Wind power generation prediction method integrated with physical knowledge

Abstract

The invention discloses a wind power generation capacity prediction method integrated with physical knowledge, and the method comprises the following steps: obtaining wind power generation original sample data, carrying out the preprocessing of the data, and dividing a data set; and constructing a simplified energy conservation equation, solving a compensation coefficient matrix, establishing a prediction model, calculating prediction residual loss by using the constructed simplified energy conservation equation, optimizing model parameters on a training set, and evaluating the model. According to the method, a simplified energy conservation equation in the wind power generation process is constructed, the energy conservation equation is used as physical knowledge to be fused into modeling, and the accuracy of predicting the wind power generation capacity by the prediction model through wind power process data is effectively improved.

Description

Technical Field

[0001] The invention belongs to the technical field of short-term wind power generation prediction, and in particular relates to a wind power generation prediction method incorporating physical knowledge. Background Art

[0002] As the installed capacity of wind turbines increases year by year, the volatility and uncertainty of wind power generation pose a challenge to maintaining the operational stability of the power grid. In order to reduce the impact of wind power generation on the power grid, accurate prediction of wind power generation is essential. Wind power generation forecasts can be divided into short-term, medium-term and long-term forecasts according to the length of the forecast period. Among them, long-term forecasts usually refer to forecasts in units of "months" and "years", which are mainly used for the design of wind farm movement plans and wind power resource assessment, and preliminary planning of wind farms; medium-term forecasts usually refer to forecasts in units of "days" and "weeks", which are mainly used for troubleshooting and maintenance of wind power equipment in the power grid; short-term forecasts are forecasts based on "hours" or "minutes", which are mainly used to adjust the power system's energy reserves, and achieve effective power grid dispatch by reducing the instability of the power grid system caused by wind power.

[0003] At present, with the development and application of data acquisition equipment, a large amount of wind power generation process data has been collected, and data-driven short-term wind power generation prediction methods have been widely used. On this basis, the use of deep learning methods to predict short-term wind power has high accuracy and robustness, and has become one of the main methods for short-term wind power prediction. However, the features extracted from the raw data by deep learning do not rely on manual design, but are directly learned from the data using a general learning process, so there is no need to rely too much on the designer's experience. Summary of the invention

[0004] In order to solve the problem of lack of interpretability of the model in the process of wind power prediction using deep learning methods, the purpose of the present invention is to provide a wind power prediction method that incorporates physical knowledge. Since the wind power generation process involves the mutual conversion of multiple energies, the present invention uses the law of conservation of energy as the physical knowledge in the wind power generation process, and designs a residual constraint method to guide the model behavior to follow the underlying physical knowledge.

[0005] The specific technical solutions are as follows:

[0006] A wind power generation prediction method incorporating physical knowledge comprises the following steps:

[0007] (1) Obtaining original sample data of wind power generation;

[0008] (2) Preprocessing of wind power generation raw sample data and data set division: The acquired raw sample data is divided into a training set and a test set; in order to speed up the model convergence and reduce the model training time, the training set data and the test set data are normalized;

[0009] (3) Construct a simplified energy conservation equation and solve the compensation coefficient matrix: Based on the principle of energy conservation, the collected data is used to calculate the various energies in the wind power generation process, and the compensation coefficient is used to correct the calculation results;

[0010] (4) Establishing a prediction model: Establish a prediction model and use the simplified energy conservation equation to calculate the prediction residual loss, and finally optimize the model parameters on the training set;

[0011] (5) Model performance evaluation: Introducing evaluation indicators such as root mean square error (RMSE), maximum absolute error (MAE), and coefficient of determination (R 2 ) to evaluate the model.

[0012] Furthermore, the specific process of step (3) is as follows:

[0013] Step 3.1: Construct the simplified energy conservation equation:

[0014] During the operation of wind turbines, wind energy is converted into other forms of energy such as mechanical energy, electrical energy, and thermal energy. However, considering the unit dimension changes between variables, in the process of formula simplification, the compensation coefficients are used to compensate for the differences in calculation results caused by inconsistent dimensions.

[0015] To calculate the wind energy captured by the wind turbine, A is used to represent the rotor swept area in m 2 , τ represents the current time, v(τ) represents the wind speed at the current moment, in m / s, ρ is the air density, and within the time period T, the wind energy captured by the wind turbine is E wind for:

[0016] E wind =1 / 3ρATv(τ) 3 (1)

[0017] At this time, the wind power per unit area and per unit time P, in units of w, can be expressed as:

[0018] P=1 / 3ρv(τ) 3 (2)

[0019] To simplify the calculation, the wind energy and the mechanical energy obtained by the wind turbine are calculated uniformly. First, the constant coefficients in the above calculation process are unified with the compensation coefficient C w At the same time, the compensation coefficient can modify the calculation results numerically, thereby reducing the impact of unified calculation. At this time, within the time period T, the difference between wind energy and mechanical energy E w for:

[0020] E w =∫ T C w v(τ) 3 dτ=C w e w (3)

[0021] Where Cw is the supplementary coefficient, e w Represents the integral term in the formula.

[0022] During the power generation process of a wind turbine, the heat energy mainly consists of two parts: the heat absorbed by the wind turbine itself and the heat dissipated by the wind turbine. When an object absorbs heat, the temperature will rise. At this time, the relationship between the absorbed heat energy and the temperature is as follows:

[0023] Q=mcΔt (4)

[0024] Where Q is the absorbed heat energy, in J, m is the mass of the object, in kg, c is the specific heat capacity of the object, in J / (kg*℃), and Δt is the change in the temperature of the object, in ℃. Therefore, in the time period T, the total amount of heat energy absorbed by the wind turbine is Q in for:

[0025]

[0026] Among them, there are N parts in the wind turbine for which temperature change data can be obtained, and δ i represents the product of the mass and specific heat capacity of the ith part, Δt i It represents the temperature change of the i-th component in the time period T, in degrees Celsius. Since the relevant data of all components cannot be obtained in the actual data set, the compensation coefficient matrix C is used when calculating the heat energy absorbed by the wind turbine. in Correct the results, e in =[Δt1,Δt2,…,Δt N ] represents the temperature variation matrix of the parts that can be obtained.

[0027] Since there is strong convection between the air and the wind turbine during the operation of the wind turbine, and the temperature difference between the wind turbine and the air is small, the radiation heat transfer is ignored and only the convection heat transfer between the wind turbine and the air is considered. The convection heat transfer formula per unit time is:

[0028] QH =αA ex (t wall -t air ) (6)

[0029] Where Q H is the convective heat transfer per unit time, α is the convective heat transfer coefficient, the unit is w / (m 2 *k), t wall is the wind turbine wall temperature, in K, t air is the air temperature in K, A ex is the heat exchange area between the wind turbine and the air, in m 2 .

[0030] Secondly, when calculating the heat dissipation of the wind turbine, in order to simplify the calculation, the wall temperature of the wind turbine is ignored, and the convection heat transfer compensation coefficient is introduced to compensate the calculation result, where C is used ex Indicates the compensation coefficient in the heat dissipation process, when the ambient temperature is a function of time, that is, the ambient temperature is t air (τ), then in the time period T, the heat exchanged between the wind turbine and the air is Q ex for:

[0031] Q ex =∫ T C ex t air (τ)dτ=C ex e ex (7)

[0032] where e ex The integral term in the formula representing the heat exchange between the wind turbine and the air.

[0033] The electric energy is divided into the electric energy output to the grid and the electric energy stored in the capacitor and inductor, which correspond to the active power P a (τ) and reactive power P r (τ), output electrical energy E a for:

[0034] E a =∫ T P a (τ)dτ (8)

[0035] Considering other losses in the circuit, C r Represents the energy compensation coefficient, so the energy E stored in the capacitor and inductor r for:

[0036] E r =∫ T C r P r (τ)dτ=Cr e r (9)

[0037] where e r Represents the integral term in the above calculation formula.

[0038] Step 3.2: Calculate simplified energy conservation compensation parameters:

[0039] Since the data are discrete values ​​sampled at fixed time intervals, this paper uses the following Newton-Cotes formula to solve the numerical integral and approximate the integral term in the formula of step 3.1. The Newton-Cotes formula is as follows:

[0040]

[0041] Where a and b are the endpoints of the quadrature region, and the quadrature region will be divided into H equal parts. i is the i-th equally divided point, and has x in value i =a+i×(ba) / H,λ i are the corresponding coefficients of the Newton-Cotes formula, and f(x) is the function whose product needs to be calculated.

[0042] Assuming that the compensation coefficient remains unchanged within the time period T, when the amount of data is sufficient, the Newton-Cotes formula can be used to construct a set of equations, and then the method of undetermined coefficients can be used to solve the corresponding compensation coefficients.

[0043] If C=[C w ,C r ,C in ,C ex ] represents the compensation coefficient matrix, E = [e w ,e r ,e in ,e ex ] represents the energy matrix. At this point, a simplified energy conservation model can be constructed as follows:

[0044] E a =CE T (11)

[0045] Furthermore, the specific process of step (4) is as follows:

[0046] Step 4.1: Get the prediction loss function of the model:

[0047] For the wind power generation data set D, assuming that there are β groups of samples, each group of samples includes auxiliary feature variables and wind power generation variables. At this time, the data set can be expressed as follows:

[0048]

[0049] Where Xi represents the i-th group of samples, x i represents the auxiliary feature variable matrix in the i-th group of samples, x ij represents the jth auxiliary variable in the i-th group of samples, p i represents the wind power generation active power in the i-th group of samples, and κ represents the number of categories of the auxiliary feature variables. In order to achieve the wind power generation prediction task, it is necessary to use other auxiliary feature variables x at the current moment to predict the wind power generation p. In order to train a neural network for the wind power generation prediction task, it is necessary to optimize the following mean square error loss function

[0050]

[0051] in To use the neural network model based on the auxiliary feature variable x i The predicted value of the output active power of wind power generation.

[0052] Step 4.2: Construct a loss function that incorporates physical knowledge:

[0053] At this time, in order to apply the energy conservation model established in step 3 to the soft sensor modeling, the following steps need to be performed. First, the original time series data needs to be divided into multiple equal-length data segments according to a fixed length, and T all ={T1,T2,…,T ξ} represents the time period T all Divide into ξ time period subspaces of the same length, ξ>κ. Since the wind power generation data set D is sampled at fixed time intervals, the corresponding numerical integral can be calculated using the Newton-Cotes formula. If Ω() is used to represent the Newton-Cotes formula, we can obtain:

[0054]

[0055] Where Z represents the integral of the variable over time, X a represents the starting point data within the time period, and θ represents the number of equally divided points required by the Newton-Cotes formula. all On the basis of the above, we can establish a set of equations ξ-κ+1. After solving the corresponding compensation coefficients for each set of equations, we can get the time period T by averaging. all The average compensation coefficient matrix

[0056] Afterwards, the average compensation coefficient matrix and auxiliary variables within the time window, the estimated value of the generated electricity can be solved At the same time, the neural network can output the active power prediction value within the current time window The power generation prediction value can be solved according to the available At this point, the residual loss function can be obtained:

[0057]

[0058] Where n is the number of solved power generation, and They represent the i-th power generation solved by using the neural network prediction value and the power generation solved by using the energy conservation model.

[0059] Therefore, the final global loss function is:

[0060] L=L MSE +ηL R (16)

[0061] Where η is a hyperparameter used to balance the loss function.

[0062] The beneficial effects of the present invention are:

[0063] The present invention proposes a wind power data-driven prediction method that incorporates physical knowledge. Since the wind power generation process involves the mutual conversion of multiple energies, the present invention uses the energy conservation equation as the prior physical knowledge in the wind power generation process and optimizes the output results of the prediction model through the residual constraint method. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of the method of the present invention;

[0065] Figure 2 It is a comparison diagram of the effects of the method of the present invention and the existing method;

[0066] Figure 3 This is a comparison chart of absolute prediction errors between the method of the present invention and the existing method. DETAILED DESCRIPTION

[0067] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but the protection scope of the present invention is not limited thereto.

[0068] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0069] like Figure 1 As shown, a wind power generation prediction method incorporating physical knowledge is taken as an example to construct active power prediction of a wind power plant, and includes the following steps:

[0070] 1) Obtain wind power generation data

[0071] The 10 process variables collected are the ambient temperature of the wind turbine, the bearing shaft temperature, the gearbox bearing temperature, the gearbox oil temperature, the generator winding 1 temperature, the generator winding 2 temperature, the hub temperature, the main box temperature, the reactive power, and the wind speed. The final predicted variable is the active power.

[0072] 2) Wind power data preprocessing and data set division

[0073] The received wind power data is divided into a training set and a test set. Then, in order to speed up the model convergence and reduce the model training time, the training data and the test data are normalized.

[0074] The specific operation process is as follows:

[0075] Step 2.1: Divide the dataset

[0076] The collected 971 wind power data are divided into two parts: 671 samples are training sets and the remaining 300 samples are test sets;

[0077] Step 2.2: Data normalization

[0078] In order to speed up the convergence of the model and reduce the training time, the data is normalized to reduce the difference between data dimensions;

[0079] 3) Construct a simplified energy conservation equation

[0080] Based on the principle of energy conservation, the collected data is used to calculate the various energies in the wind power generation process, and the compensation coefficient is used to correct the calculation results;

[0081] The specific operation process is as follows:

[0082] Step 3.1: Construct the simplified energy conservation equation

[0083] During the operation of the wind turbine, wind energy is converted into other forms of energy such as mechanical energy, electrical energy, and thermal energy. However, considering that it is impossible to obtain all the data in the actual process to construct a complete energy conservation equation, the present invention simplifies the energy conservation equation.

[0084] To calculate the wind energy captured by the wind turbine, A is used to represent the rotor swept area in m 2 ; τ represents the current time, v(τ) represents the wind speed at the current moment, in m / s; ρ represents the air density, in kg / m 3 ; In the time period T, the wind energy captured by the wind turbine is E wind for:

[0085] E wind =1 / 3ρATv(τ) 3 (1)

[0086] At this time, the wind power per unit area and per unit time P, in units of w, can be expressed as:

[0087] P=1 / 3ρv(τ) 3 (2)

[0088] To simplify the calculation, the wind energy and the mechanical energy obtained by the wind turbine are calculated uniformly. First, the constant coefficient in the calculation process is replaced by the compensation coefficient Cw. At the same time, the compensation coefficient can correct the calculation result numerically, thereby reducing the impact of the unified calculation. At this time, within the time period T, τ is used to represent the time, the wind speed is v(τ), and the difference between wind energy and mechanical energy is E w for:

[0089] E w =∫ T C w v(τ) 3 dτ=C w e w (3)

[0090] Where Cw is the supplementary coefficient, e w represents the integral term in the formula;

[0091] During the power generation process of a wind turbine, the heat energy mainly consists of two parts: the heat absorbed by the wind turbine itself and the heat dissipation of the wind turbine. When an object absorbs heat, the temperature rises. At this time, the relationship between the absorbed heat energy and the temperature is as follows:

[0092] Q=mcΔt (4)

[0093] Where Q is the absorbed heat energy, in J; m is the mass of the object, in kg; c is the specific heat capacity of the object, in J / (kg*℃); Δt is the change in the temperature of the object, in ℃. Therefore, in the time period T, the total amount of heat energy absorbed by the wind turbine is Q in for:

[0094]

[0095] Among them, there are N parts in the wind turbine for which temperature change data can be obtained, and δ i represents the product of the mass and specific heat capacity of the ith part, Δt i It represents the temperature change of the i-th component in the time period T, in degrees Celsius. Since the relevant data of all components cannot be obtained in the actual data set, the compensation coefficient matrix C is used when calculating the heat energy absorbed by the wind turbine. in Correct the results, e in=[Δt1,Δt2,…,Δt N ] represents the temperature variation matrix of the available parts;

[0096] Since there is strong convection between the air and the wind turbine during the operation of the wind turbine, and the temperature difference between the wind turbine and the air is small, the radiation heat transfer is ignored and only the convection heat transfer between the wind turbine and the air is considered. The convection heat transfer formula per unit time is:

[0097] Q H =αA ex (t wall -t air ) (6)

[0098] Where Q H is the convective heat transfer per unit time, α is the convective heat transfer coefficient, the unit is w / (m 2 *k), t wall is the wind turbine wall temperature, in K, t air is the air temperature in K, A ex is the heat exchange area between the wind turbine and the air, in m 2 ;

[0099] Secondly, when calculating the heat dissipation of the wind turbine, in order to simplify the calculation, the wall temperature of the wind turbine is ignored, and the convection heat transfer compensation coefficient is introduced to compensate the calculation result, where C is used ex Indicates the compensation coefficient in the heat dissipation process, when the ambient temperature is a function of time, that is, the ambient temperature is t air (τ), then in the time period T, the heat exchanged between the wind turbine and the air is Q ex for:

[0100] Q ex =∫ T C ex t air (τ)dτ=C ex e ex (7)

[0101] where e ex The integral term in the formula representing the heat exchange between the wind turbine and the air;

[0102] The electric energy is divided into the electric energy output to the grid and the electric energy stored in the capacitor and inductor, which correspond to the active power P a (τ) and reactive power P r (τ), output electrical energy E a for:

[0103] E a =∫ T P a (τ)dτ (8)

[0104] Considering other losses in the circuit, C r Represents the energy compensation coefficient, so the energy E stored in the capacitor and inductor r for:

[0105] E r =∫ T C r P r (τ)dτ=C r e r (9)

[0106] where e r Represents the integral term in the above calculation formula;

[0107] Step 3.2: Calculate simplified energy conservation compensation parameters

[0108] In step 3.1, a simplified energy conservation model has been constructed, but the compensation coefficient is an unknown constant, so it needs to be solved through the following steps. Since the data is a discrete value sampled at a fixed time interval, the present invention uses the following Newton-Cotes formula to solve the numerical integral and thus approximately solve the integral term in the formula of step 3.1. The Newton-Cotes formula is as follows

[0109]

[0110] Where a and b are the endpoints of the quadrature region, and the quadrature region will be divided into H equal parts. i is the i-th equally divided point, and has x in value i =a+i×(ba) / H,λ i is the corresponding coefficient of Newton-Cotes formula, f(x) is the function to be integrated;

[0111] Assuming that the compensation coefficient remains unchanged within the time period T, when the amount of data is sufficient, the Newton-Cotes formula can be used to construct the equation group, and then the method of undetermined coefficients can be used to solve the corresponding compensation coefficient;

[0112] If C=[C w ,C r ,C in ,C ex ] represents the compensation coefficient matrix, E = [e w ,e r ,e in ,e ex ] represents the energy matrix. At this time, a simplified energy conservation model can be constructed as follows:

[0113] E a =CE T (11)

[0114] 4) Establish a prediction model: Establish a prediction model and use the simplified energy conservation equation to calculate the prediction residual loss, and finally optimize the model parameters on the training set;

[0115] The specific operation process is as follows:

[0116] Step 4.1: Get the prediction loss function of the model:

[0117] For the wind power generation data set D, assuming that there are β groups of samples, each group of samples includes κ other auxiliary feature variables and wind power generation variables. At this time, the data set can be expressed as follows:

[0118]

[0119] Where X i represents the i-th group of samples, x i represents the auxiliary feature variable in the i-th group of samples, x ij represents the jth auxiliary variable in the i-th group of samples, p i represents the active power of wind power generation in the i-th group of samples, and κ represents the number of categories of auxiliary feature variables. In order to achieve the wind power generation prediction task, it is necessary to use other auxiliary feature variables x at the current moment to predict the wind power generation p;

[0120] To train a neural network for wind power generation prediction, we need to optimize the following mean square error loss function:

[0121]

[0122] in To use the neural network model based on the auxiliary feature variable x i The predicted value of the output active power of wind power generation.

[0123] Step 4.2: Construct a loss function that incorporates physical knowledge:

[0124] At this time, in order to apply the energy conservation model established in step 3 to the soft sensor modeling, the following steps need to be performed. First, the original time series data needs to be divided into multiple segments of equal length according to a fixed length L, and T all ={T1,T2,…,T ξ} represents the time period T all Divide into ξ time period subspaces of equal length, ξ>κ. Since the wind power generation data set X is sampled at fixed time intervals, the corresponding numerical integral can be calculated using the Newton-Cotes formula. If Ω() is used to represent the Newton-Cotes formula, we can obtain:

[0125]

[0126] Where Z represents the integral of the variable over time, X a represents the starting point data within the time period, and θ represents the number of equally divided points required by the Newton-Cotes formula. all On the basis of the above, we can establish a set of equations ξ-κ+1. After solving the corresponding compensation coefficients for each set of equations, we can get the time period T by averaging. all The average compensation coefficient matrix

[0127] Afterwards, the average compensation coefficient matrix and auxiliary variables within the time window, the estimated value of the generated electricity can be solved At the same time, the neural network can output the active power prediction value within the current time window The power generation prediction value can be solved according to the available At this point, the physical loss function can be obtained:

[0128]

[0129] Where n is the number of solved power generation, and They represent the i-th power generation solved by using the neural network prediction value and the power generation solved by using the energy conservation model.

[0130] Therefore, the final global loss function is:

[0131] L=L MSE +ηL R (16)

[0132] Where η is a hyperparameter used to balance the loss function.

[0133] 5) Model performance evaluation: Introduce evaluation indicators such as root mean square error (RMSE), maximum absolute error (MAE) and coefficient of determination (R 2 ) to evaluate the model.

[0134] The gated recurrent unit structure (Gated Recurrent Unit, GRU) and the convolutional neural network (Convolutional Neural Networks, CNN) model are used as the prediction model respectively. The present invention constructs a physics knowledge guided neural network (Physics Informed Neural Network, PINN) based on GRU and CNN. The models improved by the method of the present invention are PINN-GRU and PINN-CNN respectively. Given a learning rate α = 0.0001, a residual coefficient η = 0.001, and after each model is trained on the training set, the results on the test set are shown in Table 1. Specifically, after the GRU and CNN models are improved by the method of the present invention, the RMSE values ​​are reduced by 34.05% and 39.26% respectively, the MAE values ​​are reduced by 34.10% and 29.50% respectively, and the R 2 The results show that the method of the present invention can effectively improve the performance of the prediction model. Figure 2 and Figure 3 As shown, after using the method of the present invention, the predicted value of wind power generation is closer to the true value and has a lower absolute prediction error.

[0135] Table 1 Comparison of model performance

[0136] RMSE MAE <![CDATA[R 2 ]]> GRU 56.1497 33.8822 0.9847 PINN-GRU 37.0310 22.3299 0.9937 CNN 177.1956 88.4292 0.8183 PINN-CNN 107.6238 62.3409 0.9438

[0137] The present invention proposes a wind power data-driven prediction method that incorporates physical knowledge. Since the wind power generation process involves the mutual conversion of multiple energies, the present invention uses the energy conservation equation as the prior physical knowledge in the wind power generation process, optimizes the output results of the prediction model through the residual constraint method, and realizes accurate prediction of wind power generation.

[0138] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A wind power generation prediction method incorporating physical knowledge, characterized in that: The steps include: 1) Obtain original sample data of wind power generation; 2) Preprocessing of wind power generation original sample data and data set division: The acquired original sample data is divided into a training set and a test set. In order to speed up the model convergence and reduce the model training time, the training set data and the test data are normalized; 3) Construct a simplified energy conservation equation and solve the compensation coefficient matrix: Based on physical principles, use the collected data to calculate the various energies in the wind power generation process, and use the compensation coefficient to correct the calculation results; 4) Establish a prediction model: Establish a prediction model, use the simplified energy conservation equation to calculate the prediction residual loss, and optimize the model parameters on the training set; 5) Model performance evaluation: The evaluation indicators such as root mean square error, maximum absolute error and decision coefficient are introduced to evaluate the model.

2. A wind power generation prediction method incorporating physical knowledge as claimed in claim 1, characterized in that: The specific operation process of step 3) is as follows: Taking into account the unit dimension change problem between variables, in the process of formula simplification, the compensation coefficient simultaneously compensates for the difference in calculation results caused by inconsistent dimensions. Step 3.1: Construct the simplified energy conservation equation Wind energy captured by wind turbines wind for E wind =1 / 3ρATv(τ) 3 (1) Where A represents the swept area of ​​the wind wheel, in m 2 ; τ represents the current time, v(τ) represents the wind speed at the current moment, in m / s; ρ represents the air density, in kg / m 3 ; T is the time period, unit is s; At this time, the wind power per unit area and per unit time, P, in units of w, is expressed as: P=1 / 3ρv(τ) 3 (2) The difference between wind energy and mechanical energy E w for: AND w =∫ T C w v(τ) 3 dτ=C w And w (3) In the formula, C w is the supplementary coefficient, e w represents the integral term in the formula; When an object absorbs heat, the absorbed heat energy and temperature have the following relationship: Q=mcΔt (4) In the formula, Q is the absorbed heat energy, in J; m is the mass of the object, in kg; c is the specific heat capacity of the object, in J / (kg*℃); Δt is the change in the temperature of the object, in ℃; During the time period T, the total amount of heat energy absorbed by the wind turbine is Q in for: Where N represents the N parts of the fan for which temperature change data can be obtained; δ i represents the product of the mass and specific heat capacity of the i-th part; Δt i It represents the temperature change of the i-th part in the time period T, in °C; in represents the compensation coefficient matrix, e in =[Δt1,Δt2,…,Δt N ] represents the temperature variation matrix of the available parts; The formula for convective heat transfer per unit time is: Q H =αA ex (t wall -t air ) (6) Where α is the convection heat transfer coefficient, unit is w / (m 2 *k), t wall is the wind turbine wall temperature, in K, t air is the air temperature in K, A ex is the heat exchange area between the wind turbine and the air, in m 2 ; When the ambient temperature is a function of time, that is, the ambient temperature is t air (τ), then in the time period T, the heat exchanged between the wind turbine and the air is Q ex for: Q ex =∫ T C ex t air (τ)dτ=C ex e ex (7) In the formula, C ex Indicates the compensation coefficient during the heat dissipation process; e ex The integral term in the formula representing the heat exchange between the wind turbine and the air; The electric energy is divided into the electric energy output to the grid and the electric energy stored in the capacitor and inductor, which correspond to the active power P a (τ) and reactive power P r (τ), output electrical energy E a for: E a =∫ T P a (τ)dτ (8) Considering other losses in the circuit, C r Represents the energy compensation coefficient, so the energy E stored in the capacitor and inductor r for: AND r =∫ T C r P r (τ)dτ=C r And r (9) In the formula, e r Represents the integral term in the above calculation formula; Step 3.2: Calculate simplified energy conservation compensation parameters: Since the data is a discrete value obtained by sampling at fixed time intervals, the following Newton-Cotes formula is used to solve the numerical integration to approximately solve the integral term in the formula of step 3.

1. The Newton-Cotes formula is as follows Where a and b are the endpoints of the region to be quadratured. The region to be quadratured will be divided into H equal parts. i is the i-th equally divided point, and has x in value i =a+i×(ba) / H,λ i is the corresponding coefficient of Newton-Cotes formula, f(x) is the function to be integrated; Assuming that the compensation coefficient remains unchanged within the time period T, when the amount of data is sufficient, the Newton-Cotes formula is used to construct the equation group, and then the method of undetermined coefficients is used to solve the corresponding compensation coefficient. If C=[C w ,C r ,C in ,C ex ] represents the compensation coefficient matrix, E = [e w ,e r ,e in ,e ex ] represents the energy matrix. At this time, a simplified energy conservation model can be constructed as follows: It is a =EC T (11) At this point, sufficient equations are established and the compensation coefficient matrix is ​​solved using the method of undetermined coefficients.

3. A method for predicting wind power generation incorporating physical knowledge as claimed in claim 1, characterized in that: The specific operation process of step 4) is: Step 4.1: Get the prediction loss function of the model: For the wind power generation data set D, assuming that there are β groups of samples, each group of samples includes κ other auxiliary feature variables and wind power generation variables. The data set is expressed as Where, X i represents the i-th group of samples, x i represents the auxiliary feature variable in the i-th group of samples, x ij represents the jth auxiliary variable in the i-th group of samples, p i represents the active power of wind power generation in the i-th group of samples, κ represents the number of categories of the auxiliary characteristic variable; To train a neural network for wind power generation prediction, we need to optimize the following mean square error loss function: In the formula, To use the neural network model based on the auxiliary feature variable x i The predicted value of the output active power of wind power generation; Step 4.2: Construct a loss function that incorporates physical knowledge: The established energy conservation model is applied to soft sensor modeling. First, the original time series data needs to be divided into multiple segments of equal length according to a fixed length L, and then the data is divided into multiple segments of equal length using T all ={T1,T2,…,T ξ } represents the time period T all Divide into ξ time period subspaces of the same length, ξ>κ. Since the wind power generation data set X is sampled at fixed time intervals, the corresponding numerical integral can be calculated using the Newton-Cotes formula. If Ω() is used to represent the Newton-Cotes formula, we can get: In the formula, Z represents the integral of the variable over time; X a represents the starting point data within the time period, and θ represents the number of equally divided points required by the Newton-Cotes formula; At all time data T all On the basis of the equations ξ-κ+1, a set of equations is established. After solving the corresponding compensation coefficients for each set of equations, the time period T is obtained by averaging. all The average compensation coefficient matrix Using the average compensation coefficient matrix and auxiliary variables in the time window to solve the estimated value of the power generation At the same time, the neural network is used to output the active power prediction value in the current time window Solve the predicted value of power generation At this point, the physical loss function can be obtained: Where n is the number of solved power generation and They represent the i-th power generation solved by using the neural network prediction value and the power generation solved by using the energy conservation model; The final global loss function is: L=L MSE +ηL R (16) Where η is a hyperparameter used to balance the loss function.

Images