BP clustering hybrid model can automatically evolve with the seasons and photovoltaic capacity expansion, and the MAE is always <0.2% after 6 months of continuous operation, without the need for manual retraining; compared with full retraining, the time for a single incremental step is reduced from 45 minutes to 90 seconds, saving 97% of computing power.
[0046] Specific Implementation Method Eight: This implementation method further defines the active distribution network line loss assessment method based on a BP clustering hybrid model described in Specific Implementation Method One. In this implementation method, the... In each running scenario label The value is determined jointly by the elbow rule and the profile coefficient, and is taken as follows: =5.
[0047] In this embodiment, Traverse from 2 to 10, elbow point =5, The profile coefficient peaked at 0.72 and continued to increase. The silhouette coefficient no longer increases significantly, i.e. =5 makes the curve shape within the cluster most representative and the separation between clusters the greatest.
[0048] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for assessing line losses in active distribution networks based on a BP clustering hybrid model, characterized in that, Includes the following steps: Obtain the electrical characteristic parameters of the transformer area and construct the original measurement dataset; The daily load curve and distributed power output curve in the original measurement dataset are normalized and filtered smoothed to form a standardized two-dimensional curve. The standardized two-dimensional curves are jointly clustered using an improved K-Means++ algorithm to generate... The improved K-Means++ algorithm introduces a contour coefficient maximization strategy in the initial center selection stage, so that the joint cluster contour coefficient is ≥0.
72. Based on the maximum information coefficient, 11 features with the highest correlation to line loss rate were selected from the electrical characteristic parameters of the transformer area to form the input vector of the BP clustering hybrid model; Regarding the aforementioned Each of the three operating scenarios is independently trained into a BP clustering hybrid model. During online operation, the system identifies the operational scenario label of the current transformer area in real time and calls the corresponding BP clustering hybrid model. Using the operational data of the line to be evaluated as input, the system obtains the predicted line loss value for this line.
2. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 1, characterized in that, The BP clustering hybrid model is a BP sub-network with 11 input layer nodes, a 64×32 double hidden layer structure in the hidden layer, and 1 output layer node.
3. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 1, characterized in that, The normalization uses a min-max linear transformation, the filter window length is 15 points, and the polynomial order is 3.
4. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 3, characterized in that, The specific formula for normalization is: in, The data is after normalization; This is the original data; The maximum value in the original data; This is the minimum value in the original data.
5. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 1, characterized in that, The weighting ratio of the daily load curve to the distributed power generation output curve is 1:
1.
6. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 1, characterized in that, The 11-dimensional electrical characteristics include: average active power, average reactive power, three-phase voltage imbalance, average power factor, proportion of reverse power generation of distributed power sources, load factor, maximum load factor, load fluctuation coefficient, photovoltaic output fluctuation coefficient, daily power supply, and equivalent impedance modulus.
7. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 1, characterized in that, When the cumulative number of new samples in any running scenario exceeds the preset threshold N=200, the online incremental learning of the BP clustering hybrid model for that scenario is triggered. A sliding window mechanism is used, and the weights of old samples are forgotten with an exponential decay coefficient of 0.
95.
8. The method for assessing line losses in active distribution networks based on a BP clustering hybrid model according to claim 1, characterized in that, The In each running scenario label The value is determined jointly by the elbow rule and the profile coefficient, and is taken as follows: =5.