Distribution area distributed photovoltaic power identification method and system

Abstract

The invention relates to a transformer area distributed photovoltaic power identification method, and belongs to the technical field of new energy power generation, and the method comprises the steps: carrying out the missing value filling and abnormal value replacement of photovoltaic power station data, and guaranteeing the data integrity; different types of loads in a day are simulated, and are expanded into annual load data according to seasons; and selecting photovoltaic power station data as a reference power station and composing transformer area data. The fuzzy membership function of the total power and the power fluctuation degree is established, the membership degree of each weather state is calculated, and weather classification judgment is carried out according to the maximum membership principle. And establishing a mapping model of the photovoltaic variation of the reference power station and the photovoltaic variation of the solar radiation to the transformer area. The improved convolutional neural network model is used for realizing high-precision learning of a nonlinear mapping relation. According to the method, limited data of a photovoltaic power station can be efficiently utilized, and photovoltaic power fluctuation characteristics are fully considered, so that the identification accuracy and reliability are improved, and better support is provided for operation and management of a photovoltaic power generation system.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation technology, and in particular to a method and system for identifying the power of distributed photovoltaic power in a transformer substation. Background Technology

[0002] Guided by the goals of "carbon peaking and carbon neutrality," my country's energy system is undergoing a profound transformation from being dominated by traditional fossil fuels to being dominated by new energy sources. As a key support for the construction of a new power system, the development and utilization of new energy sources, represented by distributed photovoltaic (PV) power, has moved from "pilot exploration" to "comprehensive promotion." In July 2024, the National Development and Reform Commission, the National Energy Administration, and other three departments jointly released the "Action Plan for Accelerating the Construction of a New Power System (2024-2027)," which clarified the development direction of the new power system and required accelerating its construction to provide strong support for achieving the carbon peaking goal. However, with the "numerous and widespread, small and densely distributed" access characteristics of distributed PV, its deep integration with the distribution network is bringing a series of urgent technical challenges, especially the accurate identification of high-proportion distributed PV capacity, which has become a key bottleneck restricting the safe and efficient operation of the new power system.

[0003] The invention patent with publication number CN117669960A discloses a new energy power prediction method based on multivariate meteorological factors. However, photovoltaic power sources distributed in low-voltage distribution areas are subject to limitations such as technology and investment costs, making it relatively difficult to obtain their output data individually (usually only hourly power data is uploaded). Furthermore, photovoltaic power sources are located at dispersed users, with significant differences in models and characteristics, making it difficult to install meteorological and irradiance acquisition equipment for each household. This hinders the investigation, statistics, and aggregation to obtain a mapping model reflecting the entire distribution area. Therefore, classic photovoltaic power prediction methods are not applicable to low-voltage distribution areas. If the photovoltaic power generation portion is identified (separated) from the output power measurement data sequence of a distribution area using technical means, the sum of distributed photovoltaic output within the area can be obtained at a lower cost. This data also provides substantial assistance for the operation of the distribution network, demonstrating high economic efficiency and practicality. Therefore, developing a power identification method that adapts to the characteristics of distributed photovoltaic power generation is of significant research value for improving the development and utilization efficiency of photovoltaic power generation and maintaining the stable operation of the distribution network.

[0004] With a high proportion of small and distributed photovoltaic (PV) systems being integrated into regional distribution networks, traditional non-intrusive identification techniques are no longer sufficient to identify the overall PV power of the regional distribution network. Non-intrusive identification techniques rely on the detection of changes in current and voltage waveforms or frequencies, but for complex, active distribution networks, this technology struggles to accurately identify the power generation of all distributed PV systems within the area.

[0005] Photovoltaic power identification can generally be divided into traditional statistical methods and modern intelligent methods. Traditional statistical methods, such as time series analysis, regression analysis, and Kalman filtering, can effectively predict future photovoltaic power generation by analyzing the statistical relationships between historical data. However, these methods are based on linear analysis and require high stability of the original data. When the time series is highly nonlinear, they cannot obtain satisfactory identification results. In addition, with the increasing randomness and diversity of photovoltaic system access, traditional technologies are struggling to meet the demands in terms of data processing and computing power. To handle the nonlinear characteristics of photovoltaic power time series, modern intelligent methods, such as convolutional neural networks (CNNs), support vector regression, and extreme learning machines (ELM), which can fit the nonlinear relationship between input and output variables, are widely used. However, traditional CNNs mainly extract features by fixing the convolutional kernels and lack the ability to model the uncertainty and fuzziness of the input data, making it difficult to accurately characterize the fuzzy variation patterns of photovoltaic power generation affected by multiple factors such as weather and radiation intensity. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for identifying distributed photovoltaic power in a transformer substation, thereby solving the problem of accurate identification of high-proportion distributed photovoltaic capacity in a transformer substation. This invention establishes an uncertainty expression model of weather conditions by defining membership functions for total power and fluctuation, and achieves fuzzy classification based on the maximum membership principle, thus providing physically meaningful fuzzy prior information for the neural network. This fusion approach enables the convolutional network to have adaptive adjustment capabilities during feature extraction, dynamically perceiving differences in input features under different weather conditions, improving the model's identification accuracy and robustness in nonlinear fluctuation scenarios, and achieving an organic combination of data-driven and knowledge-driven approaches.

[0007] A first aspect of the present invention provides a method for identifying the power output of distributed photovoltaic (PV) systems in a transformer substation, the method comprising: Missing data and anomaly replacements are performed on the active power data of the photovoltaic grid-connected switch to form a sampling dataset. Based on the sampling dataset, photovoltaic output curves under four typical weather conditions are constructed to determine the benchmark power station and obtain a sample library of photovoltaic power in the distribution area. The four typical weather conditions are sunny, cloudy, overcast, and rainy / snowy. The photovoltaic processing curve is used to collect the power generation sequence of the photovoltaic array within a preset time window. Based on the power generation sequence, a fuzzy membership model corresponding to four typical weather conditions is constructed to achieve adaptive identification of weather conditions. The fuzzy membership model is obtained through the total power generation and power fluctuation rate. The fuzzy logic weather membership degree obtained by the fuzzy membership model is used to correct direct radiation and scattered radiation, so as to realize adaptive solar radiation estimation under different weather conditions; A photovoltaic change model for a distribution area is constructed based on the photovoltaic change of a benchmark power station and the current solar radiation. The results of the photovoltaic change model are then input into a feature extraction model. After training the feature extraction model, a feature vector is output, which corresponds to the photovoltaic power identification results of the distribution area at different times. The loss function used in the feature extraction model is obtained by weighting the uncertainty of the weather membership degree based on fuzzy logic. The trained feature extraction model is evaluated using an evaluation function, and its current accuracy and robustness are verified in different scenarios. The model is then optimized based on error analysis, and data corrections are added for abnormal scenarios.

[0008] Furthermore, including: The sample library for determining the benchmark power station and obtaining the photovoltaic power of the distribution area includes: Multiple distributed photovoltaic power stations with complementary power output characteristics were selected in a certain area, and their grid-connected switch active power data were continuously collected. Plot the actual photovoltaic output curves under four types of typical weather conditions, and select power plants with strong output stability and high matching degree with regional load characteristics under the four types of weather conditions as benchmark power plants based on the characteristics of the actual photovoltaic output curves under each type of typical weather condition. Combine power plant data with complementary power output characteristics to construct a sample library of distribution areas.

[0009] Furthermore, including: The construction of fuzzy membership models corresponding to four typical weather conditions based on the power generation sequence includes: The standard deviation of the rate of change of power generation per unit time is obtained from the power generation sequence and used as the power fluctuation. The power fluctuation and total power generation are used as independent variables to represent the fuzzy membership model. The dependent variable of this model is the fuzzy membership function corresponding to four typical weather conditions determined based on historical power generation data.

[0010] Furthermore, including: The adaptive identification process for the weather conditions includes: Based on current data of different typical weather conditions, the fuzzy logic weather membership vector of each weather category is obtained by using the corresponding fuzzy membership function, and the current weather state is determined according to the principle of maximum membership. The maximum membership principle is implemented using a rule-based fuzzy controller structure, wherein the rule base includes: The weather conditions corresponding to high total power generation and low volatility are sunny, cloudy, overcast, and rain / snow. At the end of each day, the baseline parameters of each fuzzy function are updated based on the statistical results of the total power generation and fluctuation over the past few days to achieve adaptive correction for seasonal changes. The baseline parameters include a center parameter and a width parameter.

[0011] Furthermore, including: The method employs a fuzzy logic weather membership degree correction model to adjust direct and scattered radiation, enabling adaptive solar radiation estimation under different weather conditions, including: The dynamic atmospheric transparency coefficient and dynamic scattering correction coefficient are established by using the fuzzy logic weather membership degree under different weather conditions; Direct radiation and scattered radiation are represented based on the dynamic atmospheric transparency coefficient and dynamic scattering correction coefficient; The total solar radiation is calculated using the direct radiation and the diffuse radiation. The total solar radiation is expressed as the sum of the direct radiation and the diffuse radiation, multiplied by a cloud cover correction factor. The cloud cover correction factor is obtained by weighting empirical benchmark values ​​under various weather conditions with fuzzy logic weather membership degrees.

[0012] Furthermore, including: A photovoltaic variation model for the distribution area is constructed based on the photovoltaic variation of a benchmark power station and the current solar radiation. The results of the photovoltaic variation model are then input into a feature extraction model, including: Construct a feature mapping relationship between the photovoltaic change characteristics, total solar radiation characteristics, fuzzy logic weather membership characteristics, solar altitude angle characteristics, and solar azimuth angle characteristics of the benchmark power station and the photovoltaic change of the substation area; A multi-channel structure is adopted to input features from different physical sources into the feature extraction model in parallel. The feature extraction model includes two one-dimensional convolutional layers, two max pooling layers, and a fully connected layer. Both one-dimensional convolutional layers adopt a multi-channel convolutional structure and are used to extract the temporal correlation between power change features and solar angle features, respectively.

[0013] Furthermore, including: The convolutional kernel weights of the convolutional layer are adaptively adjusted based on the prediction residuals between the baseline power plant power and the power distribution area power during the training process. If the prediction residuals between the baseline power plant power and the power distribution area power exceed a set threshold, the convolutional kernel weights of the convolutional layer are updated, as shown below: in, For learning rate, Represents a symbolic function. These are the current kernel weights of the convolutional layer. The predicted residuals are the power output of the benchmark power plant and the power output of the distribution area. Further, this includes: the loss function used in the feature extraction model is obtained by weighting the uncertainty of fuzzy logic weather membership degrees; The loss function is defined as: ;in, express t Weather clarity at any given moment, for t The change in photovoltaic power generation in the transformer substation at any given time. This represents the photovoltaic fluctuation predicted by the model.

[0014] On the other hand, the present invention also provides a distributed photovoltaic power identification system for a transformer substation, the system comprising: The data preparation module is used to fill in missing data and replace anomalies in the active power data of the photovoltaic grid-connected switch to form a sampling dataset. Based on the sampling dataset, the photovoltaic output curves under four typical weather conditions are constructed to determine the benchmark power station and obtain a sample library of photovoltaic power in the distribution area. The four typical weather conditions are sunny, cloudy, overcast, and rain / snow. The photovoltaic power identification model building module is used to collect the power generation sequence of the photovoltaic array within a preset time window through the photovoltaic processing curve, and construct fuzzy membership models corresponding to four typical weather conditions based on the power generation sequence to achieve adaptive identification of weather conditions. The fuzzy membership model is obtained through the total power generation and power fluctuation rate. The solar radiation calculation module is used to correct direct and scattered radiation using fuzzy logic weather membership degrees obtained from fuzzy membership models, thereby achieving adaptive solar radiation estimation under different weather conditions. The model training module is used to construct a photovoltaic change model for the substation area based on the photovoltaic change of the benchmark power station and the current solar radiation. The results of the photovoltaic change model are input into the feature extraction model. After training the feature extraction model, the feature vector is output to identify the photovoltaic power of the substation area at different times. The loss function used by the feature extraction model is obtained by weighting the uncertainty of the fuzzy logic weather membership degree. The model evaluation and optimization module evaluates the trained feature extraction model through an evaluation function, verifies the current accuracy and robustness in different scenarios, optimizes the model based on error analysis, and adds data correction for abnormal scenarios.

[0015] Advantages of this invention: This invention achieves dynamic classification of weather conditions by establishing a fuzzy membership function for total power and fluctuation, and adaptively corrects radiation estimation parameters by combining solar declination, hour angle, altitude angle, and azimuth angle. It establishes a photovoltaic mapping relationship using the complementary output characteristics of a benchmark power station, and the CNN neural network, guided by fuzzy logic, achieves multi-scale feature adaptive extraction, separating the photovoltaic component from the total power of complex distribution areas with high precision. For seasonal changes, date types, and abnormal weather scenarios, the load model parameters and radiation calculation coefficients are adjusted to give the system strong environmental adaptability. Simultaneously, through modular design integrating special case fusion and data correction strategies, it maintains stable output even under challenging scenarios such as grid load fluctuations and extreme weather, effectively addressing the shortcomings of insufficient robustness in existing technologies. The resulting closed-loop technology of "fuzzy recognition-radiation estimation-mapping learning-dynamic optimization" provides a reliable solution for intelligent management of distribution areas with high proportions of distributed photovoltaic access. Attached Figure Description

[0016] Figure 1 This is a flowchart of the distributed photovoltaic power identification method for transformer substations according to an embodiment of the present invention; Figure 2 This is a graph illustrating the error analysis between the photovoltaic power identification and actual power based on CNN, as described in an embodiment of the present invention. Figure 3 This is a graph illustrating the error analysis between the photovoltaic identified power and the actual power based on LSTM, as described in an embodiment of the present invention. Figure 4 This is a schematic diagram of the module composition of the distributed photovoltaic power identification system in the transformer substation according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: This example provides a method for identifying the power output of distributed photovoltaic (PV) systems in a transformer substation, including the following steps: S1 performs missing data filling and anomaly replacement on the active power data of the photovoltaic grid-connected switch to form a sampling dataset. Based on the sampling dataset, photovoltaic output curves under four typical weather conditions are constructed to determine the benchmark power station and obtain a sample library of photovoltaic power in the distribution area. The four typical weather conditions are sunny, cloudy, overcast, and rain / snow. S2 collects the power generation sequence of the photovoltaic array within a preset time window through the photovoltaic processing curve, and constructs fuzzy membership models corresponding to four typical weather conditions based on the power generation sequence to achieve adaptive identification of weather conditions. The fuzzy membership model is obtained through the total power generation and power fluctuation rate. S3 uses the fuzzy logic weather membership degree obtained by the fuzzy membership model to correct direct radiation and scattered radiation, thereby realizing adaptive solar radiation estimation under different weather conditions; S4 constructs a photovoltaic change model for the substation area based on the photovoltaic change of the benchmark power station and the current solar radiation, and inputs the results of the photovoltaic change model for the substation area into the feature extraction model. After training the feature extraction model, it outputs a feature vector, thereby identifying the photovoltaic power of the substation area at different times. The loss function used by the feature extraction model is obtained by weighting the uncertainty of the fuzzy logic weather membership degree. S5 evaluates the trained feature extraction model through an evaluation function, verifies the current accuracy and robustness in different scenarios, optimizes the model based on error analysis, and adds data correction for abnormal scenarios.

[0019] Furthermore, when performing missing value filling and outlier replacement on the active power sampling data of the grid-connected switch of the photovoltaic power station, the following steps are included: Detect outlier indexes and perform replacements based on the presence of preceding and following data: If both preceding and following data exist, replace them with the average value. Replace with the previous data only if the previous data exists; Replace with the later data only if the later data already exists; When both preceding and following data are missing, the missing values ​​are retained and filled using linear interpolation.

[0020] Furthermore, the construction of user load data for the distribution area includes: The load curve of residential users shows a double-peak and one-valley shape, with morning peak, noon peak, evening peak and nighttime trough, among which the evening peak load level is the highest. The load curve of heavy industrial users remained stable throughout the day with narrow fluctuations, only showing a slight decrease during shift changes. The load curve of light industrial users exhibits a single-peak, narrow-range shape: it forms a peak load plateau during continuous periods of daytime, and then drops to standby level at night; Seasonal adjustments: Residential users experience increased evening peak load in summer and increased morning peak load in winter; heavy industrial users experience an overall upward shift in load level in winter; the duration of peak load for light industrial users varies with the seasons. Date type adjustment: During holidays, the evening peak time for residential users is delayed and the peak value is further increased; heavy industrial users only maintain the basic load level; light industrial users' load further decreases to the valley value.

[0021] In this embodiment, determining the benchmark power station and obtaining the sample library of photovoltaic power in the distribution area includes: Multiple distributed photovoltaic power stations with complementary power output characteristics were selected in a certain area, and their grid-connected switch active power data were continuously collected. Plot the actual photovoltaic output curves under four types of typical weather conditions, and select power plants with strong output stability and high matching degree with regional load characteristics under the four types of weather conditions as benchmark power plants based on the characteristics of the actual photovoltaic output curves under each type of typical weather condition. Combine power plant data with complementary power output characteristics to construct a sample library of distribution areas.

[0022] In a preferred embodiment of this invention, the process of selecting benchmark power plants and constructing transformer area data includes: Multiple distributed photovoltaic power stations with complementary power output characteristics were selected, and their grid-connected switch active power data were continuously collected. Plot the actual photovoltaic power output curves under four typical weather conditions: Sunny day curve: It shows a significant single-peak shape, with its peak power close to the installed capacity, and a plateau period of output lasting for several hours at noon; Cloudy curve: It exhibits a wave-like undulation pattern, with its peak power level significantly lower than that under sunny conditions, and it maintains positive power output throughout the day; Cloudy day curve: The power output curve tends to flatten out, and its maximum daily power output is much lower than the installed capacity; Rain and snow curve: The output curve is basically close to the zero line, with only intermittent slight fluctuations; Selection criteria: Power plants with strong output stability under the four types of weather conditions and high matching degree with regional load characteristics are selected as benchmark power plants; Construction of a sample library for photovoltaic power distribution areas: combining power plant data with complementary power output characteristics for training of photovoltaic power identification models.

[0023] In this embodiment, fuzzy membership models corresponding to four typical weather conditions are constructed based on the power generation sequence, including: The standard deviation of the rate of change of power generation per unit time is obtained from the power generation sequence and used as the power fluctuation. The power fluctuation and total power generation are used as independent variables to represent the fuzzy membership model. The dependent variable of this model is the fuzzy membership function corresponding to four typical weather conditions determined based on historical power generation data.

[0024] In a preferred embodiment, a fuzzy membership function is established for the total power and power fluctuation, the membership degree of each weather state is calculated, and weather classification is performed based on the principle of maximum membership. Specifically, the following steps are included: Feature extraction: Collect the power generation sequence of the photovoltaic array within a preset time window. Calculate the total power generation and power fluctuation They are used as input features for weather condition recognition; Fuzzy set definition: Based on the characteristics of photovoltaic power, fuzzy sets corresponding to sunny, cloudy, overcast, and rainy / snowy weather are established, and membership functions for total power and fluctuation are defined, where the membership function for each weather state is... Represented as: (1) The The fuzzy membership function is determined based on historical power generation data; Fuzzy reasoning and decision-making: Substituting the input features into each membership function, the membership vector of each weather category is calculated. Weather conditions are determined based on the principle of maximum membership: (2) Dynamic parameter update: The baseline parameters of each fuzzy function are updated statistically based on historical periodic data, so that the system can maintain adaptive recognition capability under different seasons, regions and climate conditions.

[0025] The fuzzy membership function is represented by a Gaussian function: (3) in, For the first The central parameter for weather-like conditions, This is the width parameter.

[0026] In this embodiment, the adaptive weather condition recognition process includes: Based on current data of different typical weather conditions, the fuzzy logic weather membership vector of each weather category is obtained by using the corresponding fuzzy membership function, and the current weather state is determined according to the principle of maximum membership. The maximum membership principle is implemented using a rule-based fuzzy controller structure, wherein the rule base includes: The weather conditions corresponding to high total power generation and low volatility are sunny, cloudy, overcast, and rain / snow. At the end of each day, the baseline parameters of each fuzzy function are updated based on the statistical results of the total power generation and fluctuation over the past few days to achieve adaptive correction for seasonal changes. The baseline parameters include a center parameter and a width parameter.

[0027] In a preferred embodiment, power fluctuation Defined as the standard deviation of the rate of change of power generation per unit time: (4) in, This represents the average power.

[0028] The dynamic parameter update step includes: at the end of each day, updating the central parameters of each fuzzy function based on the statistical results of the total power generation and volatility over the past T days. and width parameter This allows for adaptive correction to seasonal changes.

[0029] Furthermore, the fuzzy reasoning and decision-making adopts a rule-based fuzzy controller structure, and the rule base takes the form of: IF (high power generation) AND (low volatility) THEN (sunny day) IF (High power generation) AND (High volatility) THEN (Partly cloudy) IF (Medium power generation capacity) THEN (Cloudy) IF (low power generation) THEN (rain / snow) Furthermore, when calculating solar angle and solar radiation, this invention does not use the traditional model with fixed parameters, but combines the fuzzy logic weather identification results to dynamically correct the solar radiation estimation parameters, thereby achieving adaptive radiation calculation for different weather and solar angle conditions.

[0030] The solar radiation estimation includes the following steps: When calculating solar angles (solar altitude and azimuth) based on latitude, longitude, and date, the following is included: Calculating true solar time LAST : (5) in, Use the local time zone. The longitude of the central time zone; Local longitude; This is the time difference correction, which is the difference between true solar time and mean solar time. This value is caused by the eccentricity of the Earth's orbit and the obliquity of the ecliptic, and can be obtained through formulas or by looking up tables.

[0031] Calculating solar declination Declination is the latitude of the point where the sun is directly overhead, and it varies with the seasons. (6) It represents the number of days in a year (January 1st is 1, December 31st is 366).

[0032] Calculate the angle The hour angle represents the sun's offset relative to local noon: (7) Calculate the solar altitude angle The angle between the sunlight and the horizon ranges from... A negative value indicates that the sun is below the horizon, which can be determined using the spherical trigonometry formula: (8) in, The latitude is the local latitude. Calculate the solar azimuth angle Starting from due north, turn clockwise to the angle formed by the projection of the sun's rays onto the horizon, ranging from... (9) because The absolute value cannot exceed 1. It may need to be adjusted due to calculation errors or polar day / polar night conditions. In practice, the direction can be determined by the quadrant (eastward in the morning, westward in the afternoon).

[0033] Furthermore, using the solar constant Atmospheric transparency coefficient With scattering correction coefficient Calculate direct radiation With scattered radiation The preliminary calculation formulas are as follows: in, For atmospheric mass number, when To avoid the denominator being zero.

[0034] In this embodiment, the fuzzy logic weather membership degree obtained by the fuzzy membership model is used to correct direct radiation and scattered radiation, thereby realizing adaptive solar radiation estimation under different weather conditions, including: The dynamic atmospheric transparency coefficient and dynamic scattering correction coefficient are established by using the fuzzy logic weather membership degree under different weather conditions; Direct radiation and scattered radiation are represented based on the dynamic atmospheric transparency coefficient and dynamic scattering correction coefficient; The total solar radiation is calculated using the direct radiation and the diffuse radiation. The total solar radiation is expressed as the sum of the direct radiation and the diffuse radiation, multiplied by a cloud cover correction factor. The cloud cover correction factor is obtained by weighting empirical benchmark values ​​under various weather conditions with fuzzy logic weather membership degrees.

[0035] In a preferred embodiment, to achieve continuous correction under different weather conditions, the present invention uses the weather membership degree output by the fuzzy logic recognition module. (Including sunny, cloudy, overcast, rainy, snowy, etc.) By introducing a weighted summation of radiation calculation parameters, the coefficients are no longer fixed, but dynamically determined by the fuzzy membership degree of the weather state: in, These are the empirical baseline values ​​for each weather condition.

[0036] This weighted approach allows the model parameters to transition continuously as the weather changes between sunny and cloudy conditions, avoiding abrupt changes caused by traditional threshold models.

[0037] Calculate the total solar radiation based on the corrected parameters: (13) in, The cloud cover correction coefficient is dynamically adjusted by the fuzzy weights.

[0038] Through the above steps, the present invention realizes a continuous adaptive coupling mechanism from weather identification to radiation estimation, which can adjust the calculation parameters in real time according to different climate conditions, time and solar altitude angle, effectively improving the accuracy of photovoltaic radiation prediction and power generation estimation.

[0039] Furthermore, establish a benchmark power station photovoltaic change. With solar radiation Changes in photovoltaic power in the Taiwan region Mapping relationship: (14) The model actually learns the mapping relationship between the photovoltaic fluctuations of the benchmark power station and the photovoltaic fluctuations of the distribution area. Therefore, when actually separating the photovoltaic power of the distribution area, the photovoltaic power in the distribution area can be obtained from the photovoltaic power of the benchmark power station through this mapping relationship.

[0040] In this embodiment, a photovoltaic variation model for the distribution area is constructed based on the photovoltaic variation of a benchmark power station and the current solar radiation. The results of the photovoltaic variation model are then input into a feature extraction model, including: Construct a feature mapping relationship between the photovoltaic change characteristics, total solar radiation characteristics, fuzzy logic weather membership characteristics, solar altitude angle characteristics, and solar azimuth angle characteristics of the benchmark power station and the photovoltaic change of the substation area; A multi-channel structure is adopted to input features from different physical sources into the feature extraction model in parallel. The feature extraction model includes two one-dimensional convolutional layers, two max pooling layers, and a fully connected layer. Both one-dimensional convolutional layers adopt a multi-channel convolutional structure and are used to extract the temporal correlation between power change features and solar angle features, respectively.

[0041] Furthermore, in a preferred embodiment, to improve the accuracy of photovoltaic power separation in a distribution area, the present invention establishes a reference power station photovoltaic variation... With solar radiation Changes in photovoltaic power in the Taiwan region To address the mapping relationship, an improved convolutional neural network (Fuzzy-Aware CNN) model integrating fuzzy logic and multi-source features is proposed for high-precision learning of nonlinear mapping relationships. Its core steps include: feature fusion input: the model input not only includes historical sequences... It also includes solar radiation at the same time. Solar altitude angle Azimuth and fuzzy logic weather membership vector The input feature combination is: (15) Through this multi-channel input, the network can simultaneously capture the temporal characteristics of photovoltaic changes and weather semantic features.

[0042] Improved Convolutional Structure: A CNN framework consisting of two convolutional layers, two pooling layers, and a fully connected layer is constructed. The convolutional layers employ a multi-channel parallel structure to extract the correlation fluctuation pattern between baseline power and meteorological features. The output of each convolutional layer is as follows: The pooling layer uses max pooling to preserve mutation features.

[0043] To address the differences in power fluctuations under different weather ambiguities, an uncertainty-weighted loss function is adopted: in, It represents the clarity of the weather, with higher weight for sunny days and lower weight for cloudy or blurry days, in order to avoid overfitting to outlier samples.

[0044] During model training, based on the power residuals between the benchmark power station and the distribution area... Perform dynamic weight adjustment: To improve the model's adaptability and real-time correction capabilities.

[0045] Finally, the feature vector H_C output by the CNN layer, after being activated by the Sigmoid function, is mapped to the photovoltaic power components of the distribution area. (20) in, The output feature vector corresponds to the photovoltaic power identification results of the transformer area at different times.

[0046] Furthermore, through mean squared error (MSE) and coefficient of determination To evaluate model performance, we divide the model into weather, power grid load, seasonal and daily variation scenarios to verify its identification accuracy and robustness. Based on the error analysis results, we optimize the model and add data correction or special case fusion strategies for abnormal scenarios.

[0047] To more clearly and completely explain this method, the following detailed description is provided: Figure 1 As shown, this method includes the following steps: Step 1: Fill in missing data and replace anomalies in the active power data of the photovoltaic grid-connected switch, simulate daily load and extend seasonally to annual data, and select the photovoltaic power station as the benchmark and data source for the distribution area.

[0048] In this embodiment, the team selects four representative photovoltaic power plants in a certain area, continuously collects and analyzes their inverter output power data, including power generation data recorded every 5 minutes for 11 months in 2024 and corresponding timestamp data, and plots a photovoltaic power output curve that reflects the actual operating characteristics, discarding power plant data with poor output data quality.

[0049] Furthermore, when performing missing value filling and outlier replacement on the active power sampling data of the grid-connected switch of the photovoltaic power station, the following steps are included: Detect outlier indexes and perform replacements based on the existence of data before and after them: If both preceding and following data exist, replace them with the average value. Replace with the previous data only if the previous data exists; Replace with the later data only if the later data already exists; When both preceding and following data are missing, the missing values ​​are retained and filled using linear interpolation.

[0050] Furthermore, the construction of user load data for the distribution area includes: The load curve of residential users shows a double-peak and one-valley pattern: morning peak (6-8 am, 120% of the daily average), noon peak (12-1 pm, 115%), evening peak (6-10 pm, 150%), and valley (0-5 pm, 60%). The load curve of heavy industrial users fluctuates by less than 10%, drops to 95% during shift change periods, and maintains a daily average of 90%-105% throughout the day. The load curve of light industrial users shows a single-peak, narrow-range shape: peak plateau (10-15 o'clock, 130%), and nighttime standby (22-6 o'clock, 30%). Seasonal adjustments: Residential users' summer evening peak load increases by 20%, and winter morning peak load increases by 15%; heavy industry's winter load increases by 5%; peak duration for light industry is adjusted by ±1 hour. Date type adjustment: During holidays, the evening peak for residents will be postponed to 8-11 pm and the peak will increase by 10%; heavy industry will maintain 80% base load; light industry will decrease to 20% trough.

[0051] Furthermore, the selection of benchmark power plants and the construction of transformer area data include: Multiple distributed photovoltaic power stations with complementary power output characteristics were selected in a certain area, and their grid-connected switch active power data were continuously collected. Plot the actual photovoltaic power output curves under four typical weather conditions: Sunny day curve: It shows a single peak shape, with peak power reaching 85%-90% of the installed capacity, and the midday plateau lasting for 4 hours; Cloudy curve: It fluctuates in a wave-like manner, with peak power at 60%-70% of that on sunny days, and maintains positive output overall; Cloudy day curve: tends to flatten out, with maximum output for the whole day less than 20% of installed capacity; Rain and snow curve: basically close to the zero value line, with only intermittent slight fluctuations; Selection criteria: Power plants with strong output stability under the four types of weather conditions and high matching degree with regional load characteristics are selected as benchmark power plants; Construction of a sample library for photovoltaic power distribution areas: combining power plant data with complementary power output characteristics for training of photovoltaic power identification models.

[0052] Step 2: Establish fuzzy membership functions for total power and fluctuation to determine weather conditions; combine solar angle calculations to dynamically correct radiation parameters, achieving adaptive radiation estimation under different weather conditions, including: Feature extraction: Collect the power generation sequence of the photovoltaic array within a preset time window. Calculate the total power generation and volatility They are used as input features for weather condition recognition; Fuzzy set definition: Fuzzy sets are established for each weather state (sunny, cloudy, overcast, rain / snow), and membership functions for total power and fluctuation are defined. For example: in, and These are baseline parameters determined through historical statistics, not fixed thresholds, and can be dynamically adjusted based on recent period data.

[0053] The fuzzy membership function is represented by a Gaussian function: (twenty three) in, For the first The central parameter for weather-like conditions, This is the width parameter.

[0054] The power fluctuation Defined as the standard deviation of the rate of change of power generation per unit time: (twenty four) in, This represents the average power.

[0055] Fuzzy Reasoning and Decision: Calculating the membership vector of each weather category based on the input features. The final weather condition is determined using the maximum membership principle or weighted fuzzy inference rules. (25) The classification of weather types is based on fuzzy logic reasoning using membership functions, enabling adaptive identification of weather conditions by photovoltaic power characteristics. This significantly improves the intelligence and applicability of the model, demonstrating outstanding technical effectiveness and creativity.

[0056] Dynamic correction: Statistical parameters are updated over time, enabling the system to adaptively adjust the shape of the fuzzy function according to seasonal changes or regional differences, thereby achieving long-term stability and robustness.

[0057] For example, on March 27, 2024, a power station in a certain location generated a total of 34.04 megawatts of electricity between 8:00 AM and 10:00 AM. More significantly, the total minute-level fluctuation in power generation during this period reached 7.76 megawatts. This combination of low power generation and high volatility clearly reflects unstable sunlight conditions. This was caused by frequent cloud movement and obstruction of sunlight; therefore, this period was accurately identified as "cloudy" weather. In stark contrast, during the same two-hour period from 10:00 AM to 12:00 PM on March 29, 2024, the power station generated a total of 110.17 megawatt-hours of electricity. Meanwhile, the total fluctuation in power generation was extremely small, only 0.93 megawatts. This characteristic of high power generation and extremely low volatility is a direct reflection of ideal sunlight conditions. The data shows that the sky was clear and cloudless, with sunlight continuously and stably shining on the photovoltaic panels; therefore, this period was clearly identified as "sunny."

[0058] Furthermore, when calculating the solar angles (solar altitude angle and azimuth angle) based on the latitude and longitude of the location in the example and the date of 2024, the following is included: Calculating true solar time LAST : (26) Use the local time zone. The longitude of the central time zone; Local longitude; This is the time difference correction, which is the difference between true solar time and mean solar time (caused by the eccentricity of the Earth's orbit and the obliquity of the ecliptic, and can be obtained through formulas or tables).

[0059] Calculating solar declination Declination is the latitude of the point where the sun is directly overhead, and it varies with the seasons. (27) It represents the number of days in a year (January 1st is 1, December 31st is 366).

[0060] Calculate the angle The hour angle represents the sun's offset relative to local noon: (28) Calculate the solar altitude angle The angle between the sunlight and the horizon ranges from... A negative value indicates that the sun is below the horizon, which can be determined using the spherical trigonometry formula: (29) in, The latitude is the local latitude. Calculate the solar azimuth angle Starting from due north, turn clockwise to the angle formed by the projection of the sun's rays onto the horizon, ranging from... (30) because The absolute value cannot exceed 1. It may need to be adjusted due to calculation errors or polar day / polar night conditions. In practice, the direction can be determined by the quadrant (eastward in the morning, westward in the afternoon).

[0061] For example, by inputting the latitude and longitude of the location in the example, we can obtain the solar altitude angle at 12:00 noon on January 2, 2024 as 35.71°, indicating that the sun is high in the sky at this time; the solar azimuth angle is 178.54°, which is the angle of the sun relative to due south. 0° is due north, 90° is due east, 180° is due south, and 270° is due west. 178.54° indicates that the sun is almost due south, slightly to the east. At 17:00 on January 2, 2024, the solar altitude angle is 1.34°, indicating that the sun is very close to the horizon, almost at sunset; the solar azimuth angle is 242.13°, meaning the sun is in the southwest direction.

[0062] Furthermore, using the solar constant Atmospheric transparency coefficient With scattering correction coefficient Calculate direct radiation With scattered radiation The preliminary calculation formulas are as follows: in, For atmospheric mass number, when To avoid the denominator being zero.

[0063] To achieve continuous correction under different weather conditions, this invention uses the weather membership degree output by the fuzzy logic recognition module. (Including sunny, cloudy, overcast, rainy, snowy, etc.) By introducing a weighted summation of radiation calculation parameters, the coefficients are no longer fixed, but dynamically determined by the fuzzy membership degree of the weather state: in, These are the empirical baseline values ​​for each weather condition.

[0064] This weighted approach allows the model parameters to transition continuously as the weather changes between sunny and cloudy conditions, avoiding abrupt changes caused by traditional threshold models.

[0065] Calculate the total solar radiation based on the corrected parameters: (34) in, The cloud cover correction coefficient is dynamically adjusted by the fuzzy weights.

[0066] Through the above steps, the present invention realizes a continuous adaptive coupling mechanism from weather identification to radiation estimation, which can adjust the calculation parameters in real time according to different climate conditions, time and solar altitude angle, effectively improving the accuracy of photovoltaic radiation prediction and power generation estimation.

[0067] This embodiment, based on the traditional solar angle and radiation calculation model, introduces fuzzy logic weather membership degree into the radiation parameter correction, and combines solar altitude angle and historical power feedback for dynamic adaptive optimization, realizing weather identification and solar radiation estimation, which has outstanding technical effects and creativity.

[0068] For example: In this case, the solar angle of a power station in this area was calculated to be 35.8° and the solar azimuth angle to be 178.4° at 12:00 on January 3, 2024. Based on the total power generation and fluctuations during that period, the weather was classified as clear, and an atmospheric transparency coefficient was assigned. Afterwards, the solar radiation was estimated to be This value represents the solar radiation power (including direct and diffuse components) received per unit horizontal area under ideal, cloudless, clear weather conditions. It is a typical radiation intensity value achievable under clear skies in mid-latitude regions at midday during winter.

[0069] Step 3: To further improve the accuracy of photovoltaic power separation in the distribution area and the generalization ability of the model, this invention establishes a benchmark power station photovoltaic change... With solar radiation Changes in photovoltaic power in the Taiwan region When determining the mapping relationship, an improved convolutional neural network model that integrates fuzzy logic and solar angle features is proposed. This embodiment is denoted as Fuzzy-Aware CNN, and its core idea is: Based on the traditional CNN structure, meteorological fuzzy membership features, solar angle information, and uncertainty weighting mechanism are added to enable the model to maintain stable recognition performance under different weather conditions.

[0070] The mapping relationship is defined as follows: (35) in, This represents the change in photovoltaic power at the benchmark power station. This refers to solar radiation. and These are the solar altitude angle and azimuth angle, respectively. This is the fuzzy logic weather membership vector output by the fuzzy logic module.

[0071] The model input employs a multi-channel structure, inputting features from different physical sources in parallel to capture the complex nonlinear relationship between solar radiation conditions and photovoltaic power variations. The input features are defined as follows: (36) Among them, the fuzzy membership vector Derived from the aforementioned weather recognition module, it represents the degree of ambiguity of the weather category at each moment.

[0072] Step 4: Train the mapping model using an improved convolutional neural network Fuzzy-AwareCNN model that integrates fuzzy logic and multi-source features.

[0073] The improved CNN model includes two one-dimensional convolutional layers, two max-pooling layers, and a fully connected layer. Convolutional layer 1 and convolutional layer 2 employ a multi-channel convolutional structure to extract the temporal correlation between power change features and solar angle features, respectively. The convolution calculation is as follows: in, and These are the outputs of convolutional layer 1 and convolutional layer 2, respectively. and These are the outputs of pooling layer 1 and pooling layer 2, respectively. This is the weight matrix; For deviation This is the convolution operation and the maximum value function.

[0074] The neural network model in this embodiment is not a simple CNN model, but an innovative design is made in the convolutional input layer, loss function and parameter update strategy: by introducing multi-source features such as fuzzy logical membership degree and solar angle, an uncertainty-weighted self-calibrated CNN structure is constructed, realizing the dynamic adaptive optimization of the photovoltaic power separation model, which has significant technical progress and creativity.

[0075] To address the issue of significant variations in sample volatility under different weather conditions, this invention designs a weighted loss function based on weather uncertainty. This function automatically reduces the loss weight under ambiguous weather conditions, thus preventing overfitting. The loss function is defined as follows: in, This indicates the membership degree of weather clarity. When the weather is clear (such as sunny), the weight is higher; under cloudy or rainy conditions, the weight automatically decreases.

[0076] To enhance the model's adaptability, this invention introduces a weight self-calibration mechanism during CNN training. This mechanism is applied when the predicted residuals of the baseline power plant power and the transformer area power are... When the threshold is exceeded, the convolution kernel weights are automatically adjusted. in, For learning rate, This represents the sign function. This mechanism enables the network to automatically compensate for errors and maintain prediction stability when the environment changes abruptly (such as rapid changes in cloud cover).

[0077] After processing by convolutional and pooling layers, the features are mapped to a fully connected layer and the photovoltaic power output of the distribution area is obtained through the Sigmoid activation function. (43) in, The output feature vector corresponds to the photovoltaic power identification results of the transformer area at different times. This is the weight matrix. This is the deviation term.

[0078] Based on an understanding of the data characteristics in this example, including the volatility of time series data and the complexity of weather impacts, an improved convolutional neural network structure integrating fuzzy logic and solar angle features was designed. This structure includes convolutional layers, pooling layers, and fully connected layers. Leveraging its powerful automatic feature extraction capabilities and spatiotemporal dependency modeling capabilities, it captures the key factors influencing photovoltaic power generation and their interactions. The Adam optimization algorithm is used to minimize the loss function MSE between the identified power and the measured power. A validation set is used for hyperparameter tuning and to prevent overfitting. Dropout and early stopping techniques are also applied.

[0079] Step 5: Evaluate model performance using mean squared error and coefficient of determination, verify accuracy and robustness in different scenarios, optimize the model based on error analysis, and add data corrections for abnormal scenarios.

[0080] In this embodiment, the error analysis-based optimization model can employ a linear model, such as by introducing a regularization term, or a piecewise linear model. Alternatively, a nonlinear model can be used: ensemble learning enhances generalization ability; residual connections and attention mechanisms are introduced into neural networks to capture long-range dependencies. Another approach is a probabilistic model: Bayesian methods quantify uncertainty, or Gaussian process regression handles small sample scenarios.

[0081] The photovoltaic power identification model constructed in this example is trained and calculated using a convolutional neural network. The identified power value is then output. The final results show a high degree of agreement between the identified power and the measured power. The identification results for two specific weather conditions are as follows: Figure 2 As shown.

[0082] Overall, the model performed well across the entire test set, exhibiting some variability under different weather conditions, consistent with the significant weather-related impact on photovoltaic power generation. Notably, the model achieved extremely high identification accuracy on sunny days, with a mean squared error (MSE) of 0.005 MW, indicating a very small mean square deviation between the identified and true values, and a low coefficient of determination. The coefficient of determination (MED) is 0.987, reflecting that under stable meteorological conditions and with highly regular changes in sunlight on sunny days, the Fuzzy-Aware CNN model in this embodiment can capture power generation patterns with extremely high accuracy. While the model's accuracy is slightly lower than on sunny days due to rapidly moving clouds and intense, irregular fluctuations in sunlight intensity during the June rainy season, it still maintains a very high level and significant practical value. At this point, the MSE only slightly increases to 0.007 MW, remaining at an extremely low level, with a coefficient of determination of [missing value]. The value is 0.867. In contrast, the identification results for the June plum rain season, obtained by training and calculating the photovoltaic power identification model using LSTM, are as follows: Figure 3 As shown, the identification accuracy is low, with an MSE of 0.022 MW and a coefficient of determination. The value is only 0.631. This fully demonstrates the strong robustness and generalization ability of the Fuzzy-Aware CNN model when facing complex and unstable weather conditions.

[0083] Example 2: Based on the same inventive concept, such as Figure 4 As shown in the figure, this embodiment also provides a distributed photovoltaic power identification system for a transformer substation, the system comprising: Data preparation module: used to fill missing values ​​and replace outliers in photovoltaic power plant data, simulate different types of loads within a day, and expand to annual load data based on seasonal characteristics, and select photovoltaic power plant data as the benchmark power plant and component area data; The photovoltaic power identification model establishment module is used to calculate the membership degree of each weather state by establishing a fuzzy membership function for total power and power fluctuation, and to classify the weather according to the maximum membership principle. It calculates solar declination, hour angle, solar altitude angle, and azimuth angle, and dynamically corrects the solar radiation estimation parameters based on the fuzzy logic weather identification results, achieving adaptive radiation calculation for different weather and solar angle conditions. It also establishes a mapping model between the photovoltaic change of the benchmark power station and the photovoltaic change of the substation area based on solar radiation. Model training module: This module is used to learn nonlinear mapping relationships with high precision by employing an improved convolutional neural network (Fuzzy-Aware CNN) model that integrates fuzzy logic and multi-source features. This enables the photovoltaic power in the distribution area to be obtained from the photovoltaic power of the benchmark power station through the mapping relationship when actually separating photovoltaic distribution areas.

[0084] Model Evaluation and Optimization Module: This module is used to evaluate model performance through evaluation functions, verify recognition accuracy and robustness by dividing different scenarios, optimize the model based on error analysis results, and add data correction or special case fusion strategies for abnormal scenarios.

[0085] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0086] 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 within the protection scope of the present invention.

Claims

1. A method for identifying the power output of distributed photovoltaic (PV) systems in a transformer substation, characterized in that, The method includes: Missing data and anomaly replacements are performed on the active power data of the photovoltaic grid-connected switch to form a sampling dataset. Based on the sampling dataset, photovoltaic output curves under four typical weather conditions are constructed to determine the benchmark power station and obtain a sample library of photovoltaic power in the distribution area. The four typical weather conditions are sunny, cloudy, overcast, and rainy / snowy. The photovoltaic processing curve is used to collect the power generation sequence of the photovoltaic array within a preset time window. Based on the power generation sequence, a fuzzy membership model corresponding to four typical weather conditions is constructed to achieve adaptive identification of weather conditions. The fuzzy membership model is obtained through the total power generation and power fluctuation rate. The fuzzy logic weather membership degree obtained by the fuzzy membership model is used to correct direct radiation and scattered radiation, so as to realize adaptive solar radiation estimation under different weather conditions; A photovoltaic change model for a distribution area is constructed based on the photovoltaic change of a benchmark power station and the current solar radiation. The results of the photovoltaic change model are then input into a feature extraction model. After training the feature extraction model, a feature vector is output, which corresponds to the photovoltaic power identification results of the distribution area at different times. The loss function used in the feature extraction model is obtained by weighting the uncertainty of the weather membership degree based on fuzzy logic. The trained feature extraction model is evaluated using an evaluation function, and its current accuracy and robustness are verified in different scenarios. The model is then optimized based on error analysis, and data corrections are added for abnormal scenarios.

2. The method for identifying distributed photovoltaic power in a transformer substation according to claim 1, characterized in that, The sample library for determining the benchmark power station and obtaining the photovoltaic power of the distribution area includes: Multiple distributed photovoltaic power stations with complementary power output characteristics were selected in a certain area, and their grid-connected switch active power data were continuously collected. Plot the actual photovoltaic output curves under four types of typical weather conditions, and select power plants with strong output stability and high matching degree with regional load characteristics under the four types of weather conditions as benchmark power plants based on the characteristics of the actual photovoltaic output curves under each type of typical weather condition. Combine power plant data with complementary power output characteristics to construct a sample library of distribution areas.

3. The method for identifying distributed photovoltaic power in a transformer substation according to claim 1, characterized in that, The construction of fuzzy membership models corresponding to four typical weather conditions based on the power generation sequence includes: The standard deviation of the rate of change of power generation per unit time is obtained from the power generation sequence and used as the power fluctuation. The power fluctuation and total power generation are used as independent variables to represent the fuzzy membership model. The dependent variable of this model is the fuzzy membership function corresponding to four typical weather conditions determined based on historical power generation data.

4. The method for identifying distributed photovoltaic power in a transformer substation according to claim 3, characterized in that, The adaptive identification process for the weather conditions includes: Based on current data of different typical weather conditions, the fuzzy logic weather membership vector of each weather category is obtained by using the corresponding fuzzy membership function, and the current weather state is determined according to the principle of maximum membership. The maximum membership principle is implemented using a rule-based fuzzy controller structure, wherein the rule base includes: The weather conditions corresponding to high total power generation and low volatility are sunny, cloudy, overcast, and rain / snow. At the end of each day, the baseline parameters of each fuzzy function are updated based on the statistical results of the total power generation and fluctuation over the past few days to achieve adaptive correction for seasonal changes. The baseline parameters include a center parameter and a width parameter.

5. The method for identifying distributed photovoltaic power in a transformer substation according to claim 4, characterized in that, The method employs a fuzzy logic weather membership degree correction model to adjust direct and scattered radiation, enabling adaptive solar radiation estimation under different weather conditions, including: The dynamic atmospheric transparency coefficient and dynamic scattering correction coefficient are established by using the fuzzy logic weather membership degree under different weather conditions; Direct radiation and scattered radiation are represented based on the dynamic atmospheric transparency coefficient and dynamic scattering correction coefficient; The total solar radiation is calculated using the direct radiation and the diffuse radiation. The total solar radiation is expressed as the sum of the direct radiation and the diffuse radiation, multiplied by a cloud cover correction factor. The cloud cover correction factor is obtained by weighting empirical benchmark values ​​under various weather conditions with fuzzy logic weather membership degrees.

6. The method for identifying distributed photovoltaic power in a transformer substation according to claim 5, characterized in that, A photovoltaic variation model for the distribution area is constructed based on the photovoltaic variation of a benchmark power station and the current solar radiation. The results of the photovoltaic variation model are then input into a feature extraction model, including: Construct a feature mapping relationship between the photovoltaic change characteristics, total solar radiation characteristics, fuzzy logic weather membership characteristics, solar altitude angle characteristics, and solar azimuth angle characteristics of the benchmark power station and the photovoltaic change of the substation area; A multi-channel structure is adopted to input features from different physical sources into the feature extraction model in parallel. The feature extraction model includes two one-dimensional convolutional layers, two max pooling layers, and a fully connected layer. Both one-dimensional convolutional layers adopt a multi-channel convolutional structure and are used to extract the temporal correlation between power change features and solar angle features, respectively.

7. The method for identifying distributed photovoltaic power in a transformer substation according to claim 6, characterized in that, The convolutional kernel weights of the convolutional layer are adaptively adjusted based on the prediction residuals between the baseline power plant power and the power distribution area power during the training process. If the prediction residuals between the baseline power plant power and the power distribution area power exceed a set threshold, the convolutional kernel weights of the convolutional layer are updated, as shown below: in, For learning rate, Represents a symbolic function. These are the current kernel weights of the convolutional layer. This represents the predicted residual between the base power plant power and the power of the distribution area.

8. The method for identifying distributed photovoltaic power in a transformer substation according to claim 7, characterized in that, The loss function used in the feature extraction model is obtained by weighting the uncertainty of the fuzzy logic weather membership degree. The loss function is defined as follows: ;in, express t Weather clarity at any given moment, for t The change in photovoltaic power generation in the transformer substation at any given time. This represents the photovoltaic fluctuation predicted by the model.

9. A distributed photovoltaic power identification system for a transformer substation, characterized in that, The system includes; The data preparation module is used to fill in missing data and replace anomalies in the active power data of the photovoltaic grid-connected switch to form a sampling dataset. Based on the sampling dataset, the photovoltaic output curves under four typical weather conditions are constructed to determine the benchmark power station and obtain a sample library of photovoltaic power in the distribution area. The four typical weather conditions are sunny, cloudy, overcast, and rain / snow. The photovoltaic power identification model building module is used to collect the power generation sequence of the photovoltaic array within a preset time window through the photovoltaic processing curve, and to build fuzzy membership models corresponding to four typical weather conditions based on the power generation sequence to achieve adaptive identification of weather conditions. The fuzzy membership model is obtained through the total power generation and power fluctuation rate. The solar radiation calculation module is used to correct direct and scattered radiation using fuzzy logic weather membership degrees obtained from fuzzy membership models, thereby achieving adaptive solar radiation estimation under different weather conditions. The model training module is used to construct a photovoltaic change model for the substation area based on the photovoltaic change of the benchmark power station and the current solar radiation. The results of the photovoltaic change model are input into the feature extraction model. After training the feature extraction model, the feature vector is output to identify the photovoltaic power of the substation area at different times. The loss function used by the feature extraction model is obtained by weighting the uncertainty of the fuzzy logic weather membership degree. The model evaluation and optimization module evaluates the trained feature extraction model through an evaluation function, verifies the current accuracy and robustness in different scenarios, optimizes the model based on error analysis, and adds data correction for abnormal scenarios.

Images