CN116742623B - Dynamic state estimation method and system based on model data dual-drive - Google Patents

Abstract

The invention discloses a dynamic state estimation method and system based on model data dual driving, which includes: obtaining real-time node power measurement information of the power system, and identifying abnormal information at the current moment, and the corresponding node is an abnormal point; using data-driven prediction The model obtains the power prediction value at the abnormal point; based on the model-driven prediction model, the power prediction value at the abnormal point is obtained; the predicted value is learned to obtain the final power prediction result at the abnormal point; the final power prediction result is used as the pseudo power prediction at the current moment. Quantity measurement, replacing the abnormal information at the current moment, using all the node power measurement information at the current moment after replacing the abnormal information for state estimation, realizing dynamic state estimation based on model data dual-driver. The invention improves the reliability of input data and the accuracy of state estimator measurement, and can better reflect the actual power system state.

Description

A dynamic state estimation method and system based on dual drive of model and data

Technical Field

The present invention belongs to the technical field of dynamic state estimation of power systems, and relates to a dynamic state estimation method and system based on dual drive of model data.

Background Art

Power system state estimation is the prerequisite for maintaining the safe operation of the power system. However, in the actual distribution network, due to the lack of measurement devices and random errors, the measurement data has problems such as missing data and bad data, which reduces the reliability of state estimation. There is currently a method for detecting bad data using quadratic linear state estimation, but the computational efficiency of this method is not high and it is difficult to meet the real-time requirements of the distribution network for power system detection quantities. Therefore, there is an urgent need to detect anomalies in real-time measurement data.

For the detected abnormal data, it is necessary to make predictions based on historical data. Conventional prediction methods include model-driven and data-driven. Model-driven is based on physical models, has strong explanatory power and causal logic, and has good generalization performance. However, since the main factors considered by model-driven are load factors and consumption requirements, the prediction accuracy of the model is low. The basic principle of data-driven deep neural networks is to learn deep features from a large number of high-dimensional samples based on deep nonlinear neural networks to achieve the approximation of complex functions. The prediction accuracy is higher than that of model-driven, but data-driven is highly dependent on data, so it has strict requirements on data quality, and its generalization performance is weaker than that of model-driven. Therefore, it is necessary to find a method that can integrate the two driving methods to achieve data prediction that takes into account both accuracy and generalization performance, and provide strong data support for the subsequent state estimation with higher accuracy.

At the same time, in order to solve the problem of non-Gaussian measurement noise in the distribution network, the conventional state estimation method based on the Kalman filter framework assumes that all noises, including conditional probability and joint probability, are Gaussian distributed. Although it is easy to obtain an analytical solution in the state estimation calculation process, for nonlinear non-Gaussian dynamic systems, all probabilities may not be Gaussian distributed, which is especially common in distribution networks. Therefore, it is impossible to obtain an analytical solution in state estimation. The conventional state framework based on Kalman filtering is not applicable to distribution networks. Therefore, in order to solve the problem of non-Gaussian noise in distribution network measurement data, it is urgent to find a model that can solve the problem that Kalman filtering is easily limited by noise models. This research has good practical application value.

Summary of the invention

In order to solve the deficiencies in the prior art, the present invention provides a dynamic state estimation method and system based on dual drive of model data, which can perform anomaly detection on real-time measurement data, integrate the model data prediction drive method to achieve accurate prediction of measurement, solve the problem of non-Gaussian noise in the distribution network, and improve the accuracy of power system state estimation.

The present invention adopts the following technical solution.

A dynamic state estimation method based on dual drive of model and data comprises the following steps:

S1: Obtain the real-time node power measurement information of the power system, and calculate the abnormal score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point;

S2: Use historical node power measurement information and a pre-trained data-driven prediction model based on an LSTM neural network to obtain the power prediction value at the abnormal point;

S3: Perform real-time load classification prediction based on the model-driven prediction model to obtain the power prediction value at the abnormal point;

S4: Combine historical node power measurement information to learn the predicted values obtained by S2 and S3 to obtain the final power prediction result at the abnormal point;

S5: The final power prediction result is used as the pseudo-measurement at the current moment to replace the abnormal information at the current moment, and the power measurement information of all nodes at the current moment after replacing the abnormal information is used for state estimation to realize dynamic state estimation based on dual drive of model data.

Preferably, in S1, the node power measurement information includes node injection power and line power, wherein the node injection power includes the injected active power and reactive power of the node, and the line power includes the active power and reactive power of the line.

Preferably, in S1, the process of identifying abnormal information includes:

1) The historical node power measurement information is divided into four measurement sets: injected active power and reactive power, line active power and reactive power. For each set, M information samples are randomly selected and used as the root node of the tree. A cutting point is randomly generated. The sample is divided into two parts based on the size of the cutting point information and placed on both sides of the cutting point. The cutting is repeated in the child nodes until only one information data is left or the child node reaches a limited height. The cutting point is randomly generated and the cutting operation is repeated in a loop to generate K trees.

2) Calculate the average height of the K trees generated by each set, and then calculate the anomaly score of the sample information in the set. If the anomaly score exceeds the set threshold, it means that the sample information is abnormal.

Preferably, for the measurement set g, its sample information anomaly score for:

(3)

(2)

Where h(g) is the average of the path length from the root node to the node where each sample point in the tree corresponding to the measurement set g passes with respect to the number of samples m;

E(h(g)) is the average value of h(g) of K trees.

is the average height of m samples;

, is Euler's constant.

Preferably, in S2, when pre-training the data-driven prediction model based on the LSTM neural network, the formula followed by the information transmission within the neurons of the LSTM neural network is as follows:

(5)

(6)

(7)

(8)

(9)

(10)

in, Represent the forget gate, input gate and output gate respectively;

, , Represent the weights of the forget gate, input gate, and output gate respectively;

, , , The gated unit biases representing the forget gate, input gate, output gate, and neural unit states, respectively;

Represents the input vector of the LSTM unit;

, Represent the states of neurons at the previous moment and the current moment respectively;

Represents the state of the input node at the current time t;

, Represent the hidden layer state variables at the previous moment and the current moment respectively;

and tanh represent Sigmoid function and Tanh function respectively.

Preferably, in S3, a model-driven prediction model is used to obtain the predicted value of injected active power at the abnormal point according to the load type of the node where the abnormal information is located. Based on the power factor, the predicted value of injected reactive power at the node where the abnormal information is located is calculated, and the predicted value of the branch power information where the abnormal point is located is obtained through flow calculation.

Preferably, the predicted value of injected active power at the abnormal point is:

(12)

(11)

in, is the injected active power of node i where the abnormal information is located at time t;

C represents the number of load types of node i where the abnormal information is located;

is the total number of nodes served by the upstream substation node r of the node i where the abnormal information is located at time t;

represents the active power of the load of type p at node i where the abnormal information is located at time t;

represents the telemetry power of the upstream substation node r of the node where the abnormal information is located at time t;

represents the load model coefficient of the load type p at time t;

represents the average value of the load model coefficient;

Represents the consumption demand of load of type p at node i.

Preferably, the predicted value calculation formula of reactive power injected into the node where the abnormal information is located is:

in, is the injected reactive power of node i where the abnormal information is located at time t;

represents the active power of the load of type p at node i where the abnormal information is located at time t;

C represents the number of load types of node i where the abnormal information is located;

is the power factor of load type p at node i.

Preferably, the specific process of S4 is:

The power prediction values at the abnormal points obtained by S2 and S3 are used to replace the abnormal information in the real-time node power measurement information of S1, and combined with the historical node power measurement information in S2 to form two large training sets. The following prediction operations are performed on each large training set: the training set is divided into several small groups for training multiple base learners, where the number of small groups is the same as the number of base learners; each small group of data is used to train the corresponding base learner, and each base learner outputs its own prediction results;

The outputs of all base learners are used as the input of the meta-learner for training, and the output of the meta-learner is used as the power prediction value at the anomaly point.

Preferably, in S5, the power measurement information of all nodes at the current moment after replacing the abnormal information is used. The model for state estimation is:

(13)

(14)

in, represents the estimated value of the state at time t;

represents the state transition function at time t;

, represent process noise and measurement noise respectively;

It is a nonlinear mapping function from the particle filter-based quantity measurement to the state quantity.

The nonlinear mapping logic from quantity measurement to state quantity is:

The state estimation result is represented by the state posterior probability, and the posterior probability is obtained by discrete sample points obtained by random sampling, and then the state estimation result represented by the particle weight polynomial is obtained; the particle weight is solved and the items with particle weight less than the set value are removed, and the remaining items are summed to obtain the final state estimation result.

A dynamic state estimation system based on dual drive of model and data, comprising:

The information acquisition and anomaly identification module is used to obtain the real-time node power measurement information of the power system, and calculate the anomaly score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point;

The data-driven prediction module is used to obtain the power prediction value at the abnormal point by using the historical node power measurement information and the pre-trained data-driven prediction model based on the LSTM neural network;

The model-driven prediction module is used to perform real-time load classification prediction based on the model-driven prediction model to obtain the power prediction value at the abnormal point;

An integrated learning module is used to combine historical node power measurement information to learn the predicted values obtained by the data-driven prediction module and the model-driven prediction module to obtain the final power prediction result at the abnormal point;

The state estimation module is used to use the final power prediction result as the pseudo-measurement at the current moment to replace the abnormal information at the current moment, and use the power measurement information of all nodes at the current moment after replacing the abnormal information for state estimation, so as to realize dynamic state estimation based on dual drive of model data.

A terminal includes a processor and a storage medium; the storage medium is used to store instructions;

The processor is configured to operate according to the instructions to execute the steps of the method.

A computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

The beneficial effects of the present invention are as follows:

1) The present invention takes into account the problem that distribution network measurement data is prone to abnormalities, and combines historical node power measurement information to calculate the abnormal score of real-time node power measurement information to detect missing data and erroneous data, thereby improving the reliability of input data;

2) The present invention integrates data-driven prediction based on LSTM neural network, model-driven prediction and integrated learning to achieve model data hybrid prediction, and predicts node power from two parts: data trend and actual physical meaning. The data-driven method of the present invention can dig deeper into the characteristics contained in the data, and finally achieve prediction, which improves the generalization potential of the model. The model-driven method of the present invention establishes a model with actual physical meaning based on the detection information of the substation, which solves the problem of high data quality requirements of traditional data-driven methods. Through the final integrated learning of the dual-driven model, better pseudo-quantity measurements are obtained, which reduces the impact of bad data on the state estimation results and improves the accuracy of state estimation quantity measurements;

3) Taking into account that the traditional Kalman filtering method is limited by the noise model and is difficult to process the distribution network measurement data under non-Gaussian noise interference, the present invention introduces a particle filter to process the noise, and performs state estimation based on the particle filter, so that the noise distribution is not limited by the model and can better reflect the actual power system state.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a dynamic state estimation method based on dual-drive of model data according to the present invention;

Figure 2 is a structural diagram of a standard IEEE 33 node distribution network system;

Figure 3 is a diagram showing the filtering results of different nodes under the influence of Laplace noise;

Figure 4 is a comparison of the estimation results of different algorithms in the presence of abnormal data.

DETAILED DESCRIPTION

In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

As shown in FIG1 , Embodiment 1 of the present invention provides a dynamic state estimation method based on dual drive of model data. In a preferred but non-limiting embodiment of the present invention, the method comprises the following steps:

S1: Obtain the real-time node power measurement information of the power system, and calculate the abnormal score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point;

S1.1: Obtain real-time node power measurement information of the power system;

This step obtains the real-time initial information of the power system, mainly the node power measurement information;

The node measurement information set can be expressed as follows:

(1)

in, is the node power measurement information set, which represents the set of all quantity measurements of node i at time 0-T;

is the line between node i and node j;

N and T are the number of nodes and measurement period, respectively;

and is the node injection power, which represents the injected active power and reactive power of node i at time t respectively;

and is the line power, respectively represents the line power at time t of active power and reactive power.

S1.2: Calculate the abnormal score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point;

Abnormal information mainly manifests itself in missing data and data anomalies.

This step can quickly detect abnormal data in big data. It is mainly divided into two stages: training and evaluation. The details are as follows:

① In offline training, the historical node power measurement information is divided into , , and Four measurement sets, for each set, randomly select M information samples from it and use it as the root node of the tree, randomly generate a cutting point o, divide the sample into two parts based on the information size and place them on both sides of the cutting point, where the node with smaller data than the cutting point is placed on the left side of the cutting point, and vice versa, on the right side, and repeat the cutting in the remaining child nodes until there is finally one information data left or the child node reaches the limited height; where all nodes except the root node are called child nodes; usually, the maximum value (limited height) can be set first, and the value is generally the average height of the tree. For mature isolation forest anomaly detection algorithm packages, the limited height can be automatically determined by the program, so it is not explained in the present invention.

Loop through the randomly generated cutting points and the above operations, that is, loop through the randomly generated cutting points, divide the sample into two parts based on the cutting point information size and place them on both sides of the cutting point, and repeat the cutting in the child nodes until only one information data is left or the child node reaches the specified height, generating K trees. According to experience, the number of trees is usually set to 100.

② In the evaluation, the real-time node power measurement information is divided into , , and The four measurement sets are substituted into the corresponding tree sets one by one to calculate the average height, and then the anomaly score of the sample information in the set is calculated. If the anomaly score exceeds the set threshold, it means that the sample information is abnormal.

The real-time node power measurement information is divided into four measurement sets: injected active power and reactive power, line active power and reactive power. For each measurement set, the operation in 1) is performed to generate four sets containing K trees. The average height of each set of trees is calculated, and then the abnormal score of the sample information in the set is calculated. If the abnormal score exceeds the set threshold, it means that the sample information is abnormal.

The closer the anomaly score of a sample is to 1, the higher the possibility that it is abnormal information; if the anomaly score is less than 0.5, it can basically be judged as normal data.

Specifically: For each measurement set g, the average height of the m samples in it With anomaly score The calculation is as follows:

(2)

(3)

in, , is Euler's constant;

h(g) is the average length of the path from the root node to the node where each sample point in the measurement set g is located with respect to the number of samples m;

E(h(g)) is the average value of the path length calculated by the measurement set g in K trees with respect to the number of trees, that is, the average value of h(g) of T trees.

S2: Use historical node power measurement information and a pre-trained data-driven prediction model based on an LSTM neural network to obtain the power prediction value at the abnormal point;

Specifically obtain historical node power measurement information and historical node status information, where the historical node status information is used for subsequent state transfer functions , the historical information is input into the data-driven prediction model based on the LSTM neural network for training, and a single-variable single-step prediction is performed to obtain the predicted value of the state variable at the current time t. The node state information includes the node voltage amplitude and phase angle, which is specifically expressed as:

, and there is (4)

in, It is a collection of node status information;

N and T are the number of nodes and measurement period, respectively;

and They represent the voltage amplitude and phase angle of node i at time t respectively.

In terms of data-driven prediction, we build an LSTM neural network, train the LSTM (Long Short Term Memory, LSTM) neural network with the historical power data of each node, predict missing and abnormal data, and obtain the predicted value of the load power at the abnormal point, so as to realize model prediction, that is, by building a long-term and short-term neural network model, we train the neural network based on historical power data to realize the prediction and correction of the abnormal node injection power in the power system state variable. Specifically:

Input the first historical node power measurement information set into a data-driven prediction model based on an LSTM neural network for pre-training, and input the second historical node power measurement information set into the pre-trained data-driven prediction model based on an LSTM neural network to obtain a power prediction value at the abnormal point;

In specific implementation, S1 can choose to use the historical data of 2010, S2 uses the historical data of 2010 as the first historical node power measurement information set for model pre-training and uses the historical data of 2011 for prediction, and S3 uses the historical data of 2010.

In formulas (5)-(10), Z historical measurement data represents input data. Usually, o in the last module represents the output abnormal information prediction value, that is, the power prediction value at the abnormal point. Formulas (5)-(10) can be used to , , and These four values are the quantity measurements at the current time t Make predictions.

When pre-training a data-driven prediction model based on an LSTM neural network, the formula followed by information transfer within the neurons of the LSTM neural network is as follows:

(5)

(6)

(7)

(8)

(9)

(10)

in, Represent the forget gate, input gate and output gate respectively;

, , Represent the weights of the forget gate, input gate, and output gate respectively;

, , , The gated unit biases representing the forget gate, input gate, output gate, and neural unit states, respectively;

Represents the input vector of the LSTM unit;

, Represent the states of neurons at the previous moment and the current moment respectively;

Represents the state of the input node at the current time t;

, Represent the hidden layer state variables at the previous moment and the current moment respectively;

and tanh represent Sigmoid function and Tanh function respectively.

S3: Perform real-time load classification prediction based on the model-driven prediction model to obtain the power prediction value at the abnormal point;

In terms of model-driven prediction, the load model coefficient curve after user classification is combined to perform real-time classification prediction of the load, and obtain the active power and reactive power of the total load at the abnormal point.

The model-driven prediction model has the relationship (11)-(12) for the injected active power of node i where the abnormal information is located at time t.

In step S3, a real-time classification model prediction model for load is constructed in combination with the user classification curve. The process can be described as follows:

By telemetering the power flow data of the upstream substation node r, the downstream abnormal information node at time t is calculated according to the load model factor (LMF) The power demand of class p load is calculated as follows:

(11)

in, is the total number of nodes served by the upstream substation node r of the node i where the abnormal information is located at time t;

It represents the active power of the load of type p at the node i where the abnormal information is located at time t. In this paper, time t represents any time. The constant relationship reflected by this physical expression has nothing to do with the specific time t.

represents the telemetry power of the upstream substation node r of the node where the abnormal information is located at time t;

represents the load model coefficient of the load type p at time t;

It represents the average value of the load model coefficient. This average value represents the daily variation of the typical load. The specific average value at time t can be obtained by consulting the improved daily load variation curve of different types of loads at time t. The improved daily load curve is obtained from statistical analysis based on historical data. It is usually a function of season, temperature, and the number of days in a week (this function is the improved daily load curve). That is, the curve is related to season, temperature, and the number of days in a week. The specific correlation cannot be described by a specific function, but is only predicted at the macro level. In actual calculation, there is no need to Calculation is performed by directly consulting the improved daily load change curves of different types of loads at time t. Because it is an improved curve, the horizontal coordinate of the corresponding curve is time t, and the vertical coordinate is E[LMF].

The load model coefficient can be obtained from a curve of daily load power variation of a certain type of load, which is usually obtained by consulting literature or statistically analyzing relevant historical data.

It represents the consumption demand of load of type p at node i, where the consumption demand is calculated as the number of kilowatt-hours per month/days of the billing cycle;

The total predicted power at node i where the abnormal information is located is the sum of the power of all types of loads, expressed as follows:

(12)

in, is the injected active power of node i where the abnormal information is located at time t;

C represents the number of load types of node i where the abnormal information is located.

Input the above-mentioned relevant load information of the node where the abnormal information is located into the above-mentioned model-driven prediction model to realize real-time load classification prediction and obtain the abnormal information prediction value ;

Based on the power factor of the corresponding load type obtained from experience or historical statistics, the abnormal information is calculated The predicted value is calculated as:

in, is the injected reactive power of node i at time t;

represents the active power of the load of type p at node i where the abnormal information is located at time t;

C represents the number of load types of node i where the abnormal information is located;

is the power factor of load type p at node i.

Finally, the abnormal branch power information is obtained through power flow calculation The predicted value of is:

In the formula, 1 j represents an imaginary number, and there is , , , ;

and are the injected active power and reactive power of node i at time t respectively;

and Line Conductance and susceptance;

and are the voltage amplitude and phase angle of node i at time t respectively;

and They represent the lines at time t of active power and reactive power.

S4: Combine historical node power measurement information to learn the predicted values obtained by S2 and S3 to obtain the final power prediction result at the abnormal point;

Construct an integrated learning model, including the base learner in the first layer and the meta learner in the second layer;

The integrated learning model uses the data obtained in step S2 and step S3 and the historical node power measurement information to train the base learner of the first layer, and uses the output of the base learner as the input of the meta learner for training. The final result predicted by the output of the meta learner is as follows:

The power prediction values at the abnormal points obtained by S2 and S3 are used to replace the abnormal information in the real-time node power measurement information of S1, and combined with the historical node power measurement information obtained in S2 (the second historical node power measurement information set) to form two large training sets, that is, the power prediction values at the abnormal points obtained by S2 replace the abnormal information in the real-time node power measurement information of S1, and combine with the historical node power measurement information obtained in S2 to form a large training set; the power prediction values at the abnormal points obtained by S3 replace the abnormal information in the real-time node power measurement information of S1, and combine with the historical node power measurement information obtained in S2 to form another large training set, and then perform the following prediction operations on each large training set:

In order to avoid overfitting, the model is usually trained by cross-validation in the base learner stage, that is, the training set is divided into several groups for training multiple base learners in the first layer. Each group of data trains the corresponding base learner, and each base learner outputs its own prediction result; in the second-layer meta-learner, the output of all base learners in the first layer is used as the input of the meta-learner for training, and the output of the meta-learner is used as the final prediction result, which is the overall prediction result based on the model and data model. The training set is divided into several groups for training multiple base learners, where the number of groups is the same as the number of base learners. This patent selects the common 5-fold cross-validation method, that is, the training set is randomly divided into 5 non-overlapping groups of basically equal size; each group of data is used to train the corresponding base learner, and each base learner outputs its own prediction result; the output of all base learners is used as the input of the meta-learner for training, and the output of the meta-learner is used as the power prediction value at the abnormal point. The multiple prediction results obtained from the two training sets are input into the meta-learner of the second layer to obtain a prediction result, which is the final prediction result.

S5: The final power prediction result is used as the pseudo-measurement at the current moment to replace the abnormal information at the current moment, and the power measurement information of all nodes at the current moment after replacing the abnormal information is used for state estimation to realize dynamic state estimation based on dual drive of model data.

All the measurement information at the current moment after replacing the abnormal information is used as the system input for state estimation. In this process, the particle filtering method is used to better fit the actual distribution network noise and achieve a more accurate state estimation of the power system.

Further preferably, in order to solve the problem that the measured data of the actual distribution network is interfered by non-Gaussian noise, the power measurement information of all nodes at the current moment after replacing the abnormal information is used. The model for state estimation is:

(13)

(14)

in, It represents the estimated value of the state at time t. In this article, time t represents any time, reflecting a unified and unchanging physical constraint, which can be narrowly understood as the prediction time.

It represents the state transfer function at time t. The process of state transfer can be expressed as follows: input the historical state data at time t-1 before into LSTM for training, and perform single-variable single-step prediction to obtain the predicted value of the state variable at the current time t;

, They represent process noise and measurement noise respectively. Process noise can be selected as Cauchy distribution noise, and measurement noise can be selected as Laplace noise in the implementation example. Process noise refers to the noise of the design system itself.

It is a nonlinear mapping function from measured quantity to state quantity.

Using particle filtering to realize nonlinear equations from quantity measurement to state quantity The specific process of modeling is as follows:

right As an important issue in state estimation, the particle filter is used here for calculation. The main idea is to simulate the probability density function of the system random variable through discrete random sampling points, and average the results to finally obtain the minimum variance estimate of the state. First, it needs to be clear that for equation (14), the known quantity is measured , and what needs to be solved is the state quantity .

Secondly, it should be noted that and The description is a mapping relationship, which is difficult to describe in mathematical language. It can be understood as a process. It can be understood as a black box processing process (the real state quantity data at time t is input, and the output is obtained after the mapping and superposition of noise ), but what the present invention actually performs is the inverse transformation of this process, that is, knowing the quantity measurement after superimposing noise, how to obtain the true state quantity, and the specific process is the following process (1)-(4).

At the same time, for the particle filtering method, the state estimation result can be used To express, that is what you want .

The nonlinear mapping logic from quantity measurement to state quantity is:

The state estimation result is represented by the state posterior probability, and the posterior probability is obtained by randomly sampling discrete sample points, and then the state estimation result represented by the particle weight polynomial is obtained; the particle weight is solved and the items with particle weight less than the set value are removed, and the remaining items are summed to obtain the final state estimation result. The specific process is (1)-(4):

(1) The state estimation result of the system can be expressed by the posterior probability of the system state, that is, it can be approximated by a set of weighted samples as follows:

(15)

in, , W is the W samples sampled in the posterior probability, Refers to the status information at time 0:T. Refers to 1: measurement information at time T, refers to the importance probability density function, refers to the wth sample of the posterior probability, for The process is to input the state information at time 0:T into LSTM to obtain the state output at the current time;

(2) In practical applications, it is difficult to directly obtain the posterior probability We extract samples from the distribution by using Bayesian importance sampling, which introduces a simple, easy-to-sample proposal distribution , that is, random sampling obtains discrete sample points to approximate the posterior probability, then the function The expectation can be rewritten as

(16)

in, represents the weight of each particle, and the solution process is as follows (3).

(3) Solve the weight of the particle in equation (16):

(3.1), first, The importance probability density function is decomposed into the following form.

(17)

(3.2), then the recursive form of the posterior probability density function can be expressed as follows:

(18)

It can be seen that the final right-hand side formula has a temporal connection with the left-hand side;

(3.3), Substituting equation (17) and equation (18) into the particle weight formula in (2), the recursive form of the particle weight can be expressed as:

(19)

(4) Through the polynomial resampling method, the particles with large weights are copied and the particles with small weights are deleted, and finally , which is the minimum variance estimate of the system state Specifically: According to formula (19) Calculate the formula (16) , replace formula (16) Items less than the set value Remove it and sum the remaining terms to get , That is , that is, the state estimation obtained in S5, which is also the state estimation value obtained by the present invention.

Application examples:

Figure 2 shows a standard IEEE 33-node distribution network system with a reference voltage of 12.66KV. The state estimation steps for this distribution network system are as follows:

1) The load data of 2010 is selected as the measurement training set. Lappass noise is superimposed on all measurement data. The parameter of the noise is 0 in mean and the variance matrix is , detect the abnormal data at the current moment according to formulas (3)-(4).

2) Based on the internal relationship of LSTM, refer to formulas (5)-(10), input the load data of 2010 as the training set into LSTM for training, and then input the measurement of 2011 as the test set into LSTM for prediction, and calculate the power prediction at the abnormal point under data-driven;

Generate daily load curves based on historical data from 2010 and obtain load model coefficients The average daily curve is substituted into formulas (11)-(12) to calculate the power prediction at the abnormal point driven by the model;

Combining the prediction results of the data and the model, the final predicted power at the abnormal point is obtained through ensemble learning, and the final predicted power data is used as a pseudo-measurement to replace the abnormal data.

3) Perform state estimation on the distribution network. The specific process refers to formulas (15)-(19), and output the final result of the state estimation. .

In the example, the root mean square error (RMSE) and absolute error (AE) are used to evaluate the results of the state error.

Figure 3 shows the filtering effect of nodes 7 and 23 under Laplace noise. It can be seen that after the method of the present invention, different noises have little effect on the final state estimation, and the calculated results are almost the same as the actual results. At the same time, Table 1 shows the estimation performance analysis under the influence of noise. It can be seen from the table that the AE value of the node is Order of magnitude, RMSE is also The accuracy of state estimation is very high and can meet the requirements of practical systems.

Table 1 Estimation results of the method proposed in this invention under the influence of Laplace noise

To further illustrate the accuracy of the proposed method when there are abnormal data, the following scenario is set for comparison: Assume that the measurement information of nodes 7, 20, and 23 is randomly missing, and the error from sample sequence 9 to sample sequence 23 is 20%. The proposed method is compared with the UKF (Unscented Kalman Filter) algorithm, and the results are shown in Figure 4.

As can be seen from Figure 4, abnormal adjustment has an impact on the AE of each algorithm. The AE fluctuation of UKF is obvious and the fluctuation amplitude is much larger than that of the proposed state estimation. The fluctuation of the proposed algorithm is also very small, and its accuracy and robustness are much better than those of the UKF algorithm.

Embodiment 2 of the present invention provides a dynamic state estimation system based on dual drive of model data, including:

The information acquisition and anomaly identification module is used to obtain the real-time node power measurement information of the power system, and calculate the anomaly score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point;

The data-driven prediction module is used to obtain the power prediction value at the abnormal point by using the historical node power measurement information and the pre-trained data-driven prediction model based on the LSTM neural network;

The model-driven prediction module is used to perform real-time load classification prediction based on the model-driven prediction model to obtain the power prediction value at the abnormal point;

An integrated learning module is used to combine historical node power measurement information to learn the predicted values obtained by the data-driven prediction module and the model-driven prediction module to obtain the final power prediction result at the abnormal point;

The state estimation module is used to use the final power prediction result as the pseudo-measurement at the current moment to replace the abnormal information at the current moment, and use the power measurement information of all nodes at the current moment after replacing the abnormal information for state estimation, so as to realize dynamic state estimation based on dual drive of model data.

A terminal includes a processor and a storage medium; the storage medium is used to store instructions;

The processor is used to operate according to the instructions to execute the steps of the method described in Example 1.

A computer-readable storage medium stores a computer program, which implements the steps of the method described in Example 1 when executed by a processor.

The beneficial effects of the present invention are as follows:

1) The present invention takes into account the problem that distribution network measurement data is prone to abnormalities, and combines historical node power measurement information to calculate the abnormal score of real-time node power measurement information to detect missing data and erroneous data, thereby improving the reliability of input data;

2) The present invention integrates data-driven prediction based on LSTM neural network, model-driven prediction and integrated learning to achieve model data hybrid prediction, predicts node power from two parts: data trend and actual physical meaning, and finally obtains better pseudo-quantity measurement through integrated learning, which reduces the impact of bad data on state estimation results and improves the accuracy of state estimation quantity measurement;

3) Taking into account that the traditional Kalman filtering method is limited by the noise model and is difficult to process the distribution network measurement data under non-Gaussian noise interference, the present invention introduces a particle filter to process the noise, and performs state estimation based on the particle filter, so that the noise distribution is not limited by the model and can better reflect the actual power system state.

The present disclosure may be a system, a method and/or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the above. The computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing/processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and/or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and/or edge servers. The network adapter card or network interface in each computing/processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing/processing device.

The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of the computer-readable program instructions to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit may execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims (13)

1. A dynamic state estimation method based on dual drive of model data, characterized by: The method comprises the following steps: S1: Obtain the real-time node power measurement information of the power system, and calculate the abnormal score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point; S2: Use historical node power measurement information and a pre-trained data-driven prediction model based on an LSTM neural network to obtain the power prediction value at the abnormal point; S3: Perform real-time load classification prediction based on the model-driven prediction model to obtain the power prediction value at the abnormal point; In S3, a model-driven prediction model is used to obtain the predicted value of injected active power at the abnormal point according to the load type of the node where the abnormal information is located; The predicted value of injected active power at the abnormal point is: (12) (11) in, is the injected active power of node i where the abnormal information is located at time t; C represents the number of load types of node i where the abnormal information is located; is the total number of nodes served by the upstream substation node r of the node i where the abnormal information is located at time t; represents the active power of the load of type p at node i where the abnormal information is located at time t; represents the telemetry power of the upstream substation node r of the node where the abnormal information is located at time t; represents the load model coefficient of the load type p at time t; represents the average value of the load model coefficient; represents the consumption demand of load of type p at node i; S4: Combine historical node power measurement information to learn the predicted values obtained by S2 and S3 to obtain the final power prediction result at the abnormal point; S5: The final power prediction result is used as the pseudo-measurement at the current moment to replace the abnormal information at the current moment, and the power measurement information of all nodes at the current moment after replacing the abnormal information is used for state estimation to realize dynamic state estimation based on dual drive of model data. 2. A dynamic state estimation method based on dual drive of model data according to claim 1, characterized in that: In S1, the node power measurement information includes node injection power and line power, wherein the node injection power includes the injected active power and reactive power of the node, and the line power includes the active power and reactive power of the line. 3. The dynamic state estimation method based on dual drive of model data according to claim 1, characterized in that: In S1, the process of identifying abnormal information includes: 1) Divide the historical node power measurement information into four measurement sets: injected active power and reactive power, line active power and reactive power. For each set, randomly select M information samples and use them as the root node of the tree. Randomly generate cutting points. Based on the information size of the cutting point, divide the sample into two parts and place them on both sides of the cutting point. Repeat the cutting in the child nodes until only one information data is left or the child node reaches the specified height. Cycle randomly generate cutting points and repeat the cutting operation to generate K trees. 2) Calculate the average height of the K trees generated by each set, and then calculate the anomaly score of the sample information in the set. If the anomaly score exceeds the set threshold, it means that the sample information is abnormal. 4. The dynamic state estimation method based on dual drive of model data according to claim 3 is characterized by: For the measurement set g, its sample information anomaly score for: (3) (2) Where h(g) is the average of the path length from the root node to the node where each sample point in the tree corresponding to the measurement set g passes with respect to the number of samples m; E(h(g)) is the average value of h(g) of K trees; is the average height of m samples; is the intermediate calculation parameter, is Euler's constant. 5. The dynamic state estimation method based on dual drive of model data according to claim 1, characterized in that: In S2, when pre-training the data-driven prediction model based on the LSTM neural network, the formula followed by the information transmission within the neurons of the LSTM neural network is as follows: (5) (6) (7) (8) (9) (10) in, Represent the forget gate, input gate and output gate respectively; , , Represent the weights of the forget gate, input gate, and output gate respectively; , , , The gated unit biases representing the forget gate, input gate, output gate, and neural unit states, respectively; Represents the input vector of the LSTM unit; , Represent the states of neurons at the previous moment and the current moment respectively; Represents the state of the input node at the current time t; , Represent the hidden layer state variables at the previous moment and the current moment respectively; and tanh represent Sigmoid function and Tanh function respectively. 6. The dynamic state estimation method based on dual drive of model data according to claim 1, characterized in that: In S3, based on the power factor, the predicted value of reactive power injected into the node where the abnormal information is located is calculated, and the predicted value of the power information of the branch where the abnormal point is located is obtained through power flow calculation. 7. A dynamic state estimation method based on dual drive of model data according to claim 6, characterized in that: The predicted value calculation formula of the reactive power injected into the node where the abnormal information is located is: in, is the injected reactive power of node i where the abnormal information is located at time t; represents the active power of the load of type p at node i where the abnormal information is located at time t; C represents the number of load types of node i where the abnormal information is located; is the power factor of load type p at node i. 8. The method for dynamic state estimation based on dual drive of model and data according to claim 1, characterized in that: The specific process of S4 is as follows: The power prediction values at the abnormal points obtained by S2 and S3 are used to replace the abnormal information in the real-time node power measurement information of S1, and combined with the historical node power measurement information in S2 to form two large training sets. The following prediction operations are performed on each large training set: the training set is divided into several small groups for training multiple base learners, where the number of small groups is the same as the number of base learners; each small group of data is used to train the corresponding base learner, and each base learner outputs its own prediction results; The outputs of all base learners are used as the input of the meta-learner for training, and the output of the meta-learner is used as the power prediction value at the anomaly point. 9. The method for dynamic state estimation based on dual drive of model and data according to claim 1, characterized in that: In S5, the power measurement information of all nodes at the current moment after replacing the abnormal information is used. The model for state estimation is: (13) (14) in, represents the estimated value of the state at time t; represents the state transition function at time t; , represent process noise and measurement noise respectively; It is a nonlinear mapping function from the particle filter-based quantity measurement to the state quantity. 10. The method for dynamic state estimation based on dual drive of model and data according to claim 9, characterized in that: The nonlinear mapping logic from quantity measurement to state quantity is: The state estimation result is represented by the state posterior probability, and the posterior probability is obtained by discrete sample points obtained by random sampling, and then the state estimation result represented by the particle weight polynomial is obtained; the particle weight is solved and the items with particle weight less than the set value are removed, and the remaining items are summed to obtain the final state estimation result. 11. A dynamic state estimation system based on dual drive of model data, used to implement the method according to any one of claims 1 to 10, characterized in that: the dynamic state estimation system comprises: The information acquisition and anomaly identification module is used to obtain the real-time node power measurement information of the power system, and calculate the anomaly score of the real-time node power measurement information in combination with the historical node power measurement information, identify the abnormal information at the current moment, and the corresponding node is the abnormal point; The data-driven prediction module is used to obtain the power prediction value at the abnormal point by using the historical node power measurement information and the pre-trained data-driven prediction model based on the LSTM neural network; The model-driven prediction module is used to perform real-time load classification prediction based on the model-driven prediction model to obtain the power prediction value at the abnormal point; An integrated learning module is used to combine historical node power measurement information to learn the predicted values obtained by the data-driven prediction module and the model-driven prediction module to obtain the final power prediction result at the abnormal point; The state estimation module is used to use the final power prediction result as the pseudo-measurement at the current moment to replace the abnormal information at the current moment, and use the power measurement information of all nodes at the current moment after replacing the abnormal information for state estimation, so as to realize dynamic state estimation based on dual drive of model data. 12. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-10. 13. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.