Power distribution network load forecasting and dynamic stability system based on big data simulation analysis

Abstract

The application provides a power distribution network load prediction and dynamic stability system based on big data simulation analysis, and relates to the technical field of power distribution network load prediction. The system comprises a data acquisition and processing module, a big data simulation analysis module and an intelligent regulation and control module. The data acquisition and processing module is used for acquiring power distribution network operation data and pre-processing the power distribution network operation data. The big data simulation analysis module performs load prediction by using a multi-model training selection mechanism based on the power distribution network operation data, and performs dynamic stability risk assessment based on the load prediction result. The intelligent regulation and control module establishes an optimization model with power supply reliability and economy as the target according to the risk assessment result, and performs collaborative regulation and control on the distributed power supply, flexible load and energy storage equipment according to the optimal solution of the optimization model.

Description

Technical Field

[0001] This invention relates to the field of power distribution network load forecasting technology, and in particular to a power distribution network load forecasting and dynamic stabilization system based on big data simulation analysis. Background Technology

[0002] The intermittent, random, and fluctuating nature of distributed energy output, along with the integration of numerous flexible loads, weakens the real-time observability of the distribution network, making it difficult to accurately monitor its operational status. This leads to problems such as line overload, node overvoltage, and degraded power quality, which seriously threaten the safe and stable operation of the distribution network.

[0003] Currently, the operation and management of power distribution networks mainly face the following technical challenges:

[0004] Data quality and integrity issues: In actual operation, incomplete measurement data is common due to incomplete coverage of measurement devices, communication interruptions, or noise interference. This makes it difficult for traditional state estimation methods that rely on comprehensive and accurate measurement data to meet accuracy requirements. Traditional "human experience" prediction models face prominent problems such as poor accuracy and low efficiency.

[0005] Insufficient prediction accuracy: Existing load forecasting methods are mostly based on a single model or limited data dimensions, which have weak prediction capabilities for scarce scenarios such as extreme weather and holidays, and cannot adapt to environments with high uncertainty on both the source and load sides.

[0006] Staticization of stability assessment: Most stability assessment methods are based on steady-state or deterministic boundary conditions, lacking the ability to fine-tune modeling and rolling correction of uncertainties in transient processes. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a distribution network load forecasting and dynamic stability system based on big data simulation analysis. By integrating multi-source heterogeneous data, it achieves high-precision load forecasting across multiple temporal and spatial scales. Based on the forecast results and real-time data, it dynamically assesses system stability risks and ultimately generates and executes collaborative control strategies to ensure the safe, stable, and efficient operation of the distribution network under high-proportion renewable energy access.

[0008] On the one hand, the present invention provides a power distribution network load forecasting and dynamic stability system based on big data simulation analysis, including: a data acquisition and processing module, a big data simulation analysis module, and an intelligent control module;

[0009] The data acquisition and processing module is used to acquire power distribution network operation data and preprocess the power distribution network operation data;

[0010] The big data simulation and analysis module uses distribution network operation data and a multi-model training and selection mechanism to perform load forecasting, and performs dynamic stability risk assessment based on the load forecasting results.

[0011] The intelligent control module establishes an optimization model based on the risk assessment results, with the goal of power supply reliability and economy, and performs coordinated control of distributed power sources, flexible loads and energy storage devices based on the optimal solution of the optimization model.

[0012] The data acquisition and processing module specifically includes a multi-dimensional data source acquisition unit and a data preprocessing unit; the multi-dimensional data source acquisition unit is used to acquire distribution network data from multi-dimensional data sources; the distribution network data includes power source side data, network side data, load side data, energy storage side data, and environmental and market side data; the data preprocessing unit is used to preprocess the acquired distribution network data;

[0013] The data preprocessing unit includes a heterogeneous data unification subunit, a multi-layer anomaly detection subunit, and an intelligent repair subunit. The heterogeneous data unification subunit is used to perform time-series alignment, spatial correlation, and unified encoding on the distribution network operation data. The multi-layer anomaly detection subunit is used to detect anomalies in the unified-encoded distribution network operation data. The intelligent repair subunit is used to repair abnormal data points.

[0014] The time alignment involves interpolating or aggregating data from different frequencies to unify them onto a reference time axis; spatial association involves establishing topological connections between data points; and unified coding involves generating a globally unique code for each data point.

[0015] The multi-layer anomaly detection subunit employs a three-layer anomaly detection mode to detect anomalies in the data. Specifically, the first-layer anomaly detection mode filters data by setting and statistically detecting outliers and constraining the rate of change. The physical limit rule is set to ±10% of the rated voltage. A sliding window is used to identify statistical outliers at abrupt changes. The rate of change of load in adjacent time periods exceeds a physically possible threshold. The second-layer anomaly detection mode uses the DBSCAN clustering algorithm, considering temporal continuity constraints, clustering data from similar days and time periods into one class, and marking isolated points or small clusters of data as anomalies. The third-layer anomaly detection mode uses an LSTM-Autoencoder reconstruction model. LSTM-Autoencoder is a long short-term memory autoencoder. The LSTM-Autoencoder reconstruction model includes a training phase, a detection phase, and a decision mechanism. During the training phase, historical normal data is used to train the autoencoder to learn the normal patterns of load, voltage, active power, and reactive power data. During the detection phase, real-time data is input, and the reconstruction error is calculated using the following formula: Where E represents the reconstruction error, and X represents the normalized real-time input data vector. This represents the reconstructed data vector generated by the LSTM-Autoencoder decoder. The decision mechanism is: if... If so, the input real-time data is determined to be abnormal; where The threshold is dynamic and determined based on historical error distribution; abnormal data is detected through a three-layer anomaly detection mode.

[0016] The intelligent repair subunit initiates a repair process for data points identified as abnormal. This process includes historical similar day matching repair, higher-order spatial interpolation repair, and generative repair. Specifically, historical similar day matching repair is as follows:

[0017] Similarity calculation: Based on multi-dimensional features such as date type, weather conditions, and holiday attributes, calculate the similarity between the current day and historical days. Similarity represents the overall similarity score between the current predicted date and historical sample dates, and w1, w2, and w3 correspond to the weight coefficients of date type, weather conditions, and holiday attributes, respectively, satisfying w1 + w2 + w3 = 1; S date S represents the date type similarity component. weather S represents the weather condition similarity component. holiday This represents the similarity component of holiday attributes;

[0018] Candidate date selection: Select the top five historical dates with the highest similarity as candidate dates;

[0019] Pattern extraction: Extract normal data from candidate days within the same time period to repair outliers;

[0020] When single-point data is missing but adjacent monitoring points are normal, it is repaired by high-order spatial interpolation. High-order spatial interpolation is based on the Kriging interpolation method with topological constraints. First, a distribution network impedance model is established and the electrical distance between each node is calculated. The estimated value of the anomaly point is obtained by weighting the adjacent nodes, and the weight is inversely proportional to the electrical distance. Independent interpolation is performed on the three phases A, B and C respectively.

[0021] Otherwise, generative repair is adopted: generative repair uses a conditional generative adversarial network (CGAN); the CGAN includes a generator, a discriminator, and a physical constraint layer; the generator's input is normal data segments before and after the abnormal period and external conditions, including weather, date, and holidays; the discriminator's input is the data generated by the generator and real historical data segments; the physical constraint layer generates data by simplifying power flow verification.

[0022] The big data simulation and analysis module includes a load prediction unit and a dynamic stability simulation unit. The load prediction unit performs load prediction based on the distribution network operation data preprocessed by the data acquisition and processing module and adopts a multi-model training and selection mechanism. The dynamic stability simulation unit uses the load prediction results as boundary conditions and inputs them into the hybrid simulation environment to generate fault scenarios. It then uses the hybrid simulation environment to output stability risk assessment indicators for each fault scenario.

[0023] The load prediction unit includes a feature vector transformation subunit, a data augmentation subunit, a multi-model training and selection subunit, and a probabilistic prediction subunit; the feature vector transformation subunit is used to transform distribution network operation data into feature vectors.

[0024] The data augmentation subunit is used to augment samples using generative adversarial networks (GANs) and temporal local compression when historical data is scarce.

[0025] The multi-model training and selection subunit is used to construct and train multiple prediction models and dynamically select multiple prediction models for load prediction; specifically, it includes heterogeneous model construction and a two-level selection mechanism.

[0026] The heterogeneous model construction is used to build a CNN-LSTM fusion model, a Transformer temporal model, and a gradient boosting tree combined model. The CNN-LSTM fusion model includes CNN layers and LSTM layers. The CNN layers are used to extract local spatial correlations of multivariate sequences, and the LSTM layers are used to capture long-term temporal dependencies. The Transformer temporal model uses a self-attention mechanism to explicitly model the dependencies between any two time points in the load sequence, which is used to handle long-period irregular events. The gradient boosting tree combined model is used to learn complex nonlinear relationships in multidimensional structured features.

[0027] The two-level selection mechanism selects the optimal model or model combination for prediction under different scenarios; the selection method of the two-level selection mechanism includes the following steps:

[0028] The first-level selection mechanism is to divide the historical data into multiple sub-validation sets according to different seasons and weather types; each model is validated on these sub-validation sets, and its scene adaptability profile is recorded.

[0029] Second-level selection mechanism: During runtime, upon receiving the latest data, the following steps are performed:

[0030] Scene matching: Based on current and forecast weather and date information, match the most similar historical scenes;

[0031] Initial model selection: Based on the scene adaptability profile, recall the top five models that have historically performed best in this type of scene;

[0032] Rolling evaluation: Using the most recent real data as a mini test set, the initial model is re-evaluated in real time, and the rolling MAPE is calculated.

[0033] Dynamic weighting or optimization: Select the single model with the best performance as the predictor for this round, or use a variable weighting method based on real-time performance to perform model combination prediction;

[0034] The probabilistic prediction subunit is used to output load prediction intervals. Specifically, model selection is based on the time scale. For ultra-short-term time scales, a CNN-LSTM fusion model is used, outputting minute-level point predictions and probability distributions based on real-time meteorological data and very short-term trends. For short-term time scales, a Transformer time series model or model combination is used, considering weekly cycles and weather forecasts, outputting hourly-level load curves and confidence intervals. For medium- and long-term time scales, a gradient boosting tree combination model is used, combining calendar and climate predictions to output interval estimates of daily maximum or minimum loads. For each selected model, Monte Carlo Dropout is used for forward propagation to obtain prediction samples, and then the prediction interval is calculated.

[0035] The dynamic stability simulation unit uses load forecast results as boundary conditions and inputs them into the hybrid simulation environment to generate fault scenarios. For each fault scenario, the hybrid simulation environment outputs stability risk assessment indicators and uses all stability risk assessment indicators to perform stability risk assessment. The hybrid simulation environment integrates a surrogate model based on neural networks and a distribution network electromagnetic transient simulation model. The stability risk assessment is the probability of overload of key equipment and voltage exceeding limits of nodes in different future time periods. Specifically, the hybrid simulation environment features a dual-channel hybrid simulation setup: a physical channel, which includes a power distribution network electromagnetic transient simulation model, and a proxy channel, which includes a neural network proxy model. The power distribution network electromagnetic transient simulation model provides training data for the proxy model and verifies its key outputs. The neural network proxy model receives verification and updates from the physical channel, selecting scenarios for the physical channel. The simulation channel is selected based on factors such as the degree of topology change, the proportion of nonlinear devices, the simulation time span, the penetration rate of new energy sources, the system short-circuit ratio, the proxy model's confidence level, and the deviation from the operating conditions. The neural network proxy model includes a GAT encoding layer, a spatiotemporal Transformer layer, and a physical information regularization layer. The GAT encoding layer models the power distribution network topology as a graph structure, learning the embedded representations of nodes and edges to capture electrical coupling relationships. The spatiotemporal Transformer layer handles time-varying inputs, capturing both temporal dependencies and spatial interactions. The physical information regularization layer adds a penalty term based on Kirchhoff's laws to the loss function, ensuring that the model output approximately satisfies fundamental physical laws.

[0036] The stability risk assessment is determined by the stability risk assessment index, R. Z Calculated by the following formula:

[0037] ;

[0038] Where Z represents the set of power grid areas for which stability risk assessment is conducted; and These are the weighting coefficients. For line or transformer L i The probability that the load rate exceeds its safety limit, i.e., the risk of equipment overload; For node V i The probability that the voltage deviates from its acceptable range;

[0039] The overload risk of the equipment is calculated by the following formula:

[0040] ;

[0041] in, For line L i The probability of overload in the future time period [t1, t2]; This represents the total number of scenes; The probability that the load rate of the power grid under test exceeds the set value at any simulation time;

[0042] The intelligent control module includes a multi-objective optimization model construction unit and a collaborative control unit. The multi-objective optimization model construction unit is used to construct the power supply reliability risk cost objective function and the system operation total cost objective function. The collaborative control unit obtains the Pareto optimal solution set by solving the power supply reliability risk cost objective function and the system operation total cost objective function, calculates the satisfaction based on the Pareto optimal solution set, and performs collaborative control of distributed power sources, flexible loads, and energy storage devices based on the comprehensive satisfaction.

[0043] The objective function for power supply reliability risk cost is expressed by the following formula:

[0044] ;

[0045] in, The risk value coefficient of equipment i in time period t; The real-time risk index for device i; The value of power shortage for load j; The load reduction is T, which is the total time length, i.e., one test cycle, and Ω. risk Ω represents the set of all devices within the system that participate in the operational risk assessment; load This represents the set of all load nodes or power users in the system;

[0046] The objective function for the total operating cost of the system is expressed by the following formula:

[0047] ;

[0048] in, The active power purchased from the upper-level power grid during time period t; The electricity purchase price for time period t; Cost of distributed power generation; Active power output for distributed power sources; A collection of distributed power sources; For energy storage cycle costs; Let m be the charging and discharging power of the m-th energy storage system during time period t; A collection of energy storage systems; Cost of switching operation; Reconfigure the network switch state; A collection of operable switches;

[0049] The coordinated control unit includes the following steps:

[0050] Normalization: Normalize the objective function values ​​for power supply reliability risk cost and total system operating cost.

[0051] ;

[0052] in, The actual calculated value of the i-th objective function, where i is 1 or 2; Let be the minimum value of the i-th objective function in the current Pareto solution set; This represents the maximum value of the i-th objective function in the current Pareto solution set;

[0053] Satisfaction level calculation: using an S-shaped membership function.

[0054] ;

[0055] in, For target satisfaction; The target value is the normalization target value; is the steepness coefficient, a parameter that controls the steepness of the S-curve; The center point parameter is the normalized target value corresponding to a satisfaction level of 0.5, which is set according to scheduling preferences.

[0056] Comprehensive selection: Choose the solution with the highest overall satisfaction.

[0057] ;

[0058] in, and These represent satisfaction with reliability risks and economic efficiency, respectively. and They are respectively and The target weight.

[0059] The beneficial effects of adopting the above technical solution are as follows:

[0060] This invention provides a distribution network load forecasting and dynamic stabilization system based on big data simulation analysis. It collects distribution network data from multiple dimensions and employs a multi-model training and selection mechanism for load forecasting, effectively overcoming the problems of data scarcity and single-model limitations, and providing a reliable basis for stability assessment and control. Furthermore, by constructing an optimization model with power supply reliability and economy as objectives, the optimal solution of the optimization model is used to coordinate the control of distributed power sources, flexible loads, and energy storage devices. This enables proactive intervention before faults or risks occur, maximizing the recovery of critical loads and reducing outage time and scope. Simultaneously, optimized scheduling reduces network losses, improves the absorption capacity of new energy sources, and achieves a balance between safety and economy. Attached Figure Description

[0061] Figure 1 Block diagram of the distribution network load prediction and dynamic stability system according to an embodiment of the present invention. Detailed Implementation

[0062] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0063] Example 1:

[0064] On the one hand, this invention provides a power distribution network load forecasting and dynamic stability system based on big data simulation analysis, such as... Figure 1 As shown, it includes: a data acquisition and processing module, a big data simulation and analysis module, and an intelligent control module;

[0065] The data acquisition and processing module is used to acquire power distribution network operation data and preprocess the power distribution network operation data;

[0066] The big data simulation and analysis module uses distribution network operation data and a multi-model training and selection mechanism to perform load forecasting, and performs dynamic stability risk assessment based on the load forecasting results.

[0067] The intelligent control module establishes an optimization model based on the risk assessment results, with the goal of power supply reliability and economy, and performs coordinated control of distributed power sources, flexible loads and energy storage devices based on the optimal solution of the optimization model.

[0068] The data acquisition and processing module specifically includes a multi-dimensional data source acquisition unit and a data preprocessing unit; the multi-dimensional data source acquisition unit is used to acquire distribution network data from multi-dimensional data sources; the distribution network data includes power source side data, network side data, load side data, energy storage side data, and environmental and market side data; the data preprocessing unit is used to preprocess the acquired distribution network data;

[0069] In this embodiment, the data acquisition and processing module specifically collects and manages multi-dimensional time-series operation data of power generation, grid, load, and storage equipment in the distribution network to construct a high-quality basic dataset. The management includes: automatically identifying data anomalies using a time-series anomaly detection algorithm and performing intelligent repair based on historical similar days and high-order interpolation algorithms; the data management process is based on a fully domestically produced software and hardware technology architecture to achieve end-to-end security and controllability.

[0070] In this embodiment, the power supply side data includes distributed photovoltaic and wind power active / reactive power output, inverter status, and power generation forecasts. Network side data includes P, Q, node voltage, current, switch status, transformer load rate, and fault recordings. Load side data includes industrial / commercial / residential load curves, interruptible load capacity, and temperature-controlled load characteristics. Energy storage side data includes energy storage SOC, charge / discharge power, health status, and cycle count. Environmental and market data includes meteorological data such as temperature, humidity, irradiance, electricity price signals, and demand response commands.

[0071] The data preprocessing unit includes a heterogeneous data unification subunit, a multi-layer anomaly detection subunit, and an intelligent repair subunit. The heterogeneous data unification subunit is used to perform time-series alignment, spatial correlation, and unified encoding on the distribution network operation data. The multi-layer anomaly detection subunit is used to detect anomalies in the unified-encoded distribution network operation data. The intelligent repair subunit is used to repair abnormal data points.

[0072] The time alignment involves interpolating or aggregating data from different frequencies to unify them onto a reference time axis; spatial association involves establishing topological connections between data points; and unified coding involves generating a globally unique code for each data point.

[0073] The multi-layer anomaly detection subunit employs a three-layer anomaly detection mode to detect anomalies in the data. Specifically, the first-layer anomaly detection mode filters data by setting and statistically detecting outliers and constraining the rate of change. The physical limit rule is set to ±10% of the rated voltage. A sliding window is used to identify statistical outliers at abrupt changes. The rate of change of load in adjacent time periods exceeds a physically possible threshold. The second-layer anomaly detection mode uses the DBSCAN clustering algorithm, considering temporal continuity constraints, clustering data from similar days and time periods into one class, and marking isolated points or small clusters of data as anomalies. The third-layer anomaly detection mode uses an LSTM-Autoencoder reconstruction model. LSTM-Autoencoder is a long short-term memory autoencoder. The LSTM-Autoencoder reconstruction model includes a training phase, a detection phase, and a decision mechanism. During the training phase, historical normal data is used to train the autoencoder to learn the normal patterns of load, voltage, active power, and reactive power data. During the detection phase, real-time data is input, and the reconstruction error is calculated using the following formula: Where E represents the reconstruction error, which measures the degree of difference between the input data and the data reconstructed by the model, and X represents the normalized real-time input data vector (including time-series features such as load and voltage). This represents the reconstructed data vector generated by the LSTM-Autoencoder decoder. The decision mechanism is: if... If so, the input real-time data is determined to be abnormal; where The threshold is dynamic and determined based on historical error distribution; abnormal data is detected through a three-layer anomaly detection mode.

[0074] The intelligent repair subunit initiates a repair process for data points identified as abnormal. This process includes historical similar day matching repair, higher-order spatial interpolation repair, and generative repair. Specifically, historical similar day matching repair is as follows:

[0075] Similarity calculation: Based on multi-dimensional features such as date type, weather conditions, and holiday attributes, calculate the similarity between the current day and historical days. Similarity represents the overall similarity score between the current prediction date (date to be predicted) and historical sample dates; a higher value indicates that the features of the two are more similar. w1, w2, and w3 correspond to the weighting coefficients of date type, weather conditions, and holiday attributes, respectively, used to measure the magnitude of the impact of different features on the load, satisfying w1 + w2 + w3 = 1; S date This represents the date type similarity component, used to assess the degree of matching between weekday types (such as Monday to Sunday) or weekday / rest day attributes. weather The weather similarity component is calculated based on the Euclidean distance or correlation coefficient of meteorological data such as temperature, humidity, and rainfall.holiday This represents the similarity component of holiday attributes, used to measure whether two events are both specific holidays such as Spring Festival and National Day, or whether their holiday adjustment patterns are consistent.

[0076] Candidate date selection: Select the top five historical dates with the highest similarity as candidate dates;

[0077] Pattern extraction: Extract normal data from candidate days within the same time period to repair outliers;

[0078] When single-point data is missing but adjacent monitoring points are normal, it is repaired by high-order spatial interpolation. High-order spatial interpolation is based on the Kriging interpolation method with topological constraints. First, a distribution network impedance model is established and the electrical distance between each node is calculated. The estimated value of the anomaly point is obtained by weighting the adjacent nodes, and the weight is inversely proportional to the electrical distance. Independent interpolation is performed on the three phases A, B and C respectively.

[0079] Otherwise, for continuous abnormal segments of data loss over long periods, such as communication interruptions lasting several hours, generative repair is employed. Generative repair uses a conditional generative adversarial network (CGAN). The CGAN consists of a generator, a discriminator, and a physical constraint layer. The generator's input consists of normal data segments before and after the abnormal period, as well as external conditions, including weather, date, and holidays. The discriminator's input consists of data generated by the generator and segments of real historical data. The physical constraint layer generates data by simplifying power flow verification to ensure numerical balance and reasonableness.

[0080] The big data simulation and analysis module includes a load prediction unit and a dynamic stability simulation unit. The load prediction unit performs load prediction based on the distribution network operation data preprocessed by the data acquisition and processing module and adopts a multi-model training and selection mechanism. The dynamic stability simulation unit uses the load prediction results as boundary conditions and inputs them into the hybrid simulation environment to generate fault scenarios. It then uses the hybrid simulation environment to output stability risk assessment indicators for each fault scenario.

[0081] The load prediction unit includes a feature vector transformation subunit, a data augmentation subunit, a multi-model training and selection subunit, and a probabilistic prediction subunit. The feature vector transformation subunit is used to transform distribution network operation data into feature vectors. In this embodiment, the feature vector transformation subunit specifically performs the following steps:

[0082] Data standardization and cleaning: For measurement data, such as power, voltage, and current, sliding window Z-score standardization is used to eliminate the influence of dimensions and retain real-time statistical characteristics. For meteorological data, such as temperature, humidity, and irradiance, physical threshold verification and interpolation are performed. For example, when the temperature value exceeds the historical extreme range, collaborative correction based on spatially adjacent meteorological station data is initiated.

[0083] Multidimensional feature construction: This involves transforming the processed data into multidimensional feature vectors. A multidimensional feature vector mainly consists of the following parts:

[0084] Intrinsic characteristics of time series: lag values, moving statistics, time series differences, and periodicity.

[0085] External driving characteristics: refined meteorological factors, date type, and macroeconomic indicators.

[0086] Domain knowledge features: generated by embedding scheduler experience and physical laws, including:

[0087] Air conditioning / heating degree-day characteristics: Based on the building thermal inertia model, temperature is converted into equivalent cooling / heating load demand.

[0088] Resumption of work and production index: constructed based on the load recovery curves of specific industry users such as industrial parks and commercial centers.

[0089] Theoretical output of distributed photovoltaic power: calculated based on geographical coordinates, tilt angle and irradiance data, serving as a reference for the decomposition of net load value.

[0090] The data augmentation subunit is used to augment samples in scenarios with scarce historical data, such as extreme heat waves, cold waves, and power supply for major events. It utilizes Generative Adversarial Networks (GANs) and temporal local compression. In this embodiment, the criteria are set as "daily load growth rate > 10%" or "recent maximum load surge > 15%". Specifically, it includes:

[0091] Scarcity scene identification and labeling: Using the isolated forest algorithm and a rule-based approach: temperature > 95th quantile and load growth rate > threshold, scarce scene fragments are screened out from historical data; the threshold is the critical judgment value for determining whether the load has undergone extreme changes, and the historical data can be actual operating data or publicly available sample datasets;

[0092] Controllable Generative Adversarial Network (GAN) Synthesis: The generator's input consists of a weather condition vector, a date type vector, and random noise that correspond to the scenario, and its output is the synthesized load curve; the discriminator calculates whether the net load corresponding to the synthesized load curve satisfies the coarse constraints of the power flow equation of the distribution network and determines the authenticity of the data.

[0093] Post-validation of generated data: Combine the composite load curve with the corresponding network topology to perform fast DC power flow calculation; if the composite load causes the power of any preset critical line to exceed the physical possibility range, the data will be discarded or corrected.

[0094] The multi-model training and selection subunit is used to construct and train multiple prediction models and dynamically select multiple prediction models for load prediction; specifically, it includes heterogeneous model construction and a two-level selection mechanism.

[0095] The heterogeneous model construction is used to build a CNN-LSTM fusion model, a Transformer time series model, and a gradient boosting tree combined model. The CNN-LSTM fusion model includes CNN layers and LSTM layers. The CNN layers are used to extract local spatial correlations in multivariate sequences such as load and weather data, including synchronous changes in load across different regions. The LSTM layers are used to capture long-term temporal dependencies. The CNN-LSTM fusion model has a strong ability to capture daily cycles and abrupt weather changes. The Transformer time series model utilizes a self-attention mechanism to explicitly model the dependencies between any two time points in the load sequence, handling complex patterns with long cycles such as weeks and months, as well as irregular events such as holidays. The gradient boosting tree combined model is used to learn complex nonlinear relationships in multidimensional structured features.

[0096] The two-level selection mechanism selects the optimal model or model combination for prediction under different scenarios; the selection method of the two-level selection mechanism includes the following steps:

[0097] The first-level selection mechanism is to divide the historical data into multiple sub-validation sets according to different seasons and weather types; each model is validated on these sub-validation sets and its scene adaptability profile is recorded; in this embodiment, the MAPE of the Transformer temporal model in the holiday scene is 3.5%, and the MAPE of the CNN-LSTM fusion model in the thunderstorm weather is 2.8%.

[0098] Second-level selection mechanism: During runtime, upon receiving the latest data, the following steps are performed:

[0099] Scene matching: Based on current and forecast weather and date information, match the most similar historical scenes;

[0100] Initial model selection: Based on the scene adaptability profile, recall the top five models that have historically performed best in this type of scene;

[0101] Rolling evaluation: Using the most recent real data as a mini test set, the initial model is re-evaluated in real time, and the rolling MAPE is calculated.

[0102] Dynamic weighting or optimization: Select the single model with the best performance as the predictor for this round, or use a variable weighting method based on real-time performance to perform model combination prediction;

[0103] The probabilistic prediction subunit is used to output load prediction intervals, providing probabilistic input for subsequent stability risk assessment. Specifically, model selection is based on the time scale; for ultra-short-term time scales, a CNN-LSTM fusion model is used, outputting minute-level point predictions and probability distributions based on real-time meteorological data and very short-term trends; for short-term time scales, a Transformer time series model or model combination is used, considering weekly cycles and weather forecasts, outputting hourly-level load curves and confidence intervals; for medium- and long-term scales, a gradient boosting tree combination model is used, combining calendar and climate predictions to output interval estimates of daily maximum or minimum loads. For each selected model, Monte Carlo Dropout is used for forward propagation to obtain prediction samples, and then the prediction interval is calculated.

[0104] The dynamic stability simulation unit uses load forecast results as boundary conditions and inputs them into the hybrid simulation environment to generate fault scenarios. For each fault scenario, the hybrid simulation environment outputs stability risk assessment indicators and uses all stability risk assessment indicators to perform stability risk assessment. The hybrid simulation environment integrates a surrogate model based on neural networks and a distribution network electromagnetic transient simulation model. The stability risk assessment is the probability of overload of key equipment and voltage exceeding limits of nodes in different future time periods. Specifically, the hybrid simulation environment is equipped with a dual-channel hybrid simulation: a physical channel, which includes a power distribution network electromagnetic transient simulation model, and a proxy channel, which includes a neural network proxy model. The power distribution network electromagnetic transient simulation model provides training data for the proxy model and verifies its key outputs. The neural network proxy model receives verification and updates from the physical channel, which is used to select scenarios for the physical channel. The simulation channel is selected based on factors such as the degree of topology change, the proportion of nonlinear devices, the simulation time span, the penetration rate of new energy sources, the system short-circuit ratio, the confidence level of the proxy model, and the deviation of the operating conditions. Among these factors, the penetration rate of new energy sources is the proportion of power electronic devices, the system short-circuit ratio represents the strength of the power grid, the confidence level of the proxy model is the prediction variance, and the deviation of the operating conditions is the characteristic distribution offset index. The neural network proxy model includes a GAT encoding layer, a spatiotemporal Transformer layer, and a physical information regularization layer. The GAT encoding layer is used to model the power distribution network topology as a graph structure, learn the embedded representations of nodes and edges, and capture electrical coupling relationships. The spatiotemporal Transformer layer is used to handle time-varying inputs, such as fluctuating loads and new energy sources, while capturing the dependencies in the time dimension and the interactions in the spatial dimension; the physical information regularization layer is used to add a penalty term based on Kirchhoff's laws to the loss function to ensure that the model output approximately satisfies the basic physical laws.

[0105] The stability risk assessment is determined by the stability risk assessment index, R. Z Calculated by the following formula:

[0106] ;

[0107] Where Z represents the power grid area for which stability risk assessment is conducted; and These are weighting coefficients, L for the line or transformer, respectively. i The weighting coefficient for the probability that the load rate exceeds its safety limit, and node V i The weighting coefficient for the probability that the voltage deviates from its acceptable range; For line or transformer L i The probability that the load rate exceeds its safety limit, i.e., the risk of equipment overload; For node V i The probability of a node voltage deviating from its acceptable range, i.e., the risk of node voltage exceeding the limit: calculate the probability of a node voltage exceeding the safe range (e.g., 0.95 pu-1.05 pu) and distinguish between high voltage risk (when photovoltaic backfeeds) and low voltage risk (when under heavy load).

[0108] The overload risk of the equipment is calculated by the following formula:

[0109] ;

[0110] in, For line L i The probability of overload in the future time period [t1, t2]; This represents the total number of scenes; To determine the probability that the load rate of the power grid under test exceeds the set value at any simulation time, this embodiment sets the number of scenarios to 105%.

[0111] The intelligent control module includes a multi-objective optimization model construction unit and a collaborative control unit. The multi-objective optimization model construction unit is used to construct the power supply reliability risk cost objective function and the system operation total cost objective function. The collaborative control unit obtains the Pareto optimal solution set by solving the power supply reliability risk cost objective function and the system operation total cost objective function, calculates the satisfaction based on the Pareto optimal solution set, and performs collaborative control of distributed power sources, flexible loads, and energy storage devices based on the comprehensive satisfaction.

[0112] The objective function for power supply reliability risk cost is expressed by the following formula:

[0113] ;

[0114] in, The risk value coefficient of equipment i in time period t; The real-time risk index for device i; The value of power shortage for load j; Ωrisk represents the load reduction amount, T represents the total time length, i.e., one test cycle, Ωrisk represents the set of all devices in the system that participate in the operational risk assessment, and Ωload represents the set of all load nodes or power users in the system.

[0115] The objective function for the total operating cost of the system is expressed by the following formula:

[0116] ;

[0117] in, The active power purchased from the upper-level power grid during time period t; The electricity purchase price for time period t; Cost of distributed power generation; Active power output for distributed power sources; A collection of distributed power sources; For energy storage cycle costs; Let m be the charging and discharging power of the m-th energy storage system during time period t; A collection of energy storage systems; Cost of switching operation; Reconfigure the network switch state; A collection of operable switches;

[0118] The coordinated control unit includes the following steps:

[0119] Normalization: Normalize the objective function values ​​for power supply reliability risk cost and total system operating cost.

[0120] ;

[0121] in, The actual calculated value of the i-th objective function, where i is 1 or 2; Let be the minimum value of the i-th objective function in the current Pareto solution set; This represents the maximum value of the i-th objective function in the current Pareto solution set;

[0122] Satisfaction level calculation: using an S-shaped membership function.

[0123] ;

[0124] in, For target satisfaction; The target value is the normalization target value; is the steepness coefficient, a parameter that controls the steepness of the S-curve; The center point parameter is the normalized target value corresponding to a satisfaction level of 0.5, which is set according to scheduling preferences.

[0125] Comprehensive selection: Choose the solution with the highest overall satisfaction.

[0126] ;

[0127] in, and These represent satisfaction with reliability risks and economic efficiency, respectively. and They are respectively and The target weight.

[0128] After selecting the solution with the highest overall satisfaction, the system will use this as a benchmark and carry out coordinated control according to the following logic:

[0129] 1. Command issuance and execution: Transform the optimal solution into physical scheduling commands to regulate unit output, energy storage actions, and flexible equipment status resources.

[0130] 2. Balancing risk and economy: By responding to target weight configuration and proactively adjusting the transfer flow, the risk of exceeding limits of critical equipment (the set of Ωrisk) is reduced, thereby minimizing the economic penalty of user-side load reduction (the set of Ωload).

[0131] 3. Rolling Status Verification: As instructions are executed, the system status is continuously refreshed; when the operating characteristics change abruptly or deviate, the "proxy model-physical simulation" dual channels will be re-linked for verification to ensure the reliability of the system under complex operating conditions.

[0132] Example 2:

[0133] In this embodiment, based on the aforementioned distribution network load prediction and dynamic stability system based on big data simulation analysis, the following method is implemented, including the following steps:

[0134] S1: Collect and process multi-dimensional time-series operation data of power sources, grids, loads, and storage devices in the distribution network to build a high-quality basic dataset;

[0135] S2: Based on the aforementioned basic dataset, load prediction is performed using data augmentation techniques that integrate domain knowledge and a multi-model parallel training and selection mechanism at multiple spatiotemporal scales.

[0136] S3: The load forecast results are used as boundary conditions and input into a hybrid simulation environment that couples the data-driven model and the physical mechanism model to perform probabilistic simulation and risk assessment of dynamic stability.

[0137] S4: Based on the risk assessment results, establish an optimization model with the goals of power supply reliability and economy, solve and generate coordinated control for distributed power sources, flexible loads and energy storage devices.

[0138] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0139] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0140] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of the methods disclosed herein and their equivalents, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A power distribution network load forecasting and dynamic stability system based on big data simulation analysis, characterized in that, include: Data acquisition and processing module, big data simulation and analysis module, and intelligent control module; The data acquisition and processing module is used to acquire power distribution network operation data and preprocess the power distribution network operation data; The big data simulation and analysis module uses distribution network operation data and a multi-model training and selection mechanism to perform load forecasting, and performs dynamic stability risk assessment based on the load forecasting results. The big data simulation and analysis module includes a load prediction unit and a dynamic stability simulation unit. The load prediction unit performs load prediction based on the distribution network operation data preprocessed by the data acquisition and processing module and adopts a multi-model training and selection mechanism. The dynamic stability simulation unit uses the load prediction results as boundary conditions and inputs them into the hybrid simulation environment to generate fault scenarios. It then uses the hybrid simulation environment to output stability risk assessment indicators for each fault scenario. The load prediction unit includes a feature vector transformation subunit, a data augmentation subunit, a multi-model training and selection subunit, and a probabilistic prediction subunit. The feature vector conversion subunit is used to convert power distribution network operation data into feature vectors; The data augmentation subunit is used to augment samples using generative adversarial networks (GANs) and temporal local compression when historical data is scarce. The multi-model training and selection subunit is used to construct and train multiple prediction models and dynamically select multiple prediction models for load prediction. The multi-model training and selection subunit specifically includes heterogeneous model construction and a two-level selection mechanism; The heterogeneous model construction is used to build a CNN-LSTM fusion model, a Transformer temporal model, and a gradient boosting tree combined model. The CNN-LSTM fusion model includes CNN layers and LSTM layers. The CNN layers are used to extract local spatial correlations of multivariate sequences, and the LSTM layers are used to capture long-term temporal dependencies. The Transformer temporal model uses a self-attention mechanism to explicitly model the dependency relationship between any two time points in the load sequence, which is used to handle long-period irregular events. The gradient boosting tree combined model is used to learn complex nonlinear relationships in multidimensional structured features. The two-level selection mechanism selects the optimal model or model combination for prediction under different scenarios; the selection method of the two-level selection mechanism includes the following steps: The first-level selection mechanism is to divide the historical data into multiple sub-validation sets according to different seasons and weather types; each model is validated on these sub-validation sets, and its scene adaptability profile is recorded. Second-level selection mechanism: During runtime, upon receiving the latest data, the following steps are performed: Scene matching: Based on current and forecast weather and date information, match the most similar historical scenes; Initial model selection: Based on the scene adaptability profile, recall the top five models that have historically performed best in this type of scene; Rolling evaluation: Using the most recent real data as a mini test set, the initial model is re-evaluated in real time, and the rolling MAPE is calculated. Dynamic weighting or optimization: Select the single model with the best performance as the predictor for this round, or use a variable weighting method based on real-time performance to perform model combination prediction; The probabilistic prediction subunit is used to output load prediction intervals. Specifically, model selection is based on the time scale. For ultra-short-term time scales, a CNN-LSTM fusion model is used, which outputs minute-level point predictions and probability distributions based on real-time meteorological data and very short-term trends. For short-term time scales, a Transformer time series model or model combination is used, considering weekly cycles and weather forecasts, to output hourly load curves and confidence intervals. For medium- and long-term time scales, a gradient boosting tree combination model is used, which combines calendar and climate predictions to output interval estimates of daily maximum or minimum loads. For each selected model, Monte Carlo Dropout is used for forward propagation to obtain prediction samples, and then the prediction interval is calculated. The intelligent control module establishes an optimization model based on the risk assessment results, with the goal of power supply reliability and economy, and performs coordinated control of distributed power sources, flexible loads and energy storage devices based on the optimal solution of the optimization model.

2. The distribution network load forecasting and dynamic stability system based on big data simulation analysis according to claim 1, characterized in that, The data acquisition and processing module specifically includes a multi-dimensional data source acquisition unit and a data preprocessing unit; The multi-dimensional data source acquisition unit is used to acquire distribution network data from multi-dimensional data sources; the distribution network data includes power source data, network data, load data, energy storage data, and environmental and market data; the data preprocessing unit is used to preprocess the acquired distribution network data.

3. The distribution network load forecasting and dynamic stability system based on big data simulation analysis according to claim 1, characterized in that, The intelligent control module includes a multi-objective optimization model construction unit and a collaborative control unit. The multi-objective optimization model construction unit is used to construct the power supply reliability risk cost objective function and the system operation total cost objective function. The collaborative control unit obtains the Pareto optimal solution set by solving the power supply reliability risk cost objective function and the system operation total cost objective function, calculates the satisfaction based on the Pareto optimal solution set, and performs collaborative control of distributed power sources, flexible loads, and energy storage devices based on the comprehensive satisfaction.

4. The distribution network load forecasting and dynamic stability system based on big data simulation analysis according to claim 2, characterized in that, The data preprocessing unit includes a heterogeneous data unification subunit, a multi-layer anomaly detection subunit, and an intelligent repair subunit; the heterogeneous data unification subunit is used to perform time-series alignment, spatial correlation, and unified encoding on the power distribution network operation data. The multi-layer anomaly detection subunit is used to detect anomalies in the unified coded distribution network operation data; the intelligent repair subunit is used to repair abnormal data points; the time alignment is to interpolate or aggregate data of different frequencies and unify them to a reference time axis; the spatial association is to establish the topological connection relationship of data points; and the unified coding is to generate a globally unique code for each data point. The multi-layer anomaly detection subunit performs anomaly detection on the data using a three-layer anomaly detection mode. Specifically, the first-layer anomaly detection mode filters data by setting and statistically detecting outliers and constraining the rate of change. The physical limit rule is set to ±10% of the rated voltage. Based on a sliding window, statistical outlier detection is used to identify abrupt changes. The rate of change of load in adjacent time periods exceeds the physical threshold. The second-layer anomaly detection mode uses the DBSCAN clustering algorithm, considering the temporal continuity constraint, to cluster data from similar days and time periods into one class, and to mark isolated points or small clusters of data as anomalies. The third-layer anomaly detection mode uses an LSTM-Autoencoder reconstructed model for detection; LSTM-Autoencoder is a long short-term memory autoencoder. The LSTM-Autoencoder reconstruction model includes a training phase, a detection phase, and a decision-making mechanism. During the training phase, the autoencoder is trained using historical normal data to learn the normal patterns of load, voltage, active power, and reactive power data. During the detection phase, real-time data is input, and the reconstruction error is calculated using the following formula: Where E represents the reconstruction error, and X represents the normalized real-time input data vector. This represents the reconstructed data vector generated by the LSTM-Autoencoder decoder output; the decision mechanism is: if If so, the input real-time data is determined to be abnormal; where The threshold is dynamic and determined based on historical error distribution; abnormal data is detected through a three-layer anomaly detection mode. The intelligent repair subunit initiates a repair process for data points identified as abnormal. The repair process includes historical similar day matching repair, higher-order spatial interpolation repair, and generative repair. Specifically, historical similar day matching repair is as follows: Similarity calculation: Based on multi-dimensional features such as date type, weather conditions, and holiday attributes, calculate the similarity between the current day and historical days. Similarity represents the overall similarity score between the current predicted date and historical sample dates, and w1, w2, and w3 correspond to the weight coefficients of date type, weather conditions, and holiday attributes, respectively, satisfying w1 + w2 + w3 = 1; S date S represents the date type similarity component. weather S represents the weather condition similarity component. holiday This represents the similarity component of holiday attributes; Candidate date selection: Select the top five historical dates with the highest similarity as candidate dates; Pattern extraction: Extract normal data from candidate days within the same time period to repair outliers; When single-point data is missing but adjacent monitoring points are normal, it is repaired by high-order spatial interpolation. High-order spatial interpolation is based on the Kriging interpolation method with topological constraints. First, a distribution network impedance model is established and the electrical distance between each node is calculated. The estimated value of the anomaly point is obtained by weighting the adjacent nodes. The weight is inversely proportional to the electrical distance. Independent interpolation is performed on the three phases A, B and C respectively. Otherwise, generative repair is adopted: generative repair uses a conditional generative adversarial network (CGAN); the CGAN includes a generator, a discriminator, and a physical constraint layer; the generator's input is normal data segments before and after the abnormal period and external conditions, including weather, date, and holidays; the discriminator's input is the data generated by the generator and real historical data segments; the physical constraint layer generates data by simplifying power flow verification.

5. The distribution network load forecasting and dynamic stability system based on big data simulation analysis according to claim 1, characterized in that, The dynamic stability simulation unit takes the load forecast results as boundary conditions and inputs them into the hybrid simulation environment to generate fault scenarios. For each fault scenario, the hybrid simulation environment outputs stability risk assessment indicators and uses all stability risk assessment indicators to perform stability risk assessment. The hybrid simulation environment integrates a neural network-based surrogate model and a power distribution network electromagnetic transient simulation model. Stability risk assessment measures the probability of overload on critical equipment and voltage exceeding limits at nodes in different future time periods. Specifically, the hybrid simulation environment has two channels: a physical channel, which includes the power distribution network electromagnetic transient simulation model, and a surrogate channel, which includes a neural network surrogate model. The power distribution network electromagnetic transient simulation model provides training data for the surrogate model and verifies its key outputs. The neural network surrogate model receives verification and updates from the physical channel, filtering scenarios based on the degree of topology change, the proportion of nonlinear equipment, etc. The simulation time span, renewable energy penetration rate, system short-circuit ratio, surrogate model confidence, and deviation from operating conditions were considered to select the simulation channel. The surrogate model of the neural network includes a GAT encoding layer, a spatiotemporal Transformer layer, and a physical information regularization layer. The GAT encoding layer is used to model the distribution network topology as a graph structure, learn the embedded representations of nodes and edges, and capture electrical coupling relationships. The spatiotemporal Transformer layer is used to handle time-varying inputs and capture the dependencies in the time dimension and the interactions in the spatial dimension. The physical information regularization layer is used to add a penalty term based on Kirchhoff's laws to the loss function to ensure that the model output approximately satisfies basic physical laws. The stability risk assessment is determined by the stability risk assessment index, R. Z Calculated by the following formula: ； Where Z represents the set of power grid areas for which stability risk assessment is conducted; and These are the weighting coefficients. For line or transformer L i The probability that the load rate exceeds its safety limit, i.e., the risk of equipment overload; For node V i The probability that the voltage deviates from its acceptable range; The overload risk of the equipment is calculated by the following formula: ； in, For line L i The probability of overload in the future time period [t1, t2]; This represents the total number of scenes; The probability that the load rate of the power grid under test exceeds the set value at any simulation time.

6. The distribution network load forecasting and dynamic stability system based on big data simulation analysis according to claim 3, characterized in that, The objective function for power supply reliability risk cost is expressed by the following formula: ； in, The risk value coefficient of equipment i in time period t; The real-time risk index for device i; The value of power shortage for load j; The load reduction is T, which is the total time length, i.e., one test cycle, and Ω. risk Ω represents the set of all devices within the system that participate in the operational risk assessment; load This represents the set of all load nodes or power users in the system; The objective function for the total operating cost of the system is expressed by the following formula: ； in, The active power purchased from the upper-level power grid during time period t; The electricity purchase price for time period t; Cost of distributed power generation; Active power output for distributed power sources; A collection of distributed power sources; For energy storage cycle costs; Let m be the charging and discharging power of the m-th energy storage system during time period t; A collection of energy storage systems; Cost of switching operation; Reconfigure the network switch state; It is a collection of operable switches.

7. A distribution network load forecasting and dynamic stability system based on big data simulation analysis according to claim 6, characterized in that, The coordinated control unit includes the following steps: Normalization: Normalize the objective function values ​​for power supply reliability risk cost and total system operating cost. ； in, The actual calculated value of the i-th objective function, where i is 1 or 2; Let be the minimum value of the i-th objective function in the current Pareto solution set; This represents the maximum value of the i-th objective function in the current Pareto solution set; Satisfaction level calculation: using an S-shaped membership function. ； in, For target satisfaction; The target value is the normalization target value; is the steepness coefficient, a parameter that controls the steepness of the S-curve; The center point parameter is the normalized target value corresponding to a satisfaction level of 0.5, which is set according to scheduling preferences. Comprehensive selection: Choose the solution with the highest overall satisfaction. ； in, and These represent satisfaction with reliability risks and economic efficiency, respectively. and They are respectively and The target weight.

Images