Fusion pre-training model optimized virtual power plant intra-day regulation reliability discrimination method

Abstract

The application discloses a virtual power plant daily regulation reliability discrimination method based on a fusion pre-training model optimization, and belongs to the technical field of power system operation and control. The steps comprise the following: collecting aggregated resource states, loading historical logs to construct a deviation probability distribution, and generating a stress test sample set; mapping daily regulation power instructions into vectors, inputting the stress test samples into a pre-training model, and outputting a multi-dimensional response sequence; performing physical superposition based on topological connections to generate a feedback power trajectory; calculating the difference between the trajectory and the instruction, screening the most extreme shortage data; and combining real-time clearing prices to calculate risk option costs and output the proportion of the risk option costs in the income. The application adopts a pre-training model to perform stress scenario deduction and physical translation, can quantify the regulation failure risk under extreme working conditions, and improves the reliability evaluation precision and economic decision level of the daily regulation of the virtual power plant.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a method for determining the intraday regulation reliability of a virtual power plant by integrating pre-trained model optimization. Background Technology

[0002] A virtual power plant is an energy management system that aggregates and coordinates distributed energy resources such as distributed power sources, energy storage systems, and controllable loads through information and communication technologies and software systems. In the operation of the electricity market, intraday regulation is a key link for virtual power plants to respond to real-time grid dispatch instructions and maintain power balance. The reliability of its regulation capability is directly related to the safe and stable operation of the power grid.

[0003] Existing research has significant shortcomings in addressing extreme operating conditions: Firstly, sample generation is mostly based on typical operating scenarios, lacking coverage of the most extreme / maximum risk power deficit scenarios, making it difficult to fully expose the weaknesses of regulation strategies under critical conditions. Secondly, the utilization of historical operating deviation logs is mostly limited to simple statistics, failing to fully explore the probability distribution characteristics of deviations, resulting in insufficient sample diversity and stress test intensity. Furthermore, the differences in resource characteristics across different regulation periods (peak, mid-load, and off-peak) are not effectively distinguished, making it impossible for the sample set to accurately match the regulation needs of each period, reducing the relevance and reference value of the test results.

[0004] The aforementioned limitations directly lead to risks in the actual operation of virtual power plant regulation strategies. For example, in extreme deficit scenarios, problems such as insufficient regulation capacity and excessive response delay may occur, affecting the safe and stable operation of the power grid. At the same time, the incompleteness of the sample set makes the strategy evaluation results too optimistic and cannot provide reliable risk warning basis for dispatch decisions.

[0005] Therefore, it is urgent to build a stress test sample generation method that can cover all time periods and all operating conditions. By integrating multi-dimensional operating data and historical deviation information, the simulation capability of extreme scenarios can be strengthened, providing technical support for improving the reliability and robustness of intraday regulation of virtual power plants. Summary of the Invention

[0006] To address the aforementioned issues, the present invention aims to provide a method for determining the reliability of intraday regulation in virtual power plants by integrating pre-trained model optimization. This method employs a pre-trained model for stress scenario simulation and physical back-translation, which can quantify the risk of regulation failure under extreme operating conditions and improve the reliability assessment accuracy and economic decision-making level of intraday regulation in virtual power plants.

[0007] The above objectives can be achieved through the following approach:

[0008] The method for determining the intraday regulation reliability of a virtual power plant by integrating pre-trained models includes the following steps:

[0009] The system utilizes virtual power plants to aggregate resources to collect data on energy storage status of charge, controllable load adjustment margin, and communication latency. It also loads historical operation deviation logs to construct a deviation probability distribution, performs random sampling and element splicing, and builds a stress test sample set.

[0010] The system acquires intraday power regulation commands and maps them to timing command vectors. It then inputs the timing command vectors and the set of stress test samples into a preset timing pre-trained model, performs forward inference, and outputs energy storage charging and discharging power sequences, controllable load response sequences, and distributed power output sequences for different stress scenarios.

[0011] By leveraging the topological connections of the virtual power plant, the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence are physically superimposed to generate a feedback power trajectory.

[0012] The timing command vector and the feedback power trajectory are differentially processed to extract the positive deviation as the power deficit sequence. The power deficit sequence is then accumulated using a stress test sample set, and the sequence with the largest accumulated value is selected as the most extreme deficit data.

[0013] By accessing real-time clearing price data through the ancillary services market, and combining the most extreme shortage data with the real-time clearing price data, risk cost accounting is performed to obtain the risk option cost, and the percentage of the risk option cost in the intraday adjustment task revenue is output.

[0014] Preferably, constructing the stress test sample set includes:

[0015] The system reads the energy storage charge status by calling the status register through the Internet of Things (IoT) gateway, reads the controllable load adjustment margin through the building temperature control system, sends heartbeat packets to the terminal equipment and records the round-trip time to generate communication latency data.

[0016] Based on historical operational deviation logs, the energy storage state of charge deviation, controllable load adjustment margin deviation, and communication delay deviation are extracted. The bin counting is performed and the cumulative frequency is calculated to generate a deviation probability distribution.

[0017] Random sampling is performed based on the deviation probability distribution to generate a disturbance noise vector. The disturbance noise vector, energy storage state of charge, controllable load adjustment margin, and communication delay data are then concatenated element by element to generate a stress test sample set.

[0018] Preferably, the output sequences of energy storage charging and discharging power, controllable load response, and distributed power generation for different pressure scenarios include:

[0019] The power value and start and end timestamps of the intraday power adjustment command are extracted by parsing the command. The power value is normalized and the start and end timestamps are mapped to time feature vectors. The vectors are then concatenated to generate a time-series command vector.

[0020] The timing instruction vector and the stress test sample set are concatenated along the feature dimension to construct a joint input tensor;

[0021] The time-series pre-trained model is invoked to perform matrix multiplication and activation operations on the joint input tensor to generate a high-dimensional feature tensor. Channel dimension slicing is then performed on the high-dimensional feature tensor to extract the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence, respectively.

[0022] Preferably, the output of energy storage charging and discharging power sequences, controllable load response sequences, and distributed power generation output sequences for different pressure scenarios also includes:

[0023] The system retrieves intraday power regulation commands from historical time periods and simultaneously extracts data on energy storage state of charge, controllable load regulation margin, and communication latency from historical time periods. It then performs time alignment and feature splicing to construct a historical training input tensor.

[0024] Based on the actual power response records of virtual power plant aggregated resources obtained from historical time periods, data cleaning and channel dimension stacking are performed to construct a multi-dimensional measured response matrix.

[0025] Using the historical training input tensor as input and the multidimensional measured response matrix as the regression objective, a temporal regression network is established and gradient descent training is performed to minimize the prediction residual and generate a temporal pre-trained model.

[0026] Preferably, generating the feedback power trajectory includes:

[0027] Based on the electrical topology of the virtual power plant, the output sequence of distributed power sources and the discharge power of energy storage are marked as positive supply components, and the response sequence of controllable loads and the charging power of energy storage are marked as negative demand components, thus establishing the power flow direction identifier.

[0028] Based on the energy storage state of charge, the energy storage charging and discharging power sequence is accumulated hourly to identify the time points when the energy is depleted or fully charged, and the power value is forced to zero based on the time points to generate an effective energy storage power sequence.

[0029] The effective energy storage power sequence, the distributed power output sequence, and the controllable load response sequence are algebraically added according to the power flow direction identifier to generate a feedback power trajectory.

[0030] Preferably, the sequence with the largest cumulative value is selected as the most extreme shortage data, including:

[0031] The timing command vector and the feedback power trajectory are subtracted at each time step to generate the original deviation trajectory. If the elements in the original deviation trajectory with values ​​less than zero are forcibly set to zero, a power deficit sequence is generated.

[0032] For intraday adjustment reading of time step parameters, the power deficit sequence and time step parameters are summed over the entire sequence to calculate the cumulative power deficit.

[0033] Based on the stress test sample set, the cumulative power deficit is sorted by value, the index of the scenario with the largest value is locked, and the power deficit sequence is extracted as the most extreme deficit data.

[0034] Preferably, the percentage of output risk option costs in the intraday adjustment task revenue includes:

[0035] Based on the disturbance noise vector and power deficit sequence, vector covariance calculation is performed to quantify the impact of disturbance amplitude on the deficit and generate a disturbance sensitivity matrix.

[0036] The time-series instruction vector and real-time clearing price data are multiplied element-wise and accumulated in the time domain to generate the intraday adjustment task profit.

[0037] Based on the perturbation sensitivity matrix, diagonal elements are extracted to construct a time-series sensitivity vector. The time-series sensitivity vector is then weighted and fused with real-time clearing price data to generate a risk premium correction sequence.

[0038] The most extreme shortage data is multiplied and summed with the risk premium correction sequence to generate the risk option cost. The risk option cost is then divided by the intraday adjustment task revenue to calculate the percentage value.

[0039] Preferably, the benefits of generating intraday adjustment tasks include:

[0040] The time-series instruction vector and real-time clearing price data are multiplied element-wise at each time step to generate an instantaneous value sequence;

[0041] The instantaneous value sequence is multiplied by the time step parameter and then summed across the entire sequence to generate the intraday adjustment task profit.

[0042] A virtual power plant intraday regulation reliability discrimination system optimized by integrating pre-trained models is used to implement the above method, including:

[0043] The stress test sample construction module is used to collect data on energy storage charge status, controllable load adjustment margin and communication latency by aggregating resources in a virtual power plant, load historical operation deviation logs to construct deviation probability distribution, perform random sampling and element splicing, and construct a stress test sample set.

[0044] The timing instruction inference output module is used to acquire intraday power regulation instructions and map them into timing instruction vectors. The timing instruction vectors and the stress test sample set are input into the preset timing pre-trained model to perform forward inference and output energy storage charging and discharging power sequences, controllable load response sequences and distributed power output sequences for different stress scenarios.

[0045] The feedback power trajectory generation module is used to physically superimpose the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence through the topological connection relationship of the virtual power plant to generate the feedback power trajectory.

[0046] The most extreme deficit screening module is used to perform differential operation on the timing command vector and the feedback power trajectory, extract the positive deviation part as the power deficit sequence, and perform accumulation on the power deficit sequence through the stress test sample set, and select the sequence with the largest accumulated value as the most extreme deficit data.

[0047] The risk option cost calculation module is used to access real-time clearing price data through the ancillary service market, combine the most extreme shortage data with the real-time clearing price data to perform risk cost calculation, obtain the risk option cost, and output the percentage of risk option cost in the intraday adjustment task revenue.

[0048] The present invention has the following advantages:

[0049] This invention solves the problem of output distortion in purely data-driven models under extreme operating conditions by constructing a stress test sample set based on historical deviation distribution and combining virtual power plant topology connections and energy conservation constraints for physical back-translation. Compared with traditional static capacity verification, it can deduce the dynamic response trajectory of energy storage and load under the dual constraints of data disturbance and physical boundaries, improving the reliability of intraday regulation commands and the accuracy of physical feasibility judgment in uncertain environments.

[0050] This invention introduces a risk option cost quantification mechanism. By screening the most extreme shortage data and combining it with real-time ancillary service market clearing prices, it transforms the abstract risk of technical-side adjustment failure into specific economic-side reserve procurement costs. This "technology-economic" coupled judgment dimension not only avoids the shortcomings of a single technical indicator in measuring market risk, but also outputs the proportion of risk cost in expected returns, providing virtual power plant operators with a basis for intraday adjustment decisions based on value maximization.

[0051] This invention leverages the powerful nonlinear feature extraction and multi-task decoupling capabilities of a time-series pre-trained model to achieve end-to-end high-dimensional mapping from total power commands to the response sequences of each subsystem (source, load, and storage). Compared to traditional step-by-step rule-based computation, it can capture the coupling characteristics and dynamic delay patterns of different aggregated resources in the time domain. While ensuring computational efficiency, it achieves refined modeling of microscopic features such as energy storage state of charge, load regulation margin, and communication latency, enhancing the model's generalization adaptability to complex heterogeneous resources. Attached Figure Description

[0052] Figure 1This is a flowchart illustrating the intraday regulation reliability discrimination method for virtual power plants that integrates pre-trained model optimization according to the present invention.

[0053] Figure 2 This is a parallel coordinate graph showing the coupling relationship between multi-source heterogeneous perturbation parameters and regulation deficit risk in Embodiment 1 of the present invention;

[0054] Figure 3 This is a probability distribution and long-tail risk characteristic diagram of power deficit under different adjustment periods in Embodiment 1 of the present invention;

[0055] Figure 4 This is a schematic diagram of the structure of the virtual power plant intraday regulation reliability discrimination system optimized by integrating pre-trained model according to the present invention. Detailed Implementation

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0057] Example 1: As Figure 2 As shown, the method for determining the intraday regulation reliability of a virtual power plant by integrating pre-trained model optimization includes the following steps:

[0058] The system utilizes virtual power plants to aggregate resources to collect data on energy storage status of charge, controllable load adjustment margin, and communication latency. It also loads historical operation deviation logs to construct a deviation probability distribution, performs random sampling and element splicing, and builds a stress test sample set.

[0059] The system acquires intraday power regulation commands and maps them to timing command vectors. It then inputs the timing command vectors and the set of stress test samples into a preset timing pre-trained model, performs forward inference, and outputs energy storage charging and discharging power sequences, controllable load response sequences, and distributed power output sequences for different stress scenarios.

[0060] By leveraging the topological connections of the virtual power plant, the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence are physically superimposed to generate a feedback power trajectory.

[0061] The timing command vector and the feedback power trajectory are differentially processed to extract the positive deviation as the power deficit sequence. The power deficit sequence is then accumulated using a stress test sample set, and the sequence with the largest accumulated value is selected as the most extreme deficit data.

[0062] By accessing real-time clearing price data through the ancillary services market, and combining the most extreme shortage data with the real-time clearing price data, risk cost accounting is performed to obtain the risk option cost, and the percentage of the risk option cost in the intraday adjustment task revenue is output.

[0063] Building the stress test sample set includes:

[0064] The system reads the energy storage charge status by calling the status register through the Internet of Things (IoT) gateway, reads the controllable load adjustment margin through the building temperature control system, sends heartbeat packets to the terminal equipment and records the round-trip time to generate communication latency data.

[0065] A data link is established with the edge IoT gateway using Modbus TCP or IEC 61850 industrial communication protocols. Read commands are sent to the energy storage converter PCS or battery management system BMS at millisecond intervals to directly obtain the remaining percentage of charge stored in a specific hexadecimal address register, i.e., the energy storage state of charge. Simultaneously, the building automation system is accessed via BACnet or OPC UA protocols to capture the current operating power setpoint and rated power limit of the central air conditioning, lighting, and electric heating equipment in real time. The difference between these two values ​​is calculated as the upward adjustment margin, and the difference between the current operating power and the minimum sustaining power is calculated as the downward adjustment margin. For stability quantification of the communication link, ICMP echo request messages are sent to the remote terminal units (RTUs) of each aggregated resource, and the timestamp of the message transmission is recorded using a high-precision system clock. and the timestamp of the received response message Communication delay data is obtained by calculating the difference. .

[0066] One application example is as follows: Assuming the current time is 10:00:00, the processor reads the value 0x55 from register 0x001A via the Modbus protocol. Converted to decimal, this is 85, representing a state of charge (SOC) of 85%. Simultaneously, it reads that the rated power of an air conditioning unit in a commercial building is 500kW, and the current operating power is 300kW. Therefore, the controllable load adjustment margin is calculated to be 200kW. A heartbeat packet is sent to the building's terminal at 10:00:00.005 and received at 10:00:00.045, resulting in a communication latency of 40ms.

[0067] Based on historical operational deviation logs, the energy storage state of charge deviation, controllable load adjustment margin deviation, and communication delay deviation are extracted. The bin counting is performed and the cumulative frequency is calculated to generate a deviation probability distribution.

[0068] Retrieve historical operation logs stored in the time-series database, such as records from the past 30 days. For each record, calculate the algebraic difference between the predicted value at the time the instruction was issued and the measured value at the subsequent actual execution time, and obtain three sets of one-dimensional data sequences: energy storage state of charge deviation, controllable load adjustment margin deviation, and communication delay deviation.

[0069] Then, histogram statistics were used to discretize each set of deviation sequences: first, the maximum value of the deviation data was determined. and minimum value , will the interval Divide the sample into K equal-width sub-intervals (bins) and count the number of samples falling into each sub-interval. And calculate the frequency of occurrence in each interval. Based on the frequency values ​​of each interval, they are accumulated sequentially in ascending order to generate a discrete distribution function reflecting the cumulative probability of the deviation. The interval frequency calculation formulas involved in this process are as follows:

[0070] ;

[0071] in, This represents the probability frequency of the i-th deviation interval, with a value range of... The physical meaning is the probability of this deviation range occurring during historical operation; This represents the number of historical deviation samples falling into the i-th interval; This represents the total number of historical samples included in the statistics; i is the interval index, with a value of This formula is used to transform discrete statistical counts into standardized probability values, providing a mathematical basis for sampling.

[0072] One application example involves extracting 1000 historical communication latency deviation data points. The deviation range is set to -10ms to +50ms, divided into 6 intervals. Statistical analysis reveals that 400 samples have a deviation in the [0ms, 10ms] interval. Therefore, the frequency of this interval... The number of samples with deviations in the range of [10ms, 20ms] is 300, with a frequency of 0.3. After sequential calculation, the cumulative frequency sequence obtained is {0.1, 0.5, 0.8, 0.9, 0.95, 1.0}, which constitutes the deviation probability distribution of the communication delay deviation.

[0073] Random sampling is performed based on the deviation probability distribution to generate a disturbance noise vector. The disturbance noise vector, energy storage state of charge, controllable load adjustment margin, and communication delay data are then concatenated element by element to generate a stress test sample set.

[0074] The inverse transform sampling method is used to simulate random disturbances. First, a pseudo-random number generator is used to generate a number that follows the... A uniformly distributed random number *u* is generated. Then, the first cumulative frequency value greater than or equal to *u* corresponding to the deviation interval is found in the deviation probability distribution. The center value of this interval is taken as the disturbance value obtained from the first sampling. This sampling operation is performed on the energy storage state of charge, controllable load regulation margin, and communication delay, respectively, to obtain three independent disturbance values, which form a disturbance noise vector. Finally, the real-time acquired raw physical quantities and the generated disturbance noise vector are concatenated along the feature channel dimension to construct a high-dimensional feature vector containing both the actual state and potential interference, i.e., the stress test sample. This concatenation and generation process can be represented by the following vector formula:

[0075] ;

[0076] in, This represents a single stress test sample vector generated, which physically represents a complete snapshot of the environment input to the pre-trained model. The state vector is acquired in real time, including the energy storage state of charge (SOC) and the controllable load regulation margin. and communication latency ; The perturbation noise vector generated by random sampling contains the corresponding bias value. , and ; This represents the concatenation or splicing operation of vectors. By repeatedly performing random sampling and splicing, a set containing M samples is generated. .

[0077] An example application is as follows: the current real-time data acquisition is: SOC = 80%, margin = 100kW, and latency = 20ms. The generated random number u = 0.85 corresponds to a high latency deviation range in the historical distribution, and the sampled SOC deviation is -5%, and the margin deviation is -10kW. The generated disturbance noise vector is then [-5, -10, 30]. After concatenation, the specific numerical vector of this stress test sample is [80, 100, 20, -5, -10, 30]. This sample will be fed into the model to test reliability under extreme conditions such as "SOC only 75%, margin only 90kW, and latency as high as 50ms".

[0078] The output sequences of energy storage charging and discharging power, controllable load response, and distributed power generation for different pressure scenarios include:

[0079] The power value and start and end timestamps of the intraday power adjustment command are extracted by parsing the command. The power value is normalized and the start and end timestamps are mapped to time feature vectors. The vectors are then concatenated to generate a time-series command vector.

[0080] The processor first receives the intraday power adjustment command in JSON or XML format from the scheduling center, and extracts the target power value through field parsing. and the start time of instruction execution and end time To eliminate the influence of dimensions and adapt to the activation function sensitivity region of the neural network, a max-min normalization method is used to map the power value to... or The dimensionless interval. At the same time, for the start and end timestamps, they are discretized into a time step sequence with minute granularity, and each time step is mapped to a multi-dimensional time feature vector using sinusoidal position coding or one-hot coding.

[0081] The normalized power value is copied and extended to the same length as the time step sequence, and then concatenated with the time feature vector on the feature channel to form the time-series command vector. The normalization process uses the following formula:

[0082] ;

[0083] in, This represents the normalized power value, which is dimensionless and typically ranges from 1 to 10. ; This represents the original power value in the intraday power adjustment command. This represents the historical maximum adjustable power limit of the aggregated resources of the virtual power plant; This represents the minimum adjustable power limit of the aggregated resources in the virtual power plant. Its physical significance lies in mapping adjustment commands of different magnitudes to a unified scale space, preventing the model gradient from vanishing or exploding due to excessive numerical differences.

[0084] One application example is assuming the total regulation capacity of the virtual power plant ranges from -10MW to +10MW, i.e. The received intraday power adjustment instruction requires a power reduction of 5MW between 14:00 and 14:15, i.e. Substitute into the formula to calculate: Meanwhile, the time step index corresponding to 14:00 is 840, which is encoded as a 4-dimensional vector [0.86, -0.5, 0, 1]. The final generated timing instruction vector for this time step is [0.25, 0.86, -0.5, 0, 1].

[0085] The timing instruction vector and the stress test sample set are concatenated along the feature dimension to construct a joint input tensor;

[0086] The timing instruction vector has a time dimension, for example, T time steps, while each sample in the stress test sample set is a static feature vector within that time period. Using a broadcast or copy mechanism, the feature vector of a single stress test sample is copied T times along the time axis, making its dimension consistent with the time dimension of the timing instruction vector. Then, in the feature dimension, the expanded stress test sample vector is concatenated with the timing instruction vector. This operation is repeated for N samples in the stress test sample set, ultimately constructing a vector of shape... The three-dimensional joint input tensor, where It is the sum of the instruction feature dimension and the stress sample feature dimension.

[0087] An application example is setting the intraday adjustment duration to 15 minutes and the time resolution to 1 minute, then the time step T=15. The feature dimension of the timing command vector is 5, consisting of 1 power value + 4 time features. The stress test sample includes 6 features: energy storage state of charge, controllable load regulation margin, communication latency and its disturbance terms. These 6 static features are copied 15 times to form... The matrix, then with By concatenating the instruction matrix, we obtain The input is a single sample. If the stress test sample set contains 1000 samples, the final constructed joint input tensor will have a dimension of [dimensionality missing]. .

[0088] The time-series pre-trained model is invoked to perform matrix multiplication and activation operations on the joint input tensor to generate a high-dimensional feature tensor. Channel dimension slicing is then performed on the high-dimensional feature tensor to extract the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence, respectively.

[0089] The joint input tensor is fed into the temporal pre-trained model. The last layer inside the model is a fully connected layer, whose core computation process involves linearly transforming the hidden state and then superimposing nonlinear activations. After the computation is complete, the model outputs a shape... The high-dimensional feature tensor is defined as follows, where "3" represents the number of channels in the output layer, corresponding to the power of the three subsystems. The processor uses tensor slicing operations to separate the tensor along the last dimension. Data with index 0 is extracted as the energy storage charging and discharging power sequence, data with index 1 is extracted as the controllable load response sequence, and data with index 2 is extracted as the distributed power output sequence. The inference process of this fully connected layer can be represented by the following matrix operation formula:

[0090] ;

[0091] in, The high-dimensional feature tensor representing the model output has the physical meaning of the predicted power response value of each aggregated resource under the current stress scenario. represents the time-series hidden layer feature matrix extracted from the first few layers of the model; W represents the weight matrix of the fully connected layer, which is obtained through training on historical data and stores the mapping relationship from input features to output power; b represents the bias vector. This represents the activation function.

[0092] One application example is where the joint input tensor is processed through multiple layers of the model before entering the output layer. Assume the shape of the weight matrix W is... Hidden features The shape is Perform matrix multiplication. get The output matrix. Assume the output vector at time t=5 is... The processor performs the following slices: extracting the first channel value of 0.4, after inverse normalization, the energy storage charging and discharging power sequence at t=5 is 4MW; extracting the second channel value of -0.2, the controllable load response sequence is reduced by 2MW; extracting the third channel value of 0.8, the distributed power output sequence is 8MW.

[0093] The output sequences for energy storage charging and discharging power, controllable load response, and distributed power generation for different pressure scenarios also include:

[0094] The system retrieves intraday power regulation commands from historical time periods and simultaneously extracts data on energy storage state of charge, controllable load regulation margin, and communication latency from historical time periods. It then performs time alignment and feature splicing to construct a historical training input tensor.

[0095] The system retrieves daily power regulation command logs within a set time span from the historical archive database of the SCADA (Supervisory Control and Data Acquisition) system using SQL queries. For each command record in the log, the server synchronously extracts a snapshot of the virtual power plant's aggregated resources at that moment based on its timestamp index, including energy storage state of charge, controllable load regulation margin, and communication latency data. To eliminate the problem of inconsistent sampling frequencies from different data sources, a time alignment operation is performed, using resampling or zero-order hold methods to uniformly map all data to a fixed time granularity. Subsequently, a feature concatenation operation is performed, concatenating the aligned daily power regulation command values ​​with the energy storage state of charge, controllable load regulation margin, and communication latency data along the feature channel dimension to construct a single-moment feature vector, which is then stacked chronologically to form the historical training input tensor.

[0096] One application example is as follows: Suppose historical data from January to June 2024 is retrieved, and the uniform time granularity is set to 1 minute. For the time 10:00 AM on January 1, 2024, the extracted data shows an intraday power regulation command of +5MW, a state of charge of energy storage of 60%, a controllable load regulation margin of 200kW, and a communication latency of 30ms. First, these values ​​are normalized. Assuming the normalized values ​​are... Concatenate these four values ​​to form the feature vector for that moment. Perform this operation on all moments within the same six-month period to construct a feature vector of shape [shape missing]. A two-dimensional tensor is used as the historical training input tensor.

[0097] Based on the actual power response records of virtual power plant aggregated resources obtained from historical time periods, data cleaning and channel dimension stacking are performed to construct a multi-dimensional measured response matrix.

[0098] The actual operating curves of each aggregated resource during historical periods are obtained from the automated metering system (AMI), including energy storage charging and discharging power sequences, controllable load response sequences, and distributed power generation output sequences. Data cleaning is performed to address packet loss or dead pixels during data transmission: null values ​​or abnormal jump values ​​exceeding physical limits in the sequences are identified and filled using linear interpolation. After cleaning, the three sets of actual power sequences are stacked along the channel dimension to construct a multidimensional measured response matrix that strictly corresponds to the historical training input tensor time series.

[0099] Using the historical training input tensor as input and the multidimensional measured response matrix as the regression objective, a temporal regression network is established and gradient descent training is performed to minimize the prediction residual and generate a temporal pre-trained model.

[0100] Initialize a temporal regression network based on a Long Short-Term Memory (LSTM) network. Divide the constructed historical training input tensor into training and validation sets, for example, in an 8:2 ratio, and input them into the network in batches. Set the network output layer to have three neurons, corresponding to the power prediction of energy storage, load, and power source, respectively. Calculate the difference between the predicted values ​​of the network output and the true values ​​in the multidimensional measured response matrix, using the mean squared error (MSE) as the loss function. Calculate the gradient using the backpropagation algorithm and update the network weights using the Adam optimizer until the loss function value of the validation set converges to below a preset threshold. Save the current weight parameters to generate a pre-trained temporal regression model.

[0101] The generated feedback power trajectory includes:

[0102] Based on the electrical topology of the virtual power plant, the output sequence of distributed power sources and the discharge power of energy storage are marked as positive supply components, and the response sequence of controllable loads and the charging power of energy storage are marked as negative demand components, thus establishing the power flow direction identifier.

[0103] The processor reads the main electrical wiring diagram or topology configuration file of the virtual power plant and parses the power transfer direction of each aggregated resource relative to the point of common coupling (PCC). A symbol mapping rule is established: any action that injects active power into the grid is defined as "positive," and any action that absorbs active power from the grid is defined as "negative." Based on this rule, the processor iterates through the output sequences of the model, marking the "discharge" portion of the distributed generation output sequence and the energy storage charging / discharging power sequence as a positive supply component with the sign +1; and marking the "charging" portion of the controllable load response sequence and the energy storage charging / discharging power sequence as a negative demand component. A power flow direction identifier vector corresponding to the time step is generated for subsequent algebraic operations.

[0104] An example application scenario is a virtual power plant comprising a photovoltaic (PV) power station, an energy storage container, and an industrial load. At 10:00 AM, the PV system generates 5MW of power, the energy storage container is discharging at 2MW, and the industrial load consumes 4MW. The PV and energy storage discharges are labeled as positive supply components, and the industrial load is labeled as a negative demand component. The established power flow is represented as: PV +1, Energy Storage Discharge +1, Load -1.

[0105] Based on the energy storage state of charge, the energy storage charging and discharging power sequence is accumulated hourly to identify the time points when the energy is depleted or fully charged, and the power value is forced to zero based on the time points to generate an effective energy storage power sequence.

[0106] Starting from the initial state of charge (SBC) of the energy storage system obtained from the stress test sample, discrete integral calculations are performed on the energy storage charge / discharge power sequence output by the model in time step sequence to deduce the theoretical SBC at each time point. During the calculation process, the deduced SBC is compared with the physical boundary of the energy storage system in real time. Once the SBC at a certain time t is detected to have reached the lower or upper limit, that time is determined as the constraint cutoff point. For the discharge scenario, if the lower limit is reached, the discharge power values ​​at time t and all subsequent times within that period are forcibly modified to 0; for the charging scenario, if the upper limit is reached, the charging power values ​​at time t and all subsequent times are forcibly modified to 0.

[0107] The sequence after this physical constraint correction is the effective energy storage power sequence. The energy accumulation and state derivation process is based on the following formula:

[0108] ;

[0109] in, The state of charge of the energy storage at the end of the t-th time step (time t); : The state of charge of the energy storage at the end of the (t-1)th time step, with the initial time being ; : The energy storage charging and discharging power at the t-th time step, where discharging is defined as a positive value and charging as a negative value; Time step, for example, 15 minutes corresponds to 0.25 hours; Rated capacity of the energy storage system. The formula originates from the law of conservation of energy in electrochemical energy storage systems, and aims to quantify the consumption or replenishment effect of power output on the remaining energy of the battery. It can map instantaneous power commands into a continuous energy state trajectory, thereby using energy boundaries to inversely constrain the feasibility of power output and prevent the physical paradox of "the battery is empty but still discharging".

[0110] One application example is: assuming the rated capacity of energy storage... Initial state of charge Time step The model predicts the discharge power in the first hour. Calculate the state at the end of the first hour: Since -40% is less than the lower limit of 0%, it was detected that the power was exhausted within the first hour. Therefore, a forced zeroing operation was performed to correct the effective power of the first hour to only be able to release the remaining 0.2MWh, or the power at that moment and subsequent moments was directly set to zero to generate an effective energy storage power sequence.

[0111] The effective energy storage power sequence, the distributed power output sequence, and the controllable load response sequence are algebraically added according to the power flow direction identifier to generate a feedback power trajectory.

[0112] Based on the power flow direction identifier, the power sequences of each subsystem are algebraically summed at the corresponding time points. The effective energy storage power sequence and the distributed generation output sequence are added as positive terms, and the controllable load response sequence and the effective energy storage power sequence are added as negative terms. The resulting net value sequence is the feedback power trajectory, which represents the net power actually injected into or absorbed by the grid at the grid connection point (PCC) of the virtual power plant as a whole.

[0113] An example application is given, at a certain time t, the following sequence values ​​are given: distributed power output 10MW, effective energy storage capacity is 5MW (discharge), and controllable load response is 12MW (consumption). The processor performs the following calculations: The result indicates that at this moment, the virtual power plant as a whole is feeding 3MW of power back to the grid. A negative result indicates that the virtual power plant is absorbing power from the grid. Arranging the calculation results at all times in chronological order yields the feedback power trajectory.

[0114] The sequence with the largest cumulative value is selected as the most extreme shortage data, including:

[0115] The timing command vector and the feedback power trajectory are subtracted at each time step to generate the original deviation trajectory. If the elements in the original deviation trajectory with values ​​less than zero are forcibly set to zero, a power deficit sequence is generated.

[0116] First, the timing instruction vector and feedback power trajectory of the intraday power adjustment instructions stored in memory are invoked. At each time step, a subtraction operation is performed, and the result forms the original deviation trajectory. Then, each element in this trajectory is traversed and a non-negativity filter is performed: if the value at a certain moment is greater than zero, it indicates that the actual response is less than the instruction requirement, and a deficit exists; this value is retained. If the value at a certain moment is less than or equal to zero, it indicates that the actual response exceeds or equals the instruction requirement, and there is no risk of a deficit; in this case, the value is forcibly set to zero. The time series after this processing is the power deficit sequence.

[0117] An application example is as follows: Suppose an intraday regulation order requires a power output of 10MW between 10:00 and 10:45. At 10:00, the feedback power trajectory shows a power output of 8MW. The calculated deviation is 10 - 8 = 2MW. At 10:15, the feedback power trajectory shows a power output of 11MW. The calculated deviation is 10 - 11 = -1MW. At 10:30, the feedback power trajectory shows a power output of 10MW. The calculated deviation is 10 - 10 = 0MW. The generated power deficit sequence values ​​at these three times are as follows: .

[0118] For intraday adjustment reading of time step parameters, the power deficit sequence and time step parameters are summed over the entire sequence to calculate the cumulative power deficit.

[0119] The intraday adjustment time resolution, i.e., the time step parameter, is read from the configuration file; for example, 15 minutes corresponds to 0.25 hours. Each power value in the power deficit sequence is multiplied by this time step parameter, converting the "power deficit" into an "energy deficit." Subsequently, a full sequence summation operation is performed on the converted sequence to obtain a scalar value, i.e., the cumulative energy deficit. This value represents the total energy that the virtual power plant failed to deliver during the entire adjustment period under the current pressure scenario. This calculation process is based on the following formula:

[0120] ;

[0121] in, : Represents the cumulative power deficit under the current pressure scenario; T: Represents the total number of time steps for the daily adjustment task; t: Represents the time step index, with a value from 1 to T; : Represents the value of the power deficit sequence at time step t.

[0122] Based on the stress test sample set, the cumulative power deficit is sorted by value, the index of the scenario with the largest value is locked, and the power deficit sequence is extracted as the most extreme deficit data.

[0123] Since the stress test sample set contains M samples, the aforementioned deficit calculation will be performed on each of these 1000 samples to obtain 1000 corresponding cumulative deficit power values. These power values ​​are then sorted in descending order using quicksort, or a maximum value search algorithm is used to locate the largest cumulative deficit power value. The original index position of this maximum value in the set is recorded. Subsequently, this index is used to retrieve the entire power deficit sequence generated under this specific scenario from memory, mark it, and extract it as the most extreme deficit data for economic quantification.

[0124] An application example is to assume that three stress test scenarios are generated. The power deficit sequence for scenario A is as follows: Time step 0.25h. Cumulative power deficit. The power deficit sequence for scenario B is as follows: Cumulative power deficit The power deficit sequence for scenario C is as follows: Cumulative power deficit Comparison revealed Therefore, scenario C is determined to be the most extreme scenario. The index of scenario C is locked, and the sequence... Extracted as the most extreme shortage data.

[0125] The percentage of risk option costs in the intraday adjustment task revenue includes:

[0126] Based on the disturbance noise vector and power deficit sequence, vector covariance is calculated to quantify the impact of disturbance amplitude on the deficit and generate a disturbance sensitivity matrix.

[0127] All sample data in the stress test sample set are structurally reorganized. The perturbation noise vectors corresponding to the M samples are stacked row-wise to construct an M-row, 3-column perturbation sample matrix. Simultaneously, the power deficit sequences corresponding to these M samples are stacked row-wise to construct an M-row, T-column deficit response matrix. Then, the covariance between the column vectors is calculated: traversing each column of the perturbation sample matrix and each column of the deficit response matrix, the statistical covariance between these two data vectors is calculated. Through this full permutation calculation, a perturbation sensitivity matrix with 3 rows and T columns is generated.

[0128] This matrix directly maps the strength of the linear correlation between input uncertainty and output moderation risk. The covariance is calculated using the following formula:

[0129] ;

[0130] in, : Represents the element value in the z-th row and t-th column of the disturbance sensitivity matrix; z: Represents the disturbance type index, where 1 represents the energy storage state of charge deviation, 2 represents the controllable load adjustment margin deviation, and 3 represents the communication delay deviation; k: Represents the sample index, with values ​​ranging from 1 to M; : Represents the value of the z-th type of perturbation noise vector in the k-th sample; : represents the arithmetic mean of the z-th type of perturbation noise vector across all M samples; : Represents the specific value of the power deficit sequence generated by the k-th sample at the t-th time step; : represents the arithmetic mean of the power deficit at time step t across all M samples.

[0131] like A large positive value means that when the z-th type of disturbance increases, the power deficit at time t tends to increase significantly, indicating that the risk at that time mainly comes from this type of disturbance; if it is close to zero, it means that the two are unrelated. Figure 2 The study demonstrates the linkage trajectory between three input dimensions—energy storage state of charge deviation, controllable load regulation margin deviation, and communication delay deviation—and the cumulative deficit power output dimension. The dark lines represent specific parameter combinations that lead to high-risk deficits, reflecting the coupling effect mechanism of multi-source heterogeneous disturbances on the regulation reliability of virtual power plants.

[0132] An application example is as follows: Suppose a stress test generates M=1000 samples. The sensitivity of the "energy storage state of charge deviation" to the "power deficit at 10:00" is being calculated. The sequence of energy storage state of charge deviation values ​​from the 1000 samples and the sequence of power deficit values ​​at 10:00 from the 1000 samples are extracted. First, the average of these two sequences is calculated. Then, for each sample, the deviation value is subtracted from the average, and multiplied by the corresponding deficit value minus the average. These 1000 products are summed and divided by 999 to obtain the final result. This negative value indicates that the more negative the energy storage state of charge deviation, the greater the power deficit, and there is a strong negative correlation between the two.

[0133] The time-series instruction vector and real-time clearing price data are multiplied element-wise and accumulated in the time domain to generate the intraday adjustment task profit.

[0134] The time-series command vector of the intraday power regulation command is aligned with the real-time clearing price data obtained from the ancillary services market interface in the time dimension. Element-wise multiplication is performed time-by-time to calculate the instantaneous theoretical revenue at each moment. Finally, combining the time step parameter, a time-domain summation operation is performed on the instantaneous theoretical revenue sequence to calculate the intraday regulation task revenue that the virtual power plant could obtain under ideal, deficit-free conditions. This calculation process is based on the following formula:

[0135] ;

[0136] in, : Represents the intraday adjustment task revenue, which is the theoretical total revenue of the virtual power plant under perfect performance conditions; : Represents the power value of the timing instruction vector at time step t; : Represents the price value of the real-time clearing price data at time step t.

[0137] An application example is as follows: Assume a daily adjustment task lasts 1 hour, divided into 4 time steps, each step lasting 15 minutes, with Δt = 0.25h. Time 1: Order 20MW, price 400 yuan / MWh. Revenue = 20 × 400 × 0.25 = 2000 yuan. Time 2: Order 20MW, price 500 yuan / MWh. Revenue = 20 × 500 × 0.25 = 2500 yuan. Assume times 3 and 4 are both 0. Then, the daily adjustment task revenue... Yuan.

[0138] Based on the perturbation sensitivity matrix, diagonal elements are extracted to construct a time-series sensitivity vector. The time-series sensitivity vector is then weighted and fused with real-time clearing price data to generate a risk premium correction sequence.

[0139] The perturbation sensitivity matrix is ​​invoked, and its main diagonal elements are extracted. These elements represent the sensitivity to fluctuations in their own state at each time step, and are arranged chronologically to construct a time-series sensitivity vector. Subsequently, this vector is used as a risk premium coefficient to perform weighted fusion on the real-time clearing price data. A premium proportional to the sensitivity is then superimposed on the original market price, thereby generating a risk premium correction sequence that implicitly incorporates risk costs.

[0140] An application example is as follows: Assume the real-time clearing price is 500 yuan / MWh at 10:00 AM. Extract the diagonal element corresponding to this moment from the disturbance sensitivity matrix, which is 0.2. Set the risk penalty weight coefficient k=1. Calculate the risk premium adjusted price: If the sensitivity coefficient is negative or zero, the original price is maintained or the minimum value is applied. This calculation is performed for all time points to generate a complete risk premium adjustment sequence.

[0141] The most extreme shortage data is multiplied and summed with the risk premium correction sequence to generate the risk option cost. The risk option cost is then divided by the intraday adjustment task revenue to calculate the percentage value.

[0142] The most extreme deficit data, i.e., the power deficit sequence under the highest risk scenario, is obtained. This sequence is then multiplied by the risk premium correction sequence at the corresponding time step, and the results are summed based on the time step. The calculated result is the risk option cost, which physically corresponds to the theoretical "option premium" or potential maximum penalty that the operator must pay to hedge against the most severe default risk. Finally, a division operation is performed to calculate the proportion of the risk option cost to the intraday adjustment task revenue. This proportion is output to the human-computer interaction interface or decision-making system as a core indicator for measuring the economic feasibility of the adjustment task. The formula for calculating the risk option cost is as follows:

[0143] ;

[0144] in, : Represents the cost of risk options; : Represents the power deficit value at the t-th time step of the most extreme deficit data; : Represents the adjusted price of the risk premium adjustment sequence at time step t. For example... Figure 3 As shown, through the multidimensional fusion of kernel density curves, box quartiles, and discrete sample points, the nonlinear clustering characteristics and extreme long-tail risk distribution of the virtual power plant regulation failure rate caused by environmental disturbances under different regulation periods are revealed.

[0145] One application example is a daily adjustment task with a profit of 4500 yuan. In the most extreme scenario, at time 1: the shortage is 5MW, and the adjusted price is 600 yuan / MWh. The potential loss = 5 × 600 × 0.25 = 750 yuan. At time 2: the shortage is 0MW, and the potential loss is 0. The risk option cost is... Yuan. Calculated percentage: Ratio = 750 / 4500 ≈ 16.67%. This figure indicates that under the most extreme operating conditions, potential risk costs will consume approximately 16.67% of the expected revenue. If the operator sets a risk threshold of 15%, the intraday adjustment task will be deemed "unreliable" or a price increase will be recommended.

[0146] The benefits of generating intraday adjustment tasks include:

[0147] The time-series instruction vector and real-time clearing price data are multiplied element-wise at each time step to generate an instantaneous value sequence;

[0148] The parsed timing instruction vector is retrieved from memory or a database, and the corresponding real-time clearing price data is simultaneously accessed through the power trading interface. These two sets of data sequences are aligned by timestamps to ensure that the power value at each discrete time step corresponds to a price value. Subsequently, element-wise multiplication is performed time-by-time, multiplying the power instruction value at each time step by the market price at that time. The resulting product sequence is the instantaneous value sequence, where each element represents the instantaneous economic value flow generated by the virtual power plant's regulatory response at the corresponding time.

[0149] The instantaneous value sequence is multiplied by the time step parameter and then summed across the entire sequence to generate the intraday adjustment task profit.

[0150] The time step parameter is read, which defines the physical time span between sampling points. Each element in the generated instantaneous value sequence is multiplied by this time step parameter, thus converting the instantaneous value flow rate into the actual monetary contribution within that time step. After the multiplication, an algebraic summation is performed on all the resulting values ​​to obtain the total amount for the entire intraday adjustment cycle, defined as the intraday adjustment task revenue. The calculation process for this intraday adjustment task revenue can be expressed by the following formula:

[0151] ;

[0152] in, : Represents the intraday adjustment task revenue, in yuan. Its physical meaning is the theoretical total economic return that the virtual power plant can obtain by completing the task under ideal conditions without considering adjustment deviation. : Represents the instruction power value corresponding to the t-th time step in the timing instruction vector, in megawatts; : Represents the market settlement price at time step t in the real-time clearing price data, in yuan / megawatt-hour, collected in real time through the ancillary services market interface; in the formula The unit is hours, representing the interval between adjacent sampling points. By discretely integrating the power command value stream over time, the dynamically changing load regulation response is quantified into static monetary value, providing benchmark data for calculating the ratio of risk cost to total revenue.

[0153] An application example is as follows: Suppose an intraday adjustment task contains two time steps, each with a time step size of 0.25 hours. In the first time step, the command power in the time-series command vector is 20 MW, and the real-time clearing price is 400 yuan / MWh, resulting in a calculated revenue of 20 × 400 × 0.25 = 2000 yuan for this period. In the second time step, the command power is 10 MW, and the market price is 500 yuan / MWh, resulting in a calculated revenue of 10 × 500 × 0.25 = 1250 yuan for this period. Through full sequence summation, the final intraday adjustment task revenue is 2000 + 1250 = 3250 yuan.

[0154] Example 2: A virtual power plant intraday regulation reliability discrimination system optimized by integrating pre-trained models, used to implement the method in Example 1, such as... Figure 4 As shown, it includes:

[0155] The stress test sample construction module is used to collect data on energy storage charge status, controllable load adjustment margin and communication latency by aggregating resources in a virtual power plant, load historical operation deviation logs to construct deviation probability distribution, perform random sampling and element splicing, and construct a stress test sample set.

[0156] The timing instruction inference output module is used to acquire intraday power regulation instructions and map them into timing instruction vectors. The timing instruction vectors and the stress test sample set are input into the preset timing pre-trained model to perform forward inference and output energy storage charging and discharging power sequences, controllable load response sequences and distributed power output sequences for different stress scenarios.

[0157] The feedback power trajectory generation module is used to physically superimpose the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence through the topological connection relationship of the virtual power plant to generate the feedback power trajectory.

[0158] The most extreme deficit screening module is used to perform differential operation on the timing command vector and the feedback power trajectory, extract the positive deviation part as the power deficit sequence, and perform accumulation on the power deficit sequence through the stress test sample set, and select the sequence with the largest accumulated value as the most extreme deficit data.

[0159] The risk option cost calculation module is used to access real-time clearing price data through the ancillary service market, combine the most extreme shortage data with the real-time clearing price data to perform risk cost calculation, obtain the risk option cost, and output the percentage of risk option cost in the intraday adjustment task revenue.

[0160] The functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between the various modules represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The implementation details of each module are the same as in Embodiment 1. The above are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for determining the intraday regulation reliability of a virtual power plant by integrating pre-trained model optimization, characterized by the following steps: include: The system utilizes virtual power plants to aggregate resources to collect data on energy storage status of charge, controllable load adjustment margin, and communication latency. It also loads historical operation deviation logs to construct a deviation probability distribution, performs random sampling and element splicing, and builds a stress test sample set. The system acquires intraday power regulation commands and maps them to timing command vectors. It then inputs the timing command vectors and the set of stress test samples into a preset timing pre-trained model, performs forward inference, and outputs energy storage charging and discharging power sequences, controllable load response sequences, and distributed power output sequences for different stress scenarios. By leveraging the topological connections of the virtual power plant, the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence are physically superimposed to generate a feedback power trajectory. The timing command vector and the feedback power trajectory are differentially processed to extract the positive deviation as the power deficit sequence. The power deficit sequence is then accumulated using a stress test sample set, and the sequence with the largest accumulated value is selected as the most extreme deficit data. By accessing real-time clearing price data through the ancillary services market, and combining the most extreme shortage data with the real-time clearing price data, risk cost accounting is performed to obtain the risk option cost, and the percentage of the risk option cost in the intraday adjustment task revenue is output.

2. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization as described in claim 1, characterized in that, Building the stress test sample set includes: The system reads the energy storage charge status by calling the status register through the Internet of Things (IoT) gateway, reads the controllable load adjustment margin through the building temperature control system, sends heartbeat packets to the terminal equipment and records the round-trip time to generate communication latency data. Based on historical operational deviation logs, the energy storage state of charge deviation, controllable load adjustment margin deviation, and communication delay deviation are extracted. The bin counting is performed and the cumulative frequency is calculated to generate a deviation probability distribution. Random sampling is performed based on the deviation probability distribution to generate a disturbance noise vector. The disturbance noise vector, energy storage state of charge, controllable load adjustment margin, and communication delay data are then concatenated element by element to generate a stress test sample set.

3. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization as described in claim 1, characterized in that, The output sequences of energy storage charging and discharging power, controllable load response, and distributed power generation for different pressure scenarios include: The power value and start and end timestamps of the intraday power adjustment command are extracted by parsing the command. The power value is normalized and the start and end timestamps are mapped to time feature vectors. The vectors are then concatenated to generate a time-series command vector. The timing instruction vector and the stress test sample set are concatenated along the feature dimension to construct a joint input tensor; The time-series pre-trained model is invoked to perform matrix multiplication and activation operations on the joint input tensor to generate a high-dimensional feature tensor. Channel dimension slicing is then performed on the high-dimensional feature tensor to extract the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence, respectively.

4. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization as described in claim 1, characterized in that, The output sequences for energy storage charging and discharging power, controllable load response, and distributed power generation for different pressure scenarios also include: The system retrieves intraday power regulation commands from historical time periods and simultaneously extracts data on energy storage state of charge, controllable load regulation margin, and communication latency from historical time periods. It then performs time alignment and feature splicing to construct a historical training input tensor. Based on the actual power response records of virtual power plant aggregated resources obtained from historical time periods, data cleaning and channel dimension stacking are performed to construct a multi-dimensional measured response matrix. Using the historical training input tensor as input and the multidimensional measured response matrix as the regression objective, a temporal regression network is established and gradient descent training is performed to minimize the prediction residual and generate a temporal pre-trained model.

5. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization as described in claim 1, characterized in that, The generated feedback power trajectory includes: Based on the electrical topology of the virtual power plant, the output sequence of distributed power sources and the discharge power of energy storage are marked as positive supply components, and the response sequence of controllable loads and the charging power of energy storage are marked as negative demand components, thus establishing the power flow direction identifier. Based on the energy storage state of charge, the energy storage charging and discharging power sequence is accumulated hourly to identify the time points when the energy is depleted or fully charged, and the power value is forced to zero based on the time points to generate an effective energy storage power sequence. The effective energy storage power sequence, the distributed power output sequence, and the controllable load response sequence are algebraically added according to the power flow direction identifier to generate a feedback power trajectory.

6. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization according to claim 2, characterized in that, The sequence with the largest cumulative value is selected as the most extreme shortage data, including: The timing command vector and the feedback power trajectory are subtracted at each time step to generate the original deviation trajectory. If the elements in the original deviation trajectory with values ​​less than zero are forcibly set to zero, a power deficit sequence is generated. For intraday adjustment reading of time step parameters, the power deficit sequence and time step parameters are summed over the entire sequence to calculate the cumulative power deficit. Based on the stress test sample set, the cumulative power deficit is sorted by value, the index of the scenario with the largest value is locked, and the power deficit sequence is extracted as the most extreme deficit data.

7. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization as described in claim 6, characterized in that, The percentage of risk option costs in the intraday adjustment task revenue includes: Based on the disturbance noise vector and power deficit sequence, vector covariance calculation is performed to quantify the impact of disturbance amplitude on the deficit and generate a disturbance sensitivity matrix. The time-series instruction vector and real-time clearing price data are multiplied element-wise and accumulated in the time domain to generate the intraday adjustment task profit. Based on the perturbation sensitivity matrix, diagonal elements are extracted to construct a time-series sensitivity vector. The time-series sensitivity vector is then weighted and fused with real-time clearing price data to generate a risk premium correction sequence. The most extreme shortage data is multiplied and summed with the risk premium correction sequence to generate the risk option cost. The risk option cost is then divided by the intraday adjustment task revenue to calculate the percentage value.

8. The method for determining the intraday regulation reliability of a virtual power plant based on the fusion of pre-trained model optimization as described in claim 7, characterized in that, The benefits of generating intraday adjustment tasks include: The time-series instruction vector and real-time clearing price data are multiplied element-wise at each time step to generate an instantaneous value sequence; The instantaneous value sequence is multiplied by the time step parameter and then summed across the entire sequence to generate the intraday adjustment task profit.

9. A virtual power plant intraday regulation reliability discrimination system based on pre-trained model optimization, used to implement the virtual power plant intraday regulation reliability discrimination method based on pre-trained model optimization as described in any one of claims 1-8, characterized in that, include: The stress test sample construction module is used to collect data on energy storage charge status, controllable load adjustment margin and communication latency by aggregating resources in a virtual power plant, load historical operation deviation logs to construct deviation probability distribution, perform random sampling and element splicing, and construct a stress test sample set. The timing instruction inference output module is used to acquire intraday power regulation instructions and map them into timing instruction vectors. The timing instruction vectors and the stress test sample set are input into the preset timing pre-trained model to perform forward inference and output energy storage charging and discharging power sequences, controllable load response sequences and distributed power output sequences for different stress scenarios. The feedback power trajectory generation module is used to physically superimpose the energy storage charging and discharging power sequence, the controllable load response sequence, and the distributed power output sequence through the topological connection relationship of the virtual power plant to generate the feedback power trajectory. The most extreme deficit screening module is used to perform differential operation on the timing command vector and the feedback power trajectory, extract the positive deviation part as the power deficit sequence, and perform accumulation on the power deficit sequence through the stress test sample set, and select the sequence with the largest accumulated value as the most extreme deficit data. The risk option cost calculation module is used to access real-time clearing price data through the ancillary service market, combine the most extreme shortage data with the real-time clearing price data to perform risk cost calculation, obtain the risk option cost, and output the percentage of risk option cost in the intraday adjustment task revenue.

Images