Virtual power plant aggregation and spot strategy optimization method based on multi-source data fusion

Abstract

The present application belongs to the technical field of virtual power plant, and particularly relates to a virtual power plant aggregation and spot strategy optimization method based on multi-source data fusion. The present application fuses distributed resource real-time operation data and power spot market price signals, constructs a hybrid-driven digital twin with physical information embedded, and outputs a time-varying polyhedron feasible region which explicitly contains climbing constraints in the time dimension and whose geometric boundary is dynamically adjusted according to the price signal; the feasible region is embedded into a day-ahead-real-time two-stage stochastic optimization model as an internal physical constraint, the day-ahead stage and the real-time stage share the same flexible domain function and are substituted into the corresponding price parameters respectively, and the real-time stage deviation penalty cost is dynamically determined by the current boundary of the feasible region and the real-time electricity price; at the same time, a closed-loop online calibration step is set to estimate the accuracy and distribution alignment of the fusion point. The present application realizes the integrated construction of flexible domain representation and multi-stage optimization decision, and improves the safety and economy of the bidding strategy.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant technology, specifically relating to a method for virtual power plant aggregation and spot market strategy optimization based on multi-source data fusion. Background Technology

[0002] For virtual power plants formed by the aggregation of distributed and controllable resources, accurate perception of the adjustable and flexible capabilities of the internal aggregated resources is a prerequisite for ensuring the economic efficiency of bidding and the safety of operation when participating in the spot market.

[0003] For the problem of virtual power plant capacity assessment and bidding decision-making, some existing technologies use data-driven methods to solve for the upper and lower limits of the overall adjustable capacity of the virtual power plant. This method can quickly quantify the adjustable range of resources, but the obtained capacity boundaries are static parameters that ignore the adjustment coupling relationship between time periods and cannot follow the dynamic changes of market prices. Using this as a constraint can easily lead to conservative or aggressive bidding strategies. Other existing technologies directly input multi-source data into deep reinforcement learning networks to generate trading instructions without explicitly constructing an aggregated flexible domain that satisfies the physical operating rules. The decision-making lacks interpretability and security barriers, and there is a risk of violating physical constraints.

[0004] In summary, existing technologies cannot simultaneously address the constraints of temporal and physical coupling with the dynamic response characteristics of market electricity prices, making it difficult to construct a dynamically adaptive aggregated flexible domain and to support multi-stage optimized bidding decisions based on compliant flexible domains. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for virtual power plant aggregation and spot market strategy optimization based on multi-source data fusion, in order to solve the problems in the background art. The specific solution is as follows: A virtual power plant aggregation and spot market strategy optimization method based on multi-source data fusion includes the following steps: Acquire real-time operational data of distributed resources within a virtual power plant and price signals from the electricity spot market; The acquired real-time operating data and price signals are fused to construct a hybrid-driven digital twin with embedded physical information. The hybrid-driven digital twin outputs a time-varying polyhedral feasible domain that reflects the aggregation flexibility of the virtual power plant. Using the time-varying polyhedral feasible region as the physical constraint for the decision variables of the virtual power plant at different time periods, a day-ahead-real-time two-stage stochastic optimization model is constructed to generate the bidding strategy for the virtual power plant to participate in the electricity spot market. The geometric boundary of the feasible region of the time-varying polyhedron is dynamically adjusted as the price signal changes.

[0006] Preferably, the hybrid-driven digital twin includes a physical equation module and a deep residual network module; the physical equation module establishes a baseline model based on the physical operation mechanism of distributed resources and outputs a baseline power vector. The input to the deep residual network module includes the time series of the price signal. and user behavior feature vector Output the boundary offset of the baseline power vector. and The output of the physical equation module and the output of the deep residual network module are superimposed in parallel to generate the time-varying polyhedral feasible region.

[0007] Furthermore, the time-varying polyhedral feasible domain The formula is: in, The time period representing the scheduling For each distributed resource in time period The active power vector, The climbing rate limit; the upper bound of the feasible region of the time-varying polyhedron. and the lower realm The baseline power vector output by the hybrid-driven digital twin is superimposed with the boundary offset to generate the vector.

[0008] Preferably, the physical equation module establishes differentiated physical baseline models for different types of distributed resources; the deep residual network module trains residual networks for different types of distributed resources to learn the boundary offset characteristics caused by market price incentives.

[0009] Furthermore, the day-ahead-real-time two-stage stochastic optimization model is constructed in two stages: day-ahead and real-time. In the day-ahead phase, with the optimization objective of maximizing expected returns, the day-ahead bid power submitted by the virtual power plant to the electricity spot market is determined. Its objective function is: The constraints are: ; In the real-time phase, for random scenarios Real-time market clearing price The goal is to minimize the sum of real-time operating costs and deviation penalties, and to solve for the actual output in each time period. Its objective function is: The constraints include: ; in, This is the market clearing price for the day before yesterday; For time period The operating cost function; For time period The deviation penalty cost function; For the scene The optimal cost in the real-time stage is determined by the expectation operator. Feedback to the objective function of the current phase; The day-ahead phase and the real-time phase use the same time-varying polyhedral feasible region. Substitute the previous day's price signal into the corresponding values. and real-time price signals As its boundary parameters.

[0010] Further, determine the deviation penalty cost. The formula is: in, The penalty coefficient is... and Substituting the real-time price into the feasible region of the time-varying polyhedron, respectively. The upper and lower bounds are then obtained.

[0011] Preferably, it also includes a closed-loop online calibration step: Based on the electricity spot market clearing results and actual operating data, the network parameters of the deep residual network module in the hybrid-driven digital twin were determined. Perform online updates; the optimization objective is to minimize the loss function consisting of the point estimation accuracy term and the distribution alignment term, wherein the loss function is: in, and These are the upper and lower bounds of the actual feasible power obtained from statistical analysis of actual operating data. and To give the price signal for time period t and network parameters The upper and lower bounds of the predicted feasible power are obtained after inputting the hybrid driving digital twin; This is an empirical distribution of the upper bound of feasible power obtained from statistics in actual operation. This represents the feasible upper bound distribution of the output power of the hybrid-driven digital twin within the corresponding price range. As a measure of the maximum mean difference, These are the weighting coefficients for the distribution alignment term.

[0012] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a hybrid-driven digital twin with embedded physical information, outputting a time-varying polyhedral feasible domain that explicitly includes ramp constraints in the time dimension and whose geometric boundaries are dynamically adjusted with spot market price signals. This feasible domain is then directly embedded as an endogenous constraint into a day-to-day and real-time two-stage stochastic optimization model, which can improve the problem of the separation between aggregated flexible domain representation and multi-stage spot market decision-making in existing technologies.

[0013] By substituting different price parameters into the same flexible domain function, this invention achieves the integrated construction of aggregated flexible domain representation and multi-stage optimization decision-making. This enables day-ahead bidding strategies to predict the impact of real-time flexible domain changes on deviation penalties, facilitating a balance between safety and economy in bidding decisions. Simultaneously, this invention designs the deviation penalty cost in the real-time stage to be dynamically determined jointly by the current boundary of the time-varying polyhedral feasible domain and the real-time market clearing price. This ensures that the penalty benchmark is updated synchronously with real-time electricity prices and the dynamic adjustability of resources, reducing the deviation assessment risk for virtual power plants participating in the spot market.

[0014] Furthermore, this invention introduces a composite loss function that combines the accuracy of fusion point estimation and the distribution alignment metric in the closed-loop online calibration step. By using the maximum mean difference metric to constrain the consistency between the predicted boundary distribution and the actual boundary distribution, the calibrated model can continuously adapt to dynamically changing user behavior and market conditions during long-term operation, thereby improving the accuracy of flexible domain representation and the robustness of bidding strategies. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the virtual power plant aggregation and spot market strategy optimization method based on multi-source data fusion of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of this application 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the scenario where virtual power plants aggregate distributed resources to participate in electricity spot market transactions, the adjustable output boundary of various internal distributed controllable resources is affected by multiple factors such as equipment physical mechanisms, market electricity price fluctuations, and users' random energy consumption behavior. Effective quantification of the flexible adjustable range plays an important role in ensuring the rationality of bidding strategies and the safety of resource operation.

[0019] To address this, this invention provides a virtual power plant aggregation and spot market strategy optimization method based on multi-source data fusion. By fusing various heterogeneous operational and market data, a hybrid-driven digital twin model combining physical mechanisms and data-driven approaches is constructed, generating a time-varying polyhedral feasible region that possesses both temporal physical constraints and dynamic electricity price response characteristics. Based on this feasible region as a physical constraint, a day-ahead-to-real-time two-stage stochastic optimization model is built, and a dynamic penalty mechanism is established whereby the deviation penalty cost is jointly determined by the feasible region boundary and the real-time electricity price. Simultaneously, a closed-loop online calibration process is added to measure the accuracy of fusion point estimation and distribution alignment. This invention, while satisfying the physical operational constraints of distributed resources, adapts to the multi-stage trading rules of the electricity spot market, effectively improving the rationality and operational profitability of virtual power plant spot market bidding strategies.

[0020] It should be noted that the virtual power plant in this invention specifically refers to a demand-side virtual power plant formed by aggregation technology of controllable demand-side resources such as air conditioning load groups, charging pile groups, and distributed energy storage. Its core characteristic is that the aggregated resources have the ability to actively adjust their output in response to market price signals, distinguishing it from generation-side virtual power plants primarily composed of distributed power sources. Furthermore, all real-time operational data collected in this invention undergoes anonymization processing before use, removing or anonymizing identifying information involving user privacy, retaining only aggregated statistical data used for feature extraction and behavioral pattern analysis, thus complying with data security and privacy protection requirements.

[0021] Specifically, such as Figure 1 As shown, the virtual power plant aggregation and spot market strategy optimization method based on multi-source data fusion of the present invention includes the following steps: S1 acquires real-time operational data of distributed resources within the virtual power plant and price signals from the electricity spot market.

[0022] First, multi-source data collection was carried out. The collected heterogeneous data included two categories: real-time operation data and price signals from the electricity spot market, which constituted the multi-source data foundation required for this invention.

[0023] Furthermore, in this embodiment, distributed resources refer to heterogeneous resources with adjustable and controllable capabilities, including but not limited to at least one of air conditioning load groups, charging pile groups, and distributed energy storage.

[0024] Real-time operational data refers to measurement data reflecting the current operational status of the aforementioned resources. For example, for air conditioning load groups, real-time operational data includes real-time power and indoor temperature; for charging pile groups, real-time operational data includes charging power and state of charge; for distributed energy storage, real-time operational data includes charging and discharging power and remaining capacity. Furthermore, the aforementioned real-time operational data contains behavioral patterns such as user temperature control, charging selection, and energy storage usage preferences. Based on this, user behavior feature vectors are extracted by performing feature engineering on the real-time operational data. Specifically, user thermal comfort preference features can be extracted from the statistical deviation between the indoor temperature setpoint and the actual temperature of the air conditioning load, as well as the temperature adjustment frequency; user travel habits and charging flexibility features can be extracted from the charging time distribution and charging amount of charging piles; and user arbitrage tendency features can be extracted from the charging and discharging time distribution and charging and discharging depth of energy storage. The extracted user behavior feature vectors are used as input to the deep residual network module in subsequent step S2.

[0025] Price signals in the electricity spot market refer to electricity price information obtained from the electricity trading center that reflects the market supply and demand relationship. Specifically, these include day-ahead market clearing prices and real-time market clearing prices. The day-ahead market clearing price is derived from the unified clearing results of the day-ahead market and is used for bidding decisions in the subsequent day-ahead phase; the real-time market clearing price is derived from the clearing results of the real-time market and is used for deviation penalty calculations in the subsequent real-time phase.

[0026] Understandably, due to the differences in the sources, formats, and dimensions of the two types of data mentioned above, the collected raw data needs to undergo fusion preprocessing operations such as null value removal, out-of-limit data cleaning, and feature normalization to form a standardized multi-source fusion dataset, which serves as the unified input data for subsequent hybrid-driven digital twins.

[0027] S2, the acquired real-time operating data and price signals are fused to construct a hybrid-driven digital twin with embedded physical information, and the hybrid-driven digital twin outputs a time-varying polyhedral feasible domain that reflects the aggregation flexibility of the virtual power plant.

[0028] In this embodiment, the multi-source fusion dataset collected and preprocessed in step S1 is input into the constructed hybrid-driven digital twin. The hybrid-driven digital twin is composed of a physical equation module and a deep residual network module connected in parallel. The two types of modules operate independently and the results are superimposed and fused to jointly complete the quantitative solution of the flexible power range of the virtual power plant.

[0029] The physical equation module relies on the inherent physical operating mechanisms, equipment rated parameters, and operating constraints of various distributed controllable resources such as air conditioning load groups, charging pile groups, and distributed energy storage to build differentiated physical baseline models for different types of resources. Taking air conditioning load groups as an example, the physical baseline model can use a thermodynamic equivalent thermal parameter model to calculate the baseline adjustable range of air conditioning cooling or heating power based on parameters such as indoor-outdoor temperature difference and building thermal resistance. Model parameters include building equivalent heat capacity and equivalent thermal resistance. Taking charging pile groups as an example, the physical baseline model can establish power-state-of-charge dynamic equations based on battery capacity and charging efficiency. Taking distributed energy storage as an example, the physical baseline model can establish a charge-discharge state transition model based on energy storage capacity and charge-discharge efficiency. Through the calculation of the physical baseline model, the baseline power vector of the virtual power plant distributed resources is output. The baseline power vector is the resource-based adjustable power range without considering market electricity price incentives and fluctuations in user behavior.

[0030] Meanwhile, the deep residual network module and the physical equation module are matched with differentiated sub-network structures. Independent deep residual networks are constructed for air conditioning load groups, charging pile groups, and distributed energy storage, respectively, and each sub-network completes training and feature learning independently. In one specific implementation, the unified input of the deep residual network module includes two types of feature data: time series of electricity spot price signals after time-series normalization. And the user behavior feature vector extracted from real-time resource operation data in S1. Each independent residual network is iteratively trained based on historical operating samples and electricity price time-series samples of the corresponding resource. During training, the loss function is to minimize the mean square error between the predicted boundary offset and the actual statistical boundary offset. The network learns the adjustable boundary offset patterns of various distributed resources under different electricity price incentives and user behavior characteristics. The final output is a bidirectional boundary offset relative to the baseline power vector, specifically the upward adjustment boundary offset. and lowering the boundary offset .

[0031] In this embodiment, the physical baseline output and data-driven offset results are fused through parallel superposition to complete the hybrid-driven modeling operation. Specifically, the baseline power vector output by the physical equation module is superimposed with the upward and downward adjustment boundary offsets output by the deep residual network module to obtain dynamically changing upper and lower bounds of the aggregated power. The upper power bound is the sum of the baseline power vector and the upward adjustment boundary offset, and the lower power bound is the sum of the baseline power vector and the downward adjustment boundary offset. Based on this, a temporal coupling constraint of the inherent ramp rate of distributed resources is introduced to limit the power fluctuation range between adjacent scheduling periods. Combined with the aforementioned dynamic upper and lower power bounds, a complete time-varying polyhedral feasible region is constructed. The boundary parameters of this feasible region are jointly controlled by price signals and user behavior characteristics, and can dynamically adjust according to changes in spot market price signals, possessing adaptive adjustment capabilities.

[0032] It should be noted that the time-varying polyhedral feasible region constructed above... Expressed using set constraint formulas, the specific expression is as follows: in, Represents the scheduling period. This represents the aggregated active power vector of all distributed resources during the corresponding scheduling period. The preset equipment ramp-up rate limit is used to restrict the maximum fluctuation range of aggregated power between adjacent scheduling periods, reflecting the temporal coupling characteristics of resource operation. The upper bound of the feasible region power satisfies... The lower power limit is satisfied. All are generated by superimposing the baseline power vector with the corresponding boundary offset, and all are price signals. The function allows the geometric boundary parameters of the feasible region to be dynamically adjusted as the input price signal changes.

[0033] In one specific implementation, the equipment ramp rate limit The ramp rate limit is determined by the inherent physical parameters of each distributed resource. Taking distributed energy storage as an example, its ramp rate limit typically depends on the rated power of the energy storage converter and the charge / discharge rate of the battery, and can generally be set to 10% to 30% of the rated power of the energy storage per scheduling period. Taking air conditioning load groups as an example, its aggregate ramp rate limit depends on the compressor start-stop time constant and building thermal inertia, and can generally be set to 5% to 15% of the aggregate rated power per scheduling period. The above parameter values ​​can be adaptively adjusted according to the actual resource type and operating scenario. The output layer of the deep residual network module uses a linear activation function, making the boundary offset... and The value can be positive or negative, thus enabling the flexible domain boundary to stretch or contract in both directions under the incentive of market prices. Specifically, when the upper boundary offset is positive, the flexible domain's upward adjustment capability expands; when the lower boundary offset is negative, the flexible domain's downward adjustment capability expands.

[0034] S3. Using the time-varying polyhedral feasible region as the physical constraint of the decision variables of the virtual power plant at each time period, a day-ahead-real-time two-stage stochastic optimization model is constructed to generate the bidding strategy for the virtual power plant to participate in the electricity spot market.

[0035] In this embodiment, a two-stage stochastic optimization model adapting to the phased trading rules of the electricity spot market is built based on the time-varying polyhedral feasible region constructed by S2. The polyhedral feasible region, which dynamically changes with electricity prices and user behavior, serves as the core physical constraint for optimization decisions at each stage. This ensures that the value range of all optimization decision variables aligns with the real-time operational capabilities of distributed controllable resources, balancing the compliance of equipment physical operation with the dynamic characteristics of the spot market. The overall model consists of two coupled stages: day-ahead market bidding optimization and real-time market output correction optimization. Both stages use the same time-varying polyhedral feasible region function, dynamically updating the feasible region boundary parameters only by substituting electricity price signals from different trading stages, thus adapting to the phased market decision-making requirements. By substituting different price parameters into the same function, an integrated construction of aggregating flexible domain representation and multi-stage optimization decision-making is achieved.

[0036] Specifically, in the day-ahead optimization phase, day-ahead market clearing price time-series data for the entire scheduling cycle is collected. This electricity price time-series data is input into the hybrid-driven digital twin built by S2, and the time-varying polyhedral feasible region corresponding to the day-ahead trading scenario is updated. The day-ahead optimization phase aims to optimize the overall expected revenue of the virtual power plant under multiple stochastic scenarios, and solves for the day-ahead bidding power submitted to the electricity spot market for each scheduling period. To achieve linkage and matching between day-ahead bidding decisions and real-time operating conditions, this embodiment introduces multiple sets of real-time electricity price stochastic scenarios, coupling the expected operating cost of the real-time stage to the day-ahead objective function. The corresponding day-ahead stage objective function is as follows: in, Let be the day-ahead bidding power decision variable for time period t. Let t be the day-ahead market clearing price. For time period The operating cost function is used to characterize the energy consumption cost and regulation cost generated during the operation of air conditioning load groups, charging pile groups, and distributed energy storage. random corresponding comprehensive , It is an expectation operator based on a random scenario set, used to calculate the average operating cost under different real-time fluctuation conditions, and realize the two-stage cost coupling.

[0037] The corresponding day-ahead optimization core constraint is: the day-ahead bidding power for each scheduling period must be within the dynamic feasible region jointly generated by the day-ahead electricity price and user behavior characteristics, satisfying: This constraint limits the day-ahead bidding power to the adjustable range of various distributed resources under current operating conditions, ensuring that the day-ahead bidding strategy aligns with the actual operating capacity of the equipment.

[0038] In the real-time optimization phase, considering the random fluctuation characteristics of real-time electricity prices in the electricity spot market, a set of random scenarios containing multiple sets of real-time electricity price time series is constructed. These random scenarios can be generated based on the real-time electricity price curves of historical operating days. Specifically, based on the statistical distribution characteristics of historical real-time electricity price curves, Monte Carlo sampling is used to repeatedly and independently sample the random variable of electricity prices, or Latin hypercube sampling is used to hierarchically sample the price distribution, generating several representative sets of real-time electricity price time series as random scenarios. For each independent random scenario, the real-time market clearing price for the corresponding time period is input into the hybrid-driven digital twin, dynamically updating the time-varying polyhedral feasible region specific to the real-time scenario, achieving adaptive updating of the constraint boundaries for each random scenario and each scheduling period. The real-time phase aims to minimize the combined cost of virtual power plant operating costs and output deviation penalty costs under a single scenario, solving for the actual operating output of distributed resources in each time period. The corresponding real-time phase objective function is as follows: in, For random scenarios The actual resource output in the next time period t. The real-time market clearing price for the corresponding scenario and time period. Let $\begin{case}$ be the cost function for the deviation penalty. The core constraint in the real-time phase is that the actual operating output for all random scenarios and all scheduling periods must match the feasible region constraint corresponding to the real-time operating condition, satisfying: .

[0039] In this embodiment, the deviation penalty cost is dynamically determined by the current boundary of the time-varying polyhedral feasible region and the real-time market clearing price. The penalty strength is related to the output over-limit deviation and the real-time electricity price level, and the specific expression is as follows: in, The preset penalty coefficient can be adaptively adjusted according to the spot market trading rules and actual operating scenarios, and its value generally ranges from 1.0 to 3.0. and Substituting real-time prices into the feasible region of the time-varying polyhedron The upper and lower bounds are then obtained. When the actual output of the resource is within the upper or lower bound of the current feasible domain, the output adapts to the real-time adjustable capability of the resource, there is no deviation beyond the boundary, and the penalty cost is zero; when the actual output exceeds the upper or lower bound of the dynamic feasible domain, the corresponding penalty cost is calculated based on the deviation value between the actual output and the boundary and the real-time market electricity price, thereby constraining and correcting unreasonable output behavior.

[0040] By combining the above two-stage stochastic optimization model with the dynamic penalty mechanism, the day-ahead bidding strategy can align with market electricity price trends and the dynamic adjustability of resources. The real-time output dynamic correction can respond to operational deviations caused by random fluctuations in electricity prices and fluctuations in user behavior, ensuring that the final generated spot bidding strategy and real-time operating output always adapt to the physical operating limits of distributed controllable resources and dynamic market conditions.

[0041] To adapt to the long-term dynamic changes in the operating status of distributed resources, user energy consumption behavior, and market environment, reduce the adaptation bias caused by fixed model parameters, and improve the long-term generation accuracy and model adaptability of the time-varying polyhedron feasible region, this embodiment also includes a closed-loop online calibration mechanism for the hybrid-driven digital twin. The model parameters are iteratively updated through actual operating data and market clearing results to continuously optimize the output effect of the feasible region.

[0042] Specifically, relying on actual clearing data from the electricity spot market and on-site measured operational data from distributed resources, we continuously accumulate real-world operating condition samples to refine the network parameters of the deep residual network module. Online iterative updates are performed to ensure that the dynamic feasible domain output by the model can closely match the actual operating characteristics of resources over a long period of time, thus avoiding the fitting bias problem caused by the long-term operation of static models.

[0043] The calibration process uses minimizing the composite loss function as the optimization criterion. This loss function integrates the point estimation accuracy constraint and the probability distribution alignment constraint, while taking into account both the fit of the single-point boundary prediction and the fitting effect of the overall boundary distribution, thus avoiding local adaptation bias caused by a single loss function. The specific expression of the loss function is as follows: in, These are the trainable network parameters for the deep residual network module. , These are the upper and lower bounds of the actual feasible power obtained from statistical analysis of on-site measured operating data; and Given the current model parameters, the price signal for time period t The upper and lower bounds of the predicted feasible power output after inputting a hybrid-driven digital twin.

[0044] The sum of squares in the first part of the formula is the point estimation accuracy term, used to reduce the error between the predicted boundary and the true boundary for each time period. The second part of the formula is the distribution alignment term, where... This represents the empirical probability distribution for the upper bound of the true power. To predict the probability distribution of the upper bound of power within the corresponding price range, It is a measure of maximum mean difference, used to measure the similarity between two probability distributions. This is a weighting coefficient used to adjust the proportion of the distribution alignment term in the overall loss function. It typically ranges from 0.1 to 1.0 and can be adjusted according to the actual training effect.

[0045] In the specific implementation process, a mini-batch gradient descent algorithm is used to iteratively optimize the network parameters. Each iteration selects a subset of historical operating condition samples, and the gradient is backpropagated based on the calculation results of the composite loss function to gradually update the network weight parameters. This process continues iterating until the loss function value tends to stabilize and converge, completing a single online calibration. Through rolling updates and continuous calibration of market operation data, the feature learning effect of the network modules can be gradually optimized, enabling the generated time-varying polyhedral feasible region to continuously adapt to dynamically changing user behavior and market conditions, thereby improving the adaptability of subsequent spot strategy optimization solutions.

[0046] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for virtual power plant aggregation and spot market strategy optimization based on multi-source data fusion, characterized in that, Includes the following steps: Acquire real-time operational data of distributed resources within a virtual power plant and price signals from the electricity spot market; The acquired real-time operating data and price signals are fused to construct a hybrid-driven digital twin with embedded physical information. The hybrid-driven digital twin outputs a time-varying polyhedral feasible domain that reflects the aggregation flexibility of the virtual power plant. Using the time-varying polyhedral feasible region as the physical constraint for the decision variables of the virtual power plant at different time periods, a day-ahead-real-time two-stage stochastic optimization model is constructed to generate the bidding strategy for the virtual power plant to participate in the electricity spot market. The geometric boundary of the feasible region of the time-varying polyhedron is dynamically adjusted as the price signal changes.

2. The method according to claim 1, characterized in that, The hybrid-driven digital twin includes a physical equation module and a deep residual network module; the physical equation module establishes a baseline model based on the physical operation mechanism of distributed resources and outputs a baseline power vector. The input to the deep residual network module includes the time series of the price signal. and user behavior feature vector Output the boundary offset of the baseline power vector. and The output of the physical equation module and the output of the deep residual network module are superimposed in parallel to generate the time-varying polyhedral feasible region.

3. The method according to claim 2, characterized in that, The time-varying polyhedron feasible region The formula is: in, The time period representing the scheduling For each distributed resource in time period The active power vector, The climbing rate limit; the upper bound of the feasible region of the time-varying polyhedron. and the lower realm The baseline power vector output by the hybrid-driven digital twin is superimposed with the boundary offset to generate the vector.

4. The method according to claim 2, characterized in that, The physical equation module establishes differentiated physical baseline models for different types of distributed resources; the deep residual network module trains residual networks for different types of distributed resources to learn the boundary offset characteristics caused by market price incentives.

5. The method according to claim 1 or 3, characterized in that, The day-ahead-real-time two-stage stochastic optimization model is constructed in two stages: day-ahead and real-time. In the day-ahead phase, with the optimization objective of maximizing expected returns, the day-ahead bid power submitted by the virtual power plant to the electricity spot market is determined. Its objective function is: The constraints are: ; In the real-time phase, for random scenarios Real-time market clearing price The goal is to minimize the sum of real-time operating costs and deviation penalties, and to solve for the actual output in each time period. Its objective function is: The constraints include: ; in, This is the market clearing price for the day before yesterday; For time period The operating cost function; For time period The deviation penalty cost function; For the scene The optimal cost in the real-time stage is determined by the expectation operator. Feedback to the objective function of the current phase; The day-ahead phase and the real-time phase use the same time-varying polyhedral feasible region. Substitute the previous day's price signal into the corresponding values. and real-time price signals As its boundary parameters.

6. The method according to claim 5, characterized in that, Determine the deviation penalty cost The formula is: in, The penalty coefficient is... and Substituting the real-time price into the feasible region of the time-varying polyhedron, respectively. The upper and lower bounds are then obtained.

7. The method according to claim 2 or 3, characterized in that, It also includes a closed-loop online calibration step: Based on the electricity spot market clearing results and actual operating data, the network parameters of the deep residual network module in the hybrid-driven digital twin were determined. Perform online updates; the optimization objective is to minimize the loss function consisting of the point estimation accuracy term and the distribution alignment term, wherein the loss function is: in, and These are the upper and lower bounds of the actual feasible power obtained from statistical analysis of actual operating data. and To give the price signal for time period t and network parameters The upper and lower bounds of the predicted feasible power are obtained after inputting the hybrid driving digital twin; This is an empirical distribution of the upper bound of feasible power obtained from statistics in actual operation. This represents the feasible upper bound distribution of the output power of the hybrid-driven digital twin within the corresponding price range. As a measure of the maximum mean difference, These are the weighting coefficients for the distribution alignment term.

Images