Virtual power plant optimization scheduling method based on composite network prediction model

By adopting a composite network model and a stepped carbon trading mechanism in virtual power plants, the limitations of a single algorithm forming a prediction model are solved, the load prediction accuracy and clean energy utilization rate are improved, and the efficient operation and stability of the virtual power plant system are achieved.

CN119030033APending Publication Date: 2024-11-26GUANGXI ELECTRICAL POLYTECHNIC INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410968381.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the prior art, the prediction model composed of a single algorithm only tends to feature extraction of local correlation or timing in the virtual power plant load prediction, resulting in low prediction accuracy and difficult for scheduling strategies to make full use of clean energy to generate electricity.

Method used

Using a prediction model based on composite network, including residual density module, bidirectional long and short-term memory network module and scaling dot-product attention module, preprocessing and feature extraction of virtual power plants, combining step-by-step carbon trading mechanism and comprehensive thermoelectric requirements In response, establish a day-and-day optimization scheduling model, and perform rolling optimization scheduling with the goal of minimizing operating costs and prediction deviations.

Benefits of technology

It improves load prediction accuracy and clean energy utilization rate, smoothes power fluctuations caused by recent prediction errors, and enhances the stability and economical system operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119030033A_ABST
    Figure CN119030033A_ABST
Patent Text Reader

Abstract

The invention discloses a virtual power plant optimization scheduling method based on a composite network prediction model. The method comprises the following steps: S1, a data collection and processing stage; s2, a day-ahead optimization scheduling stage; and S3, an intra-day optimization scheduling stage. The prediction precision of the output data and the thermoelectric load data of each generator set in the power plant system model is improved by utilizing the prediction model based on the composite network, an accurate prediction data basis is provided for accurate scheduling of the virtual power plant system model, and a stepped carbon transaction mechanism is utilized and thermoelectric load response is considered; and the virtual power plant system can perform resource allocation and energy utilization more accurately and efficiently. In combination with a day-ahead time scale scheduling method and an intra-day time scale scheduling method, prediction error penalty is added into an intra-day scheduling objective function, a calculation result of an intra-day scheduling model is optimized in a rolling manner, and the operation economy of the system is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of electric power dispatching, and in particular relates to a virtual power plant optimization dispatching method based on a composite network prediction model. Background Art

[0002] As a power coordination management system participating in the electricity market and power grid operation, the Virtual Power Plant (VPP) integrates distributed energy (such as wind turbines, photovoltaic generators, etc.), controllable loads and energy storage equipment through advanced control technology to rationally optimize the scheduling of the system's power generation and demand sides, achieve the goals of full consumption of renewable energy, peak shaving and valley filling, and carbon reduction, increase power supply reliability, and reduce the impact of load side, wind power generation, photovoltaic power generation and electricity price uncertainty on the power grid. It is one of the important measures to solve the consumption of renewable energy and the development of a high proportion of clean energy.

[0003] In order to ensure the economical and efficient operation of virtual power plants and improve the utilization rate of clean energy, accurate grid load forecasting is of great significance. In the existing technology, machine learning-based methods are widely used in energy data analysis of virtual power plants, such as support vector machines, random forests, convolutional neural networks, etc. These methods use a prediction model composed of a single algorithm within the framework of multi-task learning to handle different load forecasting tasks, and usually only tend to extract local correlation or temporal features. However, the environment in which wind turbines and photovoltaic generators in the virtual power plant system model are complex and changeable, and their loads are affected by multiple factors. The prediction model composed of a single algorithm cannot effectively utilize the characteristic information of load data samples, and the prediction accuracy is limited.

[0004] At the same time, in terms of the scheduling of virtual power plants, due to the inherent randomness and volatility of wind power and photovoltaic output, the calling strategy of virtual power plants needs to consider more factors. The calling strategies generated by existing methods often cannot fully utilize the clean energy power generation of virtual power plants.

[0005] In view of this, the present invention proposes a virtual power plant optimization scheduling method based on a composite network prediction model to solve the above technical problems. Summary of the invention

[0006] The purpose of the present invention is to provide a virtual power plant optimization scheduling method based on a composite network prediction model to solve the problems in the prior art that the prediction model composed of a single algorithm only tends to extract local correlation or temporal features, the load prediction accuracy is not high, and the existing scheduling strategy is difficult to fully utilize the clean energy power generation of the virtual power plant.

[0007] To achieve the above object, the present invention provides a virtual power plant optimization scheduling method based on a composite network prediction model. The technical solution of the present invention includes the following steps:

[0008] S1. Data collection and processing stage:

[0009] Collecting a historical data set of the operation of the virtual power plant system model, preprocessing the historical data set to obtain a first historical data set, and using a composite network-based prediction model to perform a day-ahead prediction on the first historical data set to obtain day-ahead prediction data;

[0010] S2, day-ahead optimization scheduling stage:

[0011] Considering the step-by-step carbon trading mechanism and the comprehensive demand response of thermal power, a day-ahead optimal dispatch model is established. The minimization of the operating cost of the virtual power plant is taken as the objective function. The data of the day-ahead forecast data set is substituted into the objective function to solve the day-ahead optimal dispatch plan.

[0012] S3, intraday optimization scheduling stage:

[0013] Based on the day-ahead optimal scheduling plan and considering the prediction deviation penalty, an intraday optimal scheduling model is established. The objective function is to minimize the sum of the operating cost and the prediction deviation penalty. The intraday optimal scheduling model is solved to obtain the intraday scheduling result. The rolling optimization method is used to optimize the intraday scheduling result to obtain the intraday optimal scheduling plan.

[0014] Preferably, in the above technical solution, the step S1 also includes establishing a virtual power plant system model including a cogeneration unit, an electric boiler, a photovoltaic unit, a wind turbine unit, a controllable electric thermal load, and an electric thermal energy storage device, and the virtual power plant system model generates an operation history data set when it is running.

[0015] Preferably, in the above technical solution, in step S1, the preprocessing process of the historical data set includes:

[0016] Data screening: screening the collected historical data sets to select data values ​​with specific prediction steps;

[0017] Data smoothing, due to the existence of data collection errors and outliers, a simple exponential smoothing method is used to smooth the data in the historical data set after data screening. The formula is:

[0018]

[0019] In the formula, x i+1 、x i Represent the smoothed values ​​of the i-th and (i+1)-th samples respectively; is the i-th sample value before smoothing; η is the smoothing coefficient, which ranges from 0 to 1;

[0020] Data normalization is to perform maximum and minimum normalization on the data variable values ​​in the historical data set after data smoothing, map the data in the historical data set to between (0-1), and output the first historical data set. The normalization formula is:

[0021]

[0022] In the formula, x max and x min are the maximum and minimum smoothed values ​​of the variable, y i is the normalized value of the ith sample.

[0023] Preferably, in the above technical solution, the composite network-based prediction model includes:

[0024] A residual intensive module, wherein the first historical data set is input from a data input terminal of the module, and the module forms a second historical data set with data spatial features by extracting spatial features of each data in the first historical data set;

[0025] A bidirectional long short-term memory network module, wherein a data input terminal of the bidirectional long short-term memory network module is connected to a data output terminal of the residual dense module, and a third historical data set with data time series features is formed by extracting time series features of each data in the second historical data set;

[0026] A scaled dot product attention module, wherein the data input end of the scaled dot product attention module is connected to the data output end of the bidirectional long short-term memory network module, and the day-ahead data prediction data is obtained by weighted integration of the data of the third historical data set output by the bidirectional long short-term memory network module.

[0027] Preferably, in the above technical solution, in step S2, the tiered carbon trading mechanism adopts the baseline method to determine the carbon trading quota. When the actual carbon emissions of the virtual power plant system model are less than the carbon trading quota, the remaining carbon emissions can be sold in the external carbon trading market for profit. Otherwise, the virtual power plant system model needs to purchase the corresponding carbon emissions in the external carbon trading market. The calculation formula of the tiered carbon trading cost is:

[0028]

[0029] In the formula, is the step-wise carbon trading cost at time t; λ is the predicted value of the carbon trading base price on the previous day; l is the length of the carbon emission interval; α is the price growth rate; is the predicted value of carbon emissions of the virtual power plant system model at time t; Carbon allowances owned by the virtual power plant system model; is the predicted value of carbon emissions of the virtual power plant system model at time t;

[0030] The cost calculation formula for the thermal power integrated demand response is:

[0031]

[0032] In the formula, is the comprehensive demand response cost of thermal power at time t, They are respectively the electricity load price-type demand response cost, the heat load price-type demand response cost, the electricity load incentive-type demand response cost, and the heat load incentive-type demand response cost at time t.

[0033] Preferably, in the above technical solution, in step S2, the calculation formula with minimization of the virtual power plant operating cost as the objective function is:

[0034]

[0035] In the formula, The operating cost of virtual power plants for day-ahead dispatch; They are the cost of purchasing natural gas, unit maintenance cost, electric energy storage maintenance cost, thermal energy storage maintenance cost, thermal power integrated demand response dispatch cost, and electricity purchase and sales cost at time t.

[0036] The constraints of the objective function of the day-ahead optimization model are constraints based on price demand response, constraints based on incentive demand response, output constraints of cogeneration units, output constraints of electric boilers, constraints of energy storage devices, power purchase and sales constraints, and day-ahead power balance constraints;

[0037] The day-ahead forecast data and constraint conditions are input into the objective function for calculation to obtain a day-ahead optimized scheduling plan.

[0038] Preferably, in the above technical solution, the intraday optimization model takes into account the stepped carbon trading mechanism and the comprehensive demand response for heat and electricity.

[0039] Preferably, in the above technical solution, in step S3, the calculation formula for minimizing the sum of the operating cost of the virtual power plant model and the prediction deviation penalty as the objective function is:

[0040]

[0041] In the formula, is the operating cost of the virtual power plant model within a day, They are the unit operation and maintenance cost, electricity purchase and sales cost, natural gas purchase cost, electric energy storage maintenance cost, thermal energy storage maintenance cost, carbon emission cost, daily demand response cost and output deviation penalty cost at time t within the day;

[0042] The constraints of the objective function of the intraday optimization scheduling model are constraints based on price demand response, constraints based on incentive demand response, output constraints of cogeneration units, output constraints of electric boilers, constraints of energy storage devices, power purchase and sales constraints, and day-ahead power balance constraints;

[0043] The constraint conditions of the objective function of the intraday optimization scheduling model are input into the objective function of the intraday optimization scheduling model for calculation to obtain the intraday scheduling result.

[0044] Preferably, in the above technical solution, in step S3, the process of optimizing the intraday scheduling results using the rolling optimization method includes: according to the intraday scheduling results, using the rolling optimization method, re-optimizing and decomposing the day-ahead optimized scheduling plan to obtain the intraday optimized scheduling plan.

[0045] Compared with the existing technology, the present invention has the following beneficial effects:

[0046] 1. The first historical data set is obtained by preprocessing the operation history data set of the virtual power plant system model, and the data that meets the calculation quality requirements of the subsequent model for the load forecast of the virtual power plant system and the output forecast of each unit is output to ensure the prediction accuracy of the subsequent model. The prediction model based on the composite network includes a residual dense module, a bidirectional long short-term memory network module and a scaled dot product attention module. The spatial features of each data in the first historical data set are extracted by the residual dense module to form a second historical data set with data spatial features. The temporal features of each data in the second historical data set are extracted by the bidirectional long short-term memory network module to form a third historical data set with data temporal features. The data in the third historical data set output by the bidirectional long short-term memory network module are weighted and integrated by the scaled dot product attention module to obtain the day-ahead prediction data. The spatial features, temporal features and key information of each data in the first historical data set are learned and captured, which improves the prediction accuracy of the day-ahead prediction data.

[0047] 2. By utilizing the step-by-step carbon trading mechanism to fully mobilize the enthusiasm of virtual power plants for low-carbon emission reduction, and taking into account the thermal load and power load, virtual power plants can allocate resources and utilize energy more accurately and efficiently. Combining the scheduling methods of the day-ahead and intraday time scales, the power fluctuations caused by the day-ahead forecast error can be effectively smoothed to ensure the stability of system operation. The prediction error penalty is added to the objective function of intraday scheduling, and the calculation results of the intraday scheduling model are optimized in a rolling manner, which improves the economic efficiency of system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a system structure diagram of the virtual power plant system model of the present invention.

[0049] Figure 2 It is a flow chart of the present invention.

[0050] Figure 3 It is a flowchart of the data collection and processing stage of the present invention.

[0051] Figure 4 It is a structural diagram of the residual dense module of the present invention.

[0052] Figure 5 It is a structural diagram of the bidirectional long short-term memory network module of the present invention.

[0053] Figure 6 It is a structural diagram of the scaled dot product attention module of the present invention.

[0054] Figure 7 It is a simulation flow chart of virtual power plant system optimization scheduling of the present invention. Specific embodiments

[0055] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the invention is not limited by the specific implementation modes.

[0056] refer to Figure 1-Figure 7 The specific implementation steps of the technical solution of the present invention are as follows:

[0057] S1. Establishing the virtual power plant model:

[0058] S2, data collection and processing stage:

[0059] Establish a virtual power plant system model including cogeneration units, electric boilers, photovoltaic units, wind turbines, controllable electric and thermal loads, and electric and thermal energy storage equipment. When the virtual power plant system model is running, it generates an operation history data set;

[0060] Specifically, the virtual power plant system model is also connected to the power grid market, the tiered carbon trading market, and the natural gas market to obtain real-time price data from the above markets.

[0061] Collecting a virtual power plant system model operation historical data set, pre-processing the historical data to obtain a first historical data set, and using a composite network-based prediction model to perform day-ahead prediction on the first historical data set to obtain day-ahead prediction data;

[0062] refer to Figure 3Specifically, the first historical data set includes: carbon emissions / carbon base price historical data set, wind power generation historical data set, photovoltaic power generation historical data set, natural gas price historical data set, user thermal power load historical data set and electricity purchase and sales price / electric power historical data set.

[0063] The data preprocessing process includes:

[0064] Data screening: screening the collected historical data sets to select data values ​​with specific prediction steps;

[0065] Data smoothing, due to the existence of data collection errors and outliers, a simple exponential smoothing method is used to smooth the data in the historical data set after data screening. The formula is:

[0066]

[0067] In the formula, x i+1 、x i Represent the smoothed values ​​of the i-th and (i+1)-th samples respectively; is the i-th sample value before smoothing; η is the smoothing coefficient, which ranges from 0 to 1;

[0068] Data normalization is to perform maximum and minimum normalization on the data variable values ​​in the historical data set after data smoothing, map the data in the historical data set to between (0-1), eliminate the influence of the dimension, and output the first historical data set. The normalization formula is:

[0069]

[0070] In the formula, x max and x min are the maximum and minimum smoothed values ​​of the variable, y i is the normalized value of the ith sample.

[0071] After the above preprocessing, the historical data set of the virtual power plant system model obtains the first historical data set. The obtained first historical data set effectively reduces error data and outliers, improves data quality, and outputs data that meets the subsequent model's load forecasting and each unit output forecasting calculation quality requirements for the virtual power plant system, thereby ensuring the prediction accuracy of the subsequent model.

[0072] Furthermore, the prediction model based on the composite network includes:

[0073] A residual intensive module, wherein the first historical data set is input from a data input terminal of the module, and a second historical data set with data spatial features is formed by extracting the spatial features of each data in the first historical data set;

[0074] refer to Figure 4The residual dense module is composed of multiple 3×3 and 1×1 convolutional layer neural networks. Specifically, the shallow spatial features of each data in the first historical data set are obtained by using a 3×3 convolutional layer, and then two groups of 1×1 and 3×3 convolutional layers are reused with jump connections to extract deep spatial features. The channel learning of the 1×1 convolutional layer reduces the amount of calculation, improves the model operation efficiency, and increases the network depth. The shallow spatial features and deep spatial features are fused by dense connection. The fused spatial features are operated by a 1×1 convolutional layer to enhance the nonlinear expression ability of the network. Finally, the maximum pooling is used to sample and further extract the spatial features of each data in the first historical data set to form a second historical data set with data spatial features.

[0075] A bidirectional long short-term memory network module, wherein a data input terminal of the bidirectional long short-term memory network module is connected to a data output terminal of the residual dense module, and a third historical data set with data time series features is formed by extracting time series features of each data in the second historical data set;

[0076] refer to Figure 5 Specifically, the bidirectional long short-term memory network module consists of two long short-term (LSTM) network units in opposite directions. Each LSTM network unit can selectively memorize, forget, update and transmit information through the input gate, forget gate and output gate. The module can be respectively from the forward h i1 and reverse h i2 Two hidden state sequences are calculated and concatenated to form the output sequence (h i =h i1 +h i2 ). Since the information in both the past and future directions is taken into account at the same time, the LSTM network can better capture the dependencies among the data sequences in the second historical data set and provide more comprehensive semantic understanding and expression capabilities.

[0077] A scaled dot product attention module, wherein a data input end of the scaled dot product attention module is connected to a data output end of a bidirectional long short-term memory network module, and a day-ahead data prediction data is obtained by weighted integration of data of a third historical data set output by the bidirectional long short-term memory network module.

[0078] refer to Figure 6 Specifically, the scaled dot product attention module performs weighted integration on the output sequence of each time step of the bidirectional long short-term memory network module, so that the scaled dot product attention module focuses on the feature information that has important contributions to the predicted value, and improves the perception of key information in the third historical data set. The scaled dot product attention model adopts the Query-Key-Value mode, and the model operation is When the output sequence of the bidirectional long-short time memory network module at time t is H = (h1, h2, ..., hT ), H is mapped into a query vector Q = w through three linear transformations respectively. Q H, key vector K = w K H and value vector V = w V H. Combine the query vector, key vector and its dimension to get the annotation The attention weight vector z of sequence H i =α i ·v i .

[0079] During the specific implementation, the prediction model sets the prediction step of the day-ahead data to 1 hour, and outputs the day-ahead prediction data for the next 24 moments. The day-ahead prediction data include: the day-ahead prediction value of carbon base price, the day-ahead prediction value of wind power generation, the day-ahead prediction value of photovoltaic power generation, the day-ahead prediction value of natural gas price, the day-ahead prediction value of user thermal power load, and the day-ahead prediction value of purchase and sale electricity price.

[0080] Through the above steps, the spatial features, temporal features and key information of each data in the operation history data set of the virtual power plant system model are learned and captured. By adopting a deeper and more efficient convolutional network structure, the residual dense module improves the ability of the composite network model to extract the global features of the data in the virtual power plant operation history data set, and improves the prediction accuracy of the day-ahead prediction data. It effectively solves the problem that the prediction model composed of a traditional single algorithm only tends to extract local correlation or temporal features of the operation history data of the virtual power plant system, while the environment in which the wind turbines and photovoltaic generators in the virtual power plant system model are complex and changeable, and their loads are affected by many factors. The prediction model composed of a single algorithm cannot effectively utilize the feature information of the data samples of the virtual power plant system model, resulting in low prediction accuracy.

[0081] S2, day-ahead optimization scheduling stage:

[0082] Considering the step-by-step carbon trading mechanism and the comprehensive demand response of thermal power, a day-ahead optimal dispatch model is established. The minimization of the operating cost of the virtual power plant is taken as the objective function. The data of the day-ahead forecast data set is substituted into the objective function to solve the day-ahead optimal dispatch plan.

[0083] Specifically, the tiered carbon trading mechanism adopts the baseline method to determine the carbon trading quota. The tiered carbon trading mechanism divides multiple purchase intervals. When the actual carbon emissions of the virtual power plant system model are less than the carbon trading quota, the remaining carbon emissions can be sold in the external carbon trading market for profit. Otherwise, the virtual power plant system model needs to purchase the corresponding carbon emissions in the external carbon trading market. The calculation formula is:

[0084]

[0085] In the formula, λ is the predicted value of the carbon trading base price on the previous day; l is the length of the carbon emission interval; α is the price growth rate; is the predicted value of carbon emissions of the virtual power plant system model at time t; Carbon allowances owned by the virtual power plant system model; is the predicted value of carbon emissions of the virtual power plant system model at time t;

[0086] The cost calculation formula for comprehensive heat and power demand response is:

[0087]

[0088] In the formula, is the comprehensive demand response cost of thermal power at time t, They are respectively the electricity load price-type demand response cost, the heat load price-type demand response cost, the electricity load incentive-type demand response cost, and the heat load incentive-type demand response cost at time t.

[0089] Furthermore, the calculation formula with minimization of the virtual power plant operating cost as the objective function is:

[0090]

[0091] In the formula, The day-ahead VPP operation cost; They are the cost of purchasing natural gas, unit maintenance cost, electric energy storage maintenance cost, energy storage maintenance cost, thermal power integrated demand response dispatch cost, electricity purchase and sales cost, and tiered carbon trading cost at time t.

[0092] Furthermore, the constraints of the objective function of the day-ahead optimization model are constraints based on price demand response, constraints based on incentive demand response, output constraints of cogeneration units, output constraints of electric boilers, constraints of energy storage devices, constraints on purchased and sold power, and day-ahead power balance constraints;

[0093] By inputting the day-ahead forecast data and the constraints of the objective function of the above-mentioned day-ahead optimization model into the objective function, the day-ahead optimization model is solved to obtain the day-ahead optimization scheduling plan to minimize the operating cost of the virtual power plant.

[0094] S3: Intraday optimization and scheduling stage

[0095] Based on the day-ahead optimal scheduling plan and considering the prediction deviation penalty, an intraday optimal scheduling model is established. The objective function is to minimize the sum of the operating cost and the prediction deviation penalty. The intraday optimal scheduling model is solved to obtain the intraday scheduling result. The rolling optimization method is used to optimize the intraday scheduling result to obtain the intraday optimal scheduling plan.

[0096] Specifically, the solution to the intraday optimization scheduling model takes the minimum sum of the operating cost of the virtual power plant model and the prediction deviation penalty as the objective function, and the calculation formula is:

[0097]

[0098] In the formula, is the operating cost of the virtual power plant model within a day, They are, respectively, the unit operation and maintenance cost, electricity purchase and sales cost, natural gas purchase cost, electric energy storage maintenance cost, thermal energy storage maintenance cost, carbon emission cost, daily demand response cost, and output deviation penalty cost at time t within the day.

[0099] Similar to the day-ahead dispatch stage, the constraints of the objective function of the intraday optimization dispatch model are price-based demand response constraints, incentive-based demand response constraints, cogeneration unit output constraints, electric boiler output constraints, energy storage device constraints, power purchase and sales constraints, and day-ahead power balance constraints.

[0100] Finally, the constraints of the objective function of the intraday optimization scheduling model are input into the objective function of the intraday optimization scheduling model to calculate and obtain the intraday scheduling result. According to the intraday scheduling result, the rolling optimization method is used to re-optimize and decompose the day-ahead optimization scheduling plan to obtain the intraday optimization scheduling plan. Specifically, the rolling optimization method is used to optimize the day-ahead optimization scheduling plan with a rolling optimization step of 15 minutes to obtain the intraday optimization scheduling result until the rolling optimization time is greater than 24 hours.

[0101] In summary, in terms of load data prediction of the virtual power plant system model, the present invention obtains a first historical data set by preprocessing the operating history data of the virtual power plant system model, and outputs data that meets the calculation quality requirements of the subsequent model for the load prediction of the virtual power plant system and the output prediction of each unit, thereby ensuring the prediction accuracy of the subsequent model. The prediction model based on the composite network includes a residual intensive module, a bidirectional long short-term memory network module, and a scaling dot product attention module. The spatial features of each data in the first historical data set are extracted by the residual intensive module to form a second historical data set with data spatial features. The temporal features of each data in the second historical data set are extracted by the bidirectional long short-term memory network module to form a third historical data set with data temporal features. The data in the third historical data set output by the bidirectional long short-term memory network module are weighted and integrated by the scaling dot product attention module to obtain the day-ahead prediction data. The spatial features, temporal features, and key information of each data in the first historical data set are learned and captured, thereby improving the prediction accuracy of the day-ahead prediction data.

[0102] In terms of dispatching, by utilizing the step-by-step carbon trading mechanism to fully mobilize the enthusiasm of virtual power plants for low-carbon emission reduction, and taking into account the thermal load and power load, virtual power plants can allocate resources and utilize energy more accurately and efficiently. Combining the dispatching methods of the day-ahead and intraday time scales, the power fluctuations caused by the day-ahead forecast error can be effectively smoothed to ensure the stability of system operation. The forecast error penalty is added to the objective function of intraday dispatching, and the calculation results of the intraday dispatching model are optimized in a rolling manner, which improves the economic efficiency of system operation.

Claims

1. A virtual power plant optimization scheduling method based on a composite network prediction model, characterized in that: The steps include: S1. Data collection and processing stage: Collecting a historical data set of the operation of the virtual power plant system model, preprocessing the historical data set to obtain a first historical data set, and using a composite network-based prediction model to perform a day-ahead prediction on the first historical data set to obtain day-ahead prediction data; S2, day-ahead optimization scheduling stage: Considering the step-by-step carbon trading mechanism and the comprehensive demand response of thermal power, a day-ahead optimal dispatch model is established. The minimization of the operating cost of the virtual power plant is taken as the objective function. The data of the day-ahead forecast data set is substituted into the objective function to solve the day-ahead optimal dispatch plan. S3, intraday optimization scheduling stage: Based on the day-ahead optimal scheduling plan and considering the prediction deviation penalty, an intraday optimal scheduling model is established. The objective function is to minimize the sum of the operating cost and the prediction deviation penalty. The intraday optimal scheduling model is solved to obtain the intraday scheduling result. The rolling optimization method is used to optimize the intraday scheduling result to obtain the intraday optimal scheduling plan.

2. The virtual power plant optimization scheduling method based on the composite network prediction model according to claim 1 is characterized in that: The step S1 also includes establishing a virtual power plant system model including a cogeneration unit, an electric boiler, a photovoltaic unit, a wind turbine unit, a controllable electric thermal load, and an electric thermal energy storage device. When the virtual power plant system model is running, it generates an operation history data set.

3. The virtual power plant optimization scheduling method based on the composite network prediction model according to claim 2 is characterized in that: In step S1, the preprocessing process of the historical data set includes: Data screening: screening the collected historical data sets to select data values ​​with specific prediction steps; Data smoothing uses a simple exponential smoothing method to smooth the data in the historical data set after data screening. The formula is: In the formula, x i+1 、x i Represent the smoothed values ​​of the i-th and (i+1)-th samples respectively; is the i-th sample value before smoothing; η is the smoothing coefficient, which ranges from 0 to 1; Data normalization is to perform maximum and minimum normalization on the data variable values ​​in the historical data set after data smoothing, map the data in the historical data set to between (0-1), and output the first historical data set. The normalization formula is: In the formula, x max and x min are the maximum and minimum smoothed values ​​of the variable, y i is the normalized value of the ith sample.

4. The virtual power plant optimization scheduling method based on the composite network prediction model according to claim 3 is characterized in that: The composite network-based prediction model includes: A residual intensive module, wherein the first historical data set is input from a data input terminal of the module, and the module forms a second historical data set with data spatial features by extracting spatial features of each data in the first historical data set; A bidirectional long short-term memory network module, wherein a data input terminal of the bidirectional long short-term memory network module is connected to a data output terminal of the residual dense module, and a third historical data set with data time series features is formed by extracting time series features of each data in the second historical data set; A scaled dot product attention module, wherein the data input end of the scaled dot product attention module is connected to the data output end of the bidirectional long short-term memory network module, and the day-ahead forecast data is obtained by weighted integration of the data of the third historical data set output by the bidirectional long short-term memory network module.

5. The virtual power plant optimization scheduling method based on a composite network prediction model according to claim 1 is characterized in that: In step S2, the tiered carbon trading mechanism adopts the baseline method to determine the carbon trading quota. When the actual carbon emissions of the virtual power plant system model are less than the carbon trading quota, the remaining carbon emissions can be sold in the external carbon trading market for profit. Otherwise, the virtual power plant system model needs to purchase the corresponding carbon emissions in the external carbon trading market. The calculation formula of the tiered carbon trading cost is: In the formula, is the step-wise carbon trading cost at time t; λ is the predicted value of the carbon trading base price on the previous day; l is the length of the carbon emission interval; α is the price growth rate; is the predicted value of carbon emissions of the virtual power plant system model at time t; Carbon allowances owned by the virtual power plant system model; is the predicted value of carbon emissions of the virtual power plant system model at time t; The cost calculation formula for the thermal power integrated demand response is: In the formula, is the comprehensive demand response cost of thermal power at time t, They are respectively the electricity load price-type demand response cost, the heat load price-type demand response cost, the electricity load incentive-type demand response cost, and the heat load incentive-type demand response cost at time t.

6. A virtual power plant optimization scheduling method based on a composite network prediction model according to claim 5, characterized in that: In step S2, the calculation formula with minimization of the virtual power plant operating cost as the objective function is: In the formula, The operating cost of virtual power plants for day-ahead dispatch; They are the cost of purchasing natural gas, unit maintenance cost, electric energy storage maintenance cost, thermal energy storage maintenance cost, thermal power integrated demand response dispatch cost, and electricity purchase and sales cost at time t. The constraints of the objective function of the day-ahead optimization model are constraints based on price demand response, constraints based on incentive demand response, output constraints of cogeneration units, output constraints of electric boilers, constraints on energy storage devices, constraints on purchased and sold power, and day-ahead power balance constraints; The day-ahead forecast data and constraint conditions are input into the objective function for calculation to obtain a day-ahead optimized scheduling plan.

7. According to the virtual power plant optimization scheduling method based on the composite network prediction model in claim 1, in step S3, the intraday optimization model takes into account the stepped carbon trading mechanism and the comprehensive demand response of heat and power.

8. According to the virtual power plant optimization scheduling method based on the composite network prediction model of claim 7, in the step S3, the calculation formula with the minimum sum of the operating cost of the virtual power plant model and the prediction deviation penalty as the objective function is: In the formula, is the operating cost of the virtual power plant model within a day, They are the unit operation and maintenance cost, electricity purchase and sales cost, natural gas purchase cost, electric energy storage maintenance cost, thermal energy storage maintenance cost, carbon emission cost, daily demand response cost and output deviation penalty cost at time t within the day; The constraints of the objective function of the intraday optimization scheduling model are constraints based on price demand response, constraints based on incentive demand response, output constraints of cogeneration units, output constraints of electric boilers, constraints of energy storage devices, power purchase and sales constraints, and day-ahead power balance constraints; The constraint conditions of the objective function of the intraday optimization scheduling model are input into the objective function of the intraday optimization scheduling model for calculation to obtain the intraday scheduling result.

9. According to the virtual power plant optimization scheduling method based on the composite network prediction model of claim 8, in step S3, the process of optimizing the intraday scheduling results by using the rolling optimization method includes: According to the intraday scheduling result, a rolling optimization method is adopted to re-optimize and decompose the day-ahead optimized scheduling plan to obtain an intraday optimized scheduling plan.