Scene generation method and system based on search-enhanced probability downscaling

Abstract

This invention relates to the field of power system operation and planning technology, and proposes a scenario generation method and system based on retrieval-enhanced probabilistic downscaling. It constructs wind power output conversion models and photovoltaic power output conversion models that consider the coupling of multiple meteorological variables, utilizes the TCVAE model, and performs probabilistic sampling of meteorological residuals. A multi-source condition vector is constructed, consisting of daily-scale meteorological statistics, hourly-level meteorological prior profiles of similar days, and meteorological states at the end of the previous day, serving as macroscopic constraints for intraday temporal evolution. During the extrapolation phase, the generated residuals are superimposed with the prior profiles of similar days to reconstruct the meteorological scenario, and a wedge correction algorithm is used to eliminate numerical jumps at the inter-day junctions. This invention effectively improves the physical consistency and temporal continuity of downscaled scenarios. By performing hourly-level temporal downscaling on coarse-time-resolution meteorological model data, it outputs a high-fidelity meteorological-wind-solar scenario set that meets the requirements of long-cycle simulation of new power systems.

Description

Technical Field

[0001] This invention relates to the technical field of power system operation and planning, specifically to a scenario generation method and system based on retrieval-enhanced probabilistic downscaling. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the continuous increase in the installed capacity of new energy sources such as wind power and photovoltaics, the dependence of power system operation on meteorological conditions has significantly increased. Meteorological variables such as wind speed, solar irradiance, temperature, and humidity jointly affect the intraday fluctuations, seasonal variations, and long-term evolution trends of wind and photovoltaic output, and further influence system power balance, reserve capacity configuration, energy storage capacity planning, and flexible resource dispatch. Against the backdrop of climate change, meteorological processes exhibit stronger non-stationarity and uncertainty, and extreme weather events have a more pronounced impact on the distribution of wind and solar resources and the output of new energy sources. Therefore, in the planning and operation assessment of high-proportion new energy power systems, it is necessary to construct a year-round, continuous, hourly-resolution meteorological-wind and solar output scenario that simultaneously possesses statistical realism and physical consistency to support long-term operation simulation and new energy consumption analysis.

[0004] Existing long-term climate model data typically consists mainly of daily-scale statistics, which are insufficient to directly meet the needs of hourly-scale power system simulations. Using simple interpolation or traditional statistical downscaling methods usually only yields relatively smooth hourly sequences, failing to recover true meteorological fluctuations such as abrupt changes in wind speed and rapid variations in irradiance. Conversely, directly using generative models to generate hourly sequences often lacks prior physical constraints from historically similar meteorological processes, leading to unreasonable intraday variation patterns. Furthermore, existing daily generation methods typically generate daily scenarios independently, easily causing numerical jumps between the end of the previous day and the beginning of the next, affecting the continuity of long-term scheduling simulations. Existing wind and solar power output scenario generation methods often directly generate wind and solar power or use simplified empirical formulas for conversion, failing to fully reflect the impact of physical factors such as moist air density, wind farm spatial aggregation effects, radiative decomposition, oblique projection, and battery temperature changes on wind and solar power output, resulting in insufficient physical consistency and engineering applicability of the generated results. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a scene generation method and system based on retrieval-enhanced probabilistic downscaling, achieving hourly meteorological scene generation that combines physical prior constraints, stochastic fluctuation representation, and diurnal continuity. A wedge correction algorithm ensures the diurnal statistical consistency and inter-diurnal continuity of the generated meteorological scenes. Combined with a wind-solar physical conversion model that considers moist air density, wind farm spatial aggregation effects, radiative decomposition, oblique projection, and battery temperature correction, a year-round, physically consistent meteorological-wind-solar output scene set is formed.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a scene generation method based on retrieval-enhanced probability downscaling, comprising the following steps: Based on the constructed historical observation day database, similar day sets are retrieved within the sliding time window of the target day using the rank sum mechanism, and hourly meteorological profiles are extracted and scaled to obtain hourly physical prior profiles of the target day. The daily meteorological statistics of the target day, the hourly physical prior profile, and the meteorological state at the end of the previous day are concatenated into a multi-source conditional vector. Based on the trained TCVAE model, the hourly meteorological residual is generated by latent variable sampling. The hourly meteorological residual is then superimposed with the hourly physical prior profile to generate the original hourly meteorological scene. The wedge correction algorithm is used to correct the original hourly meteorological scene, and the corrected meteorological state at the end of the current day is passed to the generation process of the next target day to obtain the corrected hourly meteorological scene that meets the target day average constraint and the inter-day continuity requirement. The corrected hourly meteorological scene is input into the wind power output conversion model and the photovoltaic output conversion model to obtain the wind power output scene and the photovoltaic output scene. They are then spliced ​​together in chronological order to obtain a continuous meteorological-wind and solar power output scene set.

[0007] A further technical solution, the process of constructing an hourly physical prior profile, includes the following steps: Historical meteorological data of the study area were acquired to construct a historical observation database D. hist And based on the target date, a sliding time window is constructed to filter candidate similar days; Based on the rank sum mechanism, multidimensional similarity retrieval is performed on candidate historical samples, and a set of similar days is selected. Hourly meteorological profiles are extracted from similar day sets and linear scaling corrections are applied to obtain hourly physical prior profiles for the target day.

[0008] A further technical solution is that the TCVAE model includes a Transformer encoder, a latent variable parameterization module, a reparameterization sampling module, and a Transformer decoder connected in sequence. The Transformer encoder is used to extract features from the input hourly multivariate meteorological sequence and multi-source conditional vector to obtain the corresponding temporal latent features; The latent variable parameterization module is used to map the temporal latent features extracted by the Transformer encoder into latent variable distribution parameters; The reparameterized sampling module is configured to sample noise ε from a standard normal distribution and generate latent variables z based on the mean vector μ and standard deviation vector σ in the latent variable distribution parameters; The Transformer decoder is used to fuse the latent variable z with the standardized multi-source condition vector c and output the hourly meteorological residual sequence.

[0009] A further technical solution involves inputting multi-source conditional vectors into a trained TCVAE model, generating hourly meteorological residuals through latent variable sampling, and then superimposing these hourly meteorological residuals with the hourly physical prior profile to generate the original hourly meteorological scene. This process includes the following steps: Based on the latent variable distribution parameters obtained during the TCVAE model training process, random sampling is performed based on the reparameterization mechanism and the latent space probability distribution to generate latent variable samples z. Based on the Transformer decoder, the temporal feature fusion and residual reconstruction are performed, and the sampled latent variable sample z and the standardized multi-source condition vector c′ are subjected to nonlinear interaction and feature diffusion to obtain the standardized hourly meteorological residual sequence. The standardized hourly meteorological residual series is inversely standardized to obtain hourly meteorological residuals with physical dimensions; The hourly physical prior profile is linearly superimposed with the hourly meteorological residual to generate the original hourly meteorological scene.

[0010] A further technical solution employs a wedge correction algorithm to correct the original hourly meteorological scene, and then transfers the corrected meteorological state at the end of the current day to the generation process of the next target day, resulting in a corrected hourly meteorological scene that meets the target day's average constraint and inter-day continuity requirements. This includes the following steps: The daily average deviation is calculated to construct a wedge-shaped correction term that changes linearly with time, and the original hourly meteorological scene is corrected to obtain the corrected hourly meteorological scene. Extract the last moment's meteorological state from the corrected hourly meteorological scene of the current target day, and use it as the initial meteorological state for the next target day. After generating the original hourly meteorological scene for the next target day, make corrections, and so on, to obtain the corrected hourly meteorological scene that meets the target day's average constraint and the requirement of inter-day continuity.

[0011] A further technical solution involves calculating the daily average deviation to construct a wedge-shaped correction term that changes linearly over time, and then correcting the original hourly meteorological scene to obtain the corrected hourly meteorological scene, including the following: x the actual daily average weather value of the original hourly weather scenario mean,i Compared with the target daily meteorological mean x target,i The difference is calculated to obtain the daily average deviation. A wedge-shaped correction term that linearly varies with the hourly sequence number within the target day is constructed. The daily average deviation is distributed to the original hourly meteorological scene at each time point to obtain the corrected hourly meteorological scene for the current target day.

[0012] Further technical solutions include wind power output conversion model that incorporates humid air density correction and wind farm spatial aggregation effect, and photovoltaic output conversion model that incorporates radiation decomposition, oblique projection and battery temperature correction. The corrected hourly meteorological scene is then input into the wind power output conversion model and the photovoltaic output conversion model to obtain wind power output and photovoltaic output.

[0013] A second aspect of the present invention provides a scene generation system based on retrieval-enhanced probabilistic downscaling, comprising: The data augmentation retrieval module is configured to retrieve a set of similar days based on the constructed historical observation day database within the sliding time window of the target day using the rank sum mechanism, extract and scale the corrected hourly meteorological profile, and obtain the hourly physical prior profile of the target day. The TCVAE extrapolation module is configured to concatenate the daily meteorological statistics of the target day, the hourly physical prior profile, and the meteorological state at the end of the previous day into a multi-source conditional vector. Based on the trained TCVAE model, it generates hourly meteorological residuals through latent variable sampling and superimposes the hourly meteorological residuals with the hourly physical prior profile to generate the original hourly meteorological scene. The correction module is configured to use a wedge correction algorithm to correct the original hourly meteorological scene and pass the corrected meteorological state at the end of the current day to the next target day generation process to obtain a corrected hourly meteorological scene that meets the target day average constraint and inter-day continuity requirements. The scene generation module is configured to input the corrected hourly meteorological scene into the wind power output conversion model and the photovoltaic output conversion model to obtain the wind power output scene and the photovoltaic output scene, and then splice them in time order to obtain a continuous meteorological-wind and solar power output scene set.

[0014] A further technical solution involves constructing an hourly-level physical prior profile in the data augmentation retrieval module, which includes the following steps: Historical meteorological data of the study area were acquired to construct a historical observation database D. hist And based on the target date, a sliding time window is constructed to filter candidate similar days; Based on the rank sum mechanism, multidimensional similarity retrieval is performed on candidate historical samples, and a set of similar days is selected. Hourly meteorological profiles are extracted from similar day sets and linear scaling corrections are applied to obtain hourly physical prior profiles for the target day.

[0015] In a further technical solution, the TCVAE simulation module is configured to perform the following steps: Based on the latent variable distribution parameters obtained during the TCVAE model training process, random sampling is performed based on the reparameterization mechanism and the latent space probability distribution to generate latent variable samples z. Based on the Transformer decoder, the temporal feature fusion and residual reconstruction are performed, and the sampled latent variable sample z and the standardized multi-source condition vector c′ are subjected to nonlinear interaction and feature diffusion to obtain the standardized hourly meteorological residual sequence. The standardized hourly meteorological residual series is inversely standardized to obtain hourly meteorological residuals with physical dimensions; The hourly physical prior profile is linearly superimposed with the hourly meteorological residual to generate the original hourly meteorological scene.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention retrieves a set of similar days using a historical observation day database and a rank-sum mechanism, and scales and corrects the hourly meteorological profiles of similar days. This allows for the introduction of intraday variations in real historical meteorological processes while satisfying the statistical constraints of the target day. This avoids the problems of overly smooth sequences generated by simple interpolation methods, which struggle to recover abrupt changes in wind speed and rapid changes in irradiance. The hourly physical prior profile, along with the meteorological statistics of the target day and the meteorological state at the end of the previous day, are used as multi-source conditional vectors input into the trained TCVAE model. Hourly meteorological residuals are generated through latent variable sampling, ensuring that the generated results retain the physical process characteristics of similar historical days while also possessing the ability to express random fluctuations using a probabilistic generation model, thus improving the realism and diversity of hourly meteorological scenarios. This addresses the issue that existing day-by-day independent generation methods are prone to scaling and correcting the intraday variations at the end of the previous day. To address the issue of numerical jumps between the initial moments of the following day, this method employs a wedge correction algorithm to correct the daily mean deviation and inter-day continuity of the original hourly meteorological scene. This ensures that the generated scene meets the statistical conditions of the target day while reducing the abrupt changes at the cross-day boundary, thereby improving the continuity and stability of the long-term meteorological sequence. The corrected hourly meteorological scene is then input into the wind power output conversion model and the photovoltaic power output conversion model, respectively. This allows the wind power output conversion to incorporate factors such as moist air density, wind turbine power curves, and the spatial aggregation effect of the wind farm, while the photovoltaic power output conversion can incorporate factors such as radiation decomposition, oblique projection, and battery temperature changes. This improves the physical consistency and engineering applicability between the meteorological scene and the wind and solar power output scene, ultimately resulting in a continuous meteorological-wind and solar power output scene set suitable for long-term scheduling simulation and new energy consumption analysis.

[0017] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is an overall flowchart of the scene generation method based on retrieval enhancement probability downscaling in Embodiment 1 of the present invention; Figure 2 This is a data flow diagram of the TCVAE inference architecture based on retrieval enhancement in Embodiment 1 of the present invention; Figure 3 This refers to the hourly meteorological sequence obtained throughout the year using the scene generation method based on retrieval-enhanced probability downscaling in Embodiment 1 of the present invention. Figure 4 This is a comparison chart of a typical 168-hour meteorological sequence between the scene generation method and the comparison method based on retrieval enhancement probability downscaling in Embodiment 1 of the present invention. Figure 5 The hourly wind and solar power output sequence for the whole year is obtained by the scene generation method based on retrieval enhancement probability downscaling in Embodiment 1 of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0023] This invention introduces a physical prior embedding mechanism. Under the macroscopic boundary constraints of daily meteorological model data, it utilizes a similar day retrieval mechanism to extract intraday evolutionary patterns with physical fidelity. A Transformer-based Conditional Variational Autoencoder (TCVAE) model is then used to stochastically model hourly meteorological residuals. This generates a high-quality meteorological-wind and solar data scene set that combines physical consistency, statistical realism, and temporal continuity, effectively compensating for the statistical structure drift and morphological distortion problems that easily occur in purely data-driven models when depicting complex meteorological processes. Specific embodiments are described below.

[0024] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 5 As shown, a scene generation method based on retrieval enhancement probability downscaling includes the following steps: Step S1: Construct hourly physical prior profile: Based on the constructed historical observation day database, retrieve the set of similar days within the sliding time window of the target day based on the rank sum mechanism, extract and scale the corrected hourly meteorological profile, and obtain the hourly physical prior profile of the target day. Step S2: Generate the original hourly meteorological scene: The daily meteorological statistics of the target day, the hourly physical prior profile and the meteorological state at the end of the previous day are concatenated into a multi-source conditional vector. Based on the trained TCVAE model, the hourly meteorological residual is generated by sampling latent variables. The hourly meteorological residual is then superimposed with the hourly physical prior profile to generate the original hourly meteorological scene. Step S3: Correct the original hourly meteorological scene: The wedge correction algorithm is used to correct the original hourly meteorological scene, and the corrected meteorological state at the end of the current day is passed to the next target day generation process to obtain the corrected hourly meteorological scene that meets the target day average constraint and the inter-day continuity requirement. Step S4: Generate a meteorological-wind and solar power output scene set: Input the corrected hourly meteorological scene into the wind power output conversion model and the photovoltaic power output conversion model to obtain the wind power output scene and the photovoltaic power output scene, and splice them in time order to obtain a continuous meteorological-wind and solar power output scene set.

[0025] In this embodiment, a set of similar days is retrieved through a historical observation day database and a rank sum mechanism. The hourly meteorological profiles of these similar days are then scaled and corrected. This allows for the introduction of intraday variations in real historical meteorological processes while satisfying the statistical constraints of the target day. This avoids the problems of overly smooth sequences generated by simple interpolation methods, which struggle to recover abrupt changes in wind speed and rapid changes in irradiance. The hourly physical prior profile, along with the meteorological statistics of the target day and the meteorological state at the end of the previous day, are used as multi-source conditional vectors input into the trained TCVAE model. Hourly meteorological residuals are generated through latent variable sampling, ensuring that the generated results retain the physical process characteristics of similar historical days while also possessing the ability to express random fluctuations using a probabilistic generation model. This improves the realism and diversity of hourly meteorological scenarios. Furthermore, this addresses the issue that existing day-by-day independent generation methods are prone to errors at the end of the previous day. To address the issue of numerical jumps between the current moment and the initial moment of the following day, this method employs a wedge correction algorithm to correct the daily mean deviation and inter-day continuity of the original hourly meteorological scene. This ensures that the generated scene meets the statistical conditions of the target day while reducing the abrupt change at the cross-day boundary, thereby improving the continuity and stability of the long-term meteorological sequence. The corrected hourly meteorological scene is then input into the wind power output conversion model and the photovoltaic power output conversion model, respectively. This allows the wind power output conversion to incorporate factors such as moist air density, wind turbine power curves, and the spatial aggregation effect of the wind farm, while the photovoltaic power output conversion can incorporate factors such as radiation decomposition, oblique projection, and battery temperature changes. This improves the physical consistency and engineering applicability between the meteorological scene and the wind and solar power output scene, ultimately resulting in a continuous meteorological-wind and solar power output scene set suitable for long-term scheduling simulation and new energy consumption analysis.

[0026] Step S1, the process of constructing the hourly physical prior profile for the target day, includes the following steps: Step S11: Obtain historical meteorological data for the study area and construct a historical observation database D. hist And based on the target date, a sliding time window is constructed to filter candidate similar days; Specifically, long-term historical hourly meteorological data of the study area are obtained, and the meteorological data are divided according to natural days to construct a historical observation day database; One specific implementation involves obtaining long-term historical hourly reanalysis datasets for the study area. These datasets include, but are not limited to, the ERA5 dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF) or the MERRA-2 dataset released by NASA.

[0027] Optionally, historical hourly meteorological data includes meteorological elements such as temperature, specific humidity, solar irradiance, and wind speed. To achieve consistency in coupling among multiple variables, this embodiment treats temperature, specific humidity, solar irradiance, and wind speed as a single meteorological vector to characterize the intraday evolution trajectory of multiple meteorological elements within a complete natural day. The historical hourly meteorological data is divided into multiple independent diurnal profile samples according to date, resulting in a historical observation daily database. Furthermore, in this step, a sliding time window corresponding to the target day is constructed in the historical observation day database, and candidate historical samples are screened based on the sliding time window; To eliminate seasonal interference and focus on similar meteorological backgrounds, this embodiment introduces a time-sliding window mechanism during retrieval, setting the window size to W. If the target day is D... target During retrieval, only the historical database D is searched. hist [D] for all years target -W,D target Samples were screened within a range of +W] days to ensure that candidate similar days have similar climate change patterns and seasonal evolution characteristics.

[0028] Step S12: Perform multi-dimensional similarity retrieval on candidate historical samples based on the rank sum mechanism, and filter the set of similar days, including: To address the Euclidean distance distortion caused by differences in the dimensions of various meteorological elements, this step employs the rank-sum algorithm for similarity quantification and extracts the intraday evolution pattern that best matches the macroscopic characteristics of the target day from historical high-resolution observation data through morphological scaling. Hourly physical prior profiles are extracted and scaled from the set of similar days, enabling daily-scale meteorological statistics to be converted into hourly meteorological base sequences with real historical intraday evolution patterns, thus achieving hourly simulation of power systems.

[0029] Step S121: Construct the daily-scale meteorological vector of the target day and calculate the single absolute deviation between the candidate historical samples and the target day; Specifically, for the target day to be generated, the average meteorological value x for the target day is constructed based on daily-scale meteorological information.target,i The four-dimensional target daily average value vector can include daily average temperature, daily average humidity, daily average irradiance, and daily average wind speed; For each candidate historical sample h within the sliding time window, calculate the historical daily average value C of the i-th meteorological variable. h,i The average weather value C corresponding to the target day d,i The absolute deviation between the two terms e i,h : ; Step S122: Sort the individual absolute deviations of each candidate historical sample to obtain the corresponding variable ranking, and obtain the total rank and score by weighted summation; Specifically, for all samples in the candidate set, e i,h Sort the samples in ascending order to obtain the rank R of sample h in the i-th dimension variable. i,h The total rank and score S are obtained by weighted summation. h : ; In the formula, w rank,i These are the retrieval weight coefficients for each variable. This mechanism enhances the robustness of retrieval results to multivariate coupling features by transforming numerical differences into ordinal differences.

[0030] This embodiment transforms the numerical differences between different meteorological variables into ordinal differences, avoiding the Euclidean distance distortion problem caused by the different dimensions and numerical ranges of variables such as temperature, humidity, irradiance, and wind speed, and improving the robustness of similar day retrieval results to multivariate coupling features.

[0031] Step S123: Filter the set of similar days based on the total rank and score; Step S13: Extract hourly meteorological profiles based on similar day sets and perform linear scaling corrections to obtain hourly physical prior profiles for the target day; Select the above S h The smallest K samples in the set S form the similarity set. d The mean of hourly meteorological profiles within the set is extracted as the original prior, and the mean of the target day's meteorological profile is used. target,i Similar daily ensemble meteorological mean x S,i The ratio is corrected by linear scaling: ; In the formula, x prior,i(t) The physical prior profile data of the i-th dimension meteorological variable at time t constitute a 96-dimensional (24-hour × 4 variables) physical prior sequence that characterizes the typical evolution of multiple meteorological elements under this climate background.

[0032] This embodiment, through scaling correction, ensures that the original physical prior profile maintains the intraday evolution pattern of similar day sets while keeping in line with the macroscopic daily meteorological characteristics of the target day. In other words, the obtained hourly physical prior profile not only preserves the relatively realistic intraday evolution pattern of historical similar days but also matches the daily meteorological level of the target day.

[0033] After the above processing, the hourly physical prior profile corresponding to the target day is obtained. The hourly physical prior profile consists of four meteorological variables within 24 hours: temperature, specific humidity, solar irradiance, and wind speed, forming a 96-dimensional physical prior sequence, i.e., 24 hours × 4 variables. This physical prior sequence is used to characterize the typical intraday evolution of multiple meteorological elements under the climate background of the target day, and serves as the physical constraint input for subsequent hourly meteorological element reconstruction or prediction models.

[0034] Simple interpolation or traditional statistical downscaling methods typically only yield relatively smooth hourly sequences, failing to recover true meteorological fluctuations such as abrupt changes in wind speed and rapid variations in irradiance. Directly generating hourly sequences using generative models often lacks physical prior constraints from historically similar meteorological processes, leading to unreasonable intraday variation patterns. To address this issue, this embodiment concatenates daily meteorological statistics, hourly physical prior profiles of similar days, and meteorological states at the end of the previous day into a multi-source conditional vector. This vector is then input into a trained Transformer-enhanced Conditional Variational Autoencoder (TCVAE). Hourly meteorological residuals are generated through latent variable sampling. These residuals are then superimposed with the hourly physical prior profiles to generate the original hourly meteorological scene, thereby supplementing random fluctuation features while preserving the physical patterns of similar days.

[0035] In step S2, the daily meteorological statistics of the target day, the hourly physical prior profile, and the meteorological state at the end of the previous day are concatenated into a multi-source conditional vector c. Specifically, the daily-scale meteorological statistics C for the target day d (4-dimensional), obtained based on the daily meteorological data corresponding to the target day, may include the daily average temperature, daily average humidity, daily average solar radiation, and daily average wind speed of the target day, denoted as C. d The vector C d Used to characterize the overall meteorological level of the target day on a daily scale, so that the subsequently generated hourly meteorological sequences can meet the daily average statistical characteristic constraints within a 24-hour cycle, and avoid the generated results deviating from the macro-climate background of the target day.

[0036] Physical prior profile x prior (96 dimensions): This refers to the similar daily hourly meteorological data obtained through enhanced retrieval in step S1. The physical prior profile provides the model with an interpretable intraday evolutionary subject, guiding the model to generate sequences that conform to the laws of meteorological physical evolution and multivariate coupling relationships; Weather conditions at the end of the previous day x anchor (4-dimensional): Obtained from meteorological observation data or reanalysis data at 23:00 the day before the target date, including temperature, humidity, solar irradiance, and wind speed at 23:00 the previous day, denoted as x. anchor By incorporating the meteorological conditions at the end of the previous day, the model can perceive the meteorological conditions at the moment before the target day when generating hourly meteorological sequences for the target day, thereby enhancing the continuity of cross-day time series evolution and reducing numerical jumps at the splicing positions of adjacent dates.

[0037] In this embodiment, a multi-source conditional vector c with dimension 104 is constructed, and the formula is expressed as follows: ; By using the above splicing method, daily-scale macroscopic statistical constraints, hourly intraday evolution patterns, and cross-day boundary continuity information are simultaneously integrated within the same feature space.

[0038] Further technical solutions, in order to eliminate the impact of differences in the dimensions and numerical ranges of different meteorological variables on model training and to adapt to the input requirements of subsequent Transformer models, this embodiment performs standardization processing on the multi-source condition vector.

[0039] This embodiment is not based on local scaling of a single sample, but rather on a historical observation database D. hist Calculate the global statistical sub-component. Let X be the spatiotemporal sample set of the i-th dimension meteorological variable in the historical observation database. i Calculate its global arithmetic mean μ hist,i Compared with the global standard deviation σ hist,i : ; In the formula, N represents the number of historical days; T=24 represents the number of hours per day; x i,n,t Let be the value of the i-th dimension meteorological variable at time t on the n-th day in history.

[0040] For each element c in the 104-dimensional multi-source conditional vector c obtained after the above concatenation... j Based on the type of meteorological variable to which it belongs, the historical global mean μ of the corresponding meteorological variable is selected. hist,i and historical global standard deviation σ hist,i The Z-Score method is used for standardization to obtain the standardized elements. : ; In the formula, The elements are in the standardized multi-source condition vector; meteorological variable types include temperature, humidity, solar radiation, and wind speed; this transformation maps the original features with physical dimensions to a dimensionless feature space with a mean of 0 and a variance of 1.

[0041] After the above processing, a standardized multi-source condition vector c' is obtained. This multi-source condition vector c' simultaneously contains the diurnal statistical constraints of the target day, the hourly physical prior evolution pattern, and the meteorological state at the end of the previous day. It can be used as input for subsequent meteorological data prediction models to generate hourly meteorological sequences that meet the requirements of diurnal statistical characteristics, intraday physical evolution patterns, and cross-day continuity.

[0042] In step S2, the multi-source conditional vector is input into the trained Transformer Enhanced Conditional Variational Autoencoder (TCVAE) model, and hourly meteorological residuals are generated through latent variable sampling. The hourly meteorological residuals are then superimposed with the hourly physical prior profile to generate the original hourly meteorological scene. Step S2 includes the TCVAE model structure construction, model training, and the generation of hourly meteorological scenes based on the trained model.

[0043] In some embodiments, the TCVAE model is a Transformer-enhanced conditional variational autoencoder used to establish a probabilistic mapping relationship between the multi-source conditional vector c and the hourly meteorological residuals; wherein, the hourly meteorological residuals are used to characterize the deviation of the real hourly meteorological observation sequence from the hourly physical prior profile. Furthermore, the TCVAE model includes a Transformer encoder, a latent variable parameterization module, a reparameterization sampling module, and a Transformer decoder connected in sequence. The Transformer encoder is used to extract features from the input hourly multivariate meteorological sequence and multi-source conditional vector to obtain the corresponding temporal latent features; The latent variable parameterization module is used to map the temporal latent features extracted by the Transformer encoder into latent variable distribution parameters; The reparameterized sampling module is configured to sample noise ε from a standard normal distribution and generate latent variables z based on the mean vector μ and standard deviation vector σ in the latent variable distribution parameters, which is used to introduce randomness while keeping the model end-to-end trainable; The Transformer decoder is used to fuse the latent variable z with the standardized multi-source conditional vector c and output the hourly meteorological residual sequence. Specifically, the Transformer encoder is used to extract features from the input hourly multivariate meteorological sequence and multi-source conditional vectors. The hourly multivariate meteorological sequence is the given data divided by hours; for a given multivariate meteorological sequence... (Where T=24 is the number of time periods, and d=4 is the number of variables), first, it is mapped to the hidden dimension d through a linear embedding layer. model The model incorporates location coding to embed intraday absolute time information, enabling it to perceive meteorological changes at different times, such as temperature changes in the morning and evening, increased solar radiation at midday, and wind speed fluctuations.

[0044] The Transformer encoder has multiple layers of self-attention modules inside, which learn the correlation between meteorological variables and between different time points in different feature subspaces through a multi-head self-attention mechanism.

[0045] By calculating the attention weights of the dot product of the query matrix Q, the key matrix K, and the value matrix V, the model can dynamically identify the temporal synchronicity between sudden changes in wind speed and a sharp decrease in irradiance. The weight allocation mechanism is expressed as follows: ; In the formula, d k The scaling factor is used to prevent gradient vanishing in the Softmax saturation region and maintain gradient stability during deep network training. This architecture effectively solves the long-term gradient vanishing problem that traditional recurrent neural networks struggle to handle, ensuring the integrity of the 24-hour weather profile.

[0046] The core task of the TCVAE model is to achieve dimensionality reduction and probabilistic modeling of the characteristics of meteorological random fluctuations by introducing a continuous latent variable z. Specifically, the latent variable parameterization module can use a fully connected layer to map the temporal latent features extracted by the Transformer encoder into a mean vector μ and a log-variance vector logσ. 2 These two vectors together define the probability distribution interval of the current weather state in the latent space.

[0047] The reparameterization sampling module introduces a reparameterization mechanism to enable end-to-end training of the model via backpropagation. This mechanism removes randomness from the computational graph by sampling a standard normally distributed noise. The latent variables are defined as random mappings of the mean vector μ and the scale vector σ: ; In the formula, This represents element-wise multiplication. The latent variable z contains all the uncertainty information of the intraday hourly fluctuations of four-dimensional meteorological elements under a given macro-weather background c, and serves as the random source for the decoder to generate the residual sequence.

[0048] The Transformer decoder fuses the latent variable z with the standardized multi-source conditional vector and outputs an hourly meteorological residual sequence. During the decoding process, the latent variable z is extended along the time dimension to T=24 time points and concatenated with the conditional features corresponding to the multi-source conditional vector. Feature interaction is performed through the multi-head self-attention module in the Transformer decoder, and finally, the standardized meteorological residual prediction sequence of 24 hours × 4 variables is obtained by mapping through the linear output layer.

[0049] Further technical solutions, the training process of the TCVAE model includes the following steps: Step 21: Construct training samples for the TCVAE model and standardize the training samples. Specifically, training samples are constructed based on a historical observation database. For each historical date sample, a real hourly meteorological observation sequence x is obtained. true The corresponding hour-level physical prior profile x prior And the multi-source conditional vector c; The difference between the actual hourly meteorological observation sequence and the hourly physical prior profile is used to obtain the actual hourly meteorological residual x. res ; Furthermore, based on global statistics in the historical observation database, the multi-source conditional vector c and the true hourly meteorological residual x are analyzed. res After standardization, we obtain the standardized multi-source condition vector c′ and the standardized true meteorological residual x. res,norm .

[0050] By standardizing the data using global statistics from a historical observation database, different meteorological variables are mapped to a unified dimensionless feature space, thereby reducing the impact of variable dimension differences on model training stability.

[0051] Step 22: Extract temporal latent features using the Transformer encoder and generate latent variables; The standardized real hourly meteorological residual sequence x res,norm The standardized multi-source conditional vector c′ is input into the Transformer encoder to extract the corresponding temporal latent feature representation.

[0052] Furthermore, the temporal latent features are mapped to latent variable distribution parameters through the latent variable parameterization module; the latent variable distribution parameters include a mean vector and a log-variance vector. Based on the mean vector and log-variance vector, latent variables are generated through reparameterized sampling. Step 23: Generate standardized hourly meteorological residual prediction sequences based on the Transformer decoder; The latent variable z and the standardized multi-source conditional vector c′ are input into the Transformer decoder, and feature fusion is performed through a multi-head self-attention mechanism to output the standardized hourly meteorological residual prediction sequence x′. res,norm ; The Transformer decoder is used to fuse the random fluctuation characteristics in latent variables and the meteorological physical constraint information in multi-source condition vectors, thereby generating hourly meteorological residual prediction results that conform to the intraday meteorological evolution pattern.

[0053] Step 24: Construct a weighted Huber loss function, calculate the loss value based on the hourly meteorological residual prediction sequence and the real meteorological residuals in the training samples, update the parameters of the TCVAE model, and iteratively train to obtain the trained TCVAE model. To improve the model's ability to fit the random fluctuation characteristics of different meteorological variables, this embodiment uses the weighted Huber loss function as the training objective function for the TCVAE model. A corresponding variable weight vector w is set according to the fluctuation characteristics of different meteorological variables, with higher weights assigned to wind speed variables, which exhibit strong random fluctuations, and lower weights assigned to solar irradiance variables, which exhibit strong regularity.

[0054] In this embodiment, the model's ability to fit different meteorological variables can be improved through a differentiated variable weighting guidance mechanism. Since temperature, specific humidity, solar irradiance, and wind speed differ in terms of volatility, dimensions, and modeling difficulty, the gradient contribution of different meteorological variables is adjusted differentially through the variable weight vector w.

[0055] The weighted Huber loss function of the TCVAE model includes a reconstruction loss term and a KL divergence constraint term; This embodiment improves the reconstruction term by using a weighted Huber loss to enhance the model's robustness against outliers. The total loss function L of the weighted Huber loss function is expressed as: ; In the formula, Let be the mathematical expectation operator, representing the variational posterior distribution defined by the encoder. Average performance under these conditions; The encoder output distribution is constrained by the multi-source condition vector c; w i The gradient weight coefficients of the i-th meteorological variable together form the aforementioned weight vector w; the Huber(·) operator uses squared loss to ensure convergence accuracy when the error is small, and uses linear loss to prevent gradient explosion when the error is large; x i x represents the actual hourly meteorological residual observation vector; i′ represents the prediction residual vector generated by the decoder; β is the regularization coefficient used to adjust the diversity of generated samples; D KL (·||·) is the Kullback-Leibler divergence, which measures the difference between the variational posterior distribution and the prior distribution p(z), and plays a role in latent space regularization; p(z) is the prior distribution of the latent variable, which is usually assumed to be a standard normal distribution.

[0056] This embodiment constrains the error between the predicted residual and the actual residual by reconstructing the loss term, and constrains the latent variable distribution to approximate the standard normal distribution by using KL divergence, thereby improving the continuity and sampleability of the latent space and overcoming the difficulties of large differences in the dimensions and strong non-stationarity of different meteorological variables.

[0057] In step S2, the multi-source conditional vector obtained in step S1 is input into the trained Transformer augmented conditional variational autoencoder model (TCVAE model). Hourly meteorological residuals are generated through latent variable sampling. The hourly meteorological residuals are then superimposed with the hourly physical prior profile to generate the original hourly meteorological scene. This process includes the following steps: Step S201: Based on the latent variable distribution parameters obtained during the TCVAE model training process, random sampling is performed based on the reparameterization mechanism and the latent space probability distribution to generate latent variable samples z; In order to characterize the uncertain fluctuations in hourly meteorological evolution under the same daily meteorological background constraints, this embodiment performs random sampling in the latent space of the trained TCVAE model.

[0058] Specifically, during the training of the TCVAE model, the Transformer encoder has learned a continuous latent space probability distribution based on a large number of historical samples, i.e., the latent space. The latent space is used to characterize the statistical distribution characteristics of different hourly meteorological evolution trajectories in a low-dimensional space, and similar meteorological evolution states have similar location distributions in the latent space.

[0059] In the generation phase, the standard normal distribution is first used. Extract L groups of random noise vectors ε of dimension dz from the data: Subsequently, based on the latent variable distribution parameters learned during the training phase, a reparameterization mapping is performed on the random noise vector ε to generate latent variable samples z: ; Where μ represents the mean vector of the latent variable distribution, and σ represents the standard deviation vector of the latent variable distribution. This represents element-wise multiplication. The mean vector μ and standard deviation vector σ are determined by the latent space probability distribution of the trained TCVAE model and are used to describe the location and fluctuation range of the probability distribution in the latent space under the current multi-source conditional vector constraints.

[0060] This embodiment, through the aforementioned reparameterization mechanism, transforms the random sampling process, which is originally not directly involved in gradient propagation, into a differentiable mapping process involving the mean vector μ, the standard deviation vector σ, and the random noise ε. This allows the random sampling in the latent space to retain its randomness while satisfying the continuous mapping characteristics of the neural network. The generated latent variable sample z does not correspond to a specific physical meteorological state, but rather represents, in a low-dimensional statistical manner, various random fluctuations that may exist in the hourly meteorological evolution trajectory under the same diurnal meteorological conditions. Different random noise vectors ε correspond to different latent variable samples z, thus enabling the generation of multiple hourly meteorological scenarios with random differences.

[0061] Step S202: Temporal feature fusion and residual reconstruction based on Transformer decoder. The sampled latent variable z and the standardized multi-source conditional vector c′ are subjected to nonlinear interaction and feature diffusion to obtain a standardized hourly meteorological residual sequence, including: Step S2021: Copy and extend the latent variable sample z along the time axis to T time periods, so that each hour corresponds to the same latent variable perturbation information; concatenate and splice the extended latent variable sample z with the standardized multi-source conditional vector c′ in the feature dimension to construct the temporal input sequence [z, c′] of the Transformer decoder. Multi-source feature dimension alignment and splicing: Since the latent variable sample z is a low-dimensional latent space vector, and the meteorological residual sequence to be generated is a 24-hour continuous time series data, the latent variable sample z is copied and extended to T time periods along the time axis to construct the time series input sequence [z, c′] of the Transformer decoder; Optional, T=24; Step S2022: Based on the multi-head self-attention mechanism of the trained Transformer decoder, nonlinear interaction and feature fusion are performed on the random fluctuation information in the latent variable sample z and the physical constraint information in the standardized multi-source condition vector c′, so that the random perturbation information in the low-dimensional latent variable is diffused into the 24-hour time-series structure and combined with the intraday evolution skeleton information represented in the physical prior profile to obtain the time-series latent representation; realizing the cross-dimensional transformation from static latent representation to dynamic time-series features.

[0062] Step S2023: Map the fused temporal implicit representation to the normalized meteorological residual prediction sequence through the linear output layer at the end of the Transformer decoder; The linear output layer at the end of the decoder maps the fused temporal implicit representation back to the physical residual space, generating a standardized meteorological residual prediction sequence x with 24 hours × 4 variables. res,norm : ; In the formula, θ dec These are the learnable weight parameters of the decoder network. The standardized meteorological residual prediction sequence reflects the degree of standardization of each meteorological variable under the current sampling from the physical prior baseline.

[0063] Step S203: Perform inverse standardization on the standardized hourly meteorological residual sequence to obtain hourly meteorological residuals with physical dimensions; Since the meteorological residual sequence output by the TCVAE model is standardized dimensionless data, in order to restore the generated results to a dimension space with actual physical meaning, this embodiment performs inverse standardization on the standardized hourly meteorological residual sequence based on a global statistical operator extracted from a historical observation database. The global statistical operator includes the aforementioned global arithmetic mean μ. hist,i Compared with the global standard deviation σ hist,i Perform the following inverse transformation standardization process: ; In the formula, x res,i (t) represents the physical residual value of the i-th dimension meteorological variable after restoration at time t, and its magnitude is strictly consistent with the original meteorological observation data; x res,norm,i (t) represents the standardized residual value of the i-th dimension meteorological variable output by the TCVAE model at time t; This embodiment uses the above-mentioned inverse standardization process to restore the dimensionless standardized residual value to a physical residual value with the same numerical order of magnitude as the original meteorological observation data, thereby ensuring that the generated residual increment can be directly superimposed on the hourly physical prior profile.

[0064] Step S204: Linearly superimpose the hourly physical prior profile with the hourly meteorological residual to generate the original hourly meteorological scene; Specifically, obtain the hourly physical prior profile x obtained in step S1. prior And the hourly meteorological residual x with physical dimensions obtained in step S203 res Among them, the hour-level physical prior profile x prior Used to characterize the basic morphology of intraday meteorological evolution on a target day under similar climatic conditions, hourly meteorological residual x res This is used to characterize the stochastic fluctuation corrections generated by the TCVAE model based on latent variable sampling. The hourly physical prior profile x... prior Compared with hourly meteorological residuals x resLinear overlay is performed to obtain the original hourly meteorological scene x. raw : ; The above-described superposition and reconstruction process, based on the non-stationary temporal reconstruction logic of probability distribution sampling, has the core mechanism of mapping the abstract latent variable z under the physical constraints of the multi-source condition vector c into a physical residual sequence with high-frequency random fluctuation characteristics through a decoder network. The original meteorological scene x generated in this embodiment... raw It retains the basic intraday evolution pattern provided by the hourly physical prior profile, and superimposes the random fluctuation residuals generated by latent variable sampling. This makes the generated results have the same characteristics as daily-scale macroscopic statistical consistency, hourly intraday pattern rationality and random fluctuation diversity, thereby reducing the problems of intraday pattern distortion, insufficient randomness and insufficient expression of multivariate fluctuations that are easy to occur in traditional generation methods.

[0065] Step S3: Correct the original hourly meteorological scene: The wedge correction algorithm is used to correct the original hourly meteorological scene, and the corrected meteorological state at the end of the current day is passed to the generation process of the next target day to obtain the corrected hourly meteorological scene that meets the target day average constraint and the inter-day continuity requirement. This includes the following steps: Step S301: Calculate the daily average deviation to construct a wedge-shaped correction term that changes linearly with time, and correct the original hourly meteorological scene to obtain the corrected hourly meteorological scene, including the following: Step S3011: x the actual daily weather average of the original hourly weather scene mean,i Compared with the target daily meteorological mean x target,i The difference is calculated to obtain the daily average deviation. Specifically, obtain the original hourly weather scene x generated in step S2. raw ; Where, x raw It includes four types of meteorological variables: temperature, specific humidity, solar irradiance, and wind speed within 24 hours of the target day.

[0066] For the i-th dimension meteorological variable, calculate the actual daily average value x of the original hourly meteorological scene over the 24 hours of the target day. mean,i ; ; In the formula, This represents the value of the i-th dimension meteorological variable in the original hourly meteorological scenario at time t on the target day; t=0 represents time 0 on the target day; t=T-1 represents time 23 on the target day.

[0067] x the actual daily average value of the original hourly weather scenario mean,i Compared with the target daily meteorological mean x target,iBy comparison, the daily average deviation Δ is obtained. x i : ; Since the TCVAE model generates hourly meteorological residuals through latent variable sampling, its generation process can characterize the random fluctuation characteristics of meteorological variables, but does not impose hard constraints on the daily average of the output sequence. Therefore, the original hourly meteorological scene may have a slight bias in daily-scale statistics.

[0068] Step S3012: Construct a wedge-shaped correction term that changes linearly with the hour number within the target day, and distribute the daily average deviation to the original hourly meteorological scene at each time point to obtain the corrected hourly meteorological scene for the current target day. To ensure that the generated results are strictly consistent with the target daily meteorological statistics, this embodiment constructs a wedge-shaped correction term that varies linearly with time, smoothly distributing the daily mean deviation along each time point within the target day to obtain the corrected hourly meteorological scene. The correction formula is as follows: ; In the formula, x corr,i (t) represents the value of the i-th dimension meteorological variable after correction at time t on the target day; This is the wedge correction term corresponding to time t.

[0069] From the above formula, we know that when t=0, the wedge correction term is zero, that is: ; Therefore, the meteorological conditions at 0:00 on the target day do not undergo additional shifts due to the daily mean correction, maintaining the continuous connection between the initial time of the target day and the last time of the previous day. As t increases, the wedge-shaped correction term increases linearly along the intraday time direction, smoothly distributing the daily mean deviation to each time of the target day, thus avoiding disturbances to the initial boundary state caused by direct overall shift correction.

[0070] Furthermore, the wedge correction weights satisfy: ; For the corrected hourly weather scene x corr,i (t) Calculating the average over the target day, we get: ; Right now: ; Therefore, it can be seen that after wedge correction, the corrected hourly meteorological scene is consistent with the target daily-scale meteorological statistics at the daily average level, thus satisfying the daily-scale macro-meteorological constraints.

[0071] Step S302: Extract the last moment's meteorological state from the corrected hourly meteorological scene of the current target day as the initial meteorological state of the next target day. After generating the original hourly meteorological scene of the next target day through the TCVAE model trained in step S2, step 301 is executed for correction. The corrected hourly meteorological scene that meets the target day's average constraint and inter-day continuity requirement is obtained by recursion. It is a long-term hourly meteorological scene, such as an hourly meteorological scene of a month or a year. In this embodiment, inter-day continuity is not achieved by forcing the meteorological values ​​at the boundary times of two adjacent days to be completely equal, but is achieved through the input of meteorological state conditions at the end of the previous day, the non-disturbance correction at the start time of the target day, and the daily recursive generation mechanism.

[0072] Specifically, in step S2, the weather status x at the end of the previous day... anchor As part of the multi-source conditional vector, x is input into the trained TCVAE model. anchor,i This represents the meteorological conditions at 23:00 the day before the target date. When generating hourly meteorological residuals for the target date, the TCVAE model inputs the meteorological boundary states at the end of the previous day, thus constraining the generated results for the initial time of the target date by the state at the end of the previous day.

[0073] Meanwhile, during the wedge correction process in step S301, the wedge correction term corresponding to time 0 of the target day is zero. Therefore, the wedge correction will not additionally change the meteorological state at the initial time of the target day, that is, the meteorological state x at the end of the previous day. anchor Based on the conditional input, wedge correction can ensure the consistency of daily averages while avoiding disrupting the continuous connection between the initial time of the target day and the last time of the previous day.

[0074] During the generation of year-round or long-term weather scenarios, steps S1 to S3 are executed daily in chronological order. For the target day, the weather state at 23:00 in the hourly weather scenario corrected the previous day is taken as the weather state at the end of the previous day as described in step S2, i.e.: ; In the formula, This represents the value of the i-th dimension meteorological variable at time 23 in the hourly meteorological scene revised the previous day.

[0075] When generating multiple weather scene paths, each scene path recursively generates the next day's scene based on the corrected weather conditions at 23:00 on the previous day. This ensures that different scene paths are recursively generated independently in time, avoiding the mixing of boundary states between different paths. This results in hourly weather scenes that are both probabilistically diverse and have cross-day continuity.

[0076] Through the above processing, the wedge correction algorithm can correct the daily average value of the original hourly meteorological scene to the meteorological statistics at the target day scale without additional disturbance to the meteorological state at the initial moment of the target day. At the same time, the meteorological state at the end of the previous day is used as a condition input and participates in the daily recursive generation, so that the long-cycle meteorological scene maintains continuous connection between adjacent dates and reduces the non-physical jumps that are easy to occur in the daily independent generation method.

[0077] Existing methods for generating wind and solar power output scenarios often directly generate wind and solar power or use simplified empirical formulas for conversion, which fails to fully reflect the impact of physical factors such as humid air density, wind farm spatial aggregation effect, radiative decomposition, oblique projection, and battery temperature changes on wind and solar power output. To address this issue, this embodiment improves the wind power output conversion model and the photovoltaic power output conversion model; Furthermore, the improved wind power output conversion model in this embodiment includes corrections for humid air density and spatial aggregation effects of wind farms; the improved photovoltaic output conversion model includes corrections for radiation decomposition, oblique projection, and battery temperature; the corrected hourly meteorological scene is input into the wind power output conversion model and the photovoltaic output conversion model to obtain wind power output and photovoltaic output. To accurately convert the generated full-element meteorological sequence into a wind-solar combined power output scenario, this embodiment establishes a power conversion model that considers the dynamic coupling effect of multiple meteorological variables, in order to characterize the actual physical impact of four-dimensional meteorological variables—temperature, humidity, irradiance, and wind speed—on wind and solar power output.

[0078] In some embodiments, considering the characteristic that wind power output is affected by the multidimensional coupling of microscale meteorological environment, this embodiment constructs a wind power output conversion model from the accurate mapping of original meteorological parameters to electricity. The specific steps are as follows: Step 41: Based on the local air pressure estimation equation and the virtual temperature correction theory, the air density of the study area is corrected according to the obtained temperature and humidity. In wind power output modeling, air density is a crucial physical parameter connecting meteorological conditions and wind turbine output power. This embodiment, based on the local air pressure estimation equation and the virtual temperature correction theory, introduces a moist air density equation that takes into account temperature and specific humidity corrections. The formula is as follows: ; In the formula, P is the local air pressure at an altitude of h; P0 is the standard atmospheric pressure; Γ is the temperature lapse rate, with a value of 0.0065 K / m; T sl It is the absolute temperature below standard sea level; n is the atmospheric pressure index constant, whose physical nature is determined by the ratio of gravitational acceleration to the dry air constant, and is taken as 5.255 under standard atmosphere; T vThe virtual temperature is used to quantify the mass displacement effect of humid air relative to dry air; T is the temperature in Celsius; λ is the virtual temperature correction factor for humid air; q is the specific humidity, i.e., the humidity sequence described in this embodiment; R d ρ is the gas constant for dry air; ρ is the density of moist air.

[0079] Step 42: Extrapolate the wind speed at the reference height to the height of the wind turbine hub based on the power law to obtain the wind speed at the actual wind-receiving height of the wind turbine. Step 43: Based on the corrected air density, perform density correction on the standard single-unit power curve to obtain the single-unit output power under actual air density conditions; Step 44: Introduce a Gaussian convolution smoothing kernel function to perform spatial aggregation smoothing on the corrected single-unit output power to simulate the aggregation effect of multiple wind turbines in the wind farm due to spatial distribution, wake disturbance and local wind speed differences. Step 45: Based on the set comprehensive loss coefficient, the output power is reduced to obtain the wind power output conversion model that takes into account hub height, air density, spatial aggregation effect and operating loss.

[0080] In this embodiment, the wind speed is extrapolated to the hub height using the power law, and the standard power curve is restored based on the corrected density. A Gaussian convolution smoothing kernel function is introduced to simulate the spatial aggregation effect at the wind farm scale. The formula for the above process is as follows: ; In the formula, v is the wind speed at the height of the wind turbine hub; v ref The wind speed at the reference height is the wind speed sequence obtained in this embodiment; H is the height of the wind turbine hub; H ref For reference height, it is 10m in this embodiment; α is the power-law exponent related to surface roughness; P std (v) represents the standard air density ρ std The single-unit power curve below can be obtained from factory data based on the brand and model of each fan; P adj The output of a single unit is calculated after considering the correction for humid air density; P(v) is the output of the wind farm after considering the spatial aggregation effect. σ is a Gaussian smoothing function with standard deviation σ, used to eliminate the single-unit flat-top effect to simulate the spatial aggregation effect of wind farms; η is the wind speed, a variable in the integration process. wind The overall system losses that need to be taken into account for the final output P.

[0081] This embodiment accurately converts the reference wind speed obtained from the meteorological field into the effective wind speed at the turbine hub height. Based on this, it corrects the standard power curve by incorporating actual air density, ensuring that the wind power calculation results reflect the actual changes in wind energy resources under different temperature, humidity, and air pressure conditions. Simultaneously, by introducing a Gaussian convolution smoothing kernel function, it effectively weakens the abrupt changes in the single-unit power curve near the cut-in wind speed, rated wind speed, and cut-out wind speed. This simulates the convergence and smoothing effect caused by the spatial distribution, local wind speed differences, and wake disturbances of multiple units within the wind farm, resulting in a more continuous, stable, and realistic wind farm output power sequence. Furthermore, by reducing the theoretical output power using the system's comprehensive loss coefficient, the engineering applicability and accuracy of the wind power prediction results are improved, providing more reliable input data for subsequent new energy output assessment, power system dispatch, and multi-energy complementary operation optimization.

[0082] Unlike wind power, which is governed by hydrodynamic mechanisms, photovoltaic (PV) output is influenced by both astronomical optics and semiconductor thermodynamics. Its core lies in the precise characterization of effective irradiance and the effect of battery temperature rise. Further technical solutions involve constructing a PV power conversion model, including the following steps: Step 401, Radiation Decomposition and Calculation: The total horizontal radiation obtained is decomposed into direct radiation and diffuse radiation. Combined with the tilt angle of the photovoltaic module, the solar incident angle and the ground albedo, the direct irradiance, diffuse irradiance and ground reflected irradiance are projected or converted to the tilt surface of the photovoltaic module respectively to obtain the total irradiance of the tilt surface actually received by the photovoltaic module. This embodiment uses the Erbs decomposition model to decompose the total horizontal radiation into direct radiation and scattered radiation.

[0083] Step 4011: Calculate the atmospheric cleanliness index k t : ; In the formula, G h The total horizontal irradiance is represented by I0, which is the irradiance sequence described in this embodiment; I0 ​​is the solar constant; θ z This is the solar zenith angle.

[0084] Step 4012: Based on the calculated atmospheric cleanliness index, calculate the scattering ratio k using the Erbs piecewise function. d : ; Step 4013: Based on the scattering ratio, separate the total irradiance of the horizontal plane to obtain the scattered irradiance G. d,h With normal direct irradiance G beam The total irradiance G received by the module on the inclined surface is obtained by projecting the irradiance onto the inclined surface. poa That is, the actual irradiance received by the photovoltaic panel: ; In the formula, G diffuse β represents the scattered irradiance on the inclined surface. pv G is the tilt angle of the photovoltaic module. ground Ground reflected irradiance; α g Ground albedo; θ aoi The angle of incidence of the sun; Step 402: Dynamic modeling of battery temperature based on Faiman core formula: Based on the Faiman temperature model, calculate the operating temperature of photovoltaic cells according to ambient temperature, total irradiance of the slope and wind speed. This embodiment takes into account the attenuation effect of extreme high temperatures on photoelectric conversion efficiency, and introduces the ambient temperature T. amb , oblique projection irradiance G poa and wind speed v w Establish the Faiman temperature model: ; In the formula, T cell U0 represents the battery operating temperature; U0 and U1 are heat dissipation coefficients related to wind speed.

[0085] Step 402, Temperature-related efficiency degradation correction: Based on the degree of deviation of the battery operating temperature from the standard test temperature, a temperature correction coefficient is constructed. Combined with the rated power curve and system comprehensive efficiency under standard test conditions, the photovoltaic output power after considering irradiation decomposition, oblique projection, module temperature rise and temperature efficiency degradation is obtained, which is the constructed photovoltaic power conversion model. Considering the impact of component temperature deviation from standard test conditions on photoelectric conversion efficiency, the rated power is corrected to determine the final photovoltaic output power P. PV : ; In the formula, f temp γ is the temperature correction factor, used to characterize the power degradation when the component temperature deviates from standard conditions; P is the power temperature coefficient. STC The rated power curve is under standard test conditions at 25℃. Factory data can be obtained according to the brand and model of each photovoltaic panel; η PV For overall system efficiency.

[0086] The hourly weather scene x after physical consistency correction in step S3 corr,i (t) Input the constructed wind power output conversion model and photovoltaic power output conversion model, and perform the final mapping from environmental parameters to power output: Multi-meteorological variable coupled power output calculation: Using the corrected wind speed, temperature and specific humidity, the wind power processing conversion model takes into account air density correction and spatial aggregation effect to output the wind power output curve; at the same time, using the corrected slope total irradiance, temperature and wind speed, the photovoltaic power output is output through the photovoltaic conversion model and Faiman temperature rise correction.

[0087] Construction of a long-term scene set throughout the year: By repeatedly executing the above steps S1 to S4, the power output segments generated each day are logically spliced ​​together in chronological order, and finally output a continuous meteorological-solar power output scene dataset with high hourly resolution and covering the hourly level throughout the year.

[0088] The resulting scene set not only strictly follows the long-term evolution trend of climate models such as CMIP6 on a macro scale, but also fully preserves the intraday morphological characteristics and random fluctuation patterns of historical observation data on a micro scale, providing reliable data support for steady-state analysis and transient simulation of high-proportion new energy power systems.

[0089] To verify the effectiveness of the meteorological-solar power output scene generation method based on retrieval enhancement probability downscaling proposed in this embodiment, simulation verification was conducted, as described below; Hourly historical observation data of a region in Europe from January 1, 2006 to December 31, 2025 were selected as the experimental sample. The data includes four core meteorological elements: temperature, specific humidity, solar irradiance, and wind speed, with a time resolution of 1 hour. The data source is the EU ERA5 meteorological data reanalysis project. 80% of the continuous data from 2006 to 2024 was used as the model training set, 20% as the validation set, and the entire year of 2025 data was used as the test set. The macroscopic constraints of the CMIP6 climate model were simulated by extracting diurnal statistics from the test set. The hourly meteorological series for the entire year of 2025 generated using the RAP downscaling method proposed in this embodiment is shown below. Figure 3 As shown, the RAP-generated data value is a set of random curves from 500 samples generated using this method. The shaded area represents the 95% confidence interval band, which is used to characterize the generation effect of this method on the random uncertainty of meteorological elements.

[0090] To demonstrate the superiority of the proposed method in terms of physical consistency and statistical fidelity, this embodiment introduces two benchmark comparison models for simulation: one is a pure physical prior method, which relies only on similar day retrieval and morphological scaling and does not have the ability to generate latent space probabilities; the other is the traditional TCVAE method without introducing a retrieval enhancement mechanism, which relies only on deep learning operators for data-driven downscaling and does not embed intraday structural prior constraints.

[0091] In terms of constructing the evaluation system, a comprehensive quantitative assessment is conducted using multi-dimensional deterministic error indicators, statistical distribution indicators, and time-series correlation indicators, among which: Deterministic error metrics include root-mean-square error (RMSE) and coefficient of determination (R²). RMSE quantifies the standard deviation of the generated and observed values ​​in Euclidean space, reflecting the absolute deviation of the reconstructed values ​​on an hourly scale. ; In the formula: N is the total number of sampling points; x obs,t x represents the historical observation value at time t; gen,t This corresponds to the generated scene value.

[0092] The coefficient of determination (R²) characterizes the extent to which the generated sequence explains the actual observed fluctuation patterns. ; In the formula: x obs,mean R is the arithmetic mean of the observed sequence. 2 The closer R is to 1, the stronger the model's ability to capture non-stationary fluctuations in meteorological elements; if R... 2 If the result is negative, it indicates that the model's generation performance is inferior to directly using the mean.

[0093] Statistical distribution indices, used to characterize the consistency of a uniform distribution, can be tested using the Kolmogorov-Smirnov test (KS). The KS statistic quantifies the maximum deviation between the generated distribution and the cumulative probability function of the true distribution, and is used to assess the distributional similarity between the generated downscaled data and the observed values. ; In the formula: F obs (x) and F gen (x) represents the cumulative empirical distribution function of the historical scene and the generated scene, respectively; sup denotes taking the supremum. D KS The smaller the value, the closer the probability distribution of the generated scene is to the real physical distribution across the entire range.

[0094] Temporal correlation indices are quantified using the autocorrelation function (ACF), which calculates the temporal correlation characteristics of continuous sequences at different lag orders k to assess the ability of generated scenarios to preserve the inertia and periodicity of meteorological processes. ; In the formula: X t X represents the power value at time t; t+kdenoted as power value after lag k; μ is the sequence mean; and E(·) is the expectation operator.

[0095] Using the above indicators, the RAP method proposed in this embodiment was evaluated against the other two comparative methods, and the calculation results are shown in Table 1. It should be noted that since the KS index evaluates the distribution of the generated ensemble, and physical prior, as a single-valued deterministic prediction method, does not have ensemble distribution characteristics, it has no KS value; the ACF index in the table is a comprehensive measure of the mean absolute error of the lag time from 0 to 7 hours.

[0096] Table 1. Comparison of hourly scene statistical indicators obtained by downscaling throughout the year;

[0097] To visually compare the precision of intraday data, this example randomly selects a typical 168-hour series for magnification, such as... Figure 4 As shown. (Combined with Table 1) Figure 4 The comparison results show that the meteorological-solar output scene generation method based on retrieval-enhanced probability downscaling proposed in this embodiment performs best in all evaluation dimensions (the method in this embodiment is referred to as RAP): In measuring the point-to-point determinism between the generated sequence and the true value, RAP performs best under all four meteorological variables, outperforming both the pure physical prior method and the TCVAE method, and R... 2 The values ​​are closer to 1, especially for specific humidity and wind speed variables, indicating that RAP has higher accuracy in characterizing hourly random fluctuations and can accurately capture the dramatic fluctuation trends of meteorological variables. Figure 4 The shaded area shows the 95% confidence interval generated based on multiple probability samplings. This interval effectively covers the fluctuation trajectory of the observed values, demonstrating the model's ability to quantitatively represent meteorological uncertainties. In contrast, the generation curves of traditional TCVAE and physical priors show significant phase shifts in some periods and obvious jumps at cross-day times, reflecting their inadequacy in characterizing temporal changes under complex meteorological conditions. They cannot reproduce the true randomness and continuity of meteorology, and their temporal continuity and deterministic error performance is poor. Meanwhile, RAP generates the sequence with the smallest ACF value, meaning the error between the generated data and the actual observations is the smallest, indicating that the generated scenario is more consistent with the temporal characteristics of actual operational data. Compared with TCVAE, which is also a probability generation method, RAP also has an advantage in KS value. The low KS value indicates that the empirical distribution constructed by RAP is closer to the distribution of the actual observations.

[0098] It should be noted that the wind and solar power output scenario finally generated in this embodiment is directly calculated by the conversion model of wind power output and photovoltaic power output that takes into account the coupling of multiple meteorological variables, which is constructed by inputting the meteorological data generated above. Therefore, by verifying the effectiveness of this method in generating four-dimensional meteorological sequences of temperature, humidity, irradiance and wind speed through the above comparison, it can be shown that the annual hourly wind and solar power output scenario sequence generated in this embodiment has the same physical fidelity and statistical authenticity. Figure 5 This demonstrates a sequence of hourly wind and solar power output scenarios for the entire year, obtained by processing hourly meteorological data obtained using the method of this embodiment through a power output conversion model. Figure 5 a) represents a sequence of wind power output scenarios. Figure 5 b) is a sequence of photovoltaic power output scenarios.

[0099] The above indicators and comparative results show that the RAP method proposed in this embodiment uses daily meteorological information as a macro-constraint and retrieves historical hourly data to construct intraday structural priors, achieving high-fidelity downscaling generation of multiple hourly meteorological variables. It exhibits good comprehensive performance in terms of deterministic error, statistical characteristics, and time series structure, and can provide a reliable scenario data foundation for multi-timescale analysis of new power systems.

[0100] Example 2 Based on Example 1, this example provides a scene generation system based on retrieval-enhanced probability downscaling, including: The data augmentation retrieval module is configured to retrieve a set of similar days based on the constructed historical observation day database within the sliding time window of the target day using the rank sum mechanism, extract and scale the corrected hourly meteorological profile, and obtain the hourly physical prior profile of the target day. The TCVAE extrapolation module is configured to concatenate the daily meteorological statistics of the target day, the hourly physical prior profile, and the meteorological state at the end of the previous day into a multi-source conditional vector. Based on the trained TCVAE model, it generates hourly meteorological residuals through latent variable sampling and superimposes the hourly meteorological residuals with the hourly physical prior profile to generate the original hourly meteorological scene. The correction module is configured to use a wedge correction algorithm to correct the original hourly meteorological scene and pass the corrected meteorological state at the end of the current day to the next target day generation process to obtain a corrected hourly meteorological scene that meets the target day average constraint and inter-day continuity requirements. The scene generation module is configured to input the corrected hourly meteorological scene into the wind power output conversion model and the photovoltaic output conversion model to obtain the wind power output scene and the photovoltaic output scene, and then splice them in time order to obtain a continuous meteorological-wind and solar power output scene set.

[0101] Furthermore, the process of constructing hourly-level physical prior profiles in the data augmentation retrieval module includes the following steps: Historical meteorological data of the study area were acquired to construct a historical observation database D. hist And based on the target date, a sliding time window is constructed to filter candidate similar days; Based on the rank sum mechanism, multidimensional similarity retrieval is performed on candidate historical samples, and a set of similar days is selected. Hourly meteorological profiles are extracted from similar day sets and linear scaling corrections are applied to obtain hourly physical prior profiles for the target day.

[0102] Furthermore, the TCVAE deduction module is configured to perform the following steps: Based on the latent variable distribution parameters obtained during the TCVAE model training process, random sampling is performed based on the reparameterization mechanism and the latent space probability distribution to generate latent variable samples z. Based on the Transformer decoder, the temporal feature fusion and residual reconstruction are performed, and the sampled latent variable sample z and the standardized multi-source condition vector c′ are subjected to nonlinear interaction and feature diffusion to obtain the standardized hourly meteorological residual sequence. The standardized hourly meteorological residual series is inversely standardized to obtain hourly meteorological residuals with physical dimensions; The hourly physical prior profile is linearly superimposed with the hourly meteorological residual to generate the original hourly meteorological scene.

[0103] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0104] Example 3 Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the scene generation method based on retrieval enhancement probability downscaling described in Embodiment 1.

[0105] Example 4 Based on Embodiment 1, this embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they complete the steps in the scene generation method based on retrieval enhancement probability downscaling described in Embodiment 1.

[0106] Example 5 Based on Embodiment 1, this embodiment provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the scene generation method based on retrieval enhancement probability downscaling described in Embodiment 1.

[0107] The steps and methods involved in Examples 2 to 5 above correspond to those in Example 1. For specific implementation details, please refer to the relevant description section of Example 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0108] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0110] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A scene generation method based on retrieval enhancement probability downscaling, characterized in that, Includes the following steps: Based on the constructed historical observation day database, similar day sets are retrieved within the sliding time window of the target day using the rank sum mechanism, and hourly meteorological profiles are extracted and scaled to obtain hourly physical prior profiles of the target day. The daily meteorological statistics of the target day, the hourly physical prior profile, and the meteorological state at the end of the previous day are concatenated into a multi-source conditional vector. Based on the trained TCVAE model, the hourly meteorological residual is generated by latent variable sampling. The hourly meteorological residual is then superimposed with the hourly physical prior profile to generate the original hourly meteorological scene. The wedge correction algorithm is used to correct the original hourly meteorological scene, and the corrected meteorological state at the end of the current day is passed to the generation process of the next target day to obtain the corrected hourly meteorological scene that meets the target day average constraint and the inter-day continuity requirement. The corrected hourly meteorological scene is input into the wind power output conversion model and the photovoltaic output conversion model to obtain the wind power output scene and the photovoltaic output scene. They are then spliced ​​together in chronological order to obtain a continuous meteorological-wind and solar power output scene set.

2. The scene generation method based on retrieval enhancement probability downscaling as described in claim 1, characterized in that, The process of constructing an hourly physical prior profile includes the following steps: Historical meteorological data of the study area were acquired to construct a historical observation database D. hist And based on the target date, a sliding time window is constructed to filter candidate similar days; Based on the rank sum mechanism, multidimensional similarity retrieval is performed on candidate historical samples, and a set of similar days is selected. Hourly meteorological profiles are extracted from similar day sets and linear scaling corrections are applied to obtain hourly physical prior profiles for the target day.

3. The scene generation method based on retrieval enhancement probability downscaling as described in claim 1, characterized in that, The TCVAE model consists of a Transformer encoder, a latent variable parameterization module, a reparameterization sampling module, and a Transformer decoder connected in sequence. The Transformer encoder is used to extract features from the input hourly multivariate meteorological sequence and multi-source conditional vector to obtain the corresponding temporal latent features; The latent variable parameterization module is used to map the temporal latent features extracted by the Transformer encoder into latent variable distribution parameters; The reparameterized sampling module is configured to sample noise ε from a standard normal distribution and generate latent variables z based on the mean vector μ and standard deviation vector σ in the latent variable distribution parameters; The Transformer decoder is used to fuse the latent variable z with the standardized multi-source condition vector c and output the hourly meteorological residual sequence.

4. The scene generation method based on retrieval enhancement probability downscaling as described in claim 1, characterized in that, The process of inputting multi-source conditional vectors into a trained TCVAE model, generating hourly meteorological residuals through latent variable sampling, and then superimposing the hourly meteorological residuals with the hourly physical prior profile to generate the original hourly meteorological scene includes the following steps: Based on the latent variable distribution parameters obtained during the TCVAE model training process, random sampling is performed based on the reparameterization mechanism and the latent space probability distribution to generate latent variable samples z. Based on the Transformer decoder, the temporal feature fusion and residual reconstruction are performed, and the sampled latent variable sample z and the standardized multi-source condition vector c′ are subjected to nonlinear interaction and feature diffusion to obtain the standardized hourly meteorological residual sequence. The standardized hourly meteorological residual series is inversely standardized to obtain hourly meteorological residuals with physical dimensions; The hourly physical prior profile is linearly superimposed with the hourly meteorological residual to generate the original hourly meteorological scene.

5. The scene generation method based on retrieval enhancement probability downscaling as described in claim 1, characterized in that, The original hourly meteorological scene is corrected using a wedge correction algorithm, and the corrected meteorological state at the end of the current day is passed to the generation process of the next target day to obtain a corrected hourly meteorological scene that meets the target day mean constraint and inter-day continuity requirements. The process includes the following steps: The daily average deviation is calculated to construct a wedge-shaped correction term that changes linearly with time, and the original hourly meteorological scene is corrected to obtain the corrected hourly meteorological scene. Extract the last moment's meteorological state from the corrected hourly meteorological scene of the current target day, and use it as the initial meteorological state for the next target day. After generating the original hourly meteorological scene for the next target day, make corrections, and so on, to obtain the corrected hourly meteorological scene that meets the target day's average constraint and the requirement of inter-day continuity.

6. The scene generation method based on retrieval enhancement probability downscaling as described in claim 1, characterized in that, The daily average deviation is calculated to construct a wedge-shaped correction term that changes linearly with time, and this term is used to correct the original hourly meteorological scene, resulting in the corrected hourly meteorological scene, which includes the following: x the actual daily average weather value of the original hourly weather scenario mean,i Compared with the target daily meteorological mean x target,i The difference is calculated to obtain the daily average deviation. A wedge-shaped correction term that linearly varies with the hourly sequence number within the target day is constructed. The daily average deviation is distributed to the original hourly meteorological scene at each time point to obtain the corrected hourly meteorological scene for the current target day.

7. The scene generation method based on retrieval enhancement probability downscaling as described in claim 1, characterized in that, The wind power output conversion model takes into account the correction of moist air density and the spatial aggregation effect of wind farms, while the photovoltaic output conversion model takes into account the correction of radiation decomposition, oblique projection and battery temperature. The corrected hourly meteorological scene is input into the wind power output conversion model and the photovoltaic output conversion model to obtain the wind power output and photovoltaic output.

8. A scene generation system based on retrieval-enhanced probabilistic downscaling, characterized in that, include: The data augmentation retrieval module is configured to retrieve a set of similar days based on the constructed historical observation day database within the sliding time window of the target day using the rank sum mechanism, extract and scale the corrected hourly meteorological profile, and obtain the hourly physical prior profile of the target day. The TCVAE extrapolation module is configured to concatenate the daily meteorological statistics of the target day, the hourly physical prior profile, and the meteorological state at the end of the previous day into a multi-source conditional vector. Based on the trained TCVAE model, it generates hourly meteorological residuals through latent variable sampling and superimposes the hourly meteorological residuals with the hourly physical prior profile to generate the original hourly meteorological scene. The correction module is configured to use a wedge correction algorithm to correct the original hourly meteorological scene and pass the corrected meteorological state at the end of the current day to the next target day generation process to obtain a corrected hourly meteorological scene that meets the target day average constraint and inter-day continuity requirements. The scene generation module is configured to input the corrected hourly meteorological scene into the wind power output conversion model and the photovoltaic output conversion model to obtain the wind power output scene and the photovoltaic output scene, and then splice them in time order to obtain a continuous meteorological-wind and solar power output scene set.

9. The scene generation system based on retrieval enhancement probability downscaling as described in claim 8, characterized in that, The process of constructing hourly physical prior profiles in the data augmentation retrieval module includes the following steps: Historical meteorological data of the study area were acquired to construct a historical observation database D. hist And based on the target date, a sliding time window is constructed to filter candidate similar days; Based on the rank sum mechanism, multidimensional similarity retrieval is performed on candidate historical samples, and a set of similar days is selected. Hourly meteorological profiles are extracted from similar day sets and linear scaling corrections are applied to obtain hourly physical prior profiles for the target day.

10. The scene generation system based on retrieval enhancement probability downscaling as described in claim 8, characterized in that, The TCVAE simulation module is configured to perform the following steps: Based on the latent variable distribution parameters obtained during the TCVAE model training process, random sampling is performed based on the reparameterization mechanism and the latent space probability distribution to generate latent variable samples z. Based on the Transformer decoder, the temporal feature fusion and residual reconstruction are performed, and the sampled latent variable sample z and the standardized multi-source condition vector c′ are subjected to nonlinear interaction and feature diffusion to obtain the standardized hourly meteorological residual sequence. The standardized hourly meteorological residual series is inversely standardized to obtain hourly meteorological residuals with physical dimensions; The hourly physical prior profile is linearly superimposed with the hourly meteorological residual to generate the original hourly meteorological scene.

Images