Electricity-carbon-energy flow collaborative economic dispatching method based on multi-objective optimization

Abstract

The invention discloses an electricity-carbon-energy flow collaborative economic dispatching method based on multi-objective optimization, and belongs to the technical field of power dispatching. The method comprises the following steps: generating local model parameters containing localized event features; constructing a global event consensus model; analyzing the power grid state data in real time through a global event consensus model, and generating an event vector; based on the change of the electric charge cost and the carbon emission cost before and after the optimal response action, forming a training sample with a value income label; and obtaining a model update amount by taking maximized value and income data as an optimization target. Through the synergistic effect of federated learning and an event-driven mechanism, the technical effects of improving scheduling real-time performance and realizing automatic response of a user side are achieved, and the problems of scheduling delay caused by a computing power bottleneck and low response efficiency caused by dependence on artificial experience in the prior art are solved.

Description

A Multi-Objective Optimization-Based Coordinated Economic Dispatch Method for Electricity-Carbon-Energy Flow Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a multi-objective optimization-based coordinated economic dispatching method for electricity-carbon-energy flow. Background Technology

[0002] With the deepening implementation of the dual-carbon strategy, the power system is transforming towards a high proportion of renewable energy integration and diversified load development. Against this backdrop, grid dispatching needs to simultaneously consider economic operating costs and carbon emission control targets, achieving multi-dimensional synergistic optimization of electricity, carbon, and energy flow. Traditional dispatching methods often employ centralized optimization models, treating the power grid as a single entity for unified calculation and command issuance.

[0003] The aforementioned disclosed technical solutions suffer from at least the following technical problems: Traditional methods, employing a centralized computing model, require synchronous carbon flow tracking and optimization solutions for tens of thousands of nodes across the entire network. The computational complexity increases exponentially with system size, resulting in severely insufficient real-time scheduling decisions and an inability to meet the dynamic balancing needs of the power grid at the second to minute level. Furthermore, the lack of an automatic decision-making mechanism on the user side to convert system-level optimization signals into specific operational instructions hinders the effective participation of distributed resources in system collaborative optimization. To address these problems, this invention proposes a solution. Summary of the Invention

[0004] This application provides a multi-objective optimization-based collaborative economic scheduling method for electricity-carbon-energy flow, which solves the problems of scheduling delay caused by centralized computing and collaborative difficulties caused by the lack of intelligent decision-making mechanisms on the user side in the prior art. It achieves the effect of improving the real-time response capability of the system through distributed parallel computing and realizing the automated and accurate response on the user side with the help of intelligent agents.

[0005] This application provides a multi-objective optimization-based coordinated economic dispatch method for electricity-carbon-energy flow, comprising: collecting local data and generating local model parameters containing localized event features; aggregating the local model parameters through federated learning to construct a global event consensus model; acquiring grid state data and parsing the grid state data in real time through the global event consensus model to generate event vectors; receiving the event vectors, executing optimal response actions, and forming training samples with value benefit labels based on the changes in electricity costs and carbon emission costs before and after the optimal response actions; obtaining model update amounts with maximizing value benefit data as the optimization objective, and iteratively updating the global event consensus model based on the model update amounts.

[0006] Furthermore, the step of generating local model parameters containing localized event features includes: receiving unstructured event seeds containing event labels broadcast by the scheduling center, extracting the collected local data from the local memory; combining the unstructured event seeds with the local data to form an initial training dataset, wherein the unstructured event seeds serve as input feature vectors and the local data serve as corresponding labels; initializing the local neural network model and executing a distributed training process; repeating the distributed training process for multiple iterations to generate local model parameters containing localized event features, and storing the local model parameters in local memory.

[0007] Furthermore, the steps for constructing a global event consensus model include: collecting uploaded local model parameters and initializing a global neural network model; obtaining a weighted average result by performing a weighted average calculation on all local model parameters; updating the weights and bias parameters of the global neural network model using the weighted average result to obtain a preliminary global event consensus model; verifying the performance of the preliminary global event consensus model and comparing its matching degree with preset event labels; if the matching degree is lower than a preset matching threshold, sending an update request to obtain a new round of local model parameters and repeating the aggregation process; after aggregation, storing the preliminary global event consensus model in a central server, the preliminary global event consensus model taking real-time power grid status data as input and outputting the corresponding event consensus representation; testing the generalization ability of the preliminary global event consensus model, and determining that the global event consensus model can map real-time power grid status data as specific events when the generalization ability is not less than a preset generalization threshold, thus completing the construction process.

[0008] Furthermore, the step of generating event vectors by real-time parsing of power grid status data through a global event consensus model includes: acquiring real-time power grid status data from the power grid sensor network and preprocessing the real-time power grid status data into a standardized input vector; inputting the standardized input vector into the input layer of the global event consensus model; the global event consensus model extracting the highest probability event type and its confidence score from the output layer; if the confidence score exceeds a preset consensus threshold, triggering the corresponding event and generating an event vector, and broadcasting the event vector to smart agents within a specific range.

[0009] Furthermore, the steps for executing the optimal response action, based on the changes in electricity costs and carbon emission costs before and after the optimal response action, include: after receiving the broadcast event vector, extracting the current user-side state from the local context information; parsing the event vector to obtain the event type, confidence level, and recommended response strategy meta-instructions; generating a candidate response action list based on the recommended response strategy meta-instructions, evaluating the expected effect of each candidate response action, and calculating and estimating electricity costs and carbon emission costs through simulation; selecting the response action with the lowest expected sum of electricity costs and carbon emission costs as the optimal response action; executing the optimal response action, including updating the local energy management system to implement adjustments; recording the electricity cost and carbon emission cost values ​​before and after executing the optimal response action; obtaining the change difference and associating the change difference with the event vector to form value benefit data; and storing the value benefit data in a local database to generate training samples labeled with value benefits.

[0010] Furthermore, the steps of obtaining the model update quantity and iteratively updating the global event consensus model based on the model update quantity include: using the generated training samples with value benefit labels as input datasets to retrain the local model, thereby obtaining the model update quantity; uploading the model update quantity to the scheduling center and transmitting it through an encrypted channel to ensure security; collecting the model update quantities of all intelligent agents through the scheduling center; the scheduling center performing a weighted summation of all model update quantities to obtain a weighted summation result; updating the weights and bias parameters of the global event consensus model through the weighted summation result; verifying the performance of the updated global model by calculating the output event vector through input test power grid state data and evaluating the matching degree with historical value benefits; if the matching degree is lower than a preset matching threshold of two, initiating a new round of retraining request; repeating the iterative update process for multiple cycles until the performance of the global model is stable.

[0011] Furthermore, the steps for testing the generalization ability of the preliminary global event consensus model include: pre-setting a test dataset with a different distribution than the training data but derived from real-world scenarios. ;Will Each sample in the process is input into the preliminary global event consensus model, and its output is obtained; the computation model is then used to... Accuracy on the surface is used as a quantitative indicator of generalization ability: In the formula, For the model's accuracy, The total number of samples in the test dataset, For the model in The model is considered to have achieved the following generalization ability: if the calculated accuracy of the model is not less than the preset generalization threshold, then the model's generalization ability is deemed to have met the standard.

[0012] Furthermore, evaluating the expected effects of each candidate response action, through simulation to calculate the estimated electricity costs and carbon emission costs, includes the following steps: for each candidate response action... Based on the described power adjustment plan, predict the power curve on the user side after implementation. Based on predicted power curves Known electricity price curve And the marginal carbon intensity curve of the power grid The estimated electricity cost is obtained through the formula for estimating electricity costs: In the formula, To execute candidate response actions The estimated electricity cost afterward, To perform the action Afterwards, during the time period Predicted power consumption For the time period The electricity price is expressed in yuan per kilowatt-hour. The time interval is used; the estimated carbon emission cost is obtained through the carbon emission cost formula: In the formula, To execute candidate response actions The estimated carbon emission costs afterward To perform the action Afterwards, during the time period Predicted power consumption For the time period The marginal carbon intensity of the power grid For time intervals, This refers to the carbon price.

[0013] Furthermore, repeating the iterative update process multiple times until the global model's performance stabilizes includes: after each round of iterative update, on a fixed validation dataset... The performance of the global model is evaluated to obtain the loss function value. : In the formula, For the first The loss value on the validation set after rounds of iteration. For the current iteration round, The preset convergence threshold, For consecutive observation rounds.

[0014] Continuous monitoring The verification loss value of the wheel.

[0015] If continuous The decrease in the validation loss for each round was less than a very small positive number. If the performance is stable, the iteration will stop.

[0016] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: Addressing the problems of high system computational complexity and slow response speed, this invention transforms centralized massive computation into distributed parallel processing by decomposing global computation tasks into multiple local computation tasks. This eliminates the dependence on the computing power of a central node, realizing a shift from unified computation and delayed results to distributed computation and real-time response, effectively ensuring the timeliness of scheduling instructions.

[0017] Secondly, addressing the issue of users struggling to make decisions when faced with specialized signals, this invention deploys intelligent units with autonomous decision-making capabilities at each user terminal, transforming abstract system-level signals into specific device-level operational instructions. This eliminates the gap between users' professional understanding and system requirements, transforming complex electric-carbon coordinated scheduling into an automated process requiring no human intervention, and ensuring the precise implementation of system optimization strategies.

[0018] Through continuous data collection, decision execution, and effect feedback, a complete optimization loop has been formed, which significantly improves the system's economic operation and environmental benefits while ensuring the stable operation of the power grid. Attached Figure Description

[0019] Figure 1 is a flowchart of a multi-objective optimization-based coordinated economic dispatching method for electricity-carbon-energy flow provided in an embodiment of this application. Detailed Implementation

[0020] This application provides a multi-objective optimization-based collaborative economic scheduling method for electricity-carbon-energy flow, which solves the problems of poor real-time performance and lack of automated decision-making capabilities on the user side in the existing centralized computing model. By using federated learning to build a global event consensus model and guiding intelligent agents to respond autonomously based on event-driven and value feedback mechanisms, the method achieves the effect of real-time and accurate issuance of scheduling instructions and automated collaborative optimization of user-side resources.

[0021] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0022] As shown in Figure 1, this application embodiment provides a multi-objective optimization-based coordinated economic dispatch method for electricity-carbon-energy flow, including: deploying intelligent agents at the user end to collect local data, and broadcasting unstructured event seeds containing event tags to the intelligent agents by the dispatch center; each intelligent agent performs distributed training on the event seeds based on local data to generate local model parameters containing localized event features; each intelligent agent forms its own local model based on the local model parameters; aggregating the local model parameters through federated learning to construct a global event consensus model that maps real-time grid status data to specific events; the step of constructing the global event consensus model is as follows: the dispatch center initializes a global model, the structure of which is the same as the local models of each agent; collects the local model parameters (such as neural network weights) uploaded by each agent; performs a weighted average of the collected model parameters according to the amount of local data of each agent, and updates the global model with the weighted average parameters; distributes the updated global model parameters to each agent for the next round of training. This process is repeated until the model performance meets the preset requirements.

[0023] The system acquires grid status data and parses it in real time using a global event consensus model. When an event that meets a preset consensus condition is identified, the event is triggered, and an event vector containing the event type, confidence level, and recommended response strategy meta-instructions is generated. This event vector is then broadcast to intelligent agents within a specific range. The preset consensus condition refers to the confidence score of a certain event type output by the global event consensus model exceeding a preset threshold. For example, if the global event consensus model determines that the confidence level of the current "sudden drop in photovoltaic output" event is 0.92, and the preset consensus threshold is 0.85, then the preset consensus condition is met, and the event is triggered.

[0024] A specific range refers to the area comprised of a group of users who are adjacent to the physical location of an event in terms of electrical distance from the power grid. Its range is dynamically determined based on the event type: for local events (such as "overload of a feeder"), the range includes all users supplied by that feeder; for global events (such as "high system carbon intensity"), the range includes all users across the entire network. The specific range can be set by the dispatcher based on the power grid topology and experience, or automatically matched to a pre-defined "event-range" mapping table by the system based on event characteristics.

[0025] The intelligent agent receives the event vector, combines it with its local context information to execute the optimal response action, and calculates and records the value benefit data associated with the event vector based on the changes in electricity cost and carbon emission cost before and after the optimal response action, forming training samples labeled with value benefits. The steps for collecting the value benefit data are as follows: record the total electricity cost in the period (e.g., 15 minutes) before executing the response action. Total carbon emission cost Record the total electricity cost within one cycle after the execution of the response action. Total carbon emission cost ; Calculate the value gain using the value gain calculation formula: In the formula, Value gain, expressed in monetary units. , These represent the electricity costs before and after the response action is executed, in monetary units. , These represent the carbon emission costs before and after the response action, respectively, in monetary units. Carbon emission cost = carbon emissions × carbon price. , These are the economic weights and the low-carbon weights, respectively, and are dimensionless constants.

[0026] The value gain data is associated with the specific event vector that triggered the response action. This association is achieved by assigning a unique identifier (EventID) to each event vector and binding that EventID when recording the value gain. Any value gain data generated by that event vector is considered to be associated with it.

[0027] With the goal of maximizing the value of data, the local model is retrained using training samples labeled with value to obtain the model update amount. Then, the global event consensus model is iteratively updated based on the model update amount through a federated evolutionary aggregation mechanism, forming a closed-loop self-optimization of scheduling strategy and user response behavior.

[0028] Furthermore, the steps for generating local model parameters containing localized event features include: after the intelligent agent receives unstructured event seeds containing event labels broadcast by the dispatch center, the intelligent agent extracts the collected local data from the local memory, the local data including user-side power consumption records, carbon emission indicators, and energy usage patterns; the intelligent agent combines the unstructured event seeds with the local data to form an initial training dataset, wherein the unstructured event seeds serve as input feature vectors and the local data serve as corresponding labels; the intelligent agent initializes a local neural network model, the local neural network model adopts a multilayer perceptron structure, including an input layer, a hidden layer, and an output layer, wherein the input layer receives the event seed vector, the hidden layer applies an activation function to process the intermediate representation, the output layer generates event feature representations, and the intelligent agent executes a distributed training process, the distributed training process first divides the initial training dataset into batches, then calculates the forward propagation to obtain the predicted output in each batch, then calculates the mean squared error loss function value between the predicted output and the local data labels, and then updates the weights and bias parameters of the local neural network model through the backpropagation algorithm; the step of calculating the forward propagation to obtain the predicted output is: input event seed vector via local neural network model Calculate and obtain the predicted output. : In the formula, The input event seed feature vector, For the weights and bias parameters of the local neural network, This represents the feature representation of the predicted events output by the model.

[0029] The steps for calculating the mean squared error loss function value between the predicted output and the local data label are as follows: Calculate the predicted output using the mean squared error. With real labels Differences (i.e., local data): In the formula, This is the mean squared error loss value. The number of samples in a batch. For the first The predicted value for each sample, For the first The actual local data label corresponding to each sample.

[0030] The intelligent agent repeats the distributed training process for multiple iterations until the loss function value converges to below a preset threshold obtained from the historical database. It then generates local model parameters containing localized event features, including an updated weight matrix and bias vector. These parameters are stored in local memory for subsequent use in forming the local model. This process ensures that the intelligent agent independently processes local data without leaking privacy information and achieves accurate extraction of localized event features through iterative optimization.

[0031] Furthermore, the steps for constructing the global event consensus model include: the scheduling center collects the local model parameters uploaded by each intelligent agent, initializes the global neural network model (the global neural network model has the same structure as the local neural network model, including an input layer, hidden layers, and an output layer); the scheduling center executes a federated learning aggregation algorithm to calculate a weighted average of all local model parameters, where the weight of each local model parameter is determined based on the number of data samples for the corresponding intelligent agent; the weights and bias parameters of the global neural network model are updated using the weighted average result to obtain a preliminary global event consensus model; the performance of the preliminary global event consensus model is verified by calculating the output event representation using input simulated power grid state data and comparing its matching degree with preset event labels; the output event representation is the probability prediction of the global event consensus model for the input power grid state data belonging to various event types. This calculation is completed through the model's forward propagation and the Softmax function.

[0032] Assuming a global event consensus model For the input power grid state vector An output layer was generated Let the original fractions be vectors. Then the event type The output event representation (i.e., probability) The calculation formula is as follows: In the formula, To represent input data Belongs to event type The predicted probability. Its range is [0,1], and the sum of the probabilities of all event types is 1, that is... , For the model output layer The raw scores of each neuron represent the model's performance on event types. The original confidence level, The total number of event types. It is a natural constant.

[0033] The steps for comparing the match with preset event labels are as follows: For an input test sample, its predicted event type... It is the type with the highest probability, that is ; this predicts the event type Compared with the real event labels of this sample (i.e., the preset event labels) Compare across the entire test set; Above, matching degree The calculation formula is: In the formula, The matching degree, expressed as a percentage, represents the accuracy of the model's predictions. The total number of test samples, For the index of the test sample, For the first The predicted event type for each sample, For the first The real event labels (preset event labels) of each sample.

[0034] If the matching degree is lower than the preset matching threshold, an update request is sent to the intelligent agent through the scheduling center to obtain a new round of local model parameters, and the aggregation process is repeated. After aggregation, the preliminary global event consensus model is stored in the central server. The preliminary global event consensus model takes real-time power grid status data as input and outputs the corresponding event consensus representation. The generalization ability of the preliminary global event consensus model is tested by observing the output stability through inputting diverse power grid status data samples. When the generalization ability is not less than the preset generalization threshold, it is determined that the global event consensus model can map real-time power grid status data into specific events, thus completing the construction process. This step ensures that the global model integrates distributed knowledge through multiple iterative aggregations.

[0035] The method for testing the generalization ability of the preliminary global event consensus model is as follows: The model is run on a diverse historical power grid dataset that was not used for training. The predicted event types are compared with actual records. Performance is quantified using two common machine learning evaluation metrics: accuracy and macro-average F1 score. When both metrics are not lower than preset thresholds, the preliminary global event consensus model is considered to have good generalization ability. Accuracy is obtained by measuring the proportion of correctly predicted results out of the total test samples, thus assessing the overall correctness of the model's judgments.

[0036] The macro average F1 score is obtained by comprehensively examining the ability of the preliminary global event consensus model to identify all event types, especially when the event type distribution is uneven, so as to fairly evaluate the recognition accuracy and recall ability of the preliminary global event consensus model for each type.

[0037] Furthermore, the steps for generating event vectors by real-time parsing of power grid state data using a global event consensus model include: acquiring real-time power grid state data from the power grid sensor network, which includes voltage levels, current intensity, load distribution, and carbon emission monitoring values; preprocessing the real-time power grid state data into a standardized input vector, including normalized numerical ranges and filling missing values; inputting the standardized input vector into the input layer of the global event consensus model; the global event consensus model performing forward propagation calculation, first passing the vector from the input layer to the hidden layer, the hidden layer applying an activation function to generate intermediate feature representations, and then the output layer calculating the event probability distribution based on the intermediate features; the global event consensus model extracting the highest probability event type and its confidence score from the output layer; the steps for obtaining the confidence score are: performing forward propagation calculation to obtain the output layer vector, applying the softmax function to convert the output layer vector into a probability distribution, and extracting the probability value of the corresponding highest probability event type from the probability distribution as the confidence score.

[0038] If the confidence score exceeds the preset consensus threshold, the dispatch center triggers the corresponding event and generates an event vector. This vector contains the event type code, confidence value, and a recommended response strategy meta-instruction sequence generated based on historical response data. The event vector is then broadcast to smart agents within a specific range and distributed via network protocol to achieve real-time synchronization. This process ensures the accuracy and timeliness of parsing power grid status data and directly generates operable event vectors for subsequent responses.

[0039] Furthermore, the steps for executing the optimal response action, based on the changes in electricity costs and carbon emission costs before and after the optimal response action, include: after receiving the broadcast event vector, the intelligent agent extracts the current user-side state from the local context information, including real-time electricity demand, carbon emission benchmark, and available energy resource allocation; parses the event vector to obtain the event type, confidence level, and recommended response strategy meta-instructions; generates a candidate response action list based on the recommended response strategy meta-instructions, where each action is described as a specific operation sequence for adjusting the electricity consumption pattern or energy allocation scheme, evaluates the expected effect of each candidate response action, and calculates and estimates electricity costs and carbon emission costs through simulation execution; selects the response action with the lowest expected sum of electricity costs and carbon emission costs as the optimal response action; the intelligent agent executes the optimal response action, including updating the local energy management system to implement the adjustment; records the electricity cost and carbon emission cost values ​​before and after executing the optimal response action; obtains the change difference, i.e., the difference between the electricity cost and carbon emission cost before and after execution, and associates the change difference with the event vector to form value benefit data; stores the value benefit data in a local database to generate training samples labeled with value benefits. This step optimizes the recording of response actions by quantifying cost changes.

[0040] Furthermore, the steps of obtaining the model update amount and iteratively updating the global event consensus model based on the model update amount include: using the generated training samples with value gain labels as input datasets to retrain the local model. The retraining steps include: first loading the current local model parameters, then calculating forward propagation in each training iteration to obtain the predicted value gain, then calculating the mean squared error loss function value between the predicted and actual value gain labels, and then calculating the gradient and updating the weights and biases of the local model through the backpropagation algorithm to obtain the model update amount, which represents the difference matrix between the weights and biases; and uploading the model update amount to the scheduling center through a smart agent, transmitting it through an encrypted channel to ensure security. The scheduling center collects model updates from all intelligent agents. It then executes a federated evolutionary aggregation mechanism, first performing a weighted summation of all updates, where weights are determined based on the value gain data magnitude of each intelligent agent. The weights and bias parameters of the global event consensus model are updated using this weighted summation. The performance of the updated global model is verified by calculating the output event vector using test grid state data and evaluating its matching degree with historical value gains. If the matching degree is lower than a preset matching threshold, a new round of retraining requests is initiated through the scheduling center. This iterative update process is repeated multiple times until the global model's performance stabilizes, thus forming a closed-loop self-optimization mechanism between the scheduling strategy and user response behavior. This process integrates distributed updates to achieve continuous model evolution.

[0041] Furthermore, the steps for testing the generalization ability of the preliminary global event consensus model include: pre-setting a test dataset with a different distribution than the training data but derived from real-world scenarios. ;Will Each sample in the process is input into the preliminary global event consensus model, and its output is obtained; the computation model is then used to... Accuracy on the surface is used as a quantitative indicator of generalization ability: In the formula, The accuracy of the model is expressed as a percentage. The total number of samples in the test dataset, For the model in The model is considered to have achieved the following generalization ability: if the calculated accuracy of the model is not less than the preset generalization threshold, then the model's generalization ability is deemed to have met the standard.

[0042] Furthermore, evaluating the expected effects of each candidate response action, through simulation to calculate the estimated electricity costs and carbon emission costs, includes the following steps: for each candidate response action... Based on the described power adjustment plan, predict the power curve on the user side after implementation. Based on predicted power curves Known electricity price curve And the marginal carbon intensity curve of the power grid The estimated electricity cost is obtained through the formula for estimating electricity costs: In the formula, To execute candidate response actions The estimated electricity cost is given below, in monetary units. To perform the action Afterwards, during the time period The predicted power consumption, in kilowatts. For the time period The electricity price is expressed in yuan per kilowatt-hour. The time interval is defined in units of time; the estimated carbon emission cost is obtained using the carbon emission cost formula: In the formula, To execute candidate response actions The estimated carbon emission costs are presented in monetary units. To perform the action Afterwards, during the time period The predicted power consumption, in kilowatts. For the time period The marginal carbon intensity of the power grid, expressed in tons of carbon dioxide per kilowatt-hour. The time interval is expressed in units of time. The price is the carbon price, expressed in yuan per ton of carbon dioxide.

[0043] Furthermore, repeating the iterative update process multiple times until the global model's performance stabilizes includes: after each round of iterative update, on a fixed validation dataset... The performance of the global model is evaluated to obtain the loss function value. : In the formula, For the first The loss value on the validation set after rounds of iteration. For the current iteration round, The preset convergence threshold, For consecutive observation rounds.

[0044] Continuous monitoring Wheel (such as) The verification loss value.

[0045] If continuous The decrease in the validation loss for each round was less than a very small positive number. If the performance is considered stable, the iteration will stop.

[0046] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0047] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0048] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0049] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0051] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-objective optimization-based coordinated economic dispatch method for electricity-carbon-energy flow, characterized in that, Includes the following steps: Collect local data and generate local model parameters that include localized event characteristics; By aggregating the parameters of each local model through federated learning, a global event consensus model is constructed; power grid state data is acquired, and the power grid state data is analyzed in real time through the global event consensus model to generate event vectors; the event vectors are received, the optimal response action is executed, and training samples with value benefit labels are formed based on the changes in electricity cost and carbon emission cost before and after the optimal response action. With the goal of maximizing the value of the data, the model update amount is obtained, and the global event consensus model is iteratively updated based on the model update amount.

2. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 1, characterized in that, The steps for generating local model parameters containing localized event features include: receiving unstructured event seeds containing event labels broadcast by the scheduling center, and extracting the collected local data from the local memory; combining the unstructured event seeds with the local data to form an initial training dataset, wherein the unstructured event seeds serve as input feature vectors and the local data serve as corresponding labels; initializing the local neural network model and executing a distributed training process; repeating the distributed training process for multiple iterations to generate local model parameters containing localized event features, and storing the local model parameters in local memory.

3. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 1, characterized in that, The steps for constructing a global event consensus model include: collecting uploaded local model parameters and initializing a global neural network model; calculating a weighted average of all local model parameters to obtain a weighted average result; updating the weights and bias parameters of the global neural network model using the weighted average result to obtain a preliminary global event consensus model; verifying the performance of the preliminary global event consensus model and comparing its matching degree with preset event labels; if the matching degree is lower than a preset matching threshold, sending an update request to obtain a new round of local model parameters and repeating the aggregation process; after aggregation, storing the preliminary global event consensus model in a central server, the preliminary global event consensus model taking real-time power grid status data as input and outputting the corresponding event consensus representation; testing the generalization ability of the preliminary global event consensus model, and when the generalization ability is not less than a preset generalization threshold, determining that the global event consensus model can map real-time power grid status data as specific events, thus completing the construction process.

4. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 1, characterized in that, The steps for generating event vectors by real-time parsing of power grid status data using a global event consensus model include: acquiring real-time power grid status data from a power grid sensor network and preprocessing the real-time power grid status data into a standardized input vector; inputting the standardized input vector into the input layer of the global event consensus model; the global event consensus model extracting the highest probability event type and its confidence score from the output layer; if the confidence score exceeds a preset consensus threshold, triggering the corresponding event and generating an event vector, and broadcasting the event vector to smart agents within a specific range.

5. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 1, characterized in that, The steps for executing the optimal response action, based on the changes in electricity costs and carbon emission costs before and after the optimal response action, include: after receiving the broadcast event vector, extracting the current user-side state from the local context information; parsing the event vector to obtain the event type, confidence level, and recommended response strategy meta-instructions; generating a candidate response action list based on the recommended response strategy meta-instructions, evaluating the expected effect of each candidate response action, and calculating and estimating electricity costs and carbon emission costs through simulation execution; selecting the response action with the lowest expected sum of electricity costs and carbon emission costs as the optimal response action; executing the optimal response action, including updating the local energy management system to implement adjustments; recording the electricity cost and carbon emission cost values ​​before and after executing the optimal response action; obtaining the change difference and associating the change difference with the event vector to form value benefit data; and storing the value benefit data in a local database to generate training samples labeled with value benefits.

6. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 1, characterized in that, The steps of obtaining the model update quantity and iteratively updating the global event consensus model based on the model update quantity include: using the generated training samples with value benefit labels as input datasets to retrain the local model to obtain the model update quantity; uploading the model update quantity to the scheduling center and transmitting it through an encrypted channel to ensure security; collecting the model update quantities of all smart agents through the scheduling center; the scheduling center performing a weighted summation of all model update quantities to obtain a weighted summation result; updating the weights and bias parameters of the global event consensus model through the weighted summation result; verifying the performance of the updated global model by calculating the output event vector through input test power grid state data and evaluating the matching degree with historical value benefits; if the matching degree is lower than a preset matching threshold of two, initiating a new round of retraining request; repeating the iterative update process for multiple cycles until the performance of the global model is stable.

7. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 3, characterized in that, The steps for testing the generalization ability of the preliminary global event consensus model include: pre-setting a test dataset with a different distribution than the training data but derived from real-world scenarios. ;Will Each sample in the process is input into the preliminary global event consensus model, and its output is obtained; the computation model is then used to... Accuracy on the surface is used as a quantitative indicator of generalization ability: In the formula, For the model's accuracy, The total number of samples in the test dataset, For the model in The model is considered to have achieved the following generalization ability: if the calculated accuracy of the model is not less than the preset generalization threshold, then the model's generalization ability is deemed to have met the standard.

8. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 5, characterized in that, The steps for evaluating the expected effects of each candidate response action, and calculating the estimated electricity and carbon emission costs through simulation, include: for each candidate response action... Based on the described power adjustment plan, predict the power curve on the user side after implementation. Based on predicted power curves Known electricity price curve And the marginal carbon intensity curve of the power grid The estimated electricity cost is obtained through the formula for estimating electricity costs: In the formula, To execute candidate response actions The estimated electricity cost afterwards To perform the action Afterwards, during the time period Predicted power consumption For the time period The electricity price is expressed in yuan per kilowatt-hour. The time interval is used; the estimated carbon emission cost is obtained through the carbon emission cost formula: In the formula, To execute candidate response actions The estimated carbon emission costs afterward To perform the action Afterwards, during the time period Predicted power consumption For the time period The marginal carbon intensity of the power grid For time intervals, This refers to the carbon price.

9. The method for coordinated economic dispatch of electricity, carbon, and energy flow based on multi-objective optimization as described in claim 6, characterized in that, Repeating the iterative update process multiple times until the global model's performance stabilizes includes the following steps: after each round of iterative updates, on a fixed validation dataset... The performance of the global model is evaluated to obtain the loss function value. ： In the formula, For the first The loss value on the validation set after rounds of iteration. For the current iteration round, The preset convergence threshold, For continuous observation rounds. Continuous monitoring. The verification loss value of the round. If continuous The decrease in the validation loss for each round was less than a very small positive number. If the performance is considered stable, the iteration will stop.

Images