FNP-DBN fusion driving-based power system frequency lowest point prediction method

Abstract

The invention belongs to the field of power system frequency stability analysis and control, and relates to a power system frequency lowest point prediction method based on FNP-DBN fusion driving. Comprising the following steps that a serial fusion mode is adopted, a physical driving link and a data driving link are combined for construction, the physical driving link is based on an FNP analysis model and used for conducting rapid initial prediction on the lowest point of the frequency of the power system, and an initial prediction value is obtained; the data driving link is based on a deep belief network (DBN) algorithm and is used for performing error correction on the initial predicted value to obtain a final predicted value; determining input characteristic quantities of the DBN algorithm, wherein the input characteristic quantities comprise a system equivalent inertia constant, active power vacancy, generator capacity, unit active output before disturbance, a reheating time constant, a high-pressure cylinder power coefficient, a damping time constant and a transient slip coefficient; and carrying out framework modeling on the FNP-DBN fusion drive, wherein the framework modeling comprises an offline training stage and an online application stage.

Description

Technical Field

[0001] This invention belongs to the field of power system frequency stability analysis and control, and relates to a power system frequency minimum point prediction method based on FNP-DBN fusion drive. Background Technology

[0002] The frequency minimum point encompasses the maximum frequency offset and its corresponding occurrence time. Achieving rapid and accurate prediction of the frequency minimum point under high power shortage scenarios not only helps in timely assessment of the system's frequency stability after a disturbance occurs, but also provides crucial information for formulating emergency frequency control strategies. This is of significant practical importance for ensuring the safe and stable operation of the power system and preventing frequency instability and even system collapse. Because it serves the dual purpose of being a key indicator for situational awareness and providing decision support for emergency control, the frequency minimum point has received widespread attention in the field of frequency stability analysis and control. Currently, methods for predicting the frequency minimum point under power shortage scenarios caused by large disturbances can be mainly categorized into three types: physical-driven methods, data-driven methods, and fusion-driven methods.

[0003] Physics-driven methods mainly include time-domain simulation and mathematical analytical methods. Time-domain simulation can achieve high computational accuracy when accurate models and parameters are available; however, as the system scales up, the order of the detailed physical model increases dramatically, significantly reducing its computational efficiency. Therefore, this method is typically suitable for offline analysis scenarios with high accuracy requirements. In contrast, mathematical analytical methods, based on the assumption of uniform network frequency, ignore the dynamic effects of the network and achieve rapid computation by deriving the analytical expression for the lowest frequency point. However, because this method involves significant aggregation and order reduction of the original model, its computational accuracy is generally lower, and it is mostly used for online analysis scenarios with high real-time requirements. In summary, there is a clear trade-off between computational accuracy and computational speed in physics-driven methods.

[0004] Data-driven methods, based on various machine learning and deep learning algorithms, learn the mapping relationship between input and output from historical or simulation data, thus completely eliminating reliance on explicit physical equations. When dealing with complex operating scenarios, this method demonstrates significant advantages in prediction speed and accuracy. Benefiting from the ever-accumulating massive amounts of measurement data in power systems, these methods typically first extract key features from historical data through offline training to build a prediction model; in the application phase, real-time data is input into the trained model to achieve rapid and accurate prediction of the lowest frequency point. However, data-driven methods inherently rely on fitting the relationships between different features using a large amount of sample data. Their predictive performance is largely limited by the scale and quality of the training data, and they are prone to weak generalization in scenarios with insufficient sample coverage. Furthermore, because they are completely detached from the constraints of physical laws, their prediction results often lack clear physical interpretation and have poor interpretability. These inherent limitations of the model restrict the widespread application of data-driven methods in frequency stability analysis.

[0005] Regarding fusion-driven methods, existing research has attempted to combine the advantages of both physics-driven and data-driven approaches to improve the performance of frequency stability analysis. Studies have shown that fusion models typically outperform single-type models in terms of prediction accuracy and robustness, effectively achieving complementary advantages. However, how to rationally select physical and data models and achieve efficient fusion between them remains a key challenge in current research. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a power system frequency minimum point prediction method based on FNP-DBN fusion, achieving efficient and high-precision prediction of the frequency minimum point. In this method, the FNP analytical model serves as the physical driving element, deriving the analytical expression for the frequency minimum point to achieve rapid initial prediction; the DBN algorithm, as the data driving element, is responsible for online correction of the initial prediction results, thereby effectively improving the accuracy of the final output. By organically combining the two through a serial fusion architecture, a frequency minimum point prediction method with both fast response and accurate output is formed.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a power system frequency minimum point prediction method based on FNP-DBN fusion drive is provided, comprising the following steps: Step 1: A serial fusion mode is adopted to combine the physical driving link and the data driving link. The physical driving link is based on the FNP analytical model and is used to quickly make an initial prediction of the lowest frequency point of the power system to obtain the initial prediction value. The data driving link is based on the Deep Belief Network (DBN) algorithm and is used to correct the error of the initial prediction value to obtain the final prediction value. Step 2: Determine the input features of the DBN algorithm, including: system equivalent inertia constant, active power deficit, generator capacity. Quantities, active power output of the unit before disturbance, reheat time constant, high-pressure cylinder power coefficient, damping time constant and transient slip coefficient; Step 3: Model the framework for FNP-DBN fusion-driven training, including offline training and online application phases; Furthermore, in step 3, during the offline training phase, generator parameters and disturbance power deficit are first extracted from historical operating data. P d And the actual lowest frequency data, including the maximum frequency deviation Δ f max and corresponding time t nadir If historical data is insufficient, simulation data generated using an average system frequency model can be used as a supplement, together forming a training sample set covering multiple operating scenarios; an FNP analytical model is constructed based on generator parameters, and... P d As its input, the initial frequency minimum point is predicted, i.e. (Δ f max,FNP , t nadir,FNP ); The predicted value of the lowest frequency point output by the FNP analytical model (Δ) f max,FNP , t nadir,FNP As input features, the actual lowest frequency point (Δ) f max , t nadir The input labels are used as the output labels to form the training samples for the DBN algorithm. All data are normalized and then input into the DBN algorithm. The training process includes two stages: unsupervised pre-training and supervised fine-tuning, in order to gradually optimize the mapping relationship between input and output. After training, the resulting DBN algorithm will be used in the online prediction stage.

[0008] Furthermore, in step 3, when an actual disturbance occurs, the disturbance power deficit is acquired and calculated online through a wide-area measurement system, and input into the FNP analytical model to obtain an initial prediction result for the frequency minimum point. This result is then input into a fully trained DBN algorithm, and the final high-precision frequency minimum point prediction value is obtained through data-driven correction. Based on this prediction result, the dynamic frequency situation of the system can be accurately assessed, and a reliable basis can be provided for the formulation of frequency emergency control strategies.

[0009] It should be noted that if the system's operating scenario changes significantly, the relevant parameters of the FNP parsing model need to be updated in a timely manner to maintain its applicability. To improve the model's generalization ability across multiple scenarios, the DBN algorithm input and output data obtained during the online application phase can be accumulated as historical samples for incremental training and optimization of the model in subsequent iterations, thereby continuously improving the scale and quality of the training data.

[0010] The beneficial effects of this invention are as follows: Compared with existing technologies, the proposed FNP-DBN fusion-driven power system frequency minimum point prediction method uses the FNP analytical model as the physical driving link, achieving rapid initial prediction by deriving the analytical expression of the frequency minimum point; and uses the DBN algorithm as the data driving link, responsible for online correction of the initial prediction results, thereby effectively improving the accuracy of the final output. By organically combining the two through a serial fusion architecture, a frequency minimum point prediction method with both fast response and accurate output is formed. Specifically, compared with a single physical driving method, this method significantly improves prediction accuracy; compared with a single data driving method, this method effectively enhances the generalization ability to different operating scenarios; and compared with existing fusion-driven methods, this method retains more comprehensive key physical influencing factors and significantly reduces the dependence on the data driving link in terms of sample quantity and feature dimensions. In summary, this method effectively overcomes the limitations of traditional physical driving and data driving methods, and can provide a more reliable technical basis for power system frequency stability analysis and control decisions. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 It is a modeling framework for the FNP-DBN fusion-driven method; Figure 2 This is a topology diagram of the New England 39-node system; Figure 3 It is the maximum frequency deviation Δ f max A comparison chart of evaluation indicators; Figure 4 The time corresponding to the maximum frequency deviation t nadirA comparison chart of evaluation indicators. Detailed Implementation

[0014] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.

[0015] refer to Figure 1-4 This invention provides a method for predicting the lowest frequency point in a power system based on FNP-DBN fusion drive, comprising the following steps: Step 1, Selection of Physical-Data Fusion Modeling Mode and Driving Method: Step 1-1, Selection of Physical-Data Fusion Modeling Mode: To meet the application needs of different scenarios in power systems, existing research has proposed four physical-data fusion modeling modes. Specifically, a serial fusion mode is proposed to address the issue of a high degree of simplification in the physical model; a parallel fusion mode is designed to address the differences in modeling focus between the physical model and the data model; an embedded fusion mode is proposed to address the complex problem of some mechanisms in the physical model being unknown or difficult to describe accurately; and a feedback fusion mode is constructed to cope with the difficulty in identifying physical model parameters.

[0016] This paper takes the prediction of the lowest frequency point of a power system after a disturbance as its research background. This scenario requires extremely high computational speed in real-time applications, so only simplified physical models based on mathematical analytical methods can usually be used, resulting in poor accuracy of the prediction results. At the same time, the massive frequency measurement data accumulated in the power system often contains a large number of invalid or redundant features, which significantly increases the difficulty of modeling data-driven models.

[0017] To address the aforementioned issues, this study employs a serial fusion mode to achieve rapid and accurate prediction of the lowest frequency point under high-power deficiency scenarios. The specific process is as follows: First, based on mathematical analytical methods, the analytical expression for the lowest frequency point in the simplified physical model is derived to quickly obtain initial predicted values. Then, using deep learning algorithms, the correlation and error patterns between the initial predicted values ​​and actual values ​​under different operating conditions are explored. The data model is then used to perform online correction of the initial prediction results, thereby obtaining a more accurate final prediction result.

[0018] Step 1-2, Selection of physical driving method: Regarding the selection of physical driving methods, the Average System Frequency (ASF) model, by assuming synchronous changes in system frequency, equates the rotor motion equations of the entire generator system to a single-machine model, and considers the dynamic response characteristics of each unit's prime mover-governor. This model significantly reduces model complexity while retaining key influencing factors of frequency dynamics, and has become the basic framework for most current simplified physical models. However, the ASF model suffers from an increasing number of order random groups and difficulty in analytically solving the frequency dynamic process, thus limiting its computational efficiency in large-scale systems.

[0019] To address the aforementioned issues, the FNP analytical model employs binomial fitting of response characteristics, frequency feedback loop decomposition, and parabolic frequency difference simulation to derive an analytical expression for the minimum frequency point, significantly improving prediction speed while maintaining a certain level of accuracy. Therefore, this invention selects the FNP analytical model as the physics-driven component to ensure the speed of the fusion-driven method in real-time applications.

[0020] Steps 1-3, Selection of data-driven methods: Regarding the selection of data-driven methods, with the rapid development of artificial intelligence technology, various machine learning and deep learning algorithms, including backpropagation neural networks, extreme learning machines, support vector machines, long short-term memory networks, and gated recurrent units, have been widely studied in power system modeling and prediction. The core of data-driven methods lies in mining the mapping relationship between input variables (such as generator parameters and disturbance scale) and output targets (the lowest frequency point) from frequency simulation data. Generally speaking, the deeper the network structure of the data model, the stronger its representational ability and the better its prediction performance. However, this also leads to higher requirements for the number and quality of training samples, as well as a significant increase in training complexity. DBN, as a typical deep learning algorithm, can effectively extract deep features from data and demonstrates superior performance in time series modeling tasks such as oil price prediction and load prediction. Therefore, this invention selects DBN as the data-driven component to correct the initial prediction error of the FNP analytical model, thereby improving the overall prediction accuracy of the fusion method.

[0021] In summary, this invention constructs an FNP-DBN fusion-driven method based on the FNP analytical model and the DBN algorithm, employing a serial fusion mode. Within this framework, the FNP analytical model provides fast and interpretable initial predictions, effectively reducing the DBN algorithm's dependence on sample data size and feature dimensions; while the DBN algorithm can perform high-precision correction on the initial prediction results in a very short time. The combination of these two methods significantly improves the prediction accuracy of the lowest frequency points while ensuring prediction efficiency and interpretability.

[0022] Step 2, Selection of input features for the DBN algorithm: The predictive performance of deep learning methods largely depends on the quantity and quality of sample data. Therefore, feature selection plays a crucial role in improving the prediction accuracy of DBN algorithms.

[0023] When applying the DBN algorithm to a purely data-driven approach, it is necessary to combine the dynamic process of frequency response, systematically analyze the key factors affecting the lowest point of the system frequency, and select features accordingly. The fundamental reason for the dynamic change of system frequency lies in the difference between electromagnetic torque and mechanical torque caused by the imbalance of active power, which in turn causes changes in generator speed. Therefore, the active power deficit of the system is the primary factor affecting the dynamic change of frequency.

[0024] In the initial stage when the power system experiences an active power deficit but the speed governor has not yet activated, the system inertia converts the kinetic energy stored in rotating equipment into mechanical power to suppress the rate of frequency decline. This indicates that the system's equivalent inertial constant has a significant impact on the dynamic frequency response.

[0025] Conventional generator sets are generally equipped with primary frequency regulation capabilities. When the system frequency drops due to disturbances, the generator sets increase active power output according to the frequency deviation to restore balance. The installed capacity and active power output of each generator in the system before the disturbance directly reflect its primary frequency regulation capability. Therefore, generator capacity and active power output should also be considered key influencing factors of dynamic frequency changes.

[0026] The primary frequency regulation capability of a generator set originates from the governor system. When the frequency drops, the governor adjusts the valve opening or guide vane position by detecting the frequency difference signal, thereby changing the prime mover's power output. Therefore, the key parameters of the governor have a significant impact on the frequency dynamic process. Specifically, for thermal power units, the governor droop coefficient, reheat time constant, and high-pressure cylinder power coefficient are selected as characteristics; for hydropower units, the damping time constant and transient slip coefficient are the main considerations.

[0027] Based on the above analysis of the dynamic frequency change mechanism, and considering key factors such as system inertia, active power deficit, unit primary frequency regulation capability, and governor parameters, the input characteristic quantities of the DBN algorithm are finally determined to include: system equivalent inertia constant, active power deficit, generator unit capacity, unit active power output before disturbance, reheat time constant, high-pressure cylinder power coefficient, damping time constant, and transient slip coefficient.

[0028] However, in the fusion-driven framework proposed in this paper, since the FNP analytical model has fully considered the key physical factors affecting the minimum frequency point, the DBN algorithm only needs to correct the error of the initial prediction results of the FNP analytical model. Therefore, in this fusion mode, the input features of the DBN algorithm can be simplified to the predicted value of the minimum frequency point calculated by the FNP analytical model. This design significantly reduces the complexity of the data model and its dependence on features, while improving the interpretability and engineering applicability of the overall method.

[0029] Step 3, FNP-DBN fusion modeling: The modeling framework of the FNP-DBN fusion-driven method is as follows: Figure 1 As shown, the overall process can be divided into two main stages: offline training and online application.

[0030] Step 3-1, Offline Training Phase: During the offline training phase, generator parameters and disturbance power deficit are first extracted from historical operating data. P d And the actual lowest frequency data (including the maximum frequency deviation Δ) f max and corresponding time t nadir If historical data is insufficient, simulation data can be generated using an average system frequency model as a supplement, together forming a training sample set covering multiple operating scenarios. An FNP analytical model is constructed based on generator parameters, and... P d As its input, the initial frequency minimum point (Δ) is predicted. f max,FNP and t nadir,FNP ).

[0031] The predicted value of the lowest frequency point output by the FNP analytical model (Δ) f max,FNP , t nadir,FNP As input features, the actual lowest frequency point (Δ) f max , t nadir The input and output labels are used as training samples for the DBN algorithm. All data is normalized before being input into the DBN algorithm. The training process includes two stages: unsupervised pre-training and supervised fine-tuning, to progressively optimize the mapping relationship between input and output. After training, the resulting DBN algorithm will be used in the online prediction stage.

[0032] Step 3-2, Online Application Stage: When an actual disturbance occurs, the disturbance power deficit can be acquired and calculated online through a wide-area measurement system. This deficit is then input into the FNP analytical model to obtain an initial prediction of the frequency minimum point. This result is subsequently input into a fully trained DBN algorithm, and data-driven corrections are used to obtain the final, high-precision prediction of the frequency minimum point. Based on this prediction, the dynamic frequency situation of the system can be accurately assessed, providing a reliable basis for the formulation of emergency frequency control strategies.

[0033] It should be noted that if the system's operating scenario changes significantly, the relevant parameters of the FNP parsing model need to be updated in a timely manner to maintain its applicability. To improve the model's generalization ability across multiple scenarios, the DBN algorithm input and output data obtained during the online application phase can be accumulated as historical samples for incremental training and optimization of the model in subsequent iterations, thereby continuously improving the scale and quality of the training data.

[0034] This study selects the Frequency Nadir Prediction (FNP) model as the physics-driven component. This model retains more key physical factors affecting frequency dynamics, thus significantly reducing the dependence of the data-driven component on the number, quality, and feature dimensions of samples, which helps improve the model's generalization ability and interpretability. Simultaneously, a Deep Belief Network (DBN) is introduced as a data-driven component to correct errors in the initial output of the frequency minimum prediction model, further improving the overall prediction accuracy. Based on this, a fusion-driven model based on FNP-DBN is constructed using a serial fusion mode. This model can achieve fast and accurate prediction of the frequency minimum, providing a more reliable basis for power system frequency stability analysis and control strategy formulation.

[0035] To verify the effectiveness and accuracy of the proposed FNP-DBN fusion-driven method in predicting the lowest frequency point, the New England 39-node system was selected as a test case, and the Monte Carlo method was used to generate the sample set required for model training and testing. To comprehensively evaluate the performance of the proposed model, single physics-driven methods (FNP analytical model), single data-driven methods (DBN algorithm), and existing fusion-driven methods (FNP-LSTM method) were used as benchmarks. Multiple sets of comparative experiments were conducted to systematically analyze the overall performance of the proposed method in terms of prediction accuracy and generalization ability.

[0036] Step 1, Introduction to the Case Study System: The New England 39-node system has a base frequency of 60 Hz, and its topology is as follows: Figure 2As shown in the figure, the system comprises 10 generators, 39 busbars, 19 loads, and 34 transmission lines. The dynamic model of this invention is constructed in the MATLAB / Simulink simulation platform. Generators G1 to G9 are thermal power units equipped with IEEE G1 type speed governors; G10 is a hydropower unit equipped with an IEEE G3 type speed governor. Specific parameters of each speed governor are shown in Table 1.

[0037] Table 1. Governor parameters for the New England 39-node system

[0038] Step 2, Data Sample Generation: To construct a high-quality sample set that meets the requirements of data-driven modeling, it is necessary to ensure that the sample space possesses basic characteristics such as comprehensive coverage, high distribution density, and good randomness. Therefore, this paper employs a combination of random sampling of system parameters and iterative calls to Simulink simulations to generate data samples in batches covering diverse operational scenarios, thus supporting the training and validation of the data model.

[0039] For the New England 39-node system, this paper constructs its ASF model as a simulation platform and simulates the changes in system operating state in the following ways: adjusting the system inertial time constant to simulate different inertial levels; changing the system active power deficit to simulate different disturbance intensities; and adjusting the key parameters of the speed governor to reflect the differences in the system's frequency regulation capability.

[0040] The frequency dynamic response of a power system after a disturbance is mainly affected by the disturbance scale, the system inertia level, and the frequency regulation capability of the governor. To improve the representativeness and coverage of the sample, it is necessary to reasonably set the variation range of each parameter so that the generated sample can comprehensively reflect the frequency response characteristics of the system under various typical and extreme operating conditions while maintaining randomness.

[0041] Based on theoretical analysis and engineering experience, and taking into account the influence of system inertia, power deficit, and speed controller parameters on frequency dynamics, reasonable variation ranges for each parameter were determined, as shown in Table 2.

[0042] Table 2 Random sampling intervals for parameters

[0043] After setting the sampling range, the program iteratively calls Simulink to perform simulation calculations. In each loop, a set of parameter values ​​is randomly selected from each parameter range for simulation, and the corresponding input features and output labels are extracted to form a sample. A total of 10,000 loops are executed to generate a total sample library, which is then proportionally divided into 7,000 training samples, 2,000 validation samples, and 1,000 test samples to ensure the statistical validity of model training and evaluation.

[0044] Step 3, Predictive performance evaluation and analysis: A systematic comparative analysis of four methods—FNP, DBN, FNP-DBN, and FNP-LSTM—was conducted from the two dimensions of prediction speed and prediction accuracy.

[0045] Step 3-1, Prediction speed comparison: The average prediction times for the FNP method, DBN method, FNP-DBN method, and FNP-LSTM method are 1.30 ms, 10.45 ms, 11.31 ms, and 16.70 ms, respectively. It can be seen that the FNP method, based on the physics-driven approach, has the fastest computation speed due to its simplified analytical form. Although the FNP-DBN method proposed in this invention is slightly slower than the pure DBN method, it still fully meets the real-time requirements of online prediction, and its prediction efficiency is superior to the FNP-LSTM method.

[0046] Step 3-2, Prediction Accuracy Comparison: To comprehensively evaluate the prediction accuracy of each method, Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE) were selected as evaluation metrics. Figure 3 and Figure 4 As shown, the FNP-DBN method has a significantly lower error in predicting the lowest frequency point than other comparative methods, indicating that it has higher prediction accuracy and better stability.

[0047] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these changes and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the lowest frequency point in a power system based on FNP-DBN fusion drive, characterized in that, Includes the following steps: Step 1: A serial fusion mode is adopted to combine the physical driving link and the data driving link. The physical driving link is based on the FNP analytical model and is used to quickly make an initial prediction of the lowest frequency point of the power system to obtain the initial prediction value. The data driving link is based on the Deep Belief Network (DBN) algorithm and is used to correct the error of the initial prediction value to obtain the final prediction value. Step 2: Determine the input features of the DBN algorithm, including: system equivalent inertia constant, active power deficit, generator capacity, active power output of the unit before disturbance, reheat time constant, high-pressure cylinder power coefficient, damping time constant, and transient slip coefficient; Step 3: Model the framework for FNP-DBN fusion-driven training, including offline training and online application phases.

2. The power system frequency minimum point prediction method based on FNP-DBN fusion drive as described in claim 1, characterized in that, In step 3, during the offline training phase, generator parameters and disturbance power deficit are first extracted from historical operating data. P d And the actual lowest frequency data, including the maximum frequency deviation Δ f max and corresponding time t nadir If historical data is insufficient, simulation data can be generated using an average system frequency model as a supplement, together forming a training sample set covering multiple operating scenarios. An analytical model of FNP is constructed based on generator parameters, and... P d As its input, the initial frequency minimum point is predicted, i.e. (Δ f max,FNP , t nadir,FNP ); The predicted value of the lowest frequency point output by the FNP analytical model (Δ) f max,FNP , t nadir,FNP As input features, the actual lowest frequency point (Δ) f max , t nadir The input labels are used as output labels to form the training samples for the DBN algorithm. All data are normalized and then input into the DBN algorithm. The training process includes two stages: unsupervised pre-training and supervised fine-tuning, in order to gradually optimize the mapping relationship between input and output. After training, the resulting DBN algorithm will be used in the online prediction phase.

3. The power system frequency minimum point prediction method based on FNP-DBN fusion drive as described in claim 1, characterized in that, In step 3, when an actual disturbance occurs, the disturbance power deficit is acquired and calculated online through a wide-area measurement system and input into the FNP analytical model to obtain the initial prediction result of the frequency minimum point. This result is then input into the fully trained DBN algorithm, and the final high-precision frequency minimum point prediction value is obtained through data-driven correction.

Images