A physical-information driven grid frequency prediction method considering the participation of new energy in frequency regulation
The integration of physical and information models using xgboost and gcforest architectures addresses the challenges of grid frequency prediction in large power systems with new energy sources, improving accuracy and efficiency.
Patent Information
- Application Number
- CN202411089399.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-09
AI Technical Summary
Existing grid frequency prediction methods struggle with accuracy and complexity in large power systems due to the integration of new energy sources, which introduce dynamic changes and nonlinearity, and lack effective models for frequency regulation.
A physical-information driven power grid frequency prediction method integrating a frequency prediction physical model and an information model using xgboost and gcforest architectures, leveraging active disturbance data for improved prediction accuracy and efficiency.
Enhances frequency prediction accuracy and reduces complexity by combining physical and information models, providing refined modeling and fast solution efficiency for grid frequency regulation.
Smart Images

Figure CN118630791B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of safe operation of electric power systems, and in particular to a physical-information driven power grid frequency prediction method taking into account the participation of new energy in frequency modulation. Background Art
[0002] Since new energy sources rely on power electronic equipment to be connected to the grid, the high proportion of power electronic equipment connected has greatly changed the stability and dynamic characteristics of the power system. Its dynamic behavior gradually shows characteristics such as control process dominance, fast change process, multi-time scale coupling, and strong nonlinearity, which makes the system synchronization and stability face unprecedented challenges. The weak / no grid-building capability of new energy sources and the high-order uncertainty of their output not only make the system frequency more likely to trigger low-frequency load reduction protection, but also increase the difficulty of regulation and control by the power system dispatching department. Therefore, it is imperative to improve the ability to perceive dynamic changes in system frequency.
[0003] Traditional grid frequency prediction methods mainly rely on time domain simulation. This method can take into account the detailed parameters of the power system, so the frequency prediction accuracy is high. However, when applied to large power systems, this method will encounter challenges in modeling complexity and solution. The single-machine equivalent physical model ignores the system network structure and only contains an equivalent generator rotor. It has a simple structure and fast calculation speed. It mainly involves the system average frequency model and the system frequency prediction physical model. However, although the physical model can describe the impact of new energy grid connection on the system frequency relatively simply, it generally ignores the comprehensive impact of different frequency regulation resources on the system frequency response. In addition, due to the simplification of system modeling, the above-mentioned physics-based models show low frequency prediction performance. The artificial intelligence method based on information model can mine the key features contained in the power system by offline training the mapping relationship between system state information and dynamic frequency. This method does not need to establish and derive complex physical formulas, has the characteristics of taking into account both prediction accuracy and calculation speed, and provides new possibilities for predicting the dynamic response of grid frequency in multi-system scenarios. However, the frequency prediction method based on information model relies on sample quality, which makes it unable to accurately respond to the time characteristics of dynamic frequency changes in the power grid. In addition, they also have problems such as slow offline training speed and weak generalization ability. Summary of the invention
[0004] The purpose of the present invention is to provide a physical-information driven grid frequency prediction method taking into account the participation of new energy in frequency regulation, so as to overcome the problems existing in the above-mentioned background technology.
[0005] The present invention solves the technical problem by the following technical solutions:
[0006] A physical-information driven power grid frequency prediction method taking into account the participation of new energy in frequency modulation is characterized by comprising the following steps:
[0007] S1, build frequency prediction physical model;
[0008] S2, build a frequency prediction information model based on xgboost;
[0009] S3, collecting training data, including the active disturbance size, active disturbance location, and frequency dynamic change data of the power grid, and using the collected training data to train the frequency prediction information model to obtain a frequency prediction result of the frequency prediction information model;
[0010] S4, integrate the frequency prediction physical model and the frequency prediction information model to build a physical-information driven power grid frequency prediction model based on the gcforest architecture;
[0011] S5, collecting sample data, including training data, frequency prediction results of the frequency prediction physical model, and frequency prediction results of the frequency prediction information model, and using the collected sample data to train the physical-information driven power grid frequency prediction model;
[0012] S6, for the trained physical-information driven grid frequency prediction model, based on the grid active disturbance information, generates a grid frequency dynamic prediction result.
[0013] Furthermore, the method of constructing a frequency prediction physical model in S1 comprises the following steps:
[0014] S1.1, build a physical model for power system frequency prediction taking into account the participation of new energy in frequency regulation;
[0015] S1.2, obtaining a time domain analytical solution of the frequency prediction physical model according to the power system frequency prediction physical model constructed in S1.1;
[0016] S1.3, based on the time domain analytical solution in S1.2, generate the frequency prediction result f of the physical model for predicting the frequency of the power grid after the disturbance according to the active power disturbance ∆P of the power grid SFR , and use the result as the frequency prediction result of the frequency prediction physical model.
[0017] Moreover, the method S2 for constructing a frequency prediction information model based on xgboost comprises the following steps:
[0018] S2.1, construct the frequency prediction information model structure based on xgboost and describe the frequency prediction problem;
[0019] S2.2, according to the frequency prediction information model structure and frequency prediction problem description constructed in step S2.1, analyze and select the key features that affect the dynamic change of power grid frequency, and construct feature engineering as the input feature X of the frequency prediction information model I ;
[0020] S2.3, based on the input feature X analyzed in step S2.2 I , construct the objective function based on the xgboost algorithm, including the loss function L I and regularization term Ω, which is used for supervised learning training of the frequency prediction information model in step S3;
[0021] S2.4, using the input feature X in S2.2 I , taking the real frequency y of the system as the label data, forming the sample set D I ={X I , y}, which is used for supervised learning training of the frequency prediction information model in step S3.
[0022] Moreover, the input features XI of the frequency prediction information model are key features that have a great impact on the dynamic change of the power grid frequency, including the system maximum active power feed-in loss P, the system active power disturbance position l, and the operating status u of the generator set.
[0023] Moreover, the method S4 for constructing a physical-information driven power grid frequency prediction model based on the gcforest architecture comprises the following steps:
[0024] S4.1, determining that the fusion mode of the frequency prediction physical model and the frequency prediction information model is parallel;
[0025] S4.2, based on the parallel fusion method of step S4.1, a physical-information driven power grid frequency prediction model is constructed using a cascaded gcforest architecture;
[0026] S4.3, based on the physical-information driven power grid frequency prediction model in S4.2, construct feature engineering X, including input features X of the frequency prediction information model I , the prediction result f of the frequency prediction information model I , the prediction result of the frequency prediction physical model f SFR ;
[0027] S4.4, based on the feature engineering X constructed in S4.3, the mean square error judgment standard is used to determine the number of cascade layers of physical-information driven power grid frequency prediction, and the physical-information driven power grid frequency prediction model structure is obtained;
[0028] S4.5, based on the physical-information driven power grid frequency prediction model structure finally determined in S4.4, the physical-information driven power grid frequency prediction model is made interpretable by using mathematical formulas;
[0029] S4.6, based on the interpretable physical-information driven grid frequency prediction model obtained in S4.5, construct a loss function L taking into account the sensitive factor θ;
[0030] S4.7, based on sample set D I and the loss function in step S4.6, using a supervised learning algorithm to train the physical-information driven grid frequency prediction model;
[0031] S4.8, the trained physical-information driven grid frequency prediction model generates the final frequency prediction result f based on feature engineering X.
[0032] Furthermore, the method for determining the fusion mode of the frequency prediction physical model and the frequency prediction information model as a parallel mode in step S4.1 is:
[0033] S4.1.1, generating grid active disturbance or fault;
[0034] S4.1.2, the frequency prediction physical model in step S1 and the frequency prediction information model in step S2 respectively obtain the power grid state information and process them respectively;
[0035] S4.1.3, the frequency prediction physical model in step S1 and the frequency prediction information model in step S2 generate respective frequency prediction results;
[0036] S4.1.4, the frequency prediction results obtained by the frequency prediction physical model and the frequency prediction information model are input into the deep forest, and then the final frequency prediction result is obtained.
[0037] Furthermore, the mathematical formula for making the physical-information driven grid frequency prediction model interpretable in step S4.5 is:
[0038]
[0039]
[0040] Where W T+1 represents the weight of the output layer, b T+1 represents the bias of the output layer, represents the intermediate variable of the frequency prediction information model of the mth unit (a total of M units) in the Tth cascade layer, represents the intermediate variable of the frequency prediction physical model of the mth unit in the Tth cascade layer, represents the frequency prediction value of the frequency prediction information model of the T-th cascade layer, represents the frequency prediction value of the frequency prediction physical model of the T-th cascade layer, represents the weight of the mth unit in the T-1 cascade layer of the frequency prediction information model, represents the weight of the mth unit in the T-1 cascade layer of the frequency prediction physical model, represents the bias of the mth unit in the T-1 cascade layer of the frequency prediction information model, Represents the bias of the mth unit in the T-1 cascade layer of the frequency prediction physical model.
[0041] The advantages and beneficial effects of the present invention are:
[0042] 1. The physical model for frequency prediction of the power system constructed by the present invention takes into account the dynamic response characteristics of the frequency of new energy sources with frequency modulation capability in the transfer function, and derives a time domain analytical solution, thereby improving the level of refined modeling of the physical model for frequency prediction.
[0043] 2. The power system frequency prediction information model constructed by the present invention has feature engineering with low sample complexity and fast solution efficiency based on a tree structure, which can reduce the difficulty of data collection for users and reduce the burden on operators.
[0044] 3. The physical-information driven power grid frequency prediction method proposed in the present invention integrates the high stability of the frequency prediction physical model and the high solution speed and prediction accuracy of the frequency prediction information model in a parallel manner based on gcforest, so as to achieve the purpose of taking into account both prediction efficiency and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a flow chart of the physical-information driven power grid frequency prediction method taking into account the participation of new energy in frequency regulation according to the present invention;
[0046] Figure 2 A structure diagram of a frequency prediction physical model taking into account the participation of new energy in frequency modulation built by the present invention;
[0047] Figure 3 This is a structural diagram of the frequency prediction information model based on xgboost constructed by the present invention;
[0048] Figure 4 It is a schematic diagram of the fusion mode of the frequency prediction physical model and the frequency prediction information model of the present invention;
[0049] Figure 5 This is a structural diagram of the deep forest in the physical-information driven power grid frequency prediction model of the present invention. DETAILED DESCRIPTION
[0050] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.
[0051] like Figure 1 As shown in FIG. 1 , the physical-information driven grid frequency prediction method considering the participation of new energy in frequency regulation is innovative in that it includes the following steps:
[0052] S1: Construct frequency prediction physical model;
[0053] S2: Build a frequency prediction information model based on xgboost;
[0054] S3: collect data for the frequency prediction information model in S2 and train it;
[0055] S4: Construct a physical-information driven grid frequency prediction (PIFP) model based on the gcforest architecture;
[0056] S5: Collect data for the physical-information driven grid frequency prediction model in S4 and train it;
[0057] S6: After training, the physical-information driven grid frequency prediction model generates a dynamic grid frequency prediction result based on the grid active disturbance information.
[0058] To construct the frequency prediction physical model in step S1, it is necessary to first determine the grid inertia center frequency, then establish the transfer function of the grid frequency response frequency prediction physical model, and finally derive the time domain analytical solution. The specific steps are as follows:
[0059] (1) Grid inertia center frequency
[0060] The inertial center frequency f of the system COI The definition is as follows:
[0061]
[0062] In the formula, H i and w i denote the inertia time constant and rotor speed of generator i, respectively, N G is the number of grid-connected units in the system. The system dynamic frequency prediction referred to in this invention refers to the system inertia center frequency f COI .
[0063] (2) Grid frequency response frequency prediction physical model
[0064] The constructed power grid frequency response frequency prediction physical model is as follows: Figure 2 shown.
[0065] In the figure, ∆P represents the active power disturbance of the power grid, T R Represents the synchronous machine equivalent reheat time constant, F H represents the proportion of high-pressure steam turbine power generation to total power generation, R is the equivalent static adjustment coefficient of the synchronous machine, D represents the equivalent damping coefficient, K m is the mechanical power gain factor, ∆f represents the grid frequency deviation, RW is the equivalent static adjustment coefficient of the fan, T W represents the fan speed regulator time constant, and s represents the pull factor.
[0066] From this we can get the transfer function:
[0067]
[0068]
[0069] In the formula, G sys represents the transfer function of the physical model for system frequency prediction, H is the equivalent inertia time constant of the power grid, is the natural vibration frequency, Represents the damping ratio, R T It represents the equivalent regulation coefficient of the power grid. Since the time constants of different speed regulators have little effect on the system frequency drop within their allowable range, the time constants of all unit speed regulators in the formula can be the same value T R Alternative.
[0070] Then, perform the pair Performing the inverse Laplace transform, we get the expression of the frequency dynamic change ∆f(t) as follows:
[0071]
[0072] Where f0 represents the grid reference frequency, t represents the time after the grid disturbance, and e represents the natural logarithm.
[0073] The built frequency prediction physical model generates the frequency prediction result f according to the power grid disturbance ∆P SFR .
[0074] The construction of step S2 is based on the frequency prediction information model of xgboost. The structure diagram of xgboost is as follows Figure 3 shown.
[0075] (1) Problem setting
[0076] The power grid frequency prediction result f of the frequency prediction information model I It is expressed as:
[0077]
[0078] In the formula, h k is the kth function of the tree model, X I Represents the input features of the frequency prediction information model.
[0079] (2) Feature selection
[0080] The main purpose of frequency prediction is to provide the dispatching department with accurate grid frequency information to ensure the stability of grid frequency when disturbances occur. This goal can be achieved by incorporating frequency constraints into the optimization dispatching model. Therefore, the input features required by the proposed frequency prediction information model should not be too many. Assuming that the governor droop control of all generators remains unchanged, the greatest impact on the system frequency should be the size of the active power disturbance, the location of the system disturbance, and the start and stop status of the generator.
[0081] The active power disturbance amplitude is expressed as:
[0082]
[0083] Where P represents the maximum power generation feed-in loss, P g represents the g-th generator, with a total of N G Then the location where the active disturbance occurs is expressed as:
[0084]
[0085] In the formula, l represents the element index of the node where the generator is located, and argmax represents the element index position where the maximum value is found.
[0086] Then, according to the maximum power generation feed-in loss P and the location l where the active power disturbance occurs, the specific disturbance information I can be obtained as follows:
[0087]
[0088] Finally, the input feature X of the frequency prediction information model I It is expressed as:
[0089] X I = [ u 1 , … , u N G , I ]
[0090] In the formula, u represents the start and stop state of the generator. u = 1 indicates that the generator is in operation, and u = 0 indicates that the generator is in shutdown.
[0091] (3) Objective function
[0092] The objective function Obj of the frequency prediction information model is expressed as:
[0093]
[0094]
[0095] Where, L I represents the loss function of the frequency prediction information model, n represents the number of samples, there are N samples in total, f I,n Represents the prediction result of the frequency prediction information model of the nth sample, y n It represents the nth system frequency dynamic curve generated by time domain simulation. In order to obtain accurate system frequency curve prediction results, the mean square absolute error is used as the loss function, abs represents the absolute value, and Ω represents the regularization term to prevent the frequency prediction information model from overfitting during the training process. It is expressed as:
[0096]
[0097] Where M represents the number of leaves, 𝜔 m represents the mth vector of the leaf, 𝛾 and 𝜆 represent penalty factors.
[0098] The step S3 of collecting data for the frequency prediction information model in S2 and training it is as follows:
[0099] The data of the frequency prediction information model is based on the input feature X in step S2 (2) I The results were constructed by using PSS / E simulation software to build an IEEE 39-node model, and PSS / E was called by python for batch training. The active disturbance of the power grid was simulated by disconnecting the generator, and the load level of the data collection was between 50% and 100% of the original load level. A generator was tripped at every 0.25% load level, and a total of 2,000 sets of data were collected, of which 1,600 sets were used for training and 400 sets were used to verify the model training results.
[0100] The frequency prediction information model is trained using supervised learning and the objective function in (3) in step S2 is used for training.
[0101] The step S4 constructs a physical-information driven power grid frequency prediction model (PIFP) based on the gcforest architecture, and integrates the physical prediction model and the frequency prediction information model in step S1 and step S2 based on the deep forest architecture, and then trains them to achieve physical-information driven power grid frequency prediction. The specific steps are as follows:
[0102] (1) Integration of frequency prediction physical model and frequency prediction information model
[0103] The fusion method of frequency prediction physical model and frequency prediction information model is as follows: Figure 4 The specific steps are:
[0104] ① Generate active disturbance or fault in the power grid;
[0105] ② The frequency prediction physical model in step S1 and the frequency prediction information model in step S2 obtain the power grid state information respectively and process them respectively;
[0106] ③ The frequency prediction physical model in step S1 and the frequency prediction information model in step S2 generate respective frequency prediction results;
[0107] ④ Input the frequency prediction results obtained by the frequency prediction physical model and the frequency prediction information model into the deep forest, and then obtain the final frequency prediction result.
[0108] (2) Establishment of deep forest structure
[0109] The structure of deep forest is as follows Figure 5 As shown. The original gcforest algorithm is constructed by cascading random forests, which has poor generalization ability due to its purely data-driven characteristics. Based on the gcforest framework, the present invention combines the SFR frequency prediction physical model and the frequency prediction information model into a unit of gcforest (similar to neurons in deep learning) and connects them in a cascade manner, thereby improving the generalization ability and prediction accuracy of PIFP.
[0110] (3) Input features
[0111] The input feature of PIFP is expressed as:
[0112] X = [ X I , f I , f SFR ]
[0113] (4) Cascade layer determination
[0114] Each cascade layer of PIFP has a set of weights corresponding to the number of combined units, and the final frequency prediction result is obtained by taking the average. The number of cascade layers should be increased or decreased according to the preset indicators, as follows:
[0115]
[0116] In the formula, f(X n ) represents the frequency prediction result of PIFP.
[0117] Whenever the PIFP model expands to a new cascade layer, it determines whether the mean square error value calculated by the current layer is greater than the mean square error value of the previous cascade layer. If the requirement is met, the cascade layer will continue to be expanded, otherwise the expansion will stop.
[0118] (5) PIFP model interpretability
[0119] The PIFP frequency prediction process can be explained by mathematical expression:
[0120]
[0121]
[0122] Where W T+1 represents the weight of the output layer, b T+1 represents the bias of the output layer, represents the intermediate variable of the frequency prediction information model of the mth unit (a total of M units) in the Tth cascade layer, represents the intermediate variable of the frequency prediction physical model of the mth unit in the Tth cascade layer, represents the frequency prediction value of the frequency prediction information model of the T-th cascade layer, represents the frequency prediction value of the frequency prediction physical model of the T-th cascade layer, represents the weight of the mth unit in the T-1 cascade layer of the frequency prediction information model, represents the weight of the mth unit in the T-1 cascade layer of the frequency prediction physical model, represents the bias of the mth unit in the T-1 cascade layer of the frequency prediction information model, Represents the bias of the mth unit in the T-1 cascade layer of the frequency prediction physical model.
[0123] (6) PIFP loss function
[0124] During the operation of the power system, the number of large-capacity active disturbance events is usually much smaller than the number of small disturbance events, which makes the distribution of the maximum frequency offset tend to present a band distribution with steep edges and higher left side than right side. This will cause PIFP to pay too much attention to the stability of frequency prediction in the case of small disturbances, while the prediction accuracy is low due to insufficient training samples in the case of large disturbances. Therefore, in order to balance the impact of training samples on prediction accuracy, a sensitivity factor θ is embedded in PIFP, which is specifically expressed as:
[0125]
[0126] In the formula, α is a compensation factor used to prevent the gradient vanishing problem of frequency-stable samples due to their high probability, p is the probability corresponding to the large disturbance sample, and f is UFLS Indicates the under-frequency load shedding (UFLS) threshold.
[0127] In the embedded sensitivity factor θ, the penalty factor λ is added to further ensure the frequency prediction accuracy of PIFP around UFLS. The final loss function L of PIFP is expressed as:
[0128]
[0129] In the formula, f t (X n ) represents the frequency prediction result of the nth sample of PIFP at time t, y n,t Represents the frequency of the nth sample generated by time domain simulation at time t.
[0130] The data used for the physical-information driven power grid frequency prediction model in step S4 and training are specifically performed in the following steps:
[0131] The data collected in step S3 is combined with the input features in step S4 (3), and the data collection method mentioned in step S3 is used as a benchmark. The amount of data collected is the same as that in step S3, and the training data of PIFP is obtained.
[0132] Based on the obtained PIFP training data, a supervised learning training method is adopted, and the loss function in (6) in step S4 is used for training.
[0133] In step S6, the trained physical-information driven grid frequency prediction model generates a grid frequency dynamic prediction result based on the grid active disturbance information.
[0134] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.
Claims
1. A physical-information driven grid frequency prediction method taking into account the participation of new energy in frequency modulation, characterized by: The steps include: S1, build frequency prediction physical model; S2, build a frequency prediction information model based on xgboost; S3, collecting training data, including the active disturbance size, active disturbance location, and frequency dynamic change data of the power grid, and using the collected training data to train the frequency prediction information model to obtain a frequency prediction result of the frequency prediction information model; S4, integrate the frequency prediction physical model and the frequency prediction information model to build a physical-information driven power grid frequency prediction model based on the gcforest architecture; S5, collecting sample data, and using the collected sample data to train a physical-information driven power grid frequency prediction model; S6, generating a dynamic prediction result of the power grid frequency based on the active power disturbance information of the power grid for the trained physical-information driven power grid frequency prediction model; The method S1 for constructing a frequency prediction physical model comprises the following steps: S1.1, build a physical model for power system frequency prediction taking into account the participation of new energy in frequency regulation; S1.2, obtaining a time domain analytical solution of the frequency prediction physical model according to the power system frequency prediction physical model constructed in S1.1; S1.3, based on the time domain analytical solution in S1.2, generate the frequency prediction result f of the physical model of the power grid after the disturbance according to the active power disturbance ΔP SFR , and use the result as the physical driving frequency prediction data; In the physical model of power system frequency prediction taking into account the participation of new energy in frequency regulation, ΔP represents the active power disturbance of the power grid, T R Represents the synchronous machine equivalent reheat time constant, F H It indicates the proportion of high-pressure steam turbine power generation to total power generation, R is the equivalent static adjustment coefficient of synchronous machine, D is the equivalent damping coefficient, K m is the mechanical power gain factor, Δf represents the grid frequency deviation, R W is the equivalent static adjustment coefficient of the fan, T W represents the fan speed regulator time constant, s represents the pull factor, The transfer function of the frequency prediction physical model is: In the formula, G sys represents the transfer function of the physical model of system frequency prediction, H is the equivalent inertia time constant of the power grid, ω n is the natural vibration frequency, ζ is the damping ratio, R T It represents the equivalent regulation coefficient of the power grid. Since the time constants of different speed regulators have little effect on the system frequency drop within their allowable range, the time constants of all unit speed regulators in the formula can be the same value T R Replacement; Then, the transfer function of the system frequency prediction physical model is subjected to Laplace inverse transformation, and the expression of the frequency dynamic change Δf is obtained as follows: In the formula, f0 represents the grid reference frequency, t represents the time after the grid disturbance, and e represents the natural logarithm; Finally, the Δf generated by the transfer function of the system frequency prediction physical model is transformed into the frequency prediction result f of the physical model of frequency prediction after the power grid is disturbed. SFR ; The method S2 for constructing a frequency prediction information model based on xgboost comprises the following steps: S2.1, construct the frequency prediction information model structure based on xgboost and describe the frequency prediction problem; S2.2, according to the model structure and frequency prediction problem description constructed in step S2.1, analyze and screen the key features that affect the dynamic changes of power grid frequency, and construct feature engineering as the input feature X of the frequency prediction information model I ; S2.3, based on the input feature X analyzed in step S2.2 I , construct the objective function based on the xgboost algorithm, including the loss function L I and regularization term Ω, which is used for supervised learning training of the frequency prediction information model in step S3; S2.4, using the input feature X in S2.2 I , taking the real frequency y of the system as the label data, forming the sample set D I ={X I ,y}, used for supervised learning training of the frequency prediction information model in step S3; The sample data includes training data, frequency prediction results of a frequency prediction physical model, and frequency prediction results of a frequency prediction information model; The input feature X of the frequency prediction information model I The key features that have a great impact on the dynamic changes of grid frequency include the system's maximum active feed-in loss P, the system's active disturbance position l, and the operating status u of the generator set; The method S4 constructs a physical-information driven power grid frequency prediction model based on the gcforest architecture, comprising the following steps: S4.1, determining that the fusion mode of the frequency prediction physical model and the frequency prediction information model is parallel; S4.2, based on the parallel fusion method of step S4.1, a physical-information driven power grid frequency prediction model is constructed using a cascaded gcforest architecture; S4.3, based on the physical-information driven power grid frequency prediction model in S4.2, construct feature engineering X, including input features X of the frequency prediction information model I , the frequency prediction result f of the frequency prediction information model I , the frequency prediction result f of the frequency prediction physical model SFR ; S4.4, based on the feature engineering X constructed in S4.3, the mean square error MSE evaluation standard is used to determine the number of cascade layers of physical-information driven power grid frequency prediction, and the physical-information driven power grid frequency prediction model structure is obtained; S4.5, based on the physical-information driven power grid frequency prediction model structure finally determined in S4.4, the physical-information driven power grid frequency prediction model is made interpretable by using mathematical formulas; S4.6, based on the interpretable physical-information driven grid frequency prediction model obtained in S4.5, construct a loss function L taking into account the sensitive factor θ; S4.7, based on sample set D I and the loss function in step S4.6, using a supervised learning algorithm to train the physical-information driven grid frequency prediction model; S4.8, the trained physical-information driven power grid frequency prediction model generates the final frequency prediction result f based on feature engineering X; The method for determining the fusion mode of the frequency prediction physical model and the frequency prediction information model as a parallel mode in step S4.1 is: S4.1.1, generating grid active disturbance or fault; S4.1.2, the frequency prediction physical model in step S1 and the frequency prediction information model in step S2 respectively obtain the power grid state information and process them respectively; S4.1.3, the frequency prediction physical model in step S1 and the frequency prediction information model in step S2 generate respective frequency prediction results; S4.1.4, inputting the frequency prediction results obtained by the frequency prediction physical model and the frequency prediction information model into the deep forest, and then obtaining the final frequency prediction result; The mathematical formula for making the physical-information driven grid frequency prediction model interpretable in step S4.5 is: Where W T+1 represents the weight of the output layer, b T+1 represents the bias of the output layer, represents the intermediate variable of the frequency prediction information model of the mth unit in the Tth cascade layer, M is the total number of units, represents the intermediate variable of the frequency prediction physical model of the mth unit in the Tth cascade layer, represents the frequency prediction value of the frequency prediction information model of the T-th cascade layer, represents the frequency prediction value of the frequency prediction physical model of the T-th cascade layer, represents the weight of the mth unit of the T-1 cascade layer of the frequency prediction information model, represents the weight of the mth unit in the T-1 cascade layer of the frequency prediction physical model, represents the bias of the mth unit in the T-1 cascade layer of the frequency prediction information model, Represents the bias of the mth unit in the T-1 cascade layer of the frequency prediction physical model.
