Virtual power plant bidding system and method for selling electricity by aggregating distributed resources

Abstract

This invention relates to the fields of electricity market trading and new energy power generation forecasting technology. Specifically, it relates to a virtual power plant auction bidding system and method that aggregates distributed resources. The system collects the actual and predicted power of each distributed photovoltaic and wind power source at preset intervals, calculates the rolling error rate, and dynamically categorizes the data into different error intervals based on an adaptive interval boundary list. A self-learning mapping table matches a basic compensation coefficient to each error interval. A risk preference adjustment coefficient is generated by combining market volatility, grid balance demand, and credit risk. A final compensation factor is obtained through nonlinear coupling. Finally, the final compensation factor is fused with the day-ahead market original bid volume to generate a corrected bid volume, which is then pushed to the electricity trading platform. This invention achieves dynamic matching between the rolling error rate and the compensation intensity, effectively reducing the risk of forecast deviation assessment and improving the accuracy of virtual power plant auction bidding.

Description

Technical Field

[0001] This invention relates to the fields of electricity market trading and new energy power generation forecasting technology, and more specifically, to a virtual power plant sales bidding system and method that aggregates distributed resources. Background Technology

[0002] In the process of virtual power plants aggregating distributed photovoltaic and wind power to participate in the electricity market bidding, the strong uncertainty and prediction deviation of new energy output have always been the core bottleneck restricting the accuracy of application.

[0003] In recent years, with the comprehensive advancement of market-oriented reforms of new energy on-grid electricity prices, various deviation assessment and reward / penalty mechanisms have become increasingly stringent. If enterprises rely solely on a single day-ahead power forecast result for bidding, the misalignment between their output curve and contracted curve can easily trigger assessment penalties, leading to a passive situation where "the higher the bid, the greater the deviation, and the heavier the assessment."

[0004] Current mainstream methods are either limited to passively compensating for prediction biases through optimized models and energy storage configurations, or to handling uncertainties based on scenario analysis and robust optimization. However, their compensation mechanisms are mostly static or open-loop, making it difficult to dynamically and adaptively adjust to real-time fluctuations in the rolling error rate and changes in market supply and demand. Meanwhile, traditional bidding correction strategies generally treat the time dimension and physical nodes as homogeneous objects, ignoring the fundamental differences in output stability across different time periods and the essential differences in the adjustment capabilities of different resource nodes. Furthermore, they lack proactive response capabilities to emergency operational risks such as grid frequency fluctuations exceeding limits. To address this issue, we provide a virtual power plant auction bidding system and method that aggregates distributed resources. Summary of the Invention

[0005] The purpose of this invention is to provide a virtual power plant auction bidding system and method for aggregating distributed resources, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, a virtual power plant auction bidding system that aggregates distributed resources is provided, including a real-time error monitoring unit, a dynamic compensation factor generation unit, and a bidding quantity correction unit.

[0007] The real-time error monitoring unit collects actual output data and corresponding predicted output data of distributed photovoltaic and wind power according to a preset cycle. It generates a rolling error rate by calculating the percentage of the absolute difference between the actual power and the predicted power to the installed capacity, and classifies the rolling error rate into different error ranges according to the preset error rate classification rules.

[0008] The dynamic compensation factor generation unit establishes a mapping relationship between error rate and compensation amount, where each error interval maps to a unique basic compensation coefficient, and generates the final compensation factor based on the real-time risk strategy of the virtual power plant and the risk preference adjustment coefficient.

[0009] The bidding volume correction unit integrates the final compensation factor with the original bid volume in the day-ahead market to generate a corrected bid volume, and then pushes the corrected bid volume to the power trading platform.

[0010] The second objective of this invention is to provide a method for implementing a virtual power plant electricity pricing system that aggregates distributed resources including any of the above-mentioned features, comprising the following steps:

[0011] S1. Collect the actual and predicted power of each distributed photovoltaic and wind power according to the preset cycle. The predicted power is generated in a rolling manner by a deep learning model based on cloud numerical weather forecast and historical big data. The actual and predicted power are cleaned and aligned. The percentage of the absolute difference between the actual and predicted power to the installed capacity is calculated. Combined with the dynamic weight determined by the prediction volatility, a rolling error rate is generated. According to the adaptive interval boundary list based on historical distribution statistics, the rolling error rate is dynamically classified into the preset error interval. When there is a risk of grid frequency exceeding the limit, the tightening mechanism of the adaptive interval boundary list is automatically triggered.

[0012] S2. Based on the error interval query, the self-learning mapping table is used. The self-learning mapping table presets the benchmark center point and sub-interval offset logic for each error interval to determine the basic compensation coefficient. At the same time, based on the real-time market volatility, power grid regulation demand and its own credit status, the comprehensive risk value is calculated through a multi-dimensional risk quantification model. Combined with the weight strategy dynamically adjusted by historical performance feedback, the risk preference adjustment coefficient is output through the risk preference curve function set by the operator. Finally, the risk preference adjustment coefficient is non-linearly coupled with the basic compensation coefficient to generate the final compensation factor.

[0013] S3. Analyze the original market bid volume of the day-ahead, decompose the original market bid volume of the day-ahead into planned output components with different time granularities and physical nodes. At the same time, expand the final compensation factor into a compensation factor matrix along the same dimension. Perform differentiated fusion operation on the compensation factor matrix according to the time distance and resource adjustment characteristics of each planned output component. Re-aggregate all planned output components after differentiated operation to generate the overall corrected bid volume.

[0014] S4. Push the revised bid amount to the power trading platform to complete the bidding submission.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] This invention addresses the problems of traditional static compensation, such as its inability to track rolling changes in output deviation and its difficulty in adapting to different error levels, by real-time monitoring of the actual and predicted output of distributed photovoltaic and wind power, calculating the rolling error rate, and dynamically classifying error intervals based on an adaptive interval boundary list. Simultaneously, it utilizes a built-in self-learning mapping table to match a basic compensation coefficient for each error interval, and combines real-time price fluctuations, grid balance requirements, and credit risk to generate a risk preference adjustment coefficient. Through nonlinear coupling, a final compensation factor is obtained, overcoming the technical shortcomings of the bidding correction strategy being disconnected from market risk and grid emergency states. Furthermore, the final compensation factor is differentially integrated with the day-ahead original bid volume along the time axis and resource nodes to generate a corrected bid volume, effectively avoiding correction distortion caused by treating different time periods and nodes as homogeneous objects. Attached Figure Description

[0017] Figure 1 This is an overall block diagram of the present invention;

[0018] Figure 2 This is the overall flowchart of the present invention.

[0019] The meanings of the labels in the diagram are as follows:

[0020] 1. Real-time error monitoring unit; 2. Dynamic compensation factor generation unit; 3. Bidding volume correction unit. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a virtual power plant electricity pricing system that aggregates distributed resources. Please refer to [link / reference]. Figure 1 As shown, it includes a real-time error monitoring unit 1, a dynamic compensation factor generation unit 2, and a bidding volume correction unit 3;

[0023] Real-time error monitoring unit 1 collects actual output data and corresponding predicted output data of distributed photovoltaic and wind power according to a preset cycle. It generates a rolling error rate by calculating the percentage of the absolute difference between actual and predicted power relative to the installed capacity. The acquisition of actual and predicted power specifically includes:

[0024] The actual power output data is collected in real time by smart meters and synchronous vector measurement devices deployed at various distributed power grid connection points, and the time stamp is marked.

[0025] Simultaneously, it connects to a cloud platform containing numerical weather forecast data and historical power output big data. Using a trained deep learning prediction model, starting from the end of the previous preset cycle, it continuously executes power output predictions for the next preset cycle, generating predicted power output data that is aligned with the actual power output data in terms of time scale. Based on this, the actual power output data and predicted power output data are cleaned and aligned, abnormal jump points are removed, and data with missing time scales is supplemented by interpolation to form actual power and predicted power.

[0026] Preferably, after the real-time error monitoring unit 1 starts the power output data acquisition, it enters the process of acquiring actual power output data of distributed power source, generating predicted power output data, cleaning data and aligning time scale. Relying on the acquisition of data by on-site metering equipment, the support of cloud big data, the prediction of deep learning model and the preprocessing of time series data, it obtains actual power and predicted power data with accurate time series matching and numerical compliance, providing a data source for subsequent rolling error rate calculation.

[0027] A distributed power grid connection point is the physical node connecting a distributed photovoltaic or wind power source to the power grid. The actual output data is the real-time active power value output by the distributed power source. The timestamp is a timestamp accurate to the second, in the format of year-month-day-hour-minute-second, used to uniquely mark the data acquisition time. During acquisition, smart meters collect the active power value of the grid connection point in real time at a preset acquisition frequency of 1 second / time. The synchronous vector measurement device synchronously collects the active power time series data of the grid connection point. The data collected by the two types of devices are transmitted to the edge computing node through the field communication module. Each data is bound to a unique timestamp to ensure that the data acquisition time is traceable. After the acquisition is completed, the edge computing node initially verifies the data integrity, removes null data generated when the device is offline, and generates the original actual output time series dataset. This dataset contains timestamps and corresponding active power values, providing raw data support for subsequent data preprocessing.

[0028] Historical power output big data comprises all historical actual power output time-series data since the deployment of the distributed power source. The deep learning prediction model is a neural network model trained on historical power output big data and meteorological data for power output time-series prediction. The preset period includes a historical backtracking period and a prediction period, with a default historical backtracking period of 24 hours and a prediction period of 15 minutes. Rolling prediction is the process of repeatedly executing predictions at fixed intervals and continuously updating the prediction results. After connecting to the cloud platform, the latest numerical weather forecast data and the historical power output big data of the target distributed power source are obtained in real time and backtracked to the end of the previous preset period at the current moment. The end of the cycle is taken as the starting point for prediction. Historical power output time series data and meteorological time series data are input into the trained deep learning prediction model. The model extracts time series features and meteorological features through multiple network layers, and performs power output prediction for the next preset cycle in a rolling manner, outputting predicted power output time series data in kilowatts. At the same time, each predicted power output data is matched with a time stamp that is consistent with the actual power output data to ensure that the predicted power output data and the actual power output data correspond one-to-one in the time dimension, generating the original predicted power output time series dataset. This dataset contains time stamps and corresponding predicted active power values, providing prediction data support for subsequent data alignment.

[0029] Abnormal jump points are invalid data points in time-series data that exhibit sudden numerical changes or deviate from the normal output fluctuation range within a short period. They are used to remove abnormal values ​​caused by equipment failure, communication interference, and transient disturbances. During the cleaning process, the original actual output time-series dataset and the original predicted output time-series dataset are traversed, and the 3σ criterion is used to identify abnormal jump points. The anomaly determination expression is as follows: In the formula The value of the current data point to be verified. The mean of the data series. To determine the standard deviation of the data sequence, the mean and standard deviation are calculated point by point during the traversal process. Data points that meet the anomaly judgment criteria are marked. After marking, all data records marked as abnormal jump points are removed to avoid abnormal data interfering with the accuracy of subsequent error rate calculations.

[0030] Missing time stamps are gaps in time series data caused by equipment offline or communication interruptions. Interpolation is a numerical estimation method based on the numerical patterns of adjacent valid data points to fill in missing data. Alignment is performed using the complete time stamp sequence of actual output data as a benchmark, matching the cleaned predicted output data, and selecting the common time stamp set of actual and predicted output data. For time stamp gaps where only a single type of data exists, linear interpolation is used to fill in the missing data. The linear interpolation expression is: In the formula Missing time stamp , For the adjacent valid time stamps before and after the missing time stamp, , The power values ​​are the values ​​corresponding to adjacent valid time points. The interpolated power values ​​are then used to generate actual power time-series datasets and predicted power time-series datasets. Actual power refers to the real-time active power value of the distributed power source after cleaning and alignment, while predicted power refers to the predicted active power value of the distributed power source after cleaning and alignment. The two types of data correspond one-to-one in time series and are compliant in value. They are directly input into the subsequent rolling error rate calculation stage, providing time-series matching and numerically reliable basic data for the accurate calculation of the error rate.

[0031] The generation of the rolling error rate specifically includes:

[0032] Take the actual power and predicted power at each collection point and several consecutive time points before it, calculate the absolute difference at each time point and divide it by the installed capacity of the corresponding distributed power source to obtain the instantaneous percentage error at each time point;

[0033] Different dynamic weights are assigned to the instantaneous percentage error at each time point. The dynamic weights are determined by the output fluctuation estimate given by the prediction model at each time point. The weighted instantaneous percentage errors at all time points are normalized and summed up, and the result is the rolling error rate at that time point.

[0034] Preferably, after completing the time-series cleaning and alignment of the actual power and the predicted power, the real-time error monitoring unit 1 performs continuous time-point data selection, instantaneous percentage error calculation, dynamic weight allocation, and rolling error rate generation. Relying on time-series window truncation, error quantification, volatility weighting, and normalized accumulation, the discrete time-point errors are integrated into a rolling error rate that can reflect the recent output deviation trend, providing a core quantitative indicator for subsequent dynamic classification of error intervals.

[0035] The acquisition time point refers to the reference time for data acquisition, corresponding to a single time scale in the cleaned and aligned actual power and predicted power time series data. Continuous time points refer to a fixed number of adjacent time series points traced back from the current acquisition time point. The fixed number defaults to 5, with a range of 3 to 7, and is used to construct the time series analysis window. Actual power is the real-time active power value of the distributed power source after cleaning and alignment, in kilowatts, denoted as . , This is the index of a time point within the time series window, with values ​​ranging from 1 to... , The total number of consecutive time points is represented by the predicted power, which is the predicted active power value of the distributed generation after cleaning and alignment, in kilowatts, denoted as . The installed capacity is the rated maximum output power of the distributed power source, in kilowatts, denoted as C, which is the base denominator parameter for error calculation.

[0036] First, using the current data collection time as the endpoint, extract the data including the current time and the data before that time. A time-series data window containing consecutive time points is used to extract the actual power corresponding to each time point within the window. With predicted power A window time-series power dataset is formed. The absolute difference at each time point is calculated and divided by the corresponding installed capacity of the distributed power source to obtain the instantaneous percentage error at each time point. The instantaneous percentage error is the proportion of the deviation between the actual output and the predicted output at a single time point to the installed capacity, used to quantify the degree of prediction deviation at a single point. The absolute difference is the absolute value of the difference between the actual power and the predicted power, avoiding the cancellation of positive and negative deviations. The expression for calculating the instantaneous percentage error is as follows: In the formula For the first The instantaneous percentage error at each point in time, expressed as a percentage. For the first The absolute difference between actual power and predicted power at each point in time. For the installed capacity of distributed power sources, iterate through all n time points within the time series window, substitute them one by one into the formula to calculate, and generate the instantaneous percentage error sequence within the window. This sequence is used for subsequent dynamic weight allocation and weighted calculation, assigning different dynamic weights to the instantaneous percentage error at each time point. The dynamic weights are weighting coefficients for the errors at each time point within the time window, ranging from 0 to 1, with the sum of all weights being 1. This reflects the degree of influence of power output fluctuations at different time points on the overall error. The predicted power output volatility at each time point is a quantitative value output by the deep learning prediction model, representing the severity of power output fluctuations at the corresponding time point, denoted as... The value is greater than 0, and the more volatile the fluctuation, the larger the estimated value. The dynamic weight is determined by this estimated value; higher weights are assigned during periods of high volatility, and lower weights are assigned during periods of stable volatility. The dynamic weight calculation expression is as follows: In the formula For the first Dynamic weights at each point in time. For the first Estimated output volatility at a given point in time The sum of the volatility estimates for all points within the window is calculated by iterating through all points within the time series window. At each time point, the formula is substituted one by one to calculate, generating a dynamic weight sequence within the window. This sequence is used for subsequent weighted error calculation.

[0037] The weighted instantaneous percentage errors at all time points are normalized and summed to obtain the rolling error rate for that data acquisition point. The weighted instantaneous percentage error is the product of the instantaneous percentage error at a single time point and its corresponding dynamic weight, used to integrate the effects of single-point deviations and fluctuations. The normalized summation is a summation of the weighted errors, eliminating the numerical scaling effect caused by weight normalization. The rolling error rate is a comprehensive quantitative value of the weighted deviations within the time series window, used to reflect the overall level of recent output prediction deviations. The expression for calculating the rolling error rate is as follows: In the formula The rolling error rate at the current data collection point, expressed as a percentage. The dynamic weight at time point i. The instantaneous percentage error at the i-th time point is substituted into the instantaneous percentage error sequence and the dynamic weight sequence to complete the calculation, obtaining the rolling error rate corresponding to the current collection time point. This rolling error rate is directly input into the subsequent error interval dynamic classification stage to match the adaptive interval boundary list to determine the error interval to which it belongs, providing an error level basis for the generation of dynamic compensation factors, ensuring that the compensation strategy fits the recent output deviation characteristics, and improving the accuracy of bidding volume correction.

[0038] The rolling error rate is then categorized into different error ranges according to a preset error rate grading rule. This error rate grading rule is a dynamic boundary rule, which includes:

[0039] An adaptive interval boundary list based on the historical rolling error rate distribution is introduced. The initial boundary of the adaptive interval boundary list is set according to the historical statistical quantile. During classification, the current rolling error rate is compared with the latest adaptive interval boundary list to determine the error interval to which it belongs. At the same time, the adaptive interval boundary list is dynamically adjusted according to the real-time operating status of the power grid. When the risk of the power grid frequency fluctuation exceeding the limit is detected, the real-time error monitoring unit 1 automatically triggers the boundary tightening mechanism, shrinking the boundary values ​​of each error interval towards zero in the next preset cycle.

[0040] Preferably, after the rolling error rate is generated, the real-time error monitoring unit 1 constructs an adaptive interval boundary list, classifies error intervals, dynamically adjusts boundaries, and tightens them due to risk triggers. Relying on historical error distribution statistics, grid status linkage adjustment, and risk emergency tightening mechanisms, it achieves dynamic adaptation of error interval boundaries, providing a hierarchical benchmark for accurate matching of compensation coefficients.

[0041] The adaptive interval boundary list is an ordered set of numerical boundaries generated based on the historical rolling error rate distribution. It is dynamically updated according to the grid status and used to classify rolling error rate levels. The historical rolling error rate distribution is the numerical distribution set of all effective rolling error rates since the target distributed power source began self-monitoring. The statistical quantile is the critical value corresponding to a preset percentage position after sorting the historical rolling error rate dataset in ascending order. It is used to classify different error levels. When setting the initial boundaries, all rolling error rate data within a preset historical period (default 30 days) are first collected to form the historical error dataset. In the formula For the k-th historical rolling error rate, Given the total number of historical data points, the dataset is sorted in ascending order. Three key quantiles are selected as initial boundaries, with the default values ​​being 0.3, 0.6, and 0.9 quantiles. The quantile calculation expression is as follows: In the formula For p quantiles, The quantile percentages (values ​​0.3, 0.6, and 0.9). This is the floor function, which generates the initial adaptive interval boundary list. The rolling error rate is divided into four levels: low error range, medium error range, high error range, and extremely high error range. The initial boundaries are set based on historical error distribution patterns to ensure that the initial classification closely matches the characteristics of normal output deviation. During classification, the current rolling error rate is compared with the latest adaptive range boundary list. The current rolling error rate is denoted as... The latest adaptive interval boundary list is denoted as In the formula , , , For each level of boundary value, The comparison logic is to judge sequentially. The interval to which it belongs, the determination expression is:

[0042]

[0043] In the formula The error intervals are numbered from 1 to 4, corresponding to the lowest to highest error intervals. After determining the interval to which the current rolling error rate belongs, this interval number is directly used as the index for querying the compensation factor, providing a hierarchical basis for the dynamic compensation factor generation unit 2. The adaptive interval boundary list is dynamically adjusted according to the real-time operating status of the power grid. The real-time operating status of the power grid includes the power grid frequency, bus voltage, and regional load rate. During adjustment, the status data issued by the power grid dispatch system is accessed in real time. A state stability threshold is set. When the power grid frequency is stable within ±0.1Hz of the rated value, the voltage is stable within ±5% of the rated value, and the load rate is in the range of 40% to 70%, it is determined to be in a stable state, and the boundaries remain unchanged. When the power grid deviates from the stable state, the state deviation degree is calculated. The greater the deviation, the higher the risk of power grid operation. The boundary adjustment expression is: In the formula To adjust the boundary, To adjust the front boundary, This is an adjustment factor (default 0.2). The deviation is the state deviation. Under stable conditions, the deviation is 0 and the boundary remains unchanged. When the risk increases, the deviation increases and the boundary contracts synchronously, so as to realize the linkage and adaptation between the boundary and the normal risk of the power grid.

[0044] When a risk of grid frequency fluctuation exceeding the limit is detected, the real-time error monitoring unit 1 automatically triggers the boundary tightening mechanism. This risk refers to an abnormal fluctuation in grid frequency exceeding the rated 50Hz by more than ±0.2Hz. During monitoring, real-time grid frequency time-series data is collected. A risk trigger is determined when the frequency exceeds the threshold for three consecutive sampling points. The boundary tightening mechanism is an emergency adjustment strategy that shrinks the boundary values ​​of each error interval towards zero within the next preset period (default 15 minutes). The shrinkage amplitude is preset to a fixed percentage (default 30%). The expression for calculating the tightened boundary is as follows: In the formula To tighten the rear boundary, To adjust the boundary, The tightening ratio (default 0.3) results in smaller boundary values ​​and narrower interval ranges for each interval after contraction. The same rolling error rate will be classified into a higher error interval. The tightened boundary list remains effective for a preset period, after which it reverts to the normal adjustment boundary. Tightening the boundaries increases the sensitivity of error grading; even minor output deviations trigger higher-level compensation, making virtual power plant output declarations more conservative and proactively reducing the impact of output fluctuations on grid frequency. This ensures operational stability under abnormal grid conditions. The error interval numbers obtained from the classification, along with the dynamically adjusted or tightened boundary list, are synchronously transmitted to the dynamic compensation factor generation unit 2 for querying the self-learning mapping table to match the corresponding basic compensation coefficient. The dynamic adjustment and tightening mechanism ensures that the error grading always aligns with the real-time risk status of the grid, allowing the compensation coefficient to adapt to the current grid regulation needs, ultimately improving the safety and adaptability of virtual power plant bidding quantity correction.

[0045] The dynamic compensation factor generation unit 2 establishes a mapping relationship between error rate and compensation amount. The dynamic compensation factor generation unit 2 has a built-in self-learning mapping table. The self-learning mapping table uses the error interval as the query key. Whenever the error interval is triggered, it records the application effect of the compensation amount generated based on the basic compensation coefficient of the error interval, that is, the degree of matching between the subsequent actual output and the corrected declared amount. The application effect is fed back to the learning parameters of the self-learning mapping table, driving the adjustment of the basic compensation amount mapped to the error interval in the next cycle. If the matching degree improves after application, the compensation amount is increased in the current direction, and vice versa.

[0046] Preferably, after the error interval is classified, the dynamic compensation factor generation unit 2 performs self-learning mapping table construction, compensation amount generation, effect recording, parameter feedback and dynamic adjustment of compensation amount. Relying on key value mapping, effect quantification and parameter iterative optimization, it realizes the adaptive correction of the basic compensation amount, so that the compensation strategy is dynamically optimized according to the actual output matching situation.

[0047] The dynamic compensation factor generation unit 2 has a built-in self-learning mapping table. This self-learning mapping table is a structured data table that stores the mapping relationship between error intervals and compensation parameters and supports dynamic updates. The query key is an index field used to uniquely identify the mapping relationship and quickly retrieve the corresponding parameter. Here, the error interval is used as the query key. The error interval is a rolling error rate level identifier obtained by the real-time error monitoring unit 1, denoted as C, with values ​​of 1, 2, 3, and 4, corresponding to low, medium, high, and extremely high error intervals, respectively. Each record in the mapping table contains the query key (error interval C) and the basic compensation coefficient. Learning parameters Historical matching degree sequence The four core fields are: the basic compensation coefficient, which is the initial compensation benchmark value corresponding to the error interval, ranging from 0.05 to 0.3, with default values ​​of 0.05 for low error interval, 0.1 for medium error interval, 0.2 for high error interval, and 0.3 for extremely high error interval; the learning parameter, which is a coefficient controlling the adjustment range of the compensation amount, ranging from 0.01 to 0.1, with a default value of 0.05; and the historical matching degree sequence, which records the effect data after each compensation application for the error interval. During mapping table initialization, corresponding records are generated according to the preset number of error intervals, and the initial basic compensation coefficient and initial learning parameter are written to establish the initial mapping relationship. Whenever an error interval is triggered, i.e., after the real-time error monitoring unit 1 determines the current rolling error rate to belong to error interval C, the dynamic compensation factor generation unit 2 retrieves the corresponding basic compensation coefficient from the self-learning mapping table using C as the query key. The compensation amount is generated, which is the product of the base compensation coefficient and the original market order volume from the previous day. It is used to adjust the original order volume. The expression for calculating the compensation amount is as follows: In the formula For compensation amount, The basic compensation coefficients corresponding to the error interval C are: The original market declaration volume is used as the basis for generating the compensation volume. After that, the compensation volume is applied to the original declaration volume to obtain the revised declaration volume. Complete the current period's compensation application, record the application effect of the compensation amount. The application effect refers to the degree of matching between the actual output after the compensation amount application and the revised declared amount. The degree of matching is an indicator that quantifies the degree of fit between the actual output and the revised declared amount, denoted as m, with a value range of 0 to 1. The closer the value is to 1, the higher the degree of matching. The degree of matching needs to be calculated in the next data period. Extract the revised declared amount. Corresponding actual output The expression for calculating the degree of matching is: In the formula The absolute difference between the actual output and the revised reported output is used. After calculation, the current matching degree m is stored in the historical matching degree sequence corresponding to the error interval C. Complete the application effect record.

[0048] The application results are fed back to the learning parameters of the self-learning mapping table. The adjustment step size for controlling the basic compensation amount is based on the current matching degree m and the historical average matching degree during feedback. The difference (where n is the number of historical matches) is used to update the learning parameters. The expression for updating the learning parameters is: In the formula To update the learning parameters, The learning rate is set to 0.02 by default. After an update, the learning parameters are synchronously written to the corresponding record in the self-learning mapping table to complete the parameter feedback.

[0049] The next cycle will adjust the base compensation amount mapped to the error range. The base compensation amount is the compensation benchmark corresponding to the base compensation coefficient. The adjustment logic is divided into two cases: if the matching degree improves after application, i.e., the current matching degree... This indicates that the current compensation direction is valid. The compensation amount is increased along the current direction, and the adjustment expression is: In the formula This is the adjusted basic compensation coefficient. Adjust the step size based on the baseline; the default value is 0.01. If the matching accuracy does not improve after application, i.e. This indicates that the current compensation direction is invalid, so the compensation amount is reduced, and the expression is adjusted as follows: The adjusted basic compensation coefficient must be limited to a preset range (0.05 to 0.3). If it exceeds the upper limit, the upper limit value will be used; if it is below the lower limit, the lower limit value will be used. After the adjustment is completed, the new basic compensation coefficient will be applied. Write the data into the self-learning mapping table, overwrite the original coefficients, and complete the iterative optimization of the basic compensation amount. The adjusted basic compensation coefficient will serve as the compensation benchmark for the error range in the next cycle. It will be used to generate the final compensation factor in combination with the risk preference adjustment coefficient, ensuring that the compensation strategy continues to optimize the matching effect with the actual output, improving the accuracy of subsequent bidding volume correction, and reducing the assessment risk caused by output deviation.

[0050] Each error interval is mapped to a unique basic compensation coefficient. The self-learning mapping table presets a reference center point for each error interval and sets sub-interval offset logic for different rolling error rate levels within the error interval.

[0051] When the rolling error rate falls into a certain error range, its specific position relative to the reference center point of that range is first determined. If it belongs to the high end of the range, the positive offset coefficient corresponding to the error range is triggered, and a positive adjustment is added to the original basic compensation coefficient. If it belongs to the low end of the range, the negative offset coefficient is triggered for negative adjustment.

[0052] Preferably, after the self-learning mapping table basic compensation coefficient is dynamically iteratively optimized, the dynamic compensation factor generation unit 2 enters the error interval benchmark center point preset, sub-interval offset logic configuration, rolling error rate position determination and basic compensation coefficient fine offset adjustment. Relying on interval benchmark anchoring, sub-interval layer division, precise position matching and coefficient fine adjustment, the basic compensation coefficient and the error rate level in the interval are accurately matched, so that the compensation amount fits the specific deviation degree in the error interval.

[0053] The self-learning mapping table is a structured data table built into the dynamic compensation factor generation unit 2, storing the mapping relationship between error intervals and compensation parameters and supporting dynamic updates. The error interval is a deviation level interval obtained by dividing the rolling error rate using an adaptive interval boundary list, denoted as C, with values ​​of 1, 2, 3, and 4, corresponding to low, medium, high, and extremely high error intervals, respectively. The benchmark center point is a preset reference value in the self-learning mapping table for each error interval, serving as the benchmark for dividing the rolling error rate into high and low levels within the interval. The rolling error rate is the comprehensive deviation value obtained by weighted normalization and accumulation at the current acquisition time point, denoted as... The sub-interval offset logic is a rule that divides different rolling error rate levels within the error interval into high and low ranges, triggers corresponding offset coefficients to fine-tune the base compensation coefficient, and the base compensation coefficient is the core compensation benchmark value corresponding to the error interval, denoted as... The positive offset coefficient is the positive adjustment ratio triggered at the high end of the interval, and the negative offset coefficient is the negative adjustment ratio triggered at the low end of the interval. The self-learning mapping table presets a reference center point for each error interval. The reference center point is taken as the arithmetic mean of the upper and lower boundaries of the corresponding error interval. The calculation expression is: In the formula The reference center point for the Cth error interval is... Let C be the lower boundary value of the error interval. This is the upper boundary value of the C-th error interval. For example, the boundary of the lower error interval C=1 is... Its reference center point The boundary of the mean error interval C=2 is Its reference center point The high error range and the extremely high error range are calculated in the same way. After the preset is completed, the reference center point of each error range is stored in the corresponding record of the self-learning mapping table as the reference for determining the rolling error rate position within the range.

[0054] The self-learning mapping table sets sub-interval offset logic for different rolling error rate levels within the error interval. First, each error interval is divided into three sub-intervals: a high-end range, a baseline range, and a low-end range. The high-end range refers to the area where the rolling error rate is greater than the midpoint between the baseline center point and the upper boundary of the interval. The low-end range refers to the area where the rolling error rate is less than the midpoint between the baseline center point and the lower boundary of the interval. The baseline range refers to the area between the high-end and low-end ranges. Then, a positive offset coefficient is preset for each error interval. With negative offset coefficient The values ​​range from 0.02 to 0.08. The default positive offset coefficient is 0.05 and the negative offset coefficient is -0.05. After the preset is completed, the two types of offset coefficients are stored in the corresponding error interval records of the self-learning mapping table as the proportional benchmark for fine-tuning the basic compensation coefficient.

[0055] When the rolling error rate falls into a certain error range, first determine its specific position relative to the reference center point of that range, and then determine the current rolling error rate. After classifying and determining the corresponding error interval C, the reference center point of that interval is extracted from the self-learning mapping table. Lower boundary of the interval Upper boundary of the interval Calculate the high-end threshold With low threshold The position determination expression is

[0056]

[0057] The rolling error rate is determined within the corresponding error interval. If it falls within the high end of the interval, a positive offset coefficient corresponding to that error interval is triggered. A positive adjustment is added to the original base compensation coefficient. The positive adjustment is the product of the original base compensation coefficient and the positive offset coefficient. The expression for calculating the positive adjustment is as follows: The adjusted basic compensation coefficient expression is as follows: If the range falls within the lower end of the interval, a negative offset coefficient is triggered for adjustment. The negative adjustment amount is the product of the original base compensation coefficient and the negative offset coefficient. The expression for calculating the negative adjustment amount is as follows: The adjusted basic compensation coefficient expression is as follows: If the value falls within the baseline range, no offset adjustment will be triggered, and the base compensation coefficient will remain unchanged after adjustment. After the adjustment is completed, the basic compensation coefficient needs to be subject to boundary restrictions. The preset value range is 0.05 to 0.3. If it exceeds the upper limit, it will be 0.3, and if it is below the lower limit, it will be 0.05 to ensure that the coefficient is within the effective range. The refined basic compensation coefficient after sub-range offset adjustment will be used as the benchmark parameter for subsequent nonlinear coupling with the risk preference adjustment coefficient to generate the final compensation factor. This will ensure that the compensation amount not only matches the level division of the error range, but also adapts to the high and low levels of the rolling error rate within the range, further improving the accuracy of the bidding volume correction and reducing the market assessment risk caused by the output deviation of distributed power sources.

[0058] Based on the real-time risk strategy of the virtual power plant and the risk preference adjustment coefficient, the final compensation factor is generated. The dynamic compensation factor generation unit 2 continuously accesses the real-time price volatility of the electricity market, the system balance demand, and the utilization rate of the virtual power plant's own credit limit. It calculates the current comprehensive risk value through a multi-dimensional risk quantification model and inputs the comprehensive risk value into a preset risk preference curve function. The risk preference curve function is set by the virtual power plant operator and outputs a risk preference adjustment coefficient in the range of zero to positive numbers. The risk preference adjustment coefficient is nonlinearly coupled with the basic compensation coefficient obtained by mapping the error interval to generate the final compensation factor.

[0059] After the basic compensation coefficient is finely adjusted by sub-interval offset, the dynamic compensation factor generation unit 2 enters the standardized implementation process of market and power grid data access, multi-dimensional risk quantification, preference coefficient generation and nonlinear coupling. This process relies on multi-source risk data fusion, quantitative modeling, preference mapping and nonlinear calculation to deeply bind the basic compensation coefficient with the real-time risk status, generate the final compensation factor that adapts to the current market and power grid risks, and provide a risk adaptive compensation benchmark for bidding volume correction.

[0060] The dynamic compensation factor generation unit 2 continuously inputs three core risk data categories: real-time price volatility in the electricity market, system balancing demand, and the utilization rate of virtual power plant credit lines. Real-time price volatility in the electricity market is a quantified value of the volatility of the real-time electricity price sequence on the electricity trading platform, denoted as... The value ranges from 0 to 1, with larger values ​​indicating more volatile market prices. Volatility is calculated using rolling standard deviation by collecting day-ahead and real-time market electricity price time-series data. The expression is: In the formula The rolling standard deviation of the electricity price series. The rolling average of the electricity price series; the system balance demand is the quantified value of regional power regulation demand released by the power grid dispatch center, denoted as... The value ranges from 0 to 1. A larger value indicates a larger peak-shaving gap in the power grid and a more urgent need for balancing. It is directly mapped from the power grid dispatch signal. The virtual power plant's own credit limit utilization rate is the ratio of the current application volume to the available credit limit, denoted as... The value ranges from 0 to 1. A higher value indicates higher credit usage and greater default risk. The calculation expression is: In the formula This represents the current number of applications. For the available credit limit of the virtual power plant, all three types of data are updated in real time at a preset period of 1 minute by default to ensure that the risk assessment is in line with the latest situation. The current comprehensive risk value is calculated through a multi-dimensional risk quantification model. The multi-dimensional risk quantification model is a mathematical model that assesses and weights market risk, technology risk, and credit risk in parallel. The model has built-in initial weights. , , These correspond to market risk, technological risk, and credit risk, respectively, with initial weights of 0.4, 0.3, and 0.3 by default, satisfying the requirements. Market risk is caused by price volatility. Direct representation Technical risks are mitigated by the system's balancing requirements. Direct representation Credit risk is determined by the credit limit utilization rate. Direct representation The overall risk value is the weighted sum of the three risk values, and the calculation expression is: In the formula The overall risk value ranges from 0 to 1. The higher the value, the higher the overall risk level. The weights can be dynamically adjusted based on the historical strategy performance. The weight allocation is optimized by regularly analyzing returns and assessment results to ensure that the risk weights are consistent with the actual impact of the risk.

[0061] The overall risk value is input into a preset risk preference curve function. This function is a monotonically increasing function defined by the virtual power plant operator to map the overall risk value to a risk preference adjustment coefficient. The function expression is as follows: In the formula This is a risk preference adjustment coefficient, with an output range from 0 to positive numbers. The function can be either a power function or an exponential function; the default is a power function. In the formula For proportionality coefficient, These are power coefficients, all set by the operator, with values ​​greater than 0. The function's characteristic is that the higher the overall risk value, the larger the output risk preference adjustment coefficient, reflecting a strategy tendency of stronger compensation adjustment for higher risks. Once the function parameters are fixed, they can be iteratively optimized based on operational performance to adapt to the operator's risk preferences. The risk preference adjustment coefficient is nonlinearly coupled with the basic compensation coefficient obtained by mapping the error interval to generate the final compensation factor. Nonlinear coupling is the process of performing nonlinear operations on the basic compensation coefficient and the risk preference adjustment coefficient, fusing deviation compensation and risk compensation. The basic compensation coefficient obtained by mapping the error interval is denoted as... The nonlinear coupling adopts a multiplicative power coupling method, and the calculation expression is as follows: In the formula As the final compensation factor, The nonlinear coupling coefficient, with a default value of 0.5, is used to control the strength of the nonlinear correlation. The final compensation factor value is greater than or equal to the basic compensation coefficient, and its value increases with the risk preference adjustment coefficient. After coupling, the final compensation factor is subject to boundary constraints, with a preset maximum value of 0.5. If the value exceeds the upper limit, it is set to 0.5 to ensure that the compensation factor is within a reasonable range. The generated final compensation factor is directly transmitted to the bidding volume correction unit, where it is used to merge with the original market bid volume to generate the corrected bid volume. The final compensation factor integrates both output deviation compensation and real-time risk compensation, making the corrected bid volume both consistent with the output deviation characteristics of distributed power sources and adapted to current market price fluctuations, grid balance needs, and its own credit risk status. This effectively reduces the risk of output deviation assessment and market transaction, and improves the security and economy of virtual power plant bidding.

[0062] The bidding volume correction unit integrates the final compensation factor with the original bid volume in the day-ahead market to generate a corrected bid volume, and then pushes the corrected bid volume to the power trading platform.

[0063] Furthermore, the multi-dimensional risk quantification model assesses market risk, technological risk, and credit risk in parallel. Market risk is represented by the product of price volatility and prediction deviation rate, technological risk is quantified by the regional regulation urgency signal issued by the power grid, and credit risk is represented by the ratio of current application volume to available credit limit.

[0064] Initial weights are assigned to these three types of risks, and the weights are dynamically adjusted based on the historical performance of the strategy. This includes periodically analyzing the market returns and deviations of the strategy under different weight combinations.

[0065] Preferably, after the dynamic compensation factor generation unit 2 completes the access of multi-source risk data, the multi-dimensional risk quantification model enters the standardized implementation process of parallel assessment of market risk, technology risk, and credit risk, initial weight configuration, and dynamic weight adjustment based on historical performance. This process relies on the independent quantification of the three types of risks, initial weight allocation, historical effect review, and weight iterative optimization to achieve dynamic adaptation between risk weights and actual operational performance, ensuring that the calculation of comprehensive risk value is consistent with the weights of real risk impact.

[0066] The multi-dimensional risk quantification model is a mathematical model that processes three independent risk indicators—market risk, technological risk, and credit risk—in parallel and outputs a weighted composite risk value. Parallel assessment means that the calculations of the three types of risks are independent, executed synchronously, and do not interfere with each other. Market risk is the trading risk arising from both electricity price fluctuations and output forecast deviations. Price volatility is the quantified value of the volatility of the real-time electricity price series in the electricity market, denoted as... The value ranges from 0 to 1, and is calculated by dividing the rolling standard deviation of real-time electricity price time-series data by the mean. The prediction deviation rate is the average degree of deviation between the predicted and actual output values, denoted as . The value ranges from 0 to 1, and is obtained by normalizing the rolling error rate. The expression for calculating market risk is: In the formula The market risk value ranges from 0 to 1, with higher values ​​indicating higher market transaction risk. Technical risk refers to the operational risk arising from grid power regulation demands. The regional regulation urgency signal is a discrete signal issued by the grid dispatch system, characterizing the urgency level of a regional power shortage, with values ​​of 1, 2, and 3, corresponding to low, medium, and high urgency levels, respectively. Technical risk is directly quantified through the signal value, expressed as: In the formula The regional adjustment of the urgency signal value, The technical risk value ranges from 0 to 1, with higher values ​​indicating higher grid regulation risk. Credit risk refers to the default risk arising from the use of credit limits by virtual power plant applications. The current application volume is the output application value submitted by virtual power plants in the market a day prior, denoted as [value missing]. The available credit limit is the maximum credit limit that a virtual power plant can claim, allocated by the power trading platform, denoted as [missing information]. The expression for credit risk calculation is: In the formula This represents the credit risk value, ranging from 0 to 1, with higher values ​​indicating a higher risk of credit default. Initial weights are assigned to these three risk categories. These initial weights are fixed values ​​preset for each risk category during the model's first run, denoted as [values ​​to be inserted here]. , , Corresponding to market risk, technological risk, and credit risk, the weights meet the normalization conditions. The default initial weight configuration is , , Initial weights are fixed and written into the multi-dimensional risk quantification model configuration file, serving as the weight benchmark for the initial comprehensive risk value calculation, ensuring reasonable risk weight allocation during the initial model runtime. Dynamic weight adjustments are made based on historical strategy execution results. These adjustments are based on operational data after strategy execution within a preset period, reverse-optimizing the iterative process of the three types of risk weights. Market returns and deviation assessment results under different weight combinations are periodically statistically analyzed. The default statistical period is set to 7 days, and market returns are the total net revenue of the virtual power plant from electricity trading within the statistical period, denoted as... The deviation assessment result is the total assessment cost incurred due to output deviation within the statistical period, denoted as... The comprehensive performance indicator is the difference between revenue and assessment fees, expressed as: It iterates through all historical weight combinations within the statistical period, records the comprehensive performance index corresponding to each weight group, and selects the weight combination with the best performance as a reference benchmark. The weight update adopts gradient descent optimization logic, and the adjustment formula is as follows: In the formula Let i be the risk weight for the i-th type in period t. For the updated weights, This is the learning rate, with a default value of 0.01. The weights are the partial derivatives of performance with respect to the weights. After the update, the weights need to be renormalized to ensure... The dynamically adjusted weights will serve as the weighting parameters for calculating the comprehensive risk value in the next cycle, allowing the risk weights to continuously optimize with actual operational performance and improving the accuracy of the comprehensive risk value. The dynamically adjusted risk weights are directly substituted into the multi-dimensional risk quantification model to recalculate the comprehensive risk value. The optimized weights are adapted to historical operational results, making the comprehensive risk value more closely match the actual risk exposure level of the virtual power plant. This provides accurate risk input for the subsequent generation of risk preference adjustment coefficients, ensuring the risk adaptability of the final compensation factor.

[0067] Furthermore, the specific process of fusion computation includes:

[0068] The bidding volume correction unit analyzes the original market bid volume of the day before and decomposes the risk appetite adjustment coefficient into planned output components with different time granularities and different physical nodes.

[0069] The final compensation factor is also expanded along the two dimensions of time axis and resource node to form a compensation factor matrix that matches the dimensions of planned output components. During fusion, differential calculations are performed based on the attributes of the planned output components. Finally, all planned output components that have undergone differential calculations are re-aggregated to generate the overall corrected quotation.

[0070] Preferably, after the dynamic compensation factor generation unit 2 completes the generation of the final compensation factor and risk preference adjustment coefficient, the bidding volume correction unit enters a standardized implementation process of original declaration volume analysis, parameter dimension decomposition, compensation factor matrix construction, differentiated fusion calculation, and component aggregation. This process relies on multi-dimensional data decomposition, dimension matching mapping, attribute differentiated calculation, and total aggregation to achieve refined correction of the original declaration volume in the time dimension and resource node dimension, generating a corrected bid volume that adapts to the output characteristics and risk status, ensuring that the bidding declaration matches the actual output capacity of distributed resources. The current market original declaration volume is the output declaration data submitted by virtual power plants to the power trading platform one day in advance, including the planned output values ​​of each time period and each physical node, i.e., each distributed photovoltaic or wind power source, under the preset time granularity of 15 minutes. During analysis, the bidding volume correction unit reads the original declaration volume structured file, decomposes it according to the two dimensions of time granularity and physical node, and extracts each time granularity. Each physical node The corresponding original planned output component is denoted as ,in The value ranges from 1 to T, where T is the total number of time granularities within a day (default 96), and n ranges from 1 to N, where N is the total number of distributed resource nodes aggregated by the virtual power plant. After parsing, a two-dimensional set of original planned output components is formed, providing a dimensional benchmark for subsequent parameter decomposition.

[0071] The risk preference adjustment coefficient is the global risk adjustment parameter output by the dynamic compensation factor generation unit 2, denoted as... The value ranges from 0 to positive numbers. During decomposition, the global coefficient is allocated to each time granularity and physical node based on time weight and node capacity weight. Time weight The weighting is based on the time period's proximity to the current time, with more recent periods having higher weighting, and node capacity also having a higher weighting. The weighting is determined based on the installed capacity ratio of each distributed resource; the larger the capacity, the higher the weighting. The weighting satisfies... , The expression for each component after decomposition is: This yields a set of risk preference adjustment coefficient components consistent with the original planned output component dimensions, used as parameter support for subsequent differential calculations. The final compensation factor is the global compensation benchmark parameter output by the dynamic compensation factor generation unit 2, denoted as... The value is greater than or equal to the basic compensation coefficient. When expanded, along both the time axis and resource node dimensions, the global factor is mapped to a two-dimensional compensation factor matrix according to the same weighting logic as the risk preference adjustment coefficient. The matrix element expression is as follows: The matrix dimension is The dimensions of the planned output components and risk preference adjustment coefficient components are fully matched to ensure dimensional alignment in subsequent fusion calculations and avoid calculation errors caused by dimensional mismatches. During fusion, differentiated calculations are performed based on the attributes of the planned output components. These attributes include three categories: time horizon, resource adjustment characteristics, and output type. Time horizon is divided into near-term and long-term periods; resource adjustment characteristics are divided into adjustable and non-adjustable resources; and output type is divided into photovoltaic and wind power. The differentiated calculation rule is that multiplicative fusion is used for near-term periods. Additive fusion is used in the long term. Adjustable resources adopt weighted fusion For unadjustable resources, a basic multiplication fusion method is used. Photovoltaic resources are prioritized for power output fluctuation compensation, and wind power resources are prioritized for wind speed fluctuation compensation. Differentiated calculations of matching attributes are performed on the planned power output components corresponding to each time granularity and physical node to generate the planned power output components corrected in each dimension. Finally, all the planned output components that have undergone differential calculations are re-aggregated to generate the overall revised quote. The aggregation uses a two-dimensional summation method, expressed as follows: In the formula To adjust the overall bid volume, the reasonableness of the aggregated values ​​is verified. The upper limit is set to the total installed capacity of the virtual power plant, and the lower limit is 0. If the value exceeds the range, it will be automatically adjusted to the boundary value to generate the final bid volume that can be submitted. This bid volume takes into account the output timing characteristics of the time dimension, the individual output differences of resource nodes, and the real-time risk compensation needs. It is directly pushed to the power trading platform to complete the bidding submission, effectively reducing the risk of output deviation assessment and improving the bidding revenue of virtual power plants.

[0072] Please see Figure 2 As shown, a second objective of this invention is to provide a method for implementing a virtual power plant e-bidding system that aggregates distributed resources, including any of the above-mentioned features, comprising the following steps:

[0073] S1. Collect the actual and predicted power of each distributed photovoltaic and wind power according to the preset cycle. The predicted power is generated in a rolling manner by a deep learning model based on cloud numerical weather forecast and historical big data. The actual and predicted power are cleaned and aligned. The percentage of the absolute difference between the actual and predicted power to the installed capacity is calculated. Combined with the dynamic weight determined by the prediction volatility, a rolling error rate is generated. According to the adaptive interval boundary list based on historical distribution statistics, the rolling error rate is dynamically classified into the preset error interval. When there is a risk of grid frequency exceeding the limit, the tightening mechanism of the adaptive interval boundary list is automatically triggered.

[0074] S2. Based on the error interval query, the self-learning mapping table is used. The self-learning mapping table presets the benchmark center point and sub-interval offset logic for each error interval to determine the basic compensation coefficient. At the same time, based on the real-time market volatility, power grid regulation demand and its own credit status, the comprehensive risk value is calculated through a multi-dimensional risk quantification model. Combined with the weight strategy dynamically adjusted by historical performance feedback, the risk preference adjustment coefficient is output through the risk preference curve function set by the operator. Finally, the risk preference adjustment coefficient is non-linearly coupled with the basic compensation coefficient to generate the final compensation factor.

[0075] S3. Analyze the original market bid volume of the day-ahead, decompose the original market bid volume of the day-ahead into planned output components with different time granularities and physical nodes. At the same time, expand the final compensation factor into a compensation factor matrix along the same dimension. Perform differentiated fusion operation on the compensation factor matrix according to the time distance and resource adjustment characteristics of each planned output component. Re-aggregate all planned output components after differentiated operation to generate the overall corrected bid volume.

[0076] S4. Push the revised bid amount to the power trading platform to complete the bidding submission.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A virtual power plant electricity pricing system that aggregates distributed resources, characterized in that: It includes a real-time error monitoring unit (1), a dynamic compensation factor generation unit (2), and a bidding volume correction unit (3). The real-time error monitoring unit (1) collects the actual output data and corresponding predicted output data of distributed photovoltaic and wind power according to a preset cycle. It generates a rolling error rate by calculating the percentage of the absolute difference between the actual power and the predicted power to the installed capacity, and classifies the rolling error rate into different error ranges according to the preset error rate classification rules. The dynamic compensation factor generation unit (2) establishes a mapping relationship between error rate and compensation amount, wherein each error interval maps a unique basic compensation coefficient, and generates the final compensation factor based on the real-time risk strategy of the virtual power plant and the risk preference adjustment coefficient. The bidding volume correction unit (3) performs a fusion calculation between the final compensation factor and the original market declaration volume, generates a corrected bid volume, and pushes the corrected bid volume to the power trading platform.

2. The virtual power plant electricity pricing system for aggregating distributed resources according to claim 1, characterized in that: The acquisition of the actual power and the predicted power specifically includes: The actual power output data is collected in real time by smart meters and synchronous vector measurement devices deployed at various distributed power grid connection points, and the time stamp is marked. Simultaneously, a cloud platform containing numerical weather forecast data and historical power output big data is accessed. Using a trained deep learning prediction model, starting from the end of the previous preset cycle, the power output prediction for the next preset cycle is executed in a rolling manner to generate predicted power output data that is aligned with the actual power output data in terms of time scale. Based on this, the actual power output data and predicted power output data are cleaned and aligned, abnormal jump points are removed, and data with missing time scales is supplemented by interpolation to form actual power and predicted power.

3. The virtual power plant electricity pricing system for aggregating distributed resources according to claim 1, characterized in that: The generation of the rolling error rate specifically includes: Take the actual power and predicted power at each collection point and several consecutive time points before it, calculate the absolute difference at each time point and divide it by the installed capacity of the corresponding distributed power source to obtain the instantaneous percentage error at each time point; Different dynamic weights are assigned to the instantaneous percentage error at each time point. The dynamic weights are determined by the output fluctuation estimate given by the prediction model at each time point. The weighted instantaneous percentage errors at all time points are normalized and summed up, and the result is the rolling error rate at that time point.

4. The virtual power plant electricity pricing system for aggregating distributed resources according to claim 1, characterized in that: The error rate grading rule is a dynamic boundary rule, including: An adaptive interval boundary list based on the historical rolling error rate distribution is introduced. The initial boundary of the adaptive interval boundary list is set according to the historical statistical quantile. When classifying, the current rolling error rate is compared with the latest adaptive interval boundary list to determine the error interval to which it belongs. At the same time, the adaptive interval boundary list is dynamically adjusted according to the real-time operating status of the power grid. When the risk of the power grid frequency fluctuation exceeding the limit is detected, the real-time error monitoring unit (1) automatically triggers the boundary tightening mechanism, and shrinks the boundary value of each error interval towards zero in the next preset period.

5. The virtual power plant electricity sales and pricing system for aggregating distributed resources according to claim 1, characterized in that: The establishment of the mapping relationship between error rate and compensation amount specifically includes: The dynamic compensation factor generation unit (2) has a built-in self-learning mapping table. The self-learning mapping table uses the error interval as the query key. Whenever the error interval is triggered, it records the application effect of the compensation amount generated based on the basic compensation coefficient of the error interval, that is, the degree of matching between the subsequent actual output and the corrected reported amount. The application effect is fed back to the learning parameters of the self-learning mapping table, driving the adjustment of the basic compensation amount mapped by the error interval in the next cycle. If the matching degree is improved after application, the compensation amount is increased in the current direction, and vice versa.

6. The virtual power plant electricity sales and pricing system for aggregating distributed resources according to claim 5, characterized in that: Each error interval maps to a unique basic compensation coefficient, specifically including: The self-learning mapping table pre-sets a reference center point for each error interval and sets sub-interval offset logic for different rolling error rate levels within the error interval: When the rolling error rate falls into a certain error range, its specific position relative to the reference center point of the range is first determined. If it belongs to the high end of the range, the positive offset coefficient corresponding to the error range is triggered, and a positive adjustment is added to the original basic compensation coefficient. If it belongs to the low end of the range, the negative offset coefficient is triggered to make a negative adjustment.

7. The virtual power plant electricity sales and pricing system for aggregating distributed resources according to claim 1, characterized in that: The method for generating the final compensation factor based on the real-time risk strategy of the virtual power plant and the risk preference adjustment coefficient specifically includes: The dynamic compensation factor generation unit (2) continuously accesses the real-time price volatility of the electricity market, the system balance demand, and the utilization rate of the virtual power plant's own credit limit. It calculates the current comprehensive risk value through a multi-dimensional risk quantification model and inputs the comprehensive risk value into a preset risk preference curve function. The risk preference curve function is set by the virtual power plant operator and outputs a risk preference adjustment coefficient in the range of zero to positive numbers. The risk preference adjustment coefficient is nonlinearly coupled with the basic compensation coefficient obtained by mapping the error interval to generate the final compensation factor.

8. The virtual power plant electricity pricing system for aggregating distributed resources according to claim 7, characterized in that: The multi-dimensional risk quantification model assesses market risk, technology risk, and credit risk in parallel. Market risk is represented by the product of price volatility and prediction deviation rate, technology risk is quantified by the regional regulation urgency signal issued by the power grid, and credit risk is represented by the ratio of current application volume to available credit limit. Initial weights are assigned to these three types of risks, and the weights are dynamically adjusted based on the historical performance of the strategy. This includes periodically analyzing the market returns and deviations of the strategy under different weight combinations.

9. The virtual power plant electricity pricing system for aggregating distributed resources according to claim 1, characterized in that: The specific process of the fusion operation includes: The bidding volume correction unit (3) analyzes the original market declaration volume of the day-ahead and decomposes the risk preference adjustment coefficient into planned output components with different time granularities and different physical nodes; The final compensation factor is also expanded along the two dimensions of time axis and resource node to form a compensation factor matrix that matches the dimension of planned output component. During fusion, differential calculation is performed according to the attributes of planned output component. Finally, all planned output components that have undergone differential calculation are re-aggregated to generate the overall corrected quotation.

10. A method for implementing a virtual power plant auction bidding system comprising aggregated distributed resources as described in any one of claims 1-9, characterized in that: Includes the following steps: S1. Collect the actual power and predicted power of each distributed photovoltaic and wind power according to a preset cycle. The predicted power is generated by a deep learning model based on cloud numerical weather forecast and historical big data. The actual power and predicted power are cleaned and aligned. The percentage of the absolute difference between the actual power and the predicted power to the installed capacity is calculated. Combined with the dynamic weight determined by the prediction volatility, a rolling error rate is generated. According to the adaptive interval boundary list based on historical distribution statistics, the rolling error rate is dynamically classified into a preset error interval. When there is a risk of grid frequency exceeding the limit, the tightening mechanism of the adaptive interval boundary list is automatically triggered. S2. Based on the error interval query, the self-learning mapping table is used. The self-learning mapping table presets the benchmark center point and sub-interval offset logic for each error interval to determine the basic compensation coefficient. At the same time, based on the real-time market volatility, power grid regulation demand and its own credit status, the comprehensive risk value is calculated through a multi-dimensional risk quantification model. Combined with the weight strategy of dynamic adjustment based on historical performance feedback, the risk preference adjustment coefficient is output through the risk preference curve function set by the operator. Finally, the risk preference adjustment coefficient is non-linearly coupled with the basic compensation coefficient to generate the final compensation factor. S3. Analyze the original market bid volume of the day-ahead, decompose the original market bid volume of the day-ahead into planned output components with different time granularities and physical nodes, and expand the final compensation factor into a compensation factor matrix along the same dimension. Perform differentiated fusion operation on the compensation factor matrix according to the time distance and resource adjustment characteristics of each planned output component. Re-aggregate all planned output components after differentiated operation to generate the overall corrected bid volume. S4. Push the revised bid amount to the power trading platform to complete the bidding submission.

Images