Prospective dispatching power balance early warning method considering wind power uncertainty and medium

Abstract

The invention discloses a prospective dispatching power balance early warning method considering wind power uncertainty and a medium. The method comprises the following steps: constructing a training data set corresponding to a power imbalance scene, a power imbalance risk level and a maximum adjustable capacity; constructing an LSTM and MDN combined model, and training the combined model by adopting the training data set; obtaining a test data set, and inputting each sample data in the test data set into the joint model to obtain each maximum adjustable capacity value of each sample data under each power imbalance risk level; and acquiring wind power prediction deviations of the sample data in the test data set, comparing the magnitude relationship between the wind power prediction deviations and the corresponding maximum adjustable capacity values to determine prediction intervals where the wind power prediction deviations fall, and determining the power imbalance risk levels of the corresponding sample data according to the prediction intervals. According to the method, the power imbalance risk level can be quickly and accurately calculated, so that power balance early warning can be quickly and accurately realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a forward-looking dispatching power balance early warning method and medium considering wind power uncertainty. Background Art

[0002] At present, the installed capacity and proportion of electricity of new energy sources are constantly increasing. Wind power generation has the characteristics of randomness and anti-peaking. Its large-scale grid connection makes the power grid's ability to reduce peaking insufficient, resulting in serious wind abandonment and the contradiction of power and electricity shortage. In response to the above problems, the traditional deterministic solution is to pre-set sufficient positive and negative rotating reserve capacity to make up for the excess or insufficient power generation of the units that may occur in the dispatch. This method can no longer fully cope with the balance and safety of the power system. Therefore, it is necessary to perceive and warn the balance risks of power grid dispatch in the next few hours through a forward-looking perspective, and provide guidance information for subsequent dispatch decisions, so that the dispatch agency can reserve climbing capacity in advance and meet the preset dispatch goals during real-time operation.

[0003] The main task of power balance risk warning is to predict and deduce the dispatching scenarios that may occur in the future, identify the dispatching scenarios with imbalance risks, and conduct risk quantitative assessment and classification of these scenarios to provide a reference for the rolling update of the daily dispatch plan. The existing risk warning work is mainly carried out from the aspects of load loss, system spare capacity and control measures. However, although the first two evaluation and grading methods can effectively reflect the degree of insecurity of the operation scenario to a certain extent, they do not reflect what type of resources need to be called in the current scenario to restore the stable operation state, and the coordination of multiple resources is not fully considered, so the results are difficult to provide more information for subsequent real-time dispatching decisions. Therefore, risk warning methods based on control measures gradually emerge. When the risk level of the operation scenario is low, fewer control methods are used; if the risk level is high and only a few control methods cannot guarantee the safe operation of the system, more and more costly control methods need to be invested.

[0004] The maximum adjustable capacity refers to the power system's ability to provide power generation capacity in a short time by reserving some units to avoid power shortages caused by source-load fluctuations. It can comprehensively characterize the system's regulation capability in the current state and serve as a criterion for judging the power balance risk level. In addition, since there is a certain time correlation between the historical sequence data of the power grid, the time series characteristics between the measured data at different times should be fully explored when constructing the early warning model. However, the existing power balance early warning method cannot quickly predict the distribution of the maximum adjustable capacity at each moment within the power grid observation time window, and cannot quickly and accurately calculate the risk level. In addition, the deep learning model used in the existing power balance early warning method has poor training effect, and lacks effectiveness and generalization due to the poor balance of training samples. Summary of the invention

[0005] The present invention aims to solve the technical problems in the related art at least to a certain extent. To this end, the first object of the present invention is to provide a forward-looking dispatching power balance early warning method considering wind power uncertainty, which can quickly and accurately calculate the risk level, thereby realizing power balance early warning quickly and accurately.

[0006] A second object of the present invention is to provide a computer-readable storage medium.

[0007] A third object of the present invention is to provide an electronic device.

[0008] To achieve the above object, the present invention is implemented through the following technical solutions:

[0009] A forward-looking dispatching power balance early warning method considering wind power uncertainty, comprising:

[0010] Step S1: Under each power imbalance scenario, each power imbalance risk level and the maximum adjustable capacity corresponding to each power imbalance risk level are calculated, and a training data set in which the power imbalance scenario, the power imbalance risk level and the maximum adjustable capacity correspond to each other is formed;

[0011] Step S2: constructing a LSTM and MDN joint model, and using the training data set to train the joint model; wherein the LSTM model is used to dynamically extract the electrical characteristics of the preset time window in the training data set to reduce the dimension of the training data set, and the reduced-dimensional training data set is input into the MDN model to obtain the maximum adjustable capacity probability distribution through the MDN model processing;

[0012] Step S3: obtaining a test data set, inputting each sample data in the test data set into the joint model, obtaining each maximum adjustable capacity probability distribution under each power imbalance risk level of each sample data, and obtaining each maximum adjustable capacity value under a preset confidence level based on each maximum adjustable capacity probability distribution;

[0013] Step S4: Obtain the wind power prediction deviation of each sample data in the test data set, compare the size relationship between each wind power prediction deviation and the corresponding maximum adjustable capacity value to determine the prediction interval in which each wind power prediction deviation falls, and determine the corresponding power imbalance risk level of each sample data according to each prediction interval, wherein the wind power prediction deviation is the difference between the wind power day-ahead predicted active power and the wind power ultra-short-term predicted active power.

[0014] Preferably, in step S1, in each power imbalance scenario, the constraints considered when calculating each power imbalance risk level and the maximum adjustable capacity corresponding to each power imbalance risk level include thermal power unit operation boundary constraints, thermal power unit climbing constraints, hydropower unit operation boundary constraints, pumped storage operation boundary constraints, photovoltaic power generation operation boundary constraints, active power constraints on DC interconnection lines, line transmission capacity constraints and total power balance constraints between the power generation side and the load side.

[0015] Preferably, before training the joint model, the method further comprises:

[0016] In the training data sets corresponding to each power imbalance risk level, a training data set containing a small amount of data is determined, and the corresponding power imbalance risk level is determined as the risk level to be enhanced with data;

[0017] The data enhancement method based on moment estimation and matrix normal distribution is used to perform data enhancement on the training data set corresponding to the risk level to be data enhanced, so as to achieve data balance of each training data set.

[0018] Preferably, the data enhancement method based on moment estimation and matrix normal distribution includes:

[0019] Calculating the covariance of the training data set corresponding to the risk level to be enhanced, and decomposing the covariance into two positive definite matrices;

[0020] Calculate the estimated values ​​of two positive definite matrices, recalculate the expectation and covariance of the corresponding training data set based on the estimated values ​​of the two positive definite matrices, and construct the matrix normal distribution function of the corresponding training data set based on the recalculated expectation and covariance;

[0021] A number of training data are sampled from the matrix normal distribution function, and corresponding power imbalance risk levels are calculated for the several training data respectively, and training data whose power imbalance risk level is the risk level to be enhanced are selected to expand the training data set corresponding to the risk level to be enhanced.

[0022] Preferably, in step S2, training the joint model includes: using the training data set, taking maximum likelihood estimation as a goal, and training the joint model through a loss function.

[0023] Preferably, in step S2, training the joint model further includes: optimizing hyperparameters of the joint model using a Bayesian optimization method to train and optimize the joint model.

[0024] Preferably, the method further comprises using a prediction evaluation index to evaluate the prediction result.

[0025] Preferably, the prediction evaluation index includes the prediction interval coverage rate, the prediction interval standard root mean square width and the coverage bandwidth comprehensive evaluation index, wherein the prediction interval coverage rate is used to measure the reliability of the prediction interval; the prediction interval standard root mean square width is used to measure the sharpness performance of the prediction interval; the coverage bandwidth comprehensive evaluation index is used to comprehensively measure the two indicators of the prediction interval coverage rate and the prediction interval standard root mean square width.

[0026] To achieve the above-mentioned purpose, the second aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned forward-looking scheduling power balance warning method considering wind power uncertainty is implemented.

[0027] To achieve the above-mentioned purpose, the third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned forward-looking scheduling power balance warning method considering the uncertainty of wind power is implemented.

[0028] The present invention has at least the following technical effects:

[0029] The present invention provides a forward-looking dispatching power balance warning method considering the uncertainty of wind power. The method uses the LSTM and MDN joint model to quickly predict the maximum adjustable capacity distribution at each moment in the power grid observation time window, characterizes the power imbalance risk by comparing the wind power prediction deviation with the maximum adjustable capacity, and quickly and accurately calculates the risk level based on the wind power prediction deviation and the maximum adjustable capacity. In addition, when training the joint model, the present invention uses the Bayesian optimization method in AutoML to find the optimal hyperparameters of the model, thereby improving the model training effect. At the same time, a data enhancement method based on moment estimation and matrix normal distribution is proposed to improve the effectiveness and generalization of the model.

[0030] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The present invention is a flowchart of a forward-looking scheduling power balance warning method considering wind power uncertainty according to an embodiment of the present invention.

[0032] Figure 2 Schematic diagram of an MDN model according to an embodiment of the present invention.

[0033] Figure 3 Schematic diagram of the relationship between wind power prediction deviation and maximum adjustable capacity according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present embodiment is described in detail below, and examples of the embodiment are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0035] A forward-looking dispatching power balance warning method and medium considering wind power uncertainty of this embodiment will be described below with reference to the accompanying drawings.

[0036] Figure 1 Flow chart of the forward-looking dispatching power balance warning method considering wind power uncertainty according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0037] Step S1: Under each power imbalance scenario, each power imbalance risk level and the maximum adjustable capacity corresponding to each power imbalance risk level are calculated, and a training data set in which the power imbalance scenario, power imbalance risk level and maximum adjustable capacity correspond to each other is constructed.

[0038] In this embodiment, power balance can be used as a safety warning object, and the maximum adjustable capacity can be used as a key indicator to build a system maximum adjustable capacity evaluation model. By calculating the maximum upward adjustment capacity and the maximum downward adjustment capacity corresponding to each power imbalance risk level in each power imbalance scenario, and calculating the power imbalance risk level of the system, a basic database of power imbalance events corresponding to power imbalance scenario data, maximum adjustable capacity and power imbalance risk level, i.e., a training data set, is constructed.

[0039] In order to fully characterize the regulation capability of the system in the current state, the maximum adjustable capacity is divided into the maximum upward regulation capacity and the maximum downward regulation capacity. The maximum upward regulation capacity is the difference between the sum of the maximum output that can be achieved by all resources that can participate in regulation under the current operating state and the sum of the current output under the constraints; the maximum downward regulation capacity is the difference between the sum of the current output and the sum of the minimum output that can be achieved. Define wind power forecast deviation ΔP wind =-(ultra-short-term forecast active power of wind power – day-ahead forecast active power of wind power). wind <0, the maximum down-regulation capacity needs to be calculated; when ΔP wind When >0, the maximum increase capacity needs to be calculated.

[0040] The power system resources considered in this embodiment include thermal power, hydropower, photovoltaic power generation, wind power, energy storage and pumped storage, and flexible loads. In this embodiment, the risk level refers to the power imbalance risk level. Risk level 1 means that after a power imbalance event occurs, the system can be restored to a power balance state by introducing the rotating reserve of the thermal power unit; risk level 2 means that the system needs to be restored to a power balance state by introducing the rotating reserve and pumped storage and energy storage of the thermal power unit; risk level 3 means that the flexible load needs to be readjusted to restore the power balance state of the system; risk level 4 means that the introduction of flexible loads cannot be restored, and wind and solar power need to be abandoned or loads need to be cut.

[0041] In each power imbalance scenario, the constraints considered when calculating each power imbalance risk level and the maximum adjustable capacity corresponding to each power imbalance risk level, namely the constraints mentioned above, include the operating boundary constraints of thermal power units, the climbing constraints of thermal power units, the operating boundary constraints of hydropower units, the operating boundary constraints of pumped storage units, the operating boundary constraints of photovoltaic power generation, the active power constraints on DC interconnection lines, the line transmission capacity constraints and the total power balance constraints between the power generation side and the load side.

[0042] Among them, the operating boundary constraints of thermal power units are:

[0043] P fmin ≤P f ≤P fmax (1)

[0044] Among them, P f For thermal power output, P fmin and P fmax They respectively represent the lower and upper limits of thermal power output.

[0045] The climbing constraints of thermal power units are:

[0046] P t -P t-1 ≤P Asc (2)

[0047] P t-1 -P t ≤P Desc (3)

[0048] Among them, P t is the active power of the thermal power unit during period t, P t-1 is the active power of the thermal power unit in period t-1, P Asc To limit the power when climbing uphill, P Desc Limit power for downhill climbs.

[0049] The operating boundary constraints of the hydropower unit are:

[0050]

[0051] in, is the power generation of the hydropower unit in period t, C w is the maximum adjustable capacity of hydropower on that day, W w is the total hydroelectric power generation on that day.

[0052] The boundary constraints for pumped storage operation are:

[0053]

[0054] in, is the operating power of pumped storage in period t (pumping is negative, power generation is positive), P ps,p , P ps,g are the maximum power of pumped storage pumping power and power generation power, is the storage capacity of pumped storage in period t, H ps,max The maximum storage capacity of pumped storage is is the storage capacity of pumped storage in the period t+1, max is the maximum value function, ΔT is the period length, η ps,p The efficiency of pumped storage.

[0055] The boundary constraints for photovoltaic power generation operation are:

[0056]

[0057] Among them, P pv (t) is the original actual output power of photovoltaic, P g is the photovoltaic installed capacity, T1 is the photovoltaic data sampling period, and Δt is the photovoltaic data collection period.

[0058] The active power constraint condition on the DC interconnection line is:

[0059]

[0060] Among them, K DC is the maximum permissible peak shaving depth, C DC is the capacity of DC interconnection lines outside the province, is the active power on the DC interconnection line during period t.

[0061] The line transmission capacity constraint is:

[0062]

[0063] in, is the maximum transmission capacity of transmission line l, k l1 , k l2 , k l3are the DC power flow transmission coefficients corresponding to conventional units, new energy fields and load nodes, respectively, m 、w t , L b,t are the powers of conventional units, new energy fields and load nodes respectively, and l represents the transmission line.

[0064] The total power balance constraint conditions between the generation side and the load side are:

[0065]

[0066] in, are the power on the generation side and the power on the load side during period t respectively.

[0067] Step S2: construct an LSTM and MDN joint model, and use the training data set to train the joint model; wherein the LSTM model is used to dynamically extract the electrical characteristics of the preset time window in the training data set to reduce the dimension of the training data set, and the reduced dimension training data set is input into the MDN model to obtain the maximum adjustable capacity probability distribution through the MDN model processing.

[0068] Among them, before training the joint model, the method also includes: determining a training data set containing a small amount of data in the training data sets corresponding to each power imbalance risk level, and determining the corresponding power imbalance risk level as the risk level to be enhanced; using a data enhancement method based on moment estimation and matrix normal distribution to enhance the training data set corresponding to the data enhancement risk level, so as to achieve data balance of each training data set and improve the generalization and effectiveness of the model.

[0069] In this embodiment, the joint model is a deep learning model, which relies on the training of massive samples, and the balance of training samples is a sufficient condition for the effectiveness and generalization of its model. The prediction target of the deep learning model in this embodiment is the distribution of the maximum upward adjustment capacity and the maximum downward adjustment capacity, and the number of samples contained in each power imbalance risk level is quite different. Therefore, it is necessary to adopt a data enhancement method to solve the problem of sample imbalance. The main implementation methods of the data enhancement method include methods based on unlabeled data, methods based on data synthesis, methods based on feature enhancement, etc. This embodiment adopts a method based on feature enhancement to achieve data enhancement, that is, to improve sample diversity by enhancing the sample feature space. In other words, the amount of data in the training data sets corresponding to each power imbalance risk level is quite different, and some training data sets have a small amount of data. In order to achieve data balance and improve the generalization and effectiveness of the model, this embodiment adopts a data enhancement method based on moment estimation and matrix normal distribution to enhance the training data set corresponding to the risk level to be enhanced. In this embodiment, it can be determined that the training data set containing a small amount of data is the training data set corresponding to risk level 3, that is, risk level 3 is the risk level to be enhanced.

[0070] In this embodiment, the data enhancement method based on moment estimation and matrix normal distribution includes: calculating the covariance of the training data set corresponding to the risk level of the data to be enhanced, and decomposing the covariance into two positive definite matrices; calculating the estimated values ​​of the two positive definite matrices, recalculating the expectation and covariance of the corresponding training data set according to the estimated values ​​of the two positive definite matrices, and constructing the matrix normal distribution function of the corresponding training data set based on the recalculated expectation and covariance; sampling a number of training data from the matrix normal distribution function, and calculating the corresponding power imbalance risk level for the several training data, respectively, and selecting the training data whose power imbalance risk level is the risk level of the data to be enhanced, so as to expand the training data set corresponding to the risk level of the data to be enhanced.

[0071] It is assumed that the minority class samples of the power grid can be described by a matrix normal distribution. This assumption is based on our prior knowledge of power grid data, which assumes that it exhibits matrix normal distribution characteristics in multidimensional space.

[0072] Assume that the expectation E(X) = μ and the covariance is:

[0073] var(X)=E[vec(XE(X))·vec T (XE(X))](11)

[0074] Where X is the training data set, μ represents the expectation, E(·) represents the expectation function, var(X) is the covariance of the training data set X, vec(·) is the matrix vectorization operator, var(·) represents the covariance function, and vec T(·) represents the transpose of vec(·), and vec(XE(X)) represents the vector obtained by performing a matrix-vector operation on XE(X).

[0075] Where vec(X) is represented as follows:

[0076]

[0077] in, For p L Line p R Column data, vec(X) represents the vector after matrix vectorization operation on X, p L 、p R Represent the number of rows and columns of the training dataset X respectively.

[0078] If the covariance of the training data set X can be decomposed into the Kronecker product of two positive definite matrices Ω and M, then the training data set X is matrix-normally distributed, and the vector vec(X) obeys a mean of vec(μ) and a covariance of The multivariate normal distribution of is expressed as The density function of the multivariate normal distribution can be written as:

[0079]

[0080] Among them, f X (X) is the density function of the multivariate normal distribution, f vec(X) represents the distribution function of the variable vec(X), tr is the matrix trace function, and exp is the exponential function.

[0081] Based on this density function, the second-order moment of the training data set X can be obtained as:

[0082] E[(X-μ)(X-μ) T ]=Mtr(Ω)(14)

[0083] E[(X-μ) T (X-μ)]=Ωtr(M) (15)

[0084] Therefore, the positive definite matrices M and Ω can be derived as:

[0085]

[0086] In order to generate a matrix normal distribution that conforms to the actual power grid data, the maximum likelihood estimation in the moment estimation method is used. By performing statistical analysis on the original data, this embodiment estimates the distribution parameters such as the mean and covariance matrix to ensure that the generated samples are consistent with the actual data.

[0087] Maximum likelihood estimation given μ yes Obviously, The parameter represents the maximum likelihood estimate parameter value of μ. When M is fixed, the maximum likelihood estimate of Ω is for:

[0088]

[0089] Among them, n is the sample data size, X p Represents the pth sample data.

[0090] Similarly, when Ω is fixed, the maximum likelihood estimate of M is for:

[0091]

[0092] Among them, X i Represents the i-th sample data.

[0093] set up The vector vec(X) is equivalent to:

[0094]

[0095] Where μ represents the sample mean, Γ2 and Γ1 are two semi-orthogonal matrices, β1 and β2 are two coefficient matrices, f(Y) is a function containing the functional form of Y, Y is a mapping of X, ε is a random error independent of Y, vec(μ) is the vector after matrix-vectorization of μ, vec(f(Y)) is the vector after matrix-vectorization of f(Y), and vec(ε) is the vector after matrix-vectorization of ε.

[0096] When the parameters Ω and M are unknown, the logarithm of the likelihood function is:

[0097]

[0098] in, Represent two column spaces respectively, represents the log-likelihood function, Y i For X i The mapping of f(Y i ) contains Y i A function of the functional form.

[0099] When Γ2 and M are fixed, the derivative of the likelihood function corresponding to β2 can be obtained as for:

[0100]

[0101] in, represents the calculation factor about X, represents the calculation factor of the function f(Y), X n represents the nth sample data, f(Y n ) contains Y n The function of the functional form, Y n For X n 's mapping.

[0102] Substitute equation (22) into equation (21), fix M after algebraic transformation, and then maximize the log-likelihood. Therefore, when Γ1, Γ2, β1, β2, M are fixed, the maximum likelihood estimate is for:

[0103]

[0104] Similarly, when Γ1, Γ2, β1, β2, Ω are fixed, the maximum likelihood estimate for:

[0105]

[0106] If and only if n ≥ max(p L / p R ,p R / p L )+1, the maximum likelihood estimates of Ω and M estimated by equations (25) and (26) are positive definite. At the same time, for high-dimensional covariance matrices can be decomposed into two smaller positive definite matrices and Therefore, the inversion of the sample covariance matrix can be avoided in parameter estimation, where Indicates dimension p L p R ×p L p R The real number, Indicates the dimension is p R ×p R The real number, Indicates the dimension is p L ×p L The above analysis proves that this method does not require a large sample size for parameter estimation.

[0107] Therefore, when n is greater than the preset value, the Ω and M algorithm is determined to be effective, and then the estimated values ​​of Ω and M can be calculated by the above formulas (25) and (26), and the expectation and covariance are constructed based on the estimated values ​​of Ω and M. Then, the matrix normal distribution function of the risk level 3 data set is constructed based on the expectation and covariance. After that, several data are sampled from the matrix normal distribution function, and the corresponding power imbalance risk levels are calculated for the several data respectively, and the data with power imbalance risk level 3 are expanded to the original risk level 3 data set.

[0108] This embodiment uses IEEE118 nodes as examples to build a database, including 15 wind turbines, 10 photovoltaic power plants, 10 hydropower units, 21 thermal power units, 6 energy storage units, 99 loads, and an out-of-region DC interconnection line. Use the planned power data of each unit and the ultra-short-term power forecast data of wind power, and then build an optimization model to solve the power imbalance risk level, and calculate the corresponding maximum increase capacity and maximum decrease capacity. Introduce meteorological data about wind power and integrate multi-source information. These variables constitute the input data set, which has 207 features. The observation time window rolls every 15 minutes to predict the risk level for the next 4 hours. Based on these settings, before data enhancement, 19,200 labeled samples were generated, including 6,430 risk level 1, 5,480 risk level 2, 973 risk level 3, and 6,317 risk level 4. After data enhancement, the number of samples from level 1 to level 4 are 6430, 5480, 3279, and 6317, respectively, alleviating the class imbalance problem.

[0109] This embodiment takes into account the lack of samples in high-risk extreme scenarios, that is, the serious imbalance of samples of different risk levels, and proposes a data enhancement method to improve the generalization and effectiveness of the deep learning risk warning model and avoid problems such as skewed learning and accuracy misleading in the model.

[0110] The LSTM (Long Short-Term Memory Network) model in this embodiment is a conventional model. First, the LSTM model is used for feature extraction. Then, for the MDN (Mixed Density Network) model, Figure 2 As shown in Figure 1, the MDN model represents the probability density of the target vector y as a linear combination of kernel functions. By parameterizing the output layer of the MDN model, a mixed distribution is used to describe the probability distribution of the maximum adjustable capacity. Consider a mixed distribution with g components. The probability density function P(y|x) of each mixed distribution can be expressed as:

[0111]

[0112] Where y is the target vector, i.e., branch flow, π q(x) is the mixing coefficient, which represents the weight of the qth component and can be regarded as the prior probability of the target vector y conditioned on the input vector x. Note that the mixing coefficient π q (x) is usually considered to be a function of the input vector x, satisfying Σπq(x)=1. is the conditional density function of the target vector y of the qth kernel. There are many choices for the kernel function, where μ q (x) represents the center of the qth kernel, σ q (x) represents the common variance. In this embodiment, Gaussian distribution is selected to represent

[0113]

[0114] Where c represents the dimension of the multivariate normal distribution. This embodiment assumes that each component of the output vector is statistically independent and can be expressed with a common variance σ q (x) to describe.

[0115] In a neural network, the parameter π q (x), μ q (x),σ q (x) is usually generated by the output layer of the network. When the training data set is input, the LSTM model will dynamically extract the electrical features of the preset time window in the training data set to reduce the dimension of the training data set, and then input the reduced-dimensional training data set into the MDN model. After processing by the MDN model, the parameter π will be output. q (x), μ q (x),σ q (x), and then substitute it into formulas (27) and (28) to obtain the probability distribution of the maximum adjustable capacity.

[0116] In this embodiment, the time series characteristics between the measured data at different times are fully exploited when constructing the early warning model, i.e., the joint model. In addition, since the prediction error of the output of new energy cannot be completely avoided, the probability prediction method used in this embodiment can provide richer and more comprehensive uncertainty information for the prediction results. In addition, the MDN model constructed in this embodiment combines the neural network with the mixed density model, and uses the nonlinear mapping ability of the neural network to extract the input data features, and the target distribution parameters can be predicted through the mixed density model.

[0117] In one embodiment of the present invention, training the joint model includes: using a training data set, taking maximum likelihood estimation as a goal, and training the joint model through a loss function. Training the joint model also includes: using a Bayesian optimization method to optimize the hyperparameters of the joint model to train and optimize the joint model.

[0118] Specifically, the joint model combining the LSTM model and the MDN model is trained with the maximum likelihood estimation as the goal, and the loss function H(X,Yθ) is defined as the negative log-likelihood function:

[0119]

[0120] Where, X={X1,X2,…,X n} is the data in the training data set; Y={Y1,Y2,…,Y n} is the target value set of the training samples, that is, the mapping of X, X n is the nth sample data, Y i For X i The mapping of X i represents the i-th sample data, Y n For X n The mapping of θ is the joint model network parameter, P(Y i |θ,X i ) indicates that the network parameters of the joint model are θ and the input is X i , the output is Y i The probability density function of .

[0121] The overall learning process of the joint model is divided into three stages: first, the LSTM model is pre-trained with equation (29) as the target to provide a set of good initialization parameters, and then all network parameters are supervised and optimized with the loss function in equation (29) as the target. Finally, for hyperparameters, the AutoML (automatic machine learning) method is used to find the optimal hyperparameters to achieve the best model effect.

[0122] This embodiment adopts the Bayesian optimization method to tune the hyperparameters of deep learning. The Bayesian optimization method in this embodiment is mainly divided into two parts: a probabilistic proxy model and an acquisition function. First, the model is initialized, some hyperparameter combinations are selected, and their performance is evaluated by actually running the deep learning model. Then, an iterative process is performed, and all current evaluation results are used to train the probabilistic proxy model, as shown in equations (30) and (31). Furthermore, the acquisition function of equation (32) is used to select the next hyperparameter combination to be evaluated, the performance of the selected hyperparameter combination is evaluated, and the probabilistic proxy model is updated. Finally, the iterative process is repeated until the maximum number of iterations is met or the performance improvement is lower than a certain threshold.

[0123] The probability proxy model of this embodiment is a Gaussian process. For a new observation point, the Gaussian process predictor variable obeys another mean and variance μ(θ) and σ respectively. 2 Normal distribution of (θ):

[0124]

[0125] Among them, v(θ,θ') is the covariance function, v * is the covariance between the current observation point and all previous observation points, V is the covariance matrix between all observation points, h is the output value of the objective function, v * The transpose of V -1 is the inverse matrix of V, and θ' is the deep learning parameter.

[0126] The acquisition function of this embodiment uses the expected gain function:

[0127]

[0128] in, represents the expected gain function, f min is the observation value that currently minimizes the objective function, represents the gain function, represents the expected gain function.

[0129] Step S3: Obtain a test data set, input each sample data in the test data set into the joint model, obtain the probability distribution of each maximum adjustable capacity under each power imbalance risk level of each sample data, and obtain each maximum adjustable capacity value under a preset confidence level based on each maximum adjustable capacity probability distribution.

[0130] In this embodiment, a test data set x, i.e., an input vector, can be obtained, where each row of x represents the output of different resources, including p L1 The planned output of wind turbines today, p L2 The planned output of photovoltaic power generation L3 The planned output of thermal power units a few days ago, p L4 The planned output of the hydropower units a day ago, p L5 The planned output of energy storage units a day ago, p L6 The planned output of the load node a day ago, p L7 Ultra-short-term power forecast value of a wind turbine, p L8 The wind speed of a wind farm, p L9 The wind direction of a wind farm, Each column represents the output of each resource at each moment in the observation time window. a represents the output of different resources, the ultra-short-term power forecast value of wind turbines, the number of wind speeds and wind directions in the wind farm, and p La Indicates the pth La Column data.

[0131] First, x passes through a k-layer LSTM model to extract the deep hidden feature vector, namely Figure 2 Then, z is input into the MDN model, and multiple sets of weight parameters of the mixed normal distribution, namely the mixing coefficient π, are output. q(x) and the distribution parameter, i.e., the center μ of the qth kernel q (x) and common variance σ q (x), the probability distribution function P(y|x) of the maximum adjustable capacity can be synthesized. Then, based on each maximum adjustable capacity probability distribution function, each maximum adjustable capacity value under a preset confidence level can be obtained, for example, Figure 3 In in, These correspond to the maximum adjustable capacity values ​​for risk levels 1 to 4 respectively.

[0132] Furthermore, in order to satisfy the constraints of the mixed distribution parameters, z is sent to the Softmax activation function to ensure that the weight parameter, i.e., the mixing coefficient π q (x) is a positive real number and its sum is 1, z is sent to the ReLU activation function to ensure that the distribution parameter μ q (x) and σ q (x) is a positive real number.

[0133] Step S4: Obtain the wind power prediction deviation of each sample data in the test data set, compare the size relationship between each wind power prediction deviation and the corresponding maximum adjustable capacity value to determine the prediction interval in which each wind power prediction deviation falls, and determine the corresponding power imbalance risk level of each sample data according to each prediction interval, wherein the wind power prediction deviation is the difference between the wind power day-ahead predicted active power and the wind power ultra-short-term predicted active power.

[0134] Specifically, the power imbalance risk level can be output by judging the wind power forecast deviation and the maximum adjustable capacity of each power imbalance risk level. The physical meaning of this judgment principle is whether the controllable resources covered by the power imbalance risk level can make up for the system imbalance caused by the wind power forecast deviation.

[0135] Among them, one test sample data corresponds to a wind power prediction deviation ΔP wind , and corresponding to four maximum adjustable capacity values ​​for four power imbalance risk levels, such as like Figure 3 As shown in the figure, if the wind power prediction deviation ΔP of a test sample data is wind lie in prediction interval, it is determined that the power imbalance risk level of the test sample data is risk level 1.

[0136] In one embodiment of the present invention, the method further comprises using a prediction evaluation index to evaluate the prediction result.

[0137] Among them, the prediction evaluation indicators include the prediction interval coverage rate, the prediction interval standard root mean square width and the coverage bandwidth comprehensive evaluation index. Among them, the prediction interval coverage rate is used to measure the reliability of the prediction interval; the prediction interval standard root mean square width is used to measure the sharpness performance of the prediction interval. The better the sharpness performance, the higher the certainty of the joint model for future prediction results; the coverage bandwidth comprehensive evaluation index is used to comprehensively measure the two indicators of prediction interval coverage rate and prediction interval standard root mean square width.

[0138] The prediction interval coverage describes the probability that the prediction interval can cover the actual observed value, which is expressed as follows:

[0139]

[0140] Among them, E PICP represents the initial calculated value of the prediction interval coverage index, N is the amount of test sample data, χ is the indicative function, BL bτ and BU bτ are the lower and upper bounds of the tidal range of the bth test sample under the premise of confidence level τ, J b is the bth calculated value, L PICP Indicates E PICP The closer it is to the confidence level, the higher the reliability, E PICP The closer the confidence level, L PICP The smaller.

[0141] The standardized RMS width of the prediction interval is expressed as follows:

[0142]

[0143] Among them, L PINRW is the standard RMS width of the prediction interval, R is the range of the power flow in the test data set, and L PINRW The smaller it is, the more accurate the prediction result is.

[0144] In this embodiment, the prediction interval standard root mean square width is used to measure the sharpness performance of the prediction interval. Among them, a smaller standard root mean square width usually indicates that the model has a higher certainty about the future observations.

[0145] The comprehensive evaluation index of coverage bandwidth is as follows:

[0146] L CWC =(1+η1L PINRW )·{1+χ(E PICP <τ)exp[-η2(E PICP -τ)]}(37)

[0147] Among them, L CWCis the comprehensive evaluation index of coverage bandwidth, η1 and η2 are the penalty coefficients of the prediction interval coverage rate and the prediction interval standard RMS width index, respectively.

[0148] In summary, the present invention can accurately warn of power imbalance events caused by uncertainty in the output of renewable energy, thereby providing guidance information for subsequent scheduling decisions and reserving sufficient decision-making time, so as to more effectively assist in resolving future safety risks; the present invention uses the LSTM-MDN joint model to quickly predict the distribution of the maximum adjustable capacity at each moment in the power grid observation time window, and characterizes the power imbalance risk by comparing the wind power prediction deviation with the maximum adjustable capacity, and can quickly and accurately calculate the power imbalance risk level based on the wind power prediction deviation and the maximum adjustable capacity. In addition, when training the joint model, the AutoML method is used to find the optimal hyperparameters of the model, which can improve the model training effect. In addition, the present invention proposes a data enhancement method based on moment estimation and matrix normal distribution. Considering the lack of samples in high-risk extreme scenarios, that is, the serious imbalance of samples with different power imbalance risk levels, the data enhancement method proposed by the present invention can improve the generalization and effectiveness of the deep learning risk warning model, and avoid problems such as skewed learning and accuracy misleading in the model.

[0149] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned forward-looking scheduling power balance warning method considering wind power uncertainty is implemented.

[0150] Furthermore, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned forward-looking scheduling power balance warning method considering wind power uncertainty is implemented.

[0151] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0152] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A forward-looking dispatching power balance early warning method considering wind power uncertainty, characterized in that: include: Step S1: Under each power imbalance scenario, each power imbalance risk level and the maximum adjustable capacity corresponding to each power imbalance risk level are calculated, and a training data set in which the power imbalance scenario, the power imbalance risk level and the maximum adjustable capacity correspond to each other is formed; Step S2: constructing a LSTM and MDN joint model, and using the training data set to train the joint model; wherein the LSTM model is used to dynamically extract the electrical characteristics of the preset time window in the training data set to reduce the dimension of the training data set, and the reduced-dimensional training data set is input into the MDN model to obtain the maximum adjustable capacity probability distribution through the MDN model processing; Step S3: obtaining a test data set, inputting each sample data in the test data set into the joint model, obtaining each maximum adjustable capacity probability distribution under each power imbalance risk level of each sample data, and obtaining each maximum adjustable capacity value under a preset confidence level based on each maximum adjustable capacity probability distribution; Step S4: Obtain the wind power prediction deviation of each sample data in the test data set, compare the size relationship between each wind power prediction deviation and the corresponding maximum adjustable capacity value to determine the prediction interval in which each wind power prediction deviation falls, and determine the corresponding power imbalance risk level of each sample data according to each prediction interval, wherein the wind power prediction deviation is the difference between the wind power day-ahead predicted active power and the wind power ultra-short-term predicted active power.

2. The forward-looking dispatching power balance early warning method considering wind power uncertainty according to claim 1 is characterized in that: In the step S1, in each power imbalance scenario, the constraints considered when calculating each power imbalance risk level and the maximum adjustable capacity corresponding to each power imbalance risk level include the operating boundary constraints of the thermal power unit, the climbing constraints of the thermal power unit, the operating boundary constraints of the hydropower unit, the operating boundary constraints of the pumped storage unit, the operating boundary constraints of the photovoltaic power generation unit, the active power constraints on the DC interconnection line, the line transmission capacity constraints and the total power balance constraints between the power generation side and the load side.

3. The forward-looking dispatching power balance early warning method considering wind power uncertainty according to claim 1 is characterized in that: Before training the joint model, the method further includes: In the training data sets corresponding to each power imbalance risk level, a training data set containing a small amount of data is determined, and the corresponding power imbalance risk level is determined as the risk level to be enhanced with data; The data enhancement method based on moment estimation and matrix normal distribution is used to perform data enhancement on the training data set corresponding to the risk level to be data enhanced, so as to achieve data balance of each training data set.

4. The forward-looking dispatching power balance early warning method considering wind power uncertainty as claimed in claim 3 is characterized in that: Data augmentation methods based on moment estimation and matrix normal distribution include: Calculating the covariance of the training data set corresponding to the risk level to be enhanced, and decomposing the covariance into two positive definite matrices; Calculate the estimated values ​​of two positive definite matrices, recalculate the expectation and covariance of the corresponding training data set based on the estimated values ​​of the two positive definite matrices, and construct the matrix normal distribution function of the corresponding training data set based on the recalculated expectation and covariance; A number of training data are sampled from the matrix normal distribution function, and corresponding power imbalance risk levels are calculated for the several training data respectively, and training data whose power imbalance risk level is the risk level to be enhanced are selected to expand the training data set corresponding to the risk level to be enhanced.

5. The forward-looking dispatching power balance early warning method considering wind power uncertainty according to claim 1 is characterized in that: In the step S2, training the joint model includes: using the training data set, taking maximum likelihood estimation as a goal, and training the joint model through a loss function.

6. The forward-looking dispatching power balance early warning method considering wind power uncertainty according to claim 1 is characterized in that: In the step S2, training the joint model further includes: optimizing the hyperparameters of the joint model using a Bayesian optimization method to train and optimize the joint model.

7. The forward-looking dispatching power balance early warning method considering wind power uncertainty according to claim 1 is characterized in that: The method also includes using a prediction evaluation index to evaluate the prediction result.

8. The forward-looking dispatching power balance early warning method considering wind power uncertainty as claimed in claim 7 is characterized in that: The prediction evaluation index includes the prediction interval coverage rate, the prediction interval standard root mean square width and the coverage bandwidth comprehensive evaluation index, wherein the prediction interval coverage rate is used to measure the reliability of the prediction interval; the prediction interval standard root mean square width is used to measure the sharpness performance of the prediction interval; the coverage bandwidth comprehensive evaluation index is used to comprehensively measure the two indicators of the prediction interval coverage rate and the prediction interval standard root mean square width.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the forward-looking scheduling power balance warning method considering wind power uncertainty according to any one of claims 1 to 8 is implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the forward-looking scheduling power balance warning method considering wind power uncertainty according to any one of claims 1 to 8 is implemented.

Images