Virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data

Abstract

The application is a virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data, belonging to the technical field of virtual power plants, which acquires structured, semi-structured and unstructured data of a virtual power plant; constructs a time sequence dynamic spectrum based on structured data, uses a BERT model to extract entity relationship in unstructured data to enhance the time sequence dynamic spectrum, and obtains an entity relationship enhanced spectrum; constructs a multi-modal performance scoring model including an XGBoost scheduling deviation scoring module, a BERT instruction response risk scoring module and a neighborhood coupling influence module, and outputs a scheduling performance risk score; uses an LSTM to perform time sequence analysis on the risk score, calculates a scheduling deviation propagation index based on the entity relationship enhanced spectrum and an attention mechanism, and performs graded early warning if the index exceeds a preset threshold. The application realizes quantitative evaluation and dynamic early warning of scheduling performance risk of a virtual power plant through the above steps.

Description

Technical Field

[0001] This invention relates to the field of virtual power plant technology, and mainly to a method for early warning of scheduling performance risks in virtual power plants based on multi-source heterogeneous data. Background Technology

[0002] Virtual power plants participate in grid dispatch by aggregating distributed energy resources, energy storage systems, and controllable loads. The reliability of their dispatch performance directly affects the stable operation of the electricity market. Due to the diverse types of units, complex response behaviors, and mutual coupling within virtual power plants, accurately assessing the dispatch performance risks of each unit and achieving early warning has become a key issue in the operation and management of virtual power plants.

[0003] For example, Chinese invention patent application CN117591584A discloses a virtual power plant optimization management and scheduling method based on multi-source heterogeneous data. This technical solution first uses a BP neural network to fuse and evaluate the accuracy of the acquired multi-source heterogeneous data (using the coefficient of determination). The system uses the mean square error (MSE) as an evaluation metric. Based on the fused data, a support vector model is used to extract energy flow bands and generate flow band vectors. Then, a preset safety vector template is used to identify vector deviations, ultimately generating an anomaly report and outputting it to the electronic control unit. However, the BP neural network fusion process in the above technical solution aims to minimize overall data error, and its evaluation metrics... The MSE (Mean Safety Estimate) reflects the fitting accuracy at the data level, but it does not construct a dedicated risk characterization model for the scheduling performance risks of virtual power plants. As a result, the fused data cannot effectively distinguish the differences in performance capabilities among different units. Secondly, this technical solution relies on a preset safety vector template to identify deviations. This template is statically set, and when the coupling relationship between units within the virtual power plant or the characteristics of scheduling instructions change, the matching accuracy of the template decreases significantly, easily leading to false alarms or missed alarms. In addition, the anomaly report generation mechanism of this technical solution is based on the deviation analysis of a single flow band vector, without considering the risk propagation effect caused by scheduling interactions between units. It cannot quantify how the performance deviation of one unit affects other units through correlation, thus making it difficult to achieve systematic risk early warning.

[0004] In summary, there is an urgent need for a virtual power plant scheduling performance risk early warning method that can integrate unstructured scheduling semantics, quantify risk propagation effects, and achieve hierarchical early warning. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this application provides a method for early warning of virtual power plant dispatch performance risks based on multi-source heterogeneous data.

[0006] The technical solution of this application is as follows: On the one hand, the present invention provides a method for early warning of virtual power plant scheduling performance risks based on multi-source heterogeneous data, the method comprising: Acquire multi-source heterogeneous data from the virtual power plant, including structured data, semi-structured data, and unstructured data; A time-series dynamic spectrogram is constructed based on structured data; based on unstructured data, the entity relationship of the time-series dynamic spectrogram is enhanced using the BERT model to obtain an entity relationship-enhanced spectrogram. A multimodal performance scoring model is constructed, which takes semi-structured data, unstructured data and entity relationship enhanced spectrum as input, and outputs the scheduling performance risk score of the virtual power plant. The dispatch deviation propagation index of the virtual power plant is calculated based on the entity relationship enhancement spectrum and the dispatch performance risk score of the virtual power plant. If the dispatch deviation propagation index exceeds the preset risk threshold, a risk warning is issued.

[0007] Preferably, the structured data includes scheduling time, scheduling instruction response value, and device type in the scheduling record; the semi-structured data includes the response rate and adjustment accuracy of the device status report; and the unstructured data includes the scheduling instruction text.

[0008] Preferably, constructing a time-series dynamic spectrum based on structured data includes defining the time-series dynamic spectrum as a weighted directed graph, where nodes are internal units of a virtual power plant, including energy storage units, power generation units, and load units, and edges represent scheduling interaction relationships. The specific steps for calculating the weight of each edge are as follows: Obtain the attenuation factor corresponding to the equipment type of the starting unit, where the attenuation factors of the energy storage unit, power generation unit, and load unit decrease sequentially; calculate the time interval between the current time and the time of the most recent dispatch command, and calculate the exponential attenuation value with the natural constant as the base and the negative time attenuation coefficient multiplied by the time interval as the exponent, where the time attenuation coefficient is a preset constant; obtain the dispatch command response value from the starting unit to the target unit, divide the dispatch command response value by the maximum dispatch command response value among all dispatch records to obtain the normalized response ratio; multiply the attenuation factor, the exponential attenuation value, and the normalized response ratio together, and the product is the weight of the corresponding edge.

[0009] Preferably, based on unstructured data, entity relationship enhancement is performed on the time-series dynamic spectrum using the BERT model. Specifically, the entity relationships in the scheduling instruction text are extracted using the BERT model, triples are generated, and the triples are injected into the time-series dynamic spectrum to obtain an entity relationship-enhanced spectrum. Preferably, the multimodal performance evaluation model includes an XGBoost scheduling deviation evaluation module, a BERT instruction response risk evaluation module, and a neighborhood coupling effect module, wherein: The XGBoost scheduling deviation scoring module is used to predict the scheduling command response deviation of the internal units of the virtual power plant. It takes semi-structured data as input data and outputs the XGBoost scheduling deviation score. The BERT instruction response risk scoring module is used to predict the degree of risk of scheduling instruction execution in the internal units of the virtual power plant. It takes unstructured data as input and outputs a BERT instruction response risk score. The neighborhood coupling effect module is used to evaluate the degree of mutual influence between units within the virtual power plant. It takes the entity relationship enhancement spectrum as input and outputs a neighborhood coupling effect score. The scheduling performance risk score of the virtual power plant is calculated based on the XGBoost scheduling deviation score, the BERT instruction response risk score, and the neighborhood coupling impact score.

[0010] Preferably, the dispatch deviation propagation index of the virtual power plant is calculated based on the entity relationship enhancement spectrum and the dispatch performance risk score of the virtual power plant, specifically as follows: A Long Short-Term Memory (LSTM) network is used to perform time-series analysis on the historical scheduling performance risk scores of each virtual power plant internal unit, and a fixed-dimensional LSTM time-series feature vector is extracted. For the target unit, select the two neighboring units with the largest edge weights in the entity relationship enhancement spectrum, obtain the LSTM temporal feature vectors of each of the three units, and horizontally concatenate the three vectors into a matrix. A trainable attention parameter vector is transposed and multiplied by the matrix to obtain an intermediate vector; the intermediate vector is multiplied by a trainable matrix to obtain a scalar value; the scalar value is divided by the square root of the LSTM feature vector dimension and then mapped by the Sigmoid function to obtain the scheduling bias propagation index, which has a value range of 0 to 1.

[0011] Preferably, the preset risk threshold includes a first threshold and a second threshold, and the first threshold is less than the second threshold; When the scheduling deviation propagation index exceeds the first threshold but does not exceed the second threshold, the reinforcement learning decision engine is triggered to make risk prevention and control decisions; when the scheduling deviation propagation index exceeds the second threshold, the reinforcement learning decision engine is triggered to take corresponding intervention measures. The reinforcement learning decision engine dynamically selects intervention actions based on the current system state.

[0012] Preferably, the method further includes introducing a reinforcement learning decision engine, which optimizes decisions based on Markov decision processes, specifically: Define the state space, including the current margin increase ratio and the virtual power plant power deviation rate; Define the action space, including increasing the margin, restricting scheduling permissions, and suspending scheduling eligibility; Define the state transition probability and the reward function, wherein the reward function comprehensively considers power deviation, intervention cost and power fluctuation; At each decision-making moment, the reinforcement learning decision engine selects the optimal action based on the current state, executes it, observes the next state and reward, and updates the decision parameters.

[0013] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data as described in any one of the present invention.

[0014] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data as described in any one of the present invention.

[0015] The present invention has the following beneficial effects: 1. This invention provides a method for early warning of scheduling performance risks in virtual power plants based on multi-source heterogeneous data. It constructs a time-series dynamic spectrum based on structured scheduling data and uses the BERT model to extract entity relationships from unstructured scheduling instruction text to enhance the spectrum, thereby obtaining an entity relationship-enhanced spectrum that includes deep relationships such as inter-unit reserve associations. This improves the accuracy of characterizing the coupling relationships between units within the virtual power plant and the completeness of risk association information. 2. This invention provides a method for early warning of virtual power plant dispatch performance risk based on multi-source heterogeneous data. By constructing a multimodal performance scoring model that includes an XGBoost dispatch deviation scoring module, a BERT command response risk scoring module, and a neighborhood coupling influence module, the method integrates semi-structured response rate and regulation accuracy, unstructured command text, and entity relationship enhanced spectrum to calculate dispatch performance risk score, thereby improving the multi-source information fusion capability and the accuracy of individual risk characterization in virtual power plant dispatch performance risk scoring. 3. This invention provides a virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data. It uses LSTM to perform time series analysis on the scheduling performance risk score of each unit and extracts time series feature vectors. It calculates the scheduling deviation propagation index based on entity relationship enhancement spectrum and graph attention mechanism, and introduces reinforcement learning decision engine to dynamically optimize hierarchical early warning actions, thereby improving the quantification ability of risk propagation effect and the adaptability of early warning decision to dynamic environment. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method corresponding to an embodiment of the present invention. Detailed Implementation

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

[0018] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0019] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0020] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0021] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0022] Example 1: This embodiment provides a method for early warning of virtual power plant dispatch performance risks based on multi-source heterogeneous data, such as... Figure 1 As shown, the method includes: S1. Obtain multi-source heterogeneous data from the virtual power plant, specifically: S11. In this embodiment, the internal units of the virtual power plant include energy storage units, power generation units, and load units, and the multi-source heterogeneous data includes structured data, semi-structured data, and unstructured data. Construct a comprehensive data acquisition network covering energy storage units, power generation units, and load units. Capture structured data such as scheduling records (including scheduling time, scheduling command response values, and equipment type) in real time through API interfaces; use ETL tools to batch extract semi-structured data such as equipment status reports (involving response rate and adjustment accuracy); and manually supplement unstructured data such as scheduling command text. Meanwhile, to ensure data integrity, this embodiment arranges for manual collection of some data that is difficult to obtain through automated means; S12. Perform preprocessing operations such as cleaning, transformation, and integration on the multi-source heterogeneous data obtained in step S11, and classify and organize it according to different sources to construct a dimensional model or relational model. Preferably, this embodiment uses the OpenRefine tool to build an automated cleaning pipeline. The specific processing steps are as follows: S121. Using big data analysis technology and data cleaning tools, deduplication and format standardization are performed on multi-source heterogeneous data. Preferably, the response data in the equipment status report is divided into different levels of excellent, good, medium and poor, and converted into a numerical form that can be processed by a computer using one-hot encoding. S122. Outliers in multi-source heterogeneous data are detected using the 3σ principle and box plot method. Simultaneously, Min-Max and Z-score dual-mode normalization is employed to normalize multi-source heterogeneous data of different dimensions to the same order of magnitude. Preferably, for data gaps resulting from outlier removal during preprocessing, the average of the same type of unit over the past three hours is used to fill the gaps, ensuring the continuity of the time-series data. The normalization operation strictly follows the principle of training set fitting and test set reuse to avoid data leakage affecting the model's generalization ability. Furthermore, this embodiment uses an actual power deviation exceeding 5% as the basis for labeling risky positive samples, and vice versa for negative samples. All data is labeled according to a 15-minute scheduling cycle. Since steps S121 and S122 are conventional data preprocessing techniques in this field, they will not be elaborated upon here. S13. Construct a time-series dynamic spectrum based on the structured data processed in step S12. Specifically: Define a time-series dynamic spectrum as a weighted directed graph. ,in, For a set of nodes, For edge set, The set is a weight set; the internal units of the virtual power plant are used as nodes. In this implementation, As an energy storage unit, For power generation units, As a load unit; the scheduling interaction relationship is taken as an edge, and the weight of the edge is expressed by the formula: ; In the formula, In time At that time, from the internal unit of the virtual power plant To the internal unit of the virtual power plant The weight of the edge; The attenuation factor is related to the device type and is determined by the starting unit. The value is determined by the type of equipment and ranges from 0 to 1. In this embodiment, the starting unit corresponds to the energy storage unit with a value of 0.9, the corresponding power generation unit with a value of 0.8, and the corresponding load unit with a value of 0.7. The time decay coefficient is fixed at 0.1 per hour in this embodiment; Current time The interval between the most recent scheduling instruction and the last time it was issued; Internal units of a virtual power plant With virtual power plant internal units The response value of the dispatch instruction between the two virtual power plants reflects the response scale of the internal units of the two virtual power plants in a single dispatch. The maximum scheduling instruction response value among all scheduling records, used to... Perform normalization processing; It is a natural constant; To enrich the relationship information between entities in the graph, this embodiment also uses the BERT model to enhance entity relationships in the time-series dynamic spectrum based on unstructured data (i.e., scheduling instruction text), obtaining an entity relationship-enhanced spectrum. Entities in this embodiment cover objects directly related to scheduling performance, such as energy storage units, power generation units, load units, scheduling instructions, power, time, and standby units. Relationships include types that reflect the interaction logic between units, such as scheduling, standby, constraint, and association. The scheduling instruction text is processed according to semantic sentence segmentation, ensuring that the length of a single sentence does not exceed the preset sentence length to adapt to the model input requirements. Triples are extracted using a Chinese BERT-base model combined with an entity relationship classification head, retaining only triples with a confidence level of not less than 0.8. Triples are output in a standard format of head entity, relation, and tail entity, for example, energy storage unit A, standby, and power generation unit B. The extracted triples are then injected into the time-series dynamic spectrum according to preset rules. Specifically: If the triple reflects a backup relationship, then add a backup relationship edge between the two corresponding cells and initialize the weight to 0.6; If the triple reflects a scheduling relationship, check if there is an existing scheduling edge between the two units: if there is, increase the weight of the original scheduling edge by 10%, but not exceeding the upper limit of 1.0; if there is no scheduling edge, add a new scheduling edge and assign an initial weight of 0.4 based on semantic confidence. The entire graph is updated every fifteen minutes in sync with the scheduling instructions, ultimately forming an entity relationship-enhanced graph that includes both scheduling interaction relationships and semantic association information.

[0023] S2. Construct a multimodal performance evaluation model, which includes an XGBoost scheduling deviation evaluation module, a BERT instruction response risk evaluation module, and a neighborhood coupling effect module. It takes semi-structured data, unstructured data, and entity relationship enhanced spectrograms as input and outputs a scheduling performance risk score for the virtual power plant. Specifically: The S21 and XGBoost scheduling deviation scoring modules use semi-structured data as input data, and the corresponding input features are response rate and regulation accuracy. The goal is to predict the scheduling command response deviation of the internal units of the virtual power plant. Furthermore, in this embodiment, the XGBoost scheduling bias scoring module includes an XGBoost model composed of multiple weak classifiers (decision trees) sequentially. Each decision tree is trained based on the residual of the previous decision tree. Specifically, the first... The training objective of the trees is to fit the prefit. The residual between the predicted value and the actual value of each tree; The number of semi-structured data samples in this embodiment is set to 1. , No. The input feature vector of each sample is The corresponding actual scheduling response deviation label is XGBoost predictions Expressed as a formula: ; In the formula, It is the number of decision trees. It is the first The predicted values ​​of each decision tree for a sample; During training, XGBoost learns the structure of the decision tree by minimizing an objective function. Expressed as a formula: ; In the formula, It is the set of all decision trees; It is the loss function; in this embodiment, mean squared error is used. This is a regularization term used to control the complexity of the decision tree and avoid overfitting. It is defined as: ; In the formula, It is the first The number of leaf nodes in a decision tree; It is the first The first decision tree The weight of each leaf node, and It is a regularization parameter. Take 0.1, Set the value to 1.0; During prediction, for the input feature vector, each decision tree is traversed, and the corresponding leaf node is found based on the path of the feature value in the decision tree. The weights of all leaf nodes are summed to obtain the final predicted value. In this embodiment, the key parameters of the XGBoost model were determined through multiple rounds of verification and optimization. Specifically: The maximum depth of a single decision tree is 6, and the learning rate is set to 0.1. In addition to response rate and adjustment accuracy, the model input also includes the three most recent historical deviation data as auxiliary features. The model output is mapped to the 0-1 interval by the Sigmoid function to form a standardized scheduling deviation score. ; The input to the S22 and BERT instruction response risk scoring modules is unstructured data, namely the scheduling instruction text. The input scheduling instruction text is segmented to obtain word sequence embeddings. Location embedding and segment embedding The transformed results are then summed to obtain the input representation. ; Will The input is fed into multiple Transformer encoder layers of the BERT model, and then... After encoding the layers, the output of the last layer is obtained. ,Pick The output vector corresponding to the CLS marker. ;Will The input is fed into a fully connected layer for nonlinear transformation to obtain the transformed vector. Expressed as a formula: ; In the formula, It is the weight matrix of the fully connected layer. It is the bias vector; Risk score of instruction response obtained through linear layer. Expressed as a formula: ; In the formula, It is the weight matrix of the linear layer. It is the bias vector; Preferably, in the training process of the BERT model in this embodiment, loss functions such as cross-entropy loss are used to optimize the model parameters; this embodiment uses a Chinese BERT-base pre-trained model for fine-tuning and adaptation, with the batch size set to 16, the training epochs being 5, and the learning rate controlled at [value missing]. At the same time, a dropout ratio of 0.1 is set to prevent the model from overfitting; S23. Enhanced spectrogram based on entity relationships for virtual power plant internal units. The neighborhood coupling effect module considers its neighboring nodes. The impact, Indicates the internal unit of the virtual power plant The set of other internal units with scheduling interaction relationships (edge ​​weights greater than 0.1), and the weights of the edges. The formula, derived from the time-series dynamic spectrum, reflects the time... Virtual power plant internal unit and The strength of the scheduling interaction relationship between them; Meanwhile, the neighborhood coupling impact module obtains the scheduling performance risk score of the neighborhood nodes at the previous moment. The weight of the edge Scheduling fulfillment risk score with neighboring nodes Multiply the results and sum them over all neighboring nodes, then normalize the results to obtain the neighborhood coupling influence score. ; S24, Combining steps S21-S23, the multimodal performance scoring model for the internal units of the virtual power plant. In time Time-based scheduling performance risk scoring Expressed as a formula: ; In the formula, and For trainable weights, and In this embodiment, the initial values ​​are set to 0.3, 0.4, and 0.3 respectively; the above trainable weights can be optimized through the training process to balance the contributions of different modules in the scheduling performance risk score. S25. Furthermore, in order to determine trainable weights and To find the optimal value, a large amount of historical data from internal units of the virtual power plant was collected, including semi-structured data, unstructured data, and scheduling records. This data was divided into training, validation, and test sets. During training, optimization algorithms such as stochastic gradient descent were used to continuously adjust the model, aiming to minimize the error between the model's predicted scheduling performance risk score and the actual scheduling performance. and The value of is used to select the optimal weight combination by evaluating the model's performance on the validation set; S26. Apply the trained multimodal performance evaluation model to the actual virtual power plant dispatch performance evaluation. For new virtual power plant internal units or new dispatch data, input the corresponding semi-structured data, unstructured data and entity relationship enhancement spectrum into the multimodal performance evaluation model to obtain the dispatch performance risk score of the virtual power plant internal unit.

[0024] S3. After calculating the scheduling performance risk score for each internal unit of the virtual power plant, quantitative indicators of the scheduling performance status of each internal unit at different time points are obtained. In this embodiment, a Long Short-Term Memory (LSTM) network is used to perform time-series analysis on the scheduling performance risk score of each internal unit to obtain the corresponding LSTM time-series feature vector. Specifically: In this embodiment, the LSTM model uses the historical risk score sequence of a single virtual power plant internal unit as input. The model adopts a single-layer structure design with a hidden layer dimension of 64. A dropout ratio of 0.2 is set to improve the robustness of the model. After the LSTM model is trained, the hidden layer state at the last moment is taken as the temporal feature vector. The dimension of this feature vector is fixed at 64. It should be noted that the training of the LSTM model is carried out independently based on the generated historical risk score sequence after the multimodal performance scoring model is trained. It belongs to a two-stage training process. For virtual power plant internal units Its time Time-based scheduling deviation propagation index Expressed as a formula: ; In the formula, The Sigmoid activation function maps the computation result to... interval; , and These are internal units of a virtual power plant. , and LSTM temporal feature vectors; For attention parameters; To perform the transpose operation; It is a trainable matrix; To promptly detect and address credit risks of electricity market participants, this embodiment establishes tiered early warning thresholds. The preset risk thresholds include a first threshold and a second threshold, with the first threshold being lower than the second threshold. Specifically, in this embodiment, the first threshold is an orange warning threshold. ,when At this time, it indicates the internal unit of the virtual power plant. Facing a certain degree of scheduling and contract fulfillment risk, corresponding risk prevention and control measures need to be taken; the second threshold is the red warning threshold. ,when At that time, explain the internal units of the virtual power plant. Faced with high scheduling and contract fulfillment risks, more stringent risk control measures need to be taken immediately. It is worth noting that the aforementioned tiered early warning thresholds are not subjectively set, but rather determined by plotting the Receiver Operating Characteristic (ROC) curve based on actual operational data from the past three months. This effectively reduces the probability of false alarms while ensuring the accuracy of the early warnings. The early warning triggering cycle is consistent with the scheduling cycle, both being 15 minutes, to ensure the timeliness of risk detection.

[0025] S4. To achieve smarter and more effective risk prevention and control decisions, this embodiment introduces a reinforcement learning decision engine, which optimizes decisions based on Markov decision processes. Specifically: Define state ,in, The current margin increase ratio (in this example, the value ranges from 0 to 0.4). The power deviation rate of the virtual power plant. ,in, This is the actual power. For planned power, The value is the rated capacity, and is truncated to [0,1]. Define Action {Increase performance bond, restrict scheduling authority, suspend scheduling eligibility} indicates the actions that can be taken at each decision point. Specifically, increasing the performance bond means increasing the performance bond ratio of the target unit. ,in To adjust the coefficient upwards, only when Execution is only permitted when an orange alert is triggered, and Truncate to [0, 0.4]; the restriction on scheduling permissions specifically involves adjusting the scheduling permission level of associated nodes that have scheduling interaction relationships with the target unit to [0, 0.4]. ,in The original permission level (initially 1, with a value range of [0,1]). The minimum retention is 0.2 (i.e., 20% of basic scheduling authority); the suspension of scheduling eligibility specifically means directly canceling the target unit's scheduling eligibility in the current scheduling cycle. This action is only permitted during a red alert ( It must be executed at any time and undergo security verification; Define state transition probability The state transition probability Describes the current state Take action below Then transition to the next state The probability of; Define reward function Used to measure the current state Take action below The goal of the reward is to maximize the value of the reward function, that is, to minimize scheduling deviation loss, intervention cost, and power fluctuation. Furthermore, The absolute value of the power deviation within the current scheduling period The historical maximum value is normalized to [0,1]; the intervention cost is the cost of the intervention measures. According to the preset action type, in this embodiment, the cost coefficient for increasing the margin is 0.2, the cost coefficient for restricting scheduling authority is 0.5, and the cost coefficient for suspending scheduling qualification is 1.0; the power fluctuation rate is the standard deviation of the net load power in the most recent 3 scheduling cycles (45 minutes), which is also normalized to [0,1]. , and Take values ​​of 0.5, 0.3, and 0.2 respectively; By each decision moment According to the current state The reinforcement learning algorithm (such as Q-learning, DeepQ-Network, etc.) is used to select the optimal action, execute the optimal action, observe the next state and reward from the environment, and update the parameters of the reinforcement learning algorithm based on the observed reward.

[0026] Example 2: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data as described in any embodiment of the present invention.

[0027] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data as described in any embodiment of the present invention.

[0028] It is worth noting that the electronic device and computer-readable storage medium described in this invention are based on the same inventive concept as the method described in Embodiment 1, and will not be repeated here.

[0029] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

[0030] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM, optical memory, etc.

[0031] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for early warning of virtual power plant dispatch performance risks based on multi-source heterogeneous data, characterized in that, The method includes: Acquire multi-source heterogeneous data from the virtual power plant, including structured data, semi-structured data, and unstructured data; A time-series dynamic spectrogram is constructed based on structured data; based on unstructured data, the entity relationship of the time-series dynamic spectrogram is enhanced using the BERT model to obtain an entity relationship-enhanced spectrogram. A multimodal performance scoring model is constructed, which takes semi-structured data, unstructured data and entity relationship enhanced spectrum as input, and outputs the scheduling performance risk score of the virtual power plant. The dispatch deviation propagation index of the virtual power plant is calculated based on the entity relationship enhancement spectrum and the dispatch performance risk score of the virtual power plant. If the dispatch deviation propagation index exceeds the preset risk threshold, a risk warning is issued.

2. The virtual power plant dispatch performance risk early warning method based on multi-source heterogeneous data according to claim 1, characterized in that, The structured data includes scheduling time, scheduling instruction response value, and equipment type in the scheduling record; the semi-structured data includes the response rate and adjustment accuracy of the equipment status report; and the unstructured data includes the scheduling instruction text.

3. The virtual power plant dispatching performance risk early warning method based on multi-source heterogeneous data according to claim 2, characterized in that, Constructing a time-series dynamic spectrum graph based on structured data includes defining the time-series dynamic spectrum graph as a weighted directed graph, where nodes are internal units of a virtual power plant, including energy storage units, generation units, and load units, and edges represent scheduling interaction relationships. The specific steps for calculating the weight of each edge are as follows: Obtain the attenuation factor corresponding to the equipment type of the starting unit, where the attenuation factors of the energy storage unit, power generation unit, and load unit decrease sequentially; calculate the time interval between the current time and the time of the most recent dispatch command, and calculate the exponential attenuation value with the natural constant as the base and the negative time attenuation coefficient multiplied by the time interval as the exponent, where the time attenuation coefficient is a preset constant; obtain the dispatch command response value from the starting unit to the target unit, divide the dispatch command response value by the maximum dispatch command response value among all dispatch records to obtain the normalized response ratio; multiply the attenuation factor, the exponential attenuation value, and the normalized response ratio together, and the product is the weight of the corresponding edge.

4. The virtual power plant dispatch performance risk early warning method based on multi-source heterogeneous data according to claim 3, characterized in that, Based on unstructured data, the BERT model is used to enhance entity relationships in the time-series dynamic spectrum. Specifically, the BERT model is used to extract entity relationships from the scheduling instruction text, generate triples, and inject the triples into the time-series dynamic spectrum to obtain an entity relationship-enhanced spectrum.

5. The virtual power plant dispatch performance risk early warning method based on multi-source heterogeneous data according to claim 4, characterized in that, The multimodal performance evaluation model includes an XGBoost scheduling deviation evaluation module, a BERT instruction response risk evaluation module, and a neighborhood coupling effect module, wherein: The XGBoost scheduling deviation scoring module is used to predict the scheduling command response deviation of the internal units of the virtual power plant. It takes semi-structured data as input data and outputs the XGBoost scheduling deviation score. The BERT instruction response risk scoring module is used to predict the degree of risk of scheduling instruction execution in the internal units of the virtual power plant. It takes unstructured data as input and outputs a BERT instruction response risk score. The neighborhood coupling effect module is used to evaluate the degree of mutual influence between units within the virtual power plant. It takes the entity relationship enhancement spectrum as input and outputs a neighborhood coupling effect score. The scheduling performance risk score of the virtual power plant is calculated based on the XGBoost scheduling deviation score, the BERT instruction response risk score, and the neighborhood coupling impact score.

6. The virtual power plant dispatch performance risk early warning method based on multi-source heterogeneous data according to claim 1, characterized in that, The dispatch deviation propagation index of the virtual power plant is calculated based on the entity relationship enhancement spectrum and the dispatch performance risk score of the virtual power plant. Specifically: A Long Short-Term Memory (LSTM) network is used to perform time-series analysis on the historical scheduling performance risk scores of each virtual power plant internal unit, and a fixed-dimensional LSTM time-series feature vector is extracted. For the target unit, select the two neighboring units with the largest edge weights in the entity relationship enhancement spectrum, obtain the LSTM temporal feature vectors of each of the three units, and horizontally concatenate the three vectors into a matrix. A trainable attention parameter vector is transposed and multiplied by the matrix to obtain an intermediate vector; the intermediate vector is then multiplied by a trainable matrix to obtain a scalar value. After dividing the scalar value by the square root of the LSTM feature vector dimension and mapping it through the Sigmoid function, the scheduling deviation propagation index, which ranges from 0 to 1, is obtained.

7. The virtual power plant dispatch performance risk early warning method based on multi-source heterogeneous data according to claim 1, characterized in that, The preset risk threshold includes a first threshold and a second threshold, and the first threshold is less than the second threshold; When the scheduling deviation propagation index exceeds the first threshold but does not exceed the second threshold, the reinforcement learning decision engine is triggered to make risk prevention and control decisions. When the scheduling deviation propagation exponent exceeds the second threshold, the reinforcement learning decision engine is triggered and corresponding intervention measures are taken. The reinforcement learning decision engine dynamically selects intervention actions based on the current system state.

8. The virtual power plant dispatch performance risk early warning method based on multi-source heterogeneous data according to claim 1, characterized in that, The method further includes introducing a reinforcement learning decision engine, which optimizes decisions based on Markov decision processes, specifically as follows: Define the state space, including the current margin increase ratio and the virtual power plant power deviation rate; Define the action space, including increasing the margin, restricting scheduling permissions, and suspending scheduling eligibility; Define the state transition probability and the reward function, wherein the reward function comprehensively considers power deviation, intervention cost and power fluctuation; At each decision-making moment, the reinforcement learning decision engine selects the optimal action based on the current state, executes it, observes the next state and reward, and updates the decision parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the virtual power plant scheduling performance risk early warning method based on multi-source heterogeneous data as described in any one of claims 1 to 8.

Images