Zone area photovoltaic consumption capability assessment method considering source load randomness

Abstract

A transformer area photovoltaic consumption capability assessment method considering source load randomness relates to the technical field of photovoltaic power generation, and comprises the steps of collecting transformer area topology, line impedance and other data in real time through a multi-dimensional perception layer, extracting and quantifying the source load randomness by using dynamic features, constructing a self-adaptive probability model, and assessing risks in combination with an extreme value theory. A photovoltaic access capacity strategy is optimized by applying reinforcement learning, and the feasibility of the scheme is ensured through physical-random consistency verification, so that the photovoltaic consumption capability is finally improved, and safe and stable operation of a power grid is ensured. According to the method, multi-dimensional perception, dynamic feature extraction, adaptive modeling and reinforcement learning optimization are fused, the photovoltaic consumption capability evaluation precision and the power grid operation stability are remarkably improved, and powerful support is provided for high-proportion renewable energy source access.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method for assessing the photovoltaic absorption capacity of a transformer substation that takes into account the randomness of source and load. Background Technology

[0002] With the large-scale application of distributed photovoltaic (PV) power generation technology, the amount of PV power connected to distribution transformer areas has been increasing year by year. However, PV output is affected by natural conditions such as weather and seasons, exhibiting high intermittency and volatility, resulting in strong randomness in PV output. At the same time, the carrying capacity of the power grid in the distribution transformer area is limited by physical conditions such as network topology, line impedance, and transformer capacity, thus constraining the maximum capacity for PV connection. How to accurately assess the PV absorption capacity of the distribution transformer area and rationally allocate capacity has become an urgent problem to be solved.

[0003] Traditional photovoltaic (PV) grid integration assessment methods are typically based on deterministic models, such as static power flow analysis or historical load data statistics. While these methods reduce computational complexity by simplifying assumptions, they fail to characterize the dynamic coupling relationship between sources and loads and extreme risks, resulting in insufficient assessment accuracy in scenarios with high PV integration ratios, making it difficult to meet the requirements for safe grid operation. Furthermore, existing assessment methods fail to fully consider the fusion of diverse information from transformer network topology, real-time meteorological conditions, and historical sample data, making them ill-suited for complex scenarios with rapidly changing source-load distribution, further exacerbating the complexity and uncertainty of PV grid integration capacity assessment. This deficiency is a problem that urgently needs to be addressed by those skilled in the art. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention discloses a method for evaluating the photovoltaic absorption capacity of a transformer substation that takes into account the randomness of source and load.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] A method for assessing the grid absorption capacity of a photovoltaic distribution area considering the randomness of source and load includes the following steps:

[0007] S1. Build an evaluation system to collect real-time data on transformer topology, line impedance, transformer capacity, meteorological information, load curves, and photovoltaic output at the second level through IoT terminals;

[0008] S2. Dynamic feature extraction: Calculate load fluctuation entropy Hload and photovoltaic intermittency index Ipv to quantify the source-load dual random intensity.

[0009] S3. Adaptive probabilistic modeling: Based on the rank correlation coefficient threshold of Cdyn, Gaussian-Cauchy mixed kernel density or generalized Pareto tail distribution is selected online as the marginal distribution; the Copula-Vine structure is used to capture high-dimensional nonlinear dependencies, and quantile regression forest is used to post-correct the marginal residuals to form a source-load joint stochastic model.

[0010] S4. Tail risk quantification: Pareto fitting of the right tail of the joint stochastic model is performed using extreme value theory to estimate the tail exponent ξ of the transformer backfeed power; the conditional value of risk CVaRα at a confidence level of α=0.99 is calculated, and the upper and lower bounds of the confidence interval are given.

[0011] S5. Reinforcement learning optimization to construct a digital twin environment, where the reward function is:

[0012] r=λ1·ΔPpv-λ2·CVaRα

[0013] Where ΔPpv represents the newly added photovoltaic capacity, and λ1 and λ2 are adjustable weights; the PPO algorithm is optimized using a near-end strategy, and the photovoltaic capacity strategy πθ is iteratively generated in a twin environment until the strategy entropy and the reward improvement converge;

[0014] S6. Consistency check: Map the strategy πθ to the actual power flow model and perform multi-constraint checks on N-1 static security, voltage deviation ±7%, and line thermal stability. If any check fails, the strategy is penalized in the twin environment and the process returns to step S5 for retraining.

[0015] Preferably, in step S1, the evaluation system includes:

[0016] The system comprises a multidimensional perception layer, a random feature extraction layer, an adaptive probabilistic modeling layer, a risk quantification layer, and a physical-random consistency verification layer. The multidimensional perception layer synchronously acquires the topology, line impedance, transformer rated capacity, real-time measurements, and historical sample data of the distribution substation. The random feature extraction layer performs time-frequency joint domain feature mining on historical source-load samples, outputting load fluctuation entropy, photovoltaic output intermittent index, and source-load dynamic coupling feature matrix. The adaptive probabilistic modeling layer, based on the dynamic coupling feature matrix, online selects and combines Gaussian-Cauchy mixed kernel density, generalized Pareto tail distribution, and distribution... A bitwise regression forest is used to construct a source-load joint stochastic model; a risk quantification layer is used to calculate the tail expected loss and its confidence interval of transformer backfeed power based on the joint stochastic model by introducing extreme value theory and conditional value of risk (CVaR); a reinforcement learning optimization layer is used to use the tail expected loss as a penalty term of the reward function, and to optimize the PPO algorithm using a near-end strategy, iteratively searching for photovoltaic access capacity strategies in a digital twin environment; a physical-stochastic consistency verification layer is used to map the optimization strategy back to the actual power flow model, perform voltage-thermal stability dual verification, and output the final limit absorption capacity and risk label.

[0017] Preferably, in step S1, the collected data is stored in a time series database in the form of a time-frequency joint domain, and outliers and missing values ​​are marked.

[0018] Preferably, in step S2, a source-load dynamic coupling feature matrix Cdyn is constructed, whose elements are the mutual information intensity and hysteresis correlation coefficient within the sliding window.

[0019] Preferably, in step S6, an interpretable risk-reward report is generated, including the limit capacity, CVaRα, sensitivity factors, and a list of key nodes; manual correction instructions from the scheduler are received, and the model parameters and policy network are updated in real time to achieve continuous learning and adaptive optimization.

[0020] By employing the technical solution described above, the present invention has the following beneficial effects:

[0021] This invention utilizes a multi-dimensional sensing layer to acquire real-time, second-level data on transformer topology, line impedance, transformer capacity, meteorological information, load curves, and photovoltaic output, providing high-quality foundational data for subsequent analysis. It efficiently mines source-load dynamic coupling characteristics, employing a time-frequency joint domain feature mining method to accurately capture the random variations in source and load, providing comprehensive information for modeling. An adaptive probabilistic modeling layer selects and combines multiple probabilistic models online based on the dynamic coupling feature matrix, dynamically adjusting parameters to adapt to changes in data characteristics and improve model accuracy. A risk quantification layer introduces extreme value theory and Conditional Value at Risk (CVaR) calculation to accurately assess the tail risk of transformer backfeed power, providing detailed uncertainty analysis results for optimization. A reinforcement learning optimization layer employs the Proximal Policy Optimization (PPO) algorithm to iteratively optimize photovoltaic access strategies in a digital twin environment, dynamically adjusting the strategy network to adapt to different target scenarios. A physical-random consistency verification layer maps the optimized strategy back to the actual power flow model, performing voltage-thermal stability dual verification to ensure the feasibility of the theoretical optimization scheme in the actual physical system, meeting key constraints, and improving photovoltaic absorption capacity and system operational reliability. Attached Figure Description

[0022] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0023] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] In the description of this invention, it should be noted that the terms "upper" and "lower" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0025] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0026] Example 1:

[0027] Combined with appendix Figure 1 A method for assessing the grid absorption capacity of a photovoltaic distribution area considering the randomness of source and load includes the following steps:

[0028] S1. Establish an evaluation system that uses IoT terminals to collect real-time data on transformer topology, line impedance, transformer capacity, meteorological information, load curves, and photovoltaic output at the second level. The aim is to inject all the raw data required for system operation into all subsequent modules in a single, comprehensive manner.

[0029] Specifically, the evaluation system comprises a multi-dimensional perception layer, a random feature extraction layer, an adaptive probability modeling layer, a risk quantification layer, and a physical-random consistency verification layer.

[0030] The multi-dimensional sensing layer is responsible for synchronously acquiring the topology, line impedance, transformer rated capacity, real-time measurements, and historical sample data of the distribution transformer area. To this end, an intelligent fusion terminal IFD-1000 is installed on the primary side of the transformer area. This terminal incorporates a 0.2S-class current transformer, a ±0.1% accuracy voltage sampling module, and a sampling frequency of 1Hz. On the secondary side, a carrier measurement unit PLC-MU is used to supplement the power of the end nodes. Meteorological data, including irradiance, temperature, and wind speed, is collected by a WXS-4C micro-weather station at 1-minute intervals and aggregated to the EdgeBox-5000 edge computing box via the MQTT-TLS protocol. These devices collectively constitute the multi-dimensional sensing layer, providing a real-time and complete raw data stream to subsequent layers.

[0031] A random feature extraction layer is used to perform time-frequency joint domain feature mining on historical source-load samples, outputting load fluctuation entropy, photovoltaic output intermittency index, and source-load dynamic coupling feature matrix. Edge nodes run a lightweight Python script locally: calculating Shannon entropy (Hload) for the load sequence within every 15-minute sliding window; calculating the coefficient of variation for the photovoltaic sequence and mapping it to the intermittency index (Ipv); and then constructing a 5×5 dynamic coupling feature matrix (Cdyn) using mutual information and correlation coefficients with lags of 1-5 minutes. The script execution time is less than 30ms, and the results are pushed to the adaptive probabilistic modeling layer via ZeroMQ.

[0032] The adaptive probabilistic modeling layer selects and combines Gaussian-Cauchy mixed kernel density, generalized Pareto tail distribution, and quantile regression forest online based on the dynamic coupling feature matrix to construct a source-load joint stochastic model. Edge nodes first detect the rank correlation coefficient τ of Cdyn: if |τ| is less than 0.3, the edge distribution adopts Gaussian-Cauchy mixed kernel density, and the bandwidth is adaptively determined by the Sheather-Jones formula; if |τ| is greater than or equal to 0.3 and the excess mean function of the tail samples is linear, the tail is switched to a generalized Pareto distribution, and the threshold is automatically given by the PBDH theorem of extreme value theory. The joint distribution is constructed using C-VineCopula, and node pairing is selected according to the minimum BIC criterion; the residual part is post-corrected using 500 quantile regression forests to ensure that the overall model error MAPE is less than 2%.

[0033] The risk quantification layer, based on a joint stochastic model, introduces extreme value theory and conditional value at risk (CVaR) to calculate the tail expected loss of transformer backfeed power and its confidence interval. A cloud-based GPU server performs Pareto fitting on the right tail of the constructed joint stochastic model to estimate the tail exponent ξ; subsequently, it calculates CVaRα and its 95% confidence interval at a confidence level of α=0.99. For example, when ξ=0.12 and CVaRα=14.7kVA, it indicates that the expected transformer backfeed power is 14.7kVA in the worst-case 1% scenario.

[0034] The reinforcement learning optimization layer uses the tail expectation loss as a penalty term in the reward function and employs a near-end policy to optimize the PPO algorithm, iteratively searching for photovoltaic access capacity strategies in a digital twin environment. The digital twin environment is built based on OpenDSS+Gym, with a 128-dimensional state space and an action space representing discrete capacity increments from 0 to 50kW. The reward function is r = λ1·ΔPpv - λ2·CVaRα, where λ1 and λ2 are initially 1.0 and 0.5, respectively; if CVaRα decreases by less than 1% for 20 consecutive rounds, λ2 is dynamically increased by 10%.

[0035] The PPO network is a four-layer MLP (256-256-128-64). After training for 3000 rounds, the policy entropy converges, and the reward improvement is less than 0.1%.

[0036] The Physical-Random Consistency Verification Layer maps the optimization strategy back to the actual power flow model, performs dual voltage-thermal stability verification, and outputs the final limit absorption capacity and risk label. After each strategy output, the cloud automatically calls the OpenDSS script: node voltage deviation ≤ ±7% of nominal voltage; line current ≤ 100% of long-term current carrying capacity; voltage drop after N-1 fault ≤ 10%. If any constraint is exceeded, a -100 penalty is returned to the PPO environment and retraining is performed until all constraints are met.

[0037] The collected data is stored in a time-series database in a time-frequency joint domain format, and outliers and missing values ​​are marked.

[0038] The primary bus voltage and current are connected to the 16-bit Δ-Σ ADC via the 0.2S class current transformer of the IFD-1000; the secondary end node power is transmitted back by the PLC-MU via OFDM carrier in the 2-30MHz frequency band; the WXS-4C weather station is connected to the IFD-1000's expansion port via RS-485. The ADC is oversampled at 1kHz, and the digital low-pass FIR is cut off from 50Hz to 100Hz for downsampling; the weather station's 1-minute average is packaged; all timestamps are uniformly synchronized by the IFD-1000's built-in GPS+PPS, with an error of less than 1μs. The IFD-1000 firmware has a built-in CRC32 check, and automatically retransmits when the packet loss rate is greater than 0.1%; missing time periods are guaranteed to be delivered "at least once" using MQTT-QoS2. Each data packet is JSON, with fields including timestamp, nodeid, Upu, IA, PW, QVar, GHIW / m², and Tamb℃. EdgeBox-5000 aggregates data via ZeroMQ and then transfers it to local InfluxDB shards, retaining a 90-day rolling window before automatically transferring it to cloud cold storage after 90 days.

[0039] S2. Dynamic feature extraction: Calculate the load fluctuation entropy Hload and the photovoltaic intermittency index Ipv to quantify the source-load dual stochastic intensity; construct the source-load dynamic coupling feature matrix Cdyn, whose elements are the mutual information intensity and hysteresis correlation coefficient within a sliding window. Real-time streams use a 15-minute sliding window with a 5-minute step; historical batch processing uses a 24-hour batch with a 1-hour step. Calculate the Shannon entropy Hload = −Σ(pilogpi), where pi is the probability that the normalized power falls within the i-th histogram interval.

[0040] The intermittent index Ipv for photovoltaic sequences is calculated as σ5min / μ5min, where σ and μ are the standard deviation and mean over 5 minutes. Mutual information MI(X;Y) is quickly estimated using Kraskovk-NN (k=3), with a lag τ ranging from 1 to 5 minutes, generating a 5×5 matrix Cdyn. If the missing rate of any window is greater than 5%, that window is skipped and a "SKIPFLAG" is written to the log. When the CPU usage of an edge node is greater than 80%, it is automatically downgraded to a 30-minute window.

[0041] S3. Adaptive probabilistic modeling: Based on the rank correlation coefficient threshold of Cdyn, Gaussian-Cauchy mixed kernel density or generalized Pareto tail distribution is selected online as the marginal distribution. A Copula-Vine structure is used to capture high-dimensional nonlinear dependencies, and quantile regression forest is used to post-correct the marginal residuals, forming a source-load joint stochastic model. If |τ_max| is less than 0.3, the Gaussian-Cauchy mixed kernel density is selected; if ξtail is greater than or equal to 0.1, the GPD tail is selected; otherwise, Beta-GEV is retained. For parameter estimation, the kernel density bandwidth h uses the Sheather-Jones plugin method; the GPD threshold u is jointly determined by the 95th quantile and the inflection point of the excess mean plot; Copula-Vine uses pyvinecopulib, and the minimum BIC principle is used to select C-Vine or D-Vine. If GPU resources are insufficient for the cloud training task, the edge nodes automatically use the previous version of the pickle model and write a "MODEL-STALE" alarm.

[0042] S4. Tail risk quantification: Pareto fitting of the right tail of the joint stochastic model is performed using extreme value theory to estimate the tail exponent ξ of the transformer backfeed power; the conditional value of risk CVaRα at a confidence level of α=0.99 is calculated, and the upper and lower bounds of the confidence interval are given. For samples whose right tail of the joint distribution exceeds the threshold u, ξ and β are estimated using the POT (Peak-Over-Threshold) method; the confidence interval is obtained using the ProfileLikelihood method. CVaRα=E[X|X>VaRα], where VaRα is given by the 99th quantile of the quantile regression forest; numerical integration uses a 10,000-times Sobol sequence MC.

[0043] S5. Reinforcement learning optimization to construct a digital twin environment, where the reward function is:

[0044] r=λ1·ΔPpv-λ2·CVaRα

[0045] Where ΔPpv represents the newly added photovoltaic capacity, and λ1 and λ2 are adjustable weights.

[0046] The PPO algorithm is optimized using a near-end policy, iteratively generating the photovoltaic capacity policy πθ in a twin environment until the policy entropy and reward improvement converge. The `gym.Env` subclass `PowerGridEnv` internally calls the `OpenDSS-COM` interface; the state space dimension is 128; the action space `a∈{0,2,4,…,50}kW`. The PPO algorithm is based on stable-baselines3, with 96 steps per round (representing 24h × 4 steps / h), batch size=512, n-epochs=10, and clip-range=0.2; training on the GPU takes 3000 rounds and 45 minutes. The policy entropy H is less than 0.05 and the average reward improvement is less than 0.1% for 100 consecutive rounds. If the twin environment and real-time measurement error (MAPE) are greater than 5% for 30 minutes, the λ1 coefficient is automatically reduced by 10% and training is restarted.

[0047] S6. Consistency Verification: Map the policy πθ to the actual power flow model and perform multi-constraint verifications for N-1 static security, voltage deviation ±7%, and line thermal stability. If any verification fails, a penalty is applied to the policy in the twin environment, and the process returns to step S5 for retraining. The Python post-processing script reads the CSV file. If any node voltage is less than 0.93 pu or greater than 1.07 pu, or any line current exceeds its rated value, a Boolean vector Violation = [True / False…] is returned. If Violation contains True, a penalty term of −100 is added to the reward function in the twin environment, and the process returns to step S5 for resampling and training.

[0048] Finally, an interpretable risk-reward report is generated, including the maximum capacity, CVaRα, sensitivity factors, and a list of key nodes. This method supports receiving manual correction instructions from schedulers, updating model parameters and the policy network in real time, and enabling continuous learning and adaptive optimization. Schedulers can drag α from 0.99 to 0.995 in the Web-UI; the system instantly calls a cloud script to recalculate and return the new capacity of 45.1 kVA, with the entire process taking less than 3 seconds. If the modification is manually confirmed, the cloud automatically packages a new policy.pt and model.pickle file, verifies it with HTTPS+SHA256 signatures, and sends it to EdgeBox, completing the closed-loop update.

[0049] The parts of this invention not described in detail are prior art. It will be apparent to those skilled in the art that this invention is not limited to the details of the above exemplary embodiments, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and are intended to encompass all changes falling within the meaning and scope of equivalents within this invention.

Claims

1. A method for assessing the photovoltaic absorption capacity of a distribution area considering the randomness of source and load, characterized in that, Includes the following steps: S1. Build an evaluation system to collect real-time data on transformer topology, line impedance, transformer capacity, meteorological information, load curves, and photovoltaic output at the second level through IoT terminals; S2. Dynamic feature extraction: Calculate load fluctuation entropy Hload and photovoltaic intermittency index Ipv to quantify the source-load dual random intensity. S3. Adaptive probabilistic modeling: Based on the rank correlation coefficient threshold of Cdyn, Gaussian-Cauchy mixed kernel density or generalized Pareto tail distribution is selected online as the marginal distribution; the Copula-Vine structure is used to capture high-dimensional nonlinear dependencies, and quantile regression forest is used to post-correct the marginal residuals to form a source-load joint stochastic model. S4. Tail risk quantification: Pareto fitting of the right tail of the joint stochastic model is performed using extreme value theory to estimate the tail exponent ξ of the transformer backfeed power; the conditional value of risk CVaRα at a confidence level of α=0.99 is calculated, and the upper and lower bounds of the confidence interval are given. S5. Reinforcement learning optimization to construct a digital twin environment, where the reward function is: r=λ1·ΔPpv-λ2·CVaRα Where ΔPpv represents the newly added photovoltaic capacity, and λ1 and λ2 are adjustable weights; The PPO algorithm is optimized using a near-end strategy. The photovoltaic capacity strategy πθ is iteratively generated in a twin environment until the strategy entropy and reward improvement converge. S6. Consistency check: Map the strategy πθ to the actual power flow model and perform multi-constraint checks on N-1 static security, voltage deviation ±7%, and line thermal stability. If any check fails, the strategy is penalized in the twin environment and the process returns to step S5 for retraining.

2. The method for assessing the photovoltaic absorption capacity of a distribution area considering the randomness of source and load as described in claim 1, characterized in that: In step S1, the evaluation system includes: The system comprises a multidimensional perception layer, a random feature extraction layer, an adaptive probabilistic modeling layer, a risk quantification layer, and a physical-random consistency verification layer. The multidimensional perception layer synchronously acquires the topology, line impedance, transformer rated capacity, real-time measurements, and historical sample data of the distribution substation. The random feature extraction layer performs time-frequency joint domain feature mining on historical source-load samples, outputting load fluctuation entropy, photovoltaic output intermittent index, and source-load dynamic coupling feature matrix. The adaptive probabilistic modeling layer, based on the dynamic coupling feature matrix, online selects and combines Gaussian-Cauchy mixed kernel density, generalized Pareto tail distribution, and distribution... A bitwise regression forest is used to construct a source-load joint stochastic model; a risk quantification layer is used to calculate the tail expected loss and its confidence interval of transformer backfeed power based on the joint stochastic model by introducing extreme value theory and conditional value of risk (CVaR); a reinforcement learning optimization layer is used to use the tail expected loss as a penalty term of the reward function, and to optimize the PPO algorithm using a near-end strategy, iteratively searching for photovoltaic access capacity strategies in a digital twin environment; a physical-stochastic consistency verification layer is used to map the optimization strategy back to the actual power flow model, perform voltage-thermal stability dual verification, and output the final limit absorption capacity and risk label.

3. The method for assessing the photovoltaic absorption capacity of a distribution area considering the randomness of source and load as described in claim 1, characterized in that: In step S1, the collected data is stored in the time series database in the form of a time-frequency joint domain, and outliers and missing values ​​are marked.

4. The method for assessing the photovoltaic absorption capacity of a distribution area considering the randomness of source and load as described in claim 1, characterized in that: In step S2, the source-load dynamic coupling feature matrix Cdyn is constructed, whose elements are the mutual information intensity and hysteresis correlation coefficient within the sliding window.

5. The method for assessing the photovoltaic absorption capacity of a distribution area considering the randomness of source and load as described in claim 1, characterized in that: In step S6, an interpretable risk-reward report is generated, including the limit capacity, CVaRα, sensitivity factors, and a list of key nodes; manual correction instructions from the scheduler are received, and the model parameters and policy network are updated in real time to achieve continuous learning and adaptive optimization.

Images