New energy power data generation method based on pre-trained physical constraint model

Abstract

The application discloses a new energy power data generation method based on a pre-training physical constraint model, relates to the technical field of new energy power generation data generation, and comprises the following steps: acquiring meteorological feature vectors and physical structure vectors of a target new energy station; inputting a sequence generation model to generate a basic power time sequence, combining a physical constraint model to calculate a physical model power value, and completing effectiveness determination through physical consistency error; constructing extreme scene disturbance parameters based on the physical structure vectors, applying disturbance to the meteorological feature vectors to generate extreme disturbance meteorological feature vectors, and generating an extreme scene power time sequence; combining extreme physical model power values to complete extreme scene effectiveness determination; and merging the determined basic power time sequence and the extreme scene power time sequence to output a pre-training data set. The application solves the technical problem that existing data generation methods cannot consider sample diversity and physical consistency, and especially cannot generate samples with insufficient authenticity under extreme weather scenes.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation data technology, specifically to a method for generating new energy power data based on a pre-trained physical constraint model. Background Technology

[0002] The output of renewable energy power generation is highly sensitive to meteorological conditions, and high-quality training data is crucial for improving the performance of power prediction models. However, historical data from renewable energy power plants suffers from problems such as a scarcity of extreme weather samples and uneven distribution across different regions, resulting in insufficient training data and limiting the generalization ability of prediction models.

[0003] Existing data generation methods mainly rely on generative adversarial networks or variational autoencoders for sample expansion, but they typically only focus on the similarity of data distribution and lack constraints on the physical mechanisms of new energy power generation, easily generating samples that do not conform to actual physical laws. Some studies have attempted to introduce physical information neural networks, such as CN116014757B which uses neural networks for frequency control of new energy power plants, and CN121503819A which applies power boundary constraints at the output of the decoding network. However, the physical constraints in the above methods are all passively applied to the output, without integrating the physical model into the intermediate stages of the generation process, and without designing a dedicated active construction and screening mechanism for the scarcity of extreme weather samples, making it difficult to guarantee the physical rationality and diversity of the generated data under extreme operating conditions. Summary of the Invention

[0004] This invention provides a new energy power data generation method based on a pre-trained physical constraint model to solve the problems of lack of full-process physical constraints, easy deviation of generated samples from the actual power generation mechanism, and insufficient ability to construct extreme operating condition samples in the existing new energy power data generation process, resulting in poor physical rationality of generated data, low scene coverage, and limited model generalization ability.

[0005] In view of the above problems, the present invention provides a method for generating new energy power data based on a pre-trained physical constraint model, the method comprising: Obtain the meteorological feature vector and physical structure vector of the target new energy power station; The meteorological feature vector and the physical structure vector are input into the sequence generation model to generate the basic power time series. The physical constraint model corresponding to the power generation type of the target new energy power station is invoked, and the power value of the physical model is calculated based on the meteorological feature vector and the physical structure vector. The validity of the basic power time series is determined based on the physical consistency error between the basic power time series and the power value of the physical model. The perturbation parameters of the extreme scenario perturbation are determined based on the physical structure vector, and the extreme scenario perturbation is applied to the meteorological feature vector based on the perturbation parameters to obtain the extreme perturbation meteorological feature vector. The extreme disturbance meteorological feature vector is input into the sequence generation model to generate a power time series of extreme scenarios; The physical constraint model is invoked, and the extreme physical model power value is calculated based on the extreme disturbance meteorological feature vector and the physical structure vector. The validity of the extreme scenario power time series is determined based on the extreme physical consistency error between the extreme scenario power time series and the extreme physical model power value. The base power time series and the extreme scenario power time series that have passed the validity determination are merged to output a pre-training dataset.

[0006] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention generates a basic power time series by inputting meteorological feature vectors and physical structure vectors into a sequence generation model, and performs physical consistency verification on the generated results using a physical constraint model. This achieves full-process physical constraints in the data generation process, effectively preventing the generated samples from deviating from the actual power generation mechanism of new energy sources, and improving the authenticity and physical rationality of the generated data. Simultaneously, by constructing extreme scenario perturbation parameters and generating extreme scenario power time series, the training samples under extreme conditions are expanded, enhancing the scenario coverage and diversity of the generated data. Furthermore, by jointly screening and fusing samples from both normal and extreme operating conditions, a high-quality pre-training dataset is constructed, thereby improving the generalization ability, stability, and robustness of the new energy power prediction model under complex environments and extreme weather conditions. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the process for generating new energy power data based on a pre-trained physical constraint model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the recursive generation structure of the sequence generation model provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the physical constraint model calculation logic provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of disturbance pre-verification and dynamic correction provided in an embodiment of the present invention. Detailed Implementation

[0008] This invention provides a method for generating new energy power data based on a pre-trained physical constraint model. It specifically addresses the technical problems in existing new energy power data generation processes, such as the lack of full-process physical constraints, the tendency of generated samples to deviate from the actual power generation mechanism, and the insufficient ability to construct extreme operating condition samples, which lead to poor physical rationality of generated data, low scenario coverage, and limited model generalization ability.

[0009] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0010] Examples, such as Figure 1 As shown, this invention provides a method for generating new energy power data based on a pre-trained physical constraint model, the method comprising: S100: Obtain the meteorological feature vector and physical structure vector of the target new energy power station; In the process of generating new energy power data, existing technologies mostly rely on single meteorological data or historical power sequences as input, making it difficult to simultaneously characterize the environmental changes of the power plant and the inherent power generation capacity of the equipment. This results in a lack of physical constraints and scenario specificity in the generated samples, affecting the data accuracy and model generalization ability. Therefore, how to integrate dynamic meteorological information with static physical structure information to construct an input feature system that can comprehensively reflect the operating status of new energy power plants has become the primary technical problem to be solved in this step.

[0011] Step S100 in the method provided in this embodiment of the invention includes: The wind speed, wind direction, temperature, humidity, atmospheric pressure, and irradiance of the target new energy power station at the current moment are obtained and added to the meteorological feature vector; Obtain the fixed physical parameters of the target new energy power station according to the power generation type of the target new energy power station. If the target new energy power station is a photovoltaic power generation station, the fixed physical parameters include the component tilt angle, azimuth angle, component temperature coefficient and installation altitude. If the target new energy power station is a wind power generation station, the fixed physical parameters include the wind turbine hub height, impeller diameter, cut-in wind speed, cut-out wind speed and rated wind speed. Add the fixed physical parameters to the physical structure vector. The hidden layer state output by the sequence generation model at the previous time step is loaded and used together with the meteorological feature vector and the physical structure vector as input to the sequence generation model.

[0012] In this embodiment, the meteorological feature vector refers to a multi-dimensional dynamic input set used to characterize the current external environmental state of the new energy power station, and is used to reflect the short-term environmental change trend; The physical structure vector refers to the set of static parameters used to characterize the fixed structural attributes of new energy power stations, and is used to reflect the inherent power generation characteristics of the power station. The hidden layer state refers to the historical time series feature information extracted by the sequence generation model at the previous time step, which is used to maintain the temporal continuity and dynamic memory capability of the power generation process.

[0013] In this step, the environmental meteorological data of the target new energy power station at the current moment is collected, and a meteorological feature vector is constructed.

[0014] Specifically, the system acquires current-moment data on wind speed, wind direction, temperature, humidity, atmospheric pressure, and irradiance through a meteorological monitoring system, and arranges these data in a unified dimensional order to form a meteorological feature vector. Among these: Meteorological characteristic vector = wind speed value + wind direction value + temperature value + humidity value + air pressure value + irradiance value.

[0015] For example, when the current monitoring data shows a wind speed of 8.2 meters per second, a wind direction of 135 degrees, a temperature of 28 degrees Celsius, a humidity of 65%, an air pressure of 100.8 kPa, and an irradiance of 820 watts per square meter, the above parameters can be combined in a preset order to form a meteorological feature vector, i.e., meteorological feature vector = 8.2 + 135 + 28 + 65 + 100.8 + 820.

[0016] Subsequently, the fixed physical parameters corresponding to the target new energy power station are extracted, and a physical structure vector is constructed.

[0017] Specifically, differentiated parameter extraction is performed based on the power generation type of the power station: If the target new energy power station is a photovoltaic power station, the fixed physical parameters include: module tilt angle, module azimuth angle, module temperature coefficient, and installation altitude.

[0018] Physical structure vector = tilt angle + azimuth angle + temperature coefficient + altitude.

[0019] For example, when the target new energy power station is a photovoltaic power station, its fixed physical parameters can be set as follows: component tilt angle 25 degrees, component azimuth angle 180 degrees, component temperature coefficient 0.42% per degree Celsius, and installation altitude 520 meters. Then the physical structure vector can be expressed as: physical structure vector = 25 + 180 + 0.42 + 520.

[0020] If the target new energy power station is a wind power station, the fixed physical parameters include: hub height, rotor diameter, cut-in wind speed, cut-out wind speed, and rated wind speed.

[0021] Physical structure vector = hub height + impeller diameter + cut-in wind speed + cut-out wind speed + rated wind speed.

[0022] For example, when the target new energy power station is a wind power station, its fixed physical parameters can be set as hub height of 110 meters, rotor diameter of 150 meters, cut-in wind speed of 3 meters per second, cut-out wind speed of 25 meters per second and rated wind speed of 12 meters per second. The corresponding physical structure vector can be expressed as: physical structure vector = 110 + 150 + 3 + 25 + 12.

[0023] Furthermore, the hidden layer states output from the previous time step of the sequence generation model are loaded, and multi-source input fusion is performed.

[0024] Specifically, the hidden layer state of the previous time step, the current meteorological feature vector, and the current physical structure vector are concatenated and fused to form the joint input vector at the current moment.

[0025] The joint input vector = the hidden layer state of the previous time step + meteorological feature vector + physical structure vector.

[0026] Among them: the hidden layer state weights are obtained automatically through backpropagation during model training; the meteorological feature weights are obtained from the historical power generation error sensitivity analysis results; and the physical structure feature weights are obtained from the long-term operational statistical contribution of the equipment.

[0027] Preferably, the initial fusion weights for each part can be set as follows: weight of the historical hidden state of the previous time step = 0.45; weight of meteorological features = 0.35; weight of physical structure = 0.20.

[0028] Then: Fusion input value = (hidden layer state × 0.45) + (meteorological feature vector × 0.35) + (physical structure vector × 0.20).

[0029] For example, when the hidden layer state composite value output by the previous time step sequence generation model is 0.72, the current meteorological feature composite value is 0.81, and the current physical structure feature composite value is 0.63, and the corresponding weights are set to 0.45, 0.35, and 0.20 respectively, the fusion input value can be calculated as: Fusion input value = (0.72 × 0.45) + (0.81 × 0.35) + (0.63 × 0.20) = 0.324 + 0.2835 + 0.126 = 0.7335.

[0030] Finally, the joint input vector is input into the sequence generation model to generate the predicted state at the current time step.

[0031] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0032] In summary, this step, by jointly modeling real-time meteorological characteristics, the fixed physical structure characteristics of the power station, and the hidden layer state of the previous time step, enables the sequence generation model to simultaneously possess environmental perception, structural constraint, and temporal memory capabilities, thereby constructing a multi-source fusion input mechanism. Compared to prediction methods based solely on single meteorological data, this approach enhances the model's ability to represent the nonlinear power generation patterns of new energy power stations, improves its adaptability to sudden meteorological disturbances, seasonal variations, and differences in equipment structure, reduces the risk of state drift and error accumulation, and improves the stability, continuity, and overall prediction accuracy of short-term power forecasting, providing a reliable state foundation for subsequent multi-time-step recursive predictions.

[0033] S200: Input the meteorological feature vector and the physical structure vector into the sequence generation model to generate the basic power time series; Step S200 in the method provided in this embodiment of the invention includes: The meteorological feature vector, the physical structure vector, and the hidden layer state output from the previous time step are concatenated to obtain a joint input vector; The joint input vector is processed through the input gate, forget gate, and output gate of the sequence generation model to update the hidden layer state; The updated hidden layer state is mapped through the fully connected output layer of the sequence generation model to obtain the base power generation value at the current moment; The updated hidden layer state is used as the input for the next time step. The process of concatenating the meteorological feature vector, the physical structure vector, and the hidden layer state of the previous time step is repeated until the basic power generation values ​​for all times within the preset time window are generated and merged into the basic power time series.

[0034] In this embodiment, the joint input vector refers to the fused input information composed of the meteorological feature vector, physical structure vector and hidden layer state output from the previous time step, which is used to drive the sequence generation model to complete the power prediction at the current time step. The input gate refers to a gating structure used to control the proportion of the current input information being written. The forget gate is a gating structure used to control the proportion of historical states retained from the previous time step; The output gate is a gating structure used to control the output ratio of the current hidden layer state; The basic power generation value refers to the basic power prediction result generated by the sequence generation model at the current time step, which is used to construct the basic power time series within the preset time window.

[0035] In this step, such as Figure 2As shown, firstly, the meteorological feature vector, physical structure vector, and hidden layer state output from the previous time step are concatenated to obtain the joint input vector. Specifically: The joint input vector = the hidden layer state of the previous time step + meteorological feature vector + physical structure vector.

[0036] Among them, the hidden layer state is used to carry the historical time series information of the previous time step, the meteorological feature vector is used to reflect the current dynamic changes of the environment, and the physical structure vector is used to provide the fixed structural constraints of the station.

[0037] Subsequently, the joint input vector is input to the sequence generation model and processed sequentially through the input gate, forget gate, and output gate to update the hidden layer state.

[0038] Specifically, the input gate is used to calculate the effective write ratio of the current input information, and its calculation method is as follows: Input gate value = joint input vector × input gate weight + input gate bias.

[0039] The forget gate is used to calculate the percentage of hidden layer states retained from the previous time step. The calculation method is as follows: Forget gate value = previous time step hidden layer state × forget gate weight + forget gate bias.

[0040] Based on the input gate and forget gate, the current internal state is updated, and the calculation method is as follows: Updated internal state = (internal state of the previous time step × forget gate value) + (current candidate state × input gate value).

[0041] The output gate is used to calculate the output ratio of the current hidden layer state, and its calculation method is as follows: Output gate value = Updated internal state × Output gate weight + Output gate bias.

[0042] The updated hidden layer state is generated based on the output gate value, and its calculation method is as follows: Updated hidden layer state = output gate value × updated internal state.

[0043] The gating weights and bias parameters were all obtained through training with historical power generation data.

[0044] Furthermore, the updated hidden layer state is mapped through the fully connected output layer of the sequence generation model to obtain the base power generation value at the current time step. Specifically: Base power generation value = Updated hidden layer state × Output mapping weight + Output bias.

[0045] The output mapping weights are trained by minimizing the prediction error, which is defined as: Prediction error = Actual power - Baseline power generation value.

[0046] For example, when the updated hidden layer state is 0.78, the output mapping weight is 95, and the output bias is 3, the base power generation value is 0.78 × 95 + 3 = 77.1 megawatts.

[0047] Finally, the updated hidden layer state is used as the input state for the next time step, and the joint input vector concatenation step is performed again until the fundamental power generation values ​​for all times within the preset time window are generated, and then merged in chronological order to form the fundamental power time series. Specifically: The basic power time series = power value at the first time step + power value at the second time step + ... + power value at the Nth time step.

[0048] For example, when the preset time window is six future time steps, generating 77.1 MW, 79.4 MW, 81.2 MW, 78.6 MW, 76.8 MW and 74.5 MW respectively, the corresponding base power time series is formed.

[0049] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0050] Step S200 in the method provided in this embodiment of the invention further includes: During the training phase of the sequence generation model, a joint loss function is constructed, which includes a data fitting term, a physical consistency term, and a smoothness constraint term. The data fitting term is used to measure the mean square error between the base power time series and the true power time series. The physical consistency term is used to measure the mean square error between the base power time series and the power values ​​of the physical model. The smoothness constraint term is used to constrain the squared difference between the power generation values ​​at adjacent time points; The model parameters of the sequence generation model are updated using the weighted sum of the data fitting term, the physical consistency term, and the smoothness constraint term as the value of the joint loss function.

[0051] In this embodiment, the joint loss function refers to the objective function used to uniformly constrain the training process of the sequence generation model. It is used to comprehensively measure the degree of fit between the model prediction results and the real observation data, the degree of consistency with the physical mechanism of new energy, and the degree of continuous smoothness in the time series change process. The data fitting term refers to the error metric between the baseline power time series and the actual power time series, which is used to reflect the direct prediction accuracy of the model. The physical consistency term refers to the error constraint term between the base power time series and the power value of the physical model calculated based on the physical mechanism of the power station, which is used to ensure that the model output conforms to the physical laws in the process of new energy power generation. The smoothness constraint is a constraint on the magnitude of change in power generation values ​​between adjacent time steps. It is used to suppress abnormal abrupt changes in the prediction results and improve the stability of the time series output.

[0052] In this step, firstly, during the training phase of the sequence generation model, a data fitting term is constructed in the joint loss function.

[0053] Specifically, the baseline power time series is compared with the corresponding real power time series time-by-time, and the mean square error is calculated as a data fitting term. Wherein: Data fitting term = Sum of squares of (actual power - base power) across all time steps ÷ Total number of time steps.

[0054] The actual power is derived from historical SCADA monitoring data of new energy power plants, while the baseline power is derived from the output results of the sequence generation model.

[0055] For example, when the actual power at three time steps is 92 MW, 95 MW and 97 MW respectively, and the corresponding base power is 90 MW, 96 MW and 94 MW respectively, then the data fitting term = ((92-90)2 + (95-96)2 + (97-94)2) ÷ 3 = (4 + 1 + 9) ÷ 3 = 4.67.

[0056] Subsequently, the physical consistency term in the joint loss function is constructed.

[0057] Specifically, a physical model power value is established based on the corresponding power generation type of the new energy power station, and the mean square error between the base power time series and the physical model power value is calculated. Wherein: Physical consistency term = Sum of squares of (physical model power - base power) across all time steps ÷ Total number of time steps.

[0058] The power values ​​in the physical model are derived from the calculation results of the new energy generator's physical properties. For photovoltaic power plants, the calculations can be based on irradiance, module temperature, and conversion efficiency; for wind power plants, the calculations can be based on wind speed, air density, and wind turbine power curves.

[0059] For example, when the physical model power at three time steps is 91 MW, 94 MW and 96 MW respectively, and the corresponding base power is 90 MW, 96 MW and 94 MW respectively, then the physical consistency term = ((91-90)2 + (94-96)2 + (96-94)2) ÷ 3 = (1 + 4 + 4) ÷ 3 = 3.

[0060] Furthermore, a smoothness constraint term is constructed in the joint loss function. Specifically, a squared constraint is applied to the difference in the change of the base power generation values ​​between adjacent time steps to limit drastic fluctuations in the sequence output. Wherein: Smoothness constraint term = Sum of squares of (current time step base power - previous time step base power) in all adjacent time steps ÷ (total number of time steps - 1).

[0061] The purpose of this constraint is to reduce high-frequency noise interference in the model prediction and enhance the continuity of the power curve.

[0062] For example, when the base power for three consecutive time steps is 90 MW, 96 MW and 94 MW respectively, the smoothness constraint term = ((96-90)² + (94-96)²) ÷ 2 = (36 + 4) ÷ 2 = 20.

[0063] Finally, the data fitting term, physical consistency term, and smoothness constraint term are weighted and summed according to preset weights to construct a joint loss function, and the model parameters of the sequence generation model are updated based on the joint loss function. Wherein: Joint loss function = (data fitting term × first loss weight) + (physical consistency term × second loss weight) + (smoothness constraint term × third loss weight).

[0064] The first loss weight, which accounts for the majority of the weights, is used to ensure the model's prediction accuracy. The second loss weight is used to strengthen the constraints imposed by physical laws. The third loss weight is used to balance the smoothness of the sequence. Preferably, the first loss weight can be set to 0.60, the second loss weight to 0.25, and the third loss weight to 0.15. The weights are derived from the cross-validation results of the historical training set, with the weight combination that results in the smallest prediction error and the most stable physical deviation being selected as the preferred configuration.

[0065] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0066] In summary, this step constructs a sequential recursive prediction mechanism by integrating meteorological features, physical structure features, and historical hidden states. It also constructs a joint loss function by combining data fitting, physical consistency, and smoothness constraints, enabling the model to ensure prediction accuracy while maintaining physical plausibility and temporal continuity. Compared to traditional single-error-driven methods, this approach effectively reduces overfitting, physical bias accumulation, and power surges, improving the stability, continuity, and reliability of short-term power prediction for renewable energy power plants.

[0067] S300: Invoke the physical constraint model corresponding to the power generation type of the target new energy power station, calculate the power value of the physical model based on the meteorological feature vector and the physical structure vector, and determine the validity of the basic power time series based on the physical consistency error between the basic power time series and the power value of the physical model; Step S300 in the method provided in this embodiment of the invention includes: If the power generation type of the target new energy power station is wind power generation, then the wind power generation physical constraint model is invoked, and the calculation is performed based on the wind speed in the meteorological feature vector and the cut-in wind speed, cut-out wind speed, and rated wind speed in the physical structure vector. When the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed, the power value of the physical model is zero. When the wind speed reaches the rated wind speed, the power value of the physical model is the upper limit of the rated power. If the power generation type of the target new energy power station is photovoltaic power generation, then the photovoltaic power generation physical constraint model is invoked, and calculations are performed based on the irradiance in the meteorological feature vector and the component tilt angle, azimuth angle, and component temperature coefficient in the physical structure vector. When the irradiance is zero, the power value of the physical model is zero.

[0068] In this embodiment, the wind power generation physical constraint model refers to the power boundary constraint model constructed based on the wind turbine power curve, which is used to characterize the physical mapping relationship between wind speed changes and wind turbine output power. The photovoltaic power generation physical constraint model refers to a power boundary constraint model built based on the solar irradiation receiving mechanism, which is used to characterize the physical mapping relationship between irradiation conditions and module output power. The power value of the physical model refers to the theoretical power value calculated based on the principle of new energy generators, which is used to constrain the physical consistency of the output results of the sequence generation model.

[0069] In this step, such as Figure 3 As shown, firstly, when the power generation type of the target new energy power station is wind power generation, the corresponding physical model power value is calculated by calling the wind power generation physical constraint model.

[0070] Specifically, based on the wind speed value in the current meteorological feature vector, and the cut-in wind speed, cut-out wind speed, and rated wind speed in the physical structure vector, segmented power calculation is performed: When wind speed < cut-in wind speed: physical model power value = 0; When wind speed ≥ cut-in wind speed and wind speed < rated wind speed: physical model power value = rated power × (current wind speed - cut-in wind speed) ÷ (rated wind speed - cut-in wind speed); When wind speed ≥ rated wind speed and wind speed ≤ cut-out wind speed: physical model power value = rated power; When the wind speed is greater than the cut-out wind speed: the physical model power value is 0.

[0071] For example, when the wind turbine's cut-in wind speed is 3 m / s, its rated wind speed is 12 m / s, its cut-out wind speed is 25 m / s, its rated power is 100 MW, and the current wind speed is 8 m / s, then the physical model power value = 100 × (8 - 3) ÷ (12 - 3) = 55.6 MW.

[0072] Subsequently, when the power generation type of the target new energy power station is photovoltaic power generation, the photovoltaic power generation physical constraint model is called to calculate the corresponding physical model power value.

[0073] Specifically, corrections are made based on the irradiance in the current meteorological feature vector, and the component tilt angle, azimuth angle, and component temperature coefficient in the physical structure vector: Physical model power value = standard rated power × irradiance correction factor × angle correction factor × temperature correction factor.

[0074] The standard rated power is derived from the nameplate parameters or rated operating configuration parameters of the target photovoltaic power station equipment. The irradiance correction coefficient is calculated based on the actual irradiance collected at the current moment and the standard irradiance under standard test conditions, specifically: Irradiance correction factor = Current irradiance ÷ Standard irradiance; The current irradiance is derived from the irradiance monitoring value in the meteorological feature vector, and the standard irradiance is preferably set to 1000 watts per square meter; The angle correction factor is derived from the cosine of the angle between the component tilt angle and the solar incidence angle; Temperature correction factor = 1 - (component temperature coefficient × temperature deviation value); The component temperature coefficient is derived from the photovoltaic module's factory technical parameters or equipment specification database, and is used to characterize the output power attenuation ratio corresponding to each unit change in component temperature. The temperature deviation value is derived from the difference between the current air temperature value in the meteorological feature vector and the preset benchmark operating temperature, specifically: Temperature deviation = Current air temperature - Reference operating temperature; The reference operating temperature is preferably set to 25 degrees Celsius. When irradiance = 0: physical model power value = 0.

[0075] For example, when the standard rated power is 80 MW, the current irradiance is 800 W / m², the standard irradiance is 1000 W / m², the angle correction factor is 0.92, the component temperature coefficient is 0.0042, the current air temperature is 35°C, and the reference operating temperature is 25°C, then the temperature deviation value = 35 - 25 = 10, the temperature correction factor = 1 - (0.0042 × 10) = 0.958, and the physical model power value = 80 × 0.8 × 0.92 × 0.958 = 56.41 MW.

[0076] Furthermore, the power value of the physical model is used as a reference value in the physical consistency term, and error calculation is performed between it and the predicted value in the base power time series. Specifically: Physical consistency error = (base power generation value - physical model power value) squared.

[0077] The basic power generation value is derived from the predicted power value of the current time step generated by the sequence generation model through the fully connected output layer; the physical model power value is derived from the theoretical power value calculated in this step based on the photovoltaic power generation physical constraint model.

[0078] By introducing physical constraint errors, the sequence generation model can avoid generating abnormal prediction results that violate the operating boundaries of the device and the laws of energy conversion during the training process.

[0079] Finally, the power values ​​of the physical model are synchronously input into the joint loss function to participate in the model parameter update, so that the data-driven prediction results are consistent with the physical mechanism constraints.

[0080] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0081] Step S300 in the method provided in this embodiment of the invention further includes: The squared difference between the base power generation value at each moment in the base power time series and the power value of the physical model at the corresponding moment is calculated to obtain the single-point physical deviation value. The physical consistency error is obtained by summing the single-point physical deviation values ​​at all times and dividing by the total number of times. If the physical consistency error is less than or equal to a preset first error threshold, the basic power time series is determined to be valid; otherwise, it is discarded.

[0082] In this embodiment, the single-point physical deviation value refers to the instantaneous deviation between the base power generation value and the corresponding physical model power value, which is used to measure the consistency between the prediction result of a single time step and the physical constraints. The physical consistency error refers to the average result of all single-point physical deviation values ​​within a preset time window, which is used to measure the overall degree of deviation between the entire base power time series and physical laws. The first error threshold is a criterion used to determine whether the base power time series meets the physical validity criteria, and its threshold value is determined based on the statistical results of historical operating errors.

[0083] In this step, firstly, the base power generation value and the physical model power value corresponding to each time step in the base power time series are compared point by point, and the single-point physical deviation value is calculated. Specifically: Single-point physical deviation value = (basic power generation value - physical model power value) squared.

[0084] For example, when the base power generation value is 82 MW at a certain moment and the corresponding physical model power value is 79 MW, then: single-point physical deviation value = (82 - 79)² = 9.

[0085] Subsequently, the single-point physical deviation values ​​of all time steps within the preset time window are accumulated and averaged to obtain the physical consistency error. Specifically: Physical consistency error = sum of all single-point physical deviation values ​​÷ total number of time steps.

[0086] For example, when the single-point physical deviation values ​​corresponding to four consecutive time steps are 9, 4, 16 and 1 respectively, then: Physical consistency error = (9 + 4 + 16 + 1) ÷ 4 = 7.5.

[0087] Furthermore, the physical consistency error is compared with a preset first error threshold to determine whether the current base power time series conforms to the physical laws of new energy power generation.

[0088] The first error threshold is preferably set as 3% of the square of the rated power, which is derived from the statistical upper bound of the physical deviation distribution in the historical training samples, and is used to balance prediction accuracy and physical rationality.

[0089] For example, when the rated power of the target renewable energy power station is 100 MW, then: the first error threshold = 1002 × 3% = 300. If the currently calculated physical consistency error is 7.5, then since 7.5 < 300, the base power time series is determined to be valid.

[0090] Finally, when the physical consistency error is less than or equal to the first error threshold, the current base power time series is retained as a valid prediction result and proceeds to subsequent power correction or multi-step recursive processing; if the physical consistency error is greater than the first error threshold, it is determined that the current base power time series has a large physical deviation, it is discarded and the sequence generation model prediction is re-executed.

[0091] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0092] In summary, this step constructs physical constraint models for wind and photovoltaic power generation, and combines them with a physical consistency error screening mechanism to ensure that the output of the sequence generation model simultaneously meets data fitting requirements and the constraints of actual power generation physical laws. By calculating the time-by-time deviation between the base power time series and the power values ​​of the physical model and performing overall consistency verification, abnormal prediction results that deviate from the equipment operating boundaries and energy conversion laws can be effectively eliminated. This reduces the prediction risks caused by extreme weather disturbances and state drift, improves the stability, reliability, and physical rationality of short-term power prediction results for new energy power plants, and provides more reliable basic data support for subsequent multi-time-step recursive prediction and power dispatch.

[0093] S400: Determine the disturbance parameters of the extreme scenario disturbance based on the physical structure vector, apply the extreme scenario disturbance to the meteorological feature vector based on the disturbance parameters, and obtain the extreme disturbance meteorological feature vector; Step S400 in the method provided in this embodiment of the invention includes: The disturbance parameters are determined, including the disturbance type and the corresponding disturbance amplitude constraint; If the power generation type of the target new energy power station is wind power generation, then the upper limit of the amplitude of the wind speed pulse disturbance is determined according to the rated wind speed in the physical structure vector, and the lower limit threshold of the wind speed reduction disturbance is determined according to the cut-in wind speed. The wind speed pulse disturbance and the wind speed reduction disturbance are taken as the disturbance type, and the upper limit of the amplitude and the lower limit threshold are taken as the corresponding disturbance amplitude constraints. If the power generation type of the target new energy power station is photovoltaic power generation, then the amplitude of the irradiance reduction disturbance is determined according to the component tilt angle and azimuth angle in the physical structure vector, the amplitude of the temperature offset disturbance is determined according to the component temperature coefficient, the irradiance reduction disturbance and the temperature offset disturbance are taken as the disturbance type, and the amplitude and the magnitude are taken as the corresponding disturbance amplitude constraints. According to the determined disturbance type and the corresponding disturbance amplitude constraint, the corresponding meteorological features in the meteorological feature vector are perturbed to obtain the extreme disturbance meteorological feature vector.

[0094] In this embodiment, the disturbance parameter refers to a set of control parameters used to simulate the magnitude of environmental changes at new energy power stations under extreme weather conditions, and is used to generate abnormal meteorological inputs that conform to physical boundary constraints; The disturbance type refers to the characteristic disturbance method set for different power generation types, which is used to construct extreme operating states under different scenarios. The disturbance amplitude constraint refers to the boundary parameter that limits the range of disturbance variation, in order to avoid generating abnormal inputs that exceed the operating limits of the equipment; The extreme perturbation meteorological feature vector refers to the abnormal scene input sequence formed by applying perturbation to the original meteorological feature vector, which is used to evaluate the robustness of the model under extreme conditions.

[0095] In this step, such as Figure 4 As shown, firstly, the disturbance parameters are determined, and then a differentiated disturbance strategy is implemented according to the power generation type of the target new energy power station.

[0096] Specifically, when the target new energy power station is wind power generation, the rated wind speed is used as the upper limit of the wind speed pulse disturbance amplitude, and the cut-in wind speed is used as the lower limit threshold of the wind speed reduction disturbance. Wherein: Wind speed pulse disturbance limit = rated wind speed × 20%; The lower limit of wind speed reduction disturbance = cut-in wind speed.

[0097] For example, when the rated wind speed is 12 m / s and the cut-in wind speed is 3 m / s, the maximum increase of the wind speed pulse disturbance is 12 × 20% = 2.4 m / s, and the minimum wind speed drop boundary is 3 m / s.

[0098] Subsequently, when the target new energy power station is photovoltaic power generation, the disturbance parameters of the photovoltaic scenario are determined based on the module tilt angle, ambient temperature, and module temperature coefficient. Specifically: Irradiance reduction disturbance amplitude = current irradiance × base attenuation coefficient × (1 + component tilt angle ÷ 180); Temperature deviation disturbance amplitude = ambient temperature × temperature disturbance proportionality coefficient.

[0099] The component tilt angle reflects the sensitivity of the component to changes in received irradiance, the component temperature coefficient reflects the impact of temperature changes on output power, and the base attenuation coefficient characterizes the degree of base irradiance attenuation caused by extreme weather. Preferably, the base attenuation coefficient ranges from 0.1 to 0.8, and its value is determined based on historical extreme weather statistics or site operation experience. The temperature disturbance proportionality coefficient characterizes the fluctuation of ambient temperature under extreme weather conditions relative to normal operating conditions, and its value is determined based on temperature change statistics in the historical operating data of the target new energy power station.

[0100] For example, when the component tilt angle is 30 degrees, the current irradiance is 800 watts per square meter, the base attenuation coefficient is 0.2, the ambient temperature is 25 degrees Celsius, and the temperature disturbance ratio coefficient is 0.1, then the irradiance reduction disturbance amplitude is 800 × 0.2 × (1 + 30 ÷ 180) = 186.7 watts per square meter, and the temperature offset disturbance amplitude is 25 × 0.1 = 2.5 degrees Celsius.

[0101] Furthermore, according to the determined disturbance type and amplitude constraints, disturbances are applied to the corresponding features in the meteorological feature vector to generate an extreme disturbance meteorological feature vector. Specifically: Extreme disturbance meteorological characteristic value = original meteorological characteristic value ± disturbance amplitude.

[0102] Positive perturbations are used to simulate sudden-increase extreme weather, while negative perturbations are used to simulate decaying extreme weather.

[0103] For example, when the original wind speed is 8 meters per second, and a positive wind speed pulse disturbance of 2 meters per second is applied, the wind speed after disturbance = 8 + 2 = 10 meters per second.

[0104] Finally, the extreme disturbance meteorological feature vector is input into the sequence generation model to generate power prediction results under the corresponding extreme conditions, which are used to evaluate the model's stability and anti-disturbance capability under abnormal environmental conditions.

[0105] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0106] Step S400 in the method provided in this embodiment of the invention further includes: Before applying the extreme scenario perturbation, the physical constraint model is invoked to pre-validate the perturbed meteorological feature vector; If the physical model power value corresponding to the perturbed meteorological feature vector exceeds the physical feasible region, the perturbed parameters are adjusted until the physical model power value is within the physical feasible region, wherein the physical feasible region is a numerical range in which the power value is greater than or equal to zero and less than or equal to the upper limit of the rated power.

[0107] In this embodiment, the physical feasible region refers to the theoretical power range that a new energy power station is allowed to output under the constraints of equipment structure and generator principle, and is used to limit the input scenario after disturbance to still meet the physical operating boundary; The pre-verification refers to the calculation of the prior power of the meteorological characteristics after the disturbance using a physical constraint model before the disturbance is formally applied, in order to verify its physical rationality. The disturbance parameter adjustment refers to the adaptive correction of the disturbance amplitude when the power value of the physical model generated after the disturbance exceeds the physical feasible region, so as to reconstruct an extreme disturbance scenario that satisfies the physical constraints.

[0108] In this step, firstly, before applying extreme scenario perturbations, the perturbed meteorological feature vectors are input into the corresponding physical constraint model for pre-verification.

[0109] Specifically, the theoretical power value is calculated by calling the corresponding physical model based on the power generation type of the new energy power station: The power value of the physical model = the physical constraint model (the meteorological feature vector after disturbance, the physical structure vector).

[0110] For wind farms, the calculation is based on wind speed, cut-in wind speed, cut-out wind speed, and rated wind speed; for photovoltaic farms, the calculation is based on irradiance, module tilt angle, and temperature coefficient.

[0111] Next, it is determined whether the power value of the physical model is within the physically feasible region. Specifically: the physically feasible region = [0, upper limit of rated power]. If this condition is met, the current disturbance parameter is deemed valid.

[0112] For example, when the rated power of the wind farm is 100 MW and the calculated power value of the physical model after the disturbance is 92 MW, the disturbance is deemed valid because 92 is within the range of [0,100].

[0113] Furthermore, if the power value of the physical model exceeds the physically feasible region, the perturbation parameters are adaptively adjusted. Specifically: Adjusted disturbance amplitude = original disturbance amplitude × correction factor.

[0114] The correction coefficient is preferably set to 0.8, which is derived from the convergence stability statistics in historical perturbation tests, and is used to ensure that the perturbation amplitude gradually shrinks and quickly returns to the physical feasible region.

[0115] For example, if the original wind speed pulse disturbance amplitude is 3 meters per second, causing the power value to exceed the rated upper limit, then the adjusted disturbance amplitude is 3 × 0.8 = 2.4 meters per second, and the physical pre-verification is re-executed.

[0116] Finally, the disturbance parameter adjustment and physical pre-verification are repeated until the power value of the physical model is within the physical feasible region. The verified disturbance result is then used as the final extreme disturbance meteorological feature vector input sequence to generate the model.

[0117] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0118] Step S400 in the method provided in this embodiment of the invention further includes: Based on the disturbance type and disturbance magnitude of the extreme scenario disturbance, the power response parameters in the physical structure vector are dynamically corrected to obtain the corrected physical structure vector. The power response parameters include the component temperature coefficient of photovoltaic power generation or the wind energy utilization coefficient of wind power generation. Replace the physical structure vector with the modified physical structure vector.

[0119] In this embodiment, the power response parameter refers to the structural parameter in the new energy power station that directly affects the power output sensitivity, and is used to reflect the dynamic response capability of the equipment under external disturbances. The dynamic correction refers to adaptively adjusting the power response parameters according to the type and magnitude of extreme disturbances, so as to enhance the physical structure parameters' ability to characterize extreme operating conditions. The modified physical structure vector refers to the set of physical structure parameters that have been dynamically modified and are used to replace the original physical structure vector in subsequent power prediction.

[0120] In this step, firstly, based on the type and magnitude of the disturbance in the extreme scenario, the corresponding power response parameter correction object is determined.

[0121] Specifically, when the target new energy power station is wind power generation, the wind energy utilization coefficient is used as a dynamic correction parameter; when the target new energy power station is photovoltaic power generation, the module temperature coefficient is used as a dynamic correction parameter.

[0122] Among them, the wind energy utilization coefficient is used to characterize the wind energy conversion efficiency, and the module temperature coefficient is used to characterize the degree of influence of temperature change on the output efficiency of photovoltaic modules.

[0123] Subsequently, the power response parameters are dynamically corrected based on the disturbance magnitude.

[0124] Specifically, in the case of wind power generation: Corrected wind energy utilization coefficient = original wind energy utilization coefficient × (1 - disturbance influence coefficient × wind speed disturbance amplitude ÷ rated wind speed).

[0125] The disturbance impact coefficient is used to characterize the degree of impact of extreme wind speed fluctuations on the energy conversion efficiency of wind turbines, and its value is determined based on the statistical results of historical operating data.

[0126] For example, when the original wind energy utilization coefficient is 0.42, the disturbance influence coefficient is 0.5, the wind speed disturbance amplitude is 2 meters per second, and the rated wind speed is 12 meters per second, then the corrected wind energy utilization coefficient = 0.42 × (1 - 0.5 × 2 ÷ 12) = 0.385.

[0127] When it is a photovoltaic power generation scenario: Corrected temperature coefficient = original temperature coefficient × (1 + temperature offset disturbance amplitude ÷ reference temperature).

[0128] The reference temperature is preferably set to 25 degrees Celsius, a value derived from the standard test conditions for photovoltaic modules.

[0129] For example, when the original temperature coefficient is 0.0042 and the temperature deviation is 5 degrees Celsius, the corrected temperature coefficient is 0.0042 × (1 + 5 ÷ 25) = 0.00504.

[0130] Furthermore, the corrected power response parameters are used to replace the corresponding parameters in the original physical structure vector to form a corrected physical structure vector.

[0131] Specifically: Corrected physical structure vector = Original physical structure vector - Original power response parameter + Corrected power response parameter.

[0132] This replacement mechanism transforms the physical structural features from static parameters into semi-dynamic parameters that can change dynamically with extreme scenarios.

[0133] Finally, the modified physical structure vector replaces the original physical structure vector and is input into the subsequent sequence generation model for power prediction, so that the model can synchronously sense changes in device response capability under extreme disturbance scenarios.

[0134] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0135] In summary, this step, by constructing a mechanism for extreme disturbance generation, physical pre-verification, and dynamic correction of structural parameters, achieves the coordinated updating of meteorological disturbance characteristics and equipment response characteristics. Compared with traditional static disturbance methods, this scheme can ensure that the disturbance scenario always meets the physical feasible region constraints and synchronously reflects the changes in the equipment's response capability under extreme operating conditions. This improves the robustness, physical consistency, and power prediction stability of the sequence generation model under abnormal weather conditions, providing a more reliable dynamic constraint basis for subsequent multi-time-step predictions.

[0136] S500: Input the extreme disturbance meteorological feature vector into the sequence generation model to generate a power time series of extreme scenarios; Step S500 in the method provided in this embodiment of the invention includes: The extreme disturbance meteorological feature vector, the physical structure vector, and the hidden layer state output from the previous time step are concatenated. The sequence generation model is invoked to generate corresponding extreme scenario power values ​​step by step over time. The extreme scenario power generation values ​​at all times are merged to form the extreme scenario power time series.

[0137] In this embodiment, the extreme disturbance meteorological feature vector refers to the abnormal environmental input feature formed by applying extreme scenario disturbances to the original meteorological feature vector, which is used to simulate the operating status of new energy power stations under sudden meteorological conditions. The extreme scenario power generation value refers to the single-time-step predicted power value generated by the sequence generation model under extreme perturbation conditions, which is used to reflect the instantaneous power generation capacity of the equipment under abnormal operating conditions. The extreme scenario power time series refers to a power change sequence formed by combining the power generation values ​​of multiple consecutive time steps in extreme scenarios in chronological order, which is used to characterize the overall power evolution trend of the station under extreme conditions.

[0138] In this step, firstly, the extreme disturbance meteorological feature vector, the physical structure vector, and the hidden layer state output from the previous time step are concatenated to construct a joint input vector for the extreme scenario. Specifically: Extreme joint input vector = hidden layer state + extreme perturbation meteorological feature vector + physical structure vector.

[0139] Among them, the hidden layer state is used to retain historical power evolution information, the extreme disturbance meteorological feature vector is used to introduce abnormal environmental change characteristics, and the physical structure vector is used to provide station fixed structure constraints.

[0140] Subsequently, the sequence generation model is invoked to perform time-step recursive calculations on the extreme joint input vector, and state updates are completed through input gates, forget gates, and output gates to generate the updated hidden layer state corresponding to the current time step. Specifically: Update hidden layer state = sequence generation model (extreme joint input vector, hidden layer state of the previous time step).

[0141] The updated hidden layer state is the output result of the internal state update of the sequence generation model, which is used to characterize the time-series power evolution characteristics under the current extreme perturbation conditions.

[0142] Furthermore, the updated hidden layer state is mapped to the fully connected output layer to obtain the extreme scene power generation value at the corresponding time. Specifically: Current extreme power generation value = Updated hidden layer state × Output mapping weight + Output bias.

[0143] For example, when the updated hidden layer state is 0.83, the output mapping weight is 96, and the output bias is 2, the current extreme scenario power generation value is 0.83 × 96 + 2 = 81.68 megawatts.

[0144] Finally, the updated hidden layer state of the current time step is used as the input state for the next time step, and the recursive prediction of the extreme scenario at the next moment is continued until the prediction generation of all time steps within the preset time window is completed. The extreme scenario power generation values ​​corresponding to all moments are then merged in chronological order to form the extreme scenario power time series. Specifically: Extreme scenario power time series = extreme power value at the first time step + extreme power value at the second time step + ... + extreme power value at the Nth time step.

[0145] For example, when the generated values ​​for the next four time steps are 79.68 MW, 75.42 MW, 71.35 MW and 68.20 MW respectively, a corresponding extreme scenario power time series is formed.

[0146] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0147] In summary, this step, by performing multi-time-step recursive power generation under extreme weather conditions, enables the sequence generation model to simulate the dynamic power change process of new energy power plants under abnormal environments. This improves the model's adaptability and robustness to extreme scenarios and provides a reliable basis for extreme condition predictions for subsequent risk assessment and power correction.

[0148] S600: Call the physical constraint model, calculate the extreme physical model power value based on the extreme disturbance meteorological feature vector and the physical structure vector, and determine the validity of the extreme scenario power time series based on the extreme physical consistency error between the extreme scenario power time series and the extreme physical model power value; Step S600 in the method provided in this embodiment of the invention includes: Calculate the extreme physical model power value by calling the physical constraint model corresponding to the power generation type; In this embodiment, the physical constraint model refers to a theoretical power calculation model pre-established based on the generator mechanism of new energy power plants, which is used to calculate the theoretical output power that meets the equipment operating boundary under extreme disturbance scenarios. The extreme physical model power value refers to the theoretical power value calculated based on the physical constraint model under extreme weather disturbance conditions, and is used to verify the physical rationality of the power prediction results for extreme scenarios.

[0149] In this step, the first step is to determine the power generation type of the target renewable energy power station and then call the corresponding physical constraint model for calculation.

[0150] Specifically, when the target new energy power station is wind power generation, the wind power generation physical constraint model is invoked, and segmented calculations are performed based on the wind speed in the extreme disturbance meteorological characteristic vector and the cut-in wind speed, cut-out wind speed, and rated wind speed in the physical structure vector: When wind speed < cut-in wind speed: extreme physical model power value = 0; When the cut-in wind speed ≤ wind speed < rated wind speed: Extreme physical model power value = rated power × (current wind speed - cut-in wind speed) ÷ (rated wind speed - cut-in wind speed); When rated wind speed ≤ wind speed ≤ cut-out wind speed: extreme physical model power value = rated power; When the wind speed is greater than the cut-out wind speed: the power value of the extreme physics model is 0.

[0151] For example, when the cut-in wind speed is 3 m / s, the rated wind speed is 12 m / s, the cut-out wind speed is 25 m / s, the rated power is 100 MW, and the wind speed after disturbance is 10 m / s, then the extreme physical model power value = 100 × (10 - 3) ÷ (12 - 3) = 77.8 MW.

[0152] Subsequently, when the target new energy power station is photovoltaic power generation, the photovoltaic power generation physical constraint model is invoked, and power is calculated based on the irradiance in the extreme disturbance meteorological characteristic vector and the module tilt angle, azimuth angle, and module temperature coefficient in the physical structure vector. Specifically: Extreme physical model power value = standard rated power × irradiation correction factor × angle correction factor × temperature correction factor.

[0153] in: Irradiation correction factor = Current irradiance ÷ Standard irradiance; The angle correction factor is derived from the cosine of the angle between the component tilt angle and the solar incidence angle; Temperature correction factor = 1 - (component temperature coefficient × temperature offset).

[0154] When irradiance = 0: the power value of the extreme physical model = 0.

[0155] For example, when the standard rated power is 80 MW, the irradiance after disturbance is 600 W / m², the angle correction factor is 0.90, and the temperature correction factor is 0.93, the extreme physical model power value is 80 × 0.6 × 0.90 × 0.93 = 40.18 MW.

[0156] Furthermore, the power value of the extreme physics model is compared with the power generation value of the extreme scene output by the sequence generation model to calculate the physics deviation value. Specifically: Physical deviation value = (Extreme scenario power generation value - Extreme physical model power value) squared.

[0157] This deviation calculation can quantify the degree of consistency between the model's predictions and the physical theoretical values.

[0158] Finally, the power value of the extreme physical model is used as a physical reference benchmark in extreme scenarios and input into the subsequent error evaluation module or joint loss function to verify the physical validity of the prediction results for extreme conditions and optimize the model parameters.

[0159] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0160] Step S600 in the method provided in this embodiment of the invention further includes: Calculate the single-point deviation between the power time series of the extreme scenario and the power value of the extreme physical model; Accumulate and average the deviation values ​​at all times to obtain the extreme physical consistency error; In this embodiment, the single-point deviation refers to the instantaneous difference between the power value output by the sequence generation model and the theoretical power value calculated by the corresponding physical constraint model under extreme scenarios, which is used to characterize the degree of deviation between the prediction result of a single time step and the physical law; The extreme physical consistency error refers to the average result of all single-point deviations within the entire extreme scenario time window, and is used to measure the physical consistency level of the model's overall prediction results under extreme conditions.

[0161] In this step, firstly, the power generation value at each time step in the extreme scenario power time series is compared with the power value of the extreme physical model at the corresponding time step, and the single-point deviation is calculated. Specifically: Single-point deviation = (Extreme scenario power generation value - Extreme physics model power value) squared.

[0162] Among these methods, squaring can amplify abnormal deviations and increase sensitivity to extreme outliers.

[0163] For example, when the extreme scenario power generation value at a certain time step is 82 MW and the corresponding extreme physical model power value is 78 MW, then the single-point deviation = (82 - 78)² = 16.

[0164] Subsequently, the single-point deviations of all time steps are accumulated and divided by the total number of time steps to obtain the extreme physical consistency error. Specifically: Extreme physical consistency error = sum of all single-point deviations ÷ total number of time steps.

[0165] The average value is used to reflect the overall deviation between the prediction results and the theoretical physical boundary under the entire extreme perturbation scenario.

[0166] Furthermore, the extreme physical consistency error can be used as a robustness evaluation index for extreme scenarios and input into the subsequent evaluation module to determine the stability of the current sequence generation model under abnormal operating conditions. The smaller the error value, the more the model output conforms to the physical operating laws of new energy equipment in extreme environments.

[0167] Finally, after completing the error statistics for all time steps, the extreme physical consistency error is output as the basis for subsequent model robustness analysis and extreme scenario performance evaluation.

[0168] For example, when the single-point deviations of four consecutive time steps are 16, 9, 4 and 1 respectively, the extreme physical consistency error = (16 + 9 + 4 + 1) ÷ 4 = 7.5.

[0169] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0170] Step S600 in the method provided in this embodiment of the invention further includes: If the extreme physical consistency error is less than or equal to a preset second error threshold, the extreme scenario power time series is determined to be valid; otherwise, it is discarded.

[0171] In this embodiment, the second error threshold refers to the error boundary value used to determine whether the power prediction result in extreme scenarios meets the physical consistency requirements, and is used to screen effective prediction sequences that still conform to the equipment operation mechanism under extreme disturbance conditions; The term "power time series for extreme scenarios is valid" means that the corresponding prediction results, as a whole, meet the acceptable deviation range of the physical constraint model and can be used for subsequent robustness analysis and risk assessment.

[0172] In this step, firstly, the extreme physical consistency error calculated in the previous steps is compared with a preset second error threshold. Specifically: Valid judgment condition = extreme physical consistency error ≤ second error threshold.

[0173] The second error threshold is preferably set to 5% of the square of the rated power. This threshold is derived from the statistical results of historical extreme working condition samples and is used to reflect the reasonable range of physical deviations that can exist under extreme conditions.

[0174] Subsequently, when the extreme physical consistency error meets the validity determination condition, the current extreme scenario power time series is determined to be valid and retained as the extreme condition prediction result.

[0175] For example, when the rated power of the target new energy power station is 100 MW, the second error threshold is 1002 × 5% = 500; if the current extreme physical consistency error is 120, then since 120 ≤ 500, the power time series of this extreme scenario is determined to be valid.

[0176] Furthermore, when the extreme physical consistency error is greater than the second error threshold, it is determined that the current extreme scenario power time series has a large physical deviation and is discarded to prevent abnormal prediction results from entering the subsequent analysis process.

[0177] For example, when the extreme physical consistency error is 620, since 620 > 500, the predicted sequence is determined to be invalid, and re-perturbation generation or re-prediction is performed.

[0178] Finally, the effective extreme scenario power time series, which has been screened through physical consistency, will be output to the subsequent robustness assessment module to analyze the stability and disturbance resistance of new energy power plants under extreme environmental conditions.

[0179] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0180] In summary, this step, by performing physical consistency verification on the power time series under extreme scenarios and combining it with a second error threshold for validity screening, can promptly eliminate abnormal prediction results that violate equipment operating boundaries or generator mechanisms, improve the physical rationality and reliability of power prediction results under extreme operating conditions, enhance the robustness and stability of the model in complex and abnormal environments, and provide reliable support for the operation analysis and risk warning of new energy power plants under extreme weather conditions.

[0181] S700: Merge the base power time series and the extreme scenario power time series that have passed the validity determination, and output the pre-training dataset.

[0182] Step S700 in the method provided in this embodiment of the invention includes: The basic power time series that has passed the validity determination is associated and stored with the corresponding meteorological feature vector and physical structure vector to construct a normal operating condition sample record; The power time series of the extreme scenarios that have passed the validity determination are associated and stored with the corresponding extreme disturbance meteorological feature vector and physical structure vector to construct extreme working condition sample records; The normal operating condition sample records and the extreme operating condition sample records are merged according to a preset ratio to output the pre-trained dataset.

[0183] In this embodiment, the conventional operating condition sample record refers to a data unit composed of a power time series generated under normal operating conditions and its corresponding input features, which is used to characterize the conventional operating pattern of new energy power plants. The extreme operating condition sample record refers to a data unit composed of the power time series generated under extreme disturbance conditions and its corresponding disturbance input features, which is used to characterize the abnormal operation pattern of new energy power stations. The pre-training dataset refers to a training sample set formed by mixing normal working condition samples and extreme working condition samples in a preset ratio, which is used to improve the generalization ability and robustness of the subsequent model in multiple scenarios.

[0184] In this step, firstly, the baseline power time series, which has passed the validity determination, is associated and stored with the corresponding meteorological feature vector and physical structure vector to construct a normal operating condition sample record. Specifically: Normal operating condition sample record = meteorological feature vector + physical structure vector + basic power time series.

[0185] For example, when a certain basic power time series is {77.1, 79.4, 81.2}, it can be combined with the corresponding input features to form a regular sample record.

[0186] Subsequently, the power time series of extreme scenarios, which have been validated for validity, will be associated and stored with the corresponding extreme disturbance meteorological feature vectors and physical structure vectors to construct extreme operating condition sample records. Specifically: Extreme operating condition sample record = extreme disturbance meteorological feature vector + physical structure vector + extreme scenario power time series.

[0187] For example, when the extreme power time series is {65.4, 61.8, 58.6}, the corresponding extreme sample record can be constructed.

[0188] Furthermore, the normal operating condition sample records and the extreme operating condition sample records are merged according to a preset ratio. Specifically: Pre-training dataset = (number of regular samples × first proportion) + (number of extreme samples × second proportion).

[0189] The first ratio and the second ratio are derived from the statistical results of historical sample distribution, and are preferably set to 7:3 to ensure the stability of learning under normal working conditions while enhancing the coverage of extreme working conditions.

[0190] For example, when the number of regular samples is 700 and the number of extreme samples is 300, a pre-training dataset containing 1000 samples can be constructed.

[0191] Finally, the pre-trained dataset is output and used for the pre-training and parameter initialization of the subsequent new energy power prediction model.

[0192] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0193] In summary, this step constructs a mixed sample dataset containing both normal and extreme operating conditions, enabling the model training process to simultaneously cover normal operation patterns and abnormal disturbance patterns. This improves the model's ability to learn the power generation characteristics of multiple scenarios and enhances its generalization ability, stability, and disturbance resistance in complex environments.

[0194] In summary, this invention constructs a power time series for extreme scenarios by jointly inputting extreme weather disturbance characteristics, power station physical structure characteristics, and historical hidden layer states, and performing multi-time-step recursive prediction based on a sequence generation model. This enables dynamic power simulation of new energy power stations under abnormal weather conditions. Compared to methods that only predict based on conventional operating conditions, this approach effectively enhances the model's responsiveness to extreme environmental changes and its ability to characterize temporal evolution. It also improves the prediction stability and robustness under sudden disturbances, severe fluctuations, and complex boundary conditions, providing a reliable data foundation for subsequent physical consistency verification and anti-disturbance performance evaluation under extreme operating conditions.

[0195] It should be noted that the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0196] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0197] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A method for generating new energy power data based on a pre-trained physical constraint model, characterized in that, include: Obtain the meteorological feature vector and physical structure vector of the target new energy power station; The meteorological feature vector and the physical structure vector are input into the sequence generation model to generate the basic power time series. The physical constraint model corresponding to the power generation type of the target new energy power station is invoked, and the power value of the physical model is calculated based on the meteorological feature vector and the physical structure vector. The validity of the basic power time series is determined based on the physical consistency error between the basic power time series and the power value of the physical model. The perturbation parameters of the extreme scenario perturbation are determined based on the physical structure vector, and the extreme scenario perturbation is applied to the meteorological feature vector based on the perturbation parameters to obtain the extreme perturbation meteorological feature vector. The extreme disturbance meteorological feature vector is input into the sequence generation model to generate a power time series of extreme scenarios; The physical constraint model is invoked, and the extreme physical model power value is calculated based on the extreme disturbance meteorological feature vector and the physical structure vector. The validity of the extreme scenario power time series is determined based on the extreme physical consistency error between the extreme scenario power time series and the extreme physical model power value. The base power time series and the extreme scenario power time series that have passed the validity determination are merged to output a pre-training dataset.

2. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, Obtain the meteorological feature vector and physical structure vector of the target renewable energy power station, including: The wind speed, wind direction, temperature, humidity, atmospheric pressure, and irradiance of the target new energy power station at the current moment are obtained and added to the meteorological feature vector; Obtain the fixed physical parameters of the target new energy power station according to the power generation type of the target new energy power station. If the target new energy power station is a photovoltaic power generation station, the fixed physical parameters include the component tilt angle, azimuth angle, component temperature coefficient and installation altitude. If the target new energy power station is a wind power generation station, the fixed physical parameters include the wind turbine hub height, impeller diameter, cut-in wind speed, cut-out wind speed and rated wind speed. Add the fixed physical parameters to the physical structure vector. The hidden layer state output by the sequence generation model at the previous time step is loaded and used together with the meteorological feature vector and the physical structure vector as input to the sequence generation model.

3. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, The meteorological feature vector and the physical structure vector are input into the sequence generation model to generate a basic power time series, including: The meteorological feature vector, the physical structure vector, and the hidden layer state output from the previous time step are concatenated to obtain a joint input vector; The joint input vector is processed through the input gate, forget gate, and output gate of the sequence generation model to update the hidden layer state; The updated hidden layer state is mapped through the fully connected output layer of the sequence generation model to obtain the base power generation value at the current moment; The updated hidden layer state is used as the input for the next time step. The process of concatenating the meteorological feature vector, the physical structure vector, and the hidden layer state of the previous time step is repeated until the basic power generation values ​​for all times within the preset time window are generated and merged into the basic power time series.

4. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, Also includes: During the training phase of the sequence generation model, a joint loss function is constructed, which includes a data fitting term, a physical consistency term, and a smoothness constraint term. The data fitting term is used to measure the mean square error between the base power time series and the true power time series. The physical consistency term is used to measure the mean square error between the base power time series and the power values ​​of the physical model. The smoothness constraint term is used to constrain the squared difference between the power generation values ​​at adjacent time points; The model parameters of the sequence generation model are updated using the weighted sum of the data fitting term, the physical consistency term, and the smoothness constraint term as the value of the joint loss function.

5. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, Invoking the physical constraint model corresponding to the power generation type of the target renewable energy power station, and calculating the power value of the physical model based on the meteorological feature vector and the physical structure vector, including: If the power generation type of the target new energy power station is wind power generation, then the wind power generation physical constraint model is invoked, and the calculation is performed based on the wind speed in the meteorological feature vector and the cut-in wind speed, cut-out wind speed, and rated wind speed in the physical structure vector. When the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed, the power value of the physical model is zero. When the wind speed reaches the rated wind speed, the power value of the physical model is the upper limit of the rated power. If the power generation type of the target new energy power station is photovoltaic power generation, then the photovoltaic power generation physical constraint model is invoked, and calculations are performed based on the irradiance in the meteorological feature vector and the component tilt angle, azimuth angle, and component temperature coefficient in the physical structure vector. When the irradiance is zero, the power value of the physical model is zero.

6. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, The validity of the base power time series is determined based on the physical consistency error between the base power time series and the power values ​​of the physical model, including: The squared difference between the base power generation value at each moment in the base power time series and the power value of the physical model at the corresponding moment is calculated to obtain the single-point physical deviation value. The physical consistency error is obtained by summing the single-point physical deviation values ​​at all times and dividing by the total number of times. If the physical consistency error is less than or equal to a preset first error threshold, the basic power time series is determined to be valid; otherwise, it is discarded.

7. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, Based on the physical structure vector, the perturbation parameters of the extreme scenario perturbation are determined. Based on the perturbation parameters, the extreme scenario perturbation is applied to the meteorological feature vector to obtain the extreme perturbation meteorological feature vector, including: The disturbance parameters are determined, including the disturbance type and the corresponding disturbance amplitude constraint; If the power generation type of the target new energy power station is wind power generation, then the upper limit of the amplitude of the wind speed pulse disturbance is determined according to the rated wind speed in the physical structure vector, and the lower limit threshold of the wind speed reduction disturbance is determined according to the cut-in wind speed. The wind speed pulse disturbance and the wind speed reduction disturbance are taken as the disturbance type, and the upper limit of the amplitude and the lower limit threshold are taken as the corresponding disturbance amplitude constraints. If the power generation type of the target new energy power station is photovoltaic power generation, then the amplitude of the irradiance reduction disturbance is determined according to the component tilt angle and azimuth angle in the physical structure vector, the amplitude of the temperature offset disturbance is determined according to the component temperature coefficient, the irradiance reduction disturbance and the temperature offset disturbance are taken as the disturbance type, and the amplitude and the magnitude are taken as the corresponding disturbance amplitude constraints. According to the determined disturbance type and the corresponding disturbance amplitude constraint, the corresponding meteorological features in the meteorological feature vector are perturbed to obtain the extreme disturbance meteorological feature vector.

8. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 7, characterized in that, Also includes: Before applying the extreme scenario perturbation, the physical constraint model is invoked to pre-validate the perturbed meteorological feature vector; If the physical model power value corresponding to the perturbed meteorological feature vector exceeds the physical feasible region, the perturbed parameters are adjusted until the physical model power value is within the physical feasible region, wherein the physical feasible region is a numerical range in which the power value is greater than or equal to zero and less than or equal to the upper limit of the rated power.

9. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, Merge the base power time series and the extreme scenario power time series that have passed the validity determination to output a pre-training dataset, including: The basic power time series that has passed the validity determination is associated and stored with the corresponding meteorological feature vector and physical structure vector to construct a normal operating condition sample record; The power time series of the extreme scenarios that have passed the validity determination are associated and stored with the corresponding extreme disturbance meteorological feature vector and physical structure vector to construct extreme working condition sample records; The normal operating condition sample records and the extreme operating condition sample records are merged according to a preset ratio to output the pre-trained dataset.

10. The method for generating new energy power data based on a pre-trained physical constraint model as described in claim 1, characterized in that, Also includes: Based on the disturbance type and disturbance magnitude of the extreme scenario disturbance, the power response parameters in the physical structure vector are dynamically corrected to obtain the corrected physical structure vector. The power response parameters include the component temperature coefficient of photovoltaic power generation or the wind energy utilization coefficient of wind power generation. Replace the physical structure vector with the modified physical structure vector.

Images