A New Energy Grid State Estimation Method Based on Deep Kalman Filter

Abstract

The present invention discloses a new energy power grid state estimation method based on deep Kalman filtering, including: obtaining a prior estimate of the system operating state at the current moment by using a long short-term memory neural network and the operating state of the power system at the previous moment; meanwhile, obtaining a system measurement estimate at the current moment according to the system measurement equation based on the AC power flow model; and obtaining a posterior estimate of the system operating state at the current moment according to the Kalman gain coefficient learned by a recurrent neural network based on deep learning, in combination with the prior estimate of the system operating state at the current moment, the system measurement estimate at the current moment, and the actual measurement value.

Description

Technical Field

[0001] The present invention relates to a new energy power grid state estimation method based on deep Kalman filtering, and belongs to the field of power system state estimation. Background Art

[0002] Power system state estimation is used to estimate and monitor the real-time operating state of a power system, and is a core component in an Energy Management System (EMS). It receives incomplete and noisy measurement data (raw data) describing the system operating state, including node voltage magnitudes, node injection powers, etc., and provides the accurate operating state of the system, i.e., node voltage magnitudes and node voltage phase angles, based on a given estimation criterion. It is the basis for the EMS to implement functions such as economic dispatch, optimal power flow, and voltage control of the power system.

[0003] Widely used power system state estimation methods include Weight Least Squares (WLS), Weight Least Absolute Values (WLAV), etc. WLS can obtain an unbiased optimal estimate on the premise that the system measurement noise follows a Gaussian distribution. However, when there are bad data in the system measurements, the estimation accuracy of WLS will be affected. WLAV can effectively identify bad data in the system measurements and ensure the accurate estimation of the system operating state in the presence of bad data. Both belong to state estimation methods for the operating state of a single time section system and do not consider the correlation between system states at adjacent times.

[0004] With the wide access of distributed new energy units on the power generation side and the increase in the proportion of flexible loads on the load side, the operating characteristics of the power system have become increasingly complex. Due to the strong volatility and randomness of the output of new energy units and the randomness of the charging demand of electric vehicle loads on the load side, the system operating state changes more frequently. To achieve accurate tracking of the system real-time operating state, state estimation methods based on Kalman filtering have been proposed, which consider the evolution relationship of the system operating state over time. For example, power system state estimation methods based on Extended Kalman Filter (EKF) and state estimation methods based on Unscented Kalman Filter (UKF). With the development of deep learning technology, power system real-time state estimation methods based on deep neural networks have been studied and applied. However, there are still the following deficiencies in power system state estimation at the present stage:

[0005] (1) The model-based state estimation method relies on the accurate modeling of the system state space model, including the system state transition matrix, measurement noise distribution, etc. However, it is difficult to accurately characterize and provide these factors for the new power system;

[0006] (2) The data-based state estimation method lacks the construction of system prior knowledge and often has a large number of parameters to be trained, which limits its interpretability and scalability. Summary of the Invention

[0007] In view of this, the present invention proposes a new energy grid state estimation method based on deep Kalman filtering to overcome the defects and deficiencies of the above-mentioned existing technologies.

[0008] A new energy grid state estimation method based on deep Kalman filtering includes: using a long short-term memory neural network and the operating state of the power system at the previous moment to obtain a prior estimate value of the system operating state at the current moment; at the same time, according to the system measurement equation based on the AC power flow model, obtain the system measurement estimate value at the current moment; according to the Kalman gain coefficient learned by the recurrent neural network based on deep learning, combine the prior estimate value of the system operating state at the current moment, the system measurement estimate value at the current moment and the actual measurement value to obtain the posterior estimate value of the system operating state at the current moment.

[0009] Further, the steps of calculating the prior estimate value of the system operating state at the current moment and the system measurement estimate value at the current moment include:

[0010] Construct a state transition equation using the system state transition matrix at the previous moment, the posterior estimate value of the system operating state at the previous moment, and the input gain term at the previous moment, and construct a system measurement estimate equation using the measurement equation and measurement noise to describe the state space model of the power system;

[0011] Use a deep neural network based on a long short-term memory neural network to learn the state transition function; at the same time, considering the strong randomness and volatility of the distributed new energy output on the system operating state, take the change in the active power injection of the node connecting the distributed new energy at the current moment as one of the inputs of the long short-term memory neural network, obtain a new state transition equation, and further obtain the prior estimate value of the system operating state at the current moment and the system measurement estimate value at the current moment.

[0012] Further, the state space model of the power system is described as:

[0013] ;

[0014] Wherein, represents tThe prior estimated value of the system operating state at the current moment; Represents the system state transition matrix, which is a diagonal matrix; Represents t The posterior estimated value of the system operating state at the previous moment (-1 moment); Represents the input gain term at the previous moment, which is a parameter vector; Represents the system measurement estimated value at the current moment, including the node voltage magnitude , the active power injected into the node , the reactive power injected into the node , the active power of the branch and the reactive power of the branch ; Represents the measurement noise; Represents the measurement equation.

[0015] Furthermore, the measurement equation is:

[0016] ;

[0017] Among them, Represents the set of nodes directly connected to node i , including node i itself; Represents t The prior estimated value of the phase angle difference of branch at moment, g ij , b ij respectively represent the conductance and susceptance of branch , g si , b si respectively represent the ground conductance and susceptance of node i , , .

[0018] Furthermore, the diagonal elements in the system state transition matrix and the elements in the parameter vector are obtained by the Holt double-parameter linear exponential smoothing method, and the solution process is:

[0019] ;

[0020] ;

[0021] ; ;​​ ;

[0022] ;

[0023] Among them, and represent horizontal term smoothing and trend term smoothing respectively, and represent horizontal smoothing factor and trend smoothing factor respectively.

[0024] Furthermore, the new state transition equation is:

[0025] , ;

[0026] Among them, f LSTM (·) represents the state transition function, represents the node i at t the change in the active power injection at time, represents the set of nodes connected to the new energy units;

[0027] Furthermore, the prior estimate value t of the system operating state at time and t the measurement estimate value of the system at time are obtained:

[0028] .

[0029] Furthermore, in the Kalman filter framework, the calculation process of the Kalman gain coefficient is:

[0030] ; ;

[0031] ;

[0032] Among them, and represent the covariance matrices of the process noise and the measurement noise respectively; and represent the state error covariance matrix and the measurement residual error covariance matrix respectively; represents the Jacobian matrix.

[0033] Furthermore, the input data of the recurrent neural network for learning the Kalman gain coefficient includes the following four categories:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] Among them, represents t the actual measurement value of the time system.

[0039] Furthermore, the Kalman gain coefficient obtained through deep learning is denoted as , then t the posterior estimate of the system operating state at time .

[0040] The present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the foregoing method can be implemented.

[0041] The beneficial effects of the technical solution of the present invention are reflected in:

[0042] 1) The foregoing technical solution of the present invention proposes a digital-analog hybrid-driven power system state estimation method that combines deep learning and Kalman filtering. Compared with traditional model-based state estimation methods, this method does not need to assume a specific distribution form of measurement noise, and directly learns the Kalman filter gain through a data-driven module, thus avoiding the calculation process of the Jacobian matrix and the Hessian matrix and reducing the dependence on the model structure and parameter settings;

[0043] 2) The present invention constructs the state transition relationship of the system through a deep neural network, explicitly introduces the influence of the volatility of distributed new energy output on the system state evolution, so that the constructed state estimation model has stronger adaptability and prediction ability in the scenario of massive distributed new energy access, and improves the estimation accuracy and engineering practicability;

[0044] 3) Compared with pure learning-based state estimation methods, the present invention integrates the update framework of Kalman filtering in the overall structure, making the estimated state results have stronger physical interpretability and facilitating meeting the actual application requirements in power system operation and dispatching. Description of the Drawings

[0045] Figure 1 is a schematic diagram of the execution process of the new energy grid state estimation method based on deep Kalman filtering in an embodiment of the present invention.

[0046] Figure 2When the embodiment of the present invention takes a 15-node power system as an example, it is the predicted value and the true value of the voltage amplitude of node 2.

[0047] Figure 3 When the embodiment of the present invention takes a 15-node power system as an example, it is the predicted value and the true value of the voltage phase angle of node 2.

[0048] Figure 4 It is the overall algorithm flow of the method of the embodiment of the present invention. Detailed implementation manners

[0049] The present invention will be further described below in conjunction with the accompanying drawings, specific implementation manners and embodiments. The purpose of providing the embodiments is only for illustration and not for any limitation.

[0050] The embodiment of the present invention proposes a digital-analog hybrid-driven power system state estimation method that combines deep learning and Kalman filtering. Under the framework of the classical Kalman filtering, by combining deep learning technology, the power system state estimation under the background of distributed new energy penetration is realized. The logic of this method is as follows: First, in the prediction stage, the prior estimate value of the system operating state is obtained through a long short-term memory neural network based on deep learning. At the same time, the system measurement estimate value is obtained from the model-based AC power flow equation combined with the prior estimate value of the system operating state obtained above; in the filtering stage, the Kalman gain coefficient is learned through a recurrent neural network based on deep learning, and then, combined with the prior estimate value of the system operating state obtained above, the posterior estimate value of the system operating state is obtained.

[0051] Figure 1 It is a schematic diagram of the execution process of the new energy power grid state estimation method based on deep Kalman filtering in the embodiment of the present invention, where represents the posterior estimate value of the system operating state at the initial moment, represents moving forward to the previous moment, and the remaining variables will be explained in the subsequent formula descriptions.

[0052] Please refer to Figure 1 The new energy power grid state estimation method proposed in the embodiment of the present invention is roughly divided into the following two stages:

[0053] Stage 1: Calculate t the prior estimate value of the system operating state and the system measurement estimate value at time t (the current time).

[0054] Using a long short-term memory neural network (LSTM) and the operating state of the power system at the previous moment ( t t - 1 moment), the current moment ( tThe prior estimated value of the system operating state at the current moment; meanwhile, according to the system measurement equation based on the AC power flow model, the system measurement estimated value at the current moment is obtained. The specific process is as follows:

[0055] First, construct the state transition equation using the system state transition matrix at the previous moment, the posterior estimated value of the system operating state at the previous moment, and the input gain term at the previous moment, and construct the system measurement estimation equation using the measurement equation and the measurement noise. Use these two equations to construct the state space model describing the power system as follows:

[0056] ;

[0057] Among them, represents t the prior estimated value of the system operating state at the current moment; represents the system state transition matrix, which is a diagonal matrix; represents t the posterior estimated value of the system operating state at the -1 moment; represents the input gain term at the previous moment, which is a parameter vector; represents the system measurement estimated value at the current moment, including the node voltage magnitude , the active power injected into the node , the reactive power injected into the node , the active power of the branch and the reactive power of the branch ; represents the measurement noise; represents the measurement equation, expressed as:

[0058] ;

[0059] Among them, represents the set of nodes directly connected to node i , including node i itself; represents t the prior estimated value of the phase angle difference of branch at the current moment, g ij , b ij respectively represent the conductance and susceptance of branch , g si , b si respectively represent the ground conductance and susceptance of node i , , .

[0060] System state transition matrix The diagonal elements in and the parameter vector The elements in are obtained by the Holt two-parameter linear exponential smoothing method. The solution process is as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] Among them, and represent the horizontal term smoothing and the trend term smoothing respectively, and represent the horizontal smoothing factor and the trend smoothing factor respectively.

[0068] The above state transition equation obtained by the Holt two-parameter linear exponential smoothing method is applicable to the scenario where the system load fluctuates little. And the present invention is more applicable to the scenario where the new energy output on the power generation side of the system fluctuates greatly.

[0069] In the context of distributed new energy penetration, in order to accurately predict the evolution of the system operating state over time, a deep neural network based on a long short-term memory neural network is used to learn the state transition function f LSTM . At the same time, considering the influence of the strong randomness and volatility of the distributed new energy output on the system operating state, the change in the active power injection d t of the node connecting the distributed new energy at the current moment is used as one of the inputs of the LSTM, and a new state transition equation is obtained as follows:

[0070] , ;

[0071] Among them, represents the change in the active power injection i of node t at time , and

[0072] represents the set of nodes connected with new energy units. t The prior estimated value of the system operating state at time and t the measured estimate of the time system :

[0073] 。

[0074] Phase 2, calculation t the posterior estimate of the operating state of the time system.

[0075] According to the Kalman gain coefficient learned by the recurrent neural network based on deep learning, combined with the calculated t the prior estimate of the operating state of the time system and t the measured estimate of the time system , calculate t the posterior estimate of the operating state of the time system. The specific process is as follows:

[0076] In the Kalman filter framework, the calculation process of the Kalman gain coefficient is as follows:

[0077] ;

[0078] ;

[0079] ;

[0080] where and represent the covariance matrices of the process noise and the measurement noise respectively; and represent the state error covariance matrix and the measurement residual error covariance matrix respectively; represents the Jacobian matrix.

[0081] Therefore, when designing the recurrent neural network for learning the Kalman gain coefficient , its input data includes the following four categories:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] where represents t the actual measurement value of the time system.

[0087] Denote the Kalman gain coefficient obtained through deep learning as . Therefore, during the filtering stage, the t posteriori estimate value of the system operating state at time can be calculated.

[0088] The overall algorithm flow of the method in the embodiment of the present invention is as Figure 4 shown, and mainly includes:

[0089] 1. Initialization: The trained long short-term memory neural network (LSTM) and recurrent neural network (RNN) structures;

[0090] 2. Input: The measurement equation of the power system , the initial state variable , and the measurement value (the node voltage amplitude U, the node injection power and the branch power );

[0091] 3. Prediction step of the Kalman filtering framework:

[0092] ;

[0093] 4. Update step of the Kalman filtering framework:

[0094] 1) Apply the trained recurrent neural network RNN to calculate the Kalman gain coefficient ;

[0095] 2) ;

[0096] 5. Output: .

[0097] The present invention is proven to be applicable to the real-time operation state tracking of the power system, can achieve accurate estimation of the power system operation state, and has the characteristics of wide compatibility and strong adaptability. Taking a 15-node power system as an example, the present invention is used to track the real-time operation state of the power system. Using the mean squared error (Mean Squared Error) and mean absolute error (Mean Absolute Error, MAE) of the node voltage amplitude and node voltage phase angle as evaluation indexes, which are defined as:

[0098] 1) Node voltage amplitude:

[0099] ;

[0100] ;

[0101] 2) Node voltage phase angle:

[0102] ;

[0103] ;

[0104] Wherein, represents the number of system nodes, represents the total number of tracking instants. and represent the estimated values of the node voltage amplitude and the node voltage phase angle obtained by the present invention for node at instant , and represent the true values of the node voltage amplitude and the node voltage phase angle for node at instant . The calculation results show that the MSE and MAE of the node voltage amplitude are and respectively, and the MSE and MAE of the node voltage phase angle are and respectively.

[0105] Meanwhile, taking node 2 as an example, the comparison between the estimated values and the true values of the voltage amplitude and phase angle of node 2 is given, see Figure 2 and Figure 3 . From the results in Figure 2 and Figure 3 , it can be seen that the estimated values obtained by the present invention are very close to the true values. The above results prove that the present invention can achieve accurate tracking of the real-time operating state of the power system.

[0106] Another embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it can implement the steps of the method in the foregoing embodiment. Based on such an understanding, the technical solution of the present invention can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including several instructions for causing a computer device (such as a PC, a server, or a network device, etc.) to execute the respective steps of the method of the present invention to implement the functions of the present invention.

[0107] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several equivalent substitutions or obvious variations can be made, and if the performance or use is the same, they should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A new energy power grid state estimation method based on deep Kalman filtering, characterized in that, Including: Using a long short-term memory neural network and the operating state of the power system at the previous moment to obtain a prior estimated value of the system operating state at the current moment; Meanwhile, according to the system measurement equation based on the AC power flow model, obtain the system measurement estimated value at the current moment; According to the Kalman gain coefficient learned by the recurrent neural network based on deep learning, combine the prior estimated value of the system operating state at the current moment, the system measurement estimated value at the current moment, and the actual measurement value to obtain the posterior estimated value of the system operating state at the current moment; The steps of calculating the prior estimated value of the system operating state at the current moment and the system measurement estimated value at the current moment include: Construct a state transition equation using the system state transition matrix at the previous moment, the posterior estimated value of the system operating state at the previous moment, and the input gain term at the previous moment, and construct a system measurement estimated equation using the measurement equation and the measurement noise to describe the state space model of the power system; Adopt a deep neural network based on a long short-term memory neural network to learn the state transition function; at the same time, considering the strong randomness and volatility of the distributed new energy output on the system operating state, take the change in the active power injected into the node connecting the distributed new energy at the current moment as one of the inputs of the long short-term memory neural network to obtain a new state transition equation, and then obtain the prior estimated value of the system operating state at the current moment and the system measurement estimated value at the current moment.

2. The new energy power grid state estimation method according to claim 1, characterized in that, The state space model of the power system is described as: ； Among them, represents the prior estimate of the system operating state at the current moment; represents the system state transition matrix, which is a diagonal matrix; represents the posterior estimate of the system operating state at the previous moment; represents the input gain term at the previous moment, which is a parameter vector; represents the system measurement estimate at the current moment, including the node voltage magnitude , the active power injected into the node , the reactive power injected into the node , the active power of the branch and the reactive power of the branch ; represents the measurement noise; represents the measurement equation.

3. The new energy grid state estimation method according to claim 2, characterized in that The measurement equation is as follows: ； Among them, represents the set of nodes directly connected to node i , including node itself; represents the prior estimated value of the phase angle difference of the branch at time , respectively represent the conductance and susceptance of branch , respectively represent the ground conductance and susceptance of node , , .

4. The new energy grid state estimation method according to claim 2, characterized in that, System state transition matrix The diagonal elements in and the parameter vector The elements in Are obtained by the Holt double-parameter linear exponential smoothing method, and the solution process is as follows: ； ； ； ； ； ； Among them, and represent horizontal term smoothing and trend term smoothing respectively, and represent horizontal smoothing factor and trend smoothing factor respectively.

5. The new energy grid state estimation method according to claim 2, wherein, The new state transition equation is: ； Among them, f LSTM (·) represents the state transition function, represents the node at the change in the active power injection at time t, represents the set of nodes connected with new energy units; Furthermore, the prior estimate of the system operating state at a certain moment and the measurement estimate of the system at a certain moment : 。 6. The new energy power grid state estimation method according to any one of claims 2-5, characterized in that In the Kalman filter framework, the Kalman gain coefficient is calculated as follows: ； ； ； wherein, and represent the covariance matrices of process noise and measurement noise respectively; and represent the state error covariance matrix and the measurement residual error covariance matrix respectively; represents the Jacobian matrix.

7. The new energy power grid state estimation method according to claim 6, wherein According to the calculation process of the Kalman gain coefficient in the Kalman filtering framework, the input data of the recurrent neural network used to learn the Kalman gain coefficient includes the following four categories: ； ； ； ； Among them, represents the actual measured value of the time system.

8. The new energy grid state estimation method according to claim 7, characterized in that, Denote the Kalman gain coefficient obtained through deep learning as , so the posterior estimate of the system operating state at time .

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it can implement the steps of the method according to any one of claims 1-8.

Images