A Method for Constructing an Agent Model of Reactive Power Support Equipment in the Near Region of DC Landing Points Based on Hybrid Machine Learning

Abstract

The present invention discloses a method for constructing an agent model of a reactive power support device in the near area of a DC landing point based on hybrid machine learning. First, a dynamic event sample set is generated and collected, and the curves before disturbance, during disturbance, and after disturbance recovery are saved. Secondly, in combination with the reactive power support device in the near area of the DC landing point, a single-machine infinite-bus test system is constructed to extract parameters. Thirdly, the input feature vector and output feature vector of the machine learning model are constructed to form a sample library. Then, different types of supervised learning model frameworks are built, and the optimal evaluation model is obtained. Finally, according to the above evaluation results, the top K models are selected, and the active power and reactive power curves of the reactive power support device in the near area of the DC landing point are output in a weighted average manner as the final agent model of the reactive power support device in the near area of the DC landing point. The present invention combines the dynamic external characteristics of the reactive power support device in the near area of the DC landing point to establish the controller agent model, which can improve the accuracy of power system simulation.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system intelligence, and particularly relates to a method for constructing an agent model of a reactive power support device in the near area of a DC landing point based on hybrid machine learning. Background Art

[0002] High-quality dynamic models of power equipment are crucial for the planning and decision-making of new power systems and their safe and economic dispatching operations. An overly optimistic model quality will lead to potential safety hazards in the power system; while an overly conservative model will result in over-investment and unnecessary resource waste. During the power system planning and dispatching operation process, it is usually necessary to perform refined modeling for various types of complex power equipment, and conduct large-scale simulation analysis on the premise of ensuring model quality, so as to make optimal planning and dispatching decisions based on the simulation deduction results. However, power equipment manufacturers usually can only provide encapsulated models and cannot provide detailed dynamic model details of the equipment, with poor transparency and flexibility, making it difficult to meet the analysis requirements during the construction of new power systems.

[0003] In the prior art, flexible AC transmission systems (FACTS) provide some technical means to improve the safety and stability of the power grid. Static var generators (SVG) are one of the core devices of FACTS. As an advanced reactive power compensation device, they are widely used in various fields such as transmission networks, distribution networks, and new energy power generation. This device can compensate for fluctuating loads, harmonics, and power factors to reduce losses and improve power quality; the accuracy of its model is related to the credibility of power system simulation, and thus affects the safe and stable operation control decisions of the power grid. However, in practice, manufacturers usually have difficulty providing refined models. In power system transient simulations, typical parameter values are mostly used, but the accuracy is difficult to meet the requirements of power system transient simulations. Therefore, it is urgent to establish an agent model for this controller by combining the dynamic external characteristics of SVG devices to improve the accuracy of power system simulation. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems raised in the above background art, and provide a method for constructing an agent model of a reactive power support device in the near area of a DC landing point based on hybrid machine learning.

[0005] To achieve the purpose of the present invention, the present invention discloses a method for constructing an agent model of a reactive power support device in the near area of a DC landing point based on hybrid machine learning, including the following steps:

[0006] Step 1: For the reactive power support device in the near area of the DC landing point (taking the static var generator, SVG as an example), use a phasor measurement unit PMU, a fault recorder, or use an encapsulated model to simulate and deduce to generate and collect a dynamic event sample set, and save the curves before disturbance, during disturbance, and after disturbance recovery;

[0007] Step 2: Combine the SVG device to build a single-machine infinite test system to extract the dynamic model parameters of the SVG device, the grid-connected bus voltage amplitude, the grid-connected bus voltage phase angle curve, the SVG device active power, and the SVG device reactive power;

[0008] Step 3: Construct the input feature vector and output feature vector of the machine learning model to form a sample library, and randomly select 80% as the training set and the remaining 20% ​​as the test set;

[0009] Step 4: Build different types of supervised learning model frameworks, select hyperparameter sets for each model, and use training sets and test sets to optimize the hyperparameter combination to obtain the optimal evaluation model; the model quality is aimed at minimizing RMSE;

[0010] Step 5: According to the above evaluation results, select the top K models, and output the SVG active power and reactive power curves in a weighted average manner as the final SVG proxy model.

[0011] Furthermore, in step 2, the single-machine infinite test system includes an infinite ideal power source, a transformer, a transmission line and an equivalent load. The ideal power source is connected to the equivalent load of the same node and the SVG through the transmission line after passing through the transformer.

[0012] Furthermore, in step 3, the input feature vector includes SVG dynamic model parameters, bus voltage amplitude curve and bus voltage phase angle curve.

[0013] Further, in step 3, the output feature vector includes an SVG active power curve and an SVG reactive power curve.

[0014] Furthermore, in step 4, the supervised learning model framework includes a deep neural network DNN, a long short-term memory network LSTM, and a gradient boosting strategy GBM.

[0015] Furthermore, in step 4, the model quality aims to minimize RMSE, and the specific formula is as follows:

[0016]

[0017]

[0018] In the formula, y i Represents the true value, represents the predicted value, N samplerepresents the number of samples; when a disturbance occurs on the system side causing voltage or frequency fluctuations, the dynamic process will stimulate the control logic action of the reactive support equipment model in the vicinity of the DC landing point to achieve the expected control target; the active power and reactive power curves calculated in the dynamic process will be used as model output to compare the model accuracy; the model deviation measurement indicators include root mean square error RMSE, mean absolute percentage error MAPE, and mean absolute error MAE.

[0019] Compared with the prior art, the significant progress of the present invention lies in: by combining multiple machine learning methods, a refined model is constructed through the external characteristics of SVG equipment. Compared with the models provided by manufacturers and the typical value models of simulation software that are difficult to meet the accuracy of transient simulation of power systems, the model provided by the present invention improves the accuracy of power system simulation and enhances the credibility of control decisions for safe and stable operation of the power grid.

[0020] In order to more clearly illustrate the functional characteristics and structural parameters of the present invention, further description is given below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0022] Figure 1 It is a schematic diagram of a single-machine infinite simulation system;

[0023] Figure 2 This is a schematic diagram of the SVG controller model;

[0024] Figure 3 It is a comparison diagram of reactive power response curves under different SVG parameters;

[0025] Figure 4 It is a flow chart of a method for constructing a static reactive generator proxy model based on hybrid machine learning;

[0026] Figure 5 It is a schematic diagram of reactive power prediction process based on SVG device agent model. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] In the prior art, to study the dynamic response characteristics of SVG devices, a single-machine infinite-bus test system is usually established, such as Figure 1 shown. The system includes an infinite ideal power source, a transformer, a transmission line, then connects to an SVG, and an equivalent load (used to simulate system disturbances. For example, short-circuiting the load to the ground can simulate the dynamic response situation after a three-phase short-circuit fault in the system. The role of the SVG is to maintain the constant voltage of the bus at the connection point. In actual engineering, the controller model of the SVG device is usually encapsulated into a "black box model", and its control logic and details are unknown. However, the operating conditions of the SVG device at the moment of grid fault can be finely simulated through a real-time digital simulator (RTDS) or a full electromagnetic transient high-precision simulation program (such as PSCAD).

[0029] Taking the SVG controller model in the BPA program for power system simulation as an example, its control block diagram is as Figure 2 shown. In the figure: V is the voltage at the output end of the SVG, V REF is the reference voltage, V SCS is the auxiliary signal, V T is the voltage on the system side, I S is the output current of the SVG, T1 is the time constant of the filter and the measurement circuit, T2 and T3 are the first-stage lead time constant and the second-stage lag time constant respectively, T4 and T5 are the second-stage lead time constant and the third-stage lag time constant respectively, T P is the time constant of the proportional link, T S is the response delay of the SVG, K P is the amplification factor of the proportional link, K i is the amplification factor of the integral link, K d is the slope of the V-I characteristic curve of the SVG, X T is the equivalent reactance between the SVG and the system. There are also six limiting links in the figure: V MAX is the upper limit of the voltage limiting link, V MIN is the lower limit of the voltage limiting link, I CMAX is the maximum capacitive current, I LMAX is the maximum inductive current, and the limits V SMAX and V SMIN of the output of the proportional and integral links are calculated as follows:

[0030] V SMAX =V T +X T *I CMAX (1)

[0031] V SMIN =V T -X T *I LMAX

[0032] Power system transient simulation method: To obtain the dynamic trajectories after grid disturbances, such as generator power angles, bus voltages, and bus frequencies, time-domain simulations are usually performed to evaluate the dynamic characteristics of the power grid. Usually, multiple safety criteria are adopted, including system synchronization, transient voltage recovery, frequency distribution, damping, etc. The models used in the simulation usually include generators, exciters, governors, PSSs, dynamic composites, DC converters, protective relays, and custom models. In electromechanical transient simulations, mathematical models are used to describe the dynamic behavior of the above-mentioned power equipment within a short period (e.g., 10 seconds to 60 seconds). The dynamic behavior (DAEs) of a multi-machine power system can be described by a set of non-linear differential and algebraic equations, as shown in Equation (2).

[0033]

[0034] where x is an n x *1-order vector, containing all the state variables of the power system, such as generator rotor speed, rotor angle, magnetic flux, and excitation system, governor, PSS, dynamic load, and DC components, etc. y is an n y *1-order vector composed of algebraic variables, that is, the magnitude and angle of the bus voltage. u is an n u *1-order vector, containing input variables, such as the voltage reference value of the automatic voltage controller (AVR). f is a set of differential equations, and the derivatives of the state variables are usually discretizable. The algebraic equation g consists of Kirchhoff's current law network equations, that is, the sum of all currents flowing into the bus must be equal to zero, providing the coupling relationship between algebraic variables and state variables. Equation (3) is used for the network equation of the power system during electromechanical transient simulation:

[0035] I = Y * V (3)

[0036] where I is the injected current, including dynamic models such as generators, dynamic loads, and user-defined models; Y represents the admittance matrix of the power network; V is the complex voltage vector of the power grid buses. To solve the above equations, the more common methods are explicit and implicit method integration, as shown in Table 1.

[0037] Table 1 Numerical integration methods

[0038]

[0039] Explicit integration methods include the Euler method, the improved Euler method, the fourth-order Runge-Kutta method, and the hybrid Adams-Bashforth method. These methods are easy to implement and can support very complex user-defined models. The discrete differential equations and algebraic equations are alternately solved multiple times within a time step to achieve convergence. Explicit integration usually requires a small time step, for example, 1 / 4 or 1 / 2 cycle, to ensure numerical stability. In North America, mainstream commercial software for electromechanical transient simulation programs all use explicit integration methods for power system planning and operation studies. In contrast, implicit integration methods can support a larger time step while maintaining numerical stability, for example, 1 cycle. The trapezoidal method is a commonly used method in commercial software packages used in China and Europe. This method uses the Newton iteration method. During the dynamic simulation process, the Jacobian matrix is solved by the iteration method or the perturbation method, and the DAEs are solved simultaneously at each time step. For large power networks, the initial guess at each time step is the key to ensuring fast convergence; otherwise, the increase in the number of iterations will significantly reduce the speed of the dynamic simulation, especially in the case where there are many switching operations.

[0040] Device model verification method based on injected signals: To effectively simulate the dynamic models of power equipment such as SVG, this invention patent uses the injected signal method to input the bus voltage amplitude and voltage phase angle (or bus frequency) at the device port side as time series signals into the transient simulation model, that is, regularly replace the bus voltage amplitude and voltage phase angle calculated in formula (3) according to the measured signals, and collect the active power (P) and reactive power curve (Q) of the SVG device, as Figure 1 、 Figure 3 shown. When a disturbance occurs on the system side, causing voltage or frequency fluctuations, this dynamic process will trigger the control logic action of the SVG device model to achieve the expected control goal. The active power and reactive power curves calculated during this process will be used as the model output to compare the model accuracy. Commonly used model deviation measurement indicators include: root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE).

[0041]

[0042] In this application, a method for constructing a proxy model of a static var generator based on hybrid machine learning is disclosed. First, a dynamic event sample set is generated and collected, and the curves before disturbance, during disturbance, and after disturbance recovery are saved. Secondly, a single-machine infinite bus test system is constructed in combination with the reactive power support equipment in the near area of the DC landing point, and parameters are extracted. Thirdly, the input feature vector and output feature vector of the machine learning model are constructed to form a sample library. Then, different types of supervised learning model frameworks are built, and a set of hyperparameters is selected for each model. The hyperparameter combination is optimized using the training set and the test set to obtain the optimal evaluation model. Finally, according to the above evaluation results, the top K models are selected, and the active power and reactive power curves of the reactive power support equipment in the near area of the DC landing point are output in a weighted average manner as the final proxy model of the reactive power support equipment in the near area of the DC landing point. The specific steps are as Figure 4 shown:

[0043] Step 1: For the SVG equipment, use a phasor measurement unit (PMU), a fault recorder, or use an encapsulated model to simulate and deduce to generate and collect a dynamic event sample set, and save the curves before disturbance, during disturbance, and after disturbance recovery.

[0044] Step 2: Combine the SVG equipment to construct the Figure 1 equivalent circuit shown, and extract the dynamic model parameters of the SVG equipment, the grid-connected bus voltage amplitude, the grid-connected bus voltage phase angle curve, the active power of the SVG equipment, and the reactive power of the SVG equipment.

[0045] Step 3: Construct the input feature vector and output vector of the machine learning model as shown in Table 2 and Table 3. Form a sample library, and randomly select 80% as the training set and the remaining 20% as the test set.

[0046] Table 2 Input Feature Vector

[0047] SVG dynamic model parameters Bus voltage amplitude curve Bus voltage phase angle curve [Par1, Par2, …, ParK] [Vm] [θm]

[0048] Table 3 Output Feature Vector

[0049] SVG active power curve SVG reactive power curve <![CDATA[[P SVG > <![CDATA[[Q SVG >

[0050] Step 4: Build different types of supervised learning model frameworks, including deep neural network (DNN), long short-term memory network (LSTM), gradient boosting strategy (GBM), etc. For each model, select a set of hyperparameters respectively, and use the training set and the test set to optimize the hyperparameter combination to obtain the optimal evaluation model. The model quality aims to minimize RMSE (formula (4)).

[0051] Step 5: According to the above evaluation results, select the top K models, and output the active power and reactive power curves of the SVG in a weighted average manner as the final SVG proxy model.

[0052] Figure 5 Shows the specific processes of information input, model construction, and active and reactive power prediction. First, input the pre-disturbance system information, bus voltage magnitude curve, bus voltage phase angle curve, and SVG model parameters. The input information passes through multiple models such as deep neural network (DNN), long short-term memory network (LSTM), and gradient boosting strategy (GBM). The specific process of the model algorithm is as follows:

[0053] 1) DNN (Deep Neural Network)

[0054] A basic feedforward neural network that includes the processes of forward propagation and backward propagation.

[0055] Forward propagation:

[0056] For a feedforward neural network with L layers, at the l-th layer, l = 1, 2,..., L, the weighted sum of the input can be defined as:

[0057] z (l) =W (l) a (l) +b (l) (5)

[0058] where z (l) is the weighted sum of the l-th layer, W (l) is the weight matrix of the l-th layer, a (l) is the activation value of the (l - 1)-th layer, and b (l) is the bias of the l-th layer. The activation function applied is:

[0059] a (l) =g(z (l) ) (6)

[0060] where g is the activation function, such as sigmoid, ReLU (Rectified Linear Unit), etc. Repeat the above steps until the output layer L.

[0061] Backward propagation:

[0062] During the backward propagation process, the weights and biases are updated by calculating the gradients of the loss function with respect to the network parameters. The error formula for calculating the output layer is:

[0063]

[0064] where J is the loss function, ⊙ represents element-wise multiplication, and g' is the derivative of the activation function of the output layer. The error propagates to the previous layer, and there is:

[0065] δ (l) =((W (l+1) ) T δ(l+1) ⊙g'(z (l) ) (8)

[0066] Then, calculate the gradients of the weights and biases:

[0067]

[0068] Finally, update the weights and biases:

[0069]

[0070] where α is the learning rate, which is used to control the step size of parameter update.

[0071] 2) LSTM (Long Short-Term Memory Network)

[0072] LSTM is a variant of the Recurrent Neural Network (RNN) that can solve the problem of long-term dependencies. Its core is the use of a special memory cell. In LSTM, the memory cell at each time step t contains a cell state C t and a hidden state h t . The calculation of the LSTM cell can be divided into the following steps:

[0073] Forget gate:

[0074] Determines which information in the previous memory cell needs to be retained or forgotten. The output of the forget gate is calculated as:

[0075] f t = σ(W t · [h t-1 , x t + b f ) (11)

[0076] where σ represents the sigmoid activation function, W t represents the weight matrix of the forget gate, [h t-1 , x t represents the hidden state h t-1 at the previous time step and the input x t at the current time step, and b f is the bias of the forget gate.

[0077] Input gate:

[0078] Determines which information in the current input x t will be added to the cell state. The calculation of the output of the input gate is:

[0079] i t = σ(W i · [h t-1 , x t + bi ) (12)

[0080] Among them, W i and b i are the weight matrix and bias of the input gate respectively, and then calculate the candidate cell state:

[0081]

[0082] Update the cell state through the outputs of the forget gate and the input gate:

[0083]

[0084] Output gate:

[0085] Determines which cell state information will be passed to the hidden state at the next time step. The calculation of the output of the output gate is:

[0086] o t = σ(W o · [h t-1 , x t + b o ) (15)

[0087] Among them, W o and b o are the weight matrix and bias of the output gate respectively, and finally calculate the hidden state:

[0088] h t = o t ⊙ tanh(C t ) (16)

[0089] 3) GBM (Gradient Boosting Strategy)

[0090] The gradient boosting strategy is an ensemble learning method that improves the prediction performance by gradually adding base models (usually decision trees). GBM optimizes the model by minimizing the gradient of the loss function.

[0091] First, initialize a constant model to minimize the initial loss function:

[0092]

[0093] For each round of iterative update with m = 1, 2,..., M, first calculate the pseudo-residual, which is the negative gradient of the loss function with respect to the current model F m-1 :

[0094]

[0095] Then fit the pseudo-residual:

[0096]

[0097] Finally, update model F and control the step size in combination with the learning rate υ:

[0098] F m (x) = F m-1 (x) + υh m (x) (20)

[0099] Integrate multiple models, and use the weighted average of the results of the top K models after evaluation as the active and reactive power results of SVG.

[0100] It should be noted that in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0101] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a proxy model of reactive support equipment in the vicinity of a DC point based on hybrid machine learning, characterized in that: The following steps are involved: Step 1: Use a phasor measurement unit (PMU), a fault recorder or a package model simulation to generate and collect dynamic event sample sets for the SVG device, and save the pre-disturbance, disturbance process and post-disturbance recovery curves; Step 2: Combine the SVG device to build a single-machine infinite test system to extract the dynamic model parameters of the SVG device, the grid-connected bus voltage amplitude, the grid-connected bus voltage phase angle curve, the SVG device active power, and the SVG device reactive power; Step 3: Construct the input feature vector and output feature vector of the machine learning model to form a sample library, and randomly select 80% as the training set and the remaining 20% ​​as the test set; In step 3, the input feature vector includes SVG dynamic model parameters, bus voltage amplitude curve and bus voltage phase angle curve; In step 3, the output feature vector includes an SVG active power curve and an SVG reactive power curve; Step 4: Build different types of supervised learning model frameworks, select hyperparameter sets for each model, and use training sets and test sets to optimize the hyperparameter combination to obtain the optimal evaluation model; the model quality is aimed at minimizing RMSE; Step 5: Evaluate the model quality with the goal of minimizing RMSE, select the top K models, and output the SVG active power and reactive power curves in a weighted average manner as the final SVG proxy model.

2. According to the method for constructing a proxy model of reactive support equipment in the vicinity of a DC point of impact based on hybrid machine learning in claim 1, it is characterized in that: In step 2, the single-machine infinite test system includes an infinite ideal power source, a transformer, a transmission line and an equivalent load. The ideal power source is connected to the equivalent load of the same node and the SVG through the transmission line after passing through the transformer.

3. The method for constructing a proxy model of reactive support equipment in the vicinity of a DC point of impact based on hybrid machine learning according to claim 1, characterized in that: In step 4, the supervised learning model framework includes a deep neural network DNN, a long short-term memory network LSTM, and a gradient boosting strategy GBM.

4. The method for constructing a proxy model of reactive support equipment in the vicinity of a DC point of impact based on hybrid machine learning according to claim 1 is characterized in that: In step 4, the model quality aims to minimize RMSE, and the specific formula is as follows: In the formula, y i Represents the true value, represents the predicted value, N sample represents the number of samples; when a disturbance occurs on the system side causing voltage or frequency fluctuations, the dynamic process will stimulate the control logic action of the reactive support equipment model in the vicinity of the DC landing point to achieve the expected control target; the active power and reactive power curves calculated in the dynamic process will be used as model output to compare the model accuracy; the model deviation measurement indicators include root mean square error RMSE, mean absolute percentage error MAPE, and mean absolute error MAE.

Images