A charging pile cluster cooperative scheduling method and system for grid interaction

Abstract

The application discloses a kind of charging pile cluster collaborative scheduling methods and systems for power grid interaction, method includes: by edge collection terminal real-time acquisition charging pile cluster's power demand data and power grid side's time-of-use electricity price, regional load and carbon intensity signal, and using long short-term memory network predicts future preset time period's power supply and demand situation;Based on power supply and demand situation, construct the space-time coupling feature matrix of charging pile cluster;Utilize distributed optimization algorithm to solve space-time coupling feature matrix, generate cluster collaborative scheduling strategy with power grid peak valley smooth and minimum carbon emission as target;According to cluster collaborative scheduling strategy, dynamically allocate the real-time power adjustment instruction of each charging pile, and based on blockchain consensus mechanism, verify instruction execution compliance.Utilize the embodiment of the application, can realize the accurate, credible collaborative interaction of charging pile cluster and power grid dynamic supply and demand, improve the economy and low carbon of power grid operation.

Description

Technical Field

[0001] This invention belongs to the field of charging pile technology, and in particular to a method and system for collaborative scheduling of charging pile clusters oriented towards grid interaction. Background Technology

[0002] With the widespread adoption of electric vehicles, the integration of charging pile clusters into the power grid presents significant challenges to power system operation. The randomness and spatiotemporal aggregation of large-scale charging loads can easily exacerbate peak-valley differences in the power grid, leading to localized overloads and increasing the carbon emission pressure on the grid. Traditional charging pile scheduling methods are mostly based on centralized control or simple electricity price responses, making it difficult to accurately characterize the differentiated adjustment capabilities and user behavior elasticity of the massive number of charging piles within the cluster, and failing to achieve deep coordination with the dynamic supply and demand situation of the power grid and low-carbon goals. Existing technologies often focus on local power balance or economic optimization, and when dealing with grid interaction scenarios, they generally suffer from problems such as single response strategies, insufficient global coordination, and a lack of reliable execution verification mechanisms, which restrict the potential of charging pile clusters as flexible resources to participate in grid regulation. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for coordinated scheduling of charging pile clusters for grid interaction, so as to overcome the shortcomings of the prior art, and realize accurate and reliable coordinated interaction between charging pile clusters and grid dynamic supply and demand, thereby improving the economy and low carbon emissions of grid operation.

[0004] One embodiment of this application provides a method for coordinated scheduling of charging pile clusters for grid interaction, the method comprising:

[0005] The power demand data of the charging pile cluster and the time-of-use electricity price, regional load and carbon intensity signals of the power grid are obtained in real time through the edge acquisition terminal, and the power grid supply and demand situation in the future preset period is predicted by the long short-term memory network.

[0006] Based on the power grid supply and demand situation, a spatiotemporal coupling feature matrix of the charging pile cluster is constructed. The spatiotemporal coupling feature matrix includes multi-dimensional indicators such as the power adjustment potential, user flexibility, and response priority of each charging pile.

[0007] The spatiotemporal coupling feature matrix is ​​solved using a distributed optimization algorithm to generate a cluster-coordinated scheduling strategy with the objectives of grid peak-valley smoothing and minimizing carbon emissions.

[0008] The real-time power adjustment instructions for each charging pile are dynamically allocated according to the cluster collaborative scheduling strategy, and the compliance of instruction execution is verified based on the blockchain consensus mechanism.

[0009] Another embodiment of this application provides a charging pile cluster collaborative scheduling system for grid interaction, the system comprising:

[0010] The acquisition module is used to acquire power demand data of charging pile clusters and time-of-use electricity prices, regional load and carbon intensity signals on the grid side in real time through edge acquisition terminals, and to use long short-term memory network to predict the grid supply and demand situation in the future preset period.

[0011] The module is used to construct a spatiotemporal coupling feature matrix of the charging pile cluster based on the power grid supply and demand situation. The spatiotemporal coupling feature matrix includes multi-dimensional indicators such as the power adjustment potential, user flexibility, and response priority of each charging pile.

[0012] The generation module is used to solve the spatiotemporal coupling feature matrix using a distributed optimization algorithm to generate a cluster collaborative scheduling strategy with the objectives of grid peak-valley smoothing and minimizing carbon emissions.

[0013] The allocation module is used to dynamically allocate real-time power adjustment instructions for each charging pile according to the cluster collaborative scheduling strategy, and verify the compliance of instruction execution based on the blockchain consensus mechanism.

[0014] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0015] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0016] Compared with existing technologies, the present invention provides a charging pile cluster collaborative scheduling method for grid interaction, which can realize accurate and reliable collaborative interaction between charging pile clusters and grid dynamic supply and demand, thereby improving the economy and low carbon emissions of grid operation. Attached Figure Description

[0017] Figure 1 A hardware structure block diagram of a computer terminal for a charging pile cluster collaborative scheduling method for grid interaction provided in an embodiment of the present invention;

[0018] Figure 2 A flowchart illustrating a collaborative scheduling method for charging pile clusters oriented towards grid interaction, provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the structure of a charging pile cluster collaborative scheduling system for grid interaction provided in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] This invention first provides a method for coordinated scheduling of charging pile clusters for grid interaction. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0022] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a charging pile cluster collaborative scheduling method for grid interaction provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0023] The non-volatile storage medium can store the operating system and computer program. The computer program includes program instructions that, when executed, cause the processor to perform any grid-interactive charging pile cluster collaborative scheduling method.

[0024] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0025] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any method of collaborative scheduling of charging pile clusters for grid interaction.

[0026] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0027] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0028] See Figure 2 The present invention provides a method for coordinated scheduling of charging pile clusters for grid interaction, which may include the following steps:

[0029] S201 acquires power demand data of charging pile clusters and time-of-use electricity prices, regional load and carbon intensity signals from the power grid in real time through edge acquisition terminals, and uses a long short-term memory network to predict the power grid supply and demand situation for a future preset period.

[0030] Specifically, edge acquisition terminals can be deployed in the charging pile cluster to collect real-time power, voltage, current, charging status and user-set charging plan data of each charging pile. At the same time, time-of-use electricity price curves, real-time regional load data and real-time carbon intensity signals can be obtained through the power grid data interface to generate multi-source heterogeneous raw data streams.

[0031] The core of this step is to build a comprehensive data acquisition system that simultaneously captures data from the equipment side, user side, and power grid side, providing a complete data foundation for subsequent analysis and prediction. The specific implementation method is as follows:

[0032] The edge acquisition terminals adopt a distributed deployment mode, with one terminal configured for each charging pile in the cluster, and a main terminal deployed at the cluster aggregation point. Each terminal has data acquisition, local caching, and preliminary forwarding capabilities. The acquisition frequency is dynamically adjusted according to the data type to ensure real-time performance and efficient resource utilization. Specifically, real-time power, voltage, and current of the charging piles are acquired at a frequency of 1Hz (once per second), with power measurement accuracy ±0.01kW, voltage accuracy ±1V, and current accuracy ±0.1A, accurately reflecting the equipment's operating status. Charging status (standby, charging, fully charged, fault) is acquired using a status-triggered acquisition method, triggering recording only when the status changes, and synchronously marking the status transition timestamp. User-defined charging plan data is acquired through the charging pile's human-machine interface, including target charging amount, estimated completion time, and charging priority preferences. After acquisition, it is bound to the corresponding charging pile ID, and updated every 5 minutes to ensure tracking of changes in user needs.

[0033] Grid-side data is obtained through a standardized grid data interface, which employs an encrypted transmission protocol to ensure data security. Time-of-use (TOU) electricity price curves are issued by the grid dispatch center and updated every 15 minutes, containing electricity price values ​​for each time period in the next 24 hours, in yuan / kWh. Peak, flat, and valley periods are also marked (peak periods: 10:00-14:00 and 17:00-21:00, with a 30% price increase; valley periods: 00:00-06:00, with a 20% price decrease; the rest are flat periods). Real-time regional load data is acquired at a 2Hz frequency, reflecting the current total electricity load of the regional grid, in MW, and is synchronously correlated with the load fluctuation rate (MW / minute). The real-time carbon intensity signal is updated hourly, representing the carbon emission intensity of the regional grid's power supply, in kg / kWh. The value is negatively correlated with the proportion of clean energy generation; the higher the proportion of clean energy, the lower the carbon intensity.

[0034] In the example, a charging pile cluster contains 20 charging piles. At time t=10:00:00, the edge acquisition terminal collects the real-time power of charging pile 1 as 7.2kW, voltage as 220V, current as 32.7A, and charging status as "charging in progress." The user sets a charging plan with a target energy of 20kWh, expected to be completed by 11:30. Simultaneously, the current time-of-use electricity price (peak hours) is obtained through the grid interface: 0.8 yuan / kWh, regional load as 125MW, load fluctuation rate as 0.3MW / minute, and carbon intensity as 0.52kg / kWh. All data is encapsulated in the format of "device ID-timestamp-data type-value," and after aggregation, a multi-source heterogeneous raw data stream is generated, containing three main categories of data: device operating parameters, user requirements, and grid status. The total data volume is controlled within 50MB per hour to avoid transmission and storage pressure.

[0035] Time alignment and missing value imputation are performed on the multi-source heterogeneous raw data streams. Sliding window smoothing filter is used to remove abnormal fluctuations and normalize to a unified dimension to generate a standardized time series dataset.

[0036] The core of this step is to eliminate data noise, bias, and dimensional differences, improve data quality, and provide standardized input for model prediction. The specific implementation method is as follows:

[0037] Time alignment is based on a unified clock from the edge acquisition terminals, with clock synchronization accuracy controlled within ±10 milliseconds to ensure the consistency of timestamps across all data. During alignment, data from different acquisition frequencies are resampled at the highest frequency (2Hz, corresponding to grid load data). Lower frequency data is supplemented with intermediate timestamp data through linear interpolation, while high frequency data retains its original sampling points. For example, charging pile power data (1Hz) has one sample value at t=10:00:00 and another at t=10:00:01. When resampled to 2Hz, interpolated data is inserted at t=10:00:00.5, using the interpolation formula x_inter=(x1+x2) / 2, where x1 and x2 are adjacent sample values, ensuring time sequence continuity. After alignment, all data is uniformly sorted by millisecond-level timestamps to form a time-series data sequence.

[0038] Missing value imputation addresses missing data points caused by data transmission or acquisition failures, employing a categorized imputation strategy: For numerical data (power, voltage, load, electricity price), a sliding window mean interpolation is used, with the window size set to 5 time steps (2.5 seconds), and the mean of the valid data within the window is used to replace the missing value; for state data (charging status), a forward imputation method is used, retaining the most recent valid state value before the missing time. In the example, power data for charging pile 3 is missing at t=10:00:03, and the adjacent valid data within the window are 7.0kW (t=10:00:01), 7.1kW (t=10:00:02), 7.3kW (t=10:00:04), and 7.2kW (t=10:00:05), with a mean of (7.0+7.1+7.3+7.2) / 4=7.15kW, which is used as the missing value imputation result; if charging status data is missing, the "charging in progress" status from the previous time step is used.

[0039] Sliding window smoothing filtering is used to remove abnormal fluctuations and noise from numerical data. A weighted sliding window algorithm is employed, with a window size of 5 time steps. The middle data point within the window has a weight of 0.4, while the two preceding and following data points have weights of 0.2 and 0.15 respectively, with a total weight of 1. This smooths noise while preserving the core trend. The filtering formula is x_filter = 0.15×x1 + 0.2×x2 + 0.4×x3 + 0.2×x4 + 0.15×x5, where x3 is the data at the current time step, x1 and x2 are the preceding data, and x4 and x5 are the following data. In the example, charging pile 1 experiences a momentary abnormal power value of 15.8kW (far exceeding the normal charging power of 5-10kW). The filtered value, calculated using the filtering algorithm, is 0.15×7.1 + 0.2×7.2 + 0.4×15.8 + 0.2×7.3 + 0.15×7.2 = 8.93kW, correcting the abnormal fluctuation while retaining the normal power change trend.

[0040] The normalization process uses the Z-score standardization algorithm to map all numerical data to the [-3,3] interval, eliminating dimensional differences. The formula is x_norm=(x-μ) / σ, where μ is the historical mean of this type of data and σ is the standard deviation, calculated based on historical data from the past 7 days. In the example, for power data, μ=6.5kW and σ=1.2kW, the normalized value for a certain data point of 7.2kW is (7.2-6.5) / 1.2≈0.58; for regional load data, μ=110MW and σ=15MW, the normalized value for the current 125MW is (125-110) / 15=1.0; for electricity price data, μ=0.6 yuan / kWh and σ=0.15 yuan / kWh, the normalized value for the current 0.8 yuan / kWh is (0.8-0.6) / 0.15≈1.33. After all data processing is completed, the data is arranged in chronological order to form a standardized time-series dataset, which includes three types of features: equipment, users, and power grid, with eight dimensions, providing high-quality input for subsequent model predictions.

[0041] The standardized time series dataset is input into a pre-trained long short-term memory network model. The model uses an attention mechanism to enhance the focus on key time steps and extracts time series features through multi-layer LSTM units to generate a preliminary prediction sequence of power grid supply and demand in the future preset time period.

[0042] The core of this step is to rely on a pre-trained LSTM model to capture the temporal correlation of data, and combine it with an attention mechanism to enhance key information, so as to achieve an accurate preliminary prediction of the power grid supply and demand situation. The specific implementation method is as follows:

[0043] The pre-trained Long Short-Term Memory (LSTM) network model adopts an architecture of "input layer - attention layer - multi-layer LSTM layer - output layer". The input layer dimension is adapted to the 8-dimensional features of the standardized time-series dataset, and the output layer dimension is 3-dimensional, corresponding to the three core indicators of regional load, time-of-use electricity price, and carbon intensity in future periods, used to characterize the power grid supply and demand situation. The model pre-training is based on historical standardized data from the past 3 months. The mean squared error (MSE) is used as the loss function during training, and the optimizer adopts the adaptive momentum optimization algorithm. The learning rate is set to 0.001, the number of iterations is 500, and the convergence threshold is training loss MSE ≤ 0.01, ensuring that the model has stable time-series prediction capabilities.

[0044] The attention mechanism employs a temporal attention mechanism, assigning different weights to each time step of the input time-series data. This enhances feature extraction for key time steps such as sudden changes in electricity prices, load peaks, and carbon intensity fluctuations. Weight calculation is based on feature importance, with key time steps having a weight ≥ 0.3 and non-key time steps having a weight ≤ 0.1, thus improving the model's sensitivity to core information. For example, time steps such as the shift from flat to peak electricity prices, regional load exceeding 120% of the historical average, and a sharp drop in carbon intensity (due to an increase in the proportion of clean energy generation) are identified as key time steps, and their attention weights are automatically increased, ensuring that the model focuses on learning the feature associations of these time steps.

[0045] The multi-layer LSTM unit contains two hidden layers, each with 128 hidden units. The first LSTM layer extracts local temporal features (such as short-term load fluctuations and short-term electricity price changes), while the second LSTM layer extracts global temporal features (such as intraday load peak-valley patterns and periodic electricity price changes). The features from the two layers are fused to form a complete temporal feature representation. The parameters of the forget gate, input gate, and output gate of the LSTM unit are optimized through pre-training. The forget gate threshold is set to 0.5 to ensure that the model forgets redundant information and retains effective temporal features. The input gate activation function uses the Sigmoid function, with an output range of [0,1], to control the intensity of feature input. The output gate combines the tanh function to generate hidden states, which are mapped to feature representations.

[0046] The projected time period is set to 2 hours, divided into 8 prediction periods with 15-minute intervals. The model predicts regional load, time-of-use electricity price, and carbon intensity for each period, generating a preliminary prediction sequence. In the example, after the standardized time-series dataset is input into the model, the attention mechanism focuses on two key time steps: t=9:50:00 (electricity price is about to enter peak period) and t=9:55:00 (load is rising rapidly), with weights of 0.32 and 0.35, respectively. After extracting time-series features, a preliminary prediction sequence is generated: regional load is 123MW, 126MW, 128MW, 124MW, 120MW, 116MW, 113MW, and 110MW respectively; time-of-use electricity price is... The prices are 0.8 yuan / kWh, 0.85 yuan / kWh, 0.9 yuan / kWh, 0.88 yuan / kWh, 0.8 yuan / kWh, 0.7 yuan / kWh, 0.65 yuan / kWh, and 0.6 yuan / kWh, respectively; the carbon intensity is 0.52 kg / kWh, 0.51 kg / kWh, 0.50 kg / kWh, 0.49 kg / kWh, 0.48 kg / kWh, 0.47 kg / kWh, 0.46 kg / kWh, and 0.45 kg / kWh, respectively, fully reflecting the changing trend of power grid supply and demand in the next 2 hours.

[0047] Uncertainty quantification is performed on the preliminary forecast sequence, and the confidence interval of the forecast results is calculated using the Monte Carlo dropout method. Combined with the output of the historical error correction model, a power grid supply and demand situation forecast report with confidence intervals is generated.

[0048] The core of this step is to quantify the uncertainty of the prediction results, improve the prediction accuracy through error correction, and generate a prediction report that combines accuracy and reliability. The specific implementation method is as follows:

[0049] Monte Carlo dropout is used for uncertainty quantification. Its core is to enable a dropout layer during model inference, generating multiple predictions through repeated inferences, statistically analyzing the distribution characteristics of the results, and calculating confidence intervals. The dropout probability is set to 0.2, meaning 20% ​​of hidden layer neurons are randomly dropped during each inference, preventing model overfitting and generating diverse predictions. The number of repeated inferences is set to 50 to ensure the stability of the result distribution. After inference, the mean and standard deviation of predictions for each prediction period and each indicator are calculated. A 95% confidence interval is set based on a normal distribution, using the formula: Confidence Interval = [μ_pred - 1.96 × σ_pred, μ_pred + 1.96 × σ_pred], where μ_pred is the mean of 50 inferences and σ_pred is the standard deviation. A narrower confidence interval indicates lower prediction uncertainty and higher reliability.

[0050] In the example, for the regional load during the first forecast period (10:00-10:15), the mean of 50 inference results is μ_pred = 123MW, the standard deviation is σ_pred = 2.1MW, and the 95% confidence interval is [123 - 1.96 × 2.1, 123 + 1.96 × 2.1] = [118.88MW, 127.12MW]; the mean time-of-use electricity price is 0.8 yuan / kWh, the standard deviation is 0.03 yuan / kWh, and the confidence interval is [0.74 yuan / kWh, 0.86 yuan / kWh]; the mean carbon intensity is 0.52 kg / kWh, the standard deviation is 0.02 kg / kWh, and the confidence interval is [0.48 kg / kWh, 0.56 kg / kWh]. Quantifying the uncertainty of each forecast result using confidence intervals provides a risk reference for subsequent dispatch strategy formulation.

[0051] The historical error correction model is built based on the predicted and actual data from the past month. Through statistical analysis of the patterns in prediction errors, an error correction function is fitted to correct the initial prediction results and reduce systematic errors. The error is defined as error = actual - pred, where actual is the actual value and pred is the initial predicted value. Based on historical error data, a linear regression fitting formula is used: pred_corrected = a × pred + b, where a and b are correction coefficients obtained through least squares fitting. In the example, the correction coefficients for the regional load are a=1.02 and b=-2.5MW. The initial forecast value for the first forecast period is 123MW, and after correction, it is 1.02×123-2.5=122.96MW≈123.0MW. The time-of-use electricity price correction coefficients are a=0.98 and b=0.02 yuan / kWh. The initial forecast value is 0.8 yuan / kWh, and after correction, it is 0.98×0.8+0.02=0.794 yuan / kWh≈0.79 yuan / kWh. The carbon intensity correction coefficients are a=1.01 and b=-0.01kg / kWh, and after correction, it is 1.01×0.52-0.01=0.5152kg / kWh≈0.52kg / kWh.

[0052] The integrated and corrected forecast results, confidence intervals, and uncertainty analysis generate a power grid supply and demand forecast report with confidence intervals. The report includes the forecast period division, the corrected values ​​of the core indicators for each period, the 95% confidence interval, and the uncertainty level (low / medium / high, determined based on the confidence interval width: ≤5% is low uncertainty, 5%-10% is medium uncertainty, and >10% is high uncertainty). An example report excerpt reads: "Power grid supply and demand forecast for the next 2 hours (10:00-12:00): From 10:00 to 10:15, the regional load correction forecast is 123.0MW, with a 95% confidence interval [118.88MW, 127.12MW], and a medium level of uncertainty; the time-of-use electricity price correction forecast is 0.79 yuan / kWh, with a confidence interval [0.74 yuan / kWh, 0.86 yuan / kWh], and a low level of uncertainty; the carbon intensity correction forecast is 0.52 kg / kWh, with a confidence interval [0.48 kg / kWh, 0.56 kg / kWh], and a low level of uncertainty." This provides a precise basis for subsequent feature matrix construction and scheduling strategy formulation.

[0053] S202, Based on the power grid supply and demand situation, construct a spatiotemporal coupling feature matrix of the charging pile cluster. The spatiotemporal coupling feature matrix includes multi-dimensional indicators such as the power adjustment potential, user flexibility, and response priority of each charging pile.

[0054] Specifically, it can analyze the power grid supply and demand forecast report, extract the predicted power surplus or shortage, electricity price and carbon intensity for each future period, and generate a power grid situation feature vector.

[0055] The core of this step is to extract key power grid operation indicators from the forecast report, construct standardized feature vectors by time period, and provide a power grid-side benchmark for subsequent calculation of multi-dimensional indicators of charging piles. The specific implementation method is as follows:

[0056] The analysis process employs structured analysis logic. First, the report content is broken down by forecast period. The projected future period is 2 hours, divided into eight 15-minute time slots (T1-T8, corresponding to 10:00-10:15 to 11:45-12:00). For each time slot, four core indicators are extracted: power surplus / shortage, time-of-use electricity price, carbon intensity, and uncertainty level. Redundant descriptive information is removed to ensure accurate correlation between indicators and time slots. Power surplus / shortage is calculated using regional load forecasts and a grid supply capacity threshold. The grid supply capacity threshold is set at 130MW based on the current grid operating status, and the calculation formula is ΔP=P_supply-P_load, where P_supply is the supply capacity threshold, and P_load is the time slot load forecast. ΔP>0 indicates power surplus, ΔP<0 indicates power shortage, and ΔP=0 indicates supply-demand balance.

[0057] In the example, analyzing the predicted data for time period T1 (10:00-10:15): regional load 123.0MW, power supply capacity threshold 130MW, calculated ΔP=130-123.0=7.0MW (power surplus); time-of-use electricity price adjusted to 0.79 yuan / kWh, carbon intensity 0.52kg / kWh, uncertainty level medium; for time period T2 (10:15-10:30), regional load 126.0MW, ΔP=130-126.0= 4.0MW (electricity surplus), electricity price 0.84 yuan / kWh, carbon intensity 0.51kg / kWh, medium uncertainty level; During the T3 period (10:30-10:45), the regional load is 128.0MW, ΔP=2.0MW (electricity surplus), electricity price 0.89 yuan / kWh, carbon intensity 0.50kg / kWh, medium uncertainty level; During the T4-T8 periods, as the load decreases, the electricity surplus gradually increases, the electricity price shows a downward trend, and the carbon intensity continues to decrease.

[0058] After extraction, a power grid situation feature vector is constructed according to the format "time period number - power surplus / shortage - electricity price - carbon intensity". Each time period corresponds to a 4-dimensional feature vector, and the 8 time periods form an 8×4 feature matrix. The vector values ​​are all standardized results (power surplus / shortage in MW, electricity price in yuan / kWh, carbon intensity in kg / kWh) to facilitate subsequent integration with charging pile indicators. In the example, the feature vector for time period T1 is [7.0, 0.79, 0.52], and for time period T2 it is [4.0, 0.84, 0.51]. The overall vector sequence clearly reflects the trend of the power grid gradually shifting from a slight surplus to a moderate surplus, the peak electricity price declining, and the carbon emission intensity decreasing over the next 2 hours.

[0059] Based on the power grid situation feature vector, combined with the historical operation data and current charging status of each charging pile, the power boundary that can be adjusted up or down in each prediction period is calculated, its dynamic power regulation potential is evaluated, and a power regulation potential index is generated.

[0060] The core of this step is to combine the power grid situation with the individual status of each charging pile to quantify the power adjustment space of each charging pile, generate dynamic potential indicators, and provide core equipment-side basis for priority calculation. The specific implementation method is as follows:

[0061] The assessment of dynamic power regulation potential needs to consider both the characteristics of the power grid during different time periods and the individual attributes of the charging piles. Individual attributes include the current charging status (standby, charging, almost fully charged), historical operating data (charging power range and adjustment response speed for the same period in the past 7 days), and the user-set charging plan (target capacity and estimated completion time). The process follows the logic of "first determining the adjustment direction, then calculating the boundary range, and finally quantifying the potential value." The adjustment direction is determined by the power grid's surplus / shortage status: when there is a power surplus, charging piles are encouraged to increase their charging power (to absorb redundant power); when there is a power shortage, charging piles are guided to decrease their charging power (to alleviate pressure on the power grid); when supply and demand are balanced, the base power is maintained, and the adjustment potential is set to medium.

[0062] Power boundary calculations are performed in two scenarios: For charging piles in the charging process, the upward power boundary is the difference between the device's rated power and the current power, with the rated power uniformly set at 10kW, and the formula is P_up = P_rated - P_current; the downward power boundary is the difference between the current power and the minimum sustaining power, with the minimum sustaining power set at 2kW based on battery charging safety, and the formula is P_down = P_current - P_min. Simultaneously, user charging plan constraints must be considered. If increasing the power would result in earlier charging (exceeding the user's expected completion time), the upward boundary is calculated backward based on the expected completion time to ensure that core user needs are not violated. If decreasing the power would prevent the target charge from being completed within the expected time, the downward boundary is appropriately reduced to balance grid demand and user experience.

[0063] In the example, charging pile 1 is currently charging, with a current power of 7.2kW, a rated power of 10kW, a minimum sustaining power of 2kW, a user-set target capacity of 20kWh, and 8.4kWh already charged. Charging is expected to be completed at 11:30 (remaining charging time is 90 minutes, covering time periods T1-T6). During period T1, the grid has a power surplus of 7.0MW, and the adjustment direction is upward. The calculated adjustment boundary P_up = 10 - 7.2 = 2.8kW. Based on the remaining power, the allowable upward adjustment power is calculated as follows: the remaining power is 11.6kWh, and the remaining time is 90 minutes. If the current power is maintained at 7.2kW, the remaining charging time will be 11.6 ÷ 7.2 ≈ 96.7 minutes, which exceeds the expected time by 6.7 minutes. Therefore, the power can be appropriately increased to 7.6kW (requiring 11.6 ÷ 7.6 ≈ 93.7 minutes). Thus, the actual adjustment boundary during period T1 is 0.4kW (7.6 - 7.2), and the downward adjustment boundary is 7.2 - 2 = 5.2kW. The comprehensive adjustment space is 5.6kW, and the quantified potential value is 0.56 (adjustment space ÷ rated power, range 0-1, the larger the value, the stronger the potential).

[0064] For charging piles in standby mode, the upper limit for power adjustment is the rated power (10kW), and the lower limit is 0kW (no charging power to adjust). The adjustment potential value is determined by a combination of grid electricity price and user charging intention. If the current electricity price is at a low point and the user has no specific charging time constraint, the potential value is set to 0.8-1.0; if the electricity price is at a high point, it is set to 0.3-0.5. In the example, charging pile 5 is in standby mode, and the user sets a charging plan with no specific time, only needing to fully charge 20kWh on the same day. The electricity price during period T1 is 0.79 yuan / kWh (late peak period), so the potential value is set to 0.4, with an upper limit of 10kW and a lower limit of 0kW.

[0065] For charging stations nearing full charge (remaining capacity ≤ 2kWh), the adjustment potential value is set to 0.1-0.2, allowing only a small adjustment margin to avoid excessive adjustment affecting battery life. By combining the adjustment potential values ​​of all charging stations across different time periods, an 8×20 power adjustment potential matrix (8 time periods × 20 charging stations) is generated. Each value corresponds to the dynamic power adjustment potential of a single charging station during the corresponding time period, providing a basis for subsequent user flexibility assessments.

[0066] Based on the power adjustment potential index and users' historical charging behavior data, we analyze users' sensitivity to the adjustment range of electricity price and power at different times, use fuzzy logic to evaluate users' resilience, and generate user resilience index.

[0067] The core of this step is to quantify users' acceptance of power adjustments using fuzzy logic algorithms, generate a resilience index, and balance grid dispatching needs with user experience. The specific implementation method is as follows:

[0068] User historical charging behavior data is extracted from local training files, covering three main categories of information over the past 30 days: user charging time preferences, electricity price sensitivity, and power adjustment response records. Charging time preferences are determined by statistically analyzing the percentage of charging times in each time period. Electricity price sensitivity is calculated by analyzing the charging power change rate in different electricity price ranges. Power adjustment response records are statistically analyzed to show the user's execution rate of past dispatch instructions (execution times ÷ instruction times). An execution rate of ≥80% is considered a high response, 60%-80% is considered a medium response, and <60% is considered a low response.

[0069] The fuzzy logic evaluation adopts a four-step process of "input-fuzzification-rule reasoning-defuzzification". Two input variables are set: electricity price sensitivity (X1, range 0-1, the larger the value, the more sensitive to electricity prices) and power adjustment tolerance (X2, range 0-1, the larger the value, the more accepting of adjustments). The output variable is user elasticity (Y, range 0-1, 0 for completely inelastic, 1 for completely elastic). Electricity price sensitivity X1 is calculated using historical data: X1 = (percentage of charging times during peak hours × 0.3 + power change rate during electricity price fluctuations × 0.7), where the power change rate during electricity price fluctuations = (off-peak electricity price power - peak electricity price power) ÷ rated power.

[0070] In the example, historical data for user 1 of charging pile: peak hours account for 30% of charging frequency, off-peak hours account for 40%, and average hours account for 30%. The power change rate during electricity price fluctuations is (8.5-6.2)÷10=0.23, so X1=0.3×0.3+0.23×0.7=0.09+0.161=0.251 (low electricity price sensitivity). The power adjustment range acceptance X2, combined with power adjustment potential and past response records: past adjustment execution rate for ≤1kW was 90%, 1-2kW was 75%, and >2kW was 50%. The current adjustment range is 0.4kW, so X2=0.9×(1-0.4 / 2)+0.75×(0.4 / 2)=0.9×0.8+0.75×0.2=0.72+0.15=0.87 (high acceptance).

[0071] The fuzzification process uses a triangular membership function to divide X1, X2, and Y into three fuzzy subsets: low (L), medium (M), and high (H). The membership function parameters for X1 are: low (0,0,0.3), medium (0.2,0.5,0.8), and high (0.7,1,1); for X2, they are: low (0,0,0.4), medium (0.3,0.6,0.9), and high (0.8,1,1); and for Y, they are: low (0,0,0.3), medium (0.2,0.5,0.8), and high (0.7,1,1). Membership degrees are calculated based on the input variable values. In the example, X1=0.251 corresponds to a low membership degree of 0.49, a medium membership degree of 0.51, and a high membership degree of 0; X2=0.87 corresponds to a low membership degree of 0, a medium membership degree of 0.1, and a high membership degree of 0.9.

[0072] The fuzzy rule inference sets nine core rules: 1. If X1 is low and X2 is low, then Y is low; 2. If X1 is low and X2 is medium, then Y is medium; 3. If X1 is low and X2 is high, then Y is medium-high; 4. If X1 is medium and X2 is low, then Y is low-medium; 5. If X1 is medium and X2 is medium, then Y is medium; 6. If X1 is medium and X2 is high, then Y is high; 7. If X1 is high and X2 is low, then Y is low; 8. If X1 is high and X2 is medium, then Y is medium; 9. If X1 is high and X2 is high, then Y is high. Based on the membership degree matching rules of the input variables, matching rule 3 in the example, the fuzzy subset of Y is determined to be medium-high.

[0073] Defuzzification employs the centroid method, calculating the centroid of the fuzzy subset as the final elasticity value. The formula is Y=∫y×μ_Y(y)dy / ∫μ_Y(y)dy, where μ_Y(y) is the membership function of Y. In the example, Y=0.68 is calculated, meaning that the elasticity index of user 1 at charging pile 1 in time period T1 is 0.68 (medium-high elasticity), indicating that the user has a high acceptance of the current power adjustment and the adjustment range can be appropriately increased. Following this process, the elasticity index of all charging piles in each time period is calculated, generating an 8×20 user elasticity matrix, consistent with the dimension of the power adjustment potential matrix, providing a user-side basis for priority calculation.

[0074] By combining the power regulation potential index, user resilience index, and the node voltage stability contribution of charging piles connected to the power grid, the response priority of each charging pile in different time periods is calculated, and a spatiotemporal coupling feature matrix is ​​constructed by arranging them in a spatiotemporal dimension.

[0075] The core of this step is to integrate multi-dimensional indicators to calculate response priority, and integrate them according to the spatiotemporal dimension to form a feature matrix, providing complete input for subsequent optimization and solution. The specific implementation method is as follows:

[0076] The response priority calculation adopts a weighted summation method to construct a priority evaluation model: P = α × P_potential + β × P_elasticity + γ × P_voltage, where P is the response priority (range 0-1, the larger the value, the higher the priority), P_potential is the power regulation potential index, P_elasticity is the user resilience index, P_voltage is the node voltage stability contribution, and α, β, and γ are weight coefficients determined based on the analytic hierarchy process. Taking into account the grid dispatching needs and user experience, α = 0.4 (equipment regulation capability has the highest weight), β = 0.3 (user acceptance has the second highest weight), and γ = 0.3 (grid security has the third highest weight), with a total weight of 1.

[0077] The node voltage stability contribution value, P_voltage, is used to assess the impact of charging pile access on the grid node voltage. Through node voltage sensitivity analysis, the reference voltage value for a charging pile connected to the grid is 220V, with an allowable voltage deviation range of ±5% (209V-231V). P_voltage = 1 - |U_actual - U_ref| / U_ref, where U_actual is the actual voltage of the node to which the charging pile is connected, and U_ref is the reference voltage value, ranging from 0 to 1. A value closer to 1 indicates a greater contribution to voltage stability and is more beneficial to the safe operation of the grid. In the example, the actual voltage of charging pile 1 connected to the node is 223V, so P_voltage = 1 - |223 - 220| / 220 = 1 - 3 / 220 ≈ 0.986; the actual voltage of charging pile 8 connected to the node is 218V, so P_voltage = 1 - |218 - 220| / 220 ≈ 0.991; and the actual voltage of charging pile 15 connected to the node is 225V, so P_voltage = 1 - 5 / 220 ≈ 0.977.

[0078] Based on the example data above, the response priority is calculated as follows: For charging pile 1 during time period T1, P_potential = 0.56, P_elasticity = 0.68, and P_voltage = 0.986. Substituting these values ​​into the formula, we get P = 0.4 × 0.56 + 0.3 × 0.68 + 0.3 × 0.986 = 0.224 + 0.204 + 0.2958 = 0.7238 (high priority); For charging pile 5 (in standby mode), during time period T1, P_potential = 0.4, P_elasticity = 0.75 (user has no time constraints, higher flexibility), and P_voltage = ... 0.982, calculated as P=0.4×0.4+0.3×0.75+0.3×0.982=0.16+0.225+0.2946=0.6796 (medium-high priority); Charging pile 18 (almost fully charged, remaining power 1.8kWh) T1 time period P_potential=0.15, P_elasticity=0.3 (low tolerance for adjustment), P_voltage=0.988, calculated as P=0.4×0.15+0.3×0.3+0.3×0.988=0.06+0.09+0.2964=0.4464 (medium-low priority).

[0079] This method calculates the response priority of all 20 charging piles across 8 time periods, generating an 8×20 priority matrix. This matrix is ​​then integrated with the previously constructed power regulation potential matrix, user flexibility matrix, and node voltage contribution matrix to build a spatiotemporal coupled feature matrix. The matrix dimensions are 8 (time periods) × 20 (charging piles) × 4 (indicators: power regulation potential, user flexibility, node voltage contribution, response priority), arranged according to the spatiotemporal dimension. Each element corresponds to a quantified value for "specific time period - specific charging pile - specific indicator." All values ​​are uniformly normalized to the 0-1 range to ensure the consistency and comparability of the matrix data.

[0080] The example spatiotemporal coupling feature matrix segment (first 3 charging piles in time period T1): Charging pile 1 corresponds to [0.56, 0.68, 0.986, 0.724]; Charging pile 2 corresponds to [0.62, 0.71, 0.983, 0.751]; Charging pile 3 corresponds to [0.48, 0.63, 0.990, 0.692]. This matrix fully integrates information from four dimensions: the spatiotemporal status of the power grid, the capabilities of the charging pile equipment, user behavior preferences, and power grid safety requirements. This provides comprehensive and accurate input data support for subsequent distributed optimization algorithm solutions, ensuring the scientific validity and feasibility of the scheduling strategy.

[0081] S203, use a distributed optimization algorithm to solve the spatiotemporal coupling feature matrix to generate a cluster collaborative scheduling strategy with the goal of smoothing power grid peak and valley and minimizing carbon emissions;

[0082] Specifically, the objective functions can be grid peak-valley smoothing and carbon emission minimization, with the charging pile power adjustment range and user charging demand as constraints, to establish a multi-objective optimization model and generate a mathematical model for the optimization problem.

[0083] The core of this step is to transform the scheduling objectives and constraints into standardized mathematical expressions, balancing power grid operating efficiency and user needs, and providing a theoretical framework for subsequent algorithm solutions. The specific implementation method is as follows:

[0084] The objective function design adopts a weighted summation method to integrate the two objectives, taking into account the priority of power grid peak-valley smoothing and carbon emission minimization. The weights are dynamically set based on the power grid dispatching needs. The peak-valley smoothing weight ω1=0.55 (to ensure the stable operation of the power grid), and the carbon emission minimization weight ω2=0.45 (to respond to the low-carbon target). The objective function as a whole is a minimization function, expressed as minF=ω1×F1+ω2×F2.

[0085] The power grid peak-valley smoothing objective F1 is characterized by minimizing the regional load variance. The smaller the variance, the smoother the load fluctuation and the smaller the peak-valley difference. The formula is F1=Var(P_load_total(t))=E[P_load_total(t)]. 2 ]-(E[P_load_total(t)]) 2 Where P_load_total(t) is the total load of the area during time period t (including the load of the charging pile cluster), and E[・] represents the expected calculation, which is based on the load forecast values ​​for the next 8 time periods. At the same time, the charging pile load adjustment amount ΔP_i(t) (the power adjustment value of the i-th charging pile during time period t) is introduced to associate the total load with the charging pile scheduling, that is, P_load_total(t)=P_load_grid(t)+ΣΔP_i(t), where P_load_grid(t) is the load forecast value of the non-charging pile area during time period t.

[0086] The minimum carbon emission target F2 is achieved by minimizing the total carbon emissions during the charging process of the charging pile cluster. The formula is F2=ΣΣ[ΔP_i(t)×Δt×C(t)], where Δt is the duration of the time period (15 minutes = 0.25 hours), C(t) is the predicted value of the grid carbon intensity during time period t (kg / kWh), and ΔP_i(t) is positive, indicating an increase in charging power (carbon emissions corresponding to electricity consumption), and negative, indicating a decrease in charging power (reducing carbon emissions).

[0087] The constraints are divided into three categories to ensure the feasibility and safety of the optimization solution. First, there is the power adjustment range constraint: P_i_min ≤ P_i(t) ≤ P_i_max, where P_i(t) is the target power of the i-th charging pile during time period t, P_i_min is the minimum sustaining power (2kW, ensuring battery charging safety), and P_i_max is the rated power (10kW). Simultaneously, ΔP_i(t) = P_i(t) - P_i_current(t) must satisfy the power adjustment boundary calculated above (upward adjustment not exceeding P_up, downward adjustment not exceeding P_down). Second, there is the user charging demand constraint: Σ[P_i(t) × Δt] ≥ Qi_i_remaining, where Qi_i_remaining is the remaining charging capacity of the i-th charging pile. The summation range covers the remaining charging period, ensuring that the target charging capacity is met within the user's expected completion time. Third, the grid voltage constraint: U_i_min≤U_i(t)≤U_i_max, where U_i(t) is the voltage of the i-th charging pile at the node t during the time period, U_i_min=209V, U_i_max=231V (voltage deviation ±5%). By analyzing the voltage sensitivity, the power adjustment amount is correlated with the voltage change to avoid voltage exceeding the limit.

[0088] In the example, a model is established for the T1 period (10:00-10:15), with ω1=0.55, ω2=0.45, P_load_grid(T1)=115.8MW (non-charging pile load), C(T1)=0.52kg / kWh, the current power of charging pile 1 is 7.2kW, the remaining power is 11.6kWh, and the remaining time period is T1-T6 (90 minutes). The power adjustment boundary is adjusted upward by 0.4kW and downward by 5.2kW. The constraints are: 2kW≤P_1(T1)≤7.6kW (upward adjustment boundary constraint), Σ[P_1(t)×0.25]≥11.6kWh (t=T1-T6), 209V≤U_1(T1)≤231V. The objective function seeks the optimal power adjustment value for this period and subsequent periods by integrating F1 and F2.

[0089] The distributed alternating direction multiplier method is used to decompose the mathematical model of the optimization problem, decomposing the global problem into sub-problems of each charging station, and generating a set of distributed sub-problems;

[0090] The core of this step is to decompose the global optimization problem using a distributed algorithm, avoiding the communication pressure and single-point failure risk of centralized solutions, allowing each charging station to solve sub-problems autonomously. The specific implementation method is as follows:

[0091] The core idea of ​​the Distributed Alternating Directional Multiplier Method (ADMM) is to decompose the global optimization problem into multiple local subproblems. By introducing dual variables and consistency constraints, it achieves a balance between local solutions and global collaboration, making it suitable for scenarios with dispersed charging pile clusters and high data privacy protection requirements. The decomposition logic is based on "global objective = Σ local objective + consistency constraints," breaking down the original multi-objective optimization model into 20 charging pile subproblems (corresponding to 20 devices). Each subproblem only relies on its own data (power boundary, user demand, access node voltage), without needing to obtain complete global data, thus reducing communication overhead.

[0092] The decomposition process consists of three steps: First, a consistency variable Z_i(t) is introduced, and Z_i(t) = P_i(t) is set to indicate that the target power of each charging pile must satisfy global consistency (i.e., the solution of the subproblem must be coordinated with the global solution); second, an augmented Lagrangian function is constructed, which combines the original objective function with the consistency constraint and the Lagrangian multiplier, and the expression is L_ρ = Σ[ω1×F1_i+ω2×F2_i] + Σλ_i(t)(Z_i(t)-P_i(t)) + (ρ / 2)Σ||Z_i(t)-P_i(t)|| 2 , where λ_i(t) is the Lagrange multiplier (dual variable), ρ is the penalty parameter (controlling the consistency constraint strength, set to 1.0), ||·|| is the Euclidean norm, and F1_i and F2_i are the local contributions of the i-th charging pile to the global objectives F1 and F2.

[0093] Subsequently, the augmented Lagrangian function is decomposed for each individual charging pile, resulting in a subproblem for each pile. The objective of each subproblem is to minimize the local augmented Lagrangian function. The constraints retain only the power adjustment range, user charging demand, and voltage constraints for that charging pile, consistent with the local constraints of the global problem. The penalty parameter ρ needs to be adaptively adjusted according to the cluster size. For a cluster of 20 charging piles, ρ is set to 1.0. If the cluster size increases, ρ can be increased to 1.2-1.5 to strengthen consistency constraints and prevent local solutions from deviating from the global optimum.

[0094] In the example, the subproblem for charging pile 1 during time T1 is: minL_ρ1=0.55×F1_1+0.45×F2_1+λ_1(T1)(Z_1(T1)-P_1(T1))+0.5×||Z_1(T1)-P_1(T1)|| 2The constraints are 2kW≤P_1(T1)≤7.6kW, P_1(T1)×0.25+Σ[P_1(t)×0.25](t=T2-T6)≥11.6kWh, and 209V≤U_1(T1)≤231V. Here, F1_1 represents the contribution of the power adjustment of charging pile 1 to the regional load variance, F2_1=P_1(T1)×0.25×0.52, λ_1(T1) is initially set to 0 (updated during iteration), and Z_1(T1) is initially set to the current power of 7.2kW.

[0095] Following this logic, all 20 charging piles are broken down into sub-problems in 8 time periods, generating an 8×20 distributed set of sub-problems. Each sub-problem includes a local objective function, constraints, and initial variable values. Sub-problems are associated with each other through consistency variables and Lagrange multipliers, laying the foundation for subsequent local iterative solutions.

[0096] Each charging pile solves its subproblem locally, exchanges intermediate results with neighboring nodes through a finite number of iterations, and gradually converges to the global optimal solution, generating a sequence of local optimal solutions;

[0097] The core of this step is that each charging pile autonomously solves the sub-problem, synchronizes intermediate results through neighbor communication, iteratively updates variable values, and achieves convergence of local solutions to the global optimal solution. The specific implementation method is as follows:

[0098] The local solution employs gradient descent, a method suitable for minimizing continuous variables. It boasts fast convergence and low computational cost, perfectly suited to the computing power requirements of charging pile controllers. During the solution process, each charging pile only needs to access its own local data (historical operational data, user charging plans, and access node voltage data), eliminating the need to upload sensitive data and ensuring data privacy. The learning rate for gradient descent is set to 0.01, and the iteration step size adaptively adjusts with the gradient. A larger absolute gradient value results in a smaller step size (avoiding overshoot), while a smaller absolute gradient value results in a larger step size (accelerating convergence).

[0099] The iterative process follows the ADMM's three-step cycle of "local update - neighbor communication - multiplier update": The first step, local update, involves each charging pile solving a subproblem based on the current Lagrange multiplier λ_i(t) and the consistency variable Z_i(t) to obtain the optimal P_i^(k)(t) (the target power for the k-th iteration); the second step, neighbor communication, uses a distributed topology (each charging pile establishes a communication link with 3-4 neighboring devices), exchanging the locally updated P_i^(k)(t) and Z_i^(k)(t), and taking the average of the P values ​​of the neighboring nodes. The value is used as a new consistency variable Z_i^(k+1)(t), i.e., Z_i^(k+1)(t)=(1 / (N+1))×[P_i^(k)(t)+ΣP_j^(k)(t)] (j is the neighbor node, N is the number of neighbors); the third step is multiplier update, which updates the Lagrange multiplier based on the deviation between the new consistency variable and the local solution, with the formula λ_i^(k+1)(t)=λ_i^(k)(t)+ρ×(Z_i^(k+1)(t)-P_i^(k)(t)), which strengthens the penalty for deviation.

[0100] Two convergence conditions are set: first, the global consistency deviation ε1 = ||Z_i^(k+1)(t) - P_i^(k)(t)|| ≤ 0.01kW (the deviation between the local solution and the consistency variable of each charging pile is less than 10W to ensure global coordination); second, the change in the objective function ε2 = |F^(k+1) - F^(k)| ≤ 0.001 (the difference in the global objective function between two iterations is less than 0.1% to ensure the stability of the solution). The maximum number of iterations is set to 50. If the convergence condition is not met after 50 iterations, the current optimal solution is taken as the local solution to avoid scheduling delays caused by excessive iteration.

[0101] In the example, the iteration process of charging pile 1 (with neighboring nodes being charging piles 2, 3, and 4) during time period T1 is as follows: When k=1, λ_1=0, Z_1=7.2kW, and the local solution yields P_1^(1)=7.3kW (close to the upper limit, balancing carbon emission reduction and user demand); after exchanging data with neighbors, the P^(1) of charging piles 2, 3, and 4 are 7.5kW, 7.1kW, and 7.4kW respectively, and Z_1^(2)=(7.3+7.5+7.1+7.4) / 4=7.325kW; the multiplier λ_1^(2) is updated to 0+1.0×(7.325-7.3)=0.025. When k=2, based on λ_1^(2)=0.025 and Z_1^(2)=7.325kW, the local solution yields P_1^(2)=7.33kW; after exchanging neighbor data, Z_1^(3)=7.34kW, and λ_1^(3)=0.025+1.0×(7.34-7.33)=0.035 is updated. The iteration is repeated until k=12, ε1=||7.35-7.35||=0kW, ε2=|F^(12)-F^(11)|=0.0008≤0.001, which meets the convergence condition, so the iteration stops, and P_1^(12)=7.35kW is the local optimization solution for time period T1.

[0102] Following this process, each charging pile independently iterates and solves the problem in each time period, recording the local solution for each iteration and generating a sequence of local optimal solutions. This sequence includes the iteration number, the target power for each iteration, the consistency variable, the Lagrange multiplier, and the objective function value. In the example, the local optimal solution sequence for charging pile 1 in time period T1 is [(k=1,7.3kW),(k=2,7.33kW),...,(k=12,7.35kW)], clearly showing the convergence process of the solution and ensuring the reliability of the local solution.

[0103] By aggregating the sequence of locally optimized solutions and performing a global consistency check, a globally optimal scheduling strategy that satisfies all constraints is obtained, and a cluster collaborative scheduling strategy is generated.

[0104] The core of this step is to integrate all local optimization solutions, verify global consistency and constraint satisfaction, and finally form an executable cluster collaborative scheduling strategy. The specific implementation method is as follows:

[0105] The aggregation of local optimization solution sequences is performed by the main terminal at the cluster convergence point. The main terminal receives the converged local solutions uploaded by each charging pile (without needing to upload iterative process data), integrates them according to the "time period-charging pile" dimension, and generates an 8×20 global solution matrix. The matrix elements are the target power P_i(t) of each charging pile in the corresponding time period. At the same time, the total power adjustment amount ΣΔP_i(t) of the cluster in each time period is calculated, and the grid supply and demand situation is correlated to evaluate the overall contribution to peak-valley smoothing and carbon emission reduction.

[0106] The global consistency verification is divided into three layers: The first layer is solution consistency verification, which calculates the average consistency deviation of all charging piles in each time period. The formula is ε_avg=Σ||Z_i(t)-P_i(t)|| / N (N is the number of charging piles). If ε_avg≤0.01kW, it means that the global solution consistency meets the standard. The second layer is constraint satisfaction verification, which checks whether the target power of each charging pile meets the power adjustment range, user charging demand and voltage constraints. If a charging pile solution violates the constraints, it is fed back to the device for re-iteration (limited to 3 supplementary iterations). If it still does not meet the requirements, the constraint boundary is adjusted (prioritizing user charging demand). The third layer is target achievement verification, which calculates the global objective function value F after optimization and compares it with the initial value before optimization. If the peak-valley smoothing effect is improved by ≥15% and carbon emissions are reduced by ≥10%, it means that the scheduling target has been achieved. Otherwise, the weights ω1 and ω2 are finely adjusted and the solution is recalculated.

[0107] In the example, after the global solution matrix aggregation in time period T1, ε_avg=0.008kW≤0.01kW, which meets the consistency target; the constraint verification found that the target power of charging pile 18, 1.8kW, was lower than the minimum maintenance power of 2kW. After feedback, it was iterated 3 times and adjusted to 2.0kW, which meets the constraint; the target achievement verification showed that after optimization, the regional load variance decreased from 12.5 before optimization to 8.2 (a reduction of 34.4%), and the carbon emissions decreased from 1.8kg before optimization to 1.5kg (a reduction of 16.7%), both of which meet the target requirements.

[0108] After successful verification, the global solution matrix is ​​transformed into a cluster collaborative scheduling strategy. The strategy content is divided into time periods, each containing the target power, adjustment range, execution timing, and safety precautions for each charging pile. The scheduling priorities for each time period are also marked (e.g., during periods of power surplus, priority is given to increasing the power of charging piles for high-elasticity users; during periods of high carbon intensity, priority is given to reducing the power of low-priority charging piles). In the example, the scheduling strategy segment for time period T1 is as follows: "Charging pile 1: Target power 7.35kW, adjustment range +0.15kW, executed promptly at 10:00, maintaining the access node voltage at 223V; Charging pile 5: Target power 8.0kW, adjustment range +8.0kW (standby start charging), executed at 10:02, absorbing redundant power from the grid; Charging pile 18: Target power 2.0kW, adjustment range -0.5kW, executed at 10:00, ensuring battery charging safety."

[0109] The scheduling content of all time periods is integrated to generate a complete cluster collaborative scheduling strategy. The strategy adopts a standardized format, which is convenient for subsequent parsing into power adjustment instructions. At the same time, it is stored in the local database as a historical scheduling record to provide data support for subsequent optimization model iteration.

[0110] S204, dynamically allocate real-time power adjustment instructions for each charging pile according to the cluster collaborative scheduling strategy, and verify the compliance of instruction execution based on the blockchain consensus mechanism.

[0111] Specifically, it can parse the cluster collaborative scheduling strategy, generate power adjustment instructions for each charging pile, including target power value, adjustment time window and execution priority, and generate a power adjustment instruction set;

[0112] The core of this step is to transform the abstract scheduling strategy into standardized, executable device instructions, clarifying the operational requirements of each charging pile, and providing a clear basis for subsequent issuance and execution. The specific implementation method is as follows:

[0113] The parsing process employs a structured mapping logic. First, the cluster collaborative scheduling strategy is split into time periods, extracting core parameters such as the target power and adjustment range of each charging pile within each time period. Then, operational details are added to form a complete instruction. The instruction must include four core elements: a unique device identifier (charging pile ID, such as pile 1-pile 20), a target power value (unit kW, accurate to 0.01kW, matching the device's adjustment precision), an adjustment time window (including start time and duration, with the start time accurate to the second and the duration consistent with the time period, which is 15 minutes), and execution priority (divided into three levels according to the response priority in the scheduling strategy: P1 high priority, P2 medium priority, and P3 low priority, with priority determining the instruction execution order and resource usage priority).

[0114] During parsing, the response priority in the spatiotemporal coupling feature matrix mentioned earlier needs to be considered. High response priority (≥0.7) corresponds to P1, medium response priority (0.4-0.7) corresponds to P2, and low response priority (<0.4) corresponds to P3. Simultaneously, it is necessary to verify whether the target power value meets the power adjustment boundary constraints. If the target power in the strategy exceeds the previously calculated upward / downward adjustment boundary, it is automatically corrected to the maximum boundary value to ensure command feasibility and avoid equipment overload or unauthorized adjustments.

[0115] In the example, the scheduling strategy for time period T1 (10:00:00-10:15:00) is analyzed as follows: Charging pile 1 has a response priority of 0.724 (P1), a target power of 7.35kW, an adjustment range of +0.15kW (current power 7.2kW), and generates the following instruction: ID=Pile 1, target power=7.35kW, adjustment time window=10:00:00-10:15:00, execution priority=P1; Charging pile 5 has a response priority of 0.6796 (P2), a target power of 8.0kW (standby start charging, upward adjustment boundary 10kW), and an adjustment range of... Charging pile 18 has a response priority of 0.4464 (P3), a target power of 2.0kW (minimum maintenance power after correction), and an adjustment time window of 10:02:00-10:17:00 (delayed start by 2 minutes to avoid peak execution of high-priority commands during the same period). The adjustment range is -0.5kW. The command is: ID=Pile 18, target power=2.0kW, adjustment time window=10:00:00-10:15:00, and execution priority=P3.

[0116] Based on this logic, the scheduling strategies for all 8 time periods and 20 charging piles are analyzed to generate a time-sequential power adjustment instruction set. The instruction set is stored in groups according to time periods. Each group of instructions contains the operation instructions for all charging piles in the corresponding time period. At the same time, an instruction index table is generated, which marks the instruction ID, device ID, time period, and priority to facilitate subsequent issuance, traceability, and verification. The instruction format adopts a standardized text format to ensure that the charging pile controller can parse it.

[0117] The power adjustment command set is sent to each charging pile controller through the edge communication network, and the command confirmation receipt is received.

[0118] The core of this step is to build a reliable edge communication link to enable efficient and secure command delivery, while using a receipt mechanism to confirm the device's reception status and prevent command loss or transmission anomalies. The specific implementation method is as follows:

[0119] The edge communication network adopts a dual-mode architecture of "WLAN + cellular network backup". Within the cluster, charging piles and edge main terminals communicate via WLAN (transmission rate ≥10Mbps, latency ≤50ms) to meet real-time scheduling requirements. When the WLAN signal is interrupted, it automatically switches to the cellular network to ensure communication continuity. The communication protocol uses an encrypted transmission protocol, encrypting command data end-to-end. The encryption key uses a dynamic generation mechanism, updated hourly to prevent commands from being tampered with or stolen.

[0120] Command issuance employs a "batch priority issuance" strategy. First, high-priority P1 commands are issued, followed by medium-priority P2 commands after a 1-second interval, and then low-priority P3 commands after a 2-second interval. This avoids communication congestion caused by simultaneous issuance. Each batch of commands is limited to no more than 5 to ensure transmission stability. During issuance, the edge master terminal monitors the transmission status in real time. Commands that fail to be issued (transmission timeout ≥ 100ms) are automatically retransmitted, with a maximum of 3 retransmissions. If 3 retransmissions fail, the command is marked as "issued abnormal," triggering a local alarm and recording the abnormal device ID and cause for subsequent manual investigation.

[0121] After receiving the instruction, the charging pile controller first verifies the instruction format and signature. Once the instruction source is confirmed to be legitimate and the format correct, it parses the instruction content and stores it in the local cache, simultaneously generating an instruction confirmation receipt. The receipt includes the device ID, instruction ID, reception timestamp, and verification result (success / failure). If verification fails, the reason must be noted (incorrect format / signature error / power boundary conflict). The receipt is fed back to the edge master terminal via the original communication link. After receiving the receipt, the master terminal compares it with the issued instruction, forming a closed-loop record of "issue-receipt".

[0122] In the example, during time period T1, priority commands P1 (stakes 1, 2, and 3) were sent precisely at 10:00:00. With a 30ms wireless LAN transmission delay, the controllers for stakes 1-3 all passed verification and returned an acknowledgment at 10:00:00.04: "ID=stake 1, command ID=CMD1001, reception time=10:00:00.03, verification result=success". The P2 command for stake 5 was sent at 10:00:01, transmission was normal, and the acknowledgment was successful. Stake 12 experienced a timeout on its first transmission due to a weak wireless signal; after two retransmissions, it was successfully received and an acknowledgment was returned at 10:00:03.2. Stake 19 failed to transmit three times, was marked as having a transmission error, and triggered a local alarm. The edge master terminal aggregated all acknowledgments and generated a command transmission status report, providing a basis for subsequent execution tracking.

[0123] The charging pile executes the power adjustment command and uploads the actual power adjustment data to the blockchain. The blockchain nodes collect the execution data and generate a power adjustment execution record.

[0124] The core of this step is to ensure that the charging piles accurately execute instructions, while using blockchain technology to solidify the execution data, achieving data immutability and traceability, and providing a real data source for compliance verification. The specific implementation method is as follows:

[0125] After receiving and verifying the command, the charging pile controller initiates power adjustment according to the adjustment time window. The adjustment process employs a closed-loop control algorithm, collecting its own output power in real time, comparing it with the target power value, and dynamically adjusting the output parameters to ensure power adjustment accuracy ≤ ±0.05kW, avoiding overshoot or excessive fluctuations. During the adjustment process, operating data is recorded synchronously, including the actual power curve (data collected every 1 second), adjustment start time, stabilization time (the time it takes for the power to reach the target value and maintain it for 3 seconds), and operating status (normal / abnormal). If an abnormality occurs (such as the power failing to reach the target value or the voltage exceeding the limit), the time of the abnormality, the abnormal code, and emergency handling measures (such as pausing adjustment and restoring the initial power) must be recorded.

[0126] The actual power adjustment data is uploaded to the blockchain using an "edge node preprocessing + batch upload" model. Three blockchain edge nodes are deployed within the cluster, each corresponding to a charging pile in a different region. Each charging pile uploads its execution data to the corresponding edge node in real time. The edge nodes preprocess the data: removing redundant data, calculating the data hash value (using the SHA-256 hash algorithm to generate a 64-bit hash value to ensure data uniqueness and integrity), and grouping and organizing the data by time window and device ID. The preprocessed data is uploaded to the blockchain in batches every 5 minutes to avoid network congestion caused by high-frequency uploads. The uploaded data includes core fields: device ID, instruction ID, actual target power, power curve summary (hash value), adjustment time window, execution status, exception records (empty if none), upload node ID, and upload timestamp.

[0127] The blockchain nodes adopt a consortium blockchain architecture, including edge nodes, power grid dispatch nodes, operation and maintenance nodes, and other consensus nodes. Edge nodes collect execution data from charging piles within their jurisdiction and synchronize it to other consensus nodes. Each node performs preliminary verification of the data (hash value verification, timestamp consistency verification), and stores the data in its local node ledger after successful verification. Simultaneously, the blockchain nodes aggregate execution data by time period, generating power adjustment execution records. Each record corresponds to the execution status of one instruction, including instruction information, execution data summary, on-chain node information, and verification results. Records are indexed by instruction ID for easy subsequent smart contract calls and queries.

[0128] In the example, charging pile 1 initiates power adjustment at 10:00:00 as instructed. The closed-loop control algorithm dynamically adjusts the power, and at 10:00:02.5 seconds, the power stabilizes at 7.35kW with an adjustment accuracy of 0.00kW, indicating normal operation. Power data is collected every second, generating a power curve based on 900 data points over 15 minutes. At 10:05:00, the corresponding edge node performs preprocessing on the execution data of charging pile 1, calculates the hash value of the power curve as "7a9f3d...", and uploads the data along with that of the other four charging piles to the blockchain in batches. After receiving data, the blockchain edge nodes verify the hash value and timestamp, and synchronize them to the power grid dispatch and maintenance nodes. Each node stores the data and sends back confirmation. The nodes aggregate the data to generate an execution record: "Instruction ID=CMD1001, Device ID=Pile 1, Actual Power=7.35kW, Execution Time Window=10:00:00-10:15:00, Execution Status=Normal, Power Curve Hash=7a9f3d..., On-Chain Node=Edge Node 1, On-Chain Time=10:05:03". During the execution of Pile 18, a brief voltage fluctuation occurred, and the exception code "U001" was recorded. The emergency handling measure was "maintain 2.0kW power operation". The relevant exception information was synchronized to the blockchain to ensure the integrity and traceability of the execution data.

[0129] Blockchain-based smart contracts automatically verify the consistency between execution records and instructions, achieve compliance verification results through consensus mechanisms, and write the verification results into an immutable distributed ledger.

[0130] The core of this step is to rely on smart contracts to achieve automated compliance verification, ensure the credibility of the verification results through a consensus mechanism, and finally solidify the results into the blockchain, forming a complete scheduling-execution-verification closed loop. The specific implementation method is as follows:

[0131] The smart contract is deployed on the blockchain using a pre-built template. The contract content includes verification rules, deviation thresholds, consensus triggering conditions, and result judgment logic, and can automatically execute verification without manual intervention. The verification rules focus on three core consistency aspects: First, power consistency, the deviation between the actual stable power and the commanded target power is ≤±5% (i.e., the allowable deviation is ≤0.37kW, adapting to the adjustment accuracy of the equipment). If the deviation exceeds this, it is judged as a power violation. Second, time consistency, the actual adjustment start time is within the command time window (allowing an advance / delay of ≤30 seconds). If it exceeds this, it is judged as a time violation. Third, state consistency, the execution state is normal, with no unreported abnormal records. If there are abnormalities and no reasonable explanation is given, it is judged as a state violation.

[0132] The smart contract triggering mechanism is "data on-chain triggering." Once a blockchain node aggregates and stores execution records, it automatically triggers the corresponding smart contract. The contract calls the instruction data and execution record data, comparing and calculating each item according to verification rules. In the example, the smart contract calls the instruction data of stub 1 (target power 7.35kW, time window 10:00:00-10:15:00) and the execution record (actual power 7.35kW, start time 10:00:00, normal status). The calculated power deviation is 0%, and the time deviation is 0 seconds, thus it is judged as "compliant." Stub 5's actual start time is 10:02:00, within the instruction time window of 10:02:00-10:17:00, and the actual power is 8.0kW. W, deviation 0%, judged as "compliant"; Stake 12 was delayed in issuance, the actual start time was 10:00:03.2, exceeding the instruction time window (10:00:00-10:15:00) by 3.2 seconds, within the allowable deviation, judged as "compliant"; Stake 19 failed to be issued and was not executed, with no execution record, judged as "non-execution violation"; Stake 18 has an abnormal record but emergency handling measures have been reported, and the power and time both meet the requirements, judged as "compliant (including abnormal reporting)".

[0133] The consensus mechanism employs the Practical Byzantine Fault Tolerance (PBFT) algorithm, suitable for scenarios with a small number of consortium blockchain nodes and high credibility requirements. The consensus nodes consist of 3 edge nodes, 1 power grid dispatch node, and 1 operations and maintenance node, totaling 5 nodes. After the smart contract generates the initial verification result, the consensus process is triggered. Each node independently recalculates the verification process and votes according to the "majority rule" principle. A consensus is reached if the votes are ≥4 / 5 (i.e., 4 or more nodes agree), determining the final compliance verification result. If the votes are insufficient, verification and consensus are retried, repeated a maximum of 2 times. If a consensus is still not reached, the power grid dispatch node exercises the final decision-making power.

[0134] After consensus is reached, compliance verification results are categorized into four types: compliance, compliance (including anomaly reporting), minor violation (deviation exceeds the threshold but ≤10%, no safety impact), and serious violation (deviation >10% or poses a safety risk). Different results correspond to different processing logics: compliance results are recorded directly, minor violations trigger alarms to prompt maintenance, and serious violations automatically suspend subsequent scheduling instructions for the charging pile until manual investigation and rectification. Verification results and verification basis (instruction data hash, execution record hash, and recalculation log) are written together into the blockchain distributed ledger. The ledger adopts a chain storage structure, with each block containing the hash value of the previous block, ensuring that the data is tamper-proof and traceable. It is also synchronized to all consensus nodes to achieve data consistency across the entire network.

[0135] In the example, after the smart contract verification of all instructions in time period T1 is completed, the PBFT consensus process is triggered. Five nodes independently recalculate and reach a consensus vote on the "compliance" and "compliance (including abnormal reporting)" results of 18 charging piles, including pile 1 and pile 5. A consensus is reached on the "non-execution violation" of pile 19 and the "minor power violation" (actual power deviation of 6%) of pile 20, generating the final verification result set. The results are written to the distributed ledger with a block hash value of "8b2e4f...", containing all verification results and verification basis. Each node updates the ledger synchronously, and the power grid dispatch node and operation and maintenance node can query the verification results in real time and take corresponding measures for non-compliant equipment to ensure that the entire dispatch process is compliant and controllable.

[0136] As can be seen, by acquiring power demand data of the charging pile cluster in real time and time-of-use electricity prices, regional loads, and carbon intensity signals from the power grid side through edge acquisition terminals, and using a long short-term memory network to predict the power grid supply and demand situation for a preset period in the future, a spatiotemporal coupling feature matrix of the charging pile cluster is constructed based on the power grid supply and demand situation. The spatiotemporal coupling feature matrix is ​​solved using a distributed optimization algorithm to generate a cluster collaborative scheduling strategy with the goal of smoothing power grid peak and valley loads and minimizing carbon emissions. Real-time power adjustment instructions for each charging pile are dynamically allocated according to the cluster collaborative scheduling strategy, and the compliance of instruction execution is verified based on a blockchain consensus mechanism. This enables accurate and reliable collaborative interaction between the charging pile cluster and the power grid's dynamic supply and demand, improving the economic efficiency and low-carbon operation of the power grid.

[0137] Another embodiment of the present invention provides a charging pile cluster collaborative scheduling system for grid interaction, see [link to relevant documentation]. Figure 3 The system may include:

[0138] The acquisition module 301 is used to acquire power demand data of the charging pile cluster and time-of-use electricity price, regional load and carbon intensity signals of the power grid in real time through the edge acquisition terminal, and to use a long short-term memory network to predict the power grid supply and demand situation in the future preset period.

[0139] The construction module 302 is used to construct a spatiotemporal coupling feature matrix of the charging pile cluster based on the power grid supply and demand situation. The spatiotemporal coupling feature matrix includes multi-dimensional indicators such as the power adjustment potential, user flexibility, and response priority of each charging pile.

[0140] The generation module 303 is used to solve the spatiotemporal coupling feature matrix using a distributed optimization algorithm to generate a cluster collaborative scheduling strategy with the objectives of grid peak-valley smoothing and minimizing carbon emissions.

[0141] The allocation module 304 is used to dynamically allocate real-time power adjustment instructions for each charging pile according to the cluster collaborative scheduling strategy, and verify the compliance of instruction execution based on the blockchain consensus mechanism.

[0142] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0143] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0144] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0145] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for coordinated scheduling of charging pile clusters oriented towards grid interaction, characterized in that, The method includes: The power demand data of the charging pile cluster and the time-of-use electricity price, regional load and carbon intensity signals of the power grid are obtained in real time through the edge acquisition terminal, and the power grid supply and demand situation in the future preset period is predicted by the long short-term memory network. Based on the power grid supply and demand situation, a spatiotemporal coupling feature matrix of the charging pile cluster is constructed. The spatiotemporal coupling feature matrix includes multi-dimensional indicators such as the power adjustment potential, user elasticity, and response priority of each charging pile. Specifically, the power grid supply and demand situation forecast report is analyzed to extract the predicted power surplus or shortage, electricity price, and carbon intensity for each future time period, generating a power grid situation feature vector. Based on the power grid situation feature vector, combined with the historical operating data and current charging status of each charging pile, the power boundary that can be adjusted upward or downward in each predicted time period is calculated to evaluate its dynamic power adjustment potential and generate a power adjustment potential index. According to the power adjustment potential index and the user's historical charging behavior data, the user's response sensitivity to the electricity price and power adjustment range in different time periods is analyzed, and the user elasticity is evaluated using fuzzy logic to generate a user elasticity index. Combining the power adjustment potential index, the user elasticity index, and the node voltage stability contribution of the charging pile connected to the power grid, the response priority of each charging pile in different time periods is calculated and arranged in a spatiotemporal dimension to construct a spatiotemporal coupling feature matrix. The spatiotemporal coupling feature matrix is ​​solved using a distributed optimization algorithm to generate a cluster-coordinated scheduling strategy with the objectives of grid peak-valley smoothing and minimizing carbon emissions. The real-time power adjustment instructions for each charging pile are dynamically allocated according to the cluster collaborative scheduling strategy, and the compliance of instruction execution is verified based on the blockchain consensus mechanism.

2. The method according to claim 1, characterized in that, The process involves acquiring real-time power demand data of the charging pile cluster, along with time-of-use electricity prices, regional load, and carbon intensity signals from the power grid via edge acquisition terminals, and using a long short-term memory network to predict the power grid supply and demand situation for a preset future period. This includes: Edge acquisition terminals are deployed in the charging pile cluster to collect real-time power, voltage, current, charging status and user-set charging plan data of each charging pile. At the same time, time-of-use electricity price curves, real-time regional load data and real-time carbon intensity signals are obtained through the power grid data interface to generate multi-source heterogeneous raw data streams. Time alignment and missing value imputation are performed on the multi-source heterogeneous raw data streams. Sliding window smoothing filter is used to remove abnormal fluctuations and normalize to a unified dimension to generate a standardized time series dataset. The standardized time series dataset is input into a pre-trained long short-term memory network model. The model uses an attention mechanism to enhance the focus on key time steps and extracts time series features through multi-layer LSTM units to generate a preliminary prediction sequence of power grid supply and demand in the future preset time period. Uncertainty quantification is performed on the preliminary forecast sequence, and the confidence interval of the forecast results is calculated using the Monte Carlo dropout method. Combined with the output of the historical error correction model, a power grid supply and demand situation forecast report with confidence intervals is generated.

3. The method according to claim 2, characterized in that, The process of solving the spatiotemporal coupling feature matrix using a distributed optimization algorithm to generate a cluster-coordinated scheduling strategy with the objectives of grid peak-valley smoothing and minimizing carbon emissions includes: With the objective functions of grid peak-valley smoothing and carbon emission minimization, and with the charging pile power adjustment range and user charging demand as constraints, a multi-objective optimization model is established to generate a mathematical model of the optimization problem. The distributed alternating direction multiplier method is used to decompose the mathematical model of the optimization problem, decomposing the global problem into sub-problems of each charging station, and generating a set of distributed sub-problems; Each charging pile solves its subproblem locally, exchanges intermediate results with neighboring nodes through a finite number of iterations, and gradually converges to the global optimal solution, generating a sequence of local optimal solutions; By aggregating the sequence of locally optimized solutions and performing a global consistency check, a globally optimal scheduling strategy that satisfies all constraints is obtained, and a cluster collaborative scheduling strategy is generated.

4. The method according to claim 3, characterized in that, The step of dynamically allocating real-time power adjustment instructions for each charging pile according to the cluster collaborative scheduling strategy, and verifying the compliance of instruction execution based on the blockchain consensus mechanism, includes: The cluster collaborative scheduling strategy is parsed to generate power adjustment instructions for each charging pile, including target power value, adjustment time window and execution priority, and a power adjustment instruction set is generated. The power adjustment command set is sent to each charging pile controller through the edge communication network, and the command confirmation receipt is received. The charging pile executes the power adjustment command and uploads the actual power adjustment data to the blockchain. The blockchain nodes collect the execution data and generate a power adjustment execution record. Blockchain-based smart contracts automatically verify the consistency between execution records and instructions, achieve compliance verification results through consensus mechanisms, and write the verification results into an immutable distributed ledger.

5. A charging pile cluster collaborative scheduling system for grid interaction, characterized in that, The system includes: The acquisition module is used to acquire power demand data of charging pile clusters and time-of-use electricity prices, regional load and carbon intensity signals on the grid side in real time through edge acquisition terminals, and to use long short-term memory network to predict the grid supply and demand situation in the future preset period. The module is used to construct a spatiotemporal coupling feature matrix of the charging pile cluster based on the power grid supply and demand situation. The spatiotemporal coupling feature matrix includes multi-dimensional indicators such as the power adjustment potential, user resilience, and response priority of each charging pile. Specifically, it analyzes the power grid supply and demand situation forecast report to extract the predicted power surplus or shortage, electricity price, and carbon intensity for each future time period, generating a power grid situation feature vector. Based on the power grid situation feature vector, combined with the historical operating data and current charging status of each charging pile, it calculates the power boundary that can be adjusted upward or downward in each predicted time period, evaluates its dynamic power adjustment potential, and generates a power adjustment potential index. Based on the power adjustment potential index and the user's historical charging behavior data, it analyzes the user's response sensitivity to the adjustment range of electricity price and power in different time periods, uses fuzzy logic to evaluate the user resilience, and generates a user resilience index. Combining the power adjustment potential index, the user resilience index, and the node voltage stability contribution of the charging pile to the power grid, it calculates the response priority of each charging pile in different time periods and arranges them in a spatiotemporal dimension to construct a spatiotemporal coupling feature matrix. The generation module is used to solve the spatiotemporal coupling feature matrix using a distributed optimization algorithm to generate a cluster collaborative scheduling strategy with the objectives of grid peak-valley smoothing and minimizing carbon emissions. The allocation module is used to dynamically allocate real-time power adjustment instructions for each charging pile according to the cluster collaborative scheduling strategy, and verify the compliance of instruction execution based on the blockchain consensus mechanism.

6. The system according to claim 5, characterized in that, The acquisition module is specifically used for: Edge acquisition terminals are deployed in the charging pile cluster to collect real-time power, voltage, current, charging status and user-set charging plan data of each charging pile. At the same time, time-of-use electricity price curves, real-time regional load data and real-time carbon intensity signals are obtained through the power grid data interface to generate multi-source heterogeneous raw data streams. Time alignment and missing value imputation are performed on the multi-source heterogeneous raw data streams. Sliding window smoothing filter is used to remove abnormal fluctuations and normalize to a unified dimension to generate a standardized time series dataset. The standardized time series dataset is input into a pre-trained long short-term memory network model. The model uses an attention mechanism to enhance the focus on key time steps and extracts time series features through multi-layer LSTM units to generate a preliminary prediction sequence of power grid supply and demand in the future preset time period. Uncertainty quantification is performed on the preliminary forecast sequence, and the confidence interval of the forecast results is calculated using the Monte Carlo dropout method. Combined with the output of the historical error correction model, a power grid supply and demand situation forecast report with confidence intervals is generated.

7. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-4 when it is run.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-4.

Images