Method for quickly quantifying scheduling uncertainty in extreme scene based on WGAN-GP and GPR-Q

Abstract

The invention relates to a WGAN-GP and GPR-Q-based scheduling uncertainty rapid quantification method in an extreme scene, comprising the following steps: obtaining a historical data set of wind, light and user side electric heating load and a corresponding meteorological data set, and preprocessing the data to determine an extreme scene set; building a WGAN-GP model to generate extreme scene data, and expanding an extreme scene set; establishing and solving an integrated energy system scheduling model in an extreme scene; importing the meteorological data set and the unit output into a GPR model to obtain a scheduling result; according to the method, a more reasonable and effective scheduling scheme can be provided for scheduling personnel, and the problem of rapid and accurate scheduling of the integrated energy system in an extreme scene can be well solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of optimized scheduling, and in particular relates to a method for quickly quantifying scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q. Background Art

[0002] The optimal dispatch of integrated energy systems is a method that realizes the unified dispatch and optimized operation of multiple energy flows by controlling the conversion between different energy sources through coupling devices. It is of great significance to expand the system operation boundary, promote the consumption of renewable energy, and realize the safe and economic operation of the system. However, global environmental changes have led to frequent extreme events, which has brought challenges to the optimal dispatch of integrated energy systems. The historical sample data of extreme events are small and lack typicality, which makes the dispatch prediction in extreme scenarios have large errors, thereby reducing the dispatch quality and seriously affecting the power supply reliability of the power system. Therefore, what means should be taken to quickly and accurately respond to extreme scenarios has become an urgent problem that the integrated energy system needs to solve.

[0003] Patent CN117674298B proposes a load dispatching method that takes into account the uncertainty of wind and solar power stations, and realizes load dispatching by estimating historical uncertainty. However, in extreme scenarios, the existing technology relies on historical data and known operating modes, and cannot effectively handle situations that have never been experienced. These situations exceed the prediction capabilities of traditional models; Patent CN117993640A proposes a virtual power plant optimization dispatching method based on wind and solar uncertainty and electricity carbon trading, and inputs the wind and solar output in typical scenarios into the final virtual power plant dispatching model to obtain the predicted dispatching results. However, in extreme scenarios, the load power of the integrated energy system often shows more frequent fluctuations, and it is difficult to predict the accurate point value. Only the load power interval at different confidence levels can be predicted to describe the fluctuation range and uncertainty of the load power of the integrated energy system. Therefore, the present invention provides a method for rapid quantification of dispatching uncertainty in extreme scenarios based on WGAN-GP and GPR-Q. Summary of the invention

[0004] To solve the above problems, the present invention discloses a method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q. This method effectively avoids wind power and photovoltaic output prediction errors by using machine learning algorithms, and ensures that the integrated energy system has sufficient reserves to respond to possible extreme output changes of new energy in a timely manner.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A method for quickly quantifying scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q of the present invention comprises the following steps:

[0007] Step S1, collecting wind and solar meteorological and load parameters to form a historical data set, selecting scenes that exceed the upper limit of the confidence interval, the lower limit of the confidence interval and large fluctuations as extreme scenes, using the maximum information coefficient to screen and obtain typical meteorological features with a high correlation with load output, and merging the data to establish a real sample data set for extreme scenes;

[0008] Step S2, based on the extreme scene real sample data set, establish an improved generative adversarial network model with gradient penalty, input the extreme scene real sample data set as a training set into the WGAN-GP model for training, and after the training is completed, obtain the enhanced extreme scene data set in batches from the model generator;

[0009] Step S3, constructing a comprehensive energy system under extreme scenarios using the enhanced extreme scenario dataset;

[0010] Step S4, taking the minimum operating cost of the integrated energy system under extreme scenarios as the objective function, bringing the enhanced extreme scenario data set into the calculation, obtaining the scheduling result of the integrated energy system under extreme scenarios, and incorporating it into the corresponding enhanced extreme scenario data set to obtain the GPR-Q model sample set;

[0011] Step S5, define the independent variables and dependent variables in the Gaussian process regression model sample set, establish a GPR-based scheduling uncertainty quantification system under extreme scenarios based on the GPR-Q model sample set obtained in step 4, divide the GPR-Q model sample set into a training set and a test set, use the enhanced extreme scenario data set as input, and the output electric power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios as output, and train the GPR-Q model to achieve scheduling uncertainty quantification.

[0012] Furthermore, in step S1, a real sample data set of extreme scenarios is constructed, including extremely high output, extremely low output, and extremely fluctuating output scenarios; after obtaining the historical data set, an interval estimation is performed on the load parameters in the historical data set at a confidence level of 1-α, and the calculation of the upper bound of the confidence interval is expressed as:

[0013]

[0014] The calculation of the lower bound of the confidence interval is expressed as:

[0015]

[0016] Among them, P Uα is the upper bound of the confidence interval, P Lα is the lower bound of the confidence interval, is the mean output value of the load parameter, σ is the standard deviation of the load parameter, Z α / 2 is the confidence factor at the significance level α;

[0017] Select 1-α confidence level as 95% confidence level. When there are wind power and photovoltaic output scenarios that exceed the upper and lower boundaries of the confidence interval, they are considered as low-probability extreme events. The scenarios that exceed the upper boundary of the confidence interval are extracted as extreme high output scenarios, and the scenarios that exceed the lower boundary of the confidence interval are extracted as extreme low output extreme scenarios. Based on the upper and lower boundaries of the confidence interval under the target confidence level, calculate the extreme fluctuation range of wind power and photovoltaic output, and select the extreme fluctuation output scenario with the largest hourly fluctuation range from the extreme fluctuation range.

[0018] The data of extremely high output scenarios, extremely low output scenarios, and extremely fluctuating output scenarios are marked, and the wind and solar meteorological data of the corresponding scenarios are added to the historical data set. The wind and solar meteorological data include temperature, air pressure, humidity, wind speed, etc. The wind and solar meteorological data are screened using the maximum information coefficient to obtain wind and solar meteorological features with high correlation with the corresponding extreme load output, forming a real sample data set for extreme scenarios.

[0019] The maximum information coefficient MIC is used to analyze the correlation between wind, solar, meteorological and load parameters. The value range of MIC is [0,1], where 0 indicates no correlation and 1 indicates complete correlation. The larger the value, the stronger the correlation between the two variables. Its calculation formula is:

[0020]

[0021]

[0022] Among them, m is the wind and solar meteorological variable, n is the load parameter variable, I(m,n) is the mutual information coefficient of m and n variables, p(m,n) is the joint probability of m and n variables, p(m) is the probability of m variable, p(n) is the probability of n variable, MIC(m,n) is the maximum information coefficient of m and n variables, a and b are the number of grids divided in the m and n directions, that is, grid distribution; B is a constant, set to 0.6 power of the data amount.

[0023] Further, in step S2, the enhanced extreme scenario data set is obtained in batches using the improved generative adversarial network model with gradient penalty, and a WGAN-GP model needs to be established. The WGAN-GP model includes a generator network, a discriminator network, and a penalty mechanism; wherein the generator network and the discriminator network both use fully connected neural networks, and the input of the generator network is random noise that obeys a standard normal distribution, and the random noise is output as data of a specified size after linear transformation of the hidden layer of the generator network; the input of the discriminator network consists of data generated by the generator network and a historical data set, and after linear transformation of the hidden layer of the discriminator network, the output is a matrix of a specified size, and the matrix represents the probability score that the generated data is real data;

[0024] The penalty mechanism is specifically as follows:

[0025]

[0026] Among them, E represents mathematical expectation, D represents discriminator, and G represents generator. Represents a single specific real data, express Sampling from real data, z represents random noise that follows a standard normal distribution, sampled from the noise set pz(z); represents the judgment result of the discriminator on the real data, G(z) represents the data generated by the generator, and D(G(z)) represents the judgment result of the discriminator on the generated data; λ represents the penalty term coefficient; ▽ represents the differential operator, p(x) represents the middle area formed by any point on the line connecting any two points in the real data space and the real data space, and x is sampled in this area; ▽ x D(x) represents the derivation of the discriminator's discrimination result D(x) in p(x); p represents the norm.

[0027] The extreme scenario real sample dataset is used as a training set to input the WGAN-GP model to train the network, which specifically includes the following steps:

[0028] S2-1 inputs random noise that obeys the standard normal distribution into the generator to obtain data;

[0029] S2-2 inputs the training set and the data generated by the generator in step S2-1 into the discriminator network for training, trains n_D times, and performs gradient updates according to the loss function. The loss function of the discriminator is:

[0030]

[0031] Among them, x i,new represents the training set, N represents the number of samples, λ×gradient_penalty represents the penalty mechanism, z i is a random noise that follows a standard normal distribution, G(z i ) represents the data generated by the generator, D(G(z i )) represents the judgment result of the discriminator on the generated data, D(x i,new ) represents the judgment result of the discriminator on the training set;

[0032] S2-3 trains the generator by inputting random noise that follows a standard normal distribution into the generator; the loss function of the generator is:

[0033]

[0034] The discriminator is trained n_D times and the generator is trained n_G times;

[0035] S2-4 uses back-propagation and gradient descent algorithms to update parameters, record the loss value, and return to S2-1 for iteration until the number of cycles reaches the set value, and outputs the sample data generated by the step generator in the last iteration to obtain an enhanced extreme scenario data set.

[0036] Furthermore, in step S3, the dispatching data needs to be obtained through calculation of the comprehensive energy system under extreme scenarios, and the system includes the following equipment models:

[0037] (1) Cogeneration unit model

[0038]

[0039]

[0040] Among them, t is the scheduling time; and are the heat production rate and gas consumption of the CHP unit respectively; and η MT,t are the power generation and power generation efficiency of CHP respectively; η L,t is the heat loss rate; Hg is the calorific value of natural gas, which is 9.88kWh / m3.

[0041] (2) Gas boiler model

[0042]

[0043] in, is the thermal power output of the gas boiler; η GB For gas boiler efficiency; is the gas consumption at time t.

[0044] (3) Heat pump model

[0045]

[0046] Where: Represents the electrical power input to the heat pump; η HP is the heating efficiency of the heat pump; Output thermal power for the heat pump.

[0047] (4) Model constraints

[0048] Electricity, heat and gas energy supply and demand balance constraints:

[0049]

[0050]

[0051]

[0052]

[0053] in, They are the charging and discharging power of the battery at time t respectively; They are respectively the charging and releasing power of the heat storage tank at time t; They are the electricity and gas purchases of IES at time t respectively; are the reducible load and the shiftable load of IES at time t respectively; are the heat-to-electricity and electricity-to-heat loads of IES at time t respectively; are the electrical load and thermal load of IES at time t respectively.

[0054] Furthermore, in step S4, a sample set of a Gaussian process regression model is constructed, and the dispatching data of the integrated energy system under the extreme scenario constructed in step 3 needs to be calculated. The system objective function is:

[0055] minC=C buy +C op +C co2

[0056] Where C is the IES operating cost, minC is the minimum IES operating cost; C buy C is the cost of purchasing energy gas; op Operation and maintenance costs of each unit; C CO2 For carbon trading costs;

[0057]

[0058] In the formula, and are the electricity purchase and sale prices of the power grid at time t; and To purchase and sell electrical power; is the natural gas price at time t; η GH,CHP and η GB are the thermal energy conversion rates of CHP and GB, respectively; is the CHP thermal power; k PV , k WT , k CHP , k GB , k HP , k cell are the unit operating costs of each unit respectively; are the unit operating power of each unit; c is the benchmark price of carbon trading; C grid , C CHP , C GB They are grid purchase, CHP and GB initial carbon quota; Cgrid,a , C CHP,a , C GB,a They are the actual carbon emissions of power purchase from the grid, CHP and GB.

[0059] Furthermore, in step S4, for the established integrated energy system under extreme scenarios, the sample data set generated by the extreme scenarios is brought into the calculation to obtain the scheduling results of the integrated energy system under extreme scenarios. The output electric power of the cogeneration unit is taken as a representative, and combined with the corresponding enhanced extreme scenario data set, sampling is used as the sample set of the Gaussian process regression model.

[0060] Furthermore, in step S5, the independent variables in the GPR-Q model sample set are defined as the meteorological and time data under extreme scenarios, and the dependent variable is the output electric power of the cogeneration unit with corresponding output, so as to establish a rapid quantification model of the scheduling uncertainty of the integrated energy system under extreme scenarios.

[0061] Further, in step S5, the GPR-based scheduling uncertainty quantification system under extreme scenarios divides the GPR-Q model sample set into a training set and a test set, with the enhanced extreme scenario data set as input and the output electric power of the cogeneration unit in the comprehensive energy system scheduling result under extreme scenarios as output:

[0062] The mapping relationship between the enhanced extreme scenario data set and the output power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios is described by the Gaussian process f(X). The parameter value of f(X) is a random variable, and the prior distribution of f(X) is a Gaussian distribution. The GPR-Q training set D = {(x i ,y i )|i=1,2,…,n}=(X,Y), where n is the number of scenes in the GPR-Q training set, X=[x 1 ,x 2 ,…,x n ] T is the input matrix of the combined heat and power unit output power training set in the comprehensive energy system dispatch results under extreme scenarios, Y = [y 1 ,y 2 ,…,y n ] T is the output matrix of the electric power output of the cogeneration unit in the comprehensive energy system dispatch results under extreme scenarios, y i is the dispatch result under scenario i. The training set D constitutes a Gaussian joint release, and its corresponding Gaussian process GP[·] is composed of the dispatch mean function μ(X) and the meteorological data covariance matrix K(x i ,x j ) uniquely defines and establishes the actual Gaussian process model:

[0063]

[0064] In the formula, σ N is the standard deviation of Gaussian distribution; δ ij is the Kroneckerdelta function; I is the identity matrix;

[0065] Select the remaining GPR-Q model sample set as the GPR-Q test set D * , the expression is as follows:

[0066] D * ={(x i ,f(x i ))|i=n+1,n+2,…,n+n *}=(X * ,f * )

[0067] Where, X * 、f * are the input matrix and output label of the GPR-Q test set respectively; n * is the number of scenes contained in the test set; the matrix form of the prior joint Gaussian distribution is:

[0068]

[0069] Where N(,) is Gaussian distribution; K(·,·) is covariance matrix. According to Bayesian theory, the output result f is obtained * The posterior distribution of is:

[0070]

[0071] Mean vector of the output results of the GPR-Q test set and the variance vector cov(f * )for:

[0072]

[0073] cov(f * )=K(X * ,X * )-K(X * ,X)(K(X,X)+σ N 2 I) -1 K(X * ,X)

[0074] On this basis, GPR-Q introduces the quantile loss function L τ , the deviation expression between the output result of the penalty test set and the training target value of the test set is:

[0075]

[0076] in, is the quantile point of the mean vector under a given quantile level τ, and y is the training target value of the GPR-Q test set.

[0077] Therefore, the goal of the GPR-Q model is to minimize the quantile loss function of the output results of the GPR-Q test set, with the enhanced extreme scenario data set as input and the output electric power of the cogeneration unit in the integrated energy system scheduling results under extreme scenarios as output. A scheduling uncertainty quantification system under extreme scenarios based on GPR is established.

[0078] The beneficial effects of the present invention are:

[0079] In step S2, the method inputs the real sample data set of extreme scenarios as a training set into the WGAN-GP model for training. After the training is completed, the enhanced extreme scenario data set is obtained in batches from the model generator. The increase in sample capacity greatly improves the prediction accuracy of machine learning. In step S5, the mapping relationship between the enhanced extreme scenario data set and the output electric power of the cogeneration unit in the integrated energy system scheduling result under the extreme scenario is directly established through the Gaussian process, which effectively avoids the prediction error of wind power and photovoltaic output, shortens the scheduling calculation time, and provides a more reasonable and fast scheduling plan for the dispatching personnel, which can well solve the problem of fast and accurate scheduling of the integrated energy system under extreme scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 A flow chart of scheduling uncertainty quantification of the method of the present invention;

[0081] Figure 2 This is a result graph of meteorological parameters screened using the maximum information coefficient by the method of the present invention.

[0082] Figure 3 It is a prediction interval diagram of the method of the present invention under 95% confidence interval. DETAILED DESCRIPTION

[0083] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0084] The present invention provides a method for quickly quantifying scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q, comprising the following steps: Figure 1 As shown:

[0085] S1, collect wind and solar meteorological and load parameters to form a historical data set, select scenes that exceed the upper limit of the confidence interval, the lower limit of the confidence interval and large fluctuations as extreme scenes, use the maximum information coefficient to screen and obtain typical meteorological characteristics with a high correlation with load output, and merge the data to establish a real sample data set for extreme scenes;

[0086] S2, based on the extreme scene real sample data set, establish an improved generative adversarial network model with gradient penalty, input the extreme scene real sample data set as a training set into the WGAN-GP model for training, and after the training is completed, obtain the enhanced extreme scene data set in batches from the model generator;

[0087] S3, uses the enhanced extreme scenario dataset to build an integrated energy system under extreme scenarios;

[0088] S4, taking the minimum operating cost of the integrated energy system under extreme scenarios as the objective function, bringing the enhanced extreme scenario data set into the calculation, obtaining the dispatch result of the integrated energy system under extreme scenarios, and incorporating it into the corresponding enhanced extreme scenario data set to obtain the GPR-Q model sample set;

[0089] S5, define the independent variables and dependent variables in the Gaussian process regression model sample set, establish a GPR-based scheduling uncertainty quantification system under extreme scenarios based on the GPR-Q model sample set obtained in step 4, divide the GPR-Q model sample set into a training set and a test set, use the enhanced extreme scenario data set as input, and the output electric power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios as output, and train the GPR-Q model to achieve scheduling uncertainty quantification.

[0090] In S1, a real sample data set of extreme scenarios is constructed, including extremely high output, extremely low output, and extremely fluctuating output scenarios. After obtaining the historical data set, the load parameters in the historical data set are estimated at a confidence level of 1-α. The calculation of the upper bound of the confidence interval is expressed as:

[0091]

[0092] The calculation of the lower bound of the confidence interval is expressed as:

[0093]

[0094] Among them, P Uα is the upper bound of the confidence interval, P Lα is the lower bound of the confidence interval, is the mean output value of the load parameter, σ is the standard deviation of the load parameter, Z α / 2 is the confidence factor at the significance level α;

[0095] Select 1-α confidence level as 95% confidence level. When there are wind power and photovoltaic output scenarios that exceed the upper and lower boundaries of the confidence interval, they are considered as low-probability extreme events. The scenarios that exceed the upper boundary of the confidence interval are extracted as extreme high output scenarios, and the scenarios that exceed the lower boundary of the confidence interval are extracted as extreme low output extreme scenarios. Based on the upper and lower boundaries of the confidence interval under the target confidence level, calculate the extreme fluctuation range of wind power and photovoltaic output, and select the extreme fluctuation output scenario with the largest hourly fluctuation range from the extreme fluctuation range.

[0096] The data of extremely high output scenes, extremely low output scenes, and extremely fluctuating output scenes are marked, and the wind and solar meteorological data of the corresponding scenes are added to the historical data set. The wind and solar meteorological data include 29 elements such as temperature, air pressure, humidity, wind speed, and zenith angle. The wind and solar meteorological data are screened using the maximum information coefficient to obtain wind and solar meteorological features with high correlation with the corresponding extreme load output, forming a real sample data set for extreme scenes. The maximum information coefficient MIC is used to analyze the correlation between wind and solar meteorology and load parameters. The value range of MIC is [0,1], where 0 means no correlation and 1 means complete correlation. The larger the value, the stronger the correlation between the two variables. Its calculation formula is:

[0097]

[0098]

[0099] Among them, m is the wind and solar meteorological variable, n is the load parameter variable, I(m,n) is the mutual information coefficient of m and n variables, p(m,n) is the joint probability of m and n variables, p(m) is the probability of m variable, p(n) is the probability of n variable, MIC(m,n) is the maximum information coefficient of m and n variables, a and b are the number of grids divided in the m and n directions, that is, grid distribution; B is a constant, set to 0.6 power of the data amount.

[0100] Elements with importance higher than 50% were screened, and the load parameters were divided into photovoltaic-related parameters and wind power-related parameters. The former included seven elements such as total solar radiation level and zenith angle, and the latter included five elements such as wind speed at 70m from the wind tower. Figure 2 shown.

[0101] In S2, the enhanced extreme scene data set is obtained in batches using the improved generative adversarial network model with gradient penalty, and a WGAN-GP model needs to be established. The WGAN-GP model includes a generator network, a discriminator network, and a penalty mechanism; wherein both the generator network and the discriminator network use fully connected neural networks, and the input of the generator network is random noise that obeys the standard normal distribution, and the random noise is output as data of a specified size after linear transformation of the hidden layer of the generator network; the input of the discriminator network consists of the data generated by the generator network and the historical data set, and after linear transformation of the hidden layer of the discriminator network, the output is a matrix of a specified size, and the matrix represents the probability score that the generated data is the real data;

[0102] The penalty mechanism is specifically as follows:

[0103]

[0104] Among them, E represents mathematical expectation, D represents discriminator, and G represents generator. Represents a single specific real data, express Sampling from real data, z represents random noise that follows a standard normal distribution, sampled from the noise set pz(z); represents the judgment result of the discriminator on the real data, G(z) represents the data generated by the generator, and D(G(z)) represents the judgment result of the discriminator on the generated data; λ represents the penalty term coefficient; ▽ represents the differential operator, p(x) represents the middle area formed by any point on the line connecting any two points in the real data space and the real data space, and x is sampled in this area; ▽ x D(x) represents the derivation of the discriminator's discrimination result D(x) in p(x); p represents the norm.

[0105] The extreme scenario real sample dataset is used as a training set to input the WGAN-GP model to train the network, which specifically includes the following steps:

[0106] S2-1 inputs random noise that obeys the standard normal distribution into the generator to obtain data;

[0107] S2-2 inputs the training set and the data generated by the generator in step S2-1 into the discriminator network for training, trains n_D times, and performs gradient updates according to the loss function. The loss function of the discriminator is:

[0108]

[0109] Among them, x i,new represents the training set, N represents the number of samples, λ×gradient_penalty represents the penalty mechanism, z iis a random noise that follows a standard normal distribution, G(z i ) represents the data generated by the generator, D(G(z i )) represents the judgment result of the discriminator on the generated data, D(x i,new ) represents the judgment result of the discriminator on the training set;

[0110] S2-3 trains the generator by inputting random noise that follows a standard normal distribution into the generator; the loss function of the generator is:

[0111]

[0112] The discriminator is trained n_D times and the generator is trained n_G times;

[0113] S2-4 uses back-propagation and gradient descent algorithms to update parameters, record the loss value, and return to 1 for iteration until the number of cycles reaches the set value, and outputs the sample data generated by the step generator in the last iteration to obtain an enhanced extreme scenario data set.

[0114] In S3, the dispatch data needs to be obtained through the calculation of the comprehensive energy system under extreme scenarios. The system includes the following equipment models:

[0115] (1) Cogeneration unit model

[0116]

[0117]

[0118] Among them, t is the scheduling time; and are the heat production rate and gas consumption of the CHP unit respectively; and η MT ,t are the power generation and power generation efficiency of CHP respectively; η L ,t is the heat loss rate; Hg is the calorific value of natural gas, which is 9.88kWh / m3.

[0119] (2) Gas boiler model

[0120]

[0121] in, is the thermal power output of the gas boiler; η GB For gas boiler efficiency; is the gas consumption at time t.

[0122] (3) Heat pump model

[0123]

[0124] Where: Represents the electrical power input to the heat pump; η HP is the heating efficiency of the heat pump; Output thermal power for the heat pump.

[0125] (4) Model constraints

[0126] Electricity, heat and gas energy supply and demand balance constraints:

[0127]

[0128]

[0129]

[0130]

[0131] in, They are the charging and discharging power of the battery at time t respectively; They are respectively the charging and releasing power of the heat storage tank at time t; They are the electricity and gas purchases of IES at time t respectively; are the reducible load and the shiftable load of IES at time t respectively; are the heat-to-electricity and electricity-to-heat loads of IES at time t respectively; are the electrical load and thermal load of IES at time t respectively.

[0132] In S4, the sample set of the Gaussian process regression model is constructed. It is necessary to calculate the comprehensive energy system dispatch data under the extreme scenario constructed in step 3. The system objective function is:

[0133] min C=C buy +C op +C co2

[0134] Where C is the IES operating cost, minC is the minimum IES operating cost; C buy C is the cost of purchasing energy gas; op Operation and maintenance costs of each unit; C CO2 For carbon trading costs;

[0135]

[0136] In the formula, and are the electricity purchase and sale prices of the power grid at time t; and To purchase and sell electrical power; is the natural gas price at time t; ηGH,CHP and η GB are the thermal energy conversion rates of CHP and GB, respectively; is the CHP thermal power; k PV , k WT , k CHP , k GB , k HP , k cell are the unit operating costs of each unit respectively; are the unit operating power of each unit; c is the benchmark price of carbon trading; C grid , C CHP , C GB They are grid purchase, CHP and GB initial carbon quota; C grid,a , C CHP,a , C GB,a They are the actual carbon emissions of power purchase from the grid, CHP and GB.

[0137] For the established integrated energy system under extreme scenarios, the sample data set generated by extreme scenarios is brought into the calculation to obtain the scheduling results of the integrated energy system under extreme scenarios. The output electric power of the cogeneration unit is taken as a representative, combined with the corresponding enhanced extreme scenario data set, and sampled as the sample set of the Gaussian process regression model.

[0138] In S5, the independent variables in the sample set of the Gaussian process regression model are defined as the meteorological and time data under extreme scenarios, and the dependent variable is the output electric power of the cogeneration unit with corresponding output, so as to establish a rapid quantitative model of the scheduling uncertainty of the integrated energy system under extreme scenarios.

[0139] The GPR-based scheduling uncertainty quantification system for extreme scenarios divides the GPR-Q model sample set into a training set and a test set, with the enhanced extreme scenario data set as input and the output power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios as output:

[0140] The mapping relationship between the enhanced extreme scenario data set and the output power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios is described by the Gaussian process f(X). The parameter value of f(X) is a random variable, and the prior distribution of f(X) is a Gaussian distribution. The GPR-Q training set D = {(x i ,y i )|i=1,2,…,n}=(X,Y), where n is the number of scenes in the GPR-Q training set, X=[x 1 ,x 2 ,…,x n ] Tis the input matrix of the combined heat and power unit output power training set in the comprehensive energy system dispatch results under extreme scenarios, Y = [y 1 ,y 2 ,…,y n ] T is the output matrix of the electric power output of the cogeneration unit in the comprehensive energy system dispatch results under extreme scenarios, y i is the dispatch result under scenario i. The training set D constitutes a Gaussian joint release, and its corresponding Gaussian process GP[·] is composed of the dispatch mean function μ(X) and the meteorological data covariance matrix K(x i ,x j ) uniquely defines and establishes the actual Gaussian process model:

[0141]

[0142] In the formula, σ N is the standard deviation of Gaussian distribution; δ ij is the Kronecker delta function; I is the identity matrix;

[0143] Select the remaining GPR-Q model sample set as the GPR-Q test set D * , the expression is as follows:

[0144] D * ={(x i ,f(x i ))|i=n+1,n+2,…,n+n *}=(X * ,f * )

[0145] Where, X * 、f * are the input matrix and output label of the GPR-Q test set respectively; n * is the number of scenes contained in the test set; the matrix form of the prior joint Gaussian distribution is:

[0146]

[0147] Where N(,) is Gaussian distribution; K(·,·) is covariance matrix. According to Bayesian theory, the output result f is obtained * The posterior distribution of is:

[0148]

[0149] Mean vector of the output results of the GPR-Q test set and the variance vector cov(f * )for:

[0150]

[0151] cov(f * )=K(X * ,X * )-K(X * ,X)(K(X,X)+σ N 2 I) -1 K(X * ,X)

[0152] On this basis, GPR-Q introduces the quantile loss function L τ , the deviation expression between the output result of the penalty test set and the training target value of the test set is:

[0153]

[0154] in, is the quantile point of the mean vector under a given quantile level τ, and y is the training target value of the GPR-Q test set.

[0155] Therefore, the goal of the GPR-Q model is to minimize the quantile loss function of the output results of the GPR-Q test set, taking the enhanced extreme scenario data set as input and the output power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios as output, and establishing a scheduling uncertainty quantification system under extreme scenarios based on GPR. Taking 24 hours as the observation period, the results of the scheduling uncertainty quantification system under extreme scenarios based on GPR are as follows: Figure 3 As shown, the blue dot is the target value, the red dot is the mean of the system output results, and the blue interval is the scheduling uncertainty quantification interval.

[0156] This method effectively avoids wind power and photovoltaic output forecast errors by utilizing machine learning algorithms, providing dispatchers with more reasonable and effective dispatch solutions, and can effectively solve the problem of fast and accurate dispatch of integrated energy systems in extreme scenarios.

[0157] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A fast quantification method for scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q, characterized in that: The following steps are involved: Step S1, collecting wind and solar meteorological and load parameters to form a historical data set, selecting scenes that exceed the upper limit of the confidence interval, the lower limit of the confidence interval and large fluctuations as extreme scenes, using the maximum information coefficient to screen and obtain typical meteorological features with a high correlation with load output, and merging the data to establish a real sample data set for extreme scenes; Step S2, based on the extreme scene real sample data set, establish an improved generative adversarial network model with gradient penalty, input the extreme scene real sample data set as a training set into the WGAN-GP model for training, and after the training is completed, obtain the enhanced extreme scene data set in batches from the model generator; Step S3, constructing a comprehensive energy system under extreme scenarios using the enhanced extreme scenario dataset; Step S4, taking the minimum operating cost of the integrated energy system under extreme scenarios as the objective function, bringing the enhanced extreme scenario data set into the calculation, obtaining the scheduling result of the integrated energy system under extreme scenarios, and incorporating it into the corresponding enhanced extreme scenario data set to obtain the GPR-Q model sample set; Step S5, define the independent variables and dependent variables in the GPR-Q model sample set, establish a GPR-based scheduling uncertainty quantification system under extreme scenarios based on the GPR-Q model sample set obtained in step S4, divide the GPR-Q model sample set into a training set and a test set, use the enhanced extreme scenario data set as input, and the output electric power of the cogeneration unit in the comprehensive energy system scheduling results under extreme scenarios as output, and train the GPR-Q model to achieve scheduling uncertainty quantification.

2. According to claim 1, a method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q, characterized in that: In step S1, the establishment of the extreme scenario real sample data set includes extremely high output, extremely low output, and extremely fluctuating output scenarios; after obtaining the historical data set, the load parameters in the historical data set are estimated at a confidence level of 1-α, and the calculation of the upper limit of the confidence interval is expressed as: The calculation of the lower bound of the confidence interval is expressed as: Among them, P Uα is the upper bound of the confidence interval, P Lα is the lower bound of the confidence interval, is the mean output value of each load parameter, σ is the standard deviation of the load parameter, Z α / 2 is the confidence factor at the significance level α; Select 1-α confidence level as 95% confidence level. When there are wind power and photovoltaic output scenarios that exceed the upper and lower boundaries of the confidence interval, they are considered as low-probability extreme events. The scenarios that exceed the upper boundary of the confidence interval are extracted as extreme high output scenarios, and the scenarios that exceed the lower boundary of the confidence interval are extracted as extreme low output extreme scenarios. Based on the upper and lower boundaries of the confidence interval under the target confidence level, calculate the extreme fluctuation range of wind power and photovoltaic output, and select the extreme fluctuation output scenario with the largest hourly fluctuation range from the extreme fluctuation range. The data of extremely high output scenarios, extremely low output scenarios, and extremely fluctuating output scenarios are marked, and the wind and solar meteorological data of the corresponding scenarios are added to the historical data set. The wind and solar meteorological data are screened using the maximum information coefficient to obtain wind and solar meteorological features with high correlation with the corresponding extreme load output, forming a real sample data set for extreme scenarios.

3. The method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q according to claim 1 is characterized in that: In step S2, the enhanced extreme scene data set is obtained in batches using the improved generative adversarial network model with gradient penalty, and a WGAN-GP model needs to be established. The WGAN-GP model includes a generator network, a discriminator network, and a penalty mechanism; wherein the generator network and the discriminator network both use fully connected neural networks, and the input of the generator network is random noise that obeys a standard normal distribution, and the random noise is output as data of a specified size after linear transformation of the hidden layer of the generator network; the input of the discriminator network consists of the data generated by the generator network and the historical data set, and after linear transformation of the hidden layer of the discriminator network, the output is a matrix of a specified size, and the matrix represents the probability score that the generated data is real data; The penalty mechanism is specifically as follows: Among them, E represents mathematical expectation, D represents discriminator, and G represents generator. Represents a single specific real data, express Sampling from real data, z represents random noise that follows a standard normal distribution, sampled from the noise set pz(z); represents the judgment result of the discriminator on the real data, G(z) represents the data generated by the generator, and D(G(z)) represents the judgment result of the discriminator on the generated data; λ represents the penalty term coefficient; represents the differential operator, p(x) represents the middle region formed by the real data space and any point on the line connecting any two points in the real data space, and x is sampled in this region; It represents the derivation of the discriminator's discriminant result D(x) in p(x); p represents the norm. The extreme scenario real sample dataset is used as a training set to input the WGAN-GP model to train the network, which specifically includes the following steps: S2-1 inputs random noise that obeys the standard normal distribution into the generator to obtain data; S2-2 inputs the training set and the data generated by the generator in step S2-1 into the discriminator network for training, trains n_D times, and performs gradient updates according to the loss function. The loss function of the discriminator is: Among them, x i,new represents the training set, N represents the number of samples, λ×gradient_penalty represents the penalty mechanism, z i is a random noise that follows a standard normal distribution, G(z i ) represents the data generated by the generator, D(G(z i )) represents the judgment result of the discriminator on the generated data, D(x i,new ) represents the judgment result of the discriminator on the training set; S2-3 trains the generator by inputting random noise that follows a standard normal distribution into the generator; the loss function of the generator is: The discriminator is trained n_D times and the generator is trained n_G times; S2-4 uses back-propagation and gradient descent algorithms to update parameters, record the loss value, and return to S2-1 for iteration until the number of cycles reaches the set value, and outputs the sample data generated by the generator in the last iteration to obtain an enhanced extreme scenario data set.

4. The method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q according to claim 1, characterized in that: In step S3, the scheduling data needs to be obtained by calculating the integrated energy system under extreme scenarios. The integrated energy system includes the following equipment models: (1) Cogeneration unit model Among them, t is the scheduling time; and are the heat production rate and gas consumption of the CHP unit respectively; and η MT,t are the power generation and power generation efficiency of CHP respectively; η L,t is the heat loss rate; Hg is the calorific value of natural gas, which is 9.88kWh / m3. (2) Gas boiler model in, is the thermal power output of the gas boiler; η GB For gas boiler efficiency; is the gas consumption at time t. (3) Heat pump model Where: Represents the electrical power input to the heat pump; η HP is the heating efficiency of the heat pump; Output thermal power for the heat pump. (4) Model constraints Electricity, heat and gas energy supply and demand balance constraints: in, They are the charging and discharging power of the battery at time t respectively; They are respectively the charging and releasing powers of the heat storage tank at time t; They are the electricity and gas purchases of IES at time t respectively; are the reducible load and the shiftable load of IES at time t respectively; are the heat-to-electricity and electricity-to-heat loads of IES at time t respectively; are the electrical load and thermal load of IES at time t respectively.

5. According to claim 4, a method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q is characterized in that: In step S4, the sample set for constructing the Gaussian process regression model needs to calculate the comprehensive energy system dispatching data under the extreme scenario constructed in step S3, and the system objective function is: minC=C buy +C op +C co2 Where C is the IES operating cost, minC is the minimum IES operating cost; C buy C is the cost of purchasing energy gas; op Operation and maintenance costs of each unit; C CO2 For carbon trading costs; In the formula, and are the electricity purchase and sale prices of the power grid at time t; and To purchase and sell electrical power; is the natural gas selling price at time t; η GH,CHP and η GB are the thermal energy conversion rates of CHP and GB, respectively; is the CHP thermal power; k PV , k WT , k CHP , k GB , k HP , k cell are the unit operating costs of each unit respectively; are the unit operating power of each unit; c is the benchmark price of carbon trading; C grid , C CHP , C GB They are grid purchase, CHP and GB initial carbon quota; C grid,a , C CHP,a , C GB,a They are the actual carbon emissions of power purchase from the grid, CHP and GB.

6. The method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q according to claim 1, characterized in that: In step S4, for the established integrated energy system under extreme scenarios, the sample data set generated by the extreme scenarios is brought into the calculation to obtain the scheduling results of the integrated energy system under extreme scenarios. The output electric power of the cogeneration unit is taken as a representative, combined with the corresponding enhanced extreme scenario data set, and sampled as the sample set of the Gaussian process regression model.

7. The method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q according to claim 6 is characterized in that: In step S5, the independent variables in the Gaussian process regression model sample set are defined as the meteorological and time data under extreme scenarios, and the dependent variable is the output electric power of the cogeneration unit corresponding to the output, so as to establish a rapid quantitative model of the scheduling uncertainty of the integrated energy system under extreme scenarios.

8. The method for rapid quantification of scheduling uncertainty in extreme scenarios based on WGAN-GP and GPR-Q according to claim 7, characterized in that: Based on the GPR-based scheduling uncertainty quantification system under extreme scenarios, the GPR-Q model sample set is divided into a training set and a test set. The enhanced extreme scenario data set is used as input, and the output power of the cogeneration unit in the scheduling result of the integrated energy system under extreme scenarios is used as output. The mapping relationship between the enhanced extreme scenario data set and the output power of the cogeneration unit in the scheduling result of the integrated energy system under extreme scenarios is described by the Gaussian process f(X). The parameter value of f(X) is a random variable, and the prior distribution of f(X) is a Gaussian distribution. The GPR-Q training set D={(x i ,y i )i=1,2,…,n}=(X,Y), where n is the number of scenes in the GPR-Q training set, X=[x1,x2,…,x n ] T is the input matrix of the combined heat and power unit output power training set in the comprehensive energy system scheduling results under extreme scenarios, Y = [y1, y2, …, y n ] T is the output matrix of the electric power output of the cogeneration unit in the comprehensive energy system dispatch results under extreme scenarios, y i is the dispatch result under scenario i. The training set D constitutes a Gaussian joint release, and its corresponding Gaussian process GP[·] is composed of the dispatch mean function μ(X) and the meteorological data covariance matrix K(x i ,x j ) uniquely defines and establishes the actual Gaussian process model: In the formula, σ N is the standard deviation of Gaussian distribution; δ ij is the Kronecker delta function; I is the identity matrix; Select the remaining GPR-Q model sample set as the GPR-Q test set D * , the expression is as follows: Where, X * 、f * are the input matrix and output label of the GPR-Q test set respectively; n * is the number of scenes contained in the test set; the matrix form of the prior joint Gaussian distribution is: Where N(,) is Gaussian distribution; K(·,·) is covariance matrix; according to Bayesian theory, the output result f is obtained * The posterior distribution of is: Mean vector of the output results of the GPR-Q test set and the variance vector cov(f * )for: those(f * )=K(X * ,X * )-K(X * ,X)(K(X,X)+σ N 2 I) -1 K(X * ,X) On this basis, GPR-Q introduces the quantile loss function L τ , the deviation expression between the output result of the penalty test set and the training target value of the test set is: in, is the quantile point of the mean vector under a given quantile level τ, and y is the training target value of the GPR-Q test set. Therefore, the goal of the GPR-Q model is to minimize the quantile loss function of the output results of the GPR-Q test set, with the enhanced extreme scenario data set as input and the output electric power of the cogeneration unit in the integrated energy system scheduling results under extreme scenarios as output. A scheduling uncertainty quantification system under extreme scenarios based on GPR is established.

Images