Two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit

Abstract

The present invention discloses a two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated cyclic unit, comprising: in the first stage, the underlying measurement data required for modeling and simulation is obtained, and a transient stability prediction model based on the improved convolutional residual memory network is established for qualitative and quantitative analysis; in the second stage, the underlying measurement data required for modeling and simulation is obtained, and a generator power angle trajectory prediction model based on a gated cyclic unit is established, and the unstable samples in the qualitative analysis and the samples of the confidence interval with a certain confidence level in the quantitative analysis are predicted for the power angle trajectory, and the final transient stability prediction result is determined by comprehensively judging the results of the qualitative analysis, the quantitative analysis and the power angle trajectory prediction. The present invention can improve the accuracy of transient stability prediction, and at the same time, improve the reliability of transient stability prediction without losing the response speed.

Description

Technical Field

[0001] The invention belongs to the technical field of transient stability prediction of power systems, and in particular relates to a two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated cyclic unit. Background Art

[0002] With the rapid advancement of the energy transformation strategy and the large-scale grid connection of new energy power plants, the mechanism model and safety and stability issues of the new generation of smart grids are becoming increasingly complex. The random volatility and uncontrollability of new energy power generation and the complexity of power system operating conditions have increased the difficulty of power system safety analysis and control, which has brought severe challenges to the reliability and accuracy of power system transient stability prediction.

[0003] The existing transient stability analysis methods of power systems mainly include time domain simulation method, direct method and artificial intelligence method. The time domain simulation method is to establish a full system model based on the topological relationship of system components to obtain a set of simultaneous algebraic equations and differential equations, and then use the steady-state power flow calculation results as the initial value to gradually integrate and calculate the system stability. Although this method has high accuracy, its calculation amount is huge and it is difficult to meet the requirements of online stability assessment. The direct method compares the transient energy with the critical energy at the moment of fault removal from the perspective of energy, and directly judges the stability of the system, overcoming the shortcomings of the low efficiency of the time domain simulation method, but its calculation results are too conservative, and there is no unified method to construct a suitable energy function, which to a certain extent hinders its development in transient stability analysis.

[0004] With the use of wide-area monitoring systems and the rapid evolution of artificial intelligence methods represented by deep learning, data-driven methods can simultaneously meet the requirements of accuracy and speed in transient stability analysis and control, or bring a new solution and technical path for the transient stability prediction of the new generation of smart grids. Among them, convolutional neural networks are widely used in transient stability assessment of power systems due to their superior feature extraction capabilities. In addition, based on the time series characteristics of data, the advantages of time correlation in a longer range can be obtained with the help of long short-term memory networks. Recurrent neural networks are also used in transient stability assessment of power systems. However, on the one hand, convolutional neural networks are not very good at extracting global features and time series correlations of data, and recurrent neural networks are not very good at extracting local features and spatial correlations of data. Most of the known literature studies only consider the spatial characteristics or temporal characteristics of spatiotemporal big data of power systems, which is obviously incomplete. On the other hand, in terms of the reliability of transient stability prediction results, all known studies sacrifice evaluation time in exchange for the reliability of prediction results. In the quantitative analysis of transient stability, there are few studies on the reliability of critical stability margin prediction results. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and propose a two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated recurrent unit to improve the accuracy of transient stability prediction while improving the reliability of transient stability prediction without sacrificing response speed.

[0006] A two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated recurrent unit of the present invention comprises the following steps:

[0007] Step 1: Based on the generator node voltage amplitude V and phase angle Φ obtained by time domain simulation as input, a transient stability prediction model based on the improved convolutional residual memory network is established, and the corresponding stable state η is used as a response to perform a qualitative analysis of the system. At the same time, the stability margin SM is used as a response to perform a quantitative analysis of the system;

[0008] Step 2: Based on the generator power angle trajectory δ obtained by time domain simulation, take 20 steps after fault removal as input, and the next 100, 200, 300, and 400 steps as responses to establish a generator power angle trajectory prediction model;

[0009] Step 3: For those samples whose results of qualitative analysis in step 1 are unstable, and the results of quantitative analysis in step 1 are confidence intervals [θ 1 ,θ 2 ] For those samples that are critically stable, the generator power angle trajectory δ of 20 steps after the fault removal in the above samples is used as the input of step 2 to predict the generator power angle trajectory;

[0010] Step 4: Based on the results of the qualitative and quantitative analysis obtained in step 1 and the generator power angle trajectory prediction results obtained in step 3, it is comprehensively determined whether the generator is unstable, so as to determine whether emergency control measures are needed.

[0011] The above-mentioned two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit, wherein: the input features of the transient stability prediction model established in step 1 are the normalized generator node voltage amplitude V and phase angle Φ, as shown below:

[0012]

[0013] The subscript G represents the index relative to the reference generator, the subscript t represents the number of samples, and the units of the normalized formula symbols are dimensionless.

[0014] The above-mentioned two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit, wherein: the transient stability prediction model established in step 1 is composed of a gated recurrent unit and a residual block:

[0015] (1) The calculation formula of the neuron internal structure of the gated recurrent unit is as follows:

[0016] r t =σ(W r ·[h t-1 ,x t ])(2)

[0017] where r t is the reset gate, which can capture the short-term dependencies in the time series data and is used to determine the degree of forgetting previous information. σ is the sigmoid activation function, and W r To reset the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless.

[0018] z t =σ(W z ·[h t-1 ,x t ])(3)

[0019] where z t is the update gate, which can capture the long-term dependencies in the data and is used to determine the extent to which the state information of the previous moment is brought to the current state. σ is the sigmoid activation function, and W z To update the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless.

[0020] H t =tanh(W·[r t *h t-1 ,x t ])(4)

[0021] Among them, H t is the candidate hidden state at time t, tanh is the hyperbolic tangent activation function, W is the weight parameter of the candidate hidden state, r t is the reset gate, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t. * represents element-wise multiplication. ˙ represents matrix multiplication. The units of the normalized formula symbols are dimensionless.

[0022] h t =(1-z t )*h t-1 +z t *Ht (5)

[0023] where h t is the hidden state at time t, z t is the update gate, h t-1 is the hidden state at the previous moment of time t, z t is the update gate, H t is the candidate hidden state at time t. * represents element-wise multiplication. The addition in formula 5 can effectively prevent gradient diffusion. The units of the normalized formula symbols are dimensionless.

[0024] (2) The calculation formula of the residual block is as follows:

[0025] x l+1 =σ(h(x l )+F(x l ,W l ))(6)

[0026] Where h has the function of identity mapping, that is, h(xl) = xl. σ is the ReLU activation function, F represents the residual calculation, and x l represents input, x l+1 Indicates output, W l Represents the convolution operation. The units of the normalized formula symbols are dimensionless.

[0027]

[0028] Where L represents a deeper layer, X L represents the output of a deeper layer, X l represents the input of the lth layer, F represents the residual calculation, i is the number of layers, and W i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless.

[0029]

[0030] The loss function is recorded as E, L represents a deeper layer, and X L represents the output of a deeper layer, X l represents the input of the lth layer, represents partial derivative, F represents residual calculation, i represents the number of layers, W represents i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless.

[0031] The above-mentioned two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit, wherein: the final loss functions of the transient stability prediction model in step 1 for qualitative analysis and quantitative analysis are respectively as follows:

[0032] (1) Calculation formula of loss function of stability discrimination model:

[0033]

[0034] Where N is the number of samples and n is the sequence number. n 0 , T n 1 are the true labels of the samples; Y n 0 , Y n 1 are the predicted stability and instability probabilities respectively; the units of the normalized formula symbols are dimensionless.

[0035] (2) Calculation formula of loss function of stability margin prediction model:

[0036]

[0037] Where N is the number of samples, n is the sequence number, Y is the predicted value of the stability margin, T is the true value of the stability margin, unit: s.

[0038] In the above-mentioned two-stage transient stability prediction method based on the improved convolutional residual memory network and the gated recurrent unit, the response value calculation and evaluation index of the transient stability prediction model for qualitative analysis and quantitative analysis in step 1 are respectively as follows:

[0039] (1) The calculation formula of the power angle stability index TSI is:

[0040]

[0041] Where η is the stability index, |△δ|max is the maximum power angle difference between any two generators, unit: degree. If η is positive, the state label is 1; if η is negative, the state label is 0.

[0042] (2) The formula for calculating the stability margin is:

[0043] SM=CCT-Δt(12)

[0044] Where SM is the stability margin, CCT is the limit clearing time, and Δt is the fault duration, unit: s.

[0045] (3) It is proposed to use the accuracy rate ACC, recall rate recall, precision Precision and F1 score as the evaluation indicators of the stability judgment model. The calculation formula is:

[0046]

[0047] The relevant symbols required for calculation come from the confusion matrix in the following table:

[0048] Table 1 Confusion matrix

[0049]

[0050] TP means that both the predicted value and the true value are stable, FP means that the predicted value is stable and the true value is unstable. FN means that the predicted value is unstable and the true value is stable, and TN means that both the predicted value and the true value are unstable. The units of the normalized formula symbols are dimensionless.

[0051] (4) The root mean square error (RMSE) is proposed as the evaluation index of the stability margin prediction model, and its calculation formula is as follows:

[0052]

[0053] Where m is the number of test samples, i is the serial number; Y i To predict the stability margin value, y i is the actual stability margin value, unit: s;

[0054] In addition, the stability margin stable state evaluation index is introduced. The formulas of missed rate Mis, false positive rate Fal, and accuracy rate Accu are as follows:

[0055]

[0056] Where N fs is the number of instability predictions that are stable, N fu For stability prediction, the number of instability is taken, N is the number of test samples, and the units of the normalized formula symbols are dimensionless.

[0057] In the above-mentioned two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit, the input feature of the generator power angle trajectory prediction model established in step 2 is the power angle trajectory δ of the generator relative to the reference generator, in degrees as shown below:

[0058]

[0059] Where G represents the index relative to the reference generator, t represents the number of samples, and the units of the normalized formula symbols are dimensionless.

[0060] In the above two-stage transient stability prediction method based on convolutional residual memory network and gated recurrent unit, the confidence interval [θ 1 ,θ 2 ] is the interval of judgment error, and the calculation formula is as follows:

[0061] P{θ 1 ≤θ≤θ 2}=1-α (22)

[0062] Where P represents the probability value and α is the significance level. 1 and θ 2 Represent the lower confidence limit and upper confidence limit respectively, unit: s.

[0063] Compared with the prior art, the present invention has obvious beneficial effects. From the above technical scheme: the modeling method of the present invention establishes a data-driven transient stability prediction model and a power angle trajectory prediction model by utilizing a large amount of historical data. By predicting the stability in the first stage, the samples predicted to be unstable and the samples in the confidence interval at a certain confidence level are predicted to predict the power angle trajectory of the generator, thereby improving the reliability of the transient stability prediction without losing the response speed. In addition, the convolutional residual memory network proposed in the present invention is a model with spatiotemporal characteristics, which can effectively extract the spatiotemporal big data features of the power system, thereby making the prediction accuracy of the first stage higher. The present invention uses a convolutional residual memory network model with spatiotemporal characteristics to extract features according to the spatiotemporal characteristics of the spatiotemporal big data of the power system, and adopts a power angle trajectory prediction method in the second stage, which has higher prediction accuracy and reliability than the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is the internal structure diagram of the gated recurrent unit;

[0065] Figure 2 is a residual block structure diagram of the present invention;

[0066] Figure 3 It is a transient stability prediction model in the embodiment analysis;

[0067] Figure 4 A modified IEEE-39 node system diagram for use in the embodiments;

[0068] Figure 5 A two-stage transient stability prediction flow chart used in the present invention;

[0069] Figure 6 The prediction result of the generator power angle trajectory of the real-time measurement data 1;

[0070] Figure 7 The prediction result of the generator power angle trajectory of the real-time measurement data 2;

[0071] Figure 8 This is the prediction result of the generator power angle trajectory of real-time measurement data 3. DETAILED DESCRIPTION

[0072] In combination with the accompanying drawings and preferred embodiments, the specific implementation method, structure, characteristics and efficacy of a two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated recurrent unit proposed in accordance with the present invention are described in detail below.

[0073] A two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated recurrent unit of the present invention comprises the following steps:

[0074] Step 1: Based on the generator node voltage amplitude V and phase angle Φ obtained from the time domain simulation as input, a transient stability prediction model based on the improved convolutional residual memory network is established. The corresponding stable state η is used as the response to conduct a qualitative analysis of the system, and the stability margin SM is used as the response to conduct a quantitative analysis of the system.

[0075] The input characteristics of the transient stability prediction model are the voltage amplitude and phase angle of the generator node, as shown below:

[0076]

[0077] Where G represents the index relative to the reference generator, t represents the number of samples, and the units of the normalized formula symbols are dimensionless.

[0078] The established transient stability prediction model consists of a gated recurrent unit and a residual block.

[0079] (1) The calculation formula of the neuron internal structure of the gated recurrent unit is as follows:

[0080] r t =σ(W r ·[h t-1 ,x t ])(2)

[0081] where r t is the reset gate, which can capture the short-term dependencies in the time series data and is used to determine the degree of forgetting previous information. σ is the sigmoid activation function, and W r To reset the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless.

[0082] z t =σ(W z ·[h t-1 ,x t ])(3)

[0083] where z tis the update gate, which can capture the long-term dependencies in the data and is used to determine the extent to which the state information of the previous moment is brought to the current state. σ is the sigmoid activation function, and W z To update the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless.

[0084] H t =tanh(W·[r t *h t-1 ,x t ]) (4)

[0085] Among them, H t is the candidate hidden state at time t, tanh is the hyperbolic tangent activation function, W is the weight parameter of the candidate hidden state, r t is the reset gate, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t. * represents element-wise multiplication. ˙ represents matrix multiplication. The units of the normalized formula symbols are dimensionless.

[0086] h t =(1-z t )*h t-1 +z t *H t (5)

[0087] where h t is the hidden state at time t, z t is the update gate, h t-1 is the hidden state at the previous moment of time t, z t is the update gate, H t is the candidate hidden state at time t. * represents element-wise multiplication. The addition in formula 5 can effectively prevent gradient diffusion. The units of the normalized formula symbols are dimensionless.

[0088] (2) The calculation formula of the residual block is as follows:

[0089] x l+1 =σ(h(x l )+F(x l ,W l )) (6)

[0090] Where h has the function of identity mapping, that is, h(xl) = xl. σ is the ReLU activation function, F represents the residual calculation, and x l represents input, x l+1 Indicates output, Wl Represents the convolution operation. The units of the normalized formula symbols are dimensionless.

[0091]

[0092] Where L represents a deeper layer, X L represents the output of a deeper layer, X l represents the input of the lth layer, F represents the residual calculation, i is the number of layers, and W i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless.

[0093]

[0094] The loss function is recorded as E, L represents a deeper layer, and X L represents the output of a deeper layer, X l represents the input of the lth layer, represents partial derivative, F represents residual calculation, i represents the number of layers, W represents i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless.

[0095] The final loss functions of the transient stability prediction model for qualitative analysis and quantitative analysis are as follows:

[0096] (1) Calculation formula of loss function of stability discrimination model:

[0097]

[0098] Where N is the number of samples and n is the sequence number. n 0 , T n 1 are the true labels of the samples; Y n 0 , Y n 1 are the predicted stability and instability probabilities respectively. The units of the normalized formula symbols are dimensionless.

[0099] (2) Calculation formula of loss function of stability margin prediction model:

[0100]

[0101] Where N is the number of samples, n is the sequence number, Y is the predicted value of the stability margin, T is the true value of the stability margin, unit: s.

[0102] The response value calculation and evaluation indexes of the transient stability prediction model for qualitative analysis and quantitative analysis are as follows:

[0103] (1) The calculation formula of the power angle stability index TSI is:

[0104]

[0105] Where η is the stability index, |△δ|max is the maximum power angle difference between any two generators, unit: degree. If η is positive, the state label is 1; if η is negative, the state label is 0.

[0106] (2) The formula for calculating the stability margin is:

[0107] SM=CCT-Δt(12)

[0108] Where SM is the stability margin, CCT is the limit clearing time, and Δt is the fault duration, unit: s.

[0109] (3) It is proposed to use the accuracy rate ACC, recall rate recall, precision Precision and F1 score as the evaluation indicators of the stability judgment model. The calculation formula is:

[0110]

[0111] The relevant symbols required for calculation come from the confusion matrix in the following table:

[0112] Table 1 Confusion matrix

[0113]

[0114] TP means that both the predicted value and the true value are stable, FP means that the predicted value is stable and the true value is unstable. FN means that the predicted value is unstable and the true value is stable, and TN means that both the predicted value and the true value are unstable. The units of the normalized formula symbols are dimensionless.

[0115] (4) The root mean square error (RMSE) is proposed as the evaluation index of the stability margin prediction model, and its calculation formula is as follows:

[0116]

[0117] Where m is the number of test samples and i is the serial number. i To predict the stability margin value, y i is the actual stability margin value, unit: s.

[0118] In addition, the stability margin stable state evaluation index is introduced. The formulas of missed rate Mis, false positive rate Fal, and accuracy rate Accu are as follows:

[0119]

[0120]

[0121] Where N fs is the number of instability predictions that are stable. fu The stable prediction is the number of instability. N is the number of test samples; the units of the normalized formula symbols are dimensionless.

[0122] Reference Figure 3 The model of the structure shown in the figure uses the obtained power system data set as the model input for training according to the matrix shown in Formula 1 and Formula 20. In order to highlight the advantages of the invention, other commonly used models are selected for comparison. Among them, LSTM, Bi-LSTM, and GRU, which belong to the same recurrent neural network, use the same parameters and the number of hidden nodes is 128. CNN finds the optimal structure after multiple experiments. The learning rate, batch size, loss function, optimization algorithm, and model input are all consistent with the improved CRMN model.

[0123] Reference for numerical simulation experiments on training sample acquisition Figure 4 The modified IEEE-39 node system shown in the figure is connected to the wind farm at nodes 22, 29 and 39, and the total system load is 5856MW. The synchronous generators all use the GENROU model, the excitation system uses the IEEEX1 model, the prime mover-speed regulator uses the IEESGO model, and the stabilizer uses the PSS2A model; the wind turbine-converter all uses the WT3G2 model, the electrical control model uses the WT3E1 model, the mechanical control uses the WT3T1 model, and the pitch control uses the WT3P1 model; when the generator output makes the power flow converge, the power penetration rate of the new energy wind power plant is considered to be 10%, 20%, ..., 50%. And the overall power generation and load levels are evenly divided into 8 levels between 85% and 120% with a step size of 5%. In terms of fault setting, a three-phase short-circuit fault is set on 34 AC lines without a transformer. The fault duration is 0.02 to 0.2 seconds (the change step size is 0.02 seconds). Among them, the fault occurs at 0.1 seconds of simulation, and the simulation time is 6 seconds. With the help of Matlab and PSSE3.4, time domain simulation was carried out to obtain a total of 12341 samples, including 7573 stable samples and 4768 unstable samples. The data set was shuffled and randomly divided into a training set and a test set, with 10000 samples in the training set and 2341 samples in the test set.

[0124] a) Qualitative analysis, the prediction results of different models are shown in the following table:

[0125] Table 2 Transient stability prediction results of different models

[0126]

[0127] From the results in Table 2, it can be seen that the proposed model combines the advantages of convolutional neural network and recurrent neural network in processing data, and can fully explore the spatiotemporal characteristics of PMU measurement data. Its various indicators are better than other models, indicating that the use of improved CRMN as a transient stability discrimination prediction model can significantly improve the classification accuracy.

[0128] b) Quantitative analysis. The prediction results of different models are shown in the following table:

[0129] Table 3 Stability margin prediction errors of different models

[0130]

[0131] As can be seen from Table 3, the RMSE of the proposed model is only 0.0023. Among the RNN variant models, the best RMSE value is 0.0075, which is still 3.26 times that of the improved CRMN. For CNN with strong spatial feature extraction ability, the RMSE is 19.26 times that of CNN, which shows that the proposed model is superior to other models. In addition, the misjudgment rate, missed judgment rate, and accuracy rate of the improved CRMN are the best.

[0132] Step 2: Based on the generator power angle trajectory δ obtained by time domain simulation, take the 20 steps after fault removal as input, and the next 100, 200, 300, and 400 steps as responses to establish a generator power angle trajectory prediction model. The input feature of the generator power angle trajectory prediction model is the power angle trajectory of the generator relative to the reference generator, as shown below:

[0133]

[0134] Where G represents the index relative to the reference generator, t represents the number of samples, and the units of the normalized formula symbols are dimensionless.

[0135] Reference Figure 5 ,In the second stage, a generator power angle trajectory prediction model based on gated cyclic units is established;

[0136] Since it is time-consuming and inaccurate to predict all generators using one gated recurrent unit model, the present invention adopts 9 parallel gated recurrent unit sub-models to predict the power angle trajectory of the 9 generators relative to the reference generator. The 20 steps of each generator after the fault is removed are used as the input of each sub-model. The sampling time is 0.01s. Predict the next 100, 200, 300, and 400 steps. This paper selects 10,000 samples as the training set of the model, and the remaining samples are used as the test set. The model results are shown in the following table:

[0137] Table 4. Power angle trajectory prediction error of gated cycle unit

[0138]

[0139] It can be seen from Table 4 that when predicting 100 steps, the maximum RMSE value is 10.9986°, and when predicting 200 steps, the maximum RMSE is 25.3734°, both of which are within the acceptable range. Although the maximum RMSE values ​​of 300 and 400 steps are 50.6178° and 65.2004° respectively, they can basically judge the instability of multiple pendulums. Therefore, although the prediction accuracy decreases with the increase of the prediction step, the prediction model of 400 steps is obviously more adaptable than that of 100 steps, so the present invention directly uses the prediction results of 400 steps in the online application stage.

[0140] Step 3: For those samples whose results of qualitative analysis in step 1 are unstable, and the results of quantitative analysis in step 1 are confidence intervals [θ 1 ,θ 2 ]For those samples that are critically stable, the generator power angle trajectory δ of 20 steps after the fault removal in the above samples is used as the input of step 2 to predict the generator power angle trajectory; The interval selection refers to the concept of confidence, that is, the interval estimation method in mathematical statistics, the interval [θ 1 ,θ 2 ] is the interval of judgment error, and the calculation formula is as follows:

[0141] P{θ 1 ≤θ≤θ 2}=1-α (22)

[0142] Where P represents the probability value and α is the significance level. 1 and θ 2 Represent the lower confidence limit and upper confidence limit respectively, unit: s.

[0143] First, for samples with unstable results in qualitative analysis and samples with a certain confidence interval in quantitative analysis, further prediction of the generator power angle trajectory is required to intuitively visualize the situation of the system's out-of-step generator rotor, thereby preventing unnecessary emergency control measures caused by misjudgment. Figure 6 , a fault occurs at 1s, the fault line is 1-39, the load level is 85%, the penetration rate of new energy electricity is 10%, the fault removal time is 0.14s, the qualitative judgment is stable, and the stability margin prediction value is -0.00156. The power angle values ​​of 20 steps after the fault removal moment are collected as the input of the power angle trajectory prediction model of the gated cycle unit. At 1.4383s, the following is obtained Figure 6 From the power angle trajectory prediction results shown, it can be seen that the power angle trajectory prediction can still predict the generator power angle trajectory well and can intuitively show the development trend of the power angle trajectory.

[0144] Reference Figure 7, a fault occurs at 1s, the fault line 1-2, the load level is 120%, the new energy power penetration level is 40%, the fault removal time is 0.02s, the qualitative judgment is stable, and the stability margin prediction value is -0.00096. The 20-step power angle value after the fault removal moment is collected as the input of the gated cycle unit power angle trajectory prediction model. At 1.3183s, the following is obtained Figure 7 The power angle trajectory prediction results are shown. It can be seen that for the samples at the stable boundary in the quantitative analysis, although there is a certain impact on the trajectory prediction, the prediction results are still the same as the actual results.

[0145] Reference Figure 8 , a fault occurs at 1s, the fault line is 28-29, the load level is 120%, the penetration rate of new energy electricity is 50%, the fault removal time is 0.2s, and it is qualitatively judged to be unstable. The stability margin prediction value is -0.0017. The power angle values ​​of 20 steps after the fault removal moment are collected as the input of the power angle trajectory prediction model of the gated cycle unit. At 1.4983s, the following is obtained: Figure 8 From the power angle trajectory prediction results shown, it can be seen that the qualitative analysis, quantitative analysis and power angle trajectory prediction are all unstable, which shows that the reliability of the prediction can be enhanced to a certain extent.

[0146] Step 4 includes three parts: data collection, model training, and online application, and two stages: qualitative and quantitative analysis and generator power angle trajectory prediction. The last step is to comprehensively judge whether the generator is unstable based on the results of qualitative and quantitative analysis obtained in step 1 and the generator power angle trajectory prediction results obtained in step 3, so as to decide whether emergency control measures are needed.

[0147] By using the convolutional residual memory network model with spatiotemporal characteristics to extract features according to the spatiotemporal characteristics of the power system spatiotemporal big data, and adopting the power angle trajectory prediction method in the second stage, it has higher prediction accuracy and reliability than the existing technology.

[0148] In summary, the present invention proposes a two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated recurrent unit, which models the characteristics of spatiotemporal big data in the power system and improves the accuracy of transient stability prediction of the power system. On the basis of the first stage, the generator power angle trajectory prediction is added, and finally a comprehensive prediction judgment is performed to ensure the prediction response speed while improving the reliability of transient stability prediction.

[0149] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A two-stage transient stability prediction method based on an improved convolutional residual memory network and a gated recurrent unit, comprising the following steps: Step 1: Based on the generator node voltage amplitude V and phase angle Φ obtained by time domain simulation as input, a transient stability prediction model based on the improved convolutional residual memory network is established. The corresponding stable state η is used as a response to perform a qualitative analysis of the system. At the same time, the stability margin SM is used as a response to perform a quantitative analysis of the system. The established transient stability prediction model consists of a gated recurrent unit and a residual block, where: (1) The calculation formula of the neuron internal structure of the gated recurrent unit is as follows: r t =σ(W r ·[h t-1 ,x t ])(2) where r t is the reset gate, which can capture the short-term dependencies in the time series data and is used to determine the degree of forgetting previous information. σ is the sigmoid activation function, and W r To reset the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless; z t =σ(W z ·[h t-1 ,x t ])(3) where z t is the update gate, which can capture the long-term dependencies in the data and is used to determine the extent to which the state information of the previous moment is brought to the current state. σ is the sigmoid activation function, and W z To update the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless; H t =tanh(W·[r t *h t-1 ,x t ])(4) Among them, H t is the candidate hidden state at time t, tanh is the hyperbolic tangent activation function, W is the weight parameter of the candidate hidden state, r t is the reset gate, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t; * represents element-wise multiplication; ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless; h t =(1-z t )*h t-1 +z t *H t (5) where h t is the hidden state at time t, z t is the update gate, h t-1 is the hidden state at the previous moment of time t, z t is the update gate, H t is the candidate hidden state at time t; * represents element-wise multiplication. The addition of formula 5 can effectively prevent gradient diffusion. The units of the normalized formula symbols are dimensionless. (2) The calculation formula of the residual block is as follows: x l+1 =σ(h(x l )+F(x l ,W l ))(6) Among them, h has the function of identity mapping, that is, h(xl) = xl, σ is the ReLU activation function, F represents the residual calculation, x l represents input, x l+1 Indicates output, W l Represents the convolution operation, and the units of the normalized formula symbols are dimensionless; Where L represents a deeper layer, X L represents the output of a deeper layer, X l represents the input of the lth layer, F represents the residual calculation, i is the number of layers, and W i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless. The loss function is recorded as E, L represents a deeper layer, and X L represents the output of a deeper layer, X l represents the input of the lth layer, represents partial derivative, F represents residual calculation, i represents the number of layers, W represents i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless. Step 2: Based on the generator power angle trajectory δ obtained by time domain simulation, take 20 steps after fault removal as input, and the next 100, 200, 300, and 400 steps as responses to establish a generator power angle trajectory prediction model; Step 3: For those samples whose results of qualitative analysis in step 1 are unstable, and those samples whose results of quantitative analysis in step 1 are critically stable in the confidence interval [θ1, θ2], the generator power angle trajectory δ of 20 steps after the fault removal in the above samples is used as the input of step 2 to predict the generator power angle trajectory; Step 4: Based on the results of the qualitative and quantitative analysis obtained in step 1 and the generator power angle trajectory prediction results obtained in step 3, it is comprehensively determined whether the generator is unstable, so as to determine whether emergency control measures are needed.

2. The two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit according to claim 1, wherein: The input features of the transient stability prediction model established in step 1 are the normalized generator node voltage amplitude V and phase angle Φ, as shown below: The subscript G represents the index relative to the reference generator, the subscript t represents the number of samples, and the units of the normalized formula symbols are dimensionless.

3. The two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit according to claim 1, wherein: The transient stability prediction model established in step 1 is composed of a gated recurrent unit and a residual block: (1) The calculation formula of the neuron internal structure of the gated recurrent unit is as follows: r t =σ(W r ·[h t-1 ,x t ])(2) where r t is the reset gate, which can capture the short-term dependencies in the time series data and is used to determine the degree of forgetting previous information. σ is the sigmoid activation function, and W r To reset the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless; z t =σ(W z ·[h t-1 ,x t ])(3) where z t is the update gate, which can capture the long-term dependencies in the data and is used to determine the extent to which the state information of the previous moment is brought to the current state. σ is the sigmoid activation function, and W z To update the gate weight, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless; H t =tanh(W·[r t *h t-1 ,x t ])(4) Among them, H t is the candidate hidden state at time t, tanh is the hyperbolic tangent activation function, W is the weight parameter of the candidate hidden state, r t is the reset gate, h t-1 is the hidden state at the previous moment of time t, x t is the time series input at time t, * represents element-wise multiplication, ˙ represents matrix multiplication, and the units of the normalized formula symbols are dimensionless; h t =(1-z t )*h t-1 +z t *H t (5) where h t is the hidden state at time t, z t is the update gate, h t-1 is the hidden state at the previous moment of time t, z t is the update gate, H t is the candidate hidden state at time t, * represents element-wise multiplication, the addition of formula 5 can effectively prevent gradient diffusion, and the units of the normalized formula symbols are dimensionless; (2) The calculation formula of the residual block is as follows: x l+1 =σ(h(x l )+F(x l ,W l ))(6) Among them, h has the function of identity mapping, that is, h(xl) = xl, σ is the ReLU activation function, F represents the residual calculation, x l represents input, x l+1 Indicates output, W l Represents the convolution operation, and the units of the normalized formula symbols are dimensionless; Where L represents a deeper layer, X L represents the output of a deeper layer, X l represents the input of the lth layer, F represents the residual calculation, i is the number of layers, W i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless. The loss function is recorded as E, L represents a deeper layer, and X L represents the output of a deeper layer, X l represents the input of the lth layer, represents partial derivative, F represents residual calculation, i represents the number of layers, W represents i Represents the convolution operation of the i-th layer. The units of the normalized formula symbols are dimensionless.

4. The two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit according to claim 1, wherein: The final loss functions of the transient stability prediction model for qualitative analysis and quantitative analysis in step 1 are respectively as follows: (1) Calculation formula of loss function of stability discrimination model: Where N is the number of samples, n represents the sequence number, are the true labels of the samples respectively; It is divided into predicted stability and instability probabilities; the units of the normalized formula symbols are dimensionless; (2) Calculation formula of loss function of stability margin prediction model: Where N is the number of samples, n represents the sequence number, Y is the predicted value of the stability margin, and T is the true value of the stability margin, unit: s.

5. The two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit according to claim 1, wherein the response value calculation and evaluation index of the transient stability prediction model for qualitative analysis and quantitative analysis in step 1 are respectively as follows: (1) The calculation formula of the power angle stability index TSI is: Where η is the stability index, |△δ|max is the maximum power angle difference between any two generators, unit: degree, if η is positive, the state label is 1; if η is negative, the state label is 0; (2) The formula for calculating the stability margin is: SM=CCT-Δt(12) Where SM is the stability margin, CCT is the limit removal time, Δt is the fault duration, unit: s; (3) It is proposed to use the accuracy rate ACC, recall rate recall, precision Precision and F1 score as the evaluation indicators of the stability judgment model. The calculation formula is: in, The relevant symbols required for calculation come from the confusion matrix in the following table: Table 1 Confusion matrix TP means that both the predicted value and the true value are stable, FP means that the predicted value is stable and the true value is unstable, FN means that the predicted value is unstable and the true value is stable, TN means that both the predicted value and the true value are unstable, and the units of all symbols are dimensionless; (4) The root mean square error (RMSE) is proposed as the evaluation index of the stability margin prediction model, and its calculation formula is as follows: Where m is the number of test samples, i is the serial number; Yi is the predicted stability margin value, yi is the actual stability margin value, unit: s; The stable state evaluation indicators of the stability margin, the missed rate Mis, the false positive rate Fal, and the accuracy rate Accu formula are as follows: Where N fs is the number of instability predictions that are stable, N fu For stability prediction, the number of instability is taken, N is the number of test samples, and the units of the normalized formula symbols are dimensionless.

6. The two-stage transient stability prediction method based on improved convolutional residual memory network and gated recurrent unit according to claim 1, wherein the input feature of the generator power angle trajectory prediction model established in step 2 is the power angle trajectory δ of the generator relative to the reference generator, unit: degree, as shown below: Where G represents the index relative to the reference generator, t represents the number of samples, and the units of the normalized formula symbols are dimensionless.

7. The two-stage transient stability prediction method based on the improved convolutional residual memory network and the gated recurrent unit according to claim 1, wherein the confidence interval [θ1, θ2] in step 3 is an interval of error judgment, and the calculation formula is as follows: P{θ1≤θ≤θ2}=1-α(22) Where P represents the probability value, α is the significance level, θ1 and θ2 represent the lower confidence limit and upper confidence limit respectively, unit: s.

Images