A method, system and application for grading the adjustable potential of a photovoltaic power station

Abstract

The application discloses a kind of station area photovoltaic adjustable potential grading method, system and application, it is related to intelligent station area and electric power big data analysis technical field.The application first obtains the source and load characteristic data of station area user, carries out net load calculation and pretreatment to source and load characteristic data;The net load indicates that user sends power to station area or takes electricity from station area;Classify after pretreatment source and load characteristic data, obtain user category and corresponding power consumption behavior characteristic curve;According to power consumption behavior characteristic curve, the adjustable load power consumption proportion of each type of user is calculated, and photovoltaic adjustable potential grading standard is constructed.The application can fully exploit the time sequence and local characteristics of source and load data, significantly improve the grading accuracy, and provide a scientific basis for the station area to develop differentiated photovoltaic consumption and regulation strategies.

Description

Technical Field

[0001] This invention relates to the field of smart grid and power big data analysis technology, specifically a method for identifying, classifying, and grading the adjustable potential of source and load characteristics of distributed photovoltaic users in low-voltage distribution areas using deep learning and hybrid clustering algorithms. Background Technology

[0002] With the high proportion of distributed photovoltaic power being connected to distribution substations, its randomness, volatility, and intermittency have brought severe challenges to low-voltage distribution networks.

[0003] Increased uncertainty in power source and load: Traditionally, the load of a distribution area is mainly affected by user behavior. However, after a high proportion of photovoltaic power is connected, the distribution area exhibits "dual uncertainty in power source and load," which can easily lead to safety hazards such as outgoing line voltage exceeding limits, transformer reverse overload, three-phase imbalance, and power flow reversal.

[0004] Limitations of existing evaluation methods: Traditional photovoltaic (PV) user profiling techniques often employ basic clustering algorithms such as K-means and DBSCAN. However, the K-means algorithm is highly sensitive to the selection of initial cluster centers; randomly selected centers often lead the algorithm into local optima. Furthermore, it lacks robustness to noisy data and outliers, making it difficult to accurately distinguish the behavior of PV users with complex temporal characteristics.

[0005] Insufficient Accuracy of Grading Models: In potential prediction and grading, existing technologies mostly employ single recurrent neural networks (RNNs) or long short-term memory networks (LSTMs). While LSTMs excel at handling long-distance dependencies in time series, they often struggle to fully extract local abrupt changes (such as sudden drops in photovoltaic power due to cloud cover) when dealing with long-sequence source-load data (e.g., 24-hour high-frequency data). Furthermore, traditional models assign equal attention to all time steps, easily overlooking the decisive weight of key time steps (e.g., the midday peak of photovoltaic power generation or the evening peak of load) in overall potential assessment. In addition, existing methods often directly use statistical indicators such as "photovoltaic penetration rate" (total power generation / total power consumption) or "peak-valley difference" for potential grading. However, the impact of distributed photovoltaic power on distribution areas is highly time-dependent, and these static indicators ignore the temporal matching degree of source-load interaction. For example, two users with the same photovoltaic penetration rate may have vastly different potential levels; one might experience severe backfeeding at midday, while the other might achieve full local consumption. Assigning them to the same potential level solely based on penetration rate would lead to the failure of control strategies and a waste of regulatory resources. Therefore, there is an urgent need for a hierarchical indicator and method that can combine user temporal behavior characteristics to accurately quantify the potential within the "effective adjustment period".

[0006] Therefore, there is an urgent need for a technical solution that can deeply mine the spatiotemporal characteristics of source data, overcome the shortcomings of traditional clustering algorithms, and achieve high-precision, automated potential classification. Summary of the Invention

[0007] The purpose of this invention is to deeply mine the spatiotemporal characteristics of source load data, overcome the shortcomings of traditional clustering algorithms, and achieve high-precision, automated potential classification.

[0008] The present invention solves the above-mentioned technical problems through the following technical means:

[0009] A method for classifying the adjustable potential of photovoltaic (PV) power in a distribution area includes: acquiring source-load characteristic data of users in the distribution area; performing net load calculation and preprocessing on the source-load characteristic data; wherein the net load represents the difference between the user's load and the user's PV output; classifying the preprocessed source-load characteristic data to obtain user categories and corresponding electricity consumption behavior characteristic curves, wherein the electricity consumption behavior characteristic curves represent the time distribution characteristics of user load and PV output; calculating the adjustable load electricity ratio of each type of user based on the electricity consumption behavior characteristic curves, and constructing a PV adjustable potential classification standard; wherein the formula for calculating the adjustable load electricity ratio is:

[0010] In the formula, For adjustable load power ratio, Let be the net load at time t, when This indicates that the user is feeding power back to the transformer area; when When this time, it indicates that the user is drawing power from the transformer substation; The user's original load at time t, This represents the cumulative sum of the original user load at all times within the time interval [1, T]. To identify the time windows during which users can achieve peak shaving / valley filling based on electricity consumption behavior characteristic curves, a set of time periods that users can adjust is defined.

[0011] Furthermore, the method for classifying the preprocessed source payload feature data is as follows: The first stage: use the Kmeans++ algorithm to calculate the initial cluster centers, that is, select sample points that are as far apart as possible from each other as the initial centers; The second stage involves mapping the initial cluster centers to particle positions in the PSO algorithm, using the reciprocal of the sum of squared errors within each cluster as the fitness function, and finding the globally optimal cluster centers through iterative updates of particle positions and velocities. The third stage: The global optimal solution output by the PSO algorithm is used as the initial center of the Kmeans algorithm, and local iterative convergence is performed until the stopping condition is met, and the final clustering result is output.

[0012] Furthermore, the user categories include: Type A: The user's net load is continuously negative during the midday period, and the proportion of photovoltaic output is greater than the first set value; Type B: Peak user load is concentrated at night, and the misalignment rate with the photovoltaic output period is greater than the second set value; the misalignment rate is the proportion of electricity consumption during peak load periods that do not overlap with photovoltaic output periods to the total peak electricity consumption, reflecting the degree of time misalignment between electricity consumption and photovoltaic output. Type C: The load curve is stable throughout the day, and the peak-to-valley difference coefficient is less than the third set value; Type D: The load peak-valley difference is large and overlaps with the photovoltaic output period. The load peak-valley difference coefficient is greater than the fourth set value, and the overlap period between the peak electricity consumption period and the photovoltaic output period is greater than the set duration. The electricity price sensitivity coefficient is greater than the fifth set value. The peak load period and photovoltaic output period are identified based on the characteristic curves of user electricity consumption behavior.

[0013] Furthermore, the photovoltaic output time misalignment rate Defined as:

[0014] in, Let be the predicted photovoltaic power output at time t. For the first The actual photovoltaic output power at any given time The time offset represents the time deviation between the prediction and the actual output; the predicted photovoltaic power output refers to the typical daily power curve extracted based on the user's historical electricity consumption data and the photovoltaic historical power output data, which respectively represent the time distribution characteristics of user electricity consumption behavior and photovoltaic power output.

[0015] Furthermore, the formula for calculating the peak-valley difference coefficient is as follows:

[0016] in, The highest load value within the statistical period. The lowest load value within the statistical period. This represents the peak-to-valley difference coefficient.

[0017] Furthermore, the electricity price sensitivity coefficient The calculation formula is:

[0018] in, This refers to the electricity consumption for the current month and the previous month. Electricity prices for the current month and the previous month

[0019] Furthermore, the photovoltaic adjustable potential grading standard is as follows: Level 1 potential: The corresponding strategy is to configure distributed energy storage; Secondary potential: The corresponding strategy is to implement time-of-use pricing to guide electricity prices. Level 3 potential: The corresponding strategy is to participate in peak shaving and valley filling auxiliary services; Level 4 potential: The corresponding strategy is not to use it as a primary adjustment resource.

[0020] The present invention also provides a grading system for the adjustable potential of photovoltaic distribution areas, comprising: Data acquisition module: used to acquire source load characteristic data of users in the transformer area, and to perform net load calculation and preprocessing on the source load characteristic data; Clustering analysis module: Classifies the preprocessed source-load characteristic data to obtain the electricity consumption behavior characteristic curves of different types of users; Standard construction module: Calculates the adjustable load power ratio of various user types based on electricity consumption behavior characteristic curves, and constructs a photovoltaic adjustable potential grading standard; the formula for calculating the adjustable load power ratio is as follows:

[0021] In the formula, For adjustable load power ratio, Let be the net load at time t, when This indicates that the user is feeding power back to the transformer area; when When this time, it indicates that the user is drawing power from the transformer substation; This represents the cumulative sum of the original user load at all times within the time interval [1, T], where T is the total number of time intervals. To identify the time windows during which users can achieve peak shaving / valley filling based on electricity consumption behavior characteristic curves, a set of time periods that users can adjust is defined.

[0022] Furthermore, the method for classifying the preprocessed source payload feature data is as follows: The first stage: use the Kmeans++ algorithm to calculate the initial cluster centers, that is, select sample points that are as far apart as possible from each other as the initial centers; The second stage involves mapping the initial cluster centers to particle positions in the PSO algorithm, using the reciprocal of the sum of squared errors within each cluster as the fitness function, and finding the globally optimal cluster centers through iterative updates of particle positions and velocities. The third stage: The global optimal solution output by the PSO algorithm is used as the initial center of the Kmeans algorithm, and local iterative convergence is performed until the stopping condition is met, and the final clustering result is output.

[0023] Furthermore, the different types of users include: Type A: The user's net load is continuously negative during the midday period, and the proportion of photovoltaic power generation is greater than the first set value; Type B: Peak user load is concentrated at night, and the misalignment rate with the photovoltaic output period is greater than the second set value; the misalignment rate is calculated as the complement of the proportion of the peak electricity consumption period and the photovoltaic output period. Type C: The load curve is stable throughout the day, and the peak-to-valley difference coefficient is less than the third set value; Type D: The load peak-valley difference is large and overlaps with the photovoltaic period. The load peak-valley difference coefficient is greater than the fourth set value, and the overlap time between the peak electricity consumption period and the photovoltaic output period is greater than the set duration. The electricity price sensitivity coefficient is greater than the fifth set value.

[0024] Furthermore, the photovoltaic output time misalignment rate Defined as:

[0025] in, This represents the total number of time periods. Let be the predicted photovoltaic power output at time t. For the first The actual photovoltaic output power at any given time This is the time offset, representing the time difference between the prediction and the actual time.

[0026] Furthermore, the formula for calculating the peak-valley difference coefficient is as follows:

[0027] in, The highest load value within the statistical period. The lowest load value within the statistical period. This represents the peak-to-valley difference coefficient.

[0028] Furthermore, the electricity price sensitivity coefficient The calculation formula is:

[0029] in, This refers to the electricity consumption for the current month and the previous month. Electricity prices for the current month and the previous month

[0030] Furthermore, the photovoltaic adjustable potential grading standard is as follows: Level 1 potential: The corresponding strategy is to configure distributed energy storage; Secondary potential: The corresponding strategy is to implement time-of-use pricing to guide electricity prices. Level 3 potential: The corresponding strategy is to participate in peak shaving and valley filling auxiliary services; Level 4 potential: The corresponding strategy is not to use it as a primary adjustment resource.

[0031] This invention also provides a method for constructing an intelligent hierarchical model of the adjustable potential of photovoltaic power generation in a transformer substation. The method employs a CNN-LSTM deep learning model with an added attention mechanism. The user type obtained by the above method is used as a label, and the source load feature data of the user is used as input to train the CNN-LSTM deep learning model to obtain an intelligent hierarchical model.

[0032] This invention also provides an application of an intelligent grading model, in which the intelligent grading model trained using the above method is used to grade the adjustable potential of newly connected photovoltaic users in a distribution area.

[0033] The advantages of this invention are: 1. Construct a multidimensional source-load feature dataset: This invention not only considers single electricity consumption, but also cleanses and normalizes the photovoltaic power generation and user electricity consumption of a distribution area within 24 hours, constructing a "net load" index that reflects the characteristics of source-load interaction, providing a high-quality data foundation for subsequent analysis. This invention establishes a quantitative potential grading standard and strategy matching: based on typical curves obtained from clustering, it quantitatively calculates the "adjustable load power ratio (…)". The system transforms abstract user behavior into four specific potential levels and matches them with differentiated control strategies such as energy storage promotion, time-of-use pricing, and demand response, which has strong practical engineering value.

[0034] This method achieves a deep integration of "profile" and "potential": it uses clustering algorithms to accurately identify user behavior patterns (such as photovoltaic-dependent and rigid load types), and then combines... By grading values ​​according to their magnitude, the classification results incorporate both temporal and numerical characteristics, significantly improving the accuracy of the assessment. Finally, this grading standard can directly correspond to differentiated control strategies, such as... For users with more than 50% photovoltaic dependence (Level 1), priority is given to configuring energy storage to solve the problem of reverse power transmission; for users with flexible and adjustable loads (Level 2), time-of-use pricing is used to guide the supply; and for rigid loads (Level 4), the focus is on ensuring supply. This "one policy for one type" management model can maximize the value of the regulation resources in the distribution area and provide a scientific basis for the distribution area to formulate refined photovoltaic consumption and demand response strategies.

[0035] 2. This invention proposes an improved PSO-Kmeans++ hybrid clustering algorithm: Addressing the problem of Kmeans clustering easily getting trapped in local optima, this invention proposes a three-stage hybrid strategy. Kmeans++ is used to optimize the initial center selection, ensuring the rationality of the initial particle distribution; the global optimization capability of Particle Swarm Optimization (PSO) is introduced, guiding particles to escape local optima through a fitness function; finally, the fast convergence characteristic of Kmeans is used for fine-tuning. This algorithm can accurately identify four typical behaviors: photovoltaic-dependent, nighttime electricity consumption, rigid load, and flexible adjustable types, significantly improving the silhouette coefficient and inter-cluster separation. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the overall technical route of the method in Embodiment 1 of the present invention.

[0037] Figure 2 This is a detailed execution logic block diagram of the improved PSO-Kmeans++ hybrid algorithm in Embodiment 1 of the present invention.

[0038] Figure 3 This is a comparison chart of source-load characteristic curves for a typical photovoltaic-dependent user in Embodiment 1 of the present invention (showing the morphological differences in the net load region).

[0039] Figure 4 This is a comparison chart of source-load characteristic curves for a typical nighttime electricity user in Embodiment 1 of the present invention (showing the morphological differences in the net load area).

[0040] Figure 5 This is a comparison diagram of the source-load characteristic curves of a typical user with rigid load in Embodiment 1 of the present invention (showing the morphological differences of the net load area).

[0041] Figure 6 This is a comparison chart of the source-load characteristic curves of a typical user with a flexible and adjustable load in Embodiment 1 of the present invention (showing the morphological differences in the net load region).

[0042] Figure 7 This is a schematic diagram of the network topology of the potential grading model based on Attention-CNN-LSTM in Embodiment 2 of the present invention.

[0043] Figure 8 The results of the clustering quality assessment index (SSE elbow method diagram and profile coefficient diagram) analysis in Embodiment 2 of the present invention are shown.

[0044] Figure 9 This is a CNN-LSTM hybrid matrix diagram in Embodiment 2 of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1 This embodiment provides a method for classifying the adjustable potential of photovoltaic (PV) power in a distribution area. First, it acquires time-series data on PV power generation and consumption of users in the distribution area, performs net load calculation, and handles outliers. Second, it proposes an improved PSO-Kmeans++ hybrid clustering algorithm, utilizing the global optimization capability of the particle swarm optimization algorithm to optimize the initial cluster centers, overcoming the shortcomings of traditional algorithms that easily get trapped in local optima. It classifies user source-load characteristic data to obtain user categories and corresponding electricity consumption behavior characteristic curves. These curves represent the time distribution characteristics of user electricity consumption behavior and PV output. Based on these curves, users are accurately classified into four categories: PV-dependent, nighttime electricity consumption, rigid load, and flexible adjustable. Third, it calculates the proportion of adjustable load electricity based on the source-load characteristic curves, constructing a PV adjustable potential classification standard. Finally, it constructs a CNN-LSTM deep learning model incorporating an attention mechanism, training it with the clustering results as labels to achieve rapid and automated classification of users' PV adjustable potential. Figure 1 As shown, the specific process is as follows: Step 1: Acquisition and Refined Preprocessing of Multidimensional Time Series Data To ensure the accuracy of the model input, this step involves rigorous cleaning of the raw data.

[0047] Data collection: Data is collected from users within the distribution area through smart converged terminals and smart meters. exist During the period ( User photovoltaic power output sequence (with a time resolution of 1 hour) and user load sequence .

[0048] Net load calculation: Constructing net load indicators that reflect the interaction characteristics between users and transformer areas. :

[0049] when This indicates that the user is feeding power back to the transformer area; when This indicates that the user is drawing power from the transformer substation. Net load. This represents the difference between a user's load and their photovoltaic output, that is, the difference between the amount of electricity a user draws from the distribution transformer and the amount of electricity they supply to the transformer. Outlier handling and normalization: Lagrange interpolation is used to fill in missing data caused by packet loss in communication.

[0050] The Min-Max normalization method is used to map all data to the interval [0, 1] to eliminate the influence of units:

[0051] in, For the raw data values ​​that need to be standardized, The minimum value in the original data. The maximum value in the original data. This is the new value after standardization.

[0052] The resulting feature vector set serves as the input for subsequent models.

[0053] Step 2: User behavior clustering analysis based on improved PSO-Kmeans++, such as... Figure 2 As shown.

[0054] 1. Kmeans++ initialization phase (optimization starting point): Traditional random initialization is prone to causing cluster center stacking. This method first randomly selects a sample from the dataset as the first center. For subsequent centers The probability of it being selected The shortest distance from this point to the existing center It is proportional to the square of:

[0055] in, For the entire dataset, For a sample point, For sample points The shortest distance to all currently selected cluster centers. It is the square of the shortest distance. For the sample points in the current step The probability of being selected as the next cluster center.

[0056] Repeat this process until a selection is made. An initial center is established. This step ensures the dispersion of the initial particles in the solution space.

[0057] 2. PSO Global Optimization Phase (Escape from Local Traps): The above K center coordinates are encoded into a particle position vector in the PSO algorithm. .

[0058] Fitness Function: Defined as the reciprocal of the sum of squared errors within each class (SSE), the goal is to find the cluster partition with the minimum SSE.

[0059] in, fitness value This is the squared error value. The preset number of clusters, The preset number of clusters, For the first A set of clusters, Belongs to cluster A sample point, For the first The center of each cluster, For vectors and The distance between them.

[0060] Particle update formula: Particles update based on their own historical best solutions. and the global optimal solution of the population Update speed v and position x:

[0061] in, The updated velocity is w, where w is the inertia weight. For the first The particle in the first During the nth iteration, at the... Speed ​​on the dimension, , For learning factors; , A random number between [0,1] No. The particle in the first The best historical position for an individual on the dimension For the entire particle swarm in the th The best position in the global history on the dimension.

[0062]

[0063] in, The updated position For the first The particle in the first During the nth iteration, at the... Current position on the dimension.

[0064] 3. Local convergence stage of K-means (refined convergence): The global optimum position output by the PSO algorithm The PSO (Programmable Optimal Search) centers are used as the initial centers for the standard K-means algorithm, and then iteratively converged. This hybrid strategy utilizes the global search capability of PSO while retaining the efficient convergence characteristics of K-means.

[0065] 4. Clustering result verification and classification definition: The silhouette coefficient and Calinski-Harabasz (CH) index were used for evaluation. A clustering pattern was considered effective when the average silhouette coefficient was >0.6. Figures 3 to 6 As shown, users are ultimately divided into the following four categories: Type A (PV-dependent): The net load of users during the midday period (11:00-14:00) is continuously negative (PV power generation far exceeds electricity consumption, and the net load value is below 0) for ≥3 hours. PV power generation accounts for >50%, and the reverse power transmission is large, with high regulation value, making it suitable for energy storage and consumption.

[0066] Type B (Nighttime Electricity Consumption): The peak load (peak electricity consumption period) is concentrated between 18:00 and 24:00, with a misalignment rate of >80% with the photovoltaic output period (8:00-18:00). This reflects that the electricity consumption period and the photovoltaic output period are completely out of sync, and load shifting in the time dimension is required.

[0067] Photovoltaic output time mismatch rate It can be defined as:

[0068] in, This represents the total number of time periods (e.g., hours or time steps). Let be the predicted photovoltaic power output at time t. For the first The actual photovoltaic output power at any given time This represents the time offset (i.e., the number of misaligned periods, which can be positive or negative, indicating the time deviation between the prediction and the actual output). Predicted photovoltaic (PV) output power refers to the typical daily power curve extracted from historical user electricity consumption data and historical PV output data, representing the time distribution characteristics of user electricity consumption behavior and PV output, respectively. The predicted power on the user side is the statistical average of electricity consumption at each moment on a typical day, and the predicted power on the PV side is the statistical average of PV output power at each moment on a typical day. Misalignment rate R m This refers to the proportion of electricity consumption that does not overlap with the photovoltaic output period during peak electricity consumption hours, reflecting the degree of time misalignment between electricity consumption and photovoltaic output.

[0069] Type C (Rigid Load): The curve is stable with small fluctuations. The load curve is stable throughout the day, the peak-to-valley difference coefficient is <1.2, the adjustment potential is the lowest, and the focus is on ensuring power supply.

[0070] Peak-valley difference coefficient The calculation formula is:

[0071] in, This refers to the highest load value within the statistical period (usually measured in days). This refers to the lowest load value within the statistical period (usually measured in days). The peak-valley difference coefficient, expressed as a percentage, reflects the degree of load fluctuation.

[0072] Type D (Flexible Adjustable): Large load peak-valley difference, with overlap with photovoltaic periods, load peak-valley difference coefficient >1.5, and the overlap between peak electricity consumption period and photovoltaic period (8:00-18:00) is >4 hours, with electricity price sensitivity coefficient >0.6, suitable for demand response.

[0073] Electricity price sensitivity coefficient The calculation formula is:

[0074] in, This refers to the electricity consumption for the current month and the previous month. This represents the electricity price for the current month and the previous month. A positive value indicates that electricity consumption and electricity price change in the same direction, while a negative value indicates that they change in opposite directions.

[0075] In this implementation, peak electricity consumption periods refer to the characteristic curves of typical daily electricity consumption behavior extracted from users' historical electricity consumption data. A threshold of 1.2 times the daily average load (or the 90th percentile load) is used, and periods exceeding this threshold consecutively are defined as peak electricity consumption periods. Photovoltaic output periods are the typical daily photovoltaic output periods extracted from users' historical photovoltaic output data.

[0076] To objectively evaluate the performance of the improved PSO-Kmeans++ hybrid clustering algorithm described in this step, we used the sum of squared errors (SSE) elbow method and the average silhouette coefficient to visualize and analyze the clustering results. The results are as follows: Figure 8 As shown.

[0077] Figure 8(a) - SSE elbow plot: This plot shows the change in the sum of squared errors (SSE) within clusters as the number of clusters (K value) increases from 1 to 8. It can be observed that when K=4, the curve shows a clear "inflection point" (elbow point), and the rate of decrease in SSE slows significantly. This indicates that selecting 4 cluster centers can effectively capture the main structure of the data while ensuring model simplicity, statistically validating the rationality of classifying users into four categories (PV-dependent, nighttime electricity-consuming, rigid load, and flexible adjustable).

[0078] Figure 8 Figure (b) – Silhouette Coefficient Plot: This plot shows the silhouette coefficient distribution of each sample and the overall average silhouette coefficient when K=4. The silhouette coefficient measures the tightness of a sample with its own cluster and its separation from other clusters. Its value ranges from -1 to 1, with a value closer to 1 indicating a better clustering effect. As shown in the figure, the silhouette coefficient distribution of the four user classes obtained by this method is generally high, and the average silhouette coefficient is greater than 0.6, which meets the preset effective clustering threshold. From the perspective of clustering quality, this confirms that the hybrid algorithm can produce user classification results with good cohesion and high separation, laying a reliable foundation for subsequent accurate classification.

[0079] Step 3: Grading and Quantifying the Adjustable Potential of Photovoltaics Based on the four typical curves obtained in step two, we define "adjustable load power ratio". As a quantitative indicator for grading.

[0080] Traditional evaluation metrics typically calculate the electricity consumption ratio across all time periods, failing to distinguish between "time periods that affect the transformer area" and "time periods that have no impact." The calculation formula proposed in this invention is as follows:

[0081] In the formula, The user's original load at time t, Let be the net load at time t. This represents the cumulative sum of the original load power of users at all times within the time interval [1, T].

[0082] The key lies in Determination of (the set of effective adjustment periods): It is not a fixed 24 hours, but rather dynamically determined based on the typical user behavior characteristic curves identified through clustering in step two. Specifically, the effective adjustment period set... This refers to identifying time windows based on the electricity consumption behavior characteristic curves of different types of users, where the load adjustment potential meets a preset threshold and does not affect the user's core electricity needs and comfort. This set needs to be dynamically updated in conjunction with the supply and demand balance requirements of the distribution area, prioritizing time periods that can achieve load peak shaving / valley filling and have high user acceptance, in order to ensure the feasibility and effectiveness of demand response.

[0083] For photovoltaic-dependent (Category A) users identified by clustering, Defined as net load The period when the power supply is less than 0 (i.e., the period of reverse power transmission). At this point, what is quantified is the proportion of its back-feeding power to the total electricity load, directly reflecting the necessity of configuring energy storage; for flexible and adjustable (Class D) users, Defined as the period when the peak load of the distribution area overlaps with the photovoltaic output. It quantifies the peak-shaving and valley-filling potential during that period.

[0084] This embodiment introduces a method defined by clustering results. This indicator essentially performs a "feature mask" processing on the source-load data. Compared with the commonly used photovoltaic penetration rate indicator in existing technologies, By eliminating interference from data during non-critical periods, the data can more accurately reflect the "adjustable value" of the distribution area to users, thus providing a precise quantitative basis for subsequent differentiated strategies (energy storage, electricity prices, demand response).

[0085] based on Based on the magnitude of the value and its cluster category, a hierarchical standard library is constructed: Level 1 potential ( Category A): Extremely high regulation value, distributed energy storage is recommended.

[0086] Secondary potential ( Category D): High regulation value, time-of-use pricing is recommended.

[0087] Level 3 potential ( Category B): Medium adjustment value, recommended for participation in peak shaving and valley filling auxiliary services.

[0088] Level 4 potential ( Category C): Low regulatory value, not used as a primary regulatory resource.

[0089] Compared to existing grading indicators, which typically involve simply calculating "total photovoltaic power generation / total electricity consumption" (i.e., penetration rate) or "maximum reverse power," these indicators are static, full-time statistics that ignore "temporal effectiveness." For example, a user may generate a large amount of electricity throughout the day, but this is concentrated during periods of high load, so no adjustment is actually needed; or a user may have a large reverse power, but the duration is extremely short, making it worthless for adjustment.

[0090] In this embodiment, the indicator (η) (time-series coupled type) introduces a "time period set" T. adjThis concept is not fixed, but dynamically determined by the "typical behavioral profile" obtained from the clustering in step two. The essential difference lies in the "profile masking effect": first, clustering identifies which behavioral pattern the user belongs to (e.g., photovoltaic dependency), and this pattern limits the specific "adjustment period" (e.g., T for users of type A). adj This refers to periods with negative net load (Category B might be nighttime peak hours). η calculates the ratio of net load to total load "within a specific effective period." This means that data interference from "ineffective adjustment periods" has been eliminated.

[0091] The photovoltaic output time periods mentioned above can be understood as the default reference time period of 8:00-18:00 in cluster analysis, which is ultimately combined with the user's actual photovoltaic output sequence P. pv The temporal characteristics of (t) are used to clarify the actual power output periods in the electricity consumption behavior characteristic curve obtained by clustering; the peak load period is represented by the user load sequence P. Load Cluster analysis of (t) identifies the periods when load peaks are concentrated from the characteristic curves of electricity consumption behavior. For example, nighttime electricity consumption type (Category B) users are clustered and identified as having peak loads concentrated between 18:00 and 24:00, while flexible adjustable type (Category D) users are identified as having peak loads that overlap with the photovoltaic output period (8:00-18:00) for more than 4 hours.

[0092] The challenge in this embodiment lies in how to automatically define the integration interval (T) through unsupervised learning (clustering). adj Traditional methods require manual setting of peak / off-peak periods, while this method utilizes the curve shape identified by PSO-Kmeans to automatically lock the "key adjustment window" for each user type, achieving adaptive feature extraction.

[0093] The method in this embodiment can avoid misjudging "pseudo-potential" users: If only the total electricity ratio is considered, a user with a large daily power generation but also a high self-consumption rate might be misjudged as having "high regulation potential." Using the η index of this application, only on T... adj Only electricity generated during specific time periods (i.e., periods when the impact on the distribution area is felt or when support is available) is counted. This allows for the precise elimination of "false potential" users, making energy storage configuration and demand response strategies more accurate and reducing waste in distribution area investment.

[0094] Example 2 Based on the data from Example 1, such as Figure 7 As shown, this embodiment designs a deep learning intelligent classification model based on Attention-CNN-LSTM. In order to achieve rapid classification of a large number of newly connected users without having to re-cluster them every time, this embodiment constructs an end-to-end deep learning classification model.

[0095] 1. Detailed network architecture design: Input layer: Input dimension is This refers to 24-hour dual-channel data for both photovoltaic and load data.

[0096] 1D-CNN layer (feature extraction): Uses a one-dimensional convolutional kernel (Kernel Size=3) to extract local morphological features of the time series.

[0097]

[0098] in, For time step or index The input vector at that point, This is the weight matrix. For bias vectors, It is a linear rectified activation function. For this neural network layer at the input The output vector after that.

[0099] Attention layer (weighted by key features): such as Figure 9 As shown, an attention mechanism is introduced between CNN and LSTM. Attention scores at each time step are learned through fully connected layers. The weights are obtained using Softmax normalization. :

[0100] in, For bias vectors, This is the weight matrix. For time step or index The input vector at that point, The hyperbolic tangent activation function is used. This is the output vector of this neural network layer.

[0101]

[0102] in, and This is the original score. The activation formula for the exponential function is... The sum of the scores for all positions. For in position Normalized attention weights at the location.

[0103]

[0104] in, At the target time step The generated context vector, The total length of the input sequence. For in position Normalized attention weights at the location, For the input sequence at position The characteristics are represented.

[0105] This mechanism can automatically assign higher weights to moments of severe photovoltaic fluctuations or sudden load changes, allowing the model to "focus" on the feature segments most critical to classification.

[0106] LSTM layer (temporal memory): Receives the weighted feature vector and uses a forgetting gate. Input gate and output gate Handling long-distance dependencies:

[0107] in, For a moment The output vector of the forget gate, For the Sigmoid activation function, Here is the weight matrix for the forget gate. For the previous moment -1 is a hidden state. For the external input information at the current moment, This is the bias vector for the forget gate.

[0108] in, For a moment The output vector of the input gate, For the Sigmoid activation function, Here is the weight matrix of the input gate. For the previous moment -1 is a hidden state. For the external input information at the current moment, is the bias vector of the input gate.

[0109]

[0110] in, For a moment The output vector of the output gate. For the Sigmoid activation function, This is the weight matrix of the output gate. For the previous moment -1 is a hidden state. For the external input information at the current moment, This is the bias vector for the output gate.

[0111] Output layer: The fully connected layer maps high-dimensional features into 4-dimensional vectors, and outputs the probability distribution of the four potential levels through Softmax.

[0112] 2. Model Training and Hyperparameter Setting: Label generation: The clustering categories (0, 1, 2, 3) obtained in step two are used as the labels (Ground Truth) of the training data.

[0113] Loss function: Categorical Cross-Entropy Loss is used.

[0114] Optimizer: Use the Adam optimizer with an initial learning rate of 0.001.

[0115] Training strategy: Set the batch size to 64, the epochs to 100, and introduce a dropout layer (dropout rate 0.2) to prevent overfitting.

[0116] This embodiment designs a cascaded deep neural network structure: the CNN layer is responsible for extracting local morphological features (such as peaks, troughs, and rates of change) of the source load curve; the Attention mechanism is responsible for weighting key time steps and suppressing background noise interference; and the LSTM layer is responsible for mining long-term temporal dependencies. This model is trained using clustering results as "soft labels," achieving rapid and intelligent identification of the potential levels of massive numbers of newly connected users, thus solving the problem of low efficiency in manual analysis.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying the adjustable potential of photovoltaic power generation in a distribution area, characterized in that, include: The source-load characteristic data of users in the transformer area is obtained, and the net load is calculated and preprocessed. The net load represents the difference between the user's load and the user's photovoltaic output. The preprocessed source-load characteristic data is classified to obtain user categories and corresponding electricity consumption behavior characteristic curves. The electricity consumption behavior characteristic curves represent the time distribution characteristics of user load and photovoltaic output. The adjustable load power ratio of each type of user is calculated based on the electricity consumption behavior characteristic curves to construct a photovoltaic adjustable potential classification standard. The formula for calculating the adjustable load power ratio is as follows: In the formula, For adjustable load power ratio, Let be the net load at time t, when This indicates that the user is feeding power back to the transformer area; when When this time, it indicates that the user is drawing power from the transformer substation; The user's original load at time t, This represents the cumulative sum of the original user load at all times within the time interval [1, T]. This represents the total number of time periods. To identify the time windows during which users can achieve peak shaving / valley filling based on electricity consumption behavior characteristic curves, a set of time periods that users can adjust is defined.

2. The method for classifying the adjustable potential of photovoltaic distribution areas according to claim 1, characterized in that, The method for classifying the preprocessed source payload feature data is as follows: The first stage: use the Kmeans++ algorithm to calculate the initial cluster centers, that is, select sample points that are as far apart as possible from each other as the initial centers; The second stage involves mapping the initial cluster centers to particle positions in the PSO algorithm, using the reciprocal of the sum of squared errors within each cluster as the fitness function, and finding the globally optimal cluster centers through iterative updates of particle positions and velocities. The third stage: The global optimal solution output by the PSO algorithm is used as the initial center of the Kmeans algorithm, and local iterative convergence is performed until the stopping condition is met, and the final clustering result is output.

3. A method for classifying the adjustable potential of photovoltaic distribution areas according to claim 1 or 2, characterized in that, The user categories include: Type A: The user's net load is continuously negative during the midday period, and the proportion of photovoltaic output is greater than the first set value; Type B: Peak user load is concentrated at night, and the misalignment rate with the photovoltaic output period is greater than the second set value; the misalignment rate is the proportion of electricity consumption during peak load periods that do not overlap with photovoltaic output periods to the total peak electricity consumption, reflecting the degree of time misalignment between electricity consumption and photovoltaic output. Type C: The load curve is stable throughout the day, and the peak-to-valley difference coefficient is less than the third set value; Type D: The load peak-valley difference is large and overlaps with the photovoltaic output period. The load peak-valley difference coefficient is greater than the fourth set value, and the overlap period between the peak electricity consumption period and the photovoltaic output period is greater than the set duration. The electricity price sensitivity coefficient is greater than the fifth set value. The peak load period and photovoltaic output period are identified based on the characteristic curves of user electricity consumption behavior.

4. The method for classifying the adjustable potential of photovoltaic distribution areas according to claim 3, characterized in that, The photovoltaic output time misalignment rate Defined as: in, Let be the predicted photovoltaic power output at time t. For the first The actual photovoltaic output power at any given time The time offset represents the time deviation between the prediction and the actual output; the predicted photovoltaic power output refers to the typical daily power curve extracted based on the user's historical electricity consumption data and the photovoltaic historical power output data, which respectively represent the time distribution characteristics of user electricity consumption behavior and photovoltaic power output.

5. The method for classifying the adjustable potential of photovoltaic distribution areas according to claim 3, characterized in that, The formula for calculating the peak-valley difference coefficient is as follows: in, The highest load value within the statistical period. This represents the lowest load value within the statistical period. This represents the peak-to-valley difference coefficient.

6. The method for classifying the adjustable potential of photovoltaic distribution areas according to claim 3, characterized in that, The electricity price sensitivity coefficient The calculation formula is: in, This refers to the electricity consumption for the current month and the previous month. This refers to the electricity price for the current month and the previous month.

7. A method for classifying the adjustable potential of photovoltaic distribution areas according to claim 1 or 2, characterized in that, The photovoltaic adjustable potential grading standard is as follows: Level 1 potential: The corresponding strategy is to configure distributed energy storage; Secondary potential: The corresponding strategy is to implement time-of-use pricing to guide electricity prices. Level 3 potential: The corresponding strategy is to participate in peak shaving and valley filling auxiliary services; Level 4 potential: The corresponding strategy is not to use it as a primary adjustment resource.

8. A tiered system for the adjustable potential of photovoltaic distribution areas, characterized in that, include: Data acquisition module: used to acquire source load characteristic data of users in the transformer area, and to perform net load calculation and preprocessing on the source load characteristic data; Clustering analysis module: Classifies the preprocessed source-load characteristic data to obtain the electricity consumption behavior characteristic curves of different types of users; Standard construction module: Calculates the adjustable load power ratio of various user types based on electricity consumption behavior characteristic curves, and constructs a photovoltaic adjustable potential grading standard; the formula for calculating the adjustable load power ratio is as follows: In the formula, For adjustable load power ratio, Let be the net load at time t, when This indicates that the user is feeding power back to the transformer area; when When this time, it indicates that the user is drawing power from the transformer substation; This represents the cumulative sum of the original user load at all times within the time interval [1, T], where T is the total number of time intervals. To identify the time windows during which users can achieve peak shaving / valley filling based on electricity consumption behavior characteristic curves, a set of time periods that users can adjust is defined.

9. A photovoltaic adjustable potential grading system for a distribution area according to claim 8, characterized in that, The method for classifying the preprocessed source payload feature data is as follows: The first stage: use the Kmeans++ algorithm to calculate the initial cluster centers, that is, select sample points that are as far apart as possible from each other as the initial centers; The second stage involves mapping the initial cluster centers to particle positions in the PSO algorithm, using the reciprocal of the sum of squared errors within each cluster as the fitness function, and finding the globally optimal cluster centers through iterative updates of particle positions and velocities. The third stage: The global optimal solution output by the PSO algorithm is used as the initial center of the Kmeans algorithm, and local iterative convergence is performed until the stopping condition is met, and the final clustering result is output.

10. A tiered system for adjustable photovoltaic potential in a distribution area according to claim 8 or 9, characterized in that, The different types of users include: Type A: The user's net load is continuously negative during the midday period, and the proportion of photovoltaic power generation is greater than the first set value; Type B: Peak user load is concentrated at night, and the misalignment rate with the photovoltaic output period is greater than the second set value; the misalignment rate is calculated as the complement of the proportion of the peak electricity consumption period and the photovoltaic output period. Type C: The load curve is stable throughout the day, and the peak-to-valley difference coefficient is less than the third set value; Type D: The load peak-valley difference is large and overlaps with the photovoltaic period. The load peak-valley difference coefficient is greater than the fourth set value, and the overlap time between the peak electricity consumption period and the photovoltaic output period is greater than the set duration. The electricity price sensitivity coefficient is greater than the fifth set value.

11. A photovoltaic adjustable potential grading system according to claim 10, characterized in that, The photovoltaic output time misalignment rate Defined as: in, Let be the predicted photovoltaic power output at time t. For the first The actual photovoltaic output power at any given time This is the time offset, representing the time difference between the prediction and the actual time.

12. A photovoltaic adjustable potential grading system for a distribution area according to claim 10, characterized in that, The formula for calculating the peak-valley difference coefficient is as follows: in, The highest load value within the statistical period. This represents the lowest load value within the statistical period. This represents the peak-to-valley difference coefficient.

13. A photovoltaic adjustable potential grading system for a distribution area according to claim 10, characterized in that, The electricity price sensitivity coefficient The calculation formula is: in, This refers to the electricity consumption for the current month and the previous month. This refers to the electricity price for the current month and the previous month.

14. A tiered system for adjustable photovoltaic potential in a distribution area according to claim 8 or 9, characterized in that, The photovoltaic adjustable potential grading standard is as follows: Level 1 potential: The corresponding strategy is to configure distributed energy storage; Secondary potential: The corresponding strategy is to implement time-of-use pricing to guide electricity prices. Level 3 potential: The corresponding strategy is to participate in peak shaving and valley filling auxiliary services; Level 4 potential: The corresponding strategy is not to use it as a primary adjustment resource.

15. A method for constructing an intelligent hierarchical model of the adjustable potential of photovoltaic power distribution areas, characterized in that, A CNN-LSTM deep learning model with added attention mechanism is used; the user type obtained by any one of the methods described in claims 1 to 7 is used as a label, and the source payload feature data of the user is used as input to train the CNN-LSTM deep learning model to obtain an intelligent classification model.

16. An application of an intelligent hierarchical model, characterized in that, The intelligent grading model trained using the method described in claim 15 is used to grade the adjustable potential of newly connected photovoltaic users in the distribution area.

Images