Method and system for evaluating photovoltaic bearing capacity of power distribution network based on multiple dimensions

Abstract

The invention relates to the field of power distribution network bearing capacity analysis, in particular to a multi-dimension-based power distribution network photovoltaic bearing capacity evaluation method and system, and aims to solve the problems that HC evaluation is inaccurate and the real-time requirement is difficult to meet in the prior art. According to the method, firstly, a multi-dimensional voltage quality index measurement space including voltage deviation, voltage unbalance degree and voltage fluctuation amplitude is constructed, and a comprehensive operation constraint model is established in combination with a total harmonic voltage distortion rate and an individual harmonic voltage distortion rate; thirdly, learning a complex nonlinear time sequence mapping relation between a time sequence of a power grid multi-dimensional state in historical operation data and the maximum allowable photovoltaic bearing capacity meeting the strict constraint conditions by using a BILSTM network; by comprehensively considering the key power quality and the harmonic constraint and applying the strong time sequence information processing capability of the BILSTM, the accuracy and robustness of HC evaluation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of distribution network carrying capacity analysis, and particularly to a method and system for evaluating the photovoltaic carrying capacity of a multi-dimensional distribution network. Background Art

[0002] With the advancement of the global energy transformation, the penetration rate of distributed renewable energy represented by photovoltaic in the distribution network is increasing day by day. However, the access of large-scale distributed photovoltaic has brought new challenges to the safe and stable operation of the distribution network. The intermittency, volatility, and randomness of photovoltaic output, as well as the characteristics of power electronic interface devices (such as inverters), may lead to problems such as voltage over-limit, voltage fluctuation, increased voltage imbalance, power flow reversal, line overload, and harmonic pollution in the distribution network. Accurately evaluating the acceptance capacity of the distribution network for distributed photovoltaic, that is, the photovoltaic carrying capacity (HC), is crucial for guiding photovoltaic planning, access approval, and optimal operation.

[0003] The photovoltaic carrying capacity is usually defined as the maximum photovoltaic installed capacity or power generation that the distribution network can accept without violating the operating safety constraints of the distribution network (such as voltage limits, thermal stability limits, protection coordination requirements, power quality standards, etc.). Traditional HC evaluation methods mainly include:

[0004] 1. Deterministic analysis method: This method usually performs power flow calculation or time-series power flow calculation based on a certain "worst" scenario (such as peak load, light load, maximum photovoltaic output, etc.) to evaluate whether the operating constraints are met in these specific scenarios. This type of method is simple to calculate, but ignores the time-varying and random nature of the load and photovoltaic output, and the evaluation results are often too conservative or optimistic and difficult to reflect the true carrying capacity of the system.

[0005] 2. Stochastic analysis method: Considering the uncertainty of the load and photovoltaic output, this method uses methods such as Monte Carlo simulation, scenario analysis, and probabilistic power flow to simulate a large number of possible operating scenarios to obtain the probability distribution of HC. This type of method can more comprehensively reflect the impact of uncertainty, but the calculation amount is huge and the time consumption is too long, making it difficult to meet the requirements of real-time or quasi-real-time evaluation.

[0006] 3. Time-series analysis method: By performing continuous time-series power flow calculations on historical or predicted data on a long time scale, the operating constraints are checked moment by moment. This method can better capture the dynamic characteristics and time-series correlations of the system, but it also has the problem of low calculation efficiency, especially when frequent evaluation or large-scale system analysis is required.

[0007] In recent years, with the development of artificial intelligence (AI) and machine learning (ML) technologies, some studies have begun to attempt to apply them to HC assessment. For example, using models such as support vector machines (SVMs), artificial neural networks (ANNs), and decision trees (DTs) to learn the relationships between factors such as power grid topology, load levels, and PV configurations and HC. These methods have improved the assessment efficiency to a certain extent.

[0008] However, the existing technologies still have the following main problems in the assessment of the HC of distribution networks:

[0009] (1) Insufficient assessment accuracy: Many methods adopt simplified constraint models, mainly focusing on voltage magnitude violations and line thermal overloads, and ignoring the comprehensive consideration of power quality (especially dynamic voltage quality, such as voltage fluctuations and voltage unbalances) and harmonic distortions. The grid connection operation of PVs, especially in weak grid or high-penetration scenarios, may significantly affect these indicators. If not constrained, the HC obtained from the assessment may lead to serious power quality problems in practical applications.

[0010] (2) Difficulty in meeting real-time requirements: Traditional simulation-based methods (stochastic analysis, time series analysis) have high computational complexity and cannot meet the needs of rapid and dynamic assessment of HC in the operation of distribution networks. For example, when formulating the day-ahead dispatch plan, performing intraday rolling optimization, or coping with sudden weather changes, it is necessary to almost real-time understand the current or future short-term HC. Although the speed of traditional AI-based methods has been improved, some models (such as standard ANNs) may be difficult to fully capture complex time series dependencies.

[0011] (3) Imperfect constraint models: Existing studies often regard voltage deviation, voltage unbalance, voltage fluctuations, etc. as independent constraint conditions, and fail to comprehensively reflect the overall situation of voltage quality from multiple dimensions. In addition, the consideration of harmonic problems is insufficient. Especially in the context of high penetration of power electronic devices, harmonic distortions (total harmonic distortion rate THD and individual harmonic distortion rate IHD) have become one of the key factors restricting HC, and it is necessary to establish a clear harmonic constraint model.

[0012] (4) Insufficient capture of temporal characteristics: The operating states of distribution networks (load, PV output, voltage, harmonic background, etc.) have significant temporal correlations. HC not only depends on the current state but is also affected by historical states and future trends. Traditional static models or machine learning models that fail to effectively handle long-term dependencies are difficult to accurately capture this complex temporal mapping relationship.

[0013] Therefore, there is an urgent need for a new method and system that can comprehensively consider multi-dimensional voltage quality indicators and harmonic distortion constraints and can effectively capture the temporal characteristics of the grid operating state, so as to achieve accurate and real-time assessment of the PV hosting capacity of distribution networks. Summary of the Invention

[0014] (1) Technical problems to be solved

[0015] The present invention aims to solve the technical problems existing in the prior art, that is, the photovoltaic carrying capacity evaluation method fails to comprehensively consider key constraints such as voltage fluctuation, voltage unbalance and harmonic distortion, resulting in inaccurate evaluation results, and the traditional evaluation method has low calculation efficiency and is difficult to meet the real-time evaluation requirements of HC in the dynamic operation of the distribution network.

[0016] (2) Technical solutions

[0017] To solve the above technical problems, the present invention proposes a multi-dimensional distribution network photovoltaic carrying capacity evaluation method. The core of this method is to construct a voltage fluctuation measure space containing multi-dimensional voltage quality indicators (voltage deviation, voltage unbalance degree, voltage fluctuation amplitude), and combine a clear harmonic distortion rate (THD and IHD) constraint model; utilize the powerful time series feature learning ability of the bidirectional long short-term memory network (BILSTM) to learn the complex non-linear time series mapping relationship between the multi-dimensional states of the power grid (load, photovoltaic output, meteorological information, node voltage, background harmonics, etc.) and the maximum allowable photovoltaic access power (i.e., HC) under the above strict constraint conditions from historical operation data; finally, use the trained BILSTM model to quickly output the current HC evaluation value according to the real-time power grid state input.

[0018] Specifically, the method proposed by the present invention includes the following steps:

[0019] Step 1: Data collection and preprocessing.

[0020] Collect the historical operation data of the distribution network, including: the time series data of the active and reactive power of each node load; the time series data of the actual output of the connected photovoltaic power stations; the time series data of relevant meteorological information (light intensity, temperature); the time series data of the three-phase voltage amplitudes of key nodes; the time series data of the background harmonic voltage or THD, IHD of key nodes; the power grid topology structure and line parameters. Clean (specifically handle missing values and outliers), synchronize (unify time stamps), and normalize or standardize the collected original data to eliminate the influence of different physical dimensions and numerical ranges, and improve the training efficiency and stability of the model.

[0021] Step 2: Voltage fluctuation measure space constraint modeling.

[0022] Define a set of multi-dimensional indicators for quantifying voltage quality and construct a voltage fluctuation measure space. This space includes the following dimensions:

[0023] 1) Voltage deviation: Evaluate the degree to which the node voltage amplitude deviates from the rated value. For any node in the distribution network At time , its voltage deviation is defined as:

[0024] ;

[0025] wherein, is the actual positive-sequence voltage amplitude of node at time ; is the rated voltage of the system.

[0026] The voltage deviation constraint is:

[0027] ;

[0028] wherein, is the maximum allowable voltage deviation percentage, set according to GB / T 12325-2008.

[0029] 2) Voltage unbalance: Evaluates the asymmetry of three-phase voltages. It is defined based on the ratio of the negative-sequence voltage amplitude to the positive-sequence voltage amplitude as:

[0030] ;

[0031] wherein, and are the amplitudes of the actual positive-sequence voltage and negative-sequence voltage of node at time , respectively, and are calculated from the three-phase voltage phasors , , by the symmetrical component method:

[0032] ;

[0033] wherein, is the complex rotation factor; is the zero-sequence voltage.

[0034] The voltage unbalance constraint is:

[0035] ;

[0036] wherein, is the maximum allowable voltage unbalance, set to 2% according to GB / T15543-2008 in this embodiment.

[0037] 3) Voltage Fluctuation Amplitude: Evaluates the rapid change degree of the voltage amplitude within a short time. This index measures the percentage of the difference between the maximum and minimum values of the voltage within a specified time window relative to the rated voltage. The voltage fluctuation amplitude is defined within the time window as:

[0038] ;

[0039] where is the positive-sequence voltage amplitude of node at time .

[0040] The voltage fluctuation amplitude constraint is:

[0041] ;

[0042] where is the maximum allowable voltage fluctuation amplitude, which is set according to the guidance of GB / T12326 - 2008 in this embodiment.

[0043] The voltage fluctuation measure space is a multi-dimensional space composed of , , these three indexes. The voltage quality of an operating state is qualified if its coordinates in this space fall within the allowable region defined by the , , constraint boundaries.

[0044] Step 3: Modeling of harmonic distortion rate constraint.

[0045] Establish a clear harmonic distortion constraint model, considering the impact of harmonics generated by non-linear devices such as photovoltaic inverters on the power grid.

[0046] 1) Total harmonic voltage distortion rate: Evaluates the overall distortion degree of the node voltage waveform. For node at time , its total harmonic voltage distortion rate is:

[0047] ;

[0048] where is the amplitude of the th harmonic voltage component of node at time ; is the amplitude of the actual positive-sequence voltage; is the highest harmonic order considered.

[0049] The total harmonic voltage distortion rate constraint is:

[0050] ;

[0051] where is the maximum allowable total harmonic voltage distortion rate, which is set according to the national standard GB / T 14549 - 1993 in this embodiment.

[0052] 2) Individual harmonic voltage distortion rate: Evaluates the relative magnitude of the harmonic voltage component of a specific order. For node at time of the st harmonic, the individual harmonic voltage distortion rate is defined as:

[0053] ;

[0054] where and are defined as above.

[0055] The individual harmonic voltage distortion rate constraint is:

[0056] ;

[0057] where is the allowable th maximum individual harmonic voltage distortion rate, and its specific limit value is set according to the regulations of the national standard GB / T 14549 - 1993 for different harmonics (odd and even).

[0058] Step Four: Generation of the maximum photovoltaic carrying capacity (HC) label based on historical data.

[0059] This step generates the training labels for the BILSTM model. For each time in the historical dataset, determine the maximum photovoltaic power that can be additionally connected under the grid state (load, background harmonics, existing photovoltaic output, etc.) at that time, such that all nodes still satisfy all the voltage quality and harmonic constraints in Step Two and Step Three after adding . This is accomplished by combining historical data and the grid model, using an iterative increase method based on power flow calculation and constraint checking:

[0060] a) Select the grid snapshot data at a historical time .

[0061] b) Set an initial photovoltaic power to be added , this power will increase at several specified PV access nodes.

[0062] c) At the current total PV power , perform power flow calculation (using the constant power factor model of PV inverters) and harmonic power flow calculation (using the harmonic current source model of PV inverters and background harmonic data).

[0063] d) Check whether all monitoring nodes meet the voltage deviation, unbalance degree, and fluctuation amplitude constraints in Step 2, and the harmonic distortion rate constraint in Step 3.

[0064] e) If all constraints are met, then increase , that is , where is a preset power increment step, and then return to Step c). If any constraint is violated, the maximum additional PV power allowed at the current moment is determined as the value of the previous step.

[0065] f) Use the calculated total maximum allowable PV power as the label for the moment . If the calculated , then is set to .

[0066] Repeat this process for all (or selected representative) moments in the historical data to generate a series of data pairs (historical state sequences, HC labels) for model training.

[0067] Step Five: Construction and Training of the BILSTM Model.

[0068] Construct a BILSTM network model to learn the mapping from the power grid state time series to the HC value that meets the constraints.

[0069] 1) Input Layer: The input is a sequence containing power grid state vectors of historical time steps. The state vector

[0070] at each time step contains the preprocessed multi-dimensional information at that moment, which specifically includes in this embodiment:

[0071] Among them, is the total active power load of the system at the moment , and is the total reactive power load of the system at the moment . is the total existing PV output of the system at time , is the average light intensity at time , is the average ambient temperature at time , is the positive-sequence voltage amplitude of node 6 at time , is the positive-sequence voltage amplitude of node 18 at time , is the positive-sequence voltage amplitude of node 32 at time , is the total harmonic distortion rate of the voltage at node 18 at time , is the individual harmonic distortion rate of the 5th harmonic voltage at node 18 at time .

[0072] 2) BILSTM layer: The input sequence is fed into the BILSTM layer. The BILSTM contains a forward LSTM and a backward LSTM.

[0073] The forward LSTM ( ) processes the sequence from to , captures the influence of past information on the current state, and outputs the forward hidden state sequence .

[0074] The backward LSTM ( ) processes the reverse sequence from to , captures the influence of subsequent information in the sequence on the current state, and outputs the backward hidden state sequence .

[0075] At each time step , the forward and backward hidden states are concatenated to form the output of the BILSTM layer at that time step .

[0076] 3) Output layer: The combined hidden state at the last time step is fed into one or more fully connected layers (Dense Layer), and finally mapped to a scalar output value, which is the predicted PV carrying capacity at the current time , .

[0077] 4) Training: Use the (input sequence, HC label) dataset generated in Step 4 to perform supervised learning training on the BILSTM model. Use the mean squared error (MSE) as the loss function and adopt the Adam optimizer to adjust the network weights. Minimize the error between the predicted by the model and the true label through the backpropagation algorithm. Use the validation set for model selection and confirmation of hyperparameter settings.

[0078] Step 6: Real-time HC evaluation.

[0079] Deploy the trained BILSTM model to the online evaluation system. The system collects the grid status data at the current moment and the previous several moments in real time to form the input sequence . Input the after the same preprocessing as during training into the trained BILSTM model, and the model quickly outputs the predicted value of the PV carrying capacity of the distribution network at the current moment . This value represents the maximum PV power that the system can accept under the current grid operating conditions and meeting the preset voltage fluctuation measure space and harmonic distortion rate constraint conditions.

[0080] In addition, the present invention also provides a real-time evaluation system for the PV carrying capacity of a distribution network based on the above method. The system at least includes:

[0081] Data acquisition and preprocessing module: Responsible for obtaining the required grid operation data from data sources such as SCADA, PMU, meteorological stations, and PV inverter monitoring systems in real time or quasi-real time, and performing preprocessing operations such as data cleaning, synchronization, and normalization.

[0082] Constraint definition module: Stores and manages the definitions and constraint limits of the voltage fluctuation measure space (voltage deviation, unbalance degree, fluctuation amplitude), as well as the constraint limits of the harmonic distortion rate (THD, IHD). These limits are configured according to relevant standards and specific requirements of the grid.

[0083] BILSTM model module: Stores the parameters of the trained BILSTM model. Receives the preprocessed real-time input sequence, performs forward calculation, and outputs the HC prediction value.

[0084] HC evaluation and output module: Invokes the BILSTM model module for calculation, obtains the HC evaluation result, and displays the result to the operator in numerical and graphical forms. At the same time, sends the HC value to the energy management system (EMS), distribution management system (DMS), or PV power station controller for guiding operation decisions.

[0085] Model Training and Update Module: It includes functions for performing Step 4 (HC Label Generation) and Step 5 (BILSTM Model Training), enabling the initial training of the offline model and regular model updates to adapt to changes in the power grid structure or new operating modes.

[0086] (III) Beneficial Effects

[0087] Compared with the prior art, the present invention has the following remarkable beneficial effects:

[0088] 1. Improved accuracy of HC assessment: By constructing a multi-dimensional voltage fluctuation measurement space including voltage deviation, unbalance degree, and fluctuation amplitude, and combining strict THD and IHD harmonic constraints, the present invention can more comprehensively and accurately reflect the key factors restricting PV access, avoiding the problem of inaccurate HC assessment results caused by ignoring important constraints, and the assessment results are closer to the actual operation limits.

[0089] 2. Achieved real-time HC assessment: Using the trained BILSTM model for online assessment, avoiding time-consuming power flow calculations or Monte Carlo simulations in traditional methods. The model inference process is completed in milliseconds, meeting the real-time requirements of HC information for the dynamic operation and control of the distribution network. Specific Embodiments

[0090] The technical solutions of the present invention will be further described in detail below in conjunction with specific embodiments.

[0091] Embodiment 1:

[0092] The implementation process of this embodiment is described using a modified IEEE 33-node radial distribution system model with a rated voltage of 12.66 kV. The nodes where distributed photovoltaic power generation units are connected to the system are 6, 12, 18, 25, and 32.

[0093] 1. Data collection and preprocessing:

[0094] Collect historical operation data covering one year with a time resolution of 15 minutes.

[0095] The collected data includes:

[0096] The active power load of each load node every 15 minutes and the reactive power load .

[0097] The actual active power output of each connected PV site every 15 minutes .

[0098] The corresponding 15-minute average light intensity and ambient temperature data.

[0099] The three-phase voltage amplitudes of each node every 15 minutes , obtained through time-sequence power flow simulation.

[0100] The background harmonic voltage spectrum of key nodes (including PV connection points and end nodes) every 15 minutes (h = 2, 3,..., 50), obtained through time-sequence harmonic power flow simulation.

[0101] System topology and line parameters (R, X).

[0102] Data preprocessing:

[0103] Cleaning: Use linear interpolation to fill in missing data; Use the 3σ principle to identify and process (such as replacing with interpolation) outliers.

[0104] Synchronization: Ensure that all data sequences are aligned in time.

[0105] Normalization: Use the min-max normalization method to scale all input features to the interval [0, 1]. For example: for the load Perform normalization:

[0106] ;

[0107] where and are respectively the minimum and maximum values of the load at node in the historical data.

[0108] 2. Definition of voltage fluctuation measure space and harmonic constraints:

[0109] Calculation of voltage quality indicators: For each historical moment and each monitoring node (including all PV connection nodes, voltage-sensitive nodes, and feeder end nodes), calculate according to the collected three-phase voltage data:

[0110] Voltage deviation (calculated using the positive-sequence voltage ).

[0111] Voltage unbalance degree .

[0112] Voltage fluctuation amplitude (calculate the voltage variation within a 15-minute window).

[0113] Calculation of harmonic indicators: Calculate the of the monitoring node and the of specific harmonics (3rd, 5th, 7th, 11th, 13th).

[0114] Constraint limit setting: Set the constraint limits according to national standards (such as GB / T 12325, 15543, 12326, 14549). The specific limits used in this embodiment are shown in Table 1.

[0115] Table 1:

[0116] ;

[0117] 3. HC label generation:

[0118] For each 15-minute time point in the historical dataset , perform the following operations:

[0119] Obtain the load at this moment , the existing PV output , and the background harmonics .

[0120] The newly added PV access capacity is proportionally allocated to the specified PV access nodes (nodes 6, 12, 18, 25, 32).

[0121] Use the iterative search method (step size = 10 kW) to find the maximum .

[0122] During the iteration process, the total PV output . The output of the newly added capacity is assumed to have a unity power factor, and its output magnitude is adjusted according to the normalized output curve determined by the current irradiance (consistent with the existing PV).

[0123] Evaluate the system status by performing Newton-Raphson power flow calculation and harmonic power flow calculation (frequency domain method).

[0124] Power flow calculation model: The PV inverter operates according to the constant power factor 1.0 model.

[0125] Harmonic power flow calculation model: The PV inverter adopts the harmonic current source model based on the IEEE 1547 standard (injecting characteristic sub-harmonics), and the power grid adopts the frequency-dependent impedance model.

[0126] Constraint check: Check whether all monitoring nodes and lines meet Table 1. The optimization problem is formulated as follows:

[0127] , is the PV power to be added;

[0128] Meet the following conditions:

[0129] , where is the node at time , the voltage deviation after increasing photovoltaic power;

[0130] , where is the node at time , the voltage unbalance degree after increasing photovoltaic power;

[0131] , where is the node at time , the voltage fluctuation amplitude after increasing photovoltaic power;

[0132] , where is the node at time , the total harmonic voltage distortion rate after increasing photovoltaic power, is the maximum allowable total harmonic voltage distortion rate;

[0133] , is the node at time , the individual harmonic voltage distortion rate of the th harmonic after increasing photovoltaic power; where each index is a function of and the state at time .

[0134] Take the calculated as the label at time . Where is the photovoltaic carrying capacity at time , is the existing photovoltaic output at the photovoltaic access point at time , is the total existing photovoltaic output of the system at time , is the maximum additional photovoltaic power that can be added at time . If , then . Execute this process on the entire historical data set to generate the training set .

[0135] 4. Construction and Training of BILSTM Model:

[0136] Input Feature Selection: State Vector It includes the following features: ;

[0137] Among them, is the total active power load of the system at time , is the total reactive power load of the system at time , is the total existing PV output of the system at time , is the average light intensity at time , is the average ambient temperature at time , is the positive sequence voltage amplitude of Node 6 at time , is the positive sequence voltage amplitude of Node 18 at time , is the positive sequence voltage amplitude of Node 32 at time , is the total harmonic distortion rate of the voltage of Node 18 at time , is the individual harmonic distortion rate of the 5th harmonic voltage of Node 18 at time .

[0138] Input Sequence uses the most recent time steps (i.e., data for the past 3 hours).

[0139] BILSTM Model Architecture:

[0140] Input Layer: The dimension is .

[0141] BILSTM Layer: Two layers of BILSTM, with 128 LSTM units in each forward and backward direction, using the tanh activation function.

[0142] Dropout Layer: A Dropout layer is connected after each BILSTM layer, with rate = 0.2.

[0143] Fully Connected Layer (Dense): The output of the last time step of the second BILSTM layer is connected to a hidden layer with 64 neurons (ReLU activation), and then connected to an output layer with a single neuron (linear activation), outputting .

[0144] Training:

[0145] Divide the generated (sequence, label) dataset into a training set (70%), a validation set (15%), and a test set (15%).

[0146] Loss function: Mean Squared Error (MSE):

[0147] ;

[0148] Optimizer: Adam optimizer, initial learning rate 0.001, using a learning rate decay strategy.

[0149] Training process: Train the model on the training set, use the validation set to monitor performance and prevent overfitting through early stopping, and save the model weights with the best performance.

[0150] Hyperparameter specifications: The hyperparameters of the BILSTM model in this embodiment are shown in Table 2.

[0151] Table 2:

[0152] ;

[0153] 5. Real-time HC evaluation:

[0154] Deploy the trained optimal BILSTM model on the evaluation server.

[0155] The system collects the grid status data required at the current moment and the previous 11 moments (a total of 12 moments) in real time every 5 minutes .

[0156] Perform the same preprocessing (normalization) on these real-time data as during training.

[0157] Input the preprocessed sequence into the loaded BILSTM model.

[0158] The model outputs the predicted .

[0159] Result display and application: The will be displayed on the monitoring interface. Calculate the real-time HC margin . This margin information is used for:

[0160] Guiding the quick preliminary evaluation of new photovoltaic grid connection applications.

[0161] Setting the upper limit of photovoltaic power output in day-ahead or intra-day scheduling.

[0162] When it is predicted that within a short period in the future When the PV output may be lower than expected, start the control strategy in advance, such as issuing a curtailment instruction, adjusting the voltage control equipment, starting the energy storage, etc.

[0163] 6. Performance evaluation and comparative analysis:

[0164] Evaluate the performance of the trained BILSTM model using the test set, and adopt the RMSE, MAE, and MAPE metrics.

[0165] Compare the method of the present invention (BILSTM) with the following methods:

[0166] a) Traditional deterministic method: Conduct power flow calculations based on the most severe operating scenarios (30% of the peak load at the lowest load, maximum PV output of 1.0 p.u., PF = 1), and check the voltage deviation (±7%).

[0167] b) ANN-based method: Use MLP (3 hidden layers: 128 - 64 - 32, ReLU activation), and the input is the state snapshot at the current moment , and the label generation is the same as that of BILSTM.

[0168] c) Unidirectional LSTM-based method: Use unidirectional LSTM (2 layers, 128 units in each layer), and the input is the historical sequence , and the label generation is the same as that of BILSTM.

[0169] Comparative experiment settings: Use the same IEEE 33-node simulation dataset and training / validation / test division. All methods use the HC labels generated by the comprehensive constraints (Table 1).

[0170] Comparative results: The performance comparison is shown in Table 3.

[0171] Table 3:

[0172] ;

[0173] Table 3 shows that the BILSTM method of the present invention is significantly superior to other methods in terms of RMSE, MAE, and MAPE metrics. The deterministic method has the largest error due to the limitations of the scenarios. ANN lacks temporal capabilities and its accuracy is limited. The unidirectional LSTM is superior to ANN but inferior to BILSTM because the latter can utilize bidirectional information flow to more comprehensively capture temporal dependencies, thus most accurately fitting the dynamic HC defined by the comprehensive constraints.

[0174] In addition, the deterministic method ignores the key power quality and harmonic constraints. All machine learning methods indirectly satisfy the constraints by learning the labels generated from comprehensive constraints. However, the method of the present invention not only uses comprehensive constraints to generate labels, but also adopts a more powerful BILSTM model, which can more accurately learn and reproduce these complex constraint boundaries. Therefore, the method of the present invention combines comprehensive constraint definition (label generation) and powerful model learning ability (model selection) in terms of constraint comprehensiveness, and has the best effect.

[0175] ANN cannot handle time series. Unidirectional LSTM deals with historical dependencies. The BILSTM of the present invention captures time series dependencies most fully and accurately by combining forward and backward information flows.

[0176] It can be seen that the BILSTM method proposed by the present invention generates high-quality labels by strictly considering multi-dimensional voltage quality and harmonic constraints in the offline stage, and uses the BILSTM model to efficiently learn complex time series relationships online, achieving the highest accuracy at the millisecond-level real-time evaluation speed, and the evaluation results can more reliably contain comprehensive operation constraints, which is the preferred HC evaluation scheme to support the operation of high-proportion renewable energy distribution networks.

[0177] Embodiment 2:

[0178] The architecture of the real-time evaluation system for implementing the method of the present invention is described below.

[0179] This system is deployed in the distribution network control center or regional dispatch center, and its main modules and functions are as follows:

[0180] Data acquisition and interface module:

[0181] Obtain real-time node voltages, line currents, switch states, and load data from the SCADA system through standard protocols.

[0182] Obtain high-precision voltage phasors and frequency data from the PMU.

[0183] Obtain real-time light intensity and temperature data from the meteorological service interface.

[0184] Obtain real-time output data of each site from the photovoltaic power plant monitoring system.

[0185] Obtain real-time harmonic voltages, currents, THD, and IHD data from the power quality monitoring device. If the number of monitoring points is insufficient, the background harmonics are estimated by combining models.

[0186] Provide a data interface layer to integrate and verify data from different sources.

[0187] Data preprocessing module:

[0188] Real-time data stream processing: Align timestamps for the incoming real-time data stream, fill in missing values (using the previous valid measurement value), and detect and handle outliers.

[0189] Feature engineering: Calculate derived features, including total load, total PV output, positive sequence voltage, voltage unbalance degree, short-term voltage fluctuation index, THD, and IHD value.

[0190] Normalization: Apply min-max normalization using the parameters determined in the training phase.

[0191] Sequence construction: Maintain a sliding window of length L = 12 and construct the input sequence for the BILSTM model. 。

[0192] Constraint definition and management module:

[0193] Provide a user interface and configuration file to allow operators to set and adjust the various constraint limits shown in Table 1 according to relevant standards and the specific conditions of the power grid.

[0194] Store these limits for offline HC label generation and online result verification.

[0195] BILSTM model inference engine module:

[0196] Load the trained BILSTM model file.

[0197] Provide an API interface to receive the real-time input sequence. 。

[0198] Output the predicted HC value. 。

[0199] HC evaluation result processing and publishing module:

[0200] Receive 。

[0201] Perform result verification (e.g., compare with historical ranges or physical limits) or calculate the confidence score.

[0202] Calculate the real-time HC margin. 。

[0203] Store the results ( , ) into the time series database.

[0204] Display the real-time HC value, historical trend, and margin status to the operator through the graphical user interface.

[0205] Push the HC results to EMS / DMS, photovoltaic power control systems, etc. through the message queue and API.

[0206] Model training and updating module:

[0207] HC label generator: Recalculate HC labels monthly using the latest historical data and run on dedicated high-performance computing resources.

[0208] BILSTM model trainer: Retrain or fine-tune the BILSTM model using the latest HC labels and historical sequence data.

[0209] Although the present invention has been described with reference to specific embodiments, those skilled in the art should understand that various changes and equivalent substitutions can be made without departing from the concept of the present invention. For example, different combinations of voltage quality indicators or harmonic constraint indicators can be adopted; therefore, the scope of the present invention should not be limited to the specific embodiments described, but should be determined by the claims.

Claims

1. A method for evaluating photovoltaic carrying capacity of a distribution network based on multiple dimensions, characterized in that: The method comprises the following steps: a) defining distribution network operation constraints, wherein the constraints at least include voltage fluctuation measurement space constraints and harmonic distortion rate constraints; the voltage fluctuation measurement space constraints cover voltage deviation, voltage imbalance and voltage fluctuation amplitude indicators within a preset time window; the harmonic distortion rate constraints cover total harmonic voltage distortion rate and individual harmonic voltage distortion rate; b) obtaining historical operation data of the distribution network, wherein the historical operation data at least includes time series data of load, connected photovoltaic output, meteorological information, node voltage and background harmonics; c) based on the historical operation data and the operation constraints defined in step a), calculating the maximum allowable distribution network photovoltaic load capacity label value corresponding to each historical moment, and generating a data set consisting of a historical state time series and its corresponding photovoltaic load capacity label value; d) constructing and training a BILSTM network model, using the data set generated in step c), to learn the temporal mapping relationship between the input grid state sequence containing multiple historical time steps and the HC tag value satisfying the operation constraint; e) Online real-time acquisition of power grid status data at the current time and several previous times to form a real-time input sequence; f) inputting the real-time input sequence into the BILSTM model trained in step d) to quickly output the photovoltaic carrying capacity assessment value of the distribution network at the current moment.

2. The method according to claim 1, characterized in that The voltage fluctuation measurement space constraints specifically include: (1) Voltage deviation constraint: ,in, For Node At the moment The positive sequence voltage amplitude, is the system rated voltage, is the maximum allowable voltage deviation percentage; (2) Voltage imbalance constraint: ,in, For Node At the moment The negative sequence voltage amplitude, is the maximum allowable voltage imbalance; (3) Voltage fluctuation amplitude constraints: ,in, To evaluate the time window length, is the maximum allowable voltage fluctuation.

3. The method according to claim 1 or 2, characterized in that: The harmonic distortion rate constraint specifically includes: (1) Total harmonic voltage distortion rate constraint: ,in, For Node At the moment of Subharmonic voltage amplitude, is the highest harmonic order considered, is the maximum allowable total harmonic voltage distortion rate; (2) Constraints on individual harmonic voltage distortion rates: ,in, For permitted The second largest individual harmonic voltage distortion rate; Furthermore, the harmonic distortion rate constraint also includes constraints on total harmonic current distortion rate and individual harmonic current distortion rate.

4. The method according to claim 1, characterized in that The method for calculating the distribution network photovoltaic load capacity tag value in step c) includes: for each historical moment , using iteration to increase the photovoltaic power to be connected In each iteration step, based on the current total photovoltaic power Perform power flow calculation and harmonic power flow calculation to check whether all monitoring nodes and lines meet all operating constraints defined in step a); if so, continue to increase ; If not, determine the maximum number of for , and As the label value of the photovoltaic carrying capacity of the distribution network at that moment.

5. The method according to claim 1, characterized in that Step d) involves building and training the BILSTM network model: The grid state vector of historical time steps The sequence of As model input, It includes normalized load, photovoltaic output, meteorology, key node voltage and harmonic indicators; the BILSTM model includes at least one bidirectional LSTM layer, which combines the hidden state of the forward LSTM output and the hidden state of the backward LSTM output; the model training adopts supervised learning method, using mean square error as the loss function and optimizer for weight update.

6. The method according to claim 1, characterized in that Step f) also includes: The actual total photovoltaic output of the current distribution network Compare and calculate the real-time available PV acceptance margin , and and / or Used for distribution network operation status monitoring, photovoltaic output control or dispatching decision-making.

7. A photovoltaic carrying capacity assessment system based on a multi-dimensional distribution network, characterized in that: The system is used to execute the method according to any one of claims 1 to 6, and the system comprises: The data acquisition and preprocessing module is used to obtain real-time distribution network operation data, including load, photovoltaic output, weather, voltage, harmonics and other information, and perform data cleaning, synchronization, and normalization to build a real-time input sequence that meets the preset length and feature dimension; A constraint definition module, used to store and manage preset voltage fluctuation measurement space constraint limits and detailed harmonic distortion rate constraint limits; A BILSTM model module stores pre-trained BILSTM model parameters, receives real-time input sequences from the data acquisition and preprocessing module, performs model inference calculations, and outputs predicted values; The distribution network photovoltaic carrying capacity evaluation and output module receives the photovoltaic carrying capacity prediction value from the BILSTM model module, processes the result, and displays the result to the user interface or sends the result to other power system applications.

8. The system according to claim 7, characterized in that The data acquisition and preprocessing module obtains data from SCADA, PMU, meteorological services, photovoltaic monitoring systems and power quality monitoring devices.

Images