Knowledge and data fusion driven photovoltaic cell system modeling identification method

Abstract

The invention discloses a knowledge and data fusion driven photovoltaic cell system modeling identification method, which comprises the following steps of: establishing a knowledge driven model, establishing a data driven model by utilizing a multivariable Hammerstein model, describing a static nonlinear module of the Hammerstein model by utilizing a bidirectional gating circulation unit neural network, and describing a dynamic linear module by utilizing an autoregressive moving average model. Constructing a photovoltaic cell system model based on knowledge and data fusion; decoupling a static nonlinear module and a dynamic linear module of the Hammerstein model by using the covariance function characteristic of the Gaussian signal; using a least square method based on a covariance function to identify parameters of the autoregression moving average model, using a dung beetle optimization algorithm to adaptively solve hyper-parameters of the bidirectional gating circulation unit neural network, and using an adaptive momentum estimation algorithm to update the weight and bias of the bidirectional gating circulation unit neural network; and taking a result obtained by the data-driven model as input of the knowledge-driven model, and identifying parameters of the knowledge-driven model by using a generalized least square method.

Description

Technical Field

[0001] The invention relates to a photovoltaic battery system modeling and identification method driven by knowledge and data fusion. Background Art

[0002] As global energy demand continues to grow, photovoltaic cells, as a green and renewable energy technology, are increasingly valued and promoted by various countries. In recent years, the conversion efficiency of photovoltaic cells has continued to increase, the cost has gradually decreased, and the application field has gradually expanded. It has not only been widely used in industrial and commercial fields, but also plays an important role in smart grids and distributed energy systems. Researching efficient and accurate photovoltaic cell models is of great significance for achieving optimal control, predicting performance and improving power generation efficiency. Therefore, establishing an extensive mathematical model of photovoltaic cell systems is the basis for solving optimization problems and implementing effective predictions, and has become a current research hotspot. Currently, the commonly used photovoltaic cell system modeling methods include physical modeling methods and data-driven modeling methods.

[0003] Physical modeling methods mainly include equivalent circuit models, empirical models, and models based on physical mechanisms. The equivalent circuit model represents the performance of photovoltaic cells by simulating circuit parameters, but it is difficult to accurately reflect the actual performance in complex environments. Although the empirical model is simple to calculate, its parameters depend on experimental data and the model generalization ability is limited. The physical mechanism model can better describe the dynamic behavior of photovoltaic cells, but it has many parameters, complex calculations, and is difficult to operate efficiently in real-time applications. In addition, photovoltaic cell systems usually have multivariable coupling characteristics, and the system output will be affected by the combined influence of multiple input variables. It is difficult to obtain an accurate mathematical model using single-input single-output modeling technology.

[0004] In recent years, with the continuous development of artificial intelligence technology, deep learning models have powerful learning and reasoning capabilities, and can solve more complex problems and tasks. Some data-driven modeling methods based on deep learning models are widely used, such as using convolutional neural networks to predict battery voltage probabilistically, or using radial basis function neural networks and convolutional neural networks to model battery systems, or using long short-term memory network models to predict voltage. These methods establish a global nonlinear mapping function based on the input and output of the system, which can better describe the nonlinear characteristics, but ignore the dynamic characteristics of the photovoltaic battery system, resulting in limited accuracy and adaptability of the model. Therefore, the existing modeling methods still have obvious deficiencies in the accuracy of model description. Summary of the invention

[0005] The present invention aims to solve the problems existing in the above-mentioned prior art and provides a photovoltaic battery system modeling and identification method driven by knowledge and data fusion.

[0006] The technical solutions adopted in the present invention are:

[0007] A photovoltaic battery system modeling and identification method driven by knowledge and data fusion includes the following steps:

[0008] S1: Analyze the mechanism characteristics of the photovoltaic battery system, establish a knowledge-driven model, and use the multivariable Hammerstein model to establish a data-driven model. In the Hammerstein model, a bidirectional gated recurrent unit neural network is used to describe the static nonlinear module of the Hammerstein model, and an autoregressive sliding average model is used to describe the dynamic linear module of the Hammerstein model, so as to construct a photovoltaic battery system model that integrates knowledge and data;

[0009] S2: Introduce Gaussian signal and use the covariance function characteristics of Gaussian signal to decouple the static nonlinear module and dynamic linear module of Hammerstein model;

[0010] S3: The parameters of the autoregressive moving average model are identified using the least squares method based on the covariance function, the hyperparameters of the bidirectional gated recurrent unit neural network are adaptively solved using the dung beetle optimization algorithm, and the weights and biases of the bidirectional gated recurrent unit neural network are updated using the adaptive momentum estimation algorithm;

[0011] S4: The results obtained by the data-driven model are used as the input of the knowledge-driven model, and the generalized least squares method is used to identify the parameters of the knowledge-driven model, so as to obtain a photovoltaic battery system model that integrates knowledge and data, and further obtain the predicted values ​​of voltage and current.

[0012] Furthermore, in S1: according to the external output characteristics of the photovoltaic battery system, a knowledge-driven model is established. For the photovoltaic battery system,

[0013] Under any environmental conditions, the maximum power point voltage U m , open circuit voltage short U oc , Maximum power point current I m , short circuit current I se It will change according to a certain rule. By introducing the corresponding compensation coefficient, the new light intensity and the new battery temperature are calculated according to the reference light intensity and the reference battery temperature. m , U o ' c , I′ m , I s ' e , the dynamic characteristics of the photovoltaic battery system, that is, the knowledge-driven model is expressed as:

[0014] U′ m =U m [1-c1(TT ref)]ln[e+b1(SS ref )]

[0015] U o ' c =U oc [1-c1(TT ref )]ln[e+b1(SS ref )]

[0016]

[0017] Where: T ref is the reference battery temperature, S ref is the standard reference light intensity, S represents the actual light intensity, T represents the actual battery temperature, and the compensation coefficients a1, b1 and c1 are the parameters to be identified.

[0018] Furthermore, a multivariable Hammerstein model is used to establish a data-driven model for the photovoltaic battery system. The Hammerstein model is:

[0019] v(t)=f(u(t))

[0020]

[0021] In the formula: u(t)=[u1(t),u2(t),…,u m (t)] T ,y(t)=[y1(t),y2(t),…,y n (t)] T ,v(t)=[v1(t),v2(t),…,v m (t)] T , q -1 represents the unit shift operator, I represents the unit matrix of appropriate dimension, n A and n B is a known order, represents the autoregressive moving average model; t represents the time.

[0022] Furthermore, the bidirectional gated recurrent unit neural network includes an input layer, two hidden layers and an output layer, and a single neuron in the network is represented as:

[0023] r t =σ(W r ⊙[h t-1 ,u t ])

[0024] z t =σ(W z ⊙[h t-1,u t ])

[0025]

[0026] Where: W r ,W z ,W h They represent the weight matrices of the reset gate, update gate, and candidate hidden state, σ represents the sigmoid activation function, and h t-1 is the hidden state of the previous moment, ⊙ represents the point-by-point multiplication of two vectors;

[0027] The hidden layer, i.e., the bidirectional gated recurrent unit, includes two-directional gated recurrent unit neural network layers. The input data is calculated in opposite directions in the two layers of gated recurrent unit neural network layers. For an input data U=u1,u2,…,u T , the forward computation of the bidirectional gated recurrent unit is expressed as:

[0028]

[0029] Where: and are the hidden states from left to right and from right to left, respectively, u t is the data at the tth moment of the input data; the hidden states in the two directions are fused to obtain the final hidden state:

[0030]

[0031] Pass the hidden state to one of the hidden layers and get the output v(t):

[0032]

[0033] Where: k (k = 1, 2, ..., n) represents the kth output of the static nonlinear module, W kγ represents the output weight from the γth neural unit in the corresponding hidden layer receiving the hidden state to the kth neural unit in the output layer, represents the output of the γth neural unit in the corresponding hidden layer, c=1,2,…,γ is the number of bidirectional gated recurrent units in the corresponding hidden layer.

[0034] Furthermore, in S2:

[0035] For a static nonlinear module, if the input is a zero-mean Gaussian signal, the product of the autocovariance function of the Gaussian input signal u(t) and the constant matrix is ​​equal to the covariance function of the intermediate variable v(t) and the input u(t), that is:

[0036] R vu (τ) = αRu (τ)

[0037] Where: α is the constant matrix, τ is the time constant,

[0038] Cross-covariance function matrix

[0039] Autocovariance function matrix

[0040] Using αR u (τ) instead of R vu (τ), realizes the decoupling of the static nonlinear module and the linear module of the Hammerstein model.

[0041] Furthermore, in S3,

[0042] The dung beetle optimization algorithm is used to adaptively solve the hyperparameters of the bidirectional gated recurrent unit neural network, including the number of hidden layer neurons, learning rate and number of iterations in the bidirectional gated recurrent unit neural network. These hyperparameters constitute the vector X. The specific solution process is expressed as:

[0043] (1) When there are no obstacles, in order to simulate the rolling behavior of the dung beetle, it moves in the search space along a given direction and navigates by the sun. Assuming that the intensity of the light source affects the path, the position of the dung beetle is updated during the rolling process as follows:

[0044] X i (t-1) = X i (t)+a×l×X i (t-1)+b×Δx

[0045] Δx=|X i (t)-X w |

[0046] Where: t represents the current iteration number, X i (t) represents the position information of the i-th dung beetle at the t-th iteration, l represents the deflection coefficient, b∈(0,1), a is the coefficient, which takes the value of -1 or 1, and X w represents the global worst position, and Δx is used to simulate the change of light intensity;

[0047] (2) When a dung beetle encounters an obstacle, it will reposition itself by “dancing” to obtain a new route. The dancing behavior obtains a new rolling direction through the tangent function, which is defined in the interval [0,π]. The new position is updated as:

[0048] X i (t-1) = X i (t)+tan(θ)|X i (t)-X i (t-1)|

[0049] Where: θ = [0, π] is the deflection angle, |X i (t)-X i (t-1)| is the difference between the position of the i-th dung beetle at the t-th iteration and the position at the t-1-th iteration. If θ is 0, π / 2 or π, the position of the dung beetle will not be updated.

[0050] Furthermore, in S3, the adaptive momentum estimation algorithm is expressed as:

[0051]

[0052] Where: and denote the first-order moment estimate and the second-order moment estimate respectively, β1 and β2 are exponential decay rates, α1 is the learning rate, ε1 is the numerical stability constant, * denotes one of z, r, h, W z ,W r ,W h is the weight matrix of the bidirectional gated recurrent unit.

[0053] Furthermore, in S4, the predicted output of the Hammerstein model is used as the input of the knowledge-driven model, and the generalized least squares method is used to identify the unknown parameters of the knowledge-driven model. The output of the knowledge-driven model is expressed as:

[0054]

[0055] Where: represents the output of the Hammerstein model, y k (t) represents the output of the knowledge-driven model, θ k is the knowledge-driven model parameter vector to be identified.

[0056] The present invention has the following beneficial effects:

[0057] 1) The method of the present invention integrates the physical characteristics and data features of the photovoltaic battery system, constructs a mathematical model driven by knowledge and data fusion, and improves the accuracy and generalization ability of the model.

[0058] 2) The method of the present invention utilizes the covariance function characteristics of Gaussian signals in a static nonlinear system to achieve the decoupling of the series modules of the Hammerstein model, thereby solving the problem that the intermediate variable information of the Hammerstein model is unmeasurable.

[0059] 3) The method of the present invention uses an adaptive dung beetle optimization algorithm to optimize the number of hidden layer neurons, learning rate and number of iterations of a bidirectional gated recurrent unit neural network, thereby improving the search efficiency, solution accuracy and convergence speed, thereby achieving better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a principle block diagram of the present invention.

[0061] Figure 2 The structure of a single neuron in a bidirectional gated recurrent unit neural network.

[0062] Figure 3 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0063] The present invention will be further described below in conjunction with the accompanying drawings.

[0064] like Figures 1 to 3 The present invention provides a photovoltaic battery system modeling and identification method driven by knowledge and data fusion, and the steps are as follows:

[0065] S1: Analyze the mechanism characteristics of the photovoltaic battery system, establish a knowledge-driven model, and use the multivariable Hammerstein model to establish a data-driven model. In the Hammerstein model, a bidirectional gated recurrent unit neural network is used to describe the static nonlinear module of the Hammerstein model, and an autoregressive sliding average model is used to describe the dynamic linear module of the Hammerstein model, so as to construct a photovoltaic battery system model that integrates knowledge and data;

[0066] S2: Introduce Gaussian signal and use the covariance function characteristics of Gaussian signal to decouple the static nonlinear module and dynamic linear module of Hammerstein model;

[0067] S3: The parameters of the autoregressive moving average model are identified using the least squares method based on the covariance function, the hyperparameters of the bidirectional gated recurrent unit neural network are adaptively solved using the dung beetle optimization algorithm, and the weights and biases of the bidirectional gated recurrent unit neural network are updated using the adaptive momentum estimation algorithm;

[0068] S4: The results obtained by the data-driven model are used as the input of the knowledge-driven model, and the generalized least squares method is used to identify the parameters of the knowledge-driven model, so as to obtain a photovoltaic battery system model that integrates knowledge and data, and further obtain the predicted values ​​of voltage and current.

[0069] Each step is described in detail below.

[0070] In S1: Based on the external output characteristics of the photovoltaic battery system, a knowledge-driven model is established.

[0071] For photovoltaic battery systems, under any environmental conditions, the maximum power point voltage U m , open circuit voltage short U oc , Maximum power point current Im , short circuit current I se It will change according to a certain rule. By introducing the corresponding compensation coefficient, the new light intensity and the new battery temperature are calculated according to the reference light intensity and the reference battery temperature. m , U o ' c , I′ m , I s ' e , the dynamic characteristics of the photovoltaic battery system, that is, the knowledge-driven model is expressed as:

[0072] U′ m =U m [1-c1(TT ref )]ln[e+b1(SS ref )]

[0073] U o ' c =U oc [1-c1(TT ref )]ln[e+b1(SS ref )]

[0074]

[0075] Where: T ref =25℃ is the reference battery temperature, S ref =1000W / m2 is the standard reference light intensity, S represents the actual light intensity, T represents the actual battery temperature, and the compensation coefficients a1, b1 and c1 are unknown parameters.

[0076] The light intensity S and battery temperature T of the photovoltaic battery system are used as the input of the knowledge-driven model, that is, u k (t)=[S(t),T(t)] T , the new maximum power point voltage U′ is obtained through the dynamic characteristics of the system m , open circuit voltage short U o ' c , maximum power point current I′ m , short circuit current I s ' e , then the output of the knowledge-driven model is:

[0077] y k (t) = [U′ m (t),U o ' c (t),I′ m (t),I s ' e (t)] T .

[0078] A data-driven model of a photovoltaic battery system is established using a two-input and four-output Hammerstein model, wherein the Hammerstein model is:

[0079] v(t)=f(u(t))

[0080]

[0081] In the formula: u(t)=u k (t)=[S(t),T(t)] T , y(t) = y d (t) = [U m (t),U oc (t),I m (t),I se (t)] T represents the output of the data-driven model, q -1 represents the unit shift operator, I represents the unit matrix of appropriate dimension, n A and n B is a known order, represents the autoregressive moving average model.

[0082] The static nonlinear module of the Hammerstein model is described using a bidirectional gated recurrent unit neural network. The bidirectional gated recurrent unit neural network includes an input layer, two hidden layers, and an output layer. In this network, a single neuron is represented as:

[0083] r t =σ(W r ⊙[h t-1 ,u t ])

[0084] z t =σ(W z ⊙[h t-1 ,u t ])

[0085]

[0086] Where: W r ,W z ,W h They represent the weight matrices of the reset gate, update gate, and candidate hidden state, σ represents the sigmoid activation function, and h t-1 is the hidden state of the previous moment, and ⊙ represents the point-by-point multiplication of two vectors.

[0087] The hidden layer, i.e., the bidirectional gated recurrent unit, includes two-directional gated recurrent unit (GRU) neural network layers. The input data is calculated in opposite directions in the two-layer gated recurrent unit neural network layers. For an input data U=u1,u2,…,u T , the forward computation of the bidirectional gated recurrent unit is expressed as:

[0088]

[0089] Where: and are the hidden states from left to right and from right to left, respectively, u t is the data at the tth moment of the input data; on this basis, the hidden states in the two directions are fused to obtain the final hidden state:

[0090]

[0091] Pass the hidden state to one of the hidden layers and get the output v(t):

[0092]

[0093] Where: k (k = 1, 2, ..., n) represents the kth output of the static nonlinearity, W kγ represents the output weight from the γth neural unit in the corresponding hidden layer receiving the hidden state to the kth neural unit in the output layer, represents the output of the γth neural unit in the hidden layer receiving the hidden state, and c=1,2,…,γ is the number of bidirectional gated recurrent units in the hidden layer receiving the hidden state.

[0094] In S2: the covariance function characteristics of the Gaussian signal are used to decouple the static nonlinear module and the dynamic linear module of the Hammerstein model, which is specifically expressed as:

[0095] For a static nonlinear module, if the input is a zero-mean Gaussian signal, then the Gaussian input signal u G The product of the autocovariance function of (t) and the constant matrix is ​​equal to the intermediate variable v(t) and the input u G The covariance function of (t) is:

[0096]

[0097] Where: α is the constant matrix, τ is the time constant,

[0098] Cross-covariance function matrix

[0099] Autocovariance function matrix

[0100] In the present invention, u G (t) consists of two independent sets of zero-mean Gaussian signals, and the two sets of signals have the same dimension, that is, u G (t)=[u G1 (t),u G2 (t)] T , after passing through the Hammerstein model, the output is y G (t) = [y G1 (t),y G2 (t)] T . According to the covariance function characteristics It can be seen that the problem of the unmeasurable intermediate variable information v(t) of the Hammerstein model has been solved, that is, using Instead of the unknown covariance matrix This achieves the decoupling of the static nonlinear module and the linear module.

[0101] In S3: the least square method based on the covariance function is used to identify the parameters of the autoregressive moving average model, that is, to identify the unknown parameters in A(q) and B(q). The specific calculation is as follows:

[0102] θ1=Rψ1 T (ψ1ψ1 T ) -1

[0103] Where: is a constant,

[0104]

[0105] The dung beetle optimization algorithm is used to adaptively solve the hyperparameters of the bidirectional gated recurrent unit neural network, including the number of neurons in the two hidden layers, i.e., h1 and h2, the learning rate β, and the number of iterations ε. These hyperparameters constitute the vector X = [h1, h2, β, ε]. The specific solution process is expressed as:

[0106] (1) When there are no obstacles, in order to simulate the rolling behavior of the dung beetle, it moves in the search space along a given direction and navigates by the sun. Assuming that the intensity of the light source affects the path, the position of the dung beetle is updated during the rolling process as follows:

[0107] X i (t-1) = X i (t)+a×l×X i (t-1)+b×Δx

[0108] Δx=|X i (t)-X w |

[0109] Where: t represents the current iteration number, X i (t) represents the position information of the i-th dung beetle at the t-th iteration, l represents the deflection coefficient, b∈(0,1), a is the coefficient, which takes the value of -1 or 1, and X w represents the global worst position, and Δx is used to simulate the change of light intensity;

[0110] (2) When the dung beetle encounters an obstacle, it will reposition itself by "dancing" to obtain a new route. The dancing behavior obtains a new rolling direction through the tangent function, which is defined in the interval [0,π]. The new position is updated as:

[0111] X i (t-1) = X i (t)+tan(θ)|X i (t)-X i (t-1)|

[0112] Where: θ = [0, π] is the deflection angle, |X i (t)-X i (t-1)| is the difference between the position of the i-th dung beetle at the t-th iteration and the position at the t-1-th iteration. If θ is 0, π / 2 or π, the position of the dung beetle will not be updated.

[0113] Combining the respective advantages of the adaptive stochastic gradient descent algorithm and the root mean square propagation algorithm, an adaptive momentum estimation algorithm is constructed to adaptively update the weights and biases of the bidirectional gated recurrent unit neural network to achieve optimal parameter search. The specific update process is as follows:

[0114] The bidirectional gated recurrent unit updates the weight matrix W of each gated recurrent unit by back-propagating the error z ,W r ,W h The input of the bidirectional gated recurrent unit neural network at time t includes two different information flows, namely h t-1 and u t , where u t In the bidirectional gated recurrent unit neural network, the input u k (t)=[S(t),T(t)] T , their weight matrices are different. Therefore, the weight matrix is ​​defined as W z =[W zh ,W zx ],W r =[W rh ,W rx ] and W h =[W hh ,W hx ].

[0115] First, update the weights and biases from the output layer to the hidden layer. The error between the actual nonlinear output and the bidirectional gated recurrent unit neural network output at time t is calculated as where e(t) = [e1(t),…,e n (t)] T For the k-th output of the bidirectional gated recurrent unit neural network, the mean square criterion function is defined as in Then calculate the error at time t

[0116] Further calculate the output weight gradient Output bias gradient The loss function for the weights and biases from the hidden layer to the output layer is given by and Get, where η is the learning rate. Then the output weight W kγ and output bias b k It can be updated to W kγ (t) = W kγ (t-1)+ΔW kγ (t) and b k (t) = b k (t-1)+Δb k (t).

[0117] Next, we will discuss the weight update of the hidden layer of the bidirectional gated recurrent unit neural network. For convenience, we take the weight update of one of the hidden layers of the bidirectional recurrent neural network as an example: First, define the criterion function Where: represents the output of the γth bidirectional gated recurrent unit in the corresponding hidden layer, and N is the total number of training data.

[0118] The error at time t can be expressed as:

[0119]

[0120] Calculate the weight gradient for each gate:

[0121]

[0122]

[0123] Where: h =1-z t .

[0124] In the process of bidirectional gated recurrent unit, an adaptive momentum estimation method is constructed as follows:

[0125]

[0126] Where: and denote the first-order moment estimate and the second-order moment estimate respectively, β1 and β2 are exponential decay rates, α1 is the learning rate, ε1 is the numerical stability constant, * denotes one of z, r, h, W z ,W r ,W h is the weight matrix of the bidirectional gated recurrent unit.

[0127] The above algorithm can be used to update the weights of all neural units in the corresponding hidden layer. Finally, repeat the above steps to update the weights of another hidden layer of the bidirectional gated recurrent network. The hyperparameter update of the bidirectional gated recurrent network is achieved through the adaptive momentum estimation algorithm.

[0128] In S4: Based on the identified Hammerstein model, the light intensity and battery temperature of the photovoltaic battery system are used as the input of the Hammerstein model to obtain the corresponding prediction output; on this basis, the prediction output of the Hammerstein model is used as the input of the knowledge-driven model, and the generalized least squares method is used to identify the unknown parameters of the knowledge-driven model. The output of the knowledge-driven model is expressed as:

[0129]

[0130] Where: represents the estimated output of the Hammerstein model, represents the output of the knowledge-driven model, θ k =[a1,b1,c1] T is the unknown parameter vector of the knowledge-driven model.

[0131] The generalized least squares method is used to identify the parameters of the knowledge-driven module. The calculation process is expressed as:

[0132] θ k =(X T Ω -1 X) -1 X T Ω -1 Z

[0133] Where: Ω is the positive definite variance matrix, X is the independent variable matrix, and Z is the bias vector.

[0134] In S4: The knowledge obtained by identification is used to drive the model with data fusion to predict the current and voltage of the photovoltaic battery system, that is, the maximum power point voltage U′ m , open circuit voltage short U o ' c , maximum power point current I′ m and short-circuit current Is ' e .

[0135] In the above modeling process, the physical characteristics and data features of the photovoltaic battery system are integrated to build a mathematical model driven by knowledge and data fusion, which not only retains the advantages of high interpretability, low data requirements and good generalization of the knowledge-driven model, but also introduces the high precision advantage of the data-driven module. Through this nested structure, the model can still improve the prediction accuracy when the data volume is limited or the quality is insufficient, and optimize the calculation efficiency and enhance the overall performance.

[0136] In the above identification process, Gaussian signals and photovoltaic cell data are used to realize the separate identification of static nonlinear module and dynamic linear module parameters of the multivariable Hammerstein model, which improves the identification accuracy and has good prediction performance.

[0137] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. A photovoltaic battery system modeling and identification method driven by knowledge and data fusion, characterized by: The following steps are involved: S1: Analyze the mechanism characteristics of the photovoltaic battery system, establish a knowledge-driven model, and use the multivariable Hammerstein model to establish a data-driven model. In the Hammerstein model, a bidirectional gated recurrent unit neural network is used to describe the static nonlinear module of the Hammerstein model, and an autoregressive sliding average model is used to describe the dynamic linear module of the Hammerstein model, so as to construct a photovoltaic battery system model that integrates knowledge and data; S2: Introduce Gaussian signal and use the covariance function characteristics of Gaussian signal to decouple the static nonlinear module and dynamic linear module of Hammerstein model; S3: The parameters of the autoregressive moving average model are identified using the least squares method based on the covariance function, the hyperparameters of the bidirectional gated recurrent unit neural network are adaptively solved using the dung beetle optimization algorithm, and the weights and biases of the bidirectional gated recurrent unit neural network are updated using the adaptive momentum estimation algorithm; S4: The results obtained by the data-driven model are used as the input of the knowledge-driven model, and the generalized least squares method is used to identify the parameters of the knowledge-driven model, so as to obtain a photovoltaic battery system model that integrates knowledge and data, and further obtain the predicted values ​​of voltage and current.

2. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 1, characterized in that: In S1: Based on the external output characteristics of the photovoltaic battery system, a knowledge-driven model is established. For the photovoltaic battery system, Under any environmental conditions, the maximum power point voltage U m , open circuit voltage short U oc , Maximum power point current I m , short circuit current I se It will change according to a certain rule. By introducing the corresponding compensation coefficient, the new light intensity and the new battery temperature are calculated according to the reference light intensity and the reference battery temperature. m , U o ' c , I′ m , I s ' e , the dynamic characteristics of the photovoltaic battery system, that is, the knowledge-driven model is expressed as: U′ m =U m [1-c1(T-T ref )]ln[e+b1(S-S ref )] U o ′ c =U oc [1-c1(T-T ref )]ln[e+b1(S-S ref )] Where: T ref is the reference battery temperature, S ref is the standard reference light intensity, S represents the actual light intensity, T represents the actual battery temperature, and the compensation coefficients a1, b1 and c1 are the parameters to be identified.

3. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 2, characterized in that: A data-driven model for a photovoltaic battery system is established using a multivariable Hammerstein model, where: v(t)=f(u(t)) In the formula: u(t)=[u1(t),u2(t),…,u m (t)] T ,y(t)=[y1(t),y2(t),…,y n (t)] T ,v(t)=[v1(t),v2(t),…,v m (t)] T , q -1 represents the unit shift operator, I represents the unit matrix of appropriate dimension, n A and n B is a known order, represents the autoregressive moving average model; t represents the time.

4. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 1, characterized in that: The bidirectional gated recurrent unit neural network includes an input layer, two hidden layers and an output layer. A single neuron in the network is represented as: r t =σ(W r ⊙[h t-1 ,u t ]) With t =σ(W z ⊙[h t-1 ,at t ]) Where: W r ,W z ,W h They represent the weight matrices of the reset gate, update gate, and candidate hidden state, σ represents the sigmoid activation function, and h t-1 is the hidden state of the previous moment, ⊙ represents the point-by-point multiplication of two vectors; The hidden layer, i.e., the bidirectional gated recurrent unit, includes two-directional gated recurrent unit neural network layers. The input data is calculated in opposite directions in the two layers of gated recurrent unit neural network layers. For an input data U=u1,u2,…,u T , the forward computation of the bidirectional gated recurrent unit is expressed as: Where: and are the hidden states from left to right and from right to left, respectively, u t is the data at the tth moment of the input data; the hidden states in the two directions are fused to obtain the final hidden state: Pass the hidden state to one of the hidden layers and get the output v(t): Where: k (k = 1, 2, ..., n) represents the kth output of the static nonlinear module, W kγ represents the output weight from the γth neural unit in the corresponding hidden layer receiving the hidden state to the kth neural unit in the output layer, represents the output of the γth neural unit in the corresponding hidden layer, c=1,2,…,γ is the number of bidirectional gated recurrent units in the corresponding hidden layer.

5. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 4, characterized in that: In S2: For a static nonlinear module, if the input is a zero-mean Gaussian signal, the product of the autocovariance function of the Gaussian input signal u(t) and the constant matrix is ​​equal to the covariance function of the intermediate variable v(t) and the input u(t), that is: R vu (τ)=αR u (t) Where: α is the constant matrix, τ is the time constant, Cross-covariance function matrix Autocovariance function matrix Using αR u (τ) instead of R vu (τ), realizes the decoupling of the static nonlinear module and the linear module of the Hammerstein model.

6. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 5, characterized in that: In S3, The dung beetle optimization algorithm is used to adaptively solve the hyperparameters of the bidirectional gated recurrent unit neural network, including the number of hidden layer neurons, learning rate and number of iterations in the bidirectional gated recurrent unit neural network. These hyperparameters constitute the vector X. The specific solution process is expressed as: (1) When there are no obstacles, in order to simulate the rolling behavior of the dung beetle, it moves in the search space along a given direction and navigates by the sun. Assuming that the intensity of the light source affects the path, the position of the dung beetle is updated during the rolling process as follows: X i (t-1)=X i (t)+a×l×X i (t-1)+b×Δx Δx=|X i (t)-X w | Where: t represents the current iteration number, X i (t) represents the position information of the i-th dung beetle at the t-th iteration, l represents the deflection coefficient, b∈(0,1), a is the coefficient, which takes the value of -1 or 1, and X w represents the global worst position, and Δx is used to simulate the change of light intensity; (2) When a dung beetle encounters an obstacle, it will reposition itself by "dancing" to obtain a new route. The dancing behavior obtains a new rolling direction through the tangent function, which is defined in the interval [0,π]. The new position is updated as: X i (t-1)=X i (t)+tan(θ)|X i (t)-X i (t-1)| Where: θ = [0, π] is the deflection angle, |X i (t)-X i (t-1)| is the difference between the position of the i-th dung beetle at the t-th iteration and the position at the t-1-th iteration. If θ is 0, π / 2 or π, the position of the dung beetle will not be updated.

7. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 6, characterized in that: In S3, the adaptive momentum estimation algorithm is expressed as: Where: and denote the first-order moment estimate and the second-order moment estimate respectively, β1 and β2 are exponential decay rates, α1 is the learning rate, ε1 is the numerical stability constant, * denotes one of z, r, h, W z ,W r ,W h is the weight matrix of the bidirectional gated recurrent unit.

8. The photovoltaic battery system modeling and identification method driven by knowledge and data fusion as claimed in claim 1, characterized in that: In S4, the predicted output of the Hammerstein model is used as the input of the knowledge-driven model, and the generalized least squares method is used to identify the unknown parameters of the knowledge-driven model. The output of the knowledge-driven model is expressed as: Where: represents the output of the Hammerstein model, y k (t) represents the output of the knowledge-driven model, θ k is the knowledge-driven model parameter vector to be identified.

Images