Coupling system frequency response model modeling method based on KAN-Informer

Abstract

The invention discloses a coupling system frequency response model modeling method based on a KAN-Informer neural network, belongs to the field of modeling of power system frequency response models, and is mainly applied to establishment of a wind-solar-thermal storage coupling system frequency response model. Aiming at complex dynamic characteristics such as time varying, nonlinearity, uncertainty and intermittency of a coupling system, an Informer neural network is utilized to capture a dependency relationship of an input quantity and an output quantity of the system in time, then a KAN neural network is utilized to process a feature map sequence output by the Informer, and meanwhile, online fine tuning is performed on KAN network parameters, so that the accuracy of the system is improved. And the model accuracy and adaptability can be effectively improved.

Description

Technical Field

[0001] The present invention is applied to the modeling field of power system frequency response (SFR) model, and specifically relates to a modeling method of a frequency response model of a wind-solar-thermal-storage coupling system. Background Art

[0002] The high proportion of renewable energy grid connection has brought certain challenges to the active power and frequency response of the power system. The output of renewable energy has strong random volatility due to its own limitations. For example, the output of photovoltaic power generation is affected by irradiance, and the output of wind power generation is affected by wind speed and wind direction. In addition, renewable energy is connected to the grid through power electronic equipment, and its power and frequency are decoupled, which causes insufficient inertia level of the system and increases the possibility of frequency instability when the system is disturbed by severe active unbalance. In order to meet the frequency requirements of the power system with a high proportion of renewable energy, it is necessary to reasonably develop and utilize grid frequency regulation resources such as wind, light, fire, and storage from the power supply side, and analyze the dynamic characteristics of the frequency response of various frequency regulation resources to the grid. Therefore, a new modeling method for the frequency response model of the wind-solar-fired-storage coupled system is proposed, which is helpful to analyze the system response characteristics under large-scale grid connection of renewable energy and formulate control algorithms for various frequency regulation resources.

[0003] System frequency response has been widely used in power systems with wind, solar, thermal and storage coupling models to analyze the frequency dynamic characteristics of complex multi-energy complementary systems. In existing literature, in view of the randomness and volatility of renewable energy sources such as wind power and photovoltaics, a frequency response model with thermal power as the main and wind power and photovoltaics as the auxiliary is proposed to describe its dynamic characteristics in the power system. At the same time, in view of the rapid response characteristics of energy storage systems, some studies have constructed frequency response models integrating wind, solar, thermal and storage, emphasizing the key role of energy storage in the corresponding frequency disturbances. However, most of the existing SFR models use mechanism modeling methods, which are difficult to fully describe the nonlinearity and randomness in the wind, solar, thermal and storage coupling system. With the continuous increase in the proportion of renewable energy, the current models have shown certain limitations when facing the problems of scale expansion and complex operating conditions.

[0004] With the continuous development of computer technology and artificial intelligence modeling technology, data-driven modeling methods have provided a new direction for the modeling of frequency responses of coupled systems. As an improved neural network algorithm of Transformer, Informer has powerful nonlinear mapping capabilities and long time series modeling capabilities. It can greatly reduce the complexity of the model while ensuring the modeling accuracy, and can effectively overcome the difficulties faced by the coupled system modeling process. However, the output layer of Informer is composed of a single-layer fully connected layer, and its processing method for the feature map sequence output by the decoder is relatively simple. Therefore, the modeling accuracy of Informer is limited to a certain extent.

[0005] Kolmogorov–Arnold Network (KAN) is a new neural network architecture. Compared with traditional fully connected neural networks, it has stronger nonlinear mapping capabilities and is suitable for dealing with high-dimensional nonlinear problems in complex systems. Moreover, its network parameters are more efficient, and it can achieve higher modeling accuracy with fewer network parameters and has stronger continuous learning capabilities. Therefore, using KAN to replace the output layer of Informer can further improve the modeling performance of Informer.

[0006] Aiming at the modeling problem of the frequency response model of the wind-solar-thermal-storage coupling system, the present invention proposes a system frequency response model modeling method based on the KAN-Informer neural network. The input of the model is the change in the power generation of the wind farm, the power generation of the photovoltaic power station, the mechanical power of the thermal power unit, the energy storage power change and the load change. The output of the model is the frequency deviation of the coupling system. The frequency response model of the wind-solar-thermal-storage system based on the KAN-Informer neural network is trained according to the input and output data of the coupling system in typical scenarios. Finally, the trained model is fine-tuned online and put into use. Summary of the invention

[0007] The present invention proposes a system frequency response model based on the KAN-Informer neural network for the complex dynamic characteristics of the wind-solar-thermal-storage coupling system, such as time-varying, nonlinear, uncertain, and intermittent. The Informer neural network is used to capture the temporal dependency between the system input and output, solving the problem that the traditional system frequency response mechanism model cannot accurately describe the nonlinearity and uncertainty of the system frequency response. The KAN neural network is then used to process the feature graph sequence output by the Informer to improve the accuracy of the model output. At the same time, based on KAN's excellent continuous learning ability, the adaptability of the model can be improved by online fine-tuning of the KAN network parameters.

[0008] The technical solution adopted by the present invention to achieve the above-mentioned purpose is to establish a system frequency response model based on the KAN-Informer neural network. The main steps are as follows:

[0009] Step (1): Collect historical input and output data of the wind-solar-thermal-storage coupled system when it is running in a specific scenario, and use it as training data after preprocessing;

[0010] Step (2): Establish a frequency response model of the wind-solar-thermal-storage coupled system based on the KAN-Informer neural network and initialize the weight parameters of the KAN-Informer neural network;

[0011] Step (3): In the offline training phase, the training data set is input into the KAN-Informer neural network for training, and the model with the best error performance index is selected as the frequency response model of the wind-solar-thermal-storage coupled system;

[0012] Step (4): In the online fine-tuning stage, the wind-solar-thermal-storage coupled frequency response model based on the KAN-Informer neural network that has been trained offline is put into online use to obtain the frequency deviation of the wind-solar-thermal-storage coupled system operating in the current scenario. At the same time, the weight parameters of the KAN-Informer neural network are updated in real time according to the range of the difference between the frequency deviation output by the model and the actual system frequency deviation.

[0013] Furthermore, in step (1), the collected coupling system data includes the wind farm power variation ΔP W , Photovoltaic power station power change ΔP PV , Mechanical power variation of thermal power plant ΔP M , Energy storage system power change ΔP ES , load change ΔP L and the coupled system frequency deviation Δf.

[0014] Furthermore, in step (2), the KAN-Informer neural network used to establish the frequency response model of the wind-solar-thermal-storage coupled system is improved by replacing the output layer of the original Informer with KAN.

[0015] Furthermore, in step (3), the output of the KAN-Informer neural network model is the frequency deviation of the coupled system at the current moment, and the input includes the time series of wind farm power change, photovoltaic power station power change, thermal power plant mechanical power change, energy storage system power change and load change, as well as the historical value sequence of the actual coupled system frequency deviation. The internal structure of the KAN-Informer neural network includes basic modules such as the KAN layer, embedding layer, multi-head attention module and attention distillation layer, and the network training is carried out using the back propagation through time (BPTT) algorithm and the recursive least squares (RLS) algorithm.

[0016] Furthermore, in step (4), the frequency response model of the wind-solar-thermal-storage coupled system put into online use is a KAN-Informer neural network model trained offline. During online use, it is necessary to determine whether the model needs to be fine-tuned in real time based on the difference between the model output and the actual system output.

[0017] Beneficial effects of the present invention: A system frequency response modeling method based on a KAN-Informer neural network is proposed. In the model, the Informer is responsible for obtaining the temporal dependency between the input sequence and the output sequence to obtain the corresponding feature graph sequence. The KAN layer is responsible for processing the feature graph sequence and calculating the output coupling system frequency deviation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the frequency response modeling method based on KAN-Informer neural network proposed in the present invention;

[0019] Figure 2 The structure diagram of the KAN-Informer neural network proposed in the present invention;

[0020] Figure 3 An overview of the frequency response modeling method for the wind-solar-thermal-storage coupled system based on KAN-Informer proposed in this invention. DETAILED DESCRIPTION

[0021] The following is a detailed description of the implementation of the present invention in conjunction with the accompanying drawings, and provides specific operation methods and implementation steps:

[0022] A system frequency response modeling method based on KAN-Informer neural network mainly includes the following steps:

[0023] Step (1): The steps for obtaining the training data of the KAN-Informer neural network are:

[0024] Step (1.1): Determine the operation scenario of the coupling system described by wind speed, solar irradiance, thermal power unit output, energy storage system output and load, and construct the operation scenario of the coupling system according to the wind speed, solar irradiance, thermal power unit output, energy storage system output and load under the typical scenario;

[0025] Step (1.2): Obtain the input and output data of the wind-solar-thermal-storage coupling system in a typical scenario. The input data includes the power change of the wind farm, the power change of the photovoltaic power station, the mechanical power change of the thermal power plant, the power change of the energy storage system, and the load change; the output data is the frequency deviation of the coupling system;

[0026] Step (1.3): Preprocess the collected data: clean invalid data, perform linear interpolation on missing data, and perform smoothing.

[0027] Step (2): The model structure of the KAN-Informer neural network used to establish the frequency response model of the wind-solar-thermal-storage coupled system is as follows: Figure 2 As shown in Figure 1, the KAN-Informer neural network is the main part of the entire model, which is improved by replacing the output layer of the original Informer with KAN. The input based on the KAN-Informer neural network includes the encoder input and the decoder side input The output is Among them, the subscript [kn i ,k-1] means from the previous n i Historical values ​​from the moment to the previous moment; is the system frequency deviation of the model output at time k, Δf is the frequency deviation of the actual coupled system; ΔP W is the power change of the wind farm; ΔP PV is the power change of the photovoltaic power station; ΔP M is the change in mechanical power of the thermal power plant; ΔP ES is the power change of the energy storage system; ΔP L is the load change; n is the sequence length, n=n1+n2+n3+n4; is an all-zero vector of length n-n5. Initialization includes initializing the hyperparameters of the KAN-Informer neural network, such as the number of model layers, hidden layer dimensions, batch size, number of attention heads, learning rate, and network weights.

[0028] Step (3): In the offline stage, use sufficient training data to pre-train the KAN-Informer neural network to obtain the frequency response model of the wind-solar-thermal-storage coupled system. The offline stage steps of the system frequency response model based on the KAN-Informer neural network are:

[0029] Step (3.1): In the KAN-Informer neural network, the Informer extracts the dependency between the input data and the output data, and obtains the corresponding feature sequence, and then uses the KAN layer to process the feature sequence to obtain the final system frequency deviation. The internal structure of the KAN-Informer neural network includes the KAN layer, the embedding layer, the multi-head attention module, and the attention distillation layer. Furthermore, the forward calculation process of each module in the KAN-Informer is as follows:

[0030] Embedding layer forward process: x em =Conv1d(x)+Pos(m,j) t Where x is the input sequence, x em is the generated embedding vector; P(m,j) t is the position encoding function, m is the position of the variable in the sequence at time t, d model The dimension of the embedding vector, j is the number, j = 1, 2, ..., d model ; Conv1d(·) is a one-dimensional convolution operation.

[0031] Probabilistic sparse attention mechanism forward process: Among them, Q, K, and V represent query vector, key vector, and value vector respectively; d k is the dimension of the vector; It is obtained by sparsely selecting multiple q in Q and substituting them into Perform sparsity evaluation and select the first u q to form a sparse matrix q i , k i are the i-th row in the query vector and key vector respectively; L K is the length of the key vector; Softmax(·) is the softmax activation function.

[0032] Multi-head attention module forward process: [x q ,x k ,x v ] AB =LayerNorm(MSA(x q ,x k ,x v )+x q ) Among them, MSA(·) is the multi-head attention mechanism, MSA(x q ,x k ,xv )=(h1,…,h n )W o ,h i is the i-th attention head, Here W i Q , W i K , W i V , W o is the corresponding linear projection matrix; is a probabilistic sparse attention mechanism; [·] AB Represents a multi-head attention module.

[0033] Self-attention distillation module forward process: x i+1 =Maxpool(ELU(Conv1d([x i ] AB ))) Among them, x i is the feature graph sequence before distillation, x i+1 is the feature map sequence after distillation; Maxpool(·) is the maximum pooling operation with a step size of 2; ELU(·) is the activation function, Here u is the input of the ELU activation function.

[0034] KAN layer forward process: Among them, n l is the number of nodes in the lth layer; x l+1,j is the output value of the jth node in the l+1th layer; φ l,j,i () is the activation function of the jth edge branched from the i-th node in the l-th layer, is the output value of the corresponding activation function, where the activation function is a univariate function. w b 、c i is the learnable coefficient, g is the grid size, B i (x) is the basis function.

[0035] Step (3.2): Input the training data into the KAN-Informer neural network for training. The trainable parameters in the KAN-Informer neural network mainly exist in the KAN layer and the multi-head attention module. The update algorithm of the weight matrix of an attention head in the multi-head attention module adopts the gradient descent method, which is as follows: in, are the query vector projection matrix, key vector projection matrix, and value vector projection matrix corresponding to the i-th head at time k respectively; η is the learning rate, and J is the performance indicator.

[0036] The update method of the learnable coefficient of the activation function in KAN uses the gradient descent method and the recursive least squares method. Then b , the update formula of y is:

[0037] Let β k =[B0(x k )…B g (x k )] Τ , Yes k =(β k ) Τ c k , then c k The update formula is as follows: Among them, α is the learning rate; ε, P, K are intermediate quantities, and P is initialized k =β k (β k ) Τ .

[0038] Step (3.3): Use a variety of error evaluation indicators to evaluate the frequency response model of the wind-solar-thermal-storage coupled system based on the KAN-Informer neural network. The model error evaluation criteria mainly use the root mean square error (RMSE), mean absolute error (MAE), standard deviation (SD) and determination coefficient (R 2 ). Let the actual frequency deviation of the coupled system at time K be Δf(k), and the output of the coupled system frequency response model based on the KAN-Informer neural network be Δf(k).

[0039] RMSE is the expected value of the square of the difference between the model frequency deviation and the actual system frequency deviation. The root mean square error is often used to evaluate the degree of change in data. Its expression is:

[0040] MAE refers to the average absolute error. The average absolute error can better reflect the actual situation of the prediction value error. Its expression is:

[0041] SD is the standard deviation. The degree of dispersion of a data set can be reflected by the standard deviation. Its expression is:

[0042] R 2 It characterizes the quality of a fit by the change of data. Its numerical value reflects the effect of the model on the data fit. Its expression is as follows in, is the average value of Δf(k).

[0043] By RMSE, SD, MAE and R 2 Four error evaluation criteria can be used to evaluate the performance of the model based on the KAN-Informer neural network.

[0044] Step (3.3): Perform multiple rounds of training on the frequency response model of the wind-solar-thermal-storage coupling system based on the KAN-Informer neural network, select the current optimal model as the final frequency response model of the wind-solar-thermal-storage coupling system according to the evaluation indicators of step (3.4), and put it into online use.

[0045] Step (4): In the online fine-tuning stage, the offline trained frequency response model of the wind-solar-thermal-storage coupled system is put into online use to obtain the frequency deviation of the coupled system running in the current scenario. At the same time, the frequency deviation output by the model is compared with the actual system frequency deviation, and the weight parameters of the KAN layer in the KAN-Informer neural network are updated in real time according to the range of the difference between the two to ensure the accuracy of the model. The specific steps are:

[0046] Step (4.1): Input the real-time input and output data of the coupled system in the current scenario into the coupled system frequency response model based on the KAN-Informer neural network, and calculate the current performance index J(k).

[0047] Step (4.2): Output the model in real time Compare with the actual system output Δf(k) to get the error E(k) at the current moment. The range of the error E(k) is divided into three intervals: when |E(k)|<E1, the error is acceptable, the model can continue to be put into use, and the model network parameters do not need to be updated; E1≤|E(k)|<E2, the error is still acceptable, the model can continue to be used, but the network parameters of the model need to be fine-tuned according to step (4.3); E2≤|E(k)|, the error is too large, and the input and output data of the current scene need to be saved and the KAN-Informer neural network needs to be retrained.

[0048] Step (4.3): Use the input and output history data saved in step (4.1) to fine-tune the KAN-Informer neural network. During the fine-tuning process, it is necessary to freeze the network parameters of the Informer module in the model, and fine-tune the network parameters of the KAN module according to the network parameter update method of step (3.2) to further improve the accuracy of the frequency response model of the coupled system.

Claims

1. A method for modeling a frequency response model of a coupled system based on KAN-Informer, characterized in that: The following steps are involved: Step (1): Collect historical input and output data of the wind-solar-thermal-storage coupled system when it is running in a specific scenario, and use them as training data after preprocessing; Step (2): Establish a frequency response model of the wind-solar-thermal-storage coupled system based on the KAN-Informer neural network and initialize the weight parameters of the KAN-Informer neural network; Step (3): In the offline training phase, the training data set is input into the KAN-Informer neural network for training, and the model with the best error performance index is selected as the frequency response model of the wind-solar-thermal-storage coupled system; Step (4): In the online fine-tuning stage, the wind-solar-thermal-storage coupled frequency response model based on KAN-Informer that has been trained offline is put into online use to obtain the frequency deviation of the wind-solar-thermal-storage coupled system in the current scenario. At the same time, the weight parameters of the KAN-Informer neural network are updated in real time according to the range of the difference between the frequency deviation output by the model and the actual system frequency deviation.

2. The system frequency response modeling method based on KAN-Informer neural network according to claim 1 is characterized in that: In the step (1), the operation scenario of the wind-solar-thermal-storage coupling system is described by wind speed, solar irradiance, thermal power unit output, energy storage system state and load change; the typical scenario of the wind-solar-thermal-storage coupling system is set according to the typical wind speed, typical solar irradiance, typical thermal power unit output, typical energy storage system state and typical load; the input and output data of the wind-solar-thermal-storage coupling system are obtained under the set typical scenario, the input data including the power change of the wind farm, the power change of the photovoltaic power station, the mechanical power change of the thermal power plant, the power change of the energy storage system and the load change in the coupling system, and the output data is the frequency deviation of the coupling system; the collected data is preprocessed: including cleaning invalid data, interpolating missing data, smoothing the original data and standardizing the data.

3. The system frequency response modeling method based on KAN-Informer neural network according to claim 1 is characterized in that: In the step (2), the KAN-Informer neural network used to establish the frequency response model of the wind-solar-thermal-storage coupled system is improved by replacing the output layer of the original Informer with KAN, and its input includes the time series of the power change of the wind farm, the power change of the photovoltaic power station, the mechanical power change of the thermal power plant, the power change of the energy storage system and the load change, as well as the historical value sequence of the actual coupled system frequency deviation, and its output is the coupled system frequency deviation.

4. The system frequency response modeling method based on KAN-Informer neural network according to claim 1 is characterized in that: In the step (3), during offline training, the KAN-Informer neural network model is trained for multiple rounds using the historical input and output data of the wind-solar-thermal-storage coupled system. The results of the multiple rounds of training are evaluated using four error evaluation criteria: root mean square error, mean absolute error, standard deviation, and coefficient of determination. The KAN-Informer neural network model with the best current performance index is selected as the frequency response model of the wind-solar-thermal-storage coupled system.

5. The system frequency response modeling method based on KAN-Informer neural network according to claim 1 is characterized in that: In step (4), the wind farm power change ΔP is obtained using Informer W , Photovoltaic power station power change ΔP PV , Mechanical power variation of thermal power plant ΔP M , Energy storage system power change ΔP ES , load change ΔP L The feature graph vector of the dependence relationship between the frequency deviation Δf and the system frequency is finally processed by KAN to obtain the frequency deviation of the current system.

6. The system frequency response modeling method based on KAN-Informer neural network according to claim 1 is characterized in that: The optimal weight parameters obtained from offline training are used to update the frequency response model of the coupled system based on the KAN-Informer neural network, and the frequency deviation of the wind-solar-thermal-storage coupled system operating in the current scenario is calculated. If the calculated frequency deviation is significantly different from the actual frequency deviation, it is necessary to fine-tune the KAN layer network parameters in the KAN-Informer neural network online according to the current performance indicators to improve the adaptability of the model; if the frequency deviation output by the model deviates seriously from the actual frequency deviation, it is necessary to re-train the KAN-Informer neural network offline according to the input and output data of the current scenario to further improve the accuracy of the frequency response model of the coupled system.

Images