Electrified flexible source load characteristic identification and adjustable potential measurement and calculation method

Abstract

The invention discloses an electrified flexible source load characteristic identification and adjustable potential measurement and calculation method, and particularly relates to the technical field of trunk line potential measurement and calculation, and the method comprises the steps: S1, multi-dimensional dynamic data collection and processing, S2, flexible source load spatial-temporal characteristic dynamic clustering, S3, adjustable potential prediction model construction of a spatial-temporal mechanism, and S4, AC trunk line power grid safety bundle potential verification. S5, dynamic potential visualization scheduling interface development; and S6, method validity verification adaptive optimization. By constructing a multi-dimensional and high-frequency dynamic data acquisition and processing system, data of different types and different scales can be effectively fused and utilized by the model, so that the defect that full utilization of the flexible load in power grid dispatching is limited due to single data source and poor quality in the prior art is overcome, and the flexible load dispatching efficiency is improved. The problems of inaccurate evaluation and poor predictability in the prior art are fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of trunk line potential calculation technology, and more specifically, to a method for identifying the characteristics of electrified flexible source loads and calculating their adjustable potential. Background Technology

[0002] In the identification of flexible source-load characteristics and the calculation of adjustable potential in electrification and in 750 kV AC transmission, power system load control has become a key technology for improving the economy, security and decarbonization of the power grid by precisely regulating flexible loads, optimizing transmission capacity and stability, and supporting the consumption of new energy sources.

[0003] Among the existing published documents, patent publication number CN120473992A discloses a system and method for rapid calculation of the adjustable load potential in port areas. This technology achieves distributed parallel calculation of the adjustable load potential; the scheduling optimization module generates a load adjustment instruction set affected by multiple factors based on the grid demand response instructions and the adjustable potential data reported by the edge nodes of each zone; it accurately calculates the adjustable load potential, improving the flexibility and reliability of power load management in port areas; however, this patent has the following defects; During the power supply process of the 750 kV AC transmission dispatch system, the adjustable potential of flexible loads changes dynamically with time, space and operating status factors. Existing technologies cannot accurately assess the dynamically changing adjustable potential in real time, and it is difficult to effectively integrate and utilize data of different types and scales by the model. Due to the single and poor quality of the data source, the effective utilization of flexible loads in power grid dispatch is limited. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of the prior art, the present invention provides a method for identifying the characteristics and calculating the adjustable potential of electrified flexible source loads, comprising the following specific steps: S1. Multi-dimensional dynamic data acquisition and processing: Data acquisition includes time dimension data and spatial dimension data. The main recording host is 10-30 units. The power grid topology of the load distribution area is recorded, and the operating status data is recorded. S2. Dynamic clustering of the spatiotemporal characteristics of flexible source loads: The load power curve is decomposed into components of different frequency bands through wavelet packet transform, and a spatial feature vector is constructed. Each load cluster corresponds to a set of flexible loads with similar spatiotemporal characteristics. S3. The adjustable potential prediction model of the spatiotemporal mechanism is constructed. The bottom layer adopts a one-dimensional CNN network, the middle layer adopts a bidirectional LSTM network, and the top layer introduces a spatiotemporal attention mechanism. S4. Verification of the safety bundle potential of AC trunk power grid: Calculate the power flow direction of 750 kV and above lines based on the DC power flow method through the power flow bundle. The time scenarios include peak load periods and fault scenarios. The adjustable potential predicted in step three is used as input. S5. Dynamic potential visualization scheduling interface development. The visualization platform adopts a layered display architecture. The interface functions include data push function, which pushes real-time adjustable potential data to the database of the intelligent scheduling system at a frequency of seconds. S6. Method effectiveness verification and adaptive optimization: Experimental tests were conducted in industrial parks and residential communities covered by 750 kV transmission networks. Data acquisition and analysis systems were deployed to predict accuracy. The feature system was optimized periodically every 5-10 hours, and new flexible load operation feature data were introduced.

[0005] In a preferred embodiment, the time-dimensional data collected in S1 includes minute-level load power curves, electricity price signals, and meteorological data such as temperature, humidity, and light intensity, as well as hourly user electricity consumption behavior characteristics such as industrial user production shifts and residential user work and rest patterns. Data is transmitted in real-time through acquisition terminals deployed at 750 kV transmission line monitoring points, substations, and user-side smart meters, occurring every 30-50 ms. The power grid topology includes node impedance matrices, line transmission capacity, and spatial clustering of load types, including densely populated industrial areas, residential areas, and commercial core areas. The data includes the output data of wind power and photovoltaic power in the area and the distributed power sources in the area. The recording time is controlled at 10-15 minutes in the early stage and updated once every 50-80 ms in the later stage. The operation status data includes the real-time operation parameters of flexible loads. Outlier removal is carried out by combining the 3σ criterion with the isolated forest algorithm to identify and remove abnormal data caused by equipment failure and metering error. Data standardization converts parameters of different dimensions such as power, temperature and electricity price to the [0,1] interval and adopts the Z-score standardization formula: x'=(x-μ) / σ, where μ is the mean and σ is the standard deviation.

[0006] In a preferred embodiment, the real-time operating parameters in S1 include the air conditioner start / stop status, electric vehicle charging power, industrial motor speed, power grid frequency, and voltage deviation safety indicators. Then, through data missing repair, an interpolation algorithm based on spatiotemporal correlation is used. By combining K-nearest neighbor interpolation with LSTM prediction, missing data caused by communication interruption during the acquisition process is completed 5-20 times to ensure the continuity of the time series.

[0007] In a preferred embodiment, the spatial feature vector in S2 is constructed primarily through a sliding window, with a window size of 15cm and a step size of 5cm. Dynamic indicators such as load fluctuation rate, peak-to-valley difference, and response delay time within the window are calculated to capture the pattern of load changes over time. Spatial feature extraction is based on power grid zoning and load type, calculating the load density of each region (total load divided by area), source-load matching degree (distributed power output divided by load demand), and grid congestion risk (actual line power divided by transmission limit). This is combined with the proportions of interruptible load (10-20%), transferable load (20-40%), and adjustable load (30-40%) in the load type. When constructing the spatial feature vector, low-frequency components reflecting the trend of the base load and high-frequency components reflecting the fluctuation characteristics of flexible loads are extracted, along with energy entropy and peak factor statistical features. A time decay term e^(-αΔt), where α is the decay coefficient and Δt is the time interval, is added to the objective function to enhance the influence of recent data on the clustering results. Simultaneously, the electrical distance between power grid nodes is controlled within 10-25 kilometers after clustering is completed.

[0008] In a preferred embodiment, the bottom layer of S3 relies on 10-20 sets of calculators, using a convolution kernel size of 3×3 to extract local correlation information from the load feature vector, and the correlation between short-term load fluctuations and meteorological factors. The middle layer uses a bidirectional LSTM network to capture the long-term dependencies of load characteristics from both the forward and backward directions of the time series. It sets 2-8 hidden layers, each containing 64 neurons. The weights of historical information are dynamically adjusted through gating units (input gate, forget gate, output gate) every 5-10 seconds. It is divided into two branches: temporal attention and spatial attention. The temporal attention branch calculates the feature vectors at different times. The weighting of quantities highlights the peak-valley switching time of electricity price and the characteristics of grid frequency fluctuations at critical moments of load response. The top-level spatiotemporal attention mechanism is introduced, and the spatial attention branch is based on the load transfer capacity under the grid safety margin of N-1 faults in each region to calculate spatial weights. The loss function adopts the weighted sum of root mean square error (RMSE) and mean absolute percentage error (MAPE), where RMSE=√[∑(y_i-ŷ_i)² / n], MAPE=(1 / n)∑|(y_i-ŷ_i) / y_i|×100%, where y_i is the actual potential value, ŷ_i is the predicted value, and n is the number of samples. The model is trained 20-50 times to complete the model construction.

[0009] In a preferred embodiment, in step S4, the power flow bundle ensures that the line power after adjusting the flexible load does not exceed the thermal stability limit P_ij ≤ P_ij^max, where P_ij is the actual line power and P_ij^max is the line transmission limit power. The calculations are performed 30-50 times. The voltage bundle calculates the voltage amplitude of each load node using the nodal voltage equation, ensuring the adjusted voltage deviation is within an allowable range of ±5%. The frequency bundle considers a range of 50 ± 0.2 Hz. The time scenarios include peak load periods of 14:00-16:00 in summer and off-peak periods of 2:00-4:00 AM. During periods of fluctuating renewable energy output, particularly wind power generation during the nighttime hours of 18:00-24:00, fault scenarios include 5-10 responses to 750 kV line N-1 faults, with each response lasting 5-15 seconds, and substation main transformer maintenance fault modes. Market scenarios include time-of-use pricing adjustments and changes in demand response incentive policies. For each scenario, the input time is controlled within 50-100 ms during the input operation. The substitution time into the grid security bundle model is controlled within 60-100 ms. By solving the optimal power flow problem, it is determined whether the adjusted grid satisfies all security bundles, with 50-100 judgments performed.

[0010] In a preferred embodiment, the bottom layer of S5 is a power grid geographic wiring diagram, which displays the real-time operating status of 750 kV transmission lines and substations based on a GIS system. The line power utilization rate is marked with different colors: green ≤60%, yellow 60%-80%, and red ≥80%. The middle layer is a heat map of flexible load potential distribution, which uses color depth to represent the adjustable potential of a region, with darker colors indicating greater potential and lighter colors indicating less potential. The update time is once every 5-10 minutes. The top layer is a key indicator dashboard, which displays 5-10 sets of total adjustable potential of the entire network, 50-100 sets of potential proportion of each load cluster, and 30-60 sets of core indicators of power grid safety margin in real time. Drill-down query is supported, and clicking on a region can view the potential details of specific loads in that region. The number of detailed display groups is 100-300.

[0011] In a preferred embodiment, step S5 provides input for dispatching decisions. The instruction response function receives the potential call instruction issued by the dispatching system to call 100-200MW of adjustable load in the region, and returns the feasibility analysis results of the instruction execution, including the expected response time and the impact on power grid security. The impact is divided into safe and unsafe. The historical data query function supports dispatchers to query the potential calculation results and call records for any time period. The calculation results are displayed in 5-10 sets at a time, and the call records are displayed in 30-50 sets at a time.

[0012] In a preferred embodiment, in step S6, the accuracy is calculated as 1 - measured value - calculated value divided by measured value × 100%, requiring an accuracy of not less than 90-98%, and a 750 kV power grid simulation model containing large-scale flexible loads is constructed. In a preferred embodiment, S6 includes prediction accuracy (MAPE), response speed data processing and calculation time (1-5s), and safety bundle satisfaction rate (95-99%). Real-time monitoring of indicator changes is conducted. When an indicator exceeds a preset threshold (MAPE 8-10%), a model update process is triggered. The model parameters are fine-tuned using the latest collected load data and actual call results. The adjustable potential calculated by this method is compared with the actual call potential value. The accuracy index is calculated based on the actual load reduction in the demand response event. Simulations of new energy output fluctuations and grid faults under different scenarios are performed to evaluate grid frequency deviation and voltage fluctuation indicators, thus constructing model performance evaluation indicators.

[0013] The technical effects and advantages of this invention are as follows: This invention significantly improves the real-time performance and accuracy of flexible load characteristic perception by constructing a multi-dimensional, high-frequency dynamic data acquisition and processing system. Traditional methods are limited by the singularity and lag of data acquisition, making it difficult to characterize the dynamic changes of load. It provides minute-level load power, meteorological factors, user behavior characteristics, and second-level power grid operation status data. Furthermore, it employs a hybrid interpolation algorithm combining K-nearest neighbors and LSTM prediction, and an anomaly data processing mechanism combining the 3σ criterion and the isolated forest algorithm. This effectively solves the data quality problems caused by communication interruptions or equipment failures. This refined data processing flow ensures the continuity and reliability of input data, laying a solid foundation for subsequent accurate analysis. At the same time, by standardizing and unifying the dimensions through Z-score, different types and scales of data can be effectively integrated and utilized by the model, thereby overcoming the limitations of existing technologies that rely on single data sources and poor quality, hindering the full utilization of flexible loads in power grid dispatch.

[0014] This invention creatively couples a one-dimensional CNN, a bidirectional LSTM, and a spatiotemporal attention mechanism in a layered manner. The CNN is responsible for capturing the short-term correlation between local load and factors such as weather and electricity prices. The bidirectional LSTM learns the long-term time dependence of load response from both positive and negative directions. The spatiotemporal attention mechanism can dynamically focus on key moments and key areas and adaptively allocate weights. This progressive analysis architecture of "local features - long-term dependence - key focus" enables the model to not only reflect the static characteristics of the load, but also accurately capture its dynamic laws of evolution with time and system state, providing highly reliable data support for scheduling decisions and fundamentally solving the problems of inaccurate assessment and poor predictability of existing technologies.

[0015] 3. This invention, by developing a dynamic potential visualization scheduling interface, presents complex calculation results to dispatchers in a layered and intuitive graphical interface, supporting drill-down queries. This greatly improves human-computer interaction efficiency and decision-making transparency, achieving second-level data interaction and command response with the intelligent scheduling system. It can directly and quickly transform the calculated adjustable potential into executable scheduling strategies, and complete multi-scenario and multi-constraint power grid safety verification before issuing commands, ensuring that every call meets the N-1 fault, voltage, and frequency safety requirements, demonstrating excellent emergency response capabilities. In addition, the introduced adaptive optimization mechanism can continuously iterate the model based on the actual call effect, forming a virtuous cycle of "application-verification-optimization," ensuring the long-term effectiveness and advancement of the method. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the operation of the electrification flexible source-load characteristic identification and adjustable potential calculation method of the present invention. Detailed Implementation

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

[0018] As attached Figure 1 This paper presents a method for identifying the characteristics of flexible power sources and loads in electrification and calculating their adjustable potential. Example

[0019] S1. Multi-dimensional dynamic data acquisition and processing: Data acquisition includes time-dimensional data such as minute-level load power curves, electricity price signals, and meteorological data (temperature, humidity, and light intensity), as well as hourly user electricity consumption behavior characteristics, such as industrial user production shifts and residential user work-rest patterns. Real-time data transmission is achieved through acquisition terminals deployed at 750 kV transmission line monitoring points, substations, and user-side smart meters, with data transmitted every 30 ms. Spatial-dimensional data is primarily collected using 10 recording hosts, recording the power grid topology of the load distribution area. The power grid topology includes node impedance matrices, line transmission capacity, load type spatial clusters (industrial dense areas, residential areas, commercial core areas), and output data of distributed power sources such as wind power and photovoltaic power within the area. Initial recording time is controlled to 10 minutes, with subsequent recording... The recording time is controlled to be updated once every 50ms. The operating status data includes the real-time operating parameters of flexible loads, including the air conditioner start / stop status, electric vehicle charging power, industrial motor speed, grid frequency, and voltage deviation safety indicators. Then, through data missing repair, an interpolation algorithm based on spatiotemporal correlation is used. Through K-nearest neighbor interpolation combined with LSTM prediction, missing data caused by communication interruption during the acquisition process is completed 5 times to ensure the continuity of the time series. Outlier removal is carried out by combining the 3σ criterion with the isolated forest algorithm to identify and remove abnormal data caused by equipment failure and metering error. Data standardization converts parameters of different dimensions such as power, temperature, and electricity price to the [0,1] interval, and adopts the Z-score standardization formula: x'=(x-μ) / σ, where μ is the mean and σ is the standard deviation.

[0020] S2. Dynamic clustering of the spatiotemporal characteristics of flexible source loads: The load power curve is decomposed into components of different frequency bands using wavelet packet transform. Low-frequency components, reflecting the trend of the base load, and high-frequency components, reflecting the fluctuation characteristics of flexible loads, are extracted. Energy entropy and peak factor statistical features are extracted. A sliding window is used, with a window size of 15cm and a step size of 5cm. Dynamic indicators such as load volatility, peak-to-valley difference, and response delay time within the window are calculated to capture the pattern of load changes over time. Spatial feature extraction is based on power grid zoning and load type, calculating the load density of each region, mainly by dividing the total load by the region area. Source-load matching is also performed. The degree is mainly calculated by dividing the output of distributed power sources by the load demand. The grid congestion risk is mainly calculated by dividing the actual power of the line by the transmission limit. Combined with the proportion of interruptible loads (10%), transferable loads (20%), and adjustable loads (30%) in the load types, a spatial feature vector is constructed. By adding a time decay term e^(-αΔt) to the objective function, where α is the decay coefficient and Δt is the time interval, the influence of recent data on the clustering results is enhanced. At the same time, the electrical distance between grid nodes is controlled within 10 kilometers. After clustering, each load cluster corresponds to a set of flexible loads with similar spatiotemporal characteristics.

[0021] S3. A spatiotemporal adjustable potential prediction model is constructed. The bottom layer uses a one-dimensional CNN network with 10 sets of data points. The convolution kernel size is set to 3×3 to extract local correlation information from the load feature vector, showing the correlation between short-term load fluctuations and meteorological factors. The middle layer uses a bidirectional LSTM network to capture the long-term dependencies of load characteristics from both the forward and backward directions of the time series. Two hidden layers are set, each containing 64 neurons. The weights of historical information are dynamically adjusted every 5 seconds through gating units (input gate, forget gate, output gate). The top layer introduces a spatiotemporal attention mechanism, divided into temporal attention and spatial attention branches. The spatial attention branch calculates the weights of feature vectors at different times to highlight the peak-valley switching time of electricity price and the characteristics of grid frequency fluctuations at critical moments of load response. The spatial attention branch calculates spatial weights based on the load transfer capacity under the grid safety margin N-1 fault in each region to complete the model construction. The loss function adopts the weighted sum of root mean square error (RMSE) and mean absolute percentage error (MAPE), where RMSE=√[∑(y_i-ŷ_i)² / n], MAPE=(1 / n)∑|(y_i-ŷ_i) / y_i|×100%, where y_i is the actual potential value, ŷ_i is the predicted value, n is the number of samples, and the training is performed 20 times.

[0022] S4. Verification of the AC trunk power grid safety potential: Using the power flow bundle, calculate the power flow direction of 750 kV and above lines based on the DC power flow method to ensure that the line power does not exceed the thermal stability limit P_ij ≤ P_ij^max after adjusting the flexible load, where P_ij is the actual line power and P_ij^max is the line transmission limit power. Perform 30 simulations. The voltage bundle calculates the voltage amplitude of each load node using the nodal voltage equation. After adjustment, the voltage deviation is within the allowable range of ±5%. The frequency bundle considers a range of 50.2 Hz. Time scenarios include peak load periods in summer (14:00-16:00) and off-peak periods. The time periods are 2:00-4:00 AM and 6:00-12:00 PM, the peak wind power generation period during periods of fluctuating renewable energy output. Fault scenarios include 5 responses to the N-1 fault on a 750 kV line, with each response taking 5 seconds, and a substation main transformer maintenance fault mode. Market scenarios include time-of-use pricing adjustments and changes in demand response incentive policies. For each scenario, the adjustable potential predicted in step 3 is used as input, with the input time controlled within 50 ms. This is then substituted into the grid security bundle model, with the substitution time controlled within 60 ms. By solving the optimal power flow problem, it is determined whether the adjusted grid satisfies all security bundles, with 50 judgments made.

[0023] S5. Development of a dynamic potential visualization scheduling interface. The visualization platform adopts a layered display architecture. The bottom layer is a power grid geographic wiring diagram, which displays the real-time operating status of 750 kV transmission lines and substations based on a GIS system, using different colors to indicate line power utilization rates: green ≤60%, yellow 60%, and red ≥80%. The middle layer is a flexible load potential distribution heat map, using color depth to represent the adjustable potential of a region, with darker colors indicating larger potential and lighter colors indicating smaller potential. The update time is once every 5 minutes. The top layer is a key indicator dashboard, displaying in real-time the total adjustable potential of the entire network (5 groups), the potential percentage of each load cluster (50-100 groups), and the core indicators of power grid safety margin (30 groups). It supports drill-down queries, allowing users to view detailed potential data for specific loads within a designated area. Up to 100 groups of details can be displayed. Interface functions include a data push function, which pushes real-time adjustable potential data to the intelligent dispatch system's database at a second-level frequency, providing input for dispatch decisions. An instruction response function receives potential call instructions from the dispatch system, calling 100MW of adjustable load in a designated area and returning feasibility analysis results for instruction execution, including estimated response time and impact on grid security (categorized as safe or unsafe). A historical data query function allows dispatchers to query potential calculation results and call records for any time period, displaying 5 groups of calculation results and 30 groups of call records at a time.

[0024] S6. Method effectiveness verification and adaptive optimization: Experimental tests were conducted in industrial parks and residential communities covered by a 750 kV transmission network. A data acquisition and analysis system was deployed, and the adjustable potential calculated by this method was compared with the actual potential value. The accuracy index was calculated based on the actual load reduction in the demand response event. Accuracy = 1 - measured value - calculated value divided by measured value × 100%, with an accuracy requirement of not less than 90%. A 750 kV power grid simulation model with large-scale flexible loads was constructed to simulate the fluctuation of new energy output and power grid faults under different scenarios. The power grid frequency deviation and voltage fluctuation index were evaluated, and model performance evaluation indexes were constructed, including prediction accuracy (MAPE), response speed (data processing and calculation time 1 second), and safety constraint satisfaction rate (95%). The index changes were monitored in real time. When the index exceeded the preset threshold (MAPE 8%), the model update process was triggered. The model parameters were fine-tuned using the latest collected load data and actual call results. The feature system was optimized every 5 hours, and new flexible load operation feature data were introduced. Example

[0025] This example illustrates how to alleviate grid congestion during the midday summer peak in an industrial park. The scenario involves a densely populated industrial area on a 750 kV transmission network. At 2:30 PM on a summer day with an ambient temperature of 38°C, the combined effects of air conditioning and production loads caused a critical 750 kV line to reach 92% of its transmission limit, posing a risk of congestion. The dispatch center needed to quickly utilize the area's flexible load resources to ensure grid safety.

[0026] S1. Multi-dimensional dynamic data acquisition and processing: Acquisition terminals deployed at substations and user sides collect minute-level load power, temperature, humidity, and real-time electricity prices at 40ms intervals. Ten recording hosts provide the power grid topology, node impedance matrix, and line transmission capacity for the area. The load types are identified as interruptible loads (non-core processes, accounting for 12%), adjustable loads, air conditioning setpoints, and motor speeds, accounting for 38%. Data packet loss due to transient communication interference is detected. The system starts the K-nearest neighbor-LSTM hybrid interpolation algorithm to successfully fill in 8 missing data points. The 3σ criterion and the isolated forest algorithm are used to remove 2 power anomalies caused by sensor failures. Z-score standardization is performed on all collected power, temperature, and electricity price data to eliminate the influence of dimensions.

[0027] S2. Flexible source load spatiotemporal characteristics dynamic clustering: Wavelet packet transform is used to decompose the total load curve of the region, extracting the low-frequency component of the basic load reflecting stable production and the high-frequency component reflecting the violent fluctuations of air conditioning group switching. A sliding window with a window size of 15cm and a step size of 5cm is used to calculate that the load fluctuation rate within the current window is 15%, and the peak-to-valley difference reaches 85MW. Spatial characteristic calculation shows that the region has high load density, low source-load matching degree, insufficient photovoltaic output, and a grid congestion risk index of 0.92. The clustering algorithm introduces a time decay term α=0.15 in the objective function and binds the electrical distance within 20 kilometers. Finally, 75% of the industrial load in the region is dynamically aggregated into a "high energy-consuming and adjustable industrial load cluster".

[0028] S3. Construction of the adjustable potential prediction model based on spatiotemporal mechanism: The feature vectors of the above load clusters are input into the prediction model. A one-dimensional CNN convolution kernel (3×3) captures the local spatiotemporal correlation between afternoon high temperature and sudden increase in air conditioning load. A bidirectional LSTM with 4 hidden layers and 64 neurons per layer learns the response pattern of this type of load under historical electricity price incentives. Its gating unit dynamically adjusts the weights every 8 seconds. Spatiotemporal attention mechanism: temporal attention focuses on 14:00, a critical period with both high electricity price and load. Spatial attention is given the maximum weight to the area with high risk of line congestion. The model predicts that the adjustable potential of this load cluster in the next 30 minutes is 180MW. After 35 training cycles, the model has the smallest weighted sum of RMSE and MAPE in the loss function. The predicted MAPE is 4.2%.

[0029] S4. AC trunk power grid safety potential verification: The predicted potential of 180MW was input into the safety verification model within 70ms. 45 simulations were performed for peak hours and N-1 fault scenarios. Core verification results: Power flow: After calling up 150MW, the power of the critical line decreased from 92% to 75%, satisfying P_ij≤P_ij^max. Voltage: The maximum voltage deviation at each node was +3.1%, within the allowable range of ±5%. Frequency: After adjustment, the system frequency stabilized at 50.05Hz, meeting the requirement of 50±0.2Hz. Verification conclusion: Calling up loads of 150MW and below is safe.

[0030] S5. Dynamic potential visualization scheduling interface development: The dispatcher sees the industrial zone displayed as a dark red heat map in the middle layer of the platform. The top-level dashboard displays 8 sets of data for "Regular Network Adjustable Potential", totaling 520MW, of which this industrial load cluster accounts for 34%. Clicking drill-down query retrieves the potential details of 150 specific factories in this area. The dispatcher issues a 120MW load reduction command, and the interface returns a feasibility analysis within 3 seconds: the estimated response time is 10 minutes, and the impact on grid security is "safe". After the command is executed, the actual load reduction is 118MW, and the line congestion warning is lifted.

[0031] S6. Method effectiveness verification and adaptive optimization: Accuracy calculation: Accuracy = 1 - |118 - 120| / 118 × 100% = 98.3%, which is higher than the 90% requirement. The entire process, from data acquisition to command confirmation, took 4.2 seconds, meeting the 1-5s response speed requirement. All safety constraints were satisfied, with a 100% safety constraint satisfaction rate. Due to the excellent performance of all indicators, no model update was triggered. Example

[0032] Application scenario of wind power consumption and transferable load dispatch in residential areas at night: This example focuses on a large residential area where wind power output continues to increase significantly at 20:00 at night, putting pressure on the power grid for peak shaving and consumption. The dispatch center hopes to assess and guide the transferable load of electric vehicles (EVs) in the area, shifting them from peak to off-peak hours to promote the consumption of clean energy.

[0033] S1. Multi-dimensional dynamic data acquisition and processing: Smart meters and EV charging piles upload hourly features of charging power, humidity, and residents' daily routines at 35ms intervals, showing that nighttime is the main charging period. In the data processing stage, the isolated forest algorithm effectively identifies and removes an abnormal charging data caused by a device communication error.

[0034] S2. Dynamic clustering of the spatiotemporal characteristics of flexible source loads. Sliding window analysis shows that the EV charging load forms an independent fluctuation peak after the evening peak. Spatially, the load density in this area is moderate, but the proportion of transferable load EV charging is as high as 32%. The clustering algorithm, combined with the electrical distance of 15 kilometers and load characteristics, aggregates the scattered residential EV charging loads into "residential EV transferable load clusters".

[0035] S3. The adjustable potential prediction model of the spatiotemporal mechanism is constructed. The model input includes historical EV charging behavior, real-time wind power output, and time-of-use electricity price signal. The bidirectional LSTM successfully captures users' response habits to the off-peak electricity price in the early morning of the next day. The time attention branch of the spatiotemporal attention mechanism significantly increases the weight of the characteristics of the off-peak electricity price period after "23:00". The model predicts that the load cluster has a transfer potential of 65MWh that night.

[0036] S4. Verification of the safety potential of the AC trunk power grid: During the off-peak hours of 02:00 and under the scenario of new energy fluctuations, the added 65MWh charging load was verified. The verification focused on the voltage bundle. The results showed that at night when wind power generation is high and the base load is low, the voltage of each node is still within the acceptable range after the additional load is connected. The substation main transformer maintenance mode was simulated, and the conclusion was still that it is safe.

[0037] S5. Dynamic potential visualization scheduling interface development: In the "Potential Ratio of Each Load Cluster" of the platform's top-level dashboard, the potential ratio of "Residential EV Transferable Load Cluster" is updated in real time. The dispatch center sends "Nighttime Charging Discount" signals to the community energy management system at a frequency of seconds through the interface's data push function. Users responded positively, and 62MWh of load transfer was actually completed. The platform's historical query function recorded all 40 call records for this event.

[0038] S6. Method effectiveness verification and adaptive optimization: accuracy = 1 - |62 - 65| / 62 × 100% = 95.2%. The system detected that the MAPE of this prediction was 6.1%, which is lower than the update threshold of 8%, so the model will not be updated for the time being. Example

[0039] Emergency Response and Adaptive Optimization of N-1 Fault in Commercial Power Grid: Application Scenario. This embodiment simulates a sudden N-1 fault on a 750 kV line in a core commercial area during off-peak hours. The system needs to call up flexible loads within seconds to provide emergency power support and prevent frequency collapse.

[0040] S1. Multi-dimensional dynamic data acquisition and processing: In the instant of failure, the acquisition terminal collects key safety indicators such as the power grid frequency starting to drop and the voltage deviation increasing at a high frequency at 50ms intervals. The failure causes a brief communication interruption. The system starts up and completes 15 missing data repairs within 100ms, ensuring the continuity of decision data during the failure.

[0041] S2. Dynamic clustering of spatiotemporal characteristics of flexible source loads. Before the fault, the area had been clustered into a "commercial adjustable load cluster". Its characteristics are that the adjustable loads of building air conditioning and lighting account for a high proportion of 38%, and are sensitive to the grid frequency.

[0042] S3. The model for predicting the adjustable potential of the spatiotemporal mechanism is constructed. Based on real-time and rapidly changing operating status data, the model can quickly predict that the load cluster can provide a second-level rapid adjustment potential of 40MW.

[0043] S4. AC trunk power grid safety potential verification: The system completes 60 rapid pre-verifications of fault scenarios within 80ms, verifying the core frequency bundle, confirming that calling 35MW load is sufficient to stabilize the frequency above 49.90Hz, meeting the requirement of 50±0.2Hz, and the power flow and voltage verifications pass simultaneously.

[0044] S5. A dynamic potential visualization scheduling interface was developed. When a fault occurs, the faulty line turns red as an alarm on the platform's underlying power grid diagram. The color depth of the commercial area's potential changes in real time on the mid-level heat map. The automatic scheduling system issues a 35MW call command, and the interface returns a successful execution signal within 1.8 seconds. The power grid frequency recovers to a stable range within 12 seconds.

[0045] S6. Method effectiveness verification: Adaptive optimization, accuracy. Actual call was 35MW, prediction was 40MW, accuracy = 1 - |35 - 40| / 35 × 100% = 85.7%. The accuracy of this emergency response was below the 90% threshold, and MAPE reached 9.8%. The system automatically triggered the model update process, optimization action, parameter fine-tuning. Using the fault data, the weights of the bidirectional LSTM network were fine-tuned. Feature optimization: In the feature system of S1 stage, a new dynamic feature, "frequency change rate", was introduced to enhance the model's ability to perceive and predict extreme faults. The entire optimization cycle was 7 hours.

[0046] The following test data table is derived from the above four sets of embodiments: Table 1. Experimental data on the predictive accuracy of adjustable potential:

[0047] Table 2. Measured data on adjustable potential for different load types:

[0048] Table 3, Data on the adaptive optimization effect of the model:

[0049] Based on the above three sets of charts, it is evident that the adjustable potential prediction accuracy is higher, the adjustable potential calculation for different load types is more excellent, and the adaptive optimization effect of the model is superior. By constructing a multi-dimensional, high-frequency dynamic data acquisition and processing system, the real-time performance and accuracy of flexible load characteristic perception are significantly improved. Traditional methods are limited by the single and lagging nature of data acquisition, making it difficult to characterize the dynamic changes of the load. By standardizing and unifying the dimensions through Z-score, data of different types and scales can be effectively integrated and utilized by the model, thereby overcoming the limitations of existing technologies in fully utilizing flexible loads in power grid dispatch due to the single and poor quality of data sources.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An electrified flexible source and load characteristics identification and adjustable potential measurement method, characterized in that: The specific steps include: S1, multi-dimensional dynamic data acquisition and processing, data acquisition time dimension data, spatial dimension data mainly use 10-30 recording host computers to record the power grid topology of load distribution area, record the running state data; S2, flexible source load space-time feature dynamic clustering, through wavelet packet transform, the load power curve is decomposed into components of different frequency bands, and a space feature vector is constructed, each load cluster corresponds to a set of flexible load sets with similar space-time characteristics; S3, adjustable potential prediction model construction of time-space mechanism, the bottom layer adopts one-dimensional CNN network, the middle layer adopts bidirectional LSTM network, and the top layer introduces time-space attention mechanism; S4, AC trunk network safety beam potential verification, through the power flow beam, based on the direct current flow method to calculate the power flow of 750kV and above lines, time scene includes load peak period, fault scene, the adjustable potential predicted in step three is used as input; S5, dynamic potential visualization scheduling interface development, the visualization platform adopts a layered display architecture, and the interface functions include data pushing function, which pushes real-time adjustable potential data to the database of intelligent dispatching system at a frequency of seconds; S6, method effectiveness verification adaptive optimization, the real test selects 750kV transmission network covering industrial parks and residential communities, deploys data acquisition and analysis system, and optimizes the feature system every 5-10h, and introduces new flexible load operation characteristic data.

2. The method of claim 1, wherein the method further comprises: The data acquisition time dimension data in S1 includes minute-level load power curve, electricity price signal, temperature, humidity and illumination in meteorological data, and hour-level user electricity behavior characteristics of industrial user production shift and resident user work and rest regularity. Through the collection terminal deployed in the 750kV transmission line monitoring point, the transformer substation and the user side intelligent electric meter, the data is realized real-time transmission, every 30-50ms, the power grid topology structure is the node impedance matrix, the line transmission capacity, the industrial intensive area, the residential area and the commercial core area in the load type space clustering, and the output data of the distributed power in the region, such as wind power and photovoltaic, the early recording time is controlled within 10-15min, and the later recording time is controlled to be updated once every 50-80ms, the running state data includes real-time running parameters of flexible load, abnormal value elimination is combined with 3σ criterion and isolated forest algorithm, abnormal data caused by equipment failure and metering error is identified and removed, and different dimension parameters such as power, temperature and electricity price are converted to [0, 1] interval through data standardization, and Z-score standardization formula is adopted: x'=(x-μ) / σ, wherein μ is mean value and σ is standard deviation.

3. The method of claim 1, wherein the method further comprises: The real-time running parameters in S1 are air conditioner start-stop state, electric vehicle charging power, industrial motor speed, power grid frequency and voltage deviation safety index, and then data missing repair is carried out, interpolation algorithm based on space-time correlation is adopted, missing data caused by communication interruption in the collection process is completed 5-20 times through K nearest neighbor interpolation combined with LSTM prediction, and the continuity of time series is ensured.

4. The method of claim 1, wherein the method further comprises: The S2 adopts the construction space feature vector mainly through the sliding window, wherein the window size is set to 15 cm, and the step is 5 cm. The load fluctuation rate, peak-valley difference, and response delay time dynamic indicators in the window are calculated to capture the law of load change over time. The spatial feature extraction is based on the power grid partition and load type. The load density of each region is calculated, mainly the total load divided by the area. The source-load matching degree is mainly the distributed power output divided by the load demand. The power grid congestion risk is mainly the actual power of the line divided by the transmission limit. Combined with the interruptible load 10-20% proportion, the transferable load 20-40% proportion, and the adjustable load 30-40% proportion in the load type, the energy entropy reflecting the basic load trend and the high-frequency component reflecting the flexible load fluctuation characteristics are extracted when constructing the spatial feature vector. The time decay term e^(-αΔt) is added to the objective function, wherein α is the decay coefficient, and Δt is the time interval. The influence of recent data on the clustering result is enhanced. At the same time, the electrical distance of the grid node is controlled within 10-25 kilometers. After clustering is completed.

5. The method of claim 1, wherein: In the S3, the bottom layer relies on 10-20 groups of calculators. The convolution kernel size is set to 3x3 to extract the local correlation information in the load feature vector. The short-term load fluctuation is associated with the meteorological factors. The middle layer adopts a bidirectional LSTM network to capture the long-term dependence of load characteristics from the forward and reverse directions of the time series. 2-8 hidden layers are set, each containing 64 neurons. The weights of historical information are dynamically adjusted through the dynamic adjustment of the input gate, forget gate, and output gate of the gating unit, which is adjusted once every 5-10 seconds. It is divided into time attention and space attention branches. The time attention branch highlights the load response key time, the peak-valley switching time of electricity price, and the grid frequency fluctuation characteristics by calculating the weights of the feature vectors at different times. The top layer introduces a spatio-temporal attention mechanism. The space attention branch calculates the spatial weight based on the load transfer capacity under the N-1 fault of the grid safety margin of each region. The loss function adopts the weighted sum of root mean square error RMSE and mean absolute percentage error MAPE, wherein RMSE = √[∑(y_i-ŷ_i)² / n], MAPE = (1 / n)∑|(y_i-ŷ_i) / y_i|×100%, wherein y_i is the actual potential value, ŷ_i is the predicted value, and n is the sample number. The model is built after 20-50 times of training.

6. The method of claim 1, wherein: The S4 ensures that the line power after adjusting the flexible load does not exceed the thermal stability limit P_ij≤P_ij^max, where P_ij is the actual line power, and P_ij^max is the line transmission limit power. The calculation simulation is performed 30-50 times, the voltage bundle calculates the voltage amplitude of each load node through the node voltage equation, the voltage deviation after adjustment is within the allowable range of ±5%, the frequency bundle is considered within the range of 50±0.2Hz, and the time scenario includes the load peak period of summer 14:00-16:00, the low valley period of early morning 2:00-4:00, and the new energy output fluctuation period of night 18:00-24:00 when wind power is large, the fault scenario includes 5-10 times of response to 750kV line N-1 fault, each response time is 5-15s, substation transformer maintenance fault mode, market scenario includes time-of-use electricity price adjustment and demand response incentive policy change. For each scenario, input the operation time within 50-100ms, substitute into the power grid safety bundle model within 60-100ms, solve the optimal power flow problem, and judge whether the adjusted power grid meets all safety bundles. The judgment is performed 50-100 times.

7. The method of claim 1, wherein the method further comprises: The S5 includes a bottom layer of a power grid geographical connection diagram, which displays the running state of 750kV transmission lines and substations in real time based on a GIS system and marks the line power utilization rate with different colors, green ≤60%, yellow 60%-80%, and red ≥80%. The middle layer is a flexible load potential distribution heat map, which represents the adjustable potential size in different regions with different colors, with dark color representing large potential and light color representing small potential. The update time is 5-10min once. The top layer is a key indicator dashboard, which displays the total adjustable potential of the whole network 5-10 groups, the potential proportion of each load cluster 50-100 groups, and the core indicators of power grid safety margin 30-60 groups in real time. It supports drilling and querying to view the potential details of specific loads in the region by clicking on the region. The detailed display group number is 100-300 groups.

8. The method of claim 1, wherein: The S5 provides input for dispatching decision-making, and the instruction response function receives the potential calling instructions issued by the dispatching system to call 100-200MW adjustable loads in the region, and returns the feasibility analysis results of instruction execution, including the expected response time and the impact on power grid safety, which is divided into safe and unsafe. The historical data query function supports dispatching personnel to query the potential calculation results and calling records in any time period. The calculation result is displayed 5-10 groups at a time, and the calling record is displayed 30-50 groups at a time.

9. The method of claim 1, wherein: The S6 includes an accuracy rate = 1- measured value - calculated value divided by measured value × 100%, with a requirement of not less than 90-98%, and a 750kV power grid simulation model containing large-scale flexible load is constructed.

10. The method of claim 1, wherein: The S6 includes prediction accuracy MAPE, response speed data processing and calculation time 1-5s, safety beam satisfaction rate 95-99%, real-time monitoring index change, when the index exceeds the preset threshold MAPE 8-10%, trigger model updating process, use the latest collected load data and actual call result to fine-tune the model parameters, compare the adjustable potential calculated by the method with the actual potential value, calculate the accuracy index through the actual load reduction in the demand response event, simulate new energy output fluctuation and power grid fault in different scenarios, evaluate power grid frequency deviation and voltage fluctuation index, and build model performance evaluation index.

Images