Statistical correction and probability separation reconstruction method and system for wind speed ensemble forecast

Abstract

The present application relates to the technical fields of weather forecast and numerical model post-processing, and particularly relates to a statistical correction and probability separation reconstruction method and system for wind speed ensemble forecast, which comprises the following steps: obtaining site wind speed observation data and multi-model wind speed prediction data; performing consistency processing on the site wind speed observation data to construct an observation reference sequence; performing ensemble statistical feature extraction on the multi-model wind speed prediction data to obtain an ensemble mean and an ensemble dispersion; taking the observation reference sequence as a reference true value, performing mean deviation correction on the ensemble mean to obtain an ensemble corrected mean; performing probability distribution form reconstruction on the ensemble dispersion to obtain a mixed probability distribution result; generating a corrected data result according to the mixed probability distribution result, and constructing a probability forecast graphic product. The present application processes the central position and the probability form of the ensemble forecast respectively, thereby improving the stability and consistency of the sub-seasonal wind speed probability prediction.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting and numerical model post-processing technology, and in particular to a statistical correction and probability separation reconstruction method and system for wind speed ensemble forecasts. Background Technology

[0002] Wind speed forecasting on a sub-seasonal timescale (2 weeks to 2 months) has significant application value in renewable energy dispatching, disaster risk prevention and control, and meteorological services. However, this timescale lies between weather and climate, with complex dynamic mechanisms and limited predictability. Single deterministic numerical forecasts are insufficient to meet operational needs. Therefore, multi-model, multi-member ensemble forecasting methods have become one of the mainstream technical approaches.

[0003] Existing research indicates that ensemble forecasts generally suffer from systematic biases and probabilistic distortions at sub-seasonal scales. On the one hand, differences in the physical process descriptions and initial condition handling among different numerical models lead to long-term deviations of the ensemble mean from observations; on the other hand, ensemble dispersion exhibits significant instability across different models and forecast lead times, making it difficult to accurately reflect the true forecast uncertainty.

[0004] To address the aforementioned issues, researchers have proposed various ensemble post-processing methods, among which Ensemble Model Output Statistics (EMOS) and Bayesian Model Averaging (BMA) are the most representative. The EMOS method corrects the predicted distribution parameters by establishing a statistical relationship between ensemble statistics and observations; the BMA method constructs the predicted probability distribution by weighted mixing of different patterns or members.

[0005] However, there are still obvious limitations in directly using existing methods for predicting daily average wind speed in the second season: on the one hand, the EMOS method usually adjusts the ensemble mean and dispersion simultaneously within the same framework. When the number of samples is limited and the dispersion is unstable, it is easy to introduce distribution distortion problems such as excessive shrinkage or excessive expansion; on the other hand, the single distribution assumption is difficult to characterize the potential multimodal structure and tail features in multi-mode ensembles.

[0006] Therefore, existing technologies still lack an ensemble post-processing method suitable for predicting daily average wind speed in the next season, which can effectively reduce systematic mean bias, avoid the problem of probability distribution distortion caused by direct correction of unstable dispersion, and improve the probability consistency and stability of prediction results through a reasonable probability distribution construction method. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a statistical correction and probabilistic separation reconstruction method and system for wind speed ensemble forecasts. While correcting systematic biases in forecasts, it maintains the relative structural characteristics of ensemble members and improves the stability and reliability of uncertainty representation through probabilistic reconstruction.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts. The method includes the following steps: acquiring station wind speed observation data and multi-model wind speed prediction data; performing consistency processing on the station wind speed observation data to construct an observation reference sequence; extracting ensemble statistical features from the multi-model wind speed prediction data to obtain an ensemble mean and ensemble dispersion; using the observation reference sequence as a reference true value, correcting the mean deviation of the ensemble mean to obtain a corrected ensemble mean; based on the corrected ensemble mean, reconstructing the probability distribution shape of the ensemble dispersion to obtain a mixed probability distribution result; generating corrected data results based on the mixed probability distribution result, and constructing a probability forecast graphical product.

[0009] Optionally, acquiring station wind speed observation data includes: acquiring high temporal resolution wind speed observation data from multiple observation stations, and using the high temporal resolution wind speed observation data as the station wind speed observation data.

[0010] Optionally, the step of performing consistency processing on the station wind speed observation data to construct an observation reference sequence includes: performing quality control processing on the station wind speed observation data to obtain wind speed preprocessed data; converting the wind speed preprocessed data into a daily average wind speed observation sequence through time scale unification processing; and performing spatiotemporal unified identification on the daily average wind speed observation sequence based on a preset station numbering system and time benchmark to construct the observation reference sequence corresponding to the multi-mode wind speed prediction data.

[0011] Optionally, the step of extracting ensemble statistical features from the multi-model wind speed prediction data to obtain the ensemble mean and ensemble dispersion includes: reading ensemble forecast results from the multi-model wind speed prediction data; obtaining ensemble member wind speed values ​​based on the ensemble forecast results; performing ensemble averaging on the ensemble member wind speed values ​​to calculate the ensemble mean; and quantifying the dispersion of the ensemble member wind speed values ​​and the ensemble mean using ensemble variance to obtain the ensemble dispersion.

[0012] Optionally, the step of using the observed reference sequence as a reference true value to correct the mean deviation of the set mean to obtain the set corrected mean includes: constructing a linear mean correction model and estimating the correction parameters of the linear mean correction model using the least squares criterion; and obtaining the set corrected mean based on the set mean using the linear mean correction model.

[0013] Optionally, constructing the linear mean correction model includes: constructing the linear mean correction model using statistical methods output by an ensemble model.

[0014] Optionally, the step of reconstructing the probability distribution shape of the set dispersion based on the set correction mean to obtain the mixed probability distribution result includes: obtaining the component variance based on historical error statistics, combining the set correction mean and the component variance to obtain the pattern prediction distribution; obtaining the pattern weights, and constructing a mixed probability distribution model based on the pattern prediction distribution to obtain the mixed probability distribution result.

[0015] Optionally, obtaining the pattern weights includes: during the training phase, estimating the pattern weights using historical samples for each forecast lead time.

[0016] Optionally, the step of generating corrected data results based on the mixed probability distribution results and constructing a probability forecast graphic product includes: acquiring the multi-model wind speed forecast data for the target reporting date and performing spatial consistency processing to obtain a set of members; obtaining the mean difference between the set mean and the corrected set mean, and shifting the set members as a whole according to the mean difference to generate the corrected data results; constructing an overall probability distribution of the predictor variables based on the corrected data results, and integrating the overall probability distribution to obtain the probability forecast graphic product.

[0017] Secondly, this invention provides a statistical correction and probabilistic separation reconstruction system for wind speed ensemble forecasts. The system executes the statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts provided by this invention. The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions, and the processor is configured to call the program instructions. This invention, through high-performance hardware collaboration, ensures stable method execution, facilitates operational deployment and expansion, and meets the high-efficiency processing requirements of next-season wind speed forecasts.

[0018] Compared with existing ensemble forecast post-processing methods, the present invention has at least the following advantages: (1) Avoiding the distribution distortion problem caused by simultaneous correction of mean and dispersion. Existing EMOS-based ensemble forecast post-processing methods usually adjust the forecast mean and dispersion simultaneously, that is, correct the center position and dispersion of the ensemble forecast through a linear relationship. In the forecast of daily average wind speed in the second season, the ensemble dispersion itself fluctuates greatly under different models and different forecast lead times, and the estimation error of the dispersion parameter is easily amplified under the condition of limited sample size, which leads to unreasonable excessive contraction or excessive expansion of the corrected probability distribution. This invention clearly separates the mean correction and probability distribution reconstruction process, and only uses the EMOS method to correct the mean of the ensemble forecast, without directly adjusting the ensemble dispersion, thereby effectively avoiding the distribution distortion problem that may occur when the mean and dispersion are corrected simultaneously.

[0019] (2) Reducing systematic bias while maintaining the internal structure of the set. In the mean correction stage, this invention employs a "global shift of set members" approach to correct bias, applying the same mean correction amount to each member of the set. This method maintains the relative difference structure among set members, shifting only the overall distribution towards the observed distribution. This approach effectively reduces the systematic bias of the numerical model in the long-term average sense without destroying the time-varying prediction information contained within the set, thus facilitating the subsequent stable characterization of uncertainty.

[0020] (3) Reconstructing a more stable probability distribution using multi-model information. After mean correction, this invention introduces the BMA method based on the concept of mixed distribution to reconstruct the probability distribution. The set of different models after mean correction is regarded as different components of the mixed distribution. By estimating the mixed weights and component variances, the complementary advantages of multiple models at different forecast lead times can be fully utilized, making the final probability distribution smoother and more stable in terms of dispersion and tail structure, avoiding the overconfidence or over-divergence problems that are prone to occur in single-model probability forecasts.

[0021] (4) Improve the reliability and interpretability of probability forecasts. Because this invention separates the "mean correction" and "distribution pattern reconstruction" processes, the physical meaning of each correction step becomes clearer: the mean correction stage is mainly responsible for correcting the prediction center position, making the prediction results statistically closer to the observations; the probability distribution reconstruction stage focuses on characterizing the uncertainty range and probability structure. This clearly defined correction strategy helps improve the reliability of probability forecasts and also enhances the interpretability of the method in operational applications, facilitating flexible parameter adjustments or functional expansion for different application needs.

[0022] (5) Applicable to operational needs of sub-seasonal wind speed forecasting. In view of the characteristics of sub-seasonal daily average wind speed forecasting, such as long sample span, many forecast lead times, and complex model error structure, this invention effectively reduces the impact of parameter instability on forecast results through a separate correction process, making the correction effect more stable and consistent across different lead times, and has good operational applicability and promotion value. Attached Figure Description

[0023] Figure 1 This is a flowchart of a statistical correction and probability separation reconstruction method for wind speed ensemble forecasting according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the comparison results of continuous rank probability scores in an embodiment of the present invention. Figure 3 This is a schematic diagram comparing the probability integral transformation statistical distribution of the next seasonal daily average wind speed probability forecast using different correction strategies in this embodiment of the invention. Figure 4 This is a schematic diagram comparing the probability distribution of predicted wind speed under different post-processing methods in an embodiment of the present invention. Figure 5 This is a schematic diagram showing the comparison of deterministic prediction errors at a sub-seasonal scale using different correction methods according to embodiments of the present invention. Figure 6 This is a system framework diagram of statistical correction and probability separation reconstruction for wind speed ensemble forecasting according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the execution flow of the statistical correction and probability separation reconstruction system for wind speed ensemble forecasting according to an embodiment of the present invention. Detailed Implementation

[0024] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0025] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0026] Please see Figure 1 An embodiment of the present invention provides a statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts, the method comprising the following steps: S1. Obtain station wind speed observation data and multi-mode wind speed prediction data.

[0027] In this embodiment, wind speed observation data from multiple observation stations are acquired. The observation data is high temporal resolution wind speed observation data, which is used as the station wind speed observation data.

[0028] Specifically, minute-level 100-meter-high wind speed observation data from 16 ground observation stations deployed within the target area were selected as the reference true values.

[0029] In this embodiment, multi-model wind speed forecast data from three numerical models—ECMWF, NCEP, and CMA—are selected as forecast inputs. For each forecast initiation time, the corresponding ensemble member wind speed forecast results are extracted, and the ensemble forecast outputs are uniformly converted into daily average wind speeds. Subsequently, the ensemble forecast results are categorized and stored according to their forecast lead time.

[0030] Given the differences in the number of ensemble members and the length of available forecast lead time among different numerical models, this embodiment processes the available forecast lead time of each model independently in subsequent processing; when multi-model joint processing is involved, only the forecast lead time range commonly covered by each model is selected to ensure the consistency and comparability of forecast information from different models in the time dimension.

[0031] S2. Perform consistency processing on the wind speed observation data of the stations to construct an observation reference sequence.

[0032] First, the wind speed observation data from the stations is subjected to quality control processing to obtain wind speed preprocessing data, including but not limited to removing missing values, outliers, and obviously unreasonable observation records.

[0033] Secondly, the station wind speed observation data (wind speed preprocessing data) that has passed quality control is subjected to time-scale unified processing to convert the high temporal resolution wind speed observation data into a daily average wind speed observation sequence.

[0034] Finally, based on the preset station numbering system and time benchmark, the daily average wind speed observation sequence is uniformly identified in space and time, and an observation reference sequence that can be directly corresponding to the multi-model wind speed prediction data (the ensemble forecast results of the daily average wind speed of the next season) is constructed.

[0035] Specifically, the study uses minute-level 100-meter-height wind speed observation data from 16 ground-based observation stations. First, the minute-level wind speed data undergoes quality control processing to remove obviously abnormal or unreasonable observations. Then, the minute-level observation data is aggregated by natural day to calculate the daily average wind speed for each station. Based on this, the daily average wind speed sequences for all stations are standardized in terms of time reference, constructing a standardized daily average wind speed observation dataset. After the above processing, a daily average wind speed observation sequence covering the period from October 2023 to May 2025 is obtained. This observation sequence serves as the reference ground truth for the training phase and subsequent evaluation of correction effectiveness.

[0036] It should be noted that the daily average wind speed observation sequence obtained through the above steps is consistent with the sub-seasonal ensemble forecast results in terms of time scale, time starting point, and spatial station definition, thus providing a unified reference true value for subsequent ensemble statistical feature extraction, mean deviation correction, and probability distribution separation and reconstruction.

[0037] In an optional embodiment, for each forecast start date, the ensemble forecast results are time-matched with the observed wind speeds on the corresponding valid dates according to the corresponding forecast lead time. For each station, each numerical model, and each forecast lead time, a corresponding "ensemble forecast-observation" training sample is constructed. Samples with missing observations or those unable to complete time matching are removed. Through the above processing, a training sample set containing multiple stations, multiple start dates, and multiple forecast lead times is finally formed, providing a unified data foundation for the subsequent training and evaluation of correction parameters.

[0038] S3. Extract the set mean and set dispersion from the multi-mode wind speed prediction data by performing set statistical feature extraction.

[0039] For the completed training sample set, the statistical characteristics of the ensemble forecast results are extracted without introducing any corrective hypotheses.

[0040] Specifically, for each numerical model and each forecast lead time, the arithmetic mean of the wind speed forecast values ​​of the corresponding ensemble members is calculated as the ensemble mean of the model under that forecast lead time; at the same time, the dispersion index of the ensemble members is calculated to characterize the uncertainty characteristics within the ensemble.

[0041] The aforementioned set of statistical characteristics serve as the basic inputs for subsequent mean correction parameter estimation and probability distribution morphology reconstruction, providing necessary statistical support for the implementation of the correction method.

[0042] In this embodiment, without applying any corrections to the original ensemble forecast results, statistical characteristics of the ensemble forecasts at different start times and forecast lead times are extracted to objectively characterize the centrality and dispersion of the original ensemble forecasts. Only statistical analysis of the ensemble forecast results is performed; observational information is not introduced, nor are the relative structural relationships between ensemble members altered. The specific processing procedure is as follows: Step 1: Collect member information. For any numerical pattern, on a given reporting date... and forecast lead time Next, read the ensemble forecast results corresponding to this model to obtain the wind speed values ​​of the ensemble members under this lead time, which satisfy the following relationship: in, For the first The daily average wind speed forecast values ​​corresponding to each set member. For the date of publication, To predict the lead time, For the index of a set member, The number of members in the set for this numerical pattern.

[0043] Step 2: Calculate the ensemble mean. Based on the ensemble member data obtained in Step 1, calculate the ensemble mean of this pattern under the corresponding reporting start date and lead time, satisfying the following relationship: in, The set mean For the date of publication, To predict the lead time, For the index of a set member, The number of members in the set for this numerical pattern. For the first The daily average wind speed forecast value corresponding to each set member.

[0044] It should be noted that the ensemble mean is used to characterize the prediction center position of the numerical model at the forecast lead time, reflecting its overall prediction level for daily average wind speed.

[0045] Step 3: Calculate the set dispersion. Calculate the degree of dispersion of set members relative to the set mean, expressed using set variance, which satisfies the following relationship: in, For set variance, For the date of publication, To predict the lead time, For the index of a set member, The number of members in the set for this numerical pattern. For the first The daily average wind speed forecast values ​​corresponding to each set member. It is the set mean.

[0046] It should be noted that ensemble variance is used to characterize the level of uncertainty within the ensemble forecast, reflecting the degree of dispersion of the wind speed forecast results by the numerical model at the corresponding lead time.

[0047] In an optional embodiment, when the input data contains forecast results from multiple sets of numerical model sets, the statistical feature extraction process described above is performed independently on different models through the above steps to obtain the statistical features corresponding to each model, satisfying the following relationship: in, It represents different numerical models (such as ECMWF, NCEP, CMA, etc.). For numerical mode The set mean, For numerical mode The set variance For the date of publication, This is to provide an advance forecast period.

[0048] It should be noted that the different models are not merged or weighted; only their respective statistical characteristics are retained.

[0049] Furthermore, the output of step S3 can be uniformly represented and satisfies the following relationship: in, For numerical mode The set mean, For numerical mode The set variance For the date of publication, To predict the lead time, Representing different numerical patterns, The number of numerical models involved in the revision.

[0050] This step involves statistically extracting the mean and dispersion of the ensemble forecast members. Without altering the original ensemble structure, it objectively characterizes the forecast centrality and uncertainty of each numerical model at different forecast lead times, providing a foundational input for subsequent mean correction and probability distribution separation and reconstruction.

[0051] S4. Using the observed reference sequence as the reference true value, the set mean is corrected by mean deviation correction to obtain the set corrected mean.

[0052] After extracting the statistical features of the ensemble, this step corrects the systematic mean bias in the ensemble forecast to adjust the central position of the ensemble forecast distribution, making it statistically closer to the observation results and eliminating the long-term accumulated systematic mean bias of the ensemble forecast at different forecast lead times. The central position of the prediction distribution is corrected without changing the relative structure and discrete characteristics of the ensemble members.

[0053] This provides a more accurate and less biased input basis for subsequent reconstruction of the probability distribution. This step only corrects the mean of the ensemble forecast, without modifying the ensemble dispersion or distribution shape.

[0054] In this embodiment, the construction of the linear mean correction model is performed for any numerical pattern. On the date of commencement of reporting The forecast lead time is Under the given conditions, let: This represents the set mean obtained in step S3; This indicates the average wind speed observed on the corresponding valid date.

[0055] Specifically, a linear mean correction model based on Ensemble Model Output Statistics (EMOS) is adopted, which satisfies the following relationship: in, For numerical mode The set-corrected mean, For the date of publication, To predict the lead time, To correct the parameters, For numerical mode The set mean.

[0056] It should be noted that the correction parameters and modes and forecast lead time Relatedly, the correction parameters are estimated using the statistical relationship between the ensemble mean and the observations in the historical samples.

[0057] In this embodiment, the correction parameters are estimated as follows: during the training phase, for a fixed pattern... With forecast lead time Using historical sample sets The correction parameters are estimated using the least squares criterion and satisfy the following relationship: in, Indicates minimization. To correct the parameters, For the date of publication, To predict the lead time, This represents the average wind speed observed on the corresponding valid dates. For numerical mode The set mean.

[0058] Through the above optimization process, the statistically optimal linear mean correction relation can be obtained.

[0059] It should be noted that, unlike the traditional EMOS method, this embodiment explicitly limits the correction to the set mean only, without making any correction to the set dispersion or distribution width. That is, the set variance obtained in step S3 is retained after correction, satisfying the following relationship: in, The variance of the set after mean correction. For numerical mode The set variance For the date of publication, This is to provide an advance forecast period.

[0060] In an optional embodiment, to address the systematic bias problem commonly found in the original ensemble forecasts, an ensemble mean correction method based on the EMOS concept is used to correct the center position of the ensemble forecasts.

[0061] Specifically, for each numerical pattern And for each forecast lead time, let the original set members be... The set mean satisfies the following relationship: in, The set mean For the number of pattern members, For the index of the pattern member, It is a member of the original set.

[0062] Furthermore, the observed daily average wind speed is By performing regression analysis on the relationship between "set mean - observed value" in historical training samples, a linear mean correction model is established, satisfying the following relationship: in, Correcting the mean of the set, These are correction parameters estimated for different models and forecast lead times. It is the set mean.

[0063] Furthermore, the corrected mean offset A uniform translation is applied to all set members under this pattern, i.e. , where are the corrected set members.

[0064] By using the above method, only the mean of the ensemble forecast is corrected without changing the relative structure among the ensemble members or their original dispersion characteristics. This effectively reduces systematic bias while avoiding introducing additional disturbances to the uncertain structure of the ensemble, providing stable and controllable input conditions for the independent reconstruction of the subsequent probability distribution.

[0065] This step, by correcting only the mean and not the dispersion, achieves a clear decoupling between mean correction and probability distribution adjustment, and also has the following technical effects: First, maintain the relative structure between set members to avoid disrupting the physical consistency of the original set; Secondly, avoid introducing new uncertainties by performing regression correction on dispersion when the sample size is limited; Third, the task of adjusting the uncertainty pattern is left to the subsequent probability distribution separation and reconstruction steps, thereby reducing the risk of distribution distortion caused by a single model simultaneously correcting the mean and dispersion.

[0066] The output of this step is the corrected ensemble mean for each model and each forecast lead time. The correction result serves as input for subsequent steps (probability distribution separation and reconstruction) to construct a more reasonable probability forecast distribution.

[0067] S5. Based on the corrected mean of the set, the probability distribution shape of the set dispersion is reconstructed to obtain the mixed probability distribution result.

[0068] In this embodiment, the ensemble forecasts of different numerical models after mean correction are treated as multiple probability distribution components, and the final predicted probability distribution is reconstructed through a mixed distribution approach. In this process: the mean information has already been corrected in step S4; the uncertainty structure is no longer determined by a single model; and the width, tail, and shape of the probability distribution are contributed by multiple models. This approach achieves functional separation between mean correction and probability distribution reconstruction.

[0069] Specifically, the hybrid distribution model is constructed, assuming the reporting time is [number]. With forecast lead time Below, the mean-corrected number of... The prediction distributions of the patterns satisfy the following relationship: in, For the first A numerical pattern in a given and Predicted distribution under certain conditions Representing different numerical patterns, This is the predicted wind speed value. For the date of publication, To predict the lead time, Indicates a normal distribution. Correcting the mean of the set, This represents the component variance of the lead time corresponding to this pattern in historical samples.

[0070] In this embodiment, to avoid introducing unstable sample-by-sample dispersion correction during probability reconstruction, the component variance is estimated using a method based on historical error statistics, satisfying the following relationship: in, For component variance, Representing different numerical patterns, To predict the lead time, For this lead time Number of available samples For sample index, For the first One sample in the lead time The observed values, For the first The pattern in the first Corrected mean of the set under a sample.

[0071] It should be noted that the component variance changes only with the pattern and lead time, and does not depend on the results of a single reporting, thus ensuring the statistical stability of the probability distribution width.

[0072] Furthermore, based on the pattern prediction distribution, a mixed probability distribution model is constructed, satisfying the following relationship: in, In a given and The mixture probability density function of the wind speed prediction under the given conditions. This is the predicted wind speed value. For the date of publication, To predict the lead time, Representing different numerical patterns, For pattern weights, For the first A numerical pattern in a given and Predicted distribution under given conditions.

[0073] It should be noted that pattern weights With forecast lead time The changes occur, and the pattern weights satisfy the constraints: .

[0074] In this embodiment, the method for estimating mode weights is as follows: During the training phase, for each forecast lead time... The weights of each pattern are estimated using historical samples. The goal of weight estimation is to make the mixture distribution as close as possible to the observed distribution in an overall sense, and the optimization method satisfies the following relationship: in, Indicates minimization. For pattern weights, To predict the lead time, For this lead time Number of available samples For sample index, For the first One sample in the lead time The observed values, Representing different numerical patterns, The number of numerical models involved in the revision. For the first The pattern in the first Corrected mean of the set under a sample.

[0075] This method allows the model weights to adaptively reflect the relative reliability of different models at different lead times.

[0076] It should be noted that this embodiment reconstructs the probability distribution shape through a multi-mode hybrid distribution, so that the prediction results can obtain a more reasonable and stable characterization of uncertainty while maintaining the bias correction effect.

[0077] In an optional embodiment, after completing the ensemble mean correction, in order to further characterize the uncertainty structure and probability distribution of the prediction results, this embodiment introduces a probability distribution shape reconstruction method based on Bayesian model averaging (BMA) to perform probability modeling on the corrected ensemble forecast results.

[0078] Specifically, the ensemble forecast results of different numerical models, after mean correction, are considered as multiple probability distribution components, with each model corresponding to one probability distribution component. For a given forecast lead time... Let there be a total of The first pattern participates in probability reconstruction, the second... The mean of the set of patterns after mean correction is: The corresponding component variance is Then the overall prediction probability distribution can be expressed as a weighted mixture of multiple component distributions, and its probability density function satisfies the following relationship: in, For predictor variables The probability density function, Representing different numerical patterns, The number of numerical models involved in the revision. Indicates the first The weights of each model in the forecast lead time satisfy the following conditions. and , Indicates For the mean, The variance is a normally distributed component. For the first The ensemble corrected mean of the patterns For the first Component variance of each pattern.

[0079] During parameter estimation, based on the error relationship between the "mode correction mean - observed value" in historical training samples, the weights of each mode are adjusted. Estimates are performed. The weights are solved using the least squares criterion, that is, by minimizing the squared error between the pooled mean and the observed values, the relative contribution weights of each model at different forecast lead times are obtained. The objective function satisfies the following relationship: in, Indicates minimization. The mode weights to be estimated The number of training samples, For the index of the training samples, Representing different numerical patterns, The number of numerical models involved in the revision. Indicates the first The weight of each model in the forecast lead time For the first The pattern in the first The corrected mean of the set of training samples. For the first The daily average wind speed values ​​observed for each training sample.

[0080] Meanwhile, the variance of each component The uncertainty level of the model at this lead time is estimated by the mean square error between the corrected forecast and the observed values. Its calculation form satisfies the following relationship: in, For component variance, The number of training samples, For the index of the training samples, For the first The pattern in the first The corrected mean of the set of training samples. For the first The daily average wind speed values ​​observed for each training sample.

[0081] It is important to emphasize that in this embodiment, the BMA method is only used for reconstructing the probability distribution morphology, and no further correction is made to the ensemble mean. The means of each model have already been corrected in the aforementioned mean correction step. The BMA process only adjusts the dispersion, tail structure, and overall morphology of the probability distribution by estimating the weights and component variances. By separating the functions of "mean correction" and "probability distribution morphology reconstruction," the parameter coupling problem caused by simultaneously adjusting the mean and dispersion in a single method can be effectively avoided, thereby reducing the risk of distribution distortion and improving the stability and consistency of probability forecast results across different forecast lead times.

[0082] The probability distribution obtained by the above-mentioned hybrid probability modeling method based on BMA can comprehensively reflect the prediction information after multi-model correction. While maintaining the reasonableness of the mean, it can achieve adaptive characterization of the uncertainty structure, providing a reliable foundation for the subsequent generation of probability forecast products and risk decision-making applications.

[0083] S6. Generate corrected data results based on the mixed probability distribution results, and construct a probability forecast graphic product.

[0084] In this embodiment, based on the completion of training of mean correction parameters and probability distribution reconstruction parameters, the forecast results of the daily average wind speed set for the next season on a given reporting date are corrected, and deterministic and probabilistic forecast products that can be directly used for business applications are generated.

[0085] First, obtain the target reporting date. The corresponding multi-model ensemble forecast results are then used to interpolate the wind speeds of the ensemble members of each model to a unified target station or target latitude and longitude location, forming a forecast based on the forecast lead time. The sequence of members of the permutation set. This step achieves spatial consistency in forecast results from different models and resolutions.

[0086] Secondly, based on the EMOS mean correction parameters obtained during the training phase, the model is adjusted for each forecast lead time. The set mean is used for correction. Suppose a certain pattern has a lead time... The mean of the original set is Then its corrected mean is ,in To correct the parameters, all set members of the pattern are then shifted globally by the mean difference, i.e. ,in The translated set member values, The values ​​of the set members before the translation are used; thus, the systematic deviation is corrected while keeping the relative structure of the set unchanged.

[0087] Subsequently, after completing the mean correction, the correction sets of different patterns are regarded as different components of a probability mixture distribution. A probability distribution separation and reconstruction method based on the BMA idea is introduced to construct the overall probability distribution of the predictor variable, which satisfies the following relationship: in, In order to predict the lead time Predictor variables under certain conditions The mixture probability density function, For pattern index, For the total number of models, Indicates the first The model in the lead time The weights below, Indicates a normal distribution. This is the mode-corrected mean after EMOS mean correction. The corresponding uncertainty parameter is used to characterize the dispersion and tail features of the probability distribution.

[0088] Finally, based on the overall probability distribution described above, the corrected deterministic and probabilistic forecast results are calculated and output, including: Deterministic forecasts are given in the form of BMA means and satisfy the following relationship: in, To correct the data results, For pattern index, For the total number of models, Indicates the first The model in the lead time The weights below, This is the mode-corrected mean after EMOS mean correction.

[0089] Probabilistic forecasting calculates different probability levels by integrating the mixture distribution. quantiles below As a probability forecasting graphic product, it is used to characterize the possible range of future wind speed values ​​and its uncertainty structure.

[0090] In this embodiment, the above correction results are organized according to the effective dates corresponding to the forecast lead time to form a standardized forecast product containing the original set information, the corrected deterministic forecast, and the probability quantile interval. The product is then output as a storable and traceable data file format, providing a unified data foundation for subsequent evaluation and analysis, visualization, and business decision-making applications.

[0091] In an optional embodiment, after completing the observation data processing, ensemble statistical feature extraction, mean correction parameter estimation, and probability distribution morphology reconstruction parameter training, multiple intermediate and final parameter results are generated and output during the training phase for subsequent rapid correction and probability forecast generation of individual reporting days.

[0092] First, the training phase outputs ensemble mean correction parameters. For each numerical model and each forecast lead time, a set of mean correction parameters is estimated based on the statistical relationship between the ensemble mean and the corresponding observed values ​​in the training samples. These mean correction parameters are stored indexed by model and forecast lead time, and are used to correct the linear deviation of new ensemble forecast means during the application phase. The mean correction relationship can be uniformly represented, satisfying the following relationship: in, Correcting the mean of the set, For the mean correction parameters under the corresponding model and forecast lead time, This is the mean of the original set.

[0093] The correction parameters are fully saved during the training phase and used as the basis for correction when directly called during the application phase.

[0094] Secondly, the training phase outputs the BMA parameters required for reconstructing the probability distribution morphology. For the forecast lead time range commonly covered by all models, based on the mean-corrected training samples, the weight parameters and component variance parameters of different models in the mixture probability distribution are estimated respectively. The BMA parameters include: the weight coefficients corresponding to each model. and component variance The data are categorized and stored according to the forecast lead time. The weights of each model are subject to the constraints described above.

[0095] The variances of each component are used to characterize the uncertainty level of different models at the corresponding forecast lead time, providing necessary parameter support for subsequent probability distribution reconstruction.

[0096] Secondly, during the training phase, statistical information on the training samples is output synchronously. This includes the number of available training samples for each forecast lead time, the time range covered by the samples, and the distribution of stations, which characterizes the sample basis and stability level of the parameter estimation. This type of statistical information can be used for subsequent quality control, parameter reliability analysis, and auxiliary judgment in operational applications.

[0097] Finally, during the training phase, the mean correction parameters, probability distribution reconstruction parameters, and training sample statistics are stored in a structured data format to form a standardized training output file. This output, serving as the core parameter set for the entire post-processing method, can be directly accessed during the application phase without retraining, thereby enabling rapid correction and probability forecast generation of new reporting day set forecasts.

[0098] By constructing and storing the output results of the training phase described above, the complex statistical modeling process is brought forward to the correction implementation phase. As a result, in actual business applications, efficient and stable ensemble forecast post-processing can be completed by relying only on the trained parameters, laying a reliable foundation for subsequent correction of individual forecast days and generation of probability products.

[0099] In an optional embodiment, based on the mean correction parameters and probability distribution reconstruction parameters obtained during the training phase, the wind speed forecast results of any given single reporting day are corrected to generate the corresponding probability forecast product.

[0100] (a) Input data, including: 1) Ensemble wind speed forecast results from ECMWF, NCEP, and CMA numerical models on the target reporting date; 2) Mean correction parameters obtained during the training phase; 3) Reconstructed parameters of the mixed probability distribution obtained during the training phase.

[0101] (ii) Set mean correction: For the ensemble forecast results on the single forecast date, the original ensemble mean of each ensemble member is calculated for each numerical model and each forecast lead time. Then, the mean correction parameters obtained during the training phase are used to linearly correct the original ensemble mean, resulting in the corrected ensemble mean. This step only adjusts the central position of the ensemble forecast to reduce systematic bias without changing the relative structure between ensemble members.

[0102] (iii) Shifting the mean of set members: After obtaining the corrected ensemble mean, a uniform mean shift is performed on the original ensemble members, ensuring that the corrected ensemble members are distributed around the corrected mean. This process corrects the mean of the ensemble forecast results while preserving the discrete structure within the ensemble.

[0103] (iv) Probability distribution separation and reconstruction: After mean correction is completed, the parameters for reconstructing the mixed probability distribution obtained during the training phase are introduced. The correction results of different numerical models are used as different components of the mixed distribution to construct a joint probability distribution model. By comprehensively applying the weights and variances of each component, the prediction uncertainty and distribution pattern are reconstructed.

[0104] This process is performed independently based on mean correction, thus avoiding the distribution distortion problem caused by adjusting the mean and dispersion simultaneously in a single correction step.

[0105] (v) Output of correction results: Based on the hybrid probability distribution model, wind speed probability forecast results for a single reporting day are generated, and post-processed forecast products including corrected ensemble forecast results, probability distribution parameters, and multiquantile forecast results are output for subsequent operational applications and effect evaluation.

[0106] S7. Verification and comparative analysis of the effectiveness of the correction method.

[0107] In this embodiment, to verify the technical effect of the ensemble forecast post-processing method based on mean correction and probability distribution separation reconstruction proposed in this invention in the prediction of daily average wind speed in the next season, the prediction performance before and after correction is compared and evaluated based on the constructed training samples and the generated corrected forecast results.

[0108] (I) Analysis of the effect of improving probability consistency: Please see Figure 2 The figure shows a schematic diagram of the comparison results of the continuous rank probability score; it presents the comparison results of the raw ensemble forecast (RAW) and the probability forecast (BMA) corrected by the method of this invention on the continuous rank probability score (CRPS) under different forecast lead times.

[0109] from Figure 2It can be seen that within the short lead period, both methods can provide relatively reasonable probabilistic forecast results, but the overall CRPS level of the original ensemble forecast is relatively high. As the forecast lead period increases, within the operationally relevant medium-term forecast interval (10 to 40 days), the CRPS of the original ensemble forecast remains basically at 1.5 m / s-1.6 m / s, showing a significant systematic probabilistic error. In contrast, after mean correction and the introduction of BMA for probability distribution reconstruction, the CRPS decreases significantly in all lead periods, remaining stable at approximately 1.12 m / s-1.16 m / s within the 10-40 day lead period, and its change with the lead period is relatively gradual. Further statistical results show that within this medium-term forecast interval, the average CRPS value of the corrected probabilistic forecast is about 20%-30% lower than that of the original ensemble forecast. This result indicates that the method of this invention can effectively improve the overall probabilistic consistency of sub-seasonal daily mean wind speed forecasts.

[0110] It should be noted that, Figure 2 This diagram illustrates the comparison of continuous rank probability scores (CRPS) between the original ensemble forecasts and the probabilistic forecasts generated by the method of this invention under different forecast lead times. The horizontal axis represents the forecast lead time (in days), and the vertical axis represents the corresponding CRPS value (in m / s). The dashed line represents the CRPS result of the original ensemble forecast under the Gaussian assumption, and the solid line represents the probabilistic forecast result generated based on the mean correction and probability distribution separation reconstruction method. The shaded area in the diagram represents the sub-seasonal forecast lead time range (lead 10-40 days) that is the focus of this study.

[0111] (II) Probabilistic Reliability Analysis: Please see Figure 3 The figure illustrates a comparison of the statistical distribution of probability integral transformations for the next seasonal daily mean wind speed probability forecasts using different correction strategies. To further analyze the impact of different correction strategies on the reliability of the probability distribution, the figure compares the performance of the original ensemble forecast (RAW), the EMOS method with mean-only correction, the EMOS method with both mean and dispersion adjustments (EMOS mean+spread), and the method of this invention (BMA mixture) in terms of probabilistic reliability. Figure 3The figures show the PIT (Probability Integral Transform) histograms for different methods. Ideally, if the predicted probability distribution matches the observed distribution, the PIT statistics should be uniformly distributed. This invention significantly improves the reliability of the sub-seasonal daily average wind speed probability forecast by separating the mean correction and probability distribution shape reconstruction of the ensemble forecast. Comparison with the original ensemble forecast reveals a significant statistical inconsistency between its predicted probability distribution and the observed distribution, manifested in the PIT statistics deviating from a uniform distribution. This indicates a systematic bias in both the mean position and distribution shape of the original ensemble forecast. By employing the EMOS mean correction method, which corrects only the ensemble mean, the central position of the predicted probability distribution is effectively corrected, and the uniformity of the PIT distribution is significantly improved. This demonstrates that this mean correction method can reduce systematic bias and improve the consistency between the predicted and observed results without disrupting the original ensemble structure. However, when the ensemble dispersion is directly adjusted using the EMOS method while simultaneously correcting the mean, the estimation error of the dispersion parameter is easily amplified due to the strong instability of the ensemble dispersion across different models and forecast lead times at the sub-seasonal scale. This leads to a decrease in the consistency of the predicted probability distribution, manifesting as the PIT distribution exhibiting non-uniform characteristics again. In contrast, this invention, based on mean correction, uses a probability distribution morphology reconstruction module to perform hybrid modeling of correction results from different models. This achieves adaptive reconstruction of the overall morphology of the probability distribution without directly correcting a single ensemble dispersion. This approach effectively avoids the distribution distortion problem that may be introduced by simultaneously correcting the mean and dispersion, making the final generated probability distribution statistically closer to the observed distribution, thereby significantly improving the reliability and stability of probability forecasts.

[0112] Figure 3 This diagram illustrates the comparison of the probability integral transform (PIT) statistical distributions of the next seasonal daily mean wind speed probability forecasts using different correction strategies. The diagram shows the PIT histogram results for the original ensemble forecast (RAW), the EMOS method with only ensemble mean correction (EMOSmean-only), the EMOS method with simultaneous adjustment of mean and dispersion (EMOSmean+spread), and the method of this invention based on mean correction and probability distribution separation reconstruction (BMA mixture). The horizontal axis represents the PIT value, and the vertical axis represents the corresponding statistical density. The dashed line represents the ideal uniform distribution reference line. The sample used is a historical sample set within the forecast lead period of 10-40 days. Figure 3 In this case, n represents the sample size.

[0113] (III) Probability Distribution Pattern and Uncertainty Characterization Analysis: Please see Figure 4The figure shows a comparison of the predicted wind speed probability distribution patterns under different post-processing methods; regarding the rationality of the probability distribution patterns... Figure 4 The paper presents a comparison of the predicted wind speed probability density function shapes corresponding to different correction methods under the same forecast conditions. It can be seen that the probability distribution corresponding to the original ensemble forecast deviates significantly from the observations in terms of center position and distribution width, and the tail decay characteristics are unstable. When the EMOS method, which adjusts both the mean and dispersion, is used, the predicted probability distribution exhibits obvious over-concentration or over-expansion in some cases, manifested as an abnormally high distribution peak or a distribution width that does not match the observations. This reflects that dispersion correction easily introduces distribution shape distortion problems at the sub-seasonal scale. In contrast, the method of this invention, based on mean correction, reconstructs the overall distribution shape through hybrid probability distribution reconstruction. Its predicted probability density function exhibits smoother and more stable morphological characteristics in terms of center position, distribution width, and tail decay, and its positional relationship with the observed values ​​is more consistent. This indicates that by separating the mean correction and probability distribution shape reconstruction processes, the instability problems of direct dispersion correction can be avoided, while obtaining a more reasonable probability distribution expression in both physical and statistical terms.

[0114] Figure 4 Taking a single forecast sample with a forecast lead of 20 days as an example, this paper compares and demonstrates the predicted wind speed probability density functions obtained by the original ensemble forecast (RAW), the EMOS mean-only method, the EMOS mean+spread method, and the method of this invention (BMA mixture). The dashed lines in the figure represent the observed daily average wind speed at the corresponding time, indicating the positional relationship between the predicted distribution and the observed values.

[0115] (iv) Analysis of the improvement in deterministic prediction accuracy (mean absolute error MAE, root mean square error RMSE, and bias): In addition to the probability consistency index, to quantitatively assess the impact of the method of this invention on the accuracy of deterministic forecasts, this embodiment further compares and analyzes the variation characteristics of mean absolute error (MAE) and root mean square error (RMSE) of the original ensemble forecast (RAW), the mean-corrected EMOS method, and the method of this invention (BMA) in the sub-seasonal scale interval of operational interest (10 to 40 days of forecast lead time).

[0116] Please see Figure 5The figure illustrates a comparison of deterministic prediction errors using different correction methods at the sub-seasonal scale. The original ensemble forecast exhibits high MAE and RMSE across the entire forecast lead time, with some fluctuations depending on the forecast lead time. After mean correction (EMOS), the overall deterministic prediction error is significantly reduced, indicating that systematic bias is effectively mitigated. Furthermore, by reconstructing the probability distribution using the method of this invention, the resulting BMA deterministic prediction results maintain or slightly outperform the EMOS method's error levels for most forecast lead times, exhibiting more stable error variation characteristics. Statistical results show that within the forecast lead time window of 10-40 days, the mean MAE corresponding to the method of this invention is approximately 1.576 m / s, and the mean RMSE is approximately 1.966 m / s, both significantly lower than the corresponding error levels of the original ensemble forecast. The above results demonstrate that by separating the mean correction and probability distribution reconstruction processes, the method of this invention improves the consistency of probabilistic forecasts without weakening the accuracy of deterministic forecasts. On the contrary, it achieves a stable reduction in deterministic forecast errors at the sub-seasonal scale, thereby improving the overall usability of the correction results in practical business applications.

[0117] Figure 5 The mean absolute error (MAE) of the original ensemble forecast (RAW), the mean-corrected EMOS method, and the method of this invention (BMA) at different forecast lead times are presented; the corresponding root mean square error (RMSE) is shown in the figure below. The shaded area represents the forecast lead time window of operational interest (days 10 to 40). The results show that within this time window, the method of this invention maintains the stability of error variation with lead time while the overall error level is significantly lower than that of the original ensemble forecast, verifying the improvement effect of the method of this invention on the deterministic prediction accuracy of daily mean wind speed in the second season.

[0118] (V) Overall Effects Explanation: comprehensive Figures 2 to 5 The analysis results show that the post-processing method for sub-seasonal daily average wind speed ensemble forecast based on mean correction and probability distribution separation reconstruction proposed in this invention can effectively reduce the systematic bias of ensemble forecasts without introducing additional distribution distortion, and significantly improve the consistency, reliability and distribution rationality of probability forecasts, providing a robust and interpretable technical solution for sub-seasonal scale wind speed probability forecasts.

[0119] Please see Figure 6In an optional embodiment, the present invention provides a statistical correction and probabilistic separation reconstruction system for wind speed ensemble forecasts. The system includes an input device, an output device, a processor, and a memory, all interconnected. The memory stores a computer program comprising program instructions. The processor is configured to invoke the program instructions to execute specific steps as described in the embodiments of the statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts provided by the present invention. The statistical correction and probabilistic separation reconstruction system for wind speed ensemble forecasts provided by the present invention is structurally complete, objective, and stable, enhancing the overall applicability and practical application capabilities of the present invention.

[0120] Please see Figure 7 The diagram illustrates the execution flow of a statistical correction and probabilistic separation reconstruction system for wind speed ensemble forecasts. In an optional embodiment, the system takes multi-model ensemble forecast data and historical observation data as input, and through the collaborative processing of multiple functional modules, outputs corrected deterministic prediction results and probabilistic forecast products. The various functional modules of the system are connected via data interfaces, and the overall processing flow has a clear data flow direction and module division of labor. The system includes the following five functional modules: (1) Observation Consistency Construction Module: Acquire ensemble wind speed forecast data and corresponding observation data. First, quality control and time scale unification are performed on multi-site minute-level wind speed observation data to construct a daily average wind speed observation sequence as a reference true value for subsequent evaluation and correction. By unifying the time base and spatial station definition, the temporal and spatial alignment between observation data and model forecasts is ensured.

[0121] (2) Ensemble Statistical Feature Extraction Module: Extracts the statistical features of the ensemble forecast, including the mean and dispersion. For the ensemble forecast outputs of different numerical models, the ensemble member information for each forecast start time and forecast lead time is extracted, and statistical features such as the ensemble mean and ensemble variance are calculated. This process objectively characterizes the central position and dispersion of the original ensemble forecast at different forecast lead times without introducing any correction assumptions, providing a basis for subsequent bias analysis and correction.

[0122] (3) Mean Deviation Correction Module: A mean correction model is used to correct the central trend of the ensemble forecast without changing the relative distribution of the ensemble members. Based on the concept of Ensemble Model Output Statistics (EMOS), the mean correction method corrects the systematic deviations of the ensemble forecasts from each model. Specifically, only the mean term of the ensemble forecast is linearly adjusted, without modifying the ensemble dispersion structure. In this way, while maintaining the relative structure within the original ensemble, the long-term accumulated systematic deviations can be effectively reduced, making the central position of the predicted distribution closer to the observed distribution.

[0123] (4) Probability Distribution Morphology Reconstruction Module: Based on the corrected ensemble forecast, a mixture probability distribution is generated through probability reconstruction methods. Building upon mean correction, a Bayesian Model Averaging (BMA) method based on the mixture distribution is introduced to reconstruct the overall morphology of the probability distribution. The ensemble information of different models after mean correction is considered as different components of the mixture distribution. By estimating the weights and component variances, an adaptive characterization of the probability distribution's dispersion and tail structure is achieved. This step-by-step strategy effectively avoids the distribution distortion problem that may be introduced by simultaneously adjusting the mean and dispersion in a single method, thereby constructing more consistent and stable probability forecast results within a continuous probability space.

[0124] (5) Corrected Data and Graphical Product Output Module: Based on the mixed probability distribution results obtained from the probability distribution morphology reconstruction module, this module provides a unified expression and productized output of the corrected forecast information. By extracting statistics from the reconstructed probability distribution, deterministic forecast results and probability forecast intervals at different quantile levels are calculated, thus providing both central trend information and uncertainty descriptions within the same forecast framework. The generated forecast results fully reflect the central location, dispersion range, and tail characteristics of the corrected distribution, avoiding information loss caused by relying solely on a single deterministic result. This module organizes the forecast results by reporting time, forecast lead time, and spatial location, and outputs them in a standard data format or graphical form, providing a unified and standardized probability forecast product for subsequent forecast capability assessment, risk analysis, and business decision-making.

[0125] In summary, the present invention provides a statistical correction and probabilistic separation reconstruction method and system for wind speed ensemble forecasts. This invention does not perform a one-time overall correction of the ensemble forecast results, but rather processes different statistical characteristics of the ensemble forecasts sequentially according to the order of "first correcting the mean, then reconstructing the distribution": First, by processing observational data and unifying the time scale, a daily average wind speed observation reference sequence matching the next season's forecast results is constructed; then, ensemble member information is extracted from the ensemble forecasts output by numerical models, and statistical characteristics such as the ensemble mean and ensemble dispersion are calculated to characterize the central location and uncertainty level of the original forecast; based on this, only the mean term of the ensemble forecast is systematically corrected to weaken the systematic error that gradually accumulates with the forecast lead time, without changing the relative structure between ensemble members; finally, based on the mean correction results, a probabilistic mixture method is introduced to reconstruct the probability distribution shape of the ensemble forecast, thereby generating a more consistent and stable probabilistic wind speed prediction result in a continuous probability space. By distinguishing and sequentially executing the mean correction and uncertainty modeling processes of ensemble forecasts, this invention effectively avoids the probability distribution distortion that may be introduced by simultaneously adjusting the mean and dispersion in a single model. It provides a clear, robust, and easily scalable technical implementation path for deterministic and probabilistic prediction of daily average wind speed in the next season. The method of this invention is easy to understand, computationally simple, requires minimal workload, and is convenient for engineering applications, providing a theoretical foundation and technical support for further development in the fields of meteorological forecasting and numerical model post-processing.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts, characterized in that, Includes the following steps: Acquire station wind speed observation data and multi-model wind speed prediction data; The wind speed observation data from the aforementioned stations are subjected to consistency processing to construct an observation reference sequence; The set mean and set dispersion are obtained by performing ensemble statistical feature extraction on the multi-mode wind speed prediction data; Using the observed reference sequence as the reference true value, the set mean is corrected by mean deviation correction to obtain the set corrected mean. Based on the corrected mean of the set, the probability distribution shape of the set dispersion is reconstructed to obtain the mixed probability distribution result; Based on the mixed probability distribution results, corrected data results are generated, and probability forecast graphic products are constructed. The step of extracting ensemble statistical features from the multi-mode wind speed prediction data to obtain the ensemble mean and ensemble dispersion includes: The ensemble forecast results are read from the multi-mode wind speed prediction data, and the wind speed values ​​of the ensemble members are obtained based on the ensemble forecast results. The set mean is calculated by averaging the wind speed values ​​of the set members. The dispersion of the set is obtained by quantifying the dispersion of the wind speed values ​​of the set members and the set mean using set variance; The set discreteness satisfies the following relationship: in, For set variance, For the date of publication, To predict the lead time, For the index of a set member, The number of members in the set for this numerical pattern. For the first The daily average wind speed forecast values ​​corresponding to each set member. The set mean; The step of using the observed reference sequence as a reference true value and performing mean deviation correction on the ensemble mean to obtain the ensemble corrected mean includes: A linear mean correction model is constructed, and the correction parameters of the linear mean correction model are estimated using the least squares criterion. Based on the set mean, the set corrected mean is obtained through the linear mean correction model; The construction of the linear mean correction model includes: The linear mean correction model is constructed using statistical methods output from an ensemble model. The linear mean correction model satisfies the following relationship: in, Correcting the mean of the set, These are correction parameters estimated for different models and forecast lead times. The set mean; The step of reconstructing the probability distribution shape of the set dispersion based on the corrected mean of the set to obtain a mixed probability distribution result includes: The component variance is obtained based on the observed values ​​and the ensemble corrected mean. The pattern prediction distribution is obtained by combining the ensemble corrected mean and the component variance. The variances of the components satisfy the following relationship: in, For component variance, Representing different numerical patterns, To predict the lead time, For this lead time Number of available samples For sample index, For the first One sample in the lead time The observed values, For the first The pattern in the first The ensemble corrected mean for each sample; The pattern prediction distribution satisfies the following relationship: in, For the first A numerical pattern in a given and Pattern prediction distribution under certain conditions Representing different numerical patterns, This is the predicted wind speed value. For the date of publication, To predict the lead time, Indicates a normal distribution. Correcting the mean of the set, This represents the component variance of the corresponding lead time for this pattern in historical samples. Obtain the pattern weights, and construct a mixed probability distribution model based on the pattern prediction distribution to obtain the mixed probability distribution result; The mixed probability distribution model satisfies the following relationship: in, In a given and The mixture probability density function of the wind speed prediction under the given conditions. This is the predicted wind speed value. For the date of publication, To predict the lead time, Representing different numerical patterns, For pattern weights, For the first A numerical pattern in a given and Pattern prediction distribution under given conditions.

2. The statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts according to claim 1, characterized in that, The acquisition of station wind speed observation data includes: High temporal resolution wind speed observation data from multiple observation stations are acquired, and the high temporal resolution wind speed observation data is used as the wind speed observation data of the stations.

3. The statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts according to claim 1, characterized in that, The process of performing consistency processing on the wind speed observation data at the stations to construct an observation reference sequence includes: The wind speed observation data from the aforementioned stations are subjected to quality control processing to obtain preprocessed wind speed data; By unifying the time scale, the wind speed preprocessing data is transformed into a daily average wind speed observation sequence; Based on a preset station numbering system and time benchmark, the daily average wind speed observation sequence is spatiotemporally and uniformly identified to construct the observation reference sequence corresponding to the multi-mode wind speed prediction data.

4. The statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts according to claim 1, characterized in that, The acquisition mode weight includes: During the training phase, for each forecast lead time, the model weights are estimated using historical samples.

5. The statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecasts according to claim 1, characterized in that, The step of generating corrected data results based on the mixed probability distribution results and constructing probability forecast graphic products includes: Obtain the multi-mode wind speed prediction data for the target reporting date and perform spatial consistency processing to obtain the set member sequence; Obtain the mean difference between the set mean and the set corrected mean, and shift the set members as a whole according to the mean difference to generate the corrected data result; Based on the corrected data results, an overall probability distribution of the predictor variables is constructed, and the overall probability distribution is integrated to obtain the probability forecast graphic product.

6. A statistical correction and probabilistic separation reconstruction system for wind speed ensemble forecasts, characterized in that, The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to call the program instructions to execute the statistical correction and probabilistic separation reconstruction method for wind speed ensemble forecast as described in any one of claims 1-5.

Images