Wind power plant dynamic equivalent modeling method, system and equipment based on residual combined neural network and medium

Abstract

The invention relates to a wind power plant dynamic equivalent modeling method based on a residual combined neural network, which comprises the following steps of: simplifying a mathematical model of a wind turbine generator, taking the wind speed of an anemometer tower and the voltage at a common coupling point as input characteristics, and taking the active power and the reactive power of a wind power plant as output characteristics; selecting an LSTM unit as a core to construct a feature memory layer, and extracting key features capable of representing the dynamic features of the wind power plant; an information flow acceleration layer with Res-GRU as a core is constructed, and the key features are optimized and refined; constructing a data relation mapping layer based on FC-BN-Relu, and carrying out integration and fine mapping on the feature information processed by the previous level; and optimizing the parameters of the previous levels by using a genetic algorithm. The residual combination network constructed by the method is used for dynamic equivalent modeling, the calculation complexity is simplified, the equivalent precision is improved, the method conforms to the actual situation of a wind power plant mathematical model, and the method is more universal than a mechanism modeling method.

Description

Technical Field

[0001] The invention relates to a wind farm modeling method, and in particular to a wind farm dynamic equivalent modeling method, system, equipment and medium based on a residual combination neural network. Background Art

[0002] As one of the most important and mature power generation methods in the field of renewable energy, wind power generation accounts for a growing proportion of the global energy supply. According to the Global Wind Energy Report 2024 released by the Global Wind Energy Council, the global wind power installed capacity will reach 117GW in 2023, an increase of 50% compared with 2022, setting a new growth record. However, the output of wind energy is inevitably intermittent and random, and the centralized grid connection of large-scale wind farms will pose serious challenges to the safe and stable operation of the power system. Therefore, it is urgent to study simulation models that can accurately reflect the dynamic characteristics of wind farms. Generally speaking, wind farms usually contain dozens or even hundreds of units. If a detailed model of each wind turbine is built, it will usually lead to high model calculation complexity and long calculation time. Therefore, it is necessary to simplify the detailed model of the wind farm to reduce the calculation scale while ensuring the simulation accuracy. Summary of the invention

[0003] In view of the shortcomings of the existing methods, the present invention provides a wind farm dynamic equivalent modeling method, system, equipment and medium based on residual combination neural network, which simplifies the calculation complexity and improves the equivalent accuracy.

[0004] To achieve the above purpose, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for dynamic equivalent modeling of a wind farm based on a residual combination neural network, comprising the following steps:

[0006] S1. Simplify the mathematical model of wind turbines, with the wind speed of the wind tower and the voltage at the common coupling point as input characteristics, and the active and reactive power of the wind farm as output characteristics;

[0007] S2, select LSTM unit as the core to build the feature memory layer and extract key features that can characterize the dynamic characteristics of the wind farm;

[0008] S3. Build an information flow acceleration layer with Res-GRU as the core to optimize and refine the key features;

[0009] S4. Construct a data relationship mapping layer based on FC-BN-Relu to integrate and finely map the key features optimized and refined by the information flow acceleration layer;

[0010] S5. Use genetic algorithm GA to optimize the parameters of feature memory layer, information flow acceleration layer, and data relationship mapping layer.

[0011] Preferably, the simplified mathematical model of the wind turbine generator set in step S1 includes:

[0012] 1) For variable speed wind turbines, the wind speed-power relationship is as follows:

[0013]

[0014] Where P mech is the output power of the wind turbine; ρ is the air density, unit is kg / m 3 ; A r is the swept area of ​​the wind turbine blade; V w is the wind speed; V in and V out are the cut-in wind speed and the cut-out wind speed respectively; C p is the wind energy utilization coefficient; λ is the tip speed ratio; θ is the pitch angle; R ae is the radius of the wind turbine, in m;

[0015] 2) Based on the wind speed-power relationship, the mathematical model of the wind turbine is simplified into a set of equations:

[0016]

[0017] Where: x is the state variable of the wind turbine; u pcc is the voltage at the PCC of the wind farm; y is the active power and reactive power output by the wind turbine; f1 and f2 are function operation symbols, which are determined by the control strategy;

[0018] 3) Differentiating the equation in step 2) yields:

[0019] x t =f3(x t-1 ,u pcc,t-1 ,u pcc,t )

[0020] In the formula, the subscript t represents the current time, t-1 represents the previous time, and f3 is the function operator.

[0021] Preferably, the feature memory layer in step S2 is composed of an LSTM unit and a leaky linear rectifier LeakyReLU function, and the LSTM unit extracts key features that can characterize the dynamic characteristics of the wind farm; the LeakyReLU function performs nonlinear activation on the key features, and the expression is as follows:

[0022]

[0023] Where: λ is a constant between 0 and 1;

[0024] The relationship between the input and output of the feature memory layer is as follows:

[0025] h t =F(h t-1 ,u t ,u t-1 ,W,b f )

[0026] Where: F is the function operator, b f is the bias, u is the input, and W is the weight vector.

[0027] Preferably, the information flow acceleration layer in step S3 is composed of two paths, the first path includes a GRU unit and a ReLU function, and the second path directly transmits the key features extracted by the feature memory layer. The two paths are added together at the end to form a residual structure, and the residual structure algorithm formula is:

[0028] G(x0)=g(x0)+x0

[0029] Where: x0 is the original input of the residual structure; g(x0) represents the output result of the residual block; G(x0) represents the output result of the entire residual structure.

[0030] Preferably, the data relationship mapping layer in step S4 is composed of a fully connected FC unit, a batch normalization BN unit and a ReLU function; the fully connected FC unit is used to linearly combine and transform the feature information transmitted from the information flow acceleration layer, and the formula is as follows:

[0031] y=Wx+b

[0032] In the formula, x is the input vector; y is the output vector; W is the weight matrix; b is the bias vector;

[0033] The batch normalization BN unit normalizes the activation values ​​of each batch of data;

[0034] The ReLU function introduces nonlinear factors and performs nonlinear activation on the normalized data;

[0035] The data relationship mapping layer expression is:

[0036]

[0037] Where: represents the activation function; x i is the input quantity; b i is the offset.

[0038] Preferably, the use of a genetic algorithm GA in step S5 to optimize the parameters of the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer includes:

[0039] 1) Determine the optimized parameter as the number of neurons, and encode the number of neurons in each layer into population chromosomes. The structure of a single chromosome is as follows:

[0040] D={a1,a2,a1',a'2,a'3}

[0041] Where: D is the chromosome structure, a1 is the number of neurons in the LSTM unit, and a2 is the number of neurons in the GRU unit.

[0042] The number of neurons, a1, a2, and a3 are the number of neurons in the first, second, and third fully connected layers respectively;

[0043] 2) Initialize the population and randomly generate a set of initial solutions as the population;

[0044] 3) Set the fitness function and select excellent chromosomes to enter the next generation according to the fitness function, and then enhance genetic diversity through crossover and mutation operations to generate new solutions;

[0045] 4) Terminate GA according to the set termination condition and output the chromosome with the highest fitness as the optimal solution.

[0046] Preferably, the fitness function is:

[0047] T=b1ξ+b2ξ'

[0048] Where: T is the fitness function; ξ is the mean square error of the training set; ξ' is the mean square error of the test set; b1 and b2 are the weight coefficients of each indicator.

[0049] In a second aspect, the present invention provides a wind farm dynamic equivalent modeling system based on a residual combination neural network, comprising:

[0050] Wind turbine mathematical model simplification module, used to simplify the mathematical model of wind turbines, with the wind speed of the wind tower and the voltage at the common coupling point as input features, and the active and reactive power of the wind farm as output features;

[0051] The feature memory layer construction module is used to select LSTM units as the core to construct the feature memory layer and extract key features that can characterize the dynamic characteristics of the wind farm;

[0052] An information flow acceleration layer construction module is used to construct an information flow acceleration layer with Res-GRU as the core, and optimize and refine the key features;

[0053] A data relationship mapping layer construction module is used to construct a data relationship mapping layer based on FC-BN-Relu, and integrate and finely map the key features optimized and refined by the information flow acceleration layer;

[0054] The parameter optimization module is used to optimize the parameters of the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer using the genetic algorithm GA.

[0055] In a third aspect, the present invention provides an electronic device, comprising a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the equivalent modeling method.

[0056] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the equivalent modeling method.

[0057] The positive beneficial effects of the present invention are:

[0058] 1. The present invention first simplifies the mathematical model of the wind turbine into a set of equations, the model input is the wind speed of the wind tower and the voltage at the common coupling point, and the model output is the active and reactive power of the wind farm; then the LSTM unit similar to it in mathematical expression is selected as the core to construct the feature memory layer, the input data is processed, and the key features that can characterize the dynamic characteristics of the wind farm are extracted; then, in order to reduce the risk of gradient disappearance and accelerate network training at the same time, an information flow acceleration layer with Res-GRU as the core is constructed, the key features are optimized and refined, the residual structure is introduced into the combined neural network and the network architecture is optimized, which can improve the equivalent accuracy of the model; then the data relationship mapping layer based on FC-BN-Relu is introduced, and the information processed by the information flow acceleration layer is integrated and finely mapped; finally, after the network model is constructed, the genetic algorithm GA is used to optimize the number of neurons in the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer, and the equivalent model is built, and the genetic algorithm is used to optimize the network parameters, which can further improve the model accuracy and calculation efficiency. The present invention adopts a residual combination network constructed with LSTM, GRU and FC as the backbone for dynamic equivalent modeling, which simplifies the calculation complexity and improves the equivalent accuracy, conforms to the actual situation of the mathematical model of the wind farm, and is more universal than the mechanism modeling method. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is the structural diagram of the wind farm;

[0060] Figure 2 It is the change graph of fitness function index;

[0061] Figure 3This is a diagram of the equivalent results (active power) of the model before and after optimization;

[0062] Figure 4 This is the diagram of equivalent results (reactive power) of the model before and after optimization;

[0063] Figure 5 Comparison of modeling effects (active power) of different models;

[0064] Figure 6 Comparison of modeling effects (reactive power) of different models;

[0065] Figure 7 This is a graph showing the change in loss function during training;

[0066] Figure 8 This is a comparison diagram (active power) between the model of the present invention and the mechanism modeling results;

[0067] Fig. 9 This is a comparison diagram (reactive power) between the model of the present invention and the mechanism modeling results. DETAILED DESCRIPTION

[0068] The present invention is further described below in conjunction with some specific embodiments.

[0069] A wind farm dynamic equivalent modeling method based on residual combination neural network includes the following steps:

[0070] S1. Simplify the mathematical model of the wind turbine into a set of equations, with the wind speed of the wind tower and the voltage at the common coupling point as input characteristics, and the active and reactive power of the wind farm as output characteristics;

[0071] S2, select LSTM units that are similar to the mathematical expression of the wind turbine model as the core to build a feature memory layer and extract key features that can characterize the dynamic characteristics of the wind farm;

[0072] S3. Build an information flow acceleration layer with Res-GRU as the core to optimize and refine the key features;

[0073] S4. Construct a data relationship mapping layer based on FC-BN-Relu to integrate and finely map the key features optimized and refined by the information flow acceleration layer;

[0074] S5. Use genetic algorithm GA to optimize the parameters of feature memory layer, information flow acceleration layer, and data relationship mapping layer.

[0075] Preferably, step S1 simplifies the mathematical model of the wind turbine into a set of equations, including:

[0076] 1) For variable speed wind turbines, the wind speed-power relationship is as follows:

[0077]

[0078] Where P mech is the output power of the wind turbine; ρ is the air density, unit is kg / m 3 ; A r is the swept area of ​​the wind turbine blade; V w is the wind speed; V in and V out are the cut-in wind speed and the cut-out wind speed respectively; C p is the wind energy utilization coefficient; λ is the tip speed ratio; θ is the pitch angle; R ae is the radius of the wind turbine, in m;

[0079] 2) Based on the wind speed-power relationship, referring to the mathematical model of wind turbines in the power system analysis and comprehensive program platform developed by the China Electric Power Research Institute, and combining the low voltage ride-through control strategy of wind turbines, the mathematical model of wind turbines is simplified into a set of equations:

[0080]

[0081] Where: x is the state variable of the wind turbine; u pcc is the voltage at the PCC of the wind farm; y is the active power and reactive power output by the wind turbine; f1 and f2 are function operation symbols, which are determined by the control strategy;

[0082] 3) Differentiating the equation in step 2) yields:

[0083] x t =f3(x t-1 ,u pcc,t-1 ,u pcc,t ) (4)

[0084] In the formula, the subscript t represents the current time, t-1 represents the previous time, and f3 is the function operator.

[0085] Preferably, the feature memory layer in step S2 is composed of LSTM units and leaky linear rectifier LeakyReLU functions. The LSTM unit is the core of the feature memory layer. It effectively processes and memorizes long-term time series dependencies through a special gating mechanism, including an input gate, a forget gate, and an output gate. The LSTM unit extracts key features that can characterize the dynamic characteristics of the wind farm; the LeakyReLU function is introduced as an activation function to perform nonlinear activation on the key features to enhance the nonlinear expression ability of the network. LeakyReLU is an improved version of ReLU, which is the same as ReLU in the positive range, but allows a small gradient in the negative range, thereby avoiding the problem of neuron death. The expression is as follows:

[0086]

[0087] Where: λ is a constant between 0 and 1;

[0088] The relationship between the input and output of the feature memory layer is as follows:

[0089] h t =F(h t-1 ,u t ,u t-1 ,W,b f ) (6)

[0090] Where: F is the function operator, b f is the bias, u is the input, and W is the weight vector.

[0091] Preferably, the information flow acceleration layer in step S3 is composed of two paths, the first path includes a GRU unit and a ReLU function, and the second path directly transmits the key features extracted by the feature memory layer, and the two paths are added together at the end to form a residual structure.

[0092] GRU is similar to LSTM and also introduces a gating mechanism to solve the gradient vanishing and gradient exploding problems. The difference is that GRU simplifies the gating mechanism and only has an update gate and a reset gate. This simplified structure effectively improves the training speed while maintaining similar accuracy to LSTM.

[0093] After GRU processes the data, the ReLU activation function is applied to the key features extracted by the feature memory layer. The characteristic of the ReLU function is that it is linear for positive input values ​​and outputs zero for negative input values. This characteristic helps to reduce the gradient vanishing problem of the network and speed up the training of the network.

[0094] The residual structure is proposed to solve the problem of network degradation when there are too many hidden layers in the neural network. The core idea is to introduce a skip layer connection for the residual block in the network, which directly adds the input of the residual block to its output. The residual structure algorithm formula is:

[0095] G(x0)=g(x0)+x0 (7)

[0096] Where: x0 is the original input of the residual structure; g(x0) represents the output result of the residual block; G(x0) represents the output result of the entire residual structure.

[0097] Preferably, the data relationship mapping layer in step S4 is composed of a full connection (FC) unit, a batch normalization (BN) unit and a ReLU function; the fully connected FC unit is the basis of the data relationship mapping layer, and is used to linearly combine and transform the feature information transmitted by the information flow acceleration layer. Each fully connected layer can learn high-level abstract representations of the input data and map these representations to a new feature space. The formula is as follows:

[0098] y=Wx+b (8)

[0099] In the formula, x is the input vector; y is the output vector; W is the weight matrix; b is the bias vector;

[0100] The BN layer follows the fully connected layer and plays a vital role in maintaining the stability of the network during training. The batch normalization BN unit normalizes the activation values ​​of each batch of data, reducing the internal covariate shift, thereby helping the network maintain a stable activation distribution during training.

[0101] After the BN layer, the ReLU function introduces nonlinear factors and performs nonlinear activation on the normalized data, thus enhancing the network’s expressive power.

[0102] In the data relationship mapping layer, the FC-BN-ReLU unit is repeated three times. Each iteration further refines and strengthens the network's understanding of the data. In order to highlight the input-output relationship, the data relationship mapping layer expression is:

[0103]

[0104] Where: represents the activation function; x i is the input quantity; b i is the offset.

[0105] Preferably, the use of a genetic algorithm GA in step S5 to optimize the parameters of the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer includes:

[0106] 1) Determine the optimized parameter as the number of neurons, and encode the number of neurons in each layer into population chromosomes. The structure of a single chromosome is as follows:

[0107] D={a1,a2,a1',a'2,a'3} (10)

[0108] Where: D is the chromosome structure, a1 is the number of neurons in the LSTM unit, and a2 is the number of neurons in the GRU unit.

[0109] The number of neurons, a1, a2, and a3 are the number of neurons in the first, second, and third fully connected layers respectively;

[0110] 2) Initialize the population and randomly generate a set of initial solutions as the population;

[0111] 3) Set the fitness function and select excellent chromosomes to enter the next generation according to the fitness function, and then enhance genetic diversity through crossover and mutation operations to generate new solutions;

[0112] 4) Terminate GA according to the set termination condition and output the chromosome with the highest fitness as the optimal solution.

[0113] Preferably, the mean square error of the training set and the mean square error of the test set are used as reference indicators of the fitness function, and the fitness function is:

[0114] T=b1ξ+b2ξ' (11)

[0115] Where: T is the fitness function; ξ is the mean square error of the training set, which is used to measure the accuracy of the model training; ξ' is the mean square error of the test set, which is used to measure the generalization ability of the model; b1 and b2 are the weight coefficients of each indicator, which are set to 0.6 and 0.4 respectively according to the task requirements.

[0116] By comparing the differential equation (4) of the simplified mathematical model of the wind turbine with the simplified expressions (6) and (7) of the feature memory layer and information flow acceleration layer in the residual combination neural network, it can be seen that the two have a high degree of similarity in terms of expression and input-output relationship. The output at the current moment is related to the input at the current moment and the input and output at the previous moment. Therefore, the feature memory layer and information flow acceleration layer can be used to simulate the differential equation in the mathematical model of the wind farm. At the same time, the algebraic equation in the mathematical model of the wind turbine (3) and the simplified expression (9) of the data relationship mapping layer also have formal similarity, that is, the input and bias in (9) are used to represent the voltage and state quantity at the PCC in the algebraic equation of (3), and the data relationship mapping layer can be used to model the differential equation of the mathematical model of the wind farm. Based on this, a combined network wind farm equivalent model containing a residual structure is constructed, thereby accurately simulating the dynamic operation characteristics of the wind farm, simplifying the calculation complexity and improving the equivalent accuracy, which conforms to the actual situation of the mathematical model of the wind farm and is more universal than the mechanism modeling method.

[0117] A typical wind farm in northern Henan Province is selected as an example for analysis and verification. Figure 1The detailed parameters of the wind farm are shown in Table 1. The wind farm contains 40 double-fed wind turbines, with a single unit capacity of 2.5MW and a total wind farm capacity of 100MW. The rated voltage of a single wind turbine is 0.69kV, and it is connected to the 35kV collection system through a 2.75MVA step-up transformer in a unit connection mode. The wind turbines in the field are connected to the 35kV collection system by four overhead lines, and each line is connected to 9, 9, 11, and 11 wind turbines respectively. The collection system is boosted to 110 kV and connected to the 110 kV busbar of the 220kV substation through a 20-kilometer transmission line. During the simulation, all wind turbines were online, and the low voltage ride-through control strategy adopted the specified current control method.

[0118] The voltage drop degree is 40%, 50%, 60%, 70%, and 80%, and the wind speed in each case changes between the cut-in wind speed of 3m / s and the cut-out wind speed of 25m / s, with a change interval of 1m / s. The total data sampling time is 5s, and the sampling interval is 0.01s. Therefore, there are 50 groups (5×10) of data, each with 500 data. At the same time, the training set and the test set are divided into 8:2.

[0119] Table 1 Detailed parameters of wind farm

[0120]

[0121]

[0122] The present invention uses the paddlepaddle framework to build a neural network, and the training and testing environment is a 32GB NVIDIA Tesla V100 GPU, a 4Cores CPU and 32GB of RAM.

[0123] The root mean square error (RMSE) is selected as the loss function, and its specific content is:

[0124]

[0125] Where: n is the number of samples, y i is the actual value of the ith sample, is the predicted value of the ith sample.

[0126] At the same time, the mean square error (MSE) is selected as the evaluation indicator, and its specific content is:

[0127]

[0128] Where: n is the number of samples, y i is the actual value of the ith sample, is the predicted value of the ith sample.

[0129] In order to achieve the best network effect, GA is used to optimize the optimization object. The settings of GA are: the number of individuals in the population is 30, the expected number of iterations is 50, the crossover rate is 0.3, and the mutation rate is 0.15.

[0130] The mean square error of the training set and the mean square error of the test set are used as reference indicators to construct the fitness function.

[0131] T=b1ξ+b2ξ'

[0132] Where: T is the fitness function; ξ is the mean square error of the training set, which is used to measure the accuracy of the model training; ξ' is the mean square error of the test set, which is used to measure the generalization ability of the model; b1 and b2 are the weight coefficients of each indicator, which are set to 0.6 and 0.4 respectively according to the task requirements.

[0133] The number of neurons in the LSTM unit varies from 200 to 300, the number of neurons in the GRU unit varies from 200 to 300, the number of neurons in the first layer FC varies from 100 to 200, the number of neurons in the second layer FC varies from 60 to 100, and the number of neurons in the third layer FC varies from 20 to 60. Due to the constraints of the network structure, the number of neurons in the LSTM unit and the GRU unit must be consistent, so when performing genetic algorithm optimization, the range of their neuron numbers is set to be consistent. The termination condition is to terminate after reaching the expected number of iterations. The optimization process is as follows: Figure 2 shown.

[0134] After GA optimization, the optimal parameters are: the number of neurons in the LSTM unit is 290, the number of neurons in the GRU unit is 290, the number of neurons in the first layer FC is 100, the number of neurons in the second layer is 65, and the number of neurons in the third layer is 25.

[0135] (1) Comparison of modeling effects before and after GA optimization

[0136] The residual combination network after GA optimization is used to carry out dynamic equivalent modeling of wind farms, and it is verified on sample data and compared with the residual combination network before optimization.

[0137] The network parameters before GA optimization were set as follows: the number of neurons in the LSTM unit was 250, the number of neurons in the GRU unit was 250, the number of neurons in the first layer FC was 150, the number of neurons in the second layer was 80, and the number of neurons in the third layer was 40.

[0138] After the network training before and after optimization was completed, the comparison was performed on the sample data when the voltage dropped to 50%, a three-phase short circuit fault occurred at 1.0s, the fault was removed at 1.3s, and the input wind speed was 3m / s. The time period from 0.8s to 1.8s was selected to improve the visualization effect of the comparison. The test results are as follows: Figure 3 and Figure 4 shown.

[0139] Depend on Figure 3 and Figure 4 It can be seen from the results that when the system is in a steady state, the difference in effect before and after the optimization of the equivalent model is not obvious, and both are approximately consistent with the output results of the detailed model; but when a three-phase short circuit fault occurs in the system and the fault is removed, the equivalent model optimized by GA can more accurately fit the power output of the detailed model, while the unoptimized model has a larger equivalent error in comparison. This result shows that the GA optimized combined residual neural network selected by the present invention has higher accuracy in the dynamic equivalent of the power output characteristics of the wind farm.

[0140] (2) Comparison with other data-driven modeling methods

[0141] In order to verify the superiority of the modeling effect of the network proposed in the present invention, a single-layer LSTM network, a single-layer GRU network, a double-layer LSTM network, a double-layer GRU network, and a GRU-LSTM-FC combination network are selected to compare the effect with the network proposed in the present invention. The comparison is carried out on the sample data of a three-phase short circuit fault occurring at 1.0s, the voltage dropping to 60%, the fault being removed at 1.5s, and the input wind speed being 3m / s. The time period from 0.8s to 1.8s is selected to improve the visualization effect of the comparison.

[0142] Figure 5 and 6 The figure shows the power output results of wind farm equivalent modeling using the above six network models. It can be seen from the figure that when the system is in a steady state, the power output results of different models are not significantly different; but in the time period of fault occurrence and recovery, only the residual combination neural network equivalent model proposed in the present invention can better fit the power output changes of the detailed model, while the output results of other network models are greatly different from the power output of the detailed model, especially at the power inflection point, the fitting effect is not ideal. The above test results show that compared with other network models, the method proposed in the present invention has higher dynamic equivalent modeling accuracy.

[0143] In order to further quantitatively evaluate the performance of different modeling methods, Tables 2 and 3 give the mean square error and training time of the power output results of wind farm equivalent modeling using different methods. From the perspective of network model training time, the time cost required for the single-layer GRU network is the smallest. The training time of the GRU-LSTM-FC combination network is reduced compared to the double-layer LSTM and double-layer GRU networks. The network proposed in the present invention has a more complex network structure, so the time cost required is the largest, but the training time is still within an acceptable range; from the mean square error calculation results, in the comparison model, the error of the single-layer GRU neural network is smaller, and the model error result of the double-layer LSTM neural network is larger, but the equivalent accuracy of the residual combination neural network model proposed in the present invention is nearly an order of magnitude higher than that of all other network models, and the modeling accuracy improvement effect is extremely obvious.

[0144] Table 2 Comparison of modeling performance of 6 models on active power data

[0145]

[0146] Table 3 Comparison of modeling performance of 6 models on reactive power data

[0147]

[0148]

[0149] Figure 7 The figure shows the changes in the loss functions of different network models during the training process. It can be seen from the figure that during the training process, the network proposed in the present invention converges first and reaches a steady state after about 50 iterations, while the number of iterations for other network models to reach a steady state is greater than 50 times, among which the single-layer LSTM neural network has the slowest convergence speed, exceeding 100 times; the values ​​of the loss functions of the network models in the other comparison groups are all stable at a relatively high level, and the model accuracy is relatively poor, while the steady-state value of the loss function of the equivalent model of the residual combination network proposed in the present invention is significantly lower than this level, which shows that the model equivalent accuracy is relatively high.

[0150] (3) Comparison with mechanism modeling methods

[0151] In order to verify the superiority of the method proposed in the present invention, the modeling effect of the wind farm equivalent model established by the method proposed in the present invention and the mechanism modeling method proposed in the article "Dynamic Equivalence of Wind Farm Based on Improved K-means Clustering Algorithm" published by Wu Hongbin et al. in the 11th issue of Volume 39 of the Acta Energiae Solaris Sinica is further compared. The specific mechanism modeling method is: firstly, the wind speed of the wind farm, the output power of the wind turbine and the slip rate are selected as clustering indicators, and then the improved K-means clustering algorithm is used to group the units, and the profile value and cluster index change curve are comprehensively analyzed. Figure 1The wind farm shown is divided into four groups, and the grouping results are shown in Table 4. Then the capacity weighted method is used to calculate the equivalent unit parameters, and the equivalent loss power method is used to calculate the collection system impedance parameters.

[0152] Table 4 Clustering results of mechanism modeling units

[0153]

[0154] Under the sample working condition that a three-phase short circuit fault occurs at 1.0s, the voltage drops to 60%, the fault is removed at 1.5s, and the input wind speed is 3m / s, the mechanism modeling method and the residual combination network equivalent model proposed in this paper are compared with the detailed model of the wind farm. The test results are as follows Figure 8 and 9 shown.

[0155] from Figure 8 and 9 The results show that when a three-phase ground short circuit occurs in the system, the modeling method proposed in the present invention and the mechanism modeling method can better fit the power change of the detailed model with high accuracy. Fig. 9 It can be seen that within 0.2 seconds after the fault occurs, the reactive power output curve of the model proposed in the present invention has a smaller fitting error than the reactive power output curve of the mechanism model, and can more accurately reflect the power output of the wind farm. Therefore, the equivalent accuracy of the residual combination network model proposed in the present invention is improved compared with the mechanism modeling.

[0156] Based on similar inventive concepts, the present invention also provides a wind farm dynamic equivalent modeling system based on residual combination neural network, comprising:

[0157] Wind turbine mathematical model simplification module, used to simplify the mathematical model of wind turbines, with the wind speed of the wind tower and the voltage at the common coupling point as input features, and the active and reactive power of the wind farm as output features;

[0158] The feature memory layer construction module is used to select LSTM units as the core to construct the feature memory layer and extract key features that can characterize the dynamic characteristics of the wind farm;

[0159] An information flow acceleration layer construction module is used to construct an information flow acceleration layer with Res-GRU as the core, and optimize and refine the key features;

[0160] A data relationship mapping layer construction module is used to construct a data relationship mapping layer based on FC-BN-Relu, and integrate and finely map the key features optimized and refined by the information flow acceleration layer;

[0161] The parameter optimization module is used to optimize the parameters of the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer using the genetic algorithm GA.

[0162] Based on similar inventive concepts, the present invention also provides an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the equivalent modeling method.

[0163] The processor and the memory may be connected via a bus or other means.

[0164] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0165] It should be noted that the power system transient stability assessment system in this embodiment may include a business processing module, an edge database, a server version information register, and a data synchronization module. When the processor executes the computer program, it implements the power grid flood control risk probability prediction method based on the improved Transformer as applied to the power system transient stability assessment system as mentioned above.

[0166] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0167] Based on similar inventive concepts, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the equivalent modeling method.

[0168] It will be appreciated by those skilled in the art that all or some of the steps and systems in the disclosed method above may be implemented as software, firmware, hardware and appropriate combinations thereof. Some physical components or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or may be implemented as hardware, or may be implemented as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or a non-transitory medium) and a communication medium (or a temporary medium). As known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that may be used to store desired information and may be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in the field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A wind farm dynamic equivalent modeling method based on residual combination neural network, characterized in that: The steps include: S1. Simplify the mathematical model of wind turbines, with the wind speed of the wind tower and the voltage at the common coupling point as input characteristics, and the active and reactive power of the wind farm as output characteristics; S2, select LSTM unit as the core to build the feature memory layer and extract key features that can characterize the dynamic characteristics of the wind farm; S3. Build an information flow acceleration layer with Res-GRU as the core to optimize and refine the key features; S4. Construct a data relationship mapping layer based on FC-BN-Relu to integrate and finely map the key features optimized and refined by the information flow acceleration layer; S5. Use genetic algorithm GA to optimize the parameters of feature memory layer, information flow acceleration layer, and data relationship mapping layer.

2. The method for dynamic equivalent modeling of a wind farm based on a residual combination neural network according to claim 1 is characterized in that: The simplified mathematical model of the wind turbine generator set in step S1 includes: 1) For variable speed wind turbines, the wind speed-power relationship is as follows: Where P mech is the output power of the wind turbine; ρ is the air density, unit is kg / m 3 ; A r is the swept area of ​​the wind turbine blade; V w is the wind speed; V in and V out are the cut-in wind speed and the cut-out wind speed respectively; C p is the wind energy utilization coefficient; λ is the tip speed ratio; θ is the pitch angle; R ae is the radius of the wind turbine, in m; 2) Based on the wind speed-power relationship, the mathematical model of the wind turbine is simplified into a set of equations: Where: x is the state variable of the wind turbine; u pcc is the voltage at the PCC of the wind farm; y is the active power and reactive power output by the wind turbine; f1 and f2 are function operation symbols, which are determined by the control strategy; 3) Differentiating the equation in step 2) yields: x t =f3(x t-1 ,u pcc,t-1 ,u pcc,t ) In the formula, the subscript t represents the current time, t-1 represents the previous time, and f3 is the function operator.

3. The method for dynamic equivalent modeling of a wind farm based on a residual combination neural network according to claim 1 is characterized in that: In step S2, the feature memory layer is composed of LSTM units and leaky linear rectification LeakyReLU functions. The LSTM units extract key features that can characterize the dynamic characteristics of the wind farm. The LeakyReLU function performs nonlinear activation on the key features, and the expression is as follows: Where: λ is a constant between 0 and 1; The relationship between the input and output of the feature memory layer is as follows: h t =F(h t-1 ,u t ,u t-1 ,W,b f ) Where: F is the function operator, b f is the bias, u is the input, and W is the weight vector.

4. The method for dynamic equivalent modeling of a wind farm based on a residual combination neural network according to claim 1 is characterized in that: The information flow acceleration layer in step S3 is composed of two paths. The first path contains the GRU unit and the ReLU function, and the second path directly transmits the key features extracted by the feature memory layer. The two paths are added together at the end to form a residual structure. The residual structure algorithm formula is: G(x0)=g(x0)+x0 Where: x0 is the original input of the residual structure; g(x0) represents the output result of the residual block; G(x0) represents the output result of the entire residual structure.

5. The method for dynamic equivalent modeling of a wind farm based on residual combination neural network according to claim 1 is characterized in that: In step S4, the data relationship mapping layer is composed of a fully connected FC unit, a batch normalization BN unit and a ReLU function; the fully connected FC unit is used to linearly combine and transform the feature information transmitted by the information flow acceleration layer, and the formula is as follows: y=Wx+b In the formula, x is the input vector; y is the output vector; W is the weight matrix; b is the bias vector; The batch normalization BN unit normalizes the activation values ​​of each batch of data; The ReLU function introduces nonlinear factors and performs nonlinear activation on the normalized data; The data relationship mapping layer expression is: Where: represents the activation function; x i is the input quantity; b i is the offset.

6. The method for dynamic equivalent modeling of a wind farm based on residual combination neural network according to claim 1 is characterized in that: Step S5 uses a genetic algorithm GA to optimize the parameters of the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer, including: 1) Determine the optimized parameter as the number of neurons, and encode the number of neurons in each layer into population chromosomes. The structure of a single chromosome is as follows: D={a1,a2,a1',a'2,a'3} Where: D is the chromosome structure, a1 is the number of neurons in the LSTM unit, and a2 is the number of neurons in the GRU unit. The number of neurons, a1, a2, and a3 are the number of neurons in the first, second, and third fully connected layers respectively; 2) Initialize the population and randomly generate a set of initial solutions as the population; 3) Set the fitness function and select excellent chromosomes to enter the next generation according to the fitness function, and then enhance genetic diversity through crossover and mutation operations to generate new solutions; 4) Terminate GA according to the set termination condition and output the chromosome with the highest fitness as the optimal solution.

7. The method for dynamic equivalent modeling of a wind farm based on residual combination neural network according to claim 6 is characterized in that: The fitness function is: T=b1ξ+b2ξ' Where: T is the fitness function; ξ is the mean square error of the training set; ξ' is the mean square error of the test set; b1 and b2 are the weight coefficients of each indicator.

8. A wind farm dynamic equivalent modeling system based on residual combination neural network, characterized in that: include: Wind turbine mathematical model simplification module, used to simplify the mathematical model of wind turbines, with the wind speed of the wind tower and the voltage at the common coupling point as input features, and the active and reactive power of the wind farm as output features; The feature memory layer construction module is used to select LSTM units as the core to construct the feature memory layer and extract key features that can characterize the dynamic characteristics of the wind farm; An information flow acceleration layer construction module is used to construct an information flow acceleration layer with Res-GRU as the core, and optimize and refine the key features; A data relationship mapping layer construction module is used to construct a data relationship mapping layer based on FC-BN-Relu, and integrate and finely map the key features optimized and refined by the information flow acceleration layer; The parameter optimization module is used to optimize the parameters of the feature memory layer, the information flow acceleration layer, and the data relationship mapping layer using the genetic algorithm GA.

9. An electronic device, characterized in that: It comprises a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the equivalent modeling method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the equivalent modeling method described in any one of claims 1-7.

Images