CN118133161A - Power system inertia interval probability prediction method based on variable decibel leaf inference - Google Patents

Power system inertia interval probability prediction method based on variable decibel leaf inference Download PDF

Info

Publication number
CN118133161A
CN118133161A CN202410559328.0A CN202410559328A CN118133161A CN 118133161 A CN118133161 A CN 118133161A CN 202410559328 A CN202410559328 A CN 202410559328A CN 118133161 A CN118133161 A CN 118133161A
Authority
CN
China
Prior art keywords
prediction
power system
inertia
probability
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202410559328.0A
Other languages
Chinese (zh)
Other versions
CN118133161B (en
Inventor
陈烜
胡俊峰
张磊
陶千惠
邱凌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
State Grid Sichuan Electric Power Service Co ltd
Wuhan Xindian Electrical Co ltd
China Three Gorges University CTGU
Original Assignee
State Grid Sichuan Electric Power Service Co ltd
Wuhan Xindian Electrical Co ltd
China Three Gorges University CTGU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by State Grid Sichuan Electric Power Service Co ltd, Wuhan Xindian Electrical Co ltd, China Three Gorges University CTGU filed Critical State Grid Sichuan Electric Power Service Co ltd
Priority to CN202410559328.0A priority Critical patent/CN118133161B/en
Publication of CN118133161A publication Critical patent/CN118133161A/en
Application granted granted Critical
Publication of CN118133161B publication Critical patent/CN118133161B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G06F18/24155—Bayesian classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Water Supply & Treatment (AREA)
  • Probability & Statistics with Applications (AREA)
  • Primary Health Care (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a power system inertia interval probability prediction method based on variable decibel leaf inference, which comprises the steps of selecting short-term inertia prediction input characteristics of a power system, establishing a priori probability model and a likelihood function model of each variable, and realizing deep learning of each variable; based on the established variable learning models, adopting KLqp variation inference algorithm learning model posterior probability distribution to construct a variation Bayesian inference learning framework suitable for inertia interval probability prediction; based on the constructed power system inertia interval probability prediction framework, a random gradient Hamiltonian Monte Carlo algorithm is adopted to realize the power system inertia interval probability prediction of the prediction model. The method provided by the invention can be used for carrying out reliable prediction and effective uncertainty estimation on the short-term inertia prediction of the power system, is beneficial to the power grid dispatching department to make a dispatching plan in advance, and reduces the accident rate of the power grid.

Description

Power system inertia interval probability prediction method based on variable decibel leaf inference
Technical Field
The invention relates to the field of power system inertia prediction, in particular to a power system inertia interval probability prediction method based on a variable decibel leaf inference.
Background
Along with transformation and upgrading of the power system, synchronous machines in the novel power system taking new energy as a main body are gradually replaced, system inertia is reduced, and the safety and stability of the system are endangered. The inertia is used as an important index for measuring the power unbalance resistance and frequency fluctuation inhibition capability of the power system, the accurate evaluation of the inertia of the system is a basis for guaranteeing the safe and stable operation of the system after the high-proportion new energy is accessed, and the early prediction of the inertia level of the system is an important premise for eliminating the weak risk of the inertia of the system.
Short-term inertia prediction is a major problem for electrical power systems, as compared to long-term inertia prediction, which ranges from half an hour to one day or one week. When the system is disturbed in a short time, the frequency modulation system does not operate, and the inertia of the system determines the change rate of the frequency, so that the inertia parameter of the unit can seriously influence the dynamic process of the frequency under the power disturbance. The effective detection of the generator frequency response characteristic plays a very important role in the safe and stable operation of the system. With the integration of asynchronous component power electronics into a high capacity ac grid, modern power system inertia compositions are complex and variable. Therefore, effectively and accurately monitoring the inertia of the system is of great importance for frequency stabilization and economical operation of the system.
At present, researches on short-term prediction methods of inertia of a power system mainly comprise a statistical method and a machine learning method. The statistical method mainly comprises regression analysis, differential autoregressive moving average model, minimum variance finite impulse response filter, bayes model and the like. The machine learning method mainly comprises linear regression, support Vector Machines (SVMs), nearest neighbors (KNNs), logistic regression, decision trees, k-means, random forests, naive Bayes, dimension reduction, gradient enhancement and the like. The statistical method-based prediction is based on the assumption that the system inertia is generated only by the synchronous generator, and the method lacks characterization of other complex nonlinear virtual inertia in the novel power system. The machine learning method avoids analysis of complex characteristics of the power grid and derivation of a high-dimensional nonlinear formula, and has better precision when predicting relative to a linear model, but the prediction process of the machine learning model is similar to a black box, the model lacks of interpretability, and the result is difficult to provide relevant interpretation information for specific decisions of a power grid dispatching department.
In order to solve the above-mentioned problems, it is needed to propose a power system inertia interval probability prediction method based on the variational bayesian inference, which helps the GB system find an improved method to estimate the system inertia and provide the integrated inertia. With the help of machine learning and python, complex big data can be well accessed and analyzed, so that the prediction result is more accurate and easier to predict.
Disclosure of Invention
In order to solve the shortages of the existing power system in short-term inertia prediction, the invention provides a power system inertia interval probability prediction method based on the variable decibel leaf inference. The method aims at obtaining the approximate distribution of posterior parameters of a high-dimensional model by adopting variable decibel leaf extrapolation, and gives the predicted probability by using random gradient Hamiltonian Monte Carlo, so as to reliably predict the short-term inertia of the power system and effectively estimate the uncertainty.
The technical scheme adopted by the invention is as follows:
The power system inertia interval probability prediction method based on the variational Bayesian inference is characterized by comprising the following steps:
Collecting historical data, wherein the historical data comprises frequency, load requirements, active output of various new energy units and conventional units, and meteorological data related to the new energy output;
And inputting the acquired data into a constructed power system inertia interval probability prediction model, and predicting the power system inertia interval probability of the prediction model based on a random gradient Hamiltonian Monte Carlo (SGHMC) algorithm.
A power system inertia interval probability prediction system based on variational bayesian inference, comprising:
The first module is configured to collect historical data, wherein the historical data comprises frequency, load requirements, active output of various new energy units and conventional units, and meteorological data related to the new energy output;
the second module is configured to input the acquired data into a constructed power system inertia interval probability prediction model, and the power system inertia interval probability prediction of the prediction model is based on a random gradient Hamiltonian Monte Carlo (SGHMC) algorithm.
A computer medium storing a computer program, the computer program being capable of executing the method steps.
The invention discloses a power system inertia interval probability prediction method based on variable decibel leaf inference, which has the following technical effects:
1) The invention utilizes the advantages of variation Bayesian inference to construct a variation Bayesian inference learning framework suitable for inertia interval probability prediction, and can more accurately estimate the probability distribution of the inertia of the power system. By selecting the characteristic data and considering uncertainty factors, the accuracy and reliability of prediction are improved.
2) The random gradient Hamiltonian Monte Carlo (SGHMC) algorithm is adopted to give interval estimation of inertia instead of single numerical prediction, so that the credibility and the interpretability of a prediction result are enhanced. The user can more clearly understand the uncertainty range of the predicted result.
3) The method has strong adaptability to inertia prediction of the power system, and can process different types of power systems and complex running conditions. Therefore, the method has higher practicability and universality.
Drawings
Fig. 1 is a schematic diagram of the inertia spatial distribution characteristic.
FIG. 2 is a graph of a power system inertia interval probability prediction framework for variational Bayesian inference.
FIG. 3 is a graph of a variational Bayesian inference learning framework for inertia interval probability prediction.
Fig. 4 is a graph of predicted results for 58 training samples on a 30 minute time scale.
Fig. 5 is a graph of predicted results for 150 training samples on a 30 minute time scale.
Fig. 6 is a graph of predicted results for 350 training samples on a 30 minute time scale.
Fig. 7 is a graph of predicted results for 550 training samples on a 30 minute time scale.
Fig. 8 is a graph of the predicted results for 58 training samples on a2 hour scale.
Fig. 9 is a graph of the predicted results of 150 training samples on a2 hour scale.
Fig. 10 is a graph of the predicted results for 350 training samples on a2 hour scale.
Fig. 11 is a graph of the predicted results for 550 training samples on a2 hour scale.
Fig. 12 is a PCC (mean) plot of 550 samples at a 30 minute time scale.
Fig. 13 is a PCC (maximum) plot of 550 samples at a 30 minute time scale.
Fig. 14 is a PCC (mean) plot of 8000 samples at the 2 hour time scale.
Fig. 15 is a PCC (maximum value) plot of 8000 samples at a2 hour time scale.
Fig. 16 is a graph of the lower bound of Evidence (ELBO) iteration progression for 550 sample data.
Fig. 17 is a graph of the lower bound of Evidence (ELBO) iteration progression for 8000 sample data.
Fig. 18 is a graph showing the effect of parameter estimation on the training set before variation estimation of 550 pieces of sample data.
Fig. 19 is a graph showing the effect of parameter estimation on the training set after the variable estimation of 550 sample data.
FIG. 20 is a graph of the effect of parameter inference on a training set before variance inference on 8000 sample data.
FIG. 21 is a graph of the effect of parameter inference on a training set after a change in 8000 sample data has been inferred.
Detailed Description
A power system inertia interval probability prediction method based on a variable decibel leaf inference comprises the following steps:
Step 1: considering inertia space-time distribution, selecting short-term inertia prediction input characteristics of a power system, and establishing a prior probability model and a likelihood function model of each variable to realize deep learning of each variable;
the magnitude of inertia is related to a plurality of factors such as a power grid topological structure, generator parameters, start-stop states, new energy permeability and the like, and influences the frequency response condition of the system. In addition, the magnitude of the disturbance and its location of occurrence and the electrical distance between the generators also affect the system frequency response. Thus, the frequency response and inertia distribution at different locations exhibit spatial variability, and FIG. 1 shows a schematic diagram of the spatial distribution characteristics of inertia.
The power system inertia interval probability prediction method based on the variational Bayesian inference is adopted, a frame diagram of the power system inertia interval probability prediction method is shown in fig. 2, in the step 1,
Firstly, historical data are collected based on a WAMS (wide area measurement management system), meteorological features, conventional units, new energy power, frequency and load requirements are extracted to serve as input features, and in a feature data collection stage, noise data such as missing values, abnormal values and the like are frequently encountered, so that accuracy and reliability of subsequent data analysis and modeling can be affected. Therefore, data preprocessing, including data cleansing and filling, is often required before the data is used. For the missing values, numerical operation and other methods can be adopted for filling so as to ensure the data integrity; for outliers, representative values such as mean values may be substituted for outlier samples to ensure an unbiased estimation of the data. And obtaining a high-quality and reliable data set through a series of preprocessing steps, and taking the data set as a data set trained by a predictive learning model.
Wherein: the sample data set description is shown in the formula (1) to the formula (2), and the input sample set elements correspond to the output sample set elements:
(1)；
(2)；
In the method, in the process of the invention, Is the input sample data set, S 1 is the input sample data 1, S 2 is the input sample data 2; /(I)Is the output sample data set, m 1 is the output sample data 1, m 2 is the output sample data 2; superscript/>Is the dimension of the input feature; subscript/>Is the number of input samples. Wherein/>Represents an input sample.
The total inertia of the power system should be quantified by both the power generation side and the demand side, but in actual cases, the inertia on the demand side is often ignored. However, the demand side inertia is also an important factor against frequency deviation, and its inertia is reflected in the response of the load demand to frequency variation.
Selecting proper characteristics according to a dynamic response equation of the system inertia center frequency, as shown in a formula (3);
(3)；
In the formula (3), the amino acid sequence of the compound, Is the center frequency of inertia of the system; /(I)And/>The inertia and load damping of the system are respectively carried out; /(I)The amount of active imbalance created for the disturbance; /(I)Active adjustment quantity is adjusted for primary frequency modulation of the unit; /(I)Is the system quota frequency; as can be seen from the formula (3), factors influencing the inertia of the system include frequency, load and active power of various energy sources, and meteorological features can influence the output of the wind turbine generator. Therefore, meteorological features, conventional units, and new energy power, frequency and load requirements are considered when selecting input features.
The prior probability model and likelihood function model of each variable are specifically as follows:
consider predicting a target variable from an input vector. Assuming a conditional probability distribution Is a Gaussian distribution, its mean/>Output with CNN neural network/>Related to and determined by. Variance is/>
(4)
In the method, in the process of the invention,Expressed at/>Under the condition of (1)/>Probability distribution of/>Representing output obeying mathematical expectations as/>Variance is/>Is a gaussian distribution of (c).
Also, the prior probability distribution of the weights is setAnd a priori probability distribution of bias/>Are all gaussian distributed in the form of:
(5)
(6)
In the method, in the process of the invention, Representing a gaussian distribution, c is a constant.
N independent observations for the same distributionThe corresponding target set is/>The likelihood function is:
(7)
In the method, in the process of the invention, For input as/>Index corresponding to time,/>In this condition, the output is/>Conditional probability of (2).
Step 2: based on the variable learning models established in the step 1, considering model learning problems, adopting KLqp variation inference algorithm to learn posterior probability distribution of the models, and constructing a variation Bayesian inference learning framework suitable for inertia interval probability prediction;
The model learning problems are mainly as follows:
1) The convergence speed is slow: the deep learning extracts optimal characteristics through forward calculation and back propagation and continuously adjusts parameters so as to achieve the purpose of prediction. The parameters that are adjusted are weight and bias, abbreviated as w and b. Almost all of the effort in deep learning training is to solve for w and b in the neural network. Model training is essentially the process of adjusting w and b, which if initialized to a reasonable value, can increase convergence rate. 2) Overfitting problem: overfitting is widely used in machine learning, and means that after a certain number of iterations, the model accuracy is better and better on the training set, but worse and worse on the test set. The reason for this is that the model learns too many extraneous features, which are considered as the features that the target should possess. 3) Gradient dispersion, unable to use deeper networks: deep learning utilizes forward propagation to extract features while utilizing backward propagation to adjust parameters. When the number of layers of the neural network is large, the gradient is close to 0 when the neural network propagates to the previous layers, and the parameters cannot be adjusted in a guiding way, so that the training effect is basically not achieved.
The construction of a variational Bayesian inference learning framework suitable for inertia interval probability prediction is shown in fig. 3, the input value of the characteristic is related to and determined by the output of a CNN neural network, then the output of the CNN is used as one model parameter affecting the inertia observation value variable of the power system, and simultaneously the variance of the output of the CNN, the weight vector of each layer of the multi-layer neural network and the three model parameters of the bias of each layer of the multi-layer neural network are added to jointly determine the observation value variable to form the inertia prediction learning framework. The method comprises the following steps:
can be obtained according to formulas (4) - (7), And/>The posterior probability distribution of (2) is:
(8)
In the method, in the process of the invention, Expressed in proportion to/>The variance of the output, weight, bias are shown, respectively.
However, the process is not limited to the above-described process,And/>The relationship between them is nonlinear, so the posterior probability is not gaussian. By using KLqp variation inference method, the/>Gaussian approximation of posterior probability distribution of (c).
Then, assume thatThe variational posterior probability distribution of (2) is/>. Decomposing the variational posterior probability distribution into:
(9)
In the method, in the process of the invention, Is the posterior probability distribution of the weights,/>Is a biased posterior probability distribution.
And (3) solving a re-estimation equation of each factor in the distribution by using the general result of the formula. For each factor, the joint probability distribution for all variables is logarithmized, and then the variables that are not in this factor are averaged. From the slaveInitially, only remain andThe term with function dependence yields:
(10)
In the method, in the process of the invention, Is a constant for noise accuracy,/>Is a constant,/>As Euler function,/>For input/>Weight/>Bias/>Hidden variable corresponding to time,/>Variance is/>Mathematical expectation of time,/>Is an identity matrix.
Since the above form is quadraticIs a gaussian distribution. The mean and covariance can be obtained using a balance method. The variational posterior probability distribution is:
(11)
Wherein: mean value is/> Covariance is/>Is a gaussian distribution of (c);
(12)
Step 3: and (3) based on the power system inertia interval probability prediction framework constructed in the step (2), adopting a random gradient Hamiltonian Monte Carlo (SGHMC) algorithm to realize power system inertia interval probability prediction of the prediction model.
The short-term inertia prediction framework of the power system is constructed, and the process of the short-term inertia prediction framework can be roughly divided into three parts, namely data preprocessing, model training and result output, and specifically comprises the following steps: based on the historical data collected by WAMS, meteorological features, conventional units, new energy power, frequency and load requirements are extracted as input features, and a data set for model training is obtained through preprocessing. And establishing a prior probability model and a likelihood function model of each variable, and realizing the deep learning of each variable. And then, adopting KLqp variation inference algorithm to learn posterior probability distribution of the model, and constructing a variation Bayesian inference learning framework suitable for inertia interval probability prediction. And finally, utilizing a random gradient Hamiltonian Monte Carlo (SGHMC) algorithm to realize the probability prediction of the inertia interval of the electric power system of the prediction model.
When training the prediction model, the optimal mode is usedTo approximate the posterior probability distribution. When given a new input, the probability of future inertia can be determined by:
(13)
In the method, in the process of the invention, Represents observations for N independent isodistributions, where/>Corresponding hidden variable,/>The probability distribution of the weight and bias corresponding to the sample D is represented. Sampling is carried out M times by adopting a sampling method, and an average value is taken as a predicted value of probability.
(14)
In the method, in the process of the invention,To change posterior probability distribution,/>Is input.
Sampling by adopting a random gradient Hamiltonian Monte Carlo:
(15)
(16)
(17)
In the method, in the process of the invention, Representing a set of weights and bias parameters employed by random gradient Hamiltonian Monte Carlo sampling,/>The output of the samples is represented, M being the number of samples.
Mean Square Error (MSE) and Mean Absolute Error (MAE) are used to verify the accuracy of the point prediction model. The methods for calculating MSE and MAE are as follows:
(18)
(19)
Where N is the number of samples, As a predicted value/>Is an actual value.
The invention also provides a power system inertia interval probability prediction system based on the variable dB leaf inference, which comprises:
The first module is configured to collect historical data, wherein the historical data comprises frequency, load requirements, active output of various new energy units and conventional units, and meteorological data related to the new energy output;
the second module is configured to input the acquired data into a constructed power system inertia interval probability prediction model, and the power system inertia interval probability prediction of the prediction model is based on a random gradient Hamiltonian Monte Carlo (SGHMC) algorithm.
The operation steps are as described above and are not described in detail herein.
The following is a specific case of using the above method.
The system inertia short-term interval probability prediction is carried out on the British power grid by adopting a power system inertia interval probability prediction model based on the variable decibel leaf inference, actual measurement data of 22 days of the year 12 and 30 days of the year 12 and the month 12 of the year 2020 of the British power grid are firstly collected based on WAMS, the time scale is 30 minutes and 2 hours, the input characteristics are selected from meteorological characteristics, conventional units, new energy power, load demands and the like, and 28-dimensional input characteristic types are collected as shown in table 1.
Table 1 shows a set of input features for system inertia prediction
The front 19-dimensional features are strong feature quantities directly related to inertia, and the rear 9-dimensional features are meteorological data. Training samples of different scales are used for learning, and then the system inertia value within 24 hours in the future is predicted. Predictive graphs using Random Forest (Random Forest) are shown in fig. 4-11. The errors of the predicted results are shown in tables 2 and 3:
TABLE 2 MSE and MAE metrics at 30min prediction scale
TABLE 3 MSE and MAE indicators at 2h prediction scale
Table 2 is the predicted results on a 30 minute sample time scale. As can be seen from table 2, the mean square error and the average absolute error of the system inertia within the future 24 hours are smallest when 275 hours of historical data samples are used for predicting, and the error of the sample prediction result is largest when less than 24 hours. In addition, as training samples increase further, the error of the samples increases. When the number of samples exceeds 5000 hours, the error is not effectively improved although there is an improvement. Table 3 shows the predicted results on the 2 hour sample time scale. From the prediction results, it can be seen that the mean square error and the mean absolute error of the system also decrease with increasing samples. From the foregoing, it can be seen that the time scale is very important for predicting the inertia of an electrical power system. For future predictions of 2 days, to ensure accuracy of the predictions, 1 week of historical data needs to be used for research to achieve better prediction accuracy. When the data acquisition scale is 2 hours, 750 days of history data are required to ensure the prediction accuracy. Thus, from the perspective of prediction accuracy, the smaller the prediction scale, the fewer historical training data sets are required.
To further evaluate the predictive effect of the model, a posterior predictive test (PPC) method was used to evaluate the fitness of the observed data. In this method, posterior predictions are made on the distribution, resulting in observed sample values. If the model matches the observed sample value, the observed sample value of the posterior predictive distribution is the same as the observed sample value. By comparing the model predictions with the observation samples, the suitability of the model for the observation data can be checked. The simplest PPC works by applying test statistics, such as T (xnew) max (xnew), to the new data generated by the posterior prediction. The distribution of test statistic PPD (T) can be obtained by applying T (xnew) to new data on multiple data replications. This distribution is compared to the test statistic applied to the original dataset T (x). PPC forms an empirical distribution of predicted differences:
(20)
In the method, in the process of the invention, Representing new data generated by posterior prediction,/>Representing an empirical distribution of the predicted differences,Representing the empirical distribution of the observation samples.
A scale set of 550 samples on a 30 minute time scale was selected for training. The PPC of the comparative dataset is shown in fig. 12-13. A set of 8000 samples was selected and trained on a 2 hour time scale. The PPC of the comparative dataset is shown in fig. 14-15. According to the graph, the probability prediction model of the inertia interval of the power system based on the power system inferred by the variable decibels has good applicability to the number of samples of both schemes.
The posterior probability distribution is then validated by inference, KLqp defaults to use a re-parameterized gradient to minimize the KL (q p) divergence metric, with the time scale and number of samples selected. Minimizing this divergence measure is equivalent to maximizing the lower bound of Evidence (ELBO). Fig. 16 shows the progress of ELBO during n=550 iterations, with the variance inference converging over about 220 iterations; fig. 17 shows the progression of ELBO in n=8000 iterations, with the variance inference converging in approximately 350 iterations. Fig. 18-21 show a comparison of probability inference results before and after inference of variation of 550 samples and 8000 samples, respectively. From the comparison graph, the two schemes have better fitting effect and are closer to the actual observed value through variation deduction. Therefore, the effectiveness and the advancement of the power system inertia interval probability prediction method based on the variable decibel leaf extrapolation provide a more effective and accurate method for short-term inertia prediction of the power system.
The specific embodiments described herein are offered by way of example only to illustrate the spirit of the invention. Those skilled in the art may make various modifications or additions to the described embodiments or substitutions thereof without departing from the spirit of the invention or exceeding the scope of the invention as defined in the accompanying claims.

Claims (9)

1. The power system inertia interval probability prediction method based on the variational Bayesian inference is characterized by comprising the following steps:
Collecting historical data, wherein the historical data comprises frequency, load requirements, active output of various new energy units and conventional units, and meteorological data related to the new energy output;
Inputting the acquired data into a constructed electric power system inertia interval probability prediction model, and predicting the electric power system inertia interval probability of the prediction model based on a random gradient Hamiltonian Monte Carlo algorithm;
The random gradient Hamiltonian Monte Carlo algorithm realizes prediction, and comprises the following steps:
S3.1: when training the prediction model, the optimal mode is used To approximate the posterior probability distribution; when given a new input, the probability of future inertia can be determined by:
(13)；
In the method, in the process of the invention, Represents observations for N independent isodistributions, where/>Corresponding hidden variable,/>A probability distribution representing the weight and bias corresponding to the sample D;
S3.2: sampling for M times by adopting a sampling method, and taking an average value as a predicted value of probability;
(14)；
In the method, in the process of the invention, To change posterior probability distribution,/>Is input;
s3.3: sampling by adopting a random gradient Hamiltonian Monte Carlo:
(15)；
(16)；
(17)；
In the method, in the process of the invention, Representing a set of weights and bias parameters employed by random gradient Hamiltonian Monte Carlo sampling,/>The output of the samples is represented, M being the number of samples.
2. The power system inertia interval probability prediction method based on variational Bayesian inference as set forth in claim 1, wherein:
Selecting short-term inertia prediction input characteristics of the power system, deeply learning each variable, and establishing a prior probability model and a likelihood function model of each variable;
Based on the established prior probability model and likelihood function model, adopting KLqp variation inference algorithm to learn posterior probability distribution of the model, and constructing a power system inertia interval probability prediction model based on the variational Bayesian inference.
3. The power system inertia interval probability prediction method based on variational Bayesian inference as set forth in claim 1, wherein: selecting short-term inertia prediction input characteristics of a power system, and performing deep learning on variables, wherein the short-term inertia prediction input characteristics comprise
Based on the historical data collected by the WAMS, the meteorological features, the power, the frequency and the load demand of the conventional unit and the new energy are extracted as the data set with input features for the training of the predictive learning model,
Wherein: the sample data set description is shown in the formula (1) to the formula (2), and the input sample set elements correspond to the output sample set elements:
(1)；
(2)；
In the method, in the process of the invention, Is the input sample data set, S 1 is the input sample data 1, S 2 is the input sample data 2; /(I)Is the output sample data set, m 1 is the output sample data 1, m 2 is the output sample data 2; superscript/>Is the dimension of the input feature; subscript/>Is the number of input samples; wherein/>Represents an input sample;
Selecting a set characteristic according to a dynamic response equation of the system inertia center frequency, as shown in a formula (3);
(3)；
In the formula (3), the amino acid sequence of the compound, Is the center frequency of inertia of the system; /(I)And/>The inertia and load damping of the system are respectively carried out; /(I)The amount of active imbalance created for the disturbance; /(I)Active adjustment quantity is adjusted for primary frequency modulation of the unit; /(I)Is the system credit frequency.
4. The power system inertia interval probability prediction method based on variational Bayesian inference as set forth in claim 1, wherein: when constructing the prior probability model and likelihood function model of each variable, consider predicting the target variable from the input vector to define a conditional probability distributionIs a Gaussian distribution, its mean/>Output with CNN neural network/>Related and determined by; variance is/>；
(4)；
In the method, in the process of the invention,Expressed at/>Under the condition of (1)/>Probability distribution of/>Representing output obeying mathematical expectations as/>Variance is/>Is a gaussian distribution of (c);
also, the prior probability distribution of the weights is set And a priori probability distribution of bias/>Are all gaussian distributed in the form of:
(5)；
(6)；
In the method, in the process of the invention, Representing a gaussian distribution, c being a constant;
n independent observations for the same distribution The corresponding target set is/>The likelihood function is:
(7)；
In the method, in the process of the invention, For input as/>Index corresponding to time,/>In this condition, the output is/>Conditional probability of (2).
5. The power system inertia interval probability prediction method based on variational Bayesian inference as set forth in claim 1, wherein:
Setting an input value of a characteristic of an inertia interval probability prediction model of the power system to be related to the output of the CNN neural network and determining the input value by the CNN neural network;
And taking the output of the CNN as a model parameter affecting the inertia observation value variable of the power system, simultaneously adding the variance of the output of the CNN, the weight vector of each layer of the multi-layer neural network and the bias three model parameters of each layer of the multi-layer neural network to jointly determine the observation value variable, and constructing a probability prediction model of the inertia interval of the power system based on the inference of the variable decibels.
6. The power system inertia interval probability prediction method based on the variational Bayesian inference as set forth in claim 1, wherein a KLqp variational inference algorithm is adopted to learn a posterior probability distribution of a model, and the method is characterized in that: comprising
S5.1: can be obtained according to formulas (4) - (7),And/>The posterior probability distribution of (2) is:
(8)；
In the method, in the process of the invention, Expressed in proportion to/>The variance of the output, weight and bias are respectively represented;
the posterior probability is not gaussian; using KLqp variation inference method to calculate Gaussian approximation of posterior probability distribution of (c);
S5.2: assume that The variational posterior probability distribution of (2) is/>; Decomposing the variational posterior probability distribution into:
(9)；
In the method, in the process of the invention, Is the posterior probability distribution of the weights,/>Is a biased posterior probability distribution;
solving a re-estimation equation of each factor in the distribution by using a general result of a formula; for each factor, taking the logarithm of the joint probability distribution of all variables, and then averaging the variables that are not in this factor;
S5.3: from the slave Initially, only AND/>The term with function dependence yields:
(10)；
In the method, in the process of the invention, Is a constant for noise accuracy,/>Is a constant,/>As Euler function,/>For input/>Weight/>Bias/>Hidden variable corresponding to time,/>Variance is/>Mathematical expectation of time,/>Is a unit matrix;
is Gaussian distribution, and a mean value and covariance are obtained by using a balance method; the variational posterior probability distribution is:
(11)；
Wherein: mean value is/> Covariance is/>Is a gaussian distribution of (c);
(12)。
7. The method for predicting the inertia interval probability of the electric power system based on the variable decibels inference according to claim 1, further comprising the step of adopting a mean square error MSE and an average absolute error MAE to verify the prediction result of a point prediction model, wherein the calculation method of the MSE and the MAE is as follows:
(18)；
(19)；
Where N is the number of samples, As a predicted value/>Is an actual value.
8. A power system inertia interval probability prediction system based on variational bayesian inference, comprising:
The first module is configured to collect historical data, wherein the historical data comprises frequency, load requirements, active output of various new energy units and conventional units, and meteorological data related to the new energy output;
The second module is configured to input the acquired data into a constructed power system inertia interval probability prediction model and predict the power system inertia interval probability of the prediction model based on a random gradient Hamiltonian Monte Carlo algorithm;
The random gradient Hamiltonian Monte Carlo algorithm realizes prediction, and comprises the following steps:
S3.1: when training the prediction model, the optimal mode is used To approximate the posterior probability distribution; when given a new input, the probability of future inertia can be determined by:
(13)；
In the method, in the process of the invention, Represents observations for N independent isodistributions, where/>Corresponding hidden variable,/>A probability distribution representing the weight and bias corresponding to the sample D;
S3.2: sampling for M times by adopting a sampling method, and taking an average value as a predicted value of probability;
(14)；
In the method, in the process of the invention, To change posterior probability distribution,/>Is input;
s3.3: sampling by adopting a random gradient Hamiltonian Monte Carlo:
(15)；
(16)；
(17)；
In the method, in the process of the invention, Representing a set of weights and bias parameters employed by random gradient Hamiltonian Monte Carlo sampling,/>The output of the samples is represented, M being the number of samples.
9. A computer medium, characterized in that it stores a computer program capable of executing the method steps of any one of claims 1-7.
CN202410559328.0A 2024-05-08 2024-05-08 Power system inertia interval probability prediction method based on variable decibel leaf inference Active CN118133161B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202410559328.0A CN118133161B (en) 2024-05-08 2024-05-08 Power system inertia interval probability prediction method based on variable decibel leaf inference

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202410559328.0A CN118133161B (en) 2024-05-08 2024-05-08 Power system inertia interval probability prediction method based on variable decibel leaf inference

Publications (2)

Publication Number Publication Date
CN118133161A true CN118133161A (en) 2024-06-04
CN118133161B CN118133161B (en) 2024-07-12

Family

ID=91231057

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202410559328.0A Active CN118133161B (en) 2024-05-08 2024-05-08 Power system inertia interval probability prediction method based on variable decibel leaf inference

Country Status (1)

Country Link
CN (1) CN118133161B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115759380A (en) * 2022-11-11 2023-03-07 三峡大学 Short-term prediction method for inertia of power system based on SHAP-XGboost algorithm
US20230213895A1 (en) * 2021-12-30 2023-07-06 Huaneng Shanghai Combined Cycle Power Co, Ltd. Method for Predicting Benchmark Value of Unit Equipment Based on XGBoost Algorithm and System thereof
CN117807879A (en) * 2023-12-29 2024-04-02 东南大学 New energy power system inertia and virtual inertia analysis method
CN117972343A (en) * 2024-01-08 2024-05-03 云南电网有限责任公司昆明供电局 Voltage sag probability assessment method and system based on Hamiltonian Monte Carlo

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20230213895A1 (en) * 2021-12-30 2023-07-06 Huaneng Shanghai Combined Cycle Power Co, Ltd. Method for Predicting Benchmark Value of Unit Equipment Based on XGBoost Algorithm and System thereof
CN115759380A (en) * 2022-11-11 2023-03-07 三峡大学 Short-term prediction method for inertia of power system based on SHAP-XGboost algorithm
CN117807879A (en) * 2023-12-29 2024-04-02 东南大学 New energy power system inertia and virtual inertia analysis method
CN117972343A (en) * 2024-01-08 2024-05-03 云南电网有限责任公司昆明供电局 Voltage sag probability assessment method and system based on Hamiltonian Monte Carlo

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
DANIELE LINARO ET AL.: "Continuous estimation of power system inertia using convolutional neural networks", NAT COMMUN 14, vol. 4440, no. 2023, 24 July 2023 (2023-07-24), pages 1 - 14 *
YANG HAOXIN ET AL.: "The Effects of Virtual Inertia Control on Power Converters in Nonideal Grid Conditions", INTERNATIONAL POWER ELECTRONICS CONFERENCE (IPEC-HIMEJI - ECCE ASIA), 19 May 2022 (2022-05-19), pages 122 - 126, XP034141619, DOI: 10.23919/IPEC-Himeji2022-ECCE53331.2022.9806895 *
张磊等: "基于可解释性XGBoost的电力系统惯量短期预测方法", 电力建设, vol. 44, no. 08, 1 August 2023 (2023-08-01), pages 22 - 30 *
杨柳等: "基于Copula函数的风-电-热相关性及其潜在不确定性分析", 电气工程学报, no. 02, 25 June 2020 (2020-06-25), pages 57 - 63 *

Also Published As

Publication number Publication date
CN118133161B (en) 2024-07-12

Similar Documents

Publication Publication Date Title
Cao et al. Switching Markov Gaussian models for dynamic power system inertia estimation
JP6645043B2 (en) Error width estimation device, error width estimation system, error width estimation method, and program
Wang et al. Multi-objective unit commitment with wind penetration and emission concerns under stochastic and fuzzy uncertainties
Zhang et al. Wind speed prediction research considering wind speed ramp and residual distribution
CN110829444B (en) Power grid receiving end system emergency load shedding method considering static frequency and voltage characteristics
CN115759380A (en) Short-term prediction method for inertia of power system based on SHAP-XGboost algorithm
CN105429134A (en) Grid voltage stability prediction method based on power big data
CN110969306A (en) Power distribution low-voltage distribution area load prediction method and device based on deep learning
CN109144986A (en) A kind of importance appraisal procedure of industrial equipment data
CN108073782A (en) A kind of data assimilation method based on the equal weight particle filter of observation window
CN112465266A (en) Bus load prediction accuracy analysis method and device and computer equipment
Li et al. Precision and accuracy co-optimization-based demand response baseline load estimation using bidirectional data
CN112149976A (en) Power grid accurate investment project decision method based on artificial intelligence
CN117851908B (en) Improved on-line low-voltage transformer area electric energy meter misalignment monitoring method and device
CN118133161B (en) Power system inertia interval probability prediction method based on variable decibel leaf inference
CN112330121B (en) Power distribution network automation construction investment benefit evaluation method
Maheu et al. How useful are historical data for forecasting the long-run equity return distribution?
CN117330963A (en) Energy storage power station fault detection method, system and equipment
CN115130743B (en) Wind turbine generator set regulation rate prediction method and system based on variation inference
CN116565861A (en) Power distribution network reliability assessment method, system, equipment and medium
CN113793077B (en) Method and system for analyzing power failure influence of power distribution network user fault
Reis et al. A fast and accurate sampler built on Bayesian inference and optimized Hamiltonian Monte Carlo for voltage sag assessment in power systems
Shi et al. Study on stability feature extraction of power system using deep learning
Artyushenko et al. Synthesis of Algorithms of Adaptive Signal Processing for Tracking Meters Using Nonlinear Blocks with Feed-Back
CN115378122B (en) Kalman filtering-based power system frequency and inertia real-time monitoring method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant