A power system typical scenario probability prediction method based on dynamic timing prediction

Abstract

The application relates to a power system typical scene probability prediction method based on dynamic timing prediction, which comprises the following steps: firstly, a dynamic timing prediction model for characteristic variables is constructed; support vector machines are used to model data knowledge and obtain data sample decision scores; a sigmoid function is used to modify the sample decision scores and map them to the interval [0, 1]; finally, the values of sigmoid parameters are determined by using maximum likelihood estimation, and quantitative probability prediction of the target typical scene is realized. The original data is reduced by using maximum mutual information, the dynamic timing prediction model is combined with the scene classification model, and the decision score probability conversion method based on maximum likelihood estimation is used to convert the classification information of the power system typical scene into probability information, which is more beneficial to dispatching operation personnel to evaluate the future risk of the system and make more accurate dispatching operation plans.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of typical scene probability prediction of power systems, and particularly relates to a typical scene probability prediction method for power systems based on dynamic time series prediction. BACKGROUND

[0002] With the rapid development of new energy and random load, when they are connected to the power grid on a large scale, their randomness and intermittency will greatly increase the difficulty of power grid planning and dispatching, and may also cause different scenarios such as serious power grid failure, important section overload, power supply and demand imbalance, and new energy accommodation obstruction. However, the current dispatching is a passive waiting mode of dispatching, which cannot meet the requirements of active dispatching, mainly manifested in the following aspects: passively waiting for power grid accidents, and there is still room for improvement in the early warning and prevention of predictable risks such as natural disasters; there is no effective monitoring means for unit power fluctuation, and it relies on the report of power plants; there is no monitoring means for sudden public opinion events in the power grid, and it relies on the report of subordinate dispatching; the information of technical support system is scattered, and there is a lack of centralized monitoring function. If the above typical operating scenarios can be effectively predicted, the dispatching personnel can be warned in advance, and controllable resources can be mobilized to participate in the active adjustment of the power system as soon as possible, which can greatly improve the monitoring and prediction ability of the power system for abnormal states and improve the safe and stable operation level of the power grid.

[0003] The existing researches on power system scene prediction mainly focus on situation awareness and stability assessment. Situation awareness refers to the extraction, understanding and prediction of the state of a large number of environmental factors in time and space to achieve decision advantage. Dr. Endsley first clearly defined situation awareness in 1988, "situation awareness is the extraction, understanding and prediction of the state of a large number of environmental factors in time and space to achieve decision advantage". He believes that situation awareness can be divided into three levels, namely "cognition, understanding and prediction". The core part of situation awareness includes: extraction of situation elements, understanding of current situation, prediction of future situation, decision and action implementation. The research contents of stability assessment mainly include static stability assessment and transient stability assessment. The existing researches on static stability prediction include line flow limit prediction, node voltage limit prediction, limit transmission capacity prediction, etc. Most of these methods are based on the measurable variables of the system, and a learning model based on data driving is constructed to achieve the purpose of prediction. However, the evaluation results of the above methods are usually the state type or the numerical value of a certain concerned index of the future power system, but in general, only the type information or the index value is difficult to reflect the criticality of the power system scene, which cannot help the dispatching personnel to accurately judge the situation, resulting in that the output results of these methods cannot play an ideal role, and the staff still need to re-judge the situation and develop strategies according to their own experience.

[0004] Therefore, based on the above analysis and elaboration, considering the current situation of the rapid development of new energy, the operation state of the system becomes more complex and variable, in order to fully master the operation development trend of the system, provide more feedforward information for the formulation of the dispatching strategy, and thus improve the situation awareness capability of the power system, the present application designs a way of modifying probability based on support vector machine (SVM) decision score output combined with sigmoid function, realizes quantitative probability early warning of the future state of the target scene or key element, and thus provides a more reliable indication for the operation dispatching of the system. In the present application, first, a time series prediction model of the key variables of the typical scene is constructed based on the long short-term memory network (LSTM), then the historical samples of the selected features are classified and learned by using the SVM, so that the SVM can accurately classify the sample scene attributes, thereby accurately predicting the future state, and based on the decision score output by the SVM model, the sigmoid function is used for probability correction learning, the parameters of the sigmoid function are determined based on the maximum likelihood method, the mapping from the decision score to the probability value is completed, and finally the quantitative probability prediction of the target typical scene of the power system is realized. SUMMARY

[0005] The present application proposes a power system typical scene probability prediction method based on dynamic time series prediction. First, according to the physical characteristics of the target typical scene of the power system, the feature subset related to it is selected based on the maximum mutual information measurement method, and the time series data of these feature variables is collected from the historical data, and a multi-dimensional time series data set is formed in combination with the state sequence of the target typical scene; based on the long short-time memory network and the historical time series data, a dynamic time series prediction model for the associated feature variables is constructed through cross-validation and grid search; the scene prediction problem is converted into a classification problem, and the support vector machine is used to model the data knowledge to obtain the decision score for each data sample; then a sigmoid function with parameters is used to modify the decision score output of the SVM, which can map the decision score value to the interval [0, 1], thereby realizing probability output; finally, the maximum likelihood estimation method is used to determine the parameter value in the sigmoid function, realizing the mapping from the decision score to the probability value, thereby finally realizing the quantitative probability prediction of the target typical scene of the power system.

[0006] The present application proposes a power system typical scene probability prediction method based on dynamic time series prediction, characterized in that it comprises the following steps:

[0007] According to the physical characteristics of the target typical scene of the power system, the feature subset related to it is selected based on the maximum mutual information measurement method, and the time series data of these feature variables is collected from the historical data, and a multi-dimensional time series data set is formed in combination with the state sequence of the target typical scene;

[0008] Based on long short-term memory network and historical time series data, a dynamic time series prediction model for associated feature variables is constructed through cross-validation and grid search, data knowledge is modeled using support vector machine, historical samples of the selected features are classified and learned to obtain a decision score for each data sample;

[0009] A sigmoid function with parameters is used to modify the decision score output of SVM, and the decision score value is mapped to the interval [0, 1]. The parameter value in the sigmoid function is determined by the maximum likelihood estimation method, the mapping from decision score to probability value is realized, and the power system target typical scene probability prediction model is obtained. Finally, the dynamic time series prediction result of the feature variable is combined to realize the quantitative probability prediction of the power system target typical scene.

[0010] In the above prediction method, the power system target typical scene to be predicted is determined, and the state sequence of the target typical scene is constructed according to historical data information N is the total number of data points, y k is the value of the target scene state at the kth time point, y k ∈{0,1}, y k =1 indicates that the target scene occurs, positive example, y k =0 indicates that the target scene does not occur, negative example. At the same time, the time series data of each feature variable recorded in the historical data information is obtained, denoted as X. Then the maximum mutual information coefficient (MIC) measurement method is used to obtain the MIC coefficient between each feature variable and the target typical scene state sequence Y, and the threshold is set to remove low-correlation feature variables to obtain the feature subset D' is equal to the total number of feature variables D minus the number of removed feature variables. Then the time series data of the feature variables contained in the feature subset Q is extracted from the historical data, and combined with the target typical scene state sequence Y to form a multi-dimensional time series data set.

[0011]

[0012] In the above formula, Ts i represents the time series data of the ith feature variable F i , y is the data measurement value of the kth time point in the time series data of the ith feature variable, D represents the total number of feature variables, and N is the total number of data points, i.e. the length of the time series data.

[0013] In the above prediction method, based on the obtained multi-dimensional time series data set, a dynamic time series prediction model for each feature variable contained in the feature subset Q is constructed using a long short-term memory network, referred to as a feature variable dynamic time series prediction model. The training input of the long short-term memory network is X' k is an input multi-dimensional time series sample, x' k+α is a regression prediction target corresponding to the sample X' k , and a is the time step of the time series prediction. Through cross-validation and grid search, the feature variable dynamic time series prediction model obtained by training can realize prediction of the feature variables in the feature subset Q for a time step of a in advance.

[0014]

[0015] In the formula, L is the length of each time series segment in the multi-dimensional time series sample.

[0016] In the above prediction method, a power system target typical scene prediction classification supervised format data set is established according to the constructed multi-dimensional time series data set y k = 1 indicates that the sample x′ k belongs to the positive example, y k = 0 indicates that the sample x′ k belongs to the negative example, and then it is divided into m groups. The m-1 groups are extracted to use a support vector machine model to construct a power system typical scene classification model, and a SVM decision function f(.) is obtained. Then the decision function f(.) is used to obtain the decision score for the remaining 1 group of samples and store it. This process is repeated m times, and each time the m-1 groups of samples extracted are different. The decision score for each data sample can be obtained, and a decision score-label set For the kth data sample x′ k , the decision score of the support vector machine model is f k , and y k is the value of the target scene state at the kth time point, k = 1, 2,..., N.

[0017]

[0018] In the above prediction method, a sigmoid function with A and B parameters is used to modify the decision score output by the SVM, map the decision score to the interval [0, 1], and based on the decision score-label set obtained in step 3, the maximum likelihood estimation method is used to determine the values of the parameters A and B of the sigmoid function, to realize the conversion from the decision score to the probability of occurrence of the target typical scene.

[0019] The sigmoid function-based decision score mapping form used is as follows:

[0020]

[0021] wherein A, B are sigmoid function parameters, f is the decision score value corresponding to the input sample x', and P(y = 1 | x') represents the probability of the input sample x' belonging to the positive example.

[0022] Based on the obtained decision score-label set The specific process of solving the parameters A and B using the maximum likelihood estimation method is as follows:

[0023]

[0024]

[0025] wherein: is the probability estimate of the kth sample belonging to the positive example. N + is the number of samples belonging to the positive example in all samples, N - is the number of samples belonging to the negative example in all samples. By solving minF(A, B), the parameters A and B can be obtained, thereby converting the decision score of the SVM model into a probability output.

[0026] In the above prediction method, in real-time state, the latest multi-dimensional time series sample X' t obtained at the current time t is dynamically obtained, which is input into the established feature variable dynamic time series prediction model to obtain the prediction value x' t+α of the D' feature variables at the advance a time step, x' t+α is input into the SVM power system typical scenario classification model established in step 3 to obtain the corresponding decision score f t+α , and the sigmoid function with the A and B parameters determined in step 4 is used to obtain the probability p t+α of the sample x' t+α belonging to the positive example, i.e., the probability of the power system target typical scenario occurring at the future a time step, to complete the probability prediction of the target typical scenario at the advance a time step.

[0027]

[0028]

[0029] The application firstly proposes a power system typical scenario probability prediction method based on dynamic time series prediction, which superimposes probability output on top of the category prediction of the power system typical scenario, thereby providing more reference information for dispatching and operating personnel and helping to establish more accurate dispatching and control strategies. First, the maximum mutual information measurement method is used to select a feature subset strongly related to the target scenario, and the time series data of these feature variables is collected from historical data, combined with the scenario state sequence to form a multi-dimensional time series data set. Then, a dynamic time series prediction model of the associated feature variables is constructed based on the long short-term memory network and the historical time series data. Then, the support vector machine is used to learn the scenario data samples, obtain the decision score for each data sample, and use a sigmoid function with parameters to modify the decision score of the SVM and map it to the interval [0, 1]. Finally, the maximum likelihood estimation method is used to determine the parameter value in the sigmoid function, and the quantitative probability prediction of the power system typical scenario is realized. The application has the following advantages: 1. The original data is reduced in dimension using the maximum mutual information coefficient measurement method, which improves the training efficiency of the subsequent feature variable dynamic time series prediction model based on the long short-term memory network and the typical scenario classification model based on the SVM, and reduces the complexity of the overall model; 2. The feature variable dynamic time series prediction model and the typical scenario classification model are combined to realize dynamic perception of the future state of the power system, which is conducive to a comprehensive understanding of the scenario trend of the power system; 3. The decision score probability conversion method based on maximum likelihood estimation is used to convert the classification information of the target typical scenario of the power system into probability information, which is more conducive to the evaluation of the future risk of the system by the dispatching and operating personnel and the development of more accurate dispatching and operation plans. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a schematic diagram of the SVM classification model.

[0031] Figure 2 is a schematic diagram of the actual load 1 fluctuation segment in the example of the application.

[0032] Figure 3 is a schematic diagram of the actual load 2 fluctuation segment in the example of the application.

[0033] Figure 4 is a schematic diagram of the key variables extracted based on MIC in the example of the application.

[0034] Figure 5 is a schematic diagram of the best prediction model training process in the example of the application.

[0035] Figure 6 is a schematic diagram of the feature 8 time series prediction effect in the example of the application.

[0036] Figure 7is the timing prediction effect schematic diagram of feature 17 in the example of the present application.

[0037] Figure 8 is the timing prediction effect schematic diagram of feature 15 in the example of the present application.

[0038] Figure 9 is the scene probability prediction effect schematic diagram in the example of the present application.

[0039] Figure 10 is the method flow schematic diagram of the present application. DETAILED DESCRIPTION

[0040] The technical solutions of the present application are further specifically described below by examples and in conjunction with the drawings.

[0041] 1. Typical scene associated feature selection based on maximum mutual information coefficient

[0042] In order to screen out the most relevant feature subset of the target typical scene of the power system and reduce the calculation complexity of subsequent learning model construction, the present application uses the maximum mutual information coefficient method to measure the correlation coefficient between the alternative features and the target typical scene, and sets a threshold to remove low correlation features.

[0043] The maximum mutual information coefficient is an effective method for measuring the correlation between two variables, which has strong robustness and can capture linear and nonlinear correlation between a wide range of two variables. The basic idea of the maximum mutual information coefficient is to divide the scatter plot based on the data of two variables into a grid, and then evaluate the correlation between the two variables based on the grid division. Therefore, in order to obtain the maximum mutual information coefficient between two variables, the grid division strategy is taken as much as possible, and the maximum correlation coefficient value calculated in all grid division cases is called the maximum mutual information coefficient.

[0044] Specifically, given a set of alternative features F i denotes the i-th feature variable, and the time series data of these feature variables collected from historical data is denoted as X:

[0045]

[0046] In the formula, Ts i denotes the time series data of the i-th feature variable F i , D denotes the total number of alternative feature variables, and N is the total number of data points, i.e. the length of the time series data.

[0047] Similarly, the state sequence of the target typical scene can be obtained​ N is the total number of data points, y k is the value of the target scene state at the kth time point, y k ∈{0,1}, y k = 1 indicates that the target scene occurs, positive example, y k = 0 indicates that the target scene does not occur, negative example. Therefore, for the ith feature variable F i , based on its historical data and the target scene state sequence, a data pair set D is obtained. Its distribution is distributed in a two-dimensional coordinate system, and a scatter distribution is obtained. The horizontal axis value is divided into a intervals, and the vertical axis value is divided into b intervals (allowing no data points in the interval). This is an a-by-b grid division, a and b are positive integers, which is denoted as G. Given a grid division G, the scatter distribution of the data points in D i | G in the grid G is represented as D i , therefore, for a certain data set D i , different grid division strategies G can lead to different scatter distributions D i | G . Therefore, for a data set D i and fixed a, b parameters, the maximum mutual information value that can be obtained on all possible grid division strategies G is represented as follows:

[0048] I * (D i ,a,b)=maxI(D i | G )

[0049] In the formula, I(D i | G ) represents the mutual information between variables calculated from the scatter distribution D i | G , and thus D i about I * characteristic matrix, as follows:

[0050]

[0051] In the formula, M(D i ) is the characteristic matrix of D i , M(D i ) a,b represents the element in the a-th row and the b-th column of the matrix, and through the denominator in the formula, the mutual information value can be normalized to the interval [0, 1], so as to ensure the fair maximum mutual information coefficient comparison between different features. Thus, the ith feature variable F iThe maximum mutual information coefficient (MIC) between the target scene label Y can be defined as the maximum value in the matrix M(D i )

[0052] MIC(F i ,Y) = max ab＜B(N) {M(D i )}

[0053] where MIC(F i ,Y) denotes the maximum mutual information coefficient between the i-th feature variable F i and the target scene label Y, and B(N) is a function of the number of data points N, which controls how many grid division strategies need to be considered, and is usually set as B(N) = N 0.6 . The MIC value is in the interval [0, 1], and the larger the value, the higher the correlation between the two. By sequentially calculating the maximum mutual information coefficient between D candidate variables and the target scene label sequence, the correlation between the D candidate variables and the target scene can be obtained, and the screening of the associated feature variables can be completed by setting a threshold to obtain a key feature variable subset D' is equal to the total number of candidate variables D minus the number of eliminated feature variables. Then the time series data of the feature variables contained in the feature subset Q can be extracted from the historical data, and combined with the target typical scene state sequence Y to form a multi-dimensional time series data set.

[0054] 2. Feature variable dynamic time series prediction model construction

[0055] (1) Long short-term memory network

[0056] With the explosive growth of data and the improvement of computer performance, traditional neural networks limit their further improvement in efficiency and performance to deal with big data problems due to their own limitations; traditional machine learning often cannot process data in raw format. For example, when classifying pictures, we do not use all the pixel values of the entire picture as input, but artificially extract the features of the picture and then convert them into digital form for network training. Deep learning is one of the important components of machine learning, which refers to a series of feature learning and processing algorithms based on deep neural networks to solve image, text detection, tracking, classification and other problems. Deep learning is a kind of representation learning, which is a learning method that enables machines to automatically detect the features of raw data. It is used in many representation layers, which are composed of simple but nonlinear modules. Each layer will pass the representation to the next layer as a higher level of abstraction. Deep learning mainly focuses on learning the features of data, and through different levels of neural networks, it obtains hierarchical data features to solve the problem of manually extracting features. There are many types and variations of neural networks used in deep learning. The two most widely used and common neural networks are RNN (Recurrent Neural Network) and CNN (Convolutional Neural Network).

[0057] CNN (Convolutional Neural Network): It can be regarded as an advanced version of the standard neural network. It contains convolutional layers, pooling layers, and full-connection layers. These results enable it to receive the complete pixel values of the picture without the need for manual feature extraction;

[0058] RNN(Recurrent Neural Network, recurrent neural network): it is a completely different model from CNN, which is used to show the dynamic behavior of sequence data. Recurrent Neural Network (RNN) is a kind of input sequence data, which combines internal state to process input sequence, which is used to show the dynamic behavior of sequence data, and all nodes (recurrent units) are connected in chain. Time series data refers to data collected at different time points, which reflects the state or degree of change of a thing, phenomenon, etc. over time. This is the definition of time series data, of course, it can not be time, such as text sequence, but the sequence data has a feature-the data after the data has a relationship with the data before it.

[0059] We know from traditional neural networks that neural networks contain input layer, hidden layer, output layer, the process is from input layer to hidden layer and then to output layer, the output is controlled by activation function, the connection between layers is full connection, and the nodes between layers are not connected. The activation function is determined in advance, so what the neural network model "learns" through training is contained in the "weight". But this basic neural network only establishes weight connection between layers, and it is powerless for many problems. For example, you want to predict the next word of a sentence, you usually need to use the previous words, because the previous and next words in a sentence are not independent. The biggest difference between RNN is that it also establishes weight connection between neurons in layers. RNN is called recurrent neural network because the current output of a sequence is related to the previous output. The specific form of expression is that the network will remember the previous information and apply it to the calculation of the current output, that is, the nodes between the hidden layers are not unconnected but connected, and the input of the hidden layer not only includes the output of the input layer but also the output of the hidden layer at the last time. In theory, RNN can process sequence data of any length. This ability to effectively represent the information between time series makes RNN widely used in natural language processing (NLP), such as speech recognition, language modeling, machine translation, etc. It is also used for various time series prediction and has achieved a series of excellent results.

[0060] The long short-term memory network (LSTM) structure is an improvement of the traditional RNN model. The long short-term memory network LSTM is a special variant of RNN, and RNN can only have short-term memory due to the gradient disappearance. The LSTM network brings the addition operation into the network through the ingenious gate control, which solves the problem of gradient disappearance to a certain extent and can learn long-term dependent information. However, too long sequence will still appear "gradient disappearance" (length more than 300 is likely to appear), so LSTM is called a "short-term memory" of long. LSTM was proposed by Hochreiter & Schmidhuber (1997) and improved and popularized by Alex Graves. In many problems, LSTM has achieved considerable success and has been widely used.

[0061] (2) Feature variable dynamic time series prediction model based on LSTM

[0062] Based on the aforementioned obtained multi-dimensional time series data set, a dynamic time series prediction model for each feature variable contained in the feature subset Q is constructed using the long short-term memory network, which is called the feature variable dynamic time series prediction model. The training input of the long short-term memory network is X' k The input multi-dimensional time series sample is as follows:

[0063]

[0064] In the formula, D' is the dimension of the feature subset, and L is the length of each time series segment in the multi-dimensional time series sample.

[0065] x' k+α The regression prediction target corresponding to the sample X' k is as follows:

[0066]

[0067] In the formula, α is the time step of the time series prediction.

[0068] Finally, through cross-validation and grid search, the feature variable dynamic time series prediction model trained can realize the prediction of the feature variables in the feature subset Q in advance by α time steps.

[0069] 3. Power system typical scenario classification model construction and probability transformation method

[0070] (1) Support vector machine model

[0071] Support Vector Machine (SVM) is widely used in power system, and has shown good prediction performance in transient stability assessment, transformer fault diagnosis, etc. It has the characteristics of less training samples and strong generalization ability. At the same time, for an arbitrary sample x, SVM can obtain the distance from the sample to the classification hyperplane, so as to define the confidence degree of the classification result, which has the meaning of probability. SVM converts the input space to a high-dimensional feature space, finds a classification hyperplane in the high-dimensional space, and maximizes the classification interval under the premise of separating the sample points without error, so as to obtain the best classification effect.

[0072] Support Vector Machine is a binary classification model, and its basic model is a linear classifier with maximum interval defined in the feature space, which makes it different from the perception machine. Support Vector Machine also includes kernel trick, which makes it a substantial nonlinear classifier. The learning strategy of support vector machine is to maximize the interval, which can be formalized as a convex quadratic programming problem, and is equivalent to the minimization problem of regularized hinge loss function. The learning algorithm of support vector machine is the optimization algorithm for solving convex quadratic programming.

[0073] The learning method of support vector machine includes constructing models from simple to complex. The simple model is the basis of the complex model and also a special case of the complex model. When the training data is linearly separable, a linear classifier is learned through hard margin maximization, which is called linearly separable support vector machine, also known as hard margin support vector machine. When the training data is approximately linearly separable, a linear classifier is learned through soft margin maximization, which is called linear support vector machine. When the training data is linearly inseparable, a nonlinear support vector machine is learned through the use of kernel trick and soft margin maximization.

[0074] When the input space is a Euclidean space or a discrete set, and the feature space is a Hilbert space, the kernel function represents the inner product between the feature vectors obtained by mapping the input from the input space to the feature space. By using kernel function, nonlinear support vector machine can be learned, which is equivalent to learning linear support vector machine in high-dimensional feature space implicitly. Such method is called kernel trick. Kernel method is a more general machine learning method than support vector machine.

[0075] (2) Typical scene classification model construction and probability conversion method based on SVM

[0076] According to the historical time series data of the feature subset and the target typical scene state sequence, the target typical scene prediction classification supervised format data set of power system can be established y k = 1 indicates that the sample x' k belongs to the positive case, y k = 0 indicates that the sample x' k belongs to the negative case, k = 1, 2,..., N:

[0077]

[0078] where x' k ∈ R D′ represents the D' dimensional input features of the kth sample, y k is the sample label value. The SVM uses a kernel function K(.) to map the original problem to a high-dimensional space, and then constructs an optimal classification hyperplane f(x') in the transformed high-dimensional space, which is the SVM decision function:

[0079]

[0080] where: a k is the Lagrange multiplier, b ∈ R is the offset. a k can be obtained by solving the following optimization problem:

[0081]

[0082] s.t. C ≥ a k ≥ 0, k = 1,..., N

[0083]

[0084] where C ∈ R is the penalty factor. The kernel function K(x' k , x' j ) has many, using the radial basis kernel function can approximate any function with arbitrarily small error, in many existing researches, the kernel function is used to construct the model, and good test performance is obtained, so the radial basis kernel function is selected in this study:

[0085]

[0086] where: γ ∈ R is the kernel parameter. Generally, the optimal parameters γ and C of the model can be obtained by grid search or heuristic algorithm.

[0087] As shown above, for any one sample x' to be classified, the decision function f(x') outputs a certain numerical value, which is the decision score. In a binary classification problem, the SVM can classify it into positive and negative classes based on whether the decision score is greater than 0, thereby completing the class prediction, but the decision score can only reflect the distance of the sample from the classification hyperplane, and does not have a probability meaning. By using a sigmoid function with parameters, the decision score output of the SVM is modified, which can map the decision function value of the SVM to the interval [0, 1], thereby realizing the probability output, and the probability output form is as follows:

[0088]

[0089] In the formula: A, B are sigmoid function parameters, f is the decision score value corresponding to the input sample x', P(y=1|x') represents the probability that the input sample x' belongs to the positive example.

[0090] A data set An SVM classifier and parameters A and B need to be trained at the same time. To avoid overfitting, a cross-validation method is adopted, and the original training samples are divided into two parts, one part is used to train the SVM classification model, and the remaining part is used to determine the A and B of the sigmoid function using the maximum likelihood estimation method. According to the above idea, the training sample set can be divided into m groups, and the SVM model is trained using m-1 groups and the decision score of the remaining one group of samples is calculated, and m times of processing can obtain the decision score of all samples, and then form the decision score-label set The specific process of solving parameters A and B based on the maximum likelihood estimation method is as follows:

[0091]

[0092]

[0093] In the formula: is the probability estimate corresponding to the kth sample belonging to the positive example. N + is the number of samples belonging to the positive example in all samples, and N- is the number of samples belonging to the negative example in all samples. By solving minF(A, B), parameters A and B can be obtained, thereby converting the decision score of the SVM model into a probability output.

[0094] The probability estimate can reflect the probability of the to-be-recognized sample x' belonging to different classes, effectively normalizing the distance of the sample from the classification hyperplane. For a common binary classification problem, the probability output p kBetween 0% and 100%. Now, in addition to the predicted label given by the classification model, the probability output gives more information to the relevant personnel, making them have a more comprehensive understanding of the situation.

[0095] 4. Simulation verification of scenario probability prediction model

[0096] In this example, we use the IEEE 300-node system to simulate the typical scenario probability prediction of the power system based on dynamic timing prediction, taking the typical scenario of the key section overload of the power system as an example. In order to construct a sufficient number of training samples with rich types for the LSTM dynamic timing prediction model and the SVM scenario probability prediction model, and have good generalization ability, we first use the actual load data collected in the Central China Power Grid to access the IEEE 300-node system to form a fluctuation source. Then, we use the power flow calculation to obtain the values of each monitoring variable in the system at each time section, thereby forming a complete data sample, which is then included in the timing prediction model and the scenario probability prediction model, so as to train a classification model that can predict the probability of future scenarios online.

[0097] Based on the above idea, this study selects 360 days of actual data from two actual load points, which are connected to two nodes of the system, with a load data sampling interval of 15 minutes. Figure 2 and Figure 3 It shows some of its characteristics, and it can be seen that it has relatively large fluctuation characteristics. After timing flow simulation calculation, a total of 34560 samples of data are formed, including 1960 alternative data characteristics. According to the maximum mutual information coefficient method, 19 key characteristics are selected, as shown in Figure 4 .

[0098] The extracted key variable data information is input into the LSTM network for training. In this study, we use the data of the previous 301 days as training and the data of the last 59 days as testing. Similarly, this scheme is also applied to the SVM scenario classification model.

[0099] After dividing the sample set into a training set and a test set, cross-validation is adopted on the training set, combined with grid search, to learn the best model and parameter setting, which makes the trained model have the largest generalization ability as possible. After determining the model parameters, the entire data set is used to train the model, so that the constructed model learns as much knowledge as possible. After constructing the time series prediction model, the data of the previous 301 days is used to train the SVM classification model. In this example, whether a certain key section is overloaded or not is taken as a label (i.e., in the learning process, the overload of the key section is taken as a positive example, and the non-overload is taken as a negative example), so that the SVM model can learn the data characteristics of the overload sample from the historical data, so as to make accurate scene classification prediction based on the LSTM time series prediction information in the future. Figure 5 The loss reduction process in the best LSTM model training process is shown. Figures 6-8 The dynamic time series prediction effect of part of the variables is shown.

[0100] Through the time series prediction effect of Figures 6-8 , we can see that the time series prediction model constructed in the present application can better predict the value of the key variable in the future short term, which also lays a solid foundation for subsequent scene probability prediction. Through the construction of the LSTM model and the SVM model, and based on the scene probability prediction method described in the foregoing based on the SVM decision score output conversion, the decision score of the SVM is converted into scene probability through the sigmoid function, and the function of quantitative probability early warning of the future power system typical scene based on dynamic time series prediction can be realized, and the probability prediction effect is shown in Figure 9 .

[0101] From Figure 9 , we can see that the SVM model better predicts the future scene category attribution of the system. In the simulation of 96 points in a day, only one time point is not predicted accurately, and under the action of the scene probability prediction model proposed in the present application, when the category information prediction is wrong, the probability prediction information can still assist the dispatch personnel to judge the situation. With scene probability prediction, the judgment of the future scene type is no longer a non-zero or one case. With probability prediction information, we can increase the certainty of our judgment, so as to assist the dispatch and operation personnel to make more accurate and effective decisions. As can be seen from the figure, the probability prediction curve (dashed line) better tracks the change of the typical scene category of the system, and provides more reference information on the category information.

[0102] The specific embodiments described herein are merely illustrative of the spirit of the application. Various modifications or changes in the specific embodiments described herein can occur to those skilled in the art to which the application pertains without departing from the spirit of the application, and it is understood that such modifications or changes are to be considered as within the scope of the application as defined by the appended claims.

Claims

1. A power system typical scenario probabilistic forecasting method based on dynamic timing prediction, characterized in that, include Based on the physical characteristics of typical target scenarios in the power system, a subset of features related to them is selected using the maximum mutual information metric method. Time series data of these feature variables are collected from historical data and combined with the state sequences of typical target scenarios to form a multidimensional time series dataset. Based on Long Short-Term Memory (LSTM) networks and historical time-series data, a dynamic time-series prediction model for associated feature variables is constructed through cross-validation and grid search. Support Vector Machines (SVM) are used to model data knowledge, and historical samples of the selected features are classified and learned to obtain a decision score for each data sample. A parameterized sigmoid function is used to modify the decision score output of the SVM and map the decision score values ​​to the interval [0,1]. The maximum likelihood estimation method is used to determine the parameter values ​​in the sigmoid function, thereby realizing the mapping from decision scores to probability values. This yields a probabilistic prediction model for typical scenarios of power system targets. Finally, the dynamic time-series prediction results of feature variables are combined to achieve quantitative probabilistic prediction of typical scenarios of power system targets. Determine the target typical scenario of the power system that needs to be predicted, and construct the state sequence of the target typical scenario according to historical data information , N is the total number of data points, is the target scenario state value of the k time point, , represents that the target scenario occurs, positive example, represents that the target scenario does not occur, negative example; meanwhile, the time sequence data of each feature variable recorded in the historical data information is obtained, denoted as X , then the MIC coefficient between each feature variable and the target typical scenario state sequence is obtained by using the maximum mutual information coefficient MIC measurement method, and the threshold is set to remove low-correlation feature variables, to obtain a feature subset , is equal to the total number of feature variables D minus the number of removed feature variables; Subsequently, a feature subset is extracted from the historical data time series data of the feature variables contained in the historical data and in combination with the target typical scenario state sequence , a multi-dimensional time series dataset is formed; A sigmoid function with parameters A and B is used to modify the decision scores output by the SVM, mapping the decision scores to the [0,1] interval, and based on the decision score-label set obtained in step 3. The maximum likelihood estimation method is used to determine the values ​​of parameters A and B of the sigmoid function, thereby realizing the transformation from decision scores to the probability of occurrence of typical target scenarios. The decision score mapping form based on the sigmoid function is as follows: In the formula: , For the sigmoid function parameter, For the corresponding input sample The decision score, Indicates input sample The probability of being a positive example.

2. The prediction method according to claim 1, characterized in that, Using Long Short-Term Memory (LSTM) networks, construct a feature subset... The dynamic time-series prediction model for each feature variable included in the network is called the feature variable dynamic time-series prediction model; the training input for the Long Short-Term Memory network is... , To input multidimensional time series samples, For the corresponding sample The regression prediction objective, This refers to the advance time steps for time series prediction; through cross-validation and grid search, the trained dynamic time series prediction model for feature variables can achieve prediction of feature subsets. Pre-emption of characteristic variables in Prediction of time steps.

3. The prediction method according to claim 1, characterized in that, Establish a supervised dataset for predicting and classifying typical scenarios of power system targets. , Indicates sample This is a positive example. Indicates sample Negative examples are then divided into m groups; m-1 groups are extracted and a typical power system scenario classification model is constructed using a support vector machine (SVM) model to obtain the SVM decision function. Then use the decision function Obtain and store the decision scores for the remaining 1 set of samples. Repeat this process m times, with each time extracting a different m-1 set of samples. This allows us to obtain the decision score for each data sample and establish a decision score-label set. For the first k Data samples The support vector machine model's decision score is , For the first k The target scene state value at each time point. .

4. The prediction method according to claim 1, characterized in that, Based on the obtained decision score-label set The parameters are solved using the maximum likelihood estimation method. , The specific process is as follows: In the formula: For the corresponding to the first A probability estimate of the number of samples that are positive examples; This represents the number of positive samples out of all samples. This represents the number of negative samples out of the total sample count; by analyzing... Solving for the parameters yields the results. , This transforms the decision scores of the SVM model into probability outputs.

5. The prediction method according to claim 1, characterized in that, Dynamically obtain the current time t The latest multidimensional time series samples that can be obtained Input it into the dynamic time series prediction model of the feature variables established in step 2 to obtain information about... Advancement of each characteristic variable Predicted values ​​at time steps ,Will Input the data into the SVM power system typical scenario classification model established in step 3, and obtain its corresponding decision score. And using the sigmoid function with parameters A and B determined in step 4, the sample is obtained. The probability of being a positive example That is, the typical target scenario of the power system in the future The probability of a time step occurring, and the typical scenario of achieving the goal ahead of schedule. Probability prediction of time steps.

Images