Weather data preprocessing methods for new energy power forecasting

Abstract

This invention discloses a weather data preprocessing method for new energy power prediction, relating to the fields of new energy power generation and meteorological data processing. This method constructs a multi-source extreme weather dataset, trains a physically constrained diffusion model to generate compliant meteorological fields, then couples historical power data to generate enhanced samples and calibrates their confidence levels. Finally, it mixes real samples to construct a robust training dataset, improving the accuracy and stability of power prediction under extreme weather conditions. This invention effectively solves the problems of scarce extreme weather samples, lack of physical consistency, and inability to quantify sample confidence levels, thereby improving the generalization ability and operational stability of new energy power prediction models under extreme weather conditions.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and meteorological data processing technology, specifically a weather data preprocessing method for new energy power prediction. Background Technology

[0002] The large-scale development of new energy power generation forms such as wind power and photovoltaics has led to a continuous increase in the proportion of new energy installed capacity in the power system, making it a core support for my country's energy structure transformation. New energy output is characterized by strong intermittency, volatility, and randomness, and its output characteristics are highly correlated with meteorological conditions. Changes in meteorological elements directly determine the output level of new energy power plants, while extreme weather events can cause significant fluctuations in new energy output, posing severe challenges to the safe and stable operation of the power system, dispatch planning, and new energy consumption. New energy power forecasting is a core technical means to smooth out output fluctuations, ensure grid security, and improve consumption levels. Meteorological data preprocessing, as a crucial pre-processing step in power forecasting, directly determines the quality of meteorological input data, thereby determining the prediction accuracy and generalization ability of the power forecasting model. With the frequent occurrence of extreme weather events, the industry's demand for new energy power forecasting under extreme weather scenarios is increasingly urgent. Meteorological data preprocessing technology has therefore become a key area restricting the improvement of power forecasting capabilities under extreme scenarios. There is an urgent need for meteorological data preprocessing methods that are adaptable to extreme weather scenarios and possess high physical fidelity to meet the practical application needs of new energy power plants and grid dispatching.

[0003] Traditional meteorological data preprocessing methods used for new energy power prediction mostly focus on processing routine meteorological observation data and general numerical weather prediction products. These methods are only suitable for power prediction needs under normal weather scenarios and have significant limitations in handling extreme weather scenarios. Traditional methods struggle to address the industry pain point of scarce real historical samples of extreme weather events. The limited number of effective samples of complete extreme weather processes cannot provide sufficient and comprehensive extreme scenario training support for power prediction models, directly resulting in severely insufficient generalization ability of prediction models under extreme weather conditions. Furthermore, traditional meteorological data augmentation methods are mostly based on statistical regularities and conventional generative networks, neglecting the importance of data augmentation in model training. Because atmospheric physics constraints are embedded in the generation process, the generated meteorological field sequences do not conform to the objective laws of atmospheric motion, resulting in insufficient spatiotemporal continuity and lack of physical compliance. They cannot truly reproduce the complete evolution process of extreme weather. In addition, traditional methods cannot achieve physical consistency coupling between meteorological elements and new energy power sequences. The generated meteorological-power pairing samples lack systematic statistics and confidence level calibration, and the sample quality cannot be quantitatively controlled and applied in a hierarchical manner. It is difficult to build a training dataset with high robustness, which ultimately leads to a decrease in the prediction accuracy of new energy power prediction models under extreme weather scenarios, and fails to meet the actual needs of safe operation and refined scheduling of power systems. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a weather data preprocessing method for new energy power prediction. This method involves constructing a multi-source extreme weather meteorological dataset; designing and training a spatiotemporal diffusion model with physical constraints; generating extreme weather meteorological fields that conform to atmospheric physical laws; constructing enhanced samples based on a meteorological-power coupling generation algorithm; and finally forming a robust training set with confidence labels. By introducing meteorological physical constraints and quantification mechanisms, this invention improves the physical authenticity and diversity of the generated samples, providing higher-quality training data support for new energy power prediction models.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a weather data preprocessing method for predicting new energy power, the specific steps of which are as follows:

[0006] S100. Dataset Construction and Labeling: Collect multi-source meteorological data and power operation data of new energy power stations in the target area, screen samples of complete processes of four types of extreme weather, complete sample preprocessing and association labeling, and construct a multi-source extreme weather meteorological dataset to provide benchmark data for subsequent algorithm training.

[0007] S200, Diffusion Model Construction and Training: Based on multi-source extreme weather meteorological datasets, a meteorological and physical constraint spatiotemporal diffusion model is built. During the model training process, three types of meteorological and physical constraints are embedded: atmospheric motion dynamics, atmospheric energy conservation, and atmospheric boundary layer physical parameterization. At the same time, the spatiotemporal continuity constraint of the meteorological field is introduced. The meteorological and physical constraint diffusion loss algorithm is used as the core optimization criterion to complete the iterative training and convergence verification of the model and output the trained diffusion model.

[0008] S300, Meteorological Field Generation Verification: Based on the trained diffusion model, the initial conditions range for extreme weather generation is set, and a spatiotemporal sequence of extreme weather meteorological fields is generated. The physical compliance verification is completed based on the meteorological physical constraint diffusion loss algorithm. The physical compliance verification uses the meteorological physical constraint diffusion loss value as the core judgment criterion. Meteorological field sequences with a loss value lower than a preset threshold are judged as qualified, and qualified meteorological field sequences and corresponding physical compliance coefficients are output.

[0009] S400, Coupled Sample Generation: Based on the historical operation data of the target new energy power station, a real meteorological-power pairing sample library is constructed, the basic power mapping relationship is pre-trained, and the meteorological-power physical coupling generation algorithm is used as the core. Combined with qualified meteorological field sequences and physical compliance coefficients, the corresponding new energy power sequences are generated, and meteorological-power pairing enhancement samples are constructed.

[0010] S500, Training Set Adaptation Construction: Complete the statistics and confidence labeling of paired samples, match the corresponding confidence labels to the samples, mix the labeled augmented samples with the original real extreme weather samples, and construct a robust training dataset.

[0011] Furthermore, the multi-source meteorological data includes global reanalysis meteorological data, measured data from ground meteorological stations, meteorological satellite remote sensing data, and weather radar detection data. Global reanalysis meteorological data is downloaded from a professional meteorological data platform to construct a spatiotemporally continuous atmospheric background field. Ground meteorological station data is collected in real-time from regional meteorological monitoring centers via wired communication to calibrate the accuracy of near-surface meteorological elements. Meteorological satellite remote sensing data is acquired through directional reception from satellite data receiving terminals to capture the overall evolution of large-scale extreme weather events. Weather radar data is retrieved in real-time through radar station data transmission interfaces to capture the changing characteristics of short-duration severe convection and other extreme weather events. The power operation data of the new energy power plants includes minute-by-minute measured output data of wind and photovoltaic units, total installed capacity data, unit operating status data, and unit maintenance and power curtailment records. All power data is collected in real-time and stored locally through a dedicated data acquisition interface at the power plant.

[0012] Furthermore, the meteorological and physical constraints spatiotemporal diffusion model adopts an encoding and decoding network architecture adapted to meteorological spatiotemporal sequences. It performs progressive noise addition and reverse restoration on the input meteorological and spatiotemporal data. During model training, three types of meteorological and physical constraints are embedded: atmospheric motion dynamics, atmospheric energy conservation, and atmospheric boundary layer physical parameterization. At the same time, the spatiotemporal continuity constraint of the meteorological field is introduced. The multi-source extreme weather meteorological dataset is divided into training set and validation set at a fixed ratio of 8:2. An adaptive optimizer is used to iteratively update the model parameters. The corresponding initial learning rate is set to 2e-4, the batch size is 16, and the maximum number of training rounds is 500. The continuous non-decreasing loss of the validation set is used as the convergence criterion. After training, the output can generate an extreme weather meteorological field model that conforms to the laws of atmospheric physics.

[0013] Furthermore, the mathematical expression for the meteorologically constrained diffusion loss algorithm is:

[0014]

[0015] in, The general constraint loss for the meteorological and physical constrained spatiotemporal diffusion model; The loss is the original noise prediction loss for the diffusion model; It is an adaptive spatiotemporal physical weighting factor with a value range of 0-1. It is dynamically adjusted with time step and spatial grid point to balance the physical constraint intensity at different spatiotemporal locations. These are the time step variables of the diffusion model, and also correspond to the time dimension of the meteorological spatiotemporal sequence. For the spatial grid coordinates of the meteorological field, corresponding to the spatial dimension of the meteorological sequence; This is a constraint term for atmospheric dynamics, which constrains the evolution of wind and pressure fields; As an atmospheric energy conservation constraint, it constrains changes in temperature and irradiance to avoid unreasonable meteorological data with sudden energy changes; This is a physical parameterization constraint term for the atmospheric boundary layer, used to constrain the distribution characteristics of near-surface meteorological elements to conform to the physical laws of the atmospheric boundary layer and match the measured characteristics of near-surface meteorology. This is a constraint term for the spatiotemporal continuity of the meteorological field, which constrains the extreme weather generation sequence to have no spatiotemporal jumps.

[0016] Furthermore, the initial conditions for the generation of extreme weather are determined based on the statistical characteristics of core meteorological elements of real extreme weather samples within the multi-source extreme weather meteorological dataset, combined with the meteorological industry's extreme weather classification standards. Specifically, this includes the extreme value range of key meteorological elements, the duration range of the process, the spatial range of impact, and the range of evolution speed. The extreme value range of key meteorological elements is taken from the 5% to 95th percentile interval of the corresponding elements in the real samples, covering core elements such as wind speed, temperature, air pressure, relative humidity, total horizontal irradiance, normal direct irradiance, and precipitation rate. The duration range of the process is set to 24 to 120 hours, matching the typical life cycle of four types of extreme weather: typhoons, sandstorms, severe convection, and cold waves. The spatial range of impact is limited to a continuous grid area centered on the target new energy power station. The range of evolution speed corresponds to the average rate of change of real extreme weather from occurrence, development to dissipation. The temporal resolution, spatial resolution, and meteorological element dimensions of the generated sequence are consistent with the multi-source extreme weather meteorological dataset.

[0017] Furthermore, the national meteorological industry's extreme weather classification standards specifically include: typhoons are classified into five levels based on the tropical cyclone classification standard, according to the average wind speed over 2 minutes at a height of 10 meters above the ground: tropical storm, severe tropical storm, typhoon, severe typhoon, and super typhoon; dust storms are classified into five levels based on the dust storm weather level, according to horizontal visibility: floating dust, blowing sand, dust storm, severe dust storm, and extremely severe dust storm; cold waves are classified into three levels based on the cold wave classification standard, according to the temperature drop within 24 hours, the temperature drop within 48 hours, and the minimum temperature: cold wave, severe cold wave, and extremely severe cold wave; severe convection is classified based on the hourly precipitation of short-duration heavy rainfall and the basic reflectivity of weather radar. The short-duration heavy rainfall level is classified according to the hourly precipitation gradient, and the radar reflectivity level is classified according to the basic reflectivity numerical gradient. The classification results of the four types of extreme weather are directly used to set the duration and intensity parameters of the corresponding initial conditions.

[0018] Furthermore, the real meteorological-power pairing sample library is constructed based on at least 5 years of continuous historical operating data of the target new energy power stations. The data time resolution is uniformly set to 1 hour, and abnormal power data during equipment failure, power curtailment, and maintenance periods are removed, while pairing data under normal power generation conditions are retained. The meteorological data and power data in the sample library correspond one-to-one. The meteorological data includes core elements such as wind speed, temperature, air pressure, relative humidity, total horizontal irradiance, direct normal irradiance, and precipitation rate. The power data includes hourly measured output data of wind power and photovoltaic units. The sample library is divided into a mapping model training set and a test set in an 8:2 ratio for pre-training of the basic power mapping relationship.

[0019] Furthermore, the mathematical expression for the meteorological-power physics coupling generation algorithm is:

[0020]

[0021] in, The new energy coupling power generated at time t; To quantify the confidence weights, they are adaptively determined from the sample confidence intervals; Based on the generated meteorological field Basic power mapping output; To generate a spatiotemporal sequence of meteorological fields, t is a time dimension variable and s is a spatial dimension variable; This is a physical fidelity power term used to ensure the physical rationality of the generated power sequence and avoid abnormal power data that does not match meteorological elements; This is the physical compliance coefficient for the meteorological field; the larger the value, the higher the physical compliance of the generated meteorological field.

[0022] Furthermore, the statistical analysis and confidence calibration of the paired samples specifically includes: analyzing the time series of core meteorological elements and the power series of new energy sources within the meteorological-power paired samples, evaluating the time series fluctuation characteristics of each element and the degree of difference between the elements and the actual samples, and judging the overall physical rationality and time series stability of the samples in conjunction with the corresponding physical compliance coefficients; based on the above analysis results, confidence calibration is completed, taking the physical compliance level, time series stability, and power-meteorological element matching degree as the core calibration basis, comprehensively evaluating the sample confidence level, dividing the confidence level into continuous intervals and discrete levels, and matching a unique confidence label for each pair of paired samples for the subsequent construction of robust training datasets and sample weight allocation.

[0023] Furthermore, the continuous interval ranges from 0 to 1, with larger values ​​representing higher sample confidence. The discrete levels are set based on segmentation of the continuous interval, specifically including: extremely high confidence level corresponding to values ​​of 0.9 to 1 within the continuous interval, high confidence level corresponding to values ​​of 0.7 to 0.9, medium confidence level corresponding to values ​​of 0.4 to 0.7, low confidence level corresponding to values ​​of 0.1 to 0.4, and extremely low confidence level corresponding to values ​​of 0 to 0.1. Each discrete level corresponds to different physical rationality, temporal stationarity, and power-meteorological matching degree of the sample. A unique discrete level label is assigned to each pair of paired samples for subsequent construction of robust training datasets and sample weight allocation.

[0024] Furthermore, the robust training dataset consists of a mixture of meteorological-power enhancement samples with confidence labels and original real extreme weather samples in a preset ratio of 7:3. The dataset is categorized and integrated according to four extreme weather types: typhoons, sandstorms, severe convection, and cold waves. Each sample in the dataset is bound to a corresponding meteorological field sequence, new energy power sequence, physical compliance coefficient, and confidence label. The temporal and spatial resolution of the samples are consistent with the multi-source extreme weather meteorological dataset. Invalid samples with missing labels or incomplete elements are removed from the dataset. The dataset is partitioned and stored according to extreme weather type and confidence level to form a standardized robust training dataset adapted to the training input format of the new energy power prediction model.

[0025] Compared with existing technologies, this weather data preprocessing method for predicting new energy power has the following advantages:

[0026] I. This invention collects multi-source meteorological data and power operation data of new energy power plants in the target area, filters complete process samples of various extreme weather events and completes multi-dimensional association annotation, constructs a multi-source extreme weather meteorological dataset, builds a spatiotemporal diffusion model embedded with meteorological physical constraints, completes model training with the meteorological physical constraint diffusion loss algorithm as the core optimization criterion, generates spatiotemporal sequences of extreme weather meteorological fields that conform to atmospheric physical laws and completes physical compliance verification. It effectively solves the problems of scarce extreme weather samples, generation sequences that do not conform to physical evolution logic, and insufficient spatiotemporal continuity in existing technologies, makes up for the shortcomings of insufficient extreme weather sample coverage and inconsistent data quality in traditional preprocessing methods, improves the ability of meteorological data to characterize extreme weather processes, provides sufficient and compliant extreme weather meteorological data sources for new energy power prediction, and lays a high-quality data foundation for the training of subsequent power prediction models.

[0027] II. This invention constructs a real meteorological-power pairing sample library for target renewable energy power plants, completes the pre-training of basic power mapping relationships, and uses a meteorological-power physical coupling generation algorithm as the core. By combining qualified meteorological field sequences with corresponding physical compliance coefficients, it generates matching renewable energy power sequences and constructs meteorological-power pairing enhancement samples. Then, it completes the statistics and confidence calibration of the paired samples, matches the samples with corresponding confidence labels, and mixes them with the original real extreme weather samples to construct a standardized robust training dataset. This method realizes full-process physical consistency control of meteorological fields and power sequences, solves the problems of scarce meteorological-power pairing samples, inability to quantify sample confidence, and insufficient robustness of training datasets under extreme weather conditions, and provides multi-dimensional and graded training data for renewable energy power prediction models, improving the generalization ability and operational stability of prediction models in response to extreme weather.

[0028] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0030] Figure 1 A flowchart of a weather data preprocessing method for predicting new energy power output;

[0031] Figure 2 A schematic diagram of data transmission for a weather data preprocessing method used for new energy power forecasting;

[0032] Figure 3 This is a schematic diagram of data transmission for generating coupled samples according to the present invention. Detailed Implementation

[0033] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0034] Example 1:

[0035] This embodiment is applied to a large-scale offshore wind farm in a typhoon-prone area along the southeast coast of my country. This region is significantly affected by tropical cyclones, and sudden changes in meteorological elements such as wind speed and air pressure during typhoons directly cause large fluctuations in wind power. This places extremely high demands on the accuracy of power prediction at the wind farm and the stability of grid dispatch. The entire data processing process is carried out using the weather data preprocessing method for new energy power prediction of this invention. Precise data preprocessing is performed specifically for extreme weather events like typhoons, providing a standardized and robust training dataset for the typhoon-specific power prediction model built for this wind farm. This ensures that the prediction model has excellent generalization ability and prediction accuracy in typhoon scenarios. Figure 1 As shown.

[0036] Multi-source meteorological data was collected from the sea area where the offshore wind farm is located and its surrounding radiation area. Simultaneously, complete wind power operation data of the farm were also collected. The multi-source meteorological data includes global reanalysis meteorological data, measured data from ground meteorological stations, meteorological satellite remote sensing data, and weather radar detection data. These various meteorological data are synergistic and complementary, enabling a comprehensive capture of the complete meteorological evolution of typhoon weather from formation to development and dissipation. The power operation data includes minute-by-minute measured output data of wind turbine units, total installed capacity data, unit operating status data, and unit maintenance and power curtailment records, completely reconstructing the actual power generation conditions of the farm under typhoon weather. Complete process samples of extreme typhoon weather were specifically selected, and the selected samples underwent standardized preprocessing to remove data noise and invalid data, ensuring the validity and consistency of the sample data. Multi-dimensional correlation annotation was then carried out to achieve accurate correspondence between meteorological elements and power data. Finally, a multi-source extreme weather meteorological dataset for typhoon scenarios was constructed, laying a solid and realistic benchmark data foundation for subsequent model training, allowing subsequent model training to accurately reflect the meteorological and power characteristics of typhoon scenarios.

[0037] Based on the constructed meteorological and physical constraint spatiotemporal diffusion model, a multi-source extreme weather meteorological dataset of typhoon scenarios is used as the core training foundation. Three types of meteorological and physical constraints—atmospheric motion dynamics, atmospheric energy conservation, and atmospheric boundary layer physical parameterization—are deeply embedded in the model architecture. Simultaneously, a spatiotemporal continuity constraint on the meteorological field is added to ensure that the meteorological field generated by the model strictly follows the laws of atmospheric physical evolution, avoiding unrealistic abrupt changes in meteorological elements. The core optimization criterion is the meteorological and physical constraint diffusion loss algorithm, whose mathematical expression is:

[0038]

[0039] in, The general constraint loss for the meteorological and physical constrained spatiotemporal diffusion model; The loss is the original noise prediction loss for the diffusion model; For adaptive spatiotemporal physical weighting factors; These are the time step variables of the diffusion model, and also correspond to the time dimension of the meteorological spatiotemporal sequence. For the spatial grid coordinates of the meteorological field, corresponding to the spatial dimension of the meteorological sequence; These are atmospheric motion dynamics constraints. This is a constraint term related to the conservation of atmospheric energy. For the physical parameterization constraints of the atmospheric boundary layer; As a constraint term for the spatiotemporal continuity of the meteorological field, this algorithm comprehensively constrains and optimizes the model training process, ensuring that the model's training direction always aligns with the meteorological and physical characteristics of typhoon scenarios. The model undergoes continuous iterative training, with the dataset divided into training and validation sets in an 8:2 ratio. Initial learning rates of 2e-4, batch sizes of 16, and a maximum training epoch of 500 are set. Convergence is determined by the validation set loss remaining unchanged for 10 consecutive epochs. Real-time convergence verification accurately assesses the model's training effectiveness and fit, ensuring a stable training state. The final output is a trained model capable of generating typhoon meteorological fields that conform to atmospheric physical laws. This model accurately simulates the spatiotemporal variations of meteorological fields during typhoon weather, providing reliable model support for subsequent meteorological field generation.

[0040] Based on a trained typhoon scenario-specific model, and combined with the core meteorological element statistical characteristics of real typhoon samples from a multi-source extreme weather meteorological dataset for typhoon scenarios, this study fully references the changing patterns of meteorological elements in real typhoon weather. Simultaneously, it strictly adheres to the national meteorological industry's tropical cyclone classification standards to determine the initial conditions for typhoon extreme weather generation. This ensures that the initial conditions are both consistent with actual typhoon scenarios and comply with industry standards and specifications. Based on this range, a spatiotemporal sequence of typhoon extreme weather meteorological fields is generated. This sequence can completely recreate the distribution and changes of meteorological elements at different times and spaces during typhoon weather. Then, a comprehensive physical compliance check is performed on the generated sequence using a meteorological physics constraint diffusion loss algorithm. This algorithm judges the rationality and compliance of the generated sequence from a meteorological physics perspective, eliminating invalid sequences that do not conform to atmospheric physics laws. After passing the check, a qualified meteorological field sequence for the typhoon scenario and its corresponding physical compliance coefficient are output. The physical compliance coefficient accurately quantifies the physical compliance level of the meteorological field sequence, providing an important quantitative reference for subsequent coupled sample generation and ensuring the physical consistency between the subsequently generated power sequence and the meteorological field sequence.

[0041] Based on at least five years of continuous historical operational data from the offshore wind farm, this long-term continuous data comprehensively covers typhoon weather of varying intensities and evolutionary processes, fully reproducing the power generation characteristics of the farm under various typhoon scenarios. A realistic meteorological-power pairing sample library under typhoon scenarios is constructed. The data in the sample library undergoes refined preprocessing, eliminating abnormal power data caused by non-typhoon factors such as equipment failure, power curtailment, and maintenance, while retaining valid pairing data under normal power generation conditions. This ensures the sample library accurately reflects the actual correlation between typhoon weather and wind power. On this basis, pre-training of the basic power mapping relationship is completed, enabling the model to accurately learn the inherent correlation between meteorological elements and wind power under typhoon scenarios. Subsequently, using a meteorological-power physical coupling generation algorithm as the core, deep physical coupling between the meteorological field and the power sequence is achieved. The mathematical expression of the meteorological-power physical coupling generation algorithm is:

[0042]

[0043] in, The new energy coupling power generated at time t; To quantify confidence weights; Based on the generated meteorological field Basic power mapping output; To generate a spatiotemporal sequence of meteorological fields, t is a time dimension variable and s is a spatial dimension variable; This is the power term for physical fidelity. To establish a physical compliance coefficient for the meteorological field, qualified meteorological field sequences for typhoon scenarios are combined with their corresponding physical compliance coefficients. This ensures that the generated power sequences strictly match the physical characteristics of the meteorological field sequences, resulting in highly matched wind power sequences. Based on these sequences, enhanced meteorological-power pairing samples for typhoon scenarios are constructed, effectively addressing the scarcity of meteorological-power pairing samples in real typhoon scenarios and significantly enriching the sample quantity and scenario coverage for typhoon scenarios. Figure 3 As shown.

[0044] Comprehensive uncertainty statistics and confidence level calibration were conducted on meteorological-power paired samples for typhoon scenarios. Uncertainty statistics were used to accurately analyze the fluctuation characteristics and data differences of the samples. Confidence level labels were then matched based on the samples' physical compliance level, time-series stability, and the degree of matching between power and meteorological elements. This ensured that each sample had a corresponding quantitative quality label, accurately distinguishing the quality level of the samples. Labeled enhanced samples were then mixed with original real typhoon extreme weather samples at a preset ratio of 7:3, achieving complementary advantages between real and enhanced samples. This preserved the actual scenario characteristics of the real samples while also enhancing their performance. The dataset is rich in strong sample data and covers a wide range of scenarios. It is then categorized and integrated according to the type of extreme typhoon weather, making the classification of the dataset more targeted. Invalid samples with missing labels or incomplete elements are removed from the dataset to ensure its completeness and effectiveness. Finally, the dataset is partitioned and stored according to confidence level, which facilitates sample weight allocation and accurate retrieval during subsequent power prediction model training. The result is a robust training dataset that is suitable for the typhoon weather power prediction model of this wind farm. This dataset allows the power prediction model to fully learn the meteorological-power correlation patterns of various typhoon scenarios, which greatly improves the accuracy and stability of the model's power prediction under typhoon weather.

[0045] This embodiment addresses the practical needs of offshore wind farms in the southeastern coastal region, which are frequently hit by typhoons. It completes the weather data preprocessing for typhoon scenarios in five steps: first, collecting multi-source data to construct a typhoon-specific meteorological dataset; second, building a spatiotemporal diffusion model embedded with meteorological physical constraints and training it with a dedicated loss algorithm; third, generating and validating typhoon meteorological field sequences; fourth, constructing enhanced samples through coupled generation algorithms; and finally, mixing the samples and calibrating the confidence levels to form a robust training dataset. The entire process closely matches the characteristics of typhoon weather, compensating for the scarcity of real typhoon samples and ensuring the physical consistency between the meteorological field and the power sequence. The constructed dataset is highly adapted to the typhoon weather power prediction model for the wind farms, significantly improving the model's prediction accuracy and stability under typhoon conditions.

[0046] Example 2:

[0047] This embodiment is applied to a wind-solar hybrid renewable energy power station in the arid and semi-arid regions of northern my country. This region has complex climatic conditions, frequently affected by alternating extreme weather events such as sandstorms and cold waves. Sandstorms reduce atmospheric visibility and irradiance, while cold waves cause drastic changes in temperature and wind speed. Both types of weather have a significant cumulative impact on wind and solar power output, placing stringent demands on the accuracy of wind-solar hybrid power prediction. This invention utilizes a weather data preprocessing method for renewable energy power prediction, specifically targeting the entire data processing process for sandstorms and cold waves. This provides a highly adaptable and robust training dataset for the wind-solar hybrid power prediction model built for this power station, helping the model accurately capture the changing patterns of wind and solar power output under these two extreme weather conditions and improving the model's predictive capabilities in complex extreme weather scenarios. Figure 2 As shown.

[0048] Multi-source meteorological data was collected from the area where the wind-solar hybrid renewable energy power station is located and its surrounding areas. Simultaneously, comprehensive power operation data of the station's wind and solar power units were also collected. The multi-source meteorological data included global reanalysis meteorological data, ground meteorological station measured data, meteorological satellite remote sensing data, and weather radar detection data. Global reanalysis meteorological data constructed a spatiotemporally continuous atmospheric background field; ground meteorological station measured data calibrated the accuracy of near-surface meteorological elements; meteorological satellite remote sensing data captured the overall evolution of large-scale dust storms and cold waves; and weather radar detection data captured short-term meteorological element changes. These various data types, each with its own focus, worked in synergy to fully reconstruct the meteorological evolution characteristics of dust storms and cold waves. The power operation data included minute-by-minute measured output data of the wind and solar power units and the total installed capacity. Based on data from unit operation status and unit maintenance and power curtailment records, this study accurately reflects the actual power generation conditions of wind and solar power plants under two types of extreme weather. It focuses on selecting complete process samples of sandstorms and cold waves, taking into account samples of different intensities and durations. The selected samples undergo standardized preprocessing to unify data format and resolution, remove noise and outliers, and ensure the standardization and validity of the sample data. Multi-dimensional association and annotation are then completed to achieve accurate correspondence between meteorological elements and wind and solar power data. A multi-source extreme weather meteorological dataset containing sandstorm and cold wave scenarios is constructed, providing benchmark data that fits the actual conditions of areas prone to sandstorms and cold waves in northern China for subsequent model training. This allows the model training to accurately match the meteorological and power characteristics of the region.

[0049] A dedicated meteorologically constrained spatiotemporal diffusion model is built based on multi-source extreme weather meteorological datasets. This model architecture adapts to the changing characteristics of meteorological spatiotemporal sequences and can accurately handle meteorological spatiotemporal data under dust storms and cold waves. During model training, three types of meteorologically constrained parameters are deeply embedded: atmospheric dynamics, atmospheric energy conservation, and atmospheric boundary layer physical parameterization. This ensures that the generated meteorological field strictly follows atmospheric physical laws. Simultaneously, a spatiotemporal continuity constraint is introduced to avoid spatiotemporal jumps in the generated meteorological field sequence, guaranteeing its continuity and rationality. A meteorologically constrained diffusion loss algorithm is used as the core optimization criterion. This algorithm comprehensively constrains and optimizes the model training loss, enabling the model to accurately learn about dust storms and cold waves. Based on the meteorological and physical characteristics and spatiotemporal variation patterns of the model, a full-process iterative training process was conducted. The dataset was divided into training and validation sets in an 8:2 ratio. The initial learning rate was set to 2e-4, the batch size to 16, and the maximum number of training epochs to 500. The convergence criterion was that the validation set loss did not decrease for 10 consecutive epochs. The convergence verification of the model was continuously completed, and the training progress and fitting effect of the model were monitored in real time to ensure that the model reached a stable and convergent training state. The final output was a trained model that could generate meteorological fields of dust storms and cold waves that conformed to the laws of atmospheric physics. This model can accurately simulate the changes of meteorological elements at different times and in different spaces under these two types of extreme weather, providing reliable model support for the subsequent generation of meteorological fields for these two types of weather, and ensuring the accuracy and rationality of the generated meteorological fields.

[0050] Based on the statistical characteristics of core meteorological elements from real samples within multi-source extreme weather meteorological datasets for dust storms and cold waves, and closely aligning with the actual meteorological element variation patterns of dust storms and cold waves in the region, this study strictly adheres to the national meteorological industry's standards for classifying dust storm and cold wave levels. Targeted initial condition ranges for extreme weather generation are set for both dust storms and cold waves, ensuring that the initial conditions not only meet industry standards but also accurately match the weather characteristics of the region. Based on these set ranges, spatiotemporal sequences of meteorological fields for both types of extreme weather are generated. These sequences can completely reconstruct the distribution and changes of meteorological elements throughout the entire process of dust storms and cold waves, from occurrence and development to dissipation, accurately capturing irradiance and visibility under dust storm conditions. The algorithm comprehensively verifies the physical compliance of the two types of sequences generated by the meteorological physics constraint diffusion loss method, based on the characteristics of change and the sudden changes in temperature and wind speed under cold wave weather. It rigorously reviews the rationality of the sequences from the perspective of atmospheric physics, eliminates invalid sequences that do not conform to physical laws, and ensures that the output meteorological field sequences all conform to the actual meteorological evolution laws of dust storms and cold waves. Finally, it outputs qualified meteorological field sequences and corresponding physical compliance coefficients for dust storm and cold wave scenarios. The physical compliance coefficient can accurately quantify the physical compliance level of each set of meteorological field sequences, providing key quantitative basis for the subsequent generation of meteorological-power coupling samples for the two types of weather, and ensuring the physical consistency between the power sequence and the meteorological field sequence.

[0051] Utilizing at least five years of continuous historical operational data from this hybrid wind-solar power station, this long-term data comprehensively covers sandstorms and cold waves of varying intensities and combinations in the region. This data fully reconstructs the actual power generation correlation characteristics of wind and solar power under various extreme weather conditions, constructing a realistic meteorological-power pairing sample library for sandstorm and cold wave scenarios. The temporal resolution of the entire sample library is standardized, ensuring a high degree of matching between meteorological and power data in terms of time dimensions. Abnormal power data from equipment failures, power curtailments, and maintenance periods are removed, as these abnormal data are not caused by extreme weather and would interfere with the true correlation between meteorological and power. Valid pairing data under normal power generation conditions are retained, allowing the sample library to accurately reflect the inherent correlation between sandstorms, cold waves, and wind and solar power. Subsequently, the sample library is divided into a mapping model training set and a test set according to a fixed ratio, used for training and validating the basic power mapping relationship, respectively. Pre-training of the basic power mapping relationship enables the model to accurately learn the correlation between meteorological elements and wind and solar power under dust storms and cold waves at the site, as well as the superposition effect of the two. With the meteorological-power physical coupling generation algorithm as the core, the algorithm achieves deep physical coupling between the meteorological field and the power sequence. It combines the qualified meteorological field sequences of dust storms and cold waves with the corresponding physical compliance coefficients, so that the generated wind and solar power sequences can strictly match the physical characteristics and change patterns of the corresponding meteorological field sequences. This generates wind and solar power sequences that are highly matched with the meteorological field, thereby constructing meteorological-power pairing enhancement samples of wind and solar hybrid under dust storm and cold wave scenarios. This effectively makes up for the scarcity of meteorological-power pairing samples of wind and solar hybrid under dust storm and cold wave weather in this region and the insufficient scene coverage. It greatly enriches the number of samples and scene types under the two types of extreme weather, and provides sufficient high-quality samples for the construction of subsequent training sets.

[0052] Comprehensive uncertainty statistics and confidence level calibration were conducted on meteorological-power paired samples for dust storm and cold wave scenarios. The fluctuation patterns and data differences of the samples were analyzed for the different characteristics of the two weather types. Physical compliance level, temporal stability, and the degree of matching between power and meteorological elements were used as core calibration criteria to comprehensively evaluate the quality of each sample group. A unique confidence level label was assigned to each paired sample group. This label accurately quantifies the quality level of the sample, allowing subsequent model training to allocate sample weights differently based on the confidence level label. Labeled augmented samples were mixed with original real dust storm and cold wave extreme weather samples at a preset ratio of 7:3, combining the actual scene characteristics of real samples with the scene coverage advantages of augmented samples. This resulted in a dataset that is both realistic and rich, categorized into dust storms and cold waves as extreme weather types. The dataset is categorized and integrated by type to make its classification more targeted, facilitating subsequent training of the power prediction model for different extreme weather conditions. Invalid samples with missing labels or incomplete features are removed to ensure the completeness, standardization, and effectiveness of the dataset. Finally, the dataset is partitioned and stored according to extreme weather type and confidence level, achieving refined management and facilitating accurate retrieval and efficient utilization during subsequent model training. The final result is a standardized and robust training dataset adapted to the extreme weather power prediction model of this wind-solar hybrid power station. This dataset allows the wind-solar hybrid power prediction model to fully learn the variation patterns and superimposed effects of wind and solar power under two types of extreme weather: dust storms and cold waves. This significantly improves the model's prediction accuracy and generalization ability under these two types of extreme weather, ensuring the stable operation of power station power generation scheduling and the power grid under extreme weather conditions.

[0053] This embodiment focuses on wind-solar hybrid renewable energy power stations in northern regions prone to sandstorms and cold waves. It completes full-process weather data preprocessing for these two extreme weather conditions. A dedicated dataset is constructed through multi-source meteorological and power data acquisition. Training is performed using a meteorologically constrained spatiotemporal diffusion model and corresponding loss algorithm. Meteorological field sequences are generated and validated for different scenarios. A coupled generation algorithm is used to construct enhanced wind-solar hybrid samples. After confidence calibration and sample mixing, a standardized robust dataset is formed. The entire process takes into account the correlation between the meteorological characteristics of both weather types and wind and solar power, enriching the scarce extreme weather paired samples and ensuring the physical rationality of the data. The dataset supports the power station's wind-solar hybrid power prediction model in accurately learning the power change patterns of the two weather types, significantly improving the model's generalization ability and prediction accuracy.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A weather data preprocessing method for predicting new energy power output, characterized in that, The specific steps of this method are as follows: S100. Dataset Construction and Labeling: Collect multi-source meteorological data and power operation data of new energy power stations in the target area, screen samples of complete processes of four types of extreme weather, complete sample preprocessing and association labeling, and construct a multi-source extreme weather meteorological dataset. S200, Diffusion Model Construction and Training: Based on multi-source extreme weather meteorological datasets, a meteorological and physical constraint spatiotemporal diffusion model is built. The core optimization criterion is the meteorological and physical constraint diffusion loss algorithm. The model iterative training and convergence verification are completed, and the trained diffusion model is output. S300, Meteorological Field Generation Verification: Based on the trained diffusion model, the initial conditions range for extreme weather generation is set, and a spatiotemporal sequence of extreme weather meteorological fields is generated. The physical verification is completed based on the meteorological physical constraint diffusion loss algorithm, and a qualified meteorological field sequence and corresponding physical coefficients are output. S400, Coupled Sample Generation: Based on the historical operation data of the target new energy power station, a real meteorological-power pairing sample library is constructed, the basic power mapping relationship is pre-trained, and the meteorological-power physical coupling generation algorithm is used as the core. Combined with qualified meteorological field sequences and physical coefficients, the corresponding new energy power sequences are generated, and meteorological-power pairing enhanced samples are constructed. S500, Training Set Adaptation Construction: Complete the statistics and confidence labeling of paired samples, match the corresponding confidence labels to the samples, mix the labeled augmented samples with the original real extreme weather samples, and construct a robust training dataset.

2. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S100, the multi-source meteorological data includes global reanalysis meteorological data, measured data from ground meteorological stations, meteorological satellite remote sensing data, and weather radar detection data. Global reanalysis meteorological data is downloaded from a professional meteorological data platform via network download. Ground meteorological station data is collected in real-time from the regional meteorological monitoring center via wired communication. Meteorological satellite remote sensing data is received directionally via a satellite data receiving terminal. Weather radar data is retrieved in real-time through the radar station's data transmission interface. The power operation data of the new energy power plant includes minute-by-minute measured output data of the wind and photovoltaic units, total installed capacity data, unit operating status data, and unit maintenance and power curtailment records. All power data is collected in real-time and stored locally through a dedicated data acquisition interface for the power plant.

3. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S200, the meteorological physical constraint spatiotemporal diffusion model adopts an encoding and decoding network architecture adapted to meteorological spatiotemporal sequences. It performs progressive noise addition and reverse restoration on the input meteorological spatiotemporal data. During model training, three types of meteorological physical constraints are embedded: atmospheric motion dynamics, atmospheric energy conservation, and atmospheric boundary layer physical parameterization. At the same time, the spatiotemporal continuity constraint of the meteorological field is introduced. The multi-source extreme weather meteorological dataset is divided into training set and validation set according to a fixed ratio. An adaptive optimizer is used to iteratively update the model parameters. The continuous non-decreasing loss of the validation set is used as the convergence criterion. After training, the output can generate an extreme weather meteorological field model that conforms to the laws of atmospheric physics.

4. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S200, the mathematical expression for the meteorological physical constraint diffusion loss algorithm is: ； in, The general constraint loss for the meteorological and physical constrained spatiotemporal diffusion model; The loss is the original noise prediction loss for the diffusion model; For adaptive spatiotemporal physical weighting factors; These are the time step variables of the diffusion model, and also correspond to the time dimension of the meteorological spatiotemporal sequence. For the spatial grid coordinates of the meteorological field, corresponding to the spatial dimension of the meteorological sequence; These are atmospheric motion dynamics constraints. This is a constraint term related to the conservation of atmospheric energy. For the physical parameterization constraints of the atmospheric boundary layer; This is a constraint term on the spatiotemporal continuity of the meteorological field.

5. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S300, the range of initial conditions for the generation of extreme weather is determined based on the statistical characteristics of core meteorological elements of real extreme weather samples in the multi-source extreme weather meteorological dataset, combined with the meteorological industry's extreme weather level classification standards. Specifically, it includes the extreme value range of key meteorological elements, the range of process duration, the range of impact space, and the range of evolution speed. The extreme value range of key meteorological elements is taken as the 5%-95% quantile range of the corresponding elements in the real samples. The duration of the process is set to be between 24 and 120 hours; the spatial range of impact is limited to a continuous grid area centered on the target new energy power station; the evolution rate range corresponds to the average rate of change of real extreme weather from occurrence, development to dissipation.

6. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S400, the real meteorological-power matching sample library is constructed based on no less than 5 years of continuous historical operating data of the target new energy power station. The meteorological data and power data in the sample library correspond one-to-one. The meteorological data includes core elements such as wind speed, temperature, air pressure, relative humidity, total horizontal irradiance, normal direct irradiance, and precipitation rate. The power data includes hourly measured output data of wind power and photovoltaic units.

7. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S400, the mathematical expression of the meteorological-power physics coupling generation algorithm is: ； in, The new energy coupling power generated at time t; To quantify confidence weights; Based on the generated meteorological field Basic power mapping output; To generate a spatiotemporal sequence of meteorological fields, t is a time dimension variable and s is a spatial dimension variable; This is the power term for physical fidelity. This is the physical compliance coefficient for the meteorological field.

8. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S500, the statistical analysis and confidence calibration of the paired samples specifically includes: analyzing the time series of core meteorological elements and the power series of new energy sources within the meteorological-power paired samples, evaluating the time series fluctuation characteristics of each element and the degree of difference between the elements and the real samples, and judging the overall physical rationality and time series stability of the samples in conjunction with the corresponding physical compliance coefficients; based on the above analysis results, confidence calibration is completed, taking the physical compliance level, time series stability, and power-meteorological element matching degree as the core calibration basis, comprehensively evaluating the sample confidence level, dividing the confidence level into continuous intervals and discrete levels, and matching a unique confidence label for each pair of paired samples for the subsequent construction of robust training datasets and sample weight allocation.

9. The weather data preprocessing method for new energy power prediction according to claim 1, characterized in that, In step S500, the robust training dataset is composed of a mixture of meteorological-power enhancement samples with confidence labels and original real extreme weather samples. The dataset is categorized and integrated according to four extreme weather types: typhoon, sandstorm, severe convection, and cold wave. Each sample in the dataset is bound to a corresponding meteorological field sequence, new energy power sequence, physical compliance coefficient, and confidence label. The temporal and spatial resolution of the samples are consistent with the multi-source extreme weather meteorological dataset. Invalid samples with missing labels or incomplete elements are removed from the dataset. The dataset is partitioned and stored according to extreme weather type and confidence level to form a standardized robust training dataset that is adapted to the training input format of the new energy power prediction model.

Images