A method and system for post-credible assessment of wind power prediction

Abstract

The present application belongs to the technical field of new energy power generation power prediction, and relates to a method and system for post-trust evaluation of wind power prediction, and the technical points are S1, collecting multi-source data; S2, determining the corresponding screening strategy according to the target prediction type and the post-processing model type, combining the preset evaluation protocol; S3, generating a native effectiveness mask and constructing a plurality of sets; S4, correcting the power prediction data; S5, generating a prediction evaluation report according to the evaluation process and the correction result; the present application collects multi-source data of a wind farm, tracks and screens the collected data according to the target prediction type and the post-processing model type, extracts an effective data segment, and then establishes a plurality of support sets based on the data obtained after screening, to provide data support for subsequent evaluation and correction, clearly defines the information boundary, service coverage and common score denominator of different tasks, so that the source, rollback, exclusion and comparison qualification of the post-processing result of the wind power prediction are traceable.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction technology, and specifically discloses a method and system for reliable evaluation after wind power prediction. Background Technology

[0002] Wind power forecasting uses technological means to predict the power generation capacity of wind farms in advance over a period of time. Its core value lies in compensating for the inherent limitation of wind power being dependent on weather conditions. Due to the volatility and uncertainty of wind power, large-scale grid connection may impact the safe and stable operation of the power grid. Accurate forecasting can assist the power grid in formulating dispatch plans in advance, maintaining the balance between power supply and demand, and promoting the better integration of wind power into the overall energy system. Wind farm operation systems typically receive day-ahead wind power forecasts, weather forecasts, and ultra-short-term forecasts simultaneously, and continuously collect SCADA measured power and meteorological observation data. To reduce the systematic bias of deployed forecasting products, post-forecast processing methods such as model output statistics, linear regression, tree models, or time series models can be used.

[0003] Various data streams arrive at different times. For the same target time, a forecast published a day before can only rely on information obtained before the publication time, while intraday updates can use actual power data from the period immediately preceding the target day; post-event diagnostics can also combine measured wind speed, power rationing status, or event tags at the target time. If data that can only be obtained after the publication time is introduced into the training or testing of the day-ahead model, it will lead to future information leakage, causing the model error to be unrealistically suppressed.

[0004] Wind farm data commonly suffers from issues such as untimely release of forecast products, communication interruptions, invalid sentinel values, abnormal engineering ranges, and missing local data. Some post-processing models can only generate predictions when auxiliary data is valid, while others can output prediction results at all target time points. If errors are calculated separately on different subsets of records, even if the model itself remains unchanged, the evaluation results will shift due to changes in the scoring denominator.

[0005] Current forecasting assessments primarily focus on error metrics, reference forecasting services, and time-series verification; while data quality tools in production environments emphasize field types, value ranges, and anomaly detection. Neither approach can simultaneously address the following key questions: Is a particular feature actually usable when the forecast is published? What fallback mechanism is used for missing data? Which records are actually covered by each model? Are multiple models scored based on the same samples? And are the reasons for removing records traceable?

[0006] Therefore, the present invention aims to provide a method and system for reliable evaluation after wind power prediction, in order to solve the above-mentioned problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for reliable evaluation of wind power prediction. By collecting relevant data such as wind farm power prediction data, actual power data, and meteorological data, the relevant data are tracked and filtered according to the target prediction type, and the predicted power is corrected according to the filtered content. This solves the problem of information forward penetration or inappropriate use caused by inconsistent feature availability time during the evaluation process. The specific plan is as follows: A method for reliable assessment after wind power prediction, the method comprising the following steps: S1. Collect multi-source data and historical multi-source data from wind farms; S2. Based on the target wind power prediction type and the type of wind power prediction post-processing model, and in conjunction with the preset evaluation protocol, determine the task time area, reference prediction service, rollback strategy and evaluation index corresponding to the target wind power prediction type, and screen out the necessary data stream, labeled data stream and candidate features from multi-source data. S3. Generate the native validity mask, admission feature set, prediction support set, label support set, and scoring support set through the required data stream, candidate feature, and label data stream. S4. Input the admission feature set into the wind power prediction post-processing model, correct the power prediction data, and perform a pairwise comparison with the reference prediction service on the common scoring support set. S5. Generate a predictive assessment report based on the above assessment process and correction results.

[0008] Furthermore, for any candidate feature For those not in the task time zone The process involves eliminating candidate features, constructing an admission feature set based on the eliminated candidate features, and constructing an inadmissible feature set and elimination log based on the eliminated features.

[0009] Furthermore, the required data stream calculates a native validity mask based on preset validity rules. and native support set :

[0010] in For necessary data streams Record index at target time The value on, For indicator functions,

[0011] in, For elements in the required data stream; The native support set Combine the necessary data stream with a fallback strategy to construct a prediction support set. At this point, under the preset validity rules, the required data stream is inherently invalid:

[0012] Building a tag support set Then, based on the prediction support set and tag support set Construct a rating support set :

[0013]

[0014] in, For elements in the label data stream.

[0015] Furthermore, if there is no corresponding backoff strategy for the target wind power prediction type, then the native support set That is, the prediction support set , If there is a corresponding backoff strategy for the target wind power prediction type, and a prediction support set is constructed based on the backoff strategy... If the required data stream is natively invalid under the preset validity rules, then the input-level backoff in the backoff strategy is triggered: data that meets the target wind power prediction type, preset validity rules, and backoff strategy is selected from historical multi-source data and used as backoff input to replace the required data stream, thus forming the prediction support set. ; If the input-level fallback meets the triggering conditions but cannot be triggered, then the service-level fallback in the fallback strategy is triggered: a prediction support set is constructed based on the reference prediction service or by calling the fallback model service. And record the construction process.

[0016] Furthermore, the data coverage of this required data stream is calculated as follows:

[0017] Native support set The data coverage rate is:

[0018] Prediction support set The data coverage rate is:

[0019] in The time index corresponding to the target wind power prediction type; When the prediction support set It can cover all At that time, the predicted support set will be used. The corresponding prediction service is marked as a fully supported service; otherwise, it is marked as a conditional service, and the uncovered time index and the reason are recorded. When there is no corresponding backoff strategy for the target wind power prediction type, and When the coverage service threshold is lower than the preset threshold, the corresponding prediction support set will be... The corresponding prediction service is marked as a restricted service.

[0020] Furthermore, the admissible feature set is input into the post-processing model for wind power prediction to calculate the residuals of the power prediction data. :

[0021] in, The actual power of the wind farm in the multi-source data. The predicted power of wind farms in multi-source data; The predicted power is corrected based on the residuals to obtain the corrected predicted power. :

[0022]

[0023] in, For training the post-processing model for wind power prediction, the target time record index is used. The admission-eligible feature vector, where and One-to-one pairing.

[0024] According to the tag support set Prediction support set The reference prediction service calculates a joint score support set. :

[0025] in, A prediction support set built for the reference prediction service. ... The prediction support set corresponding to the first wind power prediction post-processing model is then updated to the first... Prediction support set of a wind power prediction post-processing model; Based on the common score support set Calculate error index :

[0026] in, for In the wind power prediction post-processing model, the first One evaluated post-processing model for wind power prediction; The corrected predicted power for the post-processing model of wind power prediction being evaluated; The post-processing model for wind power prediction and the reference prediction service are calculated on the same common score support set. The absolute error and average loss difference are as follows:

[0027]

[0028] in, The corrected forecast power is used as a reference for the forecast service. Index for records of the same target time The absolute error of pairing on The average loss difference between the post-processing model for wind power prediction and the reference prediction is used to calculate the average loss difference. The post-processing model for wind power prediction and the corrected predicted power of the reference prediction service are evaluated.

[0029] Furthermore, when the set of features that cannot be admitted exceeds the preset rejection threshold, the data coverage of the common scoring support set is lower than the preset minimum coverage threshold, or the corrected predicted power of the wind power prediction post-processing model exceeds the preset error threshold, the corrected predicted power of the reference prediction service or the rollback model service is used as the output result of the full support service.

[0030] Furthermore, the aforementioned common scoring support set Based on seasonal risk stratification, predicted wind speed risk stratification, and power ramp risk stratification, the risk is divided into several risk subsets. Data coverage, common scoring support subsets, error indicators, and average loss differences are calculated for each risk subset. Risk stratification condition index set for:

[0031] in, Stratification can be performed for seasonal risk, predicted wind speed risk, or power ramp risk. Then the common score support subset for:

[0032] Support subsets for joint scoring of seasonal risk, joint scoring of predicted wind speed risk, or joint scoring of power ramp risk; Corresponding data coverage :

[0033] in, The number of elements in the set; Corresponding error index for:

[0034] The corresponding average loss difference is :

[0035]

[0036] The risk subset is calculated based on the average loss difference. When the upper bound of the interval is lower than the minimum value of the preset threshold of the same dimension, the risk subset is marked as a low-risk subset. When the lower bound of the interval is higher than the maximum value of the preset threshold of the same dimension, the risk subset is marked as a high-risk subset.

[0037] Furthermore, based on the revised predicted power A prediction evaluation report is generated, which also includes the task time area corresponding to the target wind power prediction type, reference prediction service, rollback strategy and evaluation index, and selects necessary data streams, candidate features, admission feature sets, prediction support sets, label support sets and scoring support sets from multi-source data, and removes logs, service tag records and rollback records.

[0038] The present invention also relates to a system for reliable assessment after wind power prediction, the system being capable of performing the method for reliable assessment after wind power prediction described above; The system includes a data acquisition module, a target input recognition module, a data processing module, a post-test module, and an evaluation output module. The acquisition module collects multi-source data and historical multi-source data from the wind farm. The target input recognition module is equipped with a preset evaluation protocol. The preset evaluation protocol determines the task time area, reference prediction service, rollback strategy and evaluation index corresponding to the target wind power prediction type based on the target wind power prediction type, and filters out the necessary data stream, label data stream and candidate features from multi-source data. The data processing module combines the data determined by the preset evaluation protocol to generate a native validity mask, an accessible feature set, a prediction support set, a label support set, a scoring support set, and a common scoring support set, and then corrects the power prediction data through a wind power prediction post-processing model. The posterior module evaluates the correction results of the wind power prediction post-processing model by calculating the loss difference of the power prediction data, and performs a pairwise comparison with the reference prediction service on the common scoring support set. The evaluation output module records the data processing, calculation, and evaluation processes of several modules and generates a predictive evaluation report.

[0039] Compared with existing technologies, the beneficial effects of this solution are: 1. This invention records the original support set, the predicted support set after rollback, the label support set, and the scoring support set respectively, and records the reasons for removal, replacement, and rollback, making the denominator of the scoring samples and the processing process traceable, reducing the comparison bias caused by different models using different scoring samples, improving data relevance, expanding data coverage, and enhancing the reliability of performance evaluation conclusions. 2. Based on the target wind power prediction type, this invention defines the task time area, feature knowledge time, reference prediction service, rollback strategy and evaluation index under a unified evaluation target framework at the beginning of the evaluation system establishment. This improves the consistency and comparability of the prediction results of different models and the comparison evaluation of the reference prediction service, and reduces the risk of future information mixing and direct ranking across tasks. Attached Figure Description

[0040] Figure 1 This is a flowchart of the method steps of the present invention.

[0041] Figure 2 This is a schematic diagram illustrating the principle of the present invention.

[0042] Figure 3 This is a schematic diagram of the information boundary and data access timing of the present invention.

[0043] Figure 4 This is a schematic diagram illustrating the support set construction and co-scoring relationship of the present invention.

[0044] Figure 5 This is a flowchart of the audit records and optional data access process of this invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0046] like Figures 1-5As shown, this invention acquires multi-source data streams through an integrated data interface, including measured wind farm power, wind power prediction, weather forecasts, and other auxiliary data streams; it configures key time attributes such as target time and earliest known time, as well as a status attribute of native validity identifier for each piece of raw data; it receives an evaluation protocol issued by the evaluation initiator, which includes task type, task time region, required data stream, rollback strategy, reference prediction service, and evaluation indicators; it filters data features that meet the admission criteria based on the earliest known time and task time region, and constructs native support sets, prediction support sets, label support sets, and scoring support sets respectively; it calculates the pairing error on the scoring support sets that are aligned with the reference prediction service for multiple post-processing models to be evaluated, and prohibits direct sorting of calculation results from different information boundaries and different scoring support sets; at the same time, it automatically generates audit records containing task time region, input version, rollback behavior, exclusion reason, coverage, and statistical results. This invention can reduce false performance improvements caused by misuse of future information and inconsistent changes in the denominator of evaluation samples in the performance evaluation of wind power prediction post-processing models, and improve the traceability and consistency of results in the evaluation, verification and service launch decisions of wind power prediction post-processing models.

[0047] Example 1: The details are as follows: A method for reliable assessment after wind power prediction, comprising the following steps: S1. Collect multi-source data and historical multi-source data of wind farms through wind farm data acquisition interface and prediction service interface, including wind farm power prediction data, actual power data, meteorological data and other related data and historical data, and record the associated target time and earliest known time of each multi-source data. S2. Pre-defined evaluation protocols include task type (e.g., day-ahead forecasting, intraday rolling forecasting, post-event diagnostic tasks) and task time range (defining the target time record index). and time index The required data streams include raw day-ahead wind power, numerical weather forecasts, and historical wind power, candidate features (such as raw forecasts, forecast wind speed, previous day's same-time deviation, and the average deviation over the past 24 hours), backoff strategies, reference forecast services (a benchmark for comparison), and evaluation metrics.

[0048] Based on the target wind power prediction type ( For tasks involving day-ahead forecasting, intraday rolling forecasting, or post-event diagnostics, and considering the types of post-processing models for wind power forecasting (post-processing models for statistical models of model output MOS, Ridge regression models, tree models, and / or time-series neural network models), a new evaluation protocol is generated by combining a pre-defined evaluation protocol. The evaluation protocol includes determining the task time zone corresponding to the target wind power forecasting type, the reference forecasting service, the rollback strategy, and the evaluation metrics, and selecting necessary data streams, candidate features, and labeled data streams from multi-source data.

[0049] Candidate post-processing models such as MOS, Ridge, tree models, or temporal neural networks can correspond to different sets of required input identifiers due to their different required inputs (model inputs). For the first... The native support set of each selected post-processing model is determined solely by the native validity of each required input on the target record; its predicted support set is constructed based on the native support set, combined with the fallback strategy pre-declared by the service. The reference prediction service is pre-determined by the evaluation protocol corresponding to the target wind power prediction type. Candidate models participating within the same target wind power prediction type use the same reference prediction service; different tasks can employ different reference prediction services. The reference prediction service can be considered the 0th independent service, containing a referable admission feature set, a prediction support set, a label support set, a scoring support set, and relevant reference values ​​such as corrected predicted power.

[0050] The evaluation protocol is a versioned, computer-readable rule statement that includes at least the following data: task type, task time zone, required data stream, candidate features, validity rules, fallback strategy, reference prediction service, evaluation metrics, coverage threshold, and time validation scheme. The executor reads this protocol before performing the evaluation and writes the record-by-record results to the audit log.

[0051] The day-ahead forecast task can only use data that was available before the day-ahead forecast release time; the intraday rolling update task can only use data that was available before the update time and after deducting the collection delay; and the post-event diagnosis task can use data that is no later than the preset diagnosis deadline.

[0052] S3. Based on the evaluation protocol, perform native validity and task time region checks on the necessary data stream and candidate features, and generate native validity mask, admission feature set, prediction support set, label support set, score support set and common score support set.

[0053] For any candidate feature According to characteristics The relevant target time and the earliest known time are used to determine the deadline for tasks not falling within the task time range. The process involves eliminating candidate features, constructing an admission feature set based on the eliminated candidate features, and building an elimination log based on the eliminated features to ensure that the reasons for elimination are traceable.

[0054] If the task belongs to {predictive safety, rolling update} and Then Features that are not available in time are therefore eliminated; if the task is diagnosis and... Then Features that are not available at the declared diagnostic deadline (i.e., the task time range determined in the pre-defined assessment agreement based on the target wind power prediction type) are therefore excluded. The earliest known time, The deadline is To evaluate the diagnostic deadlines pre-declared by the protocol for diagnostic tasks, specific information settings are shown in Table 1.

[0055] Table 1 Information Settings

[0056] D1 represents the day-ahead power forecast product released for the target date. D-1 and D-2 in the following text represent historical data for the day before the target date and the two days before the target date, respectively.

[0057] Specifically: For day-ahead missions, examples of admissions include D1 forecasts, predicted wind speeds, calendar period terms, and strict D-2 historical power and residuals. If the exact release time of D1 is unverifiable, strict D-2 history can be used as an unconditionally available option; alternatively, assumed release times of 06:00, 11:00, or 18:00 for D-1 can be declared separately, and the results can be labeled as conditional sensitivity analyses.

[0058] For intraday rolling tasks, you can use , Historical residuals and recent forecasts are considered, but the earliest known time needs to account for SCADA data collection, communication, and data entry delays. If the log lacks an actual publication timestamp, the audit log should mark the analysis as a delayed assumption or best-case scenario, rather than an unconditional date-ahead result.

[0059] ( or (etc.) indicates relative to the target record forward Each sampling step is 15 min; , , and These correspond to 15 min, 60 min, 24 h, and 48 h, respectively.

[0060] For post-event diagnosis, target time observations of wind speed, actual power changes, and ramp event labels can be used, but these fields must not be used in the day-ahead or intraday models. If the inspector finds that a target time observation field is used in a day-ahead model, it records "Time not allowed"; if it finds that an intraday model is directly ranked with a day-ahead model, it records "Cross-task not allowed".

[0061] The preset validity rules include missing value checks, sentinel value checks, engineering range checks, timestamp checks, and quality code checks, which can be set according to the wind farm's operating conditions.

[0062] The required data stream calculates the native validity mask according to the preset validity rules. and native support set :

[0063] in For necessary data streams Record index at target time The value at (indexed time point), For the indicator function, "observed" refers to the index recorded at the target time. Above, necessary data flow value It actually exists and can be read, rather than being a missing value, null value, or ungenerated value. A value of 1 indicates a required data stream. Record index at target time The above has been observed and approved by the pre-declared validity rules; if any condition is not met, It is 0.

[0064]

[0065] Indicates the target wind power forecast type The native support set. Indexed by the target time record. Corresponding required data stream Make a judgment and use the index of the target time record corresponding to the necessary data stream for the judgment. Enter .

[0066] Required data stream elements in Iterate through all required input streams for the task / candidate service declaration. Due to the various Only 0 or 1 can be selected; only the target time record index is available. The product is 1 only when every element of the required data stream is natively valid, corresponding to the target time record index. Just entered If any required input (required data stream) is invalid, the corresponding target time record index will be invalid. It will not be included in the native support set. For elements in the required data stream; Native support set Combine the necessary data stream with a fallback strategy to construct a prediction support set. :

[0067] Tag validity builds a tag support set Then, based on the prediction support set and tag support set Construct a rating support set :

[0068]

[0069] in, For elements in the label data stream.

[0070] Rating Support Set It is "a certain necessary data stream" The output is the intersection of the two conditions: "able to output a prediction" and "a valid true label exists in the index of records for the same target time". The scoring support set is a single required data stream. The valid denominator of the self-error and scoring coverage, or the basis for constructing a common scoring support set, ensures that the post-processing model of wind power prediction and the reference prediction service are paired and compared on the same target time record index. At the same time, it can also distinguish between two types of coverage loss: "the model cannot produce data" and "there are existing predictions but the true labels are missing", which facilitates auditing rollback and the reasons for the missing data.

[0071] If the target wind power prediction type does not have a corresponding fallback strategy (i.e., no fallback is declared), then the native support set That is, the prediction support set .

[0072] If the target wind power prediction type has a corresponding backoff strategy (i.e., declared backoff / already declared backoff), and a prediction support set is constructed based on the backoff strategy... If the required data stream is natively invalid under the preset validity rules, a pre-declared input-level backoff is triggered for that record: data that meets the target wind power prediction type, preset validity rules, and backoff strategy is selected from historical multi-source data and used as backoff input to replace the required data stream, forming the prediction support set. For example, replace the D-1 history (multi-source data) that is unconfirmed at the cutoff time with a strict D-2 history (multi-source data). Or use the median fitted only by the training set for the sparse auxiliary flow.

[0073] If the input-level fallback meets the triggering conditions but cannot be triggered, then the service-level fallback in the fallback strategy is triggered: a prediction support set is constructed based on the reference prediction service or by calling the fallback model service. The process of building the forecast is documented. For example, the reference forecast service for a daytime task is typically the raw D1 power forecast collected and published before the deadline; the reference forecast service for an intraday task can be a 15-minute persistent forecast calculated based on the most recent available measured power. The reference forecast service must be associated with the task type, publication time, and version.

[0074] Calculate the data coverage of this required data stream:

[0075] The data coverage of each set is:

[0076] in Desirable native support set and prediction support set .

[0077] Then native support set The data coverage rate is:

[0078] Prediction support set The data coverage rate is:

[0079] in The time index (index time period) corresponding to the target wind power prediction type. When the prediction support set covers the entire When the prediction support set does not cover all prediction services, the corresponding prediction service will be marked as a fully supported service; when the prediction support set does not cover all prediction services, the corresponding prediction service will be marked as a fully supported service. If the evaluation protocol does not request full support output or the predicted support set coverage is not lower than the pre-declared preset coverage service threshold, the corresponding predicted service will be marked as a conditional service, and the uncovered services will be recorded. And the reasons.

[0080] When the evaluation protocol requests full support output and the coverage of the prediction support set is lower than the preset coverage service threshold, the corresponding prediction service is marked as a restricted service. Restricted services retain outputs within the prediction support set for auditing or conditional evaluation, mark target time indices outside the prediction support set as non-output, and prohibit the service from entering the common ranking of fully supported services. If the evaluation protocol pre-declares an executable service-level fallback, the prediction support set is reconstructed after the fallback is executed, and the service label is re-determined according to the aforementioned coverage conditions.

[0081] S4. The post-processing model for wind power prediction includes the post-processing model of the model output statistical model MOS, the post-processing model of the Ridge regression model, the post-processing model of the tree model, and the post-processing model of the time-series neural network model. A single post-processing model can be used to correct the power prediction data, or multiple post-processing models can be used to correct the power prediction data and compare them on a common score support set.

[0082] The admissible feature set is input into the post-processing model for wind power prediction to calculate the residuals of the power prediction data. :

[0083] in, The actual power of the wind farm in the multi-source data. The predicted power of wind farms in multi-source data; The predicted power is corrected based on the residuals to obtain the corrected predicted power. :

[0084]

[0085] in, During training, the post-processing model for wind power prediction uses candidate features declared by a pre-defined evaluation protocol. These features are then checked against the task time region, validity rules, and backoff rules. The model is then processed according to the frozen field order and the time parameters, which are composed solely of preprocessed parameters fitted from the training set. The admission-eligible feature vectors are used in the training of the wind power prediction post-processing model. and One-to-one matching, through You can then find the unique corresponding one. .

[0086] Model output statistical model according to Fit, where The intercept is... and These are the regression coefficients fitted and frozen using only the training set. Record for target time MOS corrected predicted power; This is the original day-ahead power prediction; To forecast wind speed, Index the target time record. The corrected forecast power can be obtained by linearly correcting the original day-ahead forecast and predicted wind speed.

[0087] Direct Ridge regression adds hourly, intra-year, and monthly periodic features; strict historical Ridge regression adds D-2 power at the same time, D-2 residuals, and exponentially weighted bias calculated using only data prior to D-2; intraday models add recent target daily power and residuals.

[0088] According to the tag support set Prediction support set The reference prediction service calculates a joint score support set. :

[0089] in, A prediction support set built for the reference prediction service. ... The prediction support set corresponding to the first wind power prediction post-processing model is then updated to the first... Prediction support set of a post-processing model for wind power prediction. They are respectively the 1st to the 1st Each candidate model can produce a set of record indexes that produce valid predictions when its required inputs are natively valid or can perform pre-declared fallbacks, when there is only one post-processing model. Record index set .

[0090] Based on the common score support set Calculate error index : .

[0091] in, for In the wind power prediction post-processing model, the first One evaluated post-processing model for wind power prediction. This represents the mean absolute error of the current post-processing model on the same common score support set. For each Calculate actual power With model prediction The absolute differences are then averaged.

[0092] This refers to the corrected predicted power for the evaluated wind power prediction post-processing model. When only one post-processing model corrects the power prediction data... This is the set of record indexes obtained by intersecting the prediction support set corresponding to the post-processing model, the prediction support set constructed by the reference prediction service, and the label support set. When multiple post-processing models correct the power prediction data, That is, the record index set obtained by taking the intersection of the prediction support set corresponding to several processing models, the prediction support set constructed by the reference prediction service, and the label support set.

[0093] The post-processing model for calculating wind power prediction and the reference prediction service are on the same common score support set. The absolute error and average loss difference are as follows:

[0094]

[0095] in, The corrected forecast power is used as a reference for the forecast service. Index for records of the same target time The absolute error of pairing on The average loss difference between the post-processing model for wind power prediction and the reference prediction is used to calculate the average loss difference. The post-processing model for wind power prediction and the corrected predicted power of the reference prediction service are evaluated.

[0096] Employing a time-dependent overlapping move block bootstrapping method: retaining blocks in chronological order. Construct overlapping blocks of 96 records each; with 15-minute sampling, one block corresponds to 24 hours; extract blocks with replacement and splice them to the original sample length, repeating this process 2000 times, calculating each time... ;by The 2.5% and 97.5% quantiles form a 95% confidence interval. If the upper limit of the confidence interval is less than 0, the mean absolute error of the candidate model is lower than that of the reference model; if the lower limit is greater than 0, the mean absolute error of the candidate model is higher than that of the reference model; if the interval includes 0, it cannot be confirmed that there is a difference in the mean absolute errors of the two models.

[0097] Another approach is daily clustering resampling: using complete calendar days as sampling units, replacing the sampling date and retaining all records within the day, this is used for evaluating intra-day correlation structures such as directional climbs. Block length, number of repetitions, and interval methods should all be pre-declared in the evaluation protocol.

[0098] When the set of features that cannot be admitted exceeds the preset rejection threshold, the data coverage of the common scoring support set is lower than the preset minimum coverage threshold, or the corrected predicted power of the wind power prediction post-processing model exceeds the preset error threshold, the corrected predicted power of the reference prediction service or the call rollback model service will be used as the output result of the full support service.

[0099] The common score support set Based on seasonal risk stratification, predicted wind speed risk stratification, and power ramp risk stratification, the risk is divided into several risk subsets. Data coverage, common scoring support subsets, error indicators, and average loss differences are calculated for each risk subset. Risk stratification condition index set for:

[0100] in, Stratification can be performed for seasonal risk, predicted wind speed risk, or power ramp risk. Then the common score support subset for:

[0101] Support subsets for joint scoring of seasonal risk, joint scoring of predicted wind speed risk, or joint scoring of power ramp risk; Corresponding data coverage :

[0102] in, The number of elements in the set; Corresponding error index for:

[0103] The corresponding average loss difference is :

[0104]

[0105] The risk subset is calculated based on the average loss difference. When the upper bound of the interval is lower than the minimum value of the preset threshold of the same dimension, the risk subset is marked as a low-risk subset. When the lower bound of the interval is higher than the maximum value of the preset threshold of the same dimension, the risk subset is marked as a high-risk subset.

[0106] Season can be determined by the target timestamp (task time zone); the predicted wind speed layer should use the forecast wind speed available before the task time zone and should only be frozen after the quantile threshold is estimated in the training set; the power ramp layer can be defined based on the measured power changes within the evaluation interval, but it can only be used for post-event diagnosis and cannot be used as a day-ahead / intra-day model input or selector.

[0107] S5. Generate a wind power prediction assessment report based on the above assessment process and correction results, and record at least the following relevant data: task type and deadline, data and model version, native support set, prediction support set, label support set, common scoring support set, coverage, rollback strategy, removal log, reference prediction service, evaluation indicators, confidence interval and comparison qualification identifier.

[0108] The present invention also relates to a system for reliable assessment after wind power prediction, which can perform the above-described method.

[0109] The system mainly includes the following core functional modules: acquisition module, target input recognition module, data processing module, post-test module, and evaluation output module.

[0110] The data acquisition module is responsible for collecting various real-time data from multiple sources generated during the operation of the wind farm, as well as collecting relevant historical data from multiple sources, providing a comprehensive data foundation for subsequent analysis.

[0111] The target input recognition module has a set of preset evaluation protocols. These protocols can clearly define the task time area corresponding to the type of wind power forecast to be evaluated (such as day-ahead forecast), the reference forecast service to be adopted, the rollback strategy when data is missing, and the core evaluation index system, based on the type of wind power forecast to be evaluated (such as day-ahead forecast). Based on these rules, the necessary data stream, the labeled data stream, and the candidate feature set are accurately selected from the multi-source data stream.

[0112] The data processing module, based on the data requirements defined by the pre-defined evaluation protocol, performs a series of key data preprocessing and organization tasks. These include generating a mask reflecting the original validity of the data, constructing an accessible feature set for subsequent analysis, determining the dataset actually supported by the model predictions, establishing a label support set for model validation, and finally forming a scoring support set for model performance evaluation. Building upon this, the module refines the original power prediction data by calling a specialized wind power prediction post-processing model to improve its accuracy.

[0113] The posterior module calculates the pairwise loss difference between each candidate post-processing model and the reference prediction service on the common scoring support set, forms confidence intervals using a pre-declared resampling method, and generates comparison qualification labels based on whether the task type, the common scoring support set, and the reference prediction service are consistent.

[0114] The evaluation output module records the task deadline, data and model version, each support set, coverage, rollback strategy, removal logs, reference prediction service, evaluation metrics, confidence intervals, and comparison qualification identifiers, and generates a prediction evaluation report; the results of different tasks or different score support are only used for comparison and are not included in the joint ranking.

[0115] Example 2: Verification Example Using Anonymized Industrial Wind Power Logs The following data is used to illustrate how this invention exposes changes in information sets and scoring support, and does not constitute a limitation on the scope of protection. Anonymized industrial data covers the period from January 1, 2024 to April 30, 2026, totaling 81,696 15-minute timestamps. Each data stream has varying degrees of missing or invalid values; the measured power tag coverage is 94.5%, the D1 predicted and forecasted wind speed coverage is 100%, the second-path observed wind speed coverage is 35.0%, and the ultra-short-term forecast coverage is 31.8%.

[0116] As shown in Table 2, for the fixed winter / spring testing period, the D1 prediction MAE for the day-ahead task, as verified by this invention, is 51.26, the bivariate MOS is 40.53, the direct Ridge regression value is 41.44, and the strict D-2 historical Ridge regression value is 39.25. Given that the 24-hour moving block interval of the strict D-2 historical relative MOS includes zero values, this uncertainty is retained in the audit log, rather than just recording the point estimate.

[0117] The intraday update task uses recent target daily power, with a 15-minute persistence prediction MAE of 7.43, and Ridge autoregressive and Extra Trees autoregressive MAEs of 7.33 and 7.31, respectively. Because these models use recent power that was not available at the time of the previous day's release, the system labels them as intraday update tasks and ranks them only within the same task as the persistence reference, not directly with the previous day's D1 or previous day's Ridge.

[0118] The rolling start time intervals show that MOS improved D1 in three out of four time intervals and deteriorated in one time interval; Direct Ridge deteriorated in the first two time intervals and improved in the last two. The audit log therefore includes time intervals, ensuring that the time-varying nature of the model's returns is not masked by a single test interval.

[0119] Table 2 Summary of Example Information

[0120] The above are merely preferred embodiments of the present invention and are not intended to limit the present 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.

Claims

1. A method for reliable assessment after wind power prediction, characterized in that, The method includes the following steps: S1. Collect multi-source data and historical multi-source data from wind farms; S2. Based on the target wind power prediction type and the type of wind power prediction post-processing model, and in conjunction with the preset evaluation protocol, determine the task time area, reference prediction service, rollback strategy and evaluation index corresponding to the target wind power prediction type, and screen out the necessary data stream, labeled data stream and candidate features from multi-source data. S3. Generate the native validity mask, admission feature set, prediction support set, label support set, and scoring support set through the required data stream, candidate feature, and label data stream. S4. Input the admission feature set into the wind power prediction post-processing model, correct the power prediction data, and perform a pairwise comparison with the reference prediction service on the common scoring support set. S5. Generate a predictive assessment report based on the above assessment process and correction results.

2. The method for reliable assessment after wind power prediction according to claim 1, characterized in that, For any candidate feature For those not in the task time zone The process involves eliminating candidate features, constructing an admission feature set based on the eliminated candidate features, and constructing an inadmissible feature set and elimination log based on the eliminated features.

3. The method for reliable assessment after wind power prediction according to claim 2, characterized in that, The required data stream calculates a native validity mask according to preset validity rules. and native support set : in For necessary data streams Record index at target time The value on, It is an indicator function; in, For elements in the required data stream; The native support set Combine the necessary data stream with a fallback strategy to construct a prediction support set. At this point, under the preset validity rules, the required data stream is inherently invalid: Building a tag support set Then, based on the prediction support set and tag support set Construct a rating support set : in, For elements in the label data stream.

4. The method for reliable assessment after wind power prediction according to claim 3, characterized in that, If there is no corresponding fallback strategy for the target wind power prediction type, then the native support set That is, the prediction support set , If there is a corresponding backoff strategy for the target wind power prediction type, and a prediction support set is constructed based on the backoff strategy... If the required data stream is natively invalid under the preset validity rules, then the input-level backoff in the backoff strategy is triggered: data that meets the target wind power prediction type, preset validity rules, and backoff strategy is selected from historical multi-source data and used as backoff input to replace the required data stream, thus forming the prediction support set. ; If the input-level fallback meets the triggering conditions but cannot be triggered, then the service-level fallback in the fallback strategy is triggered: a prediction support set is constructed based on the reference prediction service or by calling the fallback model service. And record the construction process.

5. The method for reliable assessment after wind power prediction according to claim 4, characterized in that, Calculate the data coverage of this required data stream: Native support set The data coverage rate is: Prediction support set The data coverage rate is: in The time index corresponding to the target wind power prediction type; When the prediction support set It can cover all At that time, the predicted support set will be used. The corresponding prediction service is marked as a fully supported service; otherwise, it is marked as a conditional service, and the uncovered time index and the reason are recorded. When there is no corresponding backoff strategy for the target wind power prediction type, and When the coverage service threshold is lower than the preset threshold, the corresponding prediction support set will be... The corresponding prediction service is marked as a restricted service.

6. The method for reliable assessment after wind power prediction according to claim 5, characterized in that, The admissible feature set is input into the post-processing model for wind power prediction to calculate the residuals of the power prediction data. : in, The actual power of the wind farm in the multi-source data. The predicted power of wind farms in multi-source data; The predicted power is corrected based on the residuals to obtain the corrected predicted power. : in, For training the post-processing model for wind power prediction, the target time record index is used. The admission-eligible feature vector, where and One-to-one pairing.

7. The method for reliable assessment after wind power prediction according to claim 6, characterized in that, According to the tag support set Prediction support set The reference prediction service calculates a joint score support set. : in, A prediction support set built for the reference prediction service. ... The prediction support set corresponding to the first wind power prediction post-processing model is then updated to the first... Prediction support set of a wind power prediction post-processing model; Based on the common score support set Calculate error index : in, for In the wind power prediction post-processing model, the first One evaluated post-processing model for wind power prediction; The corrected predicted power for the post-processing model of wind power prediction being evaluated; The post-processing model for wind power prediction and the reference prediction service are calculated on the same common score support set. The absolute error and average loss difference are as follows: in, The corrected forecast power is used as a reference for the forecast service. Index for records of the same target time The absolute error of pairing on The average loss difference between the post-processing model for wind power prediction and the reference prediction is used to calculate the average loss difference. The post-processing model for wind power prediction and the corrected predicted power of the reference prediction service are evaluated.

8. The method for reliable assessment after wind power prediction according to claim 7, characterized in that, When the set of features that cannot be admitted exceeds the preset rejection threshold, the data coverage of the common scoring support set is lower than the preset minimum coverage threshold, or the corrected predicted power of the wind power prediction post-processing model exceeds the preset error threshold, the corrected predicted power of the reference prediction service or the call rollback model service will be used as the output result of the full support service.

9. The method for reliable assessment after wind power prediction according to claim 8, characterized in that, The common score support set Based on seasonal risk stratification, predicted wind speed risk stratification, and power ramp risk stratification, the risk is divided into several risk subsets. Data coverage, common scoring support subsets, error indicators, and average loss differences are calculated for each risk subset. Risk stratification condition index set for: in, Stratification can be performed for seasonal risk, predicted wind speed risk, or power ramp risk. Then the common score support subset for: Support subsets for joint scoring of seasonal risk, joint scoring of predicted wind speed risk, or joint scoring of power ramp risk; Corresponding data coverage : in, The number of elements in the set; Corresponding error index for: The corresponding average loss difference is : The risk subset is calculated based on the average loss difference. When the upper bound of the interval is lower than the minimum value of the preset threshold of the same dimension, the risk subset is marked as a low-risk subset. When the lower bound of the interval is higher than the maximum value of the preset threshold of the same dimension, the risk subset is marked as a high-risk subset.

10. A system for reliable assessment after wind power prediction, characterized in that, The system can execute a method for reliable assessment of wind power prediction as described in any one of claims 1-9; The system includes a data acquisition module, a target input recognition module, a data processing module, a post-test module, and an evaluation output module. The acquisition module collects multi-source data and historical multi-source data from the wind farm. The target input recognition module is equipped with a preset evaluation protocol. The preset evaluation protocol determines the task time area, reference prediction service, rollback strategy and evaluation index corresponding to the target wind power prediction type based on the target wind power prediction type, and filters out the necessary data stream, label data stream and candidate features from multi-source data. The data processing module combines the data determined by the preset evaluation protocol to generate a native validity mask, an accessible feature set, a prediction support set, a label support set, a scoring support set, and a common scoring support set, and then corrects the power prediction data through a wind power prediction post-processing model. The posterior module evaluates the correction results of the wind power prediction post-processing model by calculating the loss difference of the power prediction data, and performs a pairwise comparison with the reference prediction service on the common scoring support set. The evaluation output module records the data processing, calculation, and evaluation processes of several modules and generates a predictive evaluation report.

Images