A power curtailment sample identification and training weighting method for renewable energy power prediction
By collecting multi-source data in renewable energy power prediction and utilizing gradient boosting tree model and regional resource collaborative correction factor, the problem of power curtailment sample identification and training weighting was solved, achieving high-precision and robust power curtailment sample identification and training, and enhancing the adaptability and accuracy of the prediction model.
Patent Information
- Application Number
- CN202610952446.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-29
- Publication Date
- 2026-11-10
AI Technical Summary
Existing new energy power prediction models have difficulty effectively distinguishing between low output caused by power curtailment, equipment malfunctions, and local weather deviations during the training process. This leads to systematic underestimation errors in high-resource scenarios, and the models lack regional-scale resource information collaborative correction and continuous training weight output.
By collecting multi-source heterogeneous data from the target station and its area, theoretical power reference values and their deviations from actual power are generated. Combined with local meteorological elements and equipment status, conditional probability estimation is performed using a gradient boosting tree model. A collaborative correction factor for the overall abundance of regional wind and solar resources is introduced to correct the judgment results of individual stations. Continuous sample credibility weights and power curtailment level labels are output to constrain the training of the prediction model.
It significantly improves the accuracy and robustness of power rationing sample identification, enhances the output estimation accuracy and generalization ability of the prediction model in complex operating scenarios, avoids training label contamination, and strengthens the robustness and adaptability of the model.
Smart Images

Figure CN123068947A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy power generation prediction technology, and in particular to a method and apparatus for identifying and training weighted samples for power curtailment prediction of new energy power. Background Technology
[0002] The training effectiveness of renewable energy power prediction models is highly dependent on the quality of the training sample labels. In actual operation, a typical type of sample frequently occurs: meteorological resources are favorable, yet the measured power output of the target power station is consistently significantly lower than the theoretical output level. The causes of such samples are complex, potentially stemming from power curtailment under grid dispatch instructions, equipment failure, abnormal wind or irradiance data, or discrepancies between local micro-meteorological conditions and wide-area weather forecasts. If such "potentially generating but not generating" samples are directly used as labels for normal output in model training, the prediction model will mislearn the output pattern under constrained conditions as the true wind and solar resource conversion pattern, leading to systematic underestimation errors in future high-resource scenarios, severely compromising the accuracy and reliability of power prediction.
[0003] To address the aforementioned issues, existing technologies primarily follow two technical routes. The first type of method uses the deviation between the theoretical and actual power output of a single power station, setting fixed or dynamic thresholds for discrimination. Deviations exceeding the threshold are considered anomalous samples and are removed or downweighted. The limitation of this method is that relying solely on the station's own information cannot distinguish the true causes of low power output—power curtailment, equipment malfunctions, and local meteorological deviations often exhibit similar numerical characteristics on a single station's power trajectory, making reliable discrimination based solely on the deviation amount. The second type of method smooths the power curve or performs regularized cleaning and downweighting based on the abnormal shape of the power output curve. The problem with this method is that it only performs single-point data cleaning and fails to incorporate regional-scale resource information as a basis for cross-validation. When regional power curtailment occurs, the power curve of a single station remains smooth in shape, making it difficult for curve shape anomaly detection algorithms to effectively identify it.
[0004] Observations from actual operational scenarios show that the limitation of renewable energy output is often not an isolated event. Regional photovoltaic power curtailment frequently occurs even when the overall regional irradiance is good and multiple photovoltaic power plants have sufficient resources. Similarly, when regional wind resources are generally abundant, multiple wind farms may experience simultaneous curtailment. This indicates that power curtailment events exhibit significant regional synchronicity in their spatiotemporal distribution. Relying solely on local information from a single station for judgment is inherently insufficient, necessitating the introduction of regional-scale resource adequacy information to collaboratively correct the judgment results for individual stations.
[0005] The existing technologies have the following shortcomings: First, the criteria for judgment are limited to the power and meteorological information of a single power station, failing to utilize the overall conditions of regional wind and solar resources to enhance the reliability and robustness of power curtailment judgment. Second, the identification output is mostly in the form of binary labels or regularized weights, lacking continuous probability output, making it difficult to organically integrate and smoothly weight them with gradient-based prediction model training frameworks. Third, it cannot explicitly encode the typical regional power curtailment scenario of "good meteorological resources at the power station and abundant overall regional resources" into a computable criterion, resulting in insufficient identification sensitivity in such scenarios. Fourth, it lacks a systematic mechanism for distinguishing between power curtailment, equipment anomalies, and data anomalies, easily conflating various abnormal output samples, causing pollution of training labels and degradation of model performance.
[0006] Therefore, there is an urgent need for a new energy power station power curtailment sample identification method that can integrate single-station posterior probability inference with regional resource collaborative correction and output continuous training weights, so as to improve the training quality and generalization performance of the prediction model in complex operating scenarios. Summary of the Invention
[0007] The main objective of this invention is to provide a method for identifying and weighting training samples for power curtailment prediction of new energy sources.
[0008] Another objective of this invention is to propose a power curtailment sample identification and training weighting device for predicting renewable energy power.
[0009] The third objective of this invention is to provide an electronic device.
[0010] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.
[0011] To achieve the above objectives, a first aspect of the present invention proposes a method for identifying and weighting training curtailment samples for new energy power prediction, comprising:
[0012] Collect multi-source heterogeneous data of the target site and wind farms and photovoltaic power stations in the area. After time alignment and quality control, analyze the physical output capacity of the target site under the current resource endowment and operating conditions to generate theoretical power reference values and the deviation between them and actual power. The theoretical power reference value, the deviation, the local meteorological elements of the target station, the power time-series dynamic characteristics, and the equipment status are combined to form a feature vector, which is then mapped by a gradient boosting tree model to a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station. By utilizing meteorological and theoretical power data from similar stations within the region, the overall abundance of wind and solar resources in the region is quantified, and a collaborative correction factor independent of individual station judgments is constructed. The conditional probability estimate is corrected using the collaborative correction factor. Sample confidence weights and power rationing level labels are generated based on the corrected final probability to constrain the training of the prediction model, so that the model output converges with the recognition result, thereby realizing the final recognition of the power rationing state and the generation of training constraints.
[0013] Optionally, multi-source heterogeneous data from the target site and wind farms and photovoltaic power stations within the area are collected. After time alignment and quality control, the physical output capacity of the target site under the current resource endowment and operating conditions is analyzed to generate theoretical power reference values and the deviation between them and actual power, including: Collect historical actual power generation, meteorological data, and equipment status data of the target wind farm, as well as power data and meteorological data of other wind farms and photovoltaic power stations in the region, and simultaneously acquire time characteristics; Perform unified time alignment on the collected multi-source heterogeneous data to ensure that all data sequences are consistent in the time dimension; Missing data can be filled using linear interpolation or nearest-neighbor filling. Perform initial screening of abnormal data to identify and mark power jumps and meteorological extremes that exceed reasonable ranges; Based on the meteorological data of the target power station, the current wind and solar resource conditions are analyzed, the equipment availability is determined by combining the equipment status data, and the historical output characteristics are referenced to estimate the theoretical power generation capacity that the power station should have under unrestricted conditions. The theoretical power generation capacity is then compared with the actual grid-connected power collected to obtain the deviation.
[0014] Optionally, the theoretical power reference value, the deviation, and the local meteorological elements, power time-series dynamic characteristics, and equipment status of the target station are combined to form a feature vector, which is then mapped through a gradient boosting tree model to a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station, including: The theoretical power reference value and the deviation amount are used as target feature inputs to characterize the degree of power output deviation. The local meteorological elements of the target station, the power time-series change rate, and the equipment operating status are used as auxiliary interpretation feature inputs. The power time-series change rate is used to capture the dynamic characteristics of power output fluctuations, and the equipment operating status is used to distinguish the equipment performance under power rationing and non-power rationing conditions. All features are merged into a multidimensional feature vector; A gradient boosting tree model is used as a probabilistic inference framework. The feature vector is used as input and the power curtailment probability is used as output. The gradient boosting tree model is trained to perform probabilistic mapping of the power curtailment state of a single station and output the conditional probability estimate.
[0015] Optionally, by utilizing meteorological and theoretical power data from similar stations within the region, the overall abundance of wind and solar resources in the region can be quantified, and a collaborative correction factor independent of individual station judgments can be constructed, including: Based on meteorological and theoretical power data of all or sampled stations in the region, the regional wind resource index and the regional solar resource index are calculated respectively. The wind resource index is determined by the regional wind power meteorological characteristics and the regional wind power theoretical power generation, while the solar resource index is determined by the regional photovoltaic meteorological characteristics and the regional photovoltaic theoretical power generation. The regional wind resource index and the regional light resource index are integrated into the overall abundance of regional resources, which represents the comprehensive abundance level of wind and light resources at the regional scale. Using the overall abundance of regional resources as the basic component, the regional wind resource index and the regional light resource index are weighted and combined to construct the collaborative correction factor.
[0016] Optionally, the conditional probability estimate is corrected using the collaborative correction factor, including: The collaborative correction factor is used as a prior weight and applied to the conditional probability estimate. When the local meteorological conditions at the target station are favorable, and the overall abundance of regional resources indicates that the regional scenic resources are abundant, and the deviation on which the conditional probability estimate is based is still significantly positive, the final probability value after correction is increased. The inference results of a single station are jointly corrected by a consistency test based on the sufficiency of regional resources, and the final probability is output.
[0017] Optionally, sample confidence weights and power rationing level labels are generated based on the corrected final probabilities, including: Based on the numerical range of the final probability, the corresponding samples are divided into four categories of power rationing levels: normal, suspected, relatively high, and high. For samples whose final probability value is lower than the first threshold, they are identified as normal samples and assigned a high confidence weight; For samples whose final probability value is higher than the second threshold, they are identified as power-limited samples and their training weights are reduced. For samples falling between these two values, their credibility is assessed based on a combination of their final probability value and data integrity, and they are assigned a moderate weight. The power rationing level label and the sample confidence weight are output for differential control of sample contribution in subsequent prediction model training.
[0018] Optional, the process of training the prediction model by constraints includes: During the training of the prediction model, the contribution of each sample to the loss function is weighted by the sample confidence weight, so that normal samples dominate the model learning process and the impact of power-limited samples on the model parameter update is suppressed. In addition to the conventional prediction error constraint, a posterior consistency constraint term is added to ensure that the predicted power output of the model is consistent with the power curtailment status indicated by the power curtailment level label, thereby preventing the model from mislearning the limited power output sample as a normal power generation pattern. Based on the corrected final probability and the sample confidence weights, the parameters of the prediction model are updated in a rolling manner using a sliding window or online learning method, so that the recognition and prediction results can maintain adaptive adjustment capabilities over time.
[0019] To achieve the above objectives, a second aspect of the present invention provides a power curtailment sample identification and training weighting device for new energy power prediction, comprising: The data foundation modeling unit is used to collect multi-source heterogeneous data of the target site and wind farms and photovoltaic power stations in the region. After time alignment and quality control, it generates theoretical power reference values and the deviation between actual power by analyzing the physical output capacity of the target site under the current resource endowment and operating conditions. The single-station probability inference unit is used to combine the theoretical power reference value, the deviation, and the local meteorological elements, power time-series dynamic characteristics, and equipment status of the target station to form a feature vector, which is then mapped by a gradient boosting tree model into a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station. Regional collaborative construction unit is used to quantify the overall abundance of regional wind and solar resources by utilizing meteorological and theoretical power data from similar stations within the region, and to construct a collaborative correction factor independent of single-station judgment. The joint correction training unit is used to correct the conditional probability estimate using the collaborative correction factor, generate sample confidence weights and power rationing level labels based on the corrected final probability, constrain the training of the prediction model, make the model output converge with the recognition result, and realize the final recognition of the power rationing state and the generation of training constraints.
[0020] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0021] To achieve the above objectives, a third aspect of this application provides an electronic device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a power curtailment sample identification and training weighting method for new energy power prediction as described in the first aspect embodiment.
[0022] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a power curtailment sample identification and training weighting method for new energy power prediction as described in the first aspect embodiment.
[0023] The embodiments of the present invention have the following beneficial effects: (1) Compared with the discrimination method that relies solely on the deviation threshold of a single station, this method constructs a collaborative correction factor by introducing the quantitative results of the overall abundance of regional wind and solar resources, and jointly corrects the inference results of a single station, which significantly improves the accuracy and robustness of power curtailment sample identification, and is especially suitable for regional common power curtailment scenarios.
[0024] (2) Compared with the traditional method of outputting binary labels or regularized weights, this method uses a gradient boosting tree model to output continuous posterior probabilities and directly maps the final probability to sample confidence weights and power limit level labels, realizing end-to-end probabilistic output from identification to weighting, which can be smoothly connected with gradient-based prediction model training framework.
[0025] (3) In the correction stage, this method can improve the judgment strength of power rationing in the scenario where regional resources are abundant and the station has low output by joint consistency test of regional resource abundance and local conditions. At the same time, it can reduce the judgment strength when the low output can be explained by local resource conditions or equipment abnormality. This method realizes the systematic distinction of the three types of situations: power rationing, equipment abnormality and data abnormality, and effectively avoids the pollution of training labels.
[0026] (4) This method transforms the power curtailment identification results into training weighting factors and posterior consistency constraints, so that normal samples dominate the model learning process and power curtailment samples are suppressed. This avoids the prediction model from mislearning the limited power output as normal power generation law, and significantly improves the power output estimation accuracy and generalization ability of the prediction model in high resource scenarios. Attached Figure Description
[0027] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of a power curtailment sample identification and training weighting method for predicting renewable energy power is provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a hierarchical architecture for a power curtailment sample identification and training weighted system for predicting renewable energy power, provided by an embodiment of the present invention. Figure 3 This is a structural diagram of a power curtailment sample identification and training weighting device for predicting renewable energy power, provided in an embodiment of the present invention. Detailed Implementation
[0028] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0030] The following describes, with reference to the accompanying drawings, a method and apparatus for identifying and training weighting power curtailment samples for new energy power prediction, according to an embodiment of the present invention.
[0031] Example 1 This invention provides a method for identifying and weighting training samples for power curtailment prediction of renewable energy sources. Figure 1 This is a flowchart illustrating a power curtailment sample identification and training weighting method for predicting renewable energy power, provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step S1: Collect multi-source heterogeneous data of the target site and wind farms and photovoltaic power stations in the area. After time alignment and quality control, analyze the physical output capacity of the target site under the current resource endowment and operating conditions to generate theoretical power reference values and the deviation between them and actual power.
[0032] The primary foundational step in this application's embodiments is the acquisition and standardized quality control processing of multi-source heterogeneous data across the entire region. Simultaneously, it involves the calculation and extrapolation of the theoretically achievable power generation of the target wind farm and the quantification of power deviation, providing authentic, effective, and well-organized basic data support for subsequent power curtailment probability determination, regional resource correction, and model training. The data types collected in this step cover both single-site and regional dimensions, specifically including historical actual power generation data of the target wind farm, time-series meteorological data of the target wind farm, grid-connected power data of other wind farms and photovoltaic power stations within the region, regional wind and solar corresponding meteorological data, real-time operating status data of new energy equipment, and full-dimensional time-series characteristic data. The collaborative acquisition of multi-dimensional data ensures the completeness and accuracy of subsequent single-site feature modeling and regional resource feature analysis.
[0033] To address the issue of multi-source, heterogeneous, and time-inconsistent raw data collected, this application employs a standardized data quality control process for unified preprocessing, specifically including four core operations: time alignment, missing data completion, initial anomaly screening, and data standardization. Specifically, the time alignment operation unifies the power sequences, meteorological sequences, and equipment status sequences of all stations to the same time scale, completely eliminating data misalignment caused by inconsistent timing of multi-source data collection. For the common problem of short-term data loss during data acquisition, this application's embodiments preferably use linear interpolation or neighboring value filling to complete data completion, maximizing the preservation of the variation patterns and characteristics of the original time-series data. For abnormal and dirty data such as power jumps and meteorological extremes that exceed the physically reasonable range of wind and solar power generation, this step performs preliminary screening and labeling to prevent abnormal data from interfering with the accuracy of subsequent modeling calculations.
[0034] After completing all data preprocessing, for each model calculation moment, the theoretical power generation capacity of the target power station is estimated based on the quality-controlled and valid data. Specifically, by combining the real-time meteorological resource endowment of the target power station, the real-time operating conditions of the equipment, and the long-term historical output patterns and characteristics of the station, the maximum power output value that the target power station can generate under ideal operating conditions—without grid connection restrictions, equipment failures, or maintenance shutdowns—is comprehensively deduced. This value is the theoretical power generation capacity of the station. This theoretical power generation capacity can objectively reflect the station's maximum power generation capacity under current wind and solar resource conditions and equipment status, and can effectively isolate output fluctuations caused by non-human-caused power curtailment factors such as natural resource fluctuations and equipment failures.
[0035] Based on this, the calculated theoretical generating power is compared with the actual grid-connected power of the target power station collected in real time on-site to quantify the real-time power deficit of the power station, which is the output deviation mentioned in this application. The embodiments of this application complete the standardized calculation of the power deficit using a fixed formula, the specific calculation formula of which is as follows:
[0036] in, The power deficit of the target power station at time t is the real-time output deviation of the power station. This represents the theoretical power generation capacity of the target power station at time t; This represents the actual grid-connected power generation of the target power station at time t. It should be clarified that this power deficit cannot be directly equated with the grid curtailment result. It serves only as the core basis for subsequent curtailment probability judgment, used to quantitatively determine whether there is an abnormal deviation between the theoretical power generation capacity of the power station and the actual grid-connected output, and to provide quantitative support for curtailment probability inference at the single-station level.
[0037] Step S2: The theoretical power reference value, the deviation, and the local meteorological elements, power time-series dynamic characteristics, and equipment status of the target station are combined to form a feature vector, which is then mapped by a gradient boosting tree model to a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station.
[0038] Based on the theoretical power reference value, power deviation, and preprocessed standardized basic data output from step S1, this application further integrates and constructs multi-dimensional feature vectors, and uses a machine learning model to accurately infer the posterior probability of power curtailment based solely on single-station data. Compared to traditional single-indicator discrimination methods, this application's embodiments integrate multi-dimensional operational features, which can comprehensively distinguish the causes of low power output at power stations and effectively improve the accuracy of single-station power curtailment discrimination.
[0039] This step constructs a feature system encompassing six major categories: local meteorological elements of the target power station, theoretical generating power, actual generating power deviation, power time-series change rate, low-output characteristics of the platform, and equipment operating status. These multiple features complement and verify each other, accurately distinguishing whether low power output at the power station stems from normal resource fluctuations, abnormal equipment failures, or deliberate grid-connected power curtailment. In the feature combination process, the theoretical power reference value and power deviation are used as core discriminative features to directly quantify the deviation between the actual power output and the theoretical maximum generating capacity of the power station. Local meteorological elements and power time-series change rate are used as dynamic auxiliary features, relying on the power time-series change rate to capture short-term power output fluctuation trends and quickly identify abnormal power output surges. Simultaneously, combined with equipment operating status characteristics, non-curtailment-related low-output scenarios such as fault shutdowns and equipment maintenance are eliminated from the equipment operation perspective.
[0040] This embodiment integrates all the aforementioned core and auxiliary features into a unified multi-dimensional feature vector, which serves as the standard input data source for the probabilistic inference model. Regarding model selection, this embodiment preferably uses a gradient boosting tree model to complete the mapping calculation of the posterior probability of power curtailment at a single station. This model possesses excellent nonlinear fitting capabilities and complex feature association mining capabilities, and can well adapt to the industry characteristics of strong randomness and volatility in wind and solar power output. The model ultimately outputs a probability value in the range of 0 to 1, which is the posterior probability of power curtailment at the station itself. The core physical meaning of this probability is: based solely on the target station's own operational data, meteorological data, and equipment status data, to determine the probability distribution of the station's current operating status as normal operation, equipment malfunction, or grid connection restriction.
[0041] This application's embodiments establish a clear probability discrimination mapping logic, and the probability value can dynamically and adaptively change with operating conditions and resource status. When the local meteorological resource conditions of the target station are excellent, the theoretical power generation is significantly higher than the real-time actual power generation, and the equipment is operating without faults or maintenance, it can be determined that the station's low output cannot be reasonably explained by natural resource conditions or equipment operating conditions, and the posterior probability of the station's own power curtailment increases accordingly. Conversely, when the station's actual power generation is low, it can be reasonably explained by objective factors such as weak wind and solar resources, low irradiance, equipment failure, and equipment maintenance, the possibility of deliberate power curtailment can be ruled out, and the posterior probability of the station's own power curtailment decreases accordingly, thereby completing the preliminary accurate discrimination of power curtailment probability at the single-station level.
[0042] Step S3: Using meteorological and theoretical power data from similar stations within the region, quantify the overall abundance of wind and solar resources in the region and construct a collaborative correction factor independent of individual station judgments.
[0043] To overcome the technical shortcomings of single-station independent judgment, which is easily affected by the individual operating conditions of the stations and has a high misjudgment rate, this application goes beyond the limitations of single-station data analysis. It constructs a collaborative correction factor from the perspective of regional resources to assist in the verification and optimization of single-station judgment results. This step does not rely on external operation and maintenance data such as regional synchronous operating status and transmission section load pressure. It only relies on meteorological data and theoretical power data of each new energy station in the region to complete the quantitative assessment of regional resource adequacy, effectively improving the generalization ability and adaptability of the overall algorithm to complex scenarios.
[0044] This application employs a hierarchical and progressive computational logic to construct the regional resource synergistic correction factor. First, the regional wind resource index and the regional solar resource index are quantified and solved separately. Then, the two indices are fused to obtain the overall sufficiency of regional resources. Finally, a standardized regional resource synergistic correction factor is generated through a weighted combination method. Specifically, the regional wind resource index is calculated jointly from the regional wind power meteorological characteristics and the overall theoretical generating capacity of regional wind power, objectively and quantitatively representing the abundance of overall regional wind resources. The regional solar resource index is calculated jointly from the regional photovoltaic meteorological characteristics and the overall theoretical generating capacity of regional photovoltaic power, accurately reflecting the sufficiency level of overall regional solar resources.
[0045] After obtaining the regional wind resource index and the regional solar resource index separately, the two resource indices are fused to obtain the overall regional resource sufficiency that comprehensively reflects the overall wind and solar resource level of the region, thus achieving a standardized and quantitative representation of the regional resource status. The embodiment of this application uses the following formula to calculate the overall regional resource sufficiency:
[0046] in, This indicates the overall sufficiency of regional resources at time t. It is the weighting coefficient for wind and solar integration, with a fixed value range of 0 to 1. It can be adaptively adjusted according to the proportion of wind power and photovoltaic installed capacity in the region to adapt to the power structure characteristics of different regions. This represents the regional wind resource index at time t; This represents the regional light resource index at time t.
[0047] Based on this, this application provides a preferred implementation method, which constructs a regional resource synergy correction factor through a multi-component weighted combination formula. The specific calculation formula is as follows:
[0048] in, The regional resource coordination correction factor at time t is used to characterize the correction weight and correction intensity of the overall wind and solar resource status of the region on the judgment of power curtailment of the target station. The region's wind resource index at time t represents the abundance of wind power resources in the region, with a larger value indicating more abundant wind power resources. The region's solar resource index at time t represents the abundance of photovoltaic solar resources in the region, with a larger value indicating more abundant solar resources. t represents the overall resource sufficiency of the region at time t, used to comprehensively characterize the overall abundance level of the region's scenic resources; This is the weighting coefficient corresponding to the regional wind resource index, used to adjust the contribution of wind power resources to the correction factor; This is the weighting coefficient corresponding to the regional solar resource index, used to adjust the contribution of photovoltaic resources to the correction factor; The weighting coefficient corresponding to the overall sufficiency of regional resources is used to regulate the overall impact of comprehensive resource level on the correction factor, and All are non-negative real numbers.
[0049] Regarding the aforementioned weighting coefficients, this application provides two feasible parameter calibration methods: manual calibration based on massive historical operating samples, or embedding the weighting parameters into the model training process and automatically optimizing the optimal weight values through iterative learning. This ensures that the correction factor can adapt to the characteristics of wind and solar resources in different regions and seasons. This regional resource collaborative correction factor can accurately reflect the overall status of regional wind and solar resources. When the overall regional wind and solar resources are abundant, but the target power station continues to experience significantly low power output, it can corroborate that the target power station is likely experiencing grid-connected power curtailment, providing a reliable regional dimension for subsequent joint probability correction.
[0050] Step S4: The conditional probability estimate is corrected using the collaborative correction factor. Sample confidence weights and power rationing level labels are generated based on the corrected final probability to constrain the training of the prediction model, so that the model output converges with the recognition result, thereby realizing the final recognition of the power rationing state and the generation of training constraints.
[0051] After obtaining the initial posterior probability of power curtailment at the single-station level and the regional resource collaborative correction factor, this embodiment combines the consistency relationship between the single-station output characteristics and the overall regional resource characteristics. Through joint correction of the posterior probability after regional resource consistency verification, it overcomes the limitations of relying solely on single-station data and outputs an accurate final posterior probability of power curtailment. Based on this, power curtailment level classification and sample training weight assignment are performed using the final probability, constructing a constrained model training mechanism to achieve accurate iteration and online rolling updates of the prediction model.
[0052] The core correction logic in this step is a two-dimensional consistency verification of single-station and regional conditions, taking into account both the local operating conditions of the target station and the overall resource status of the region, rather than simply relying on the output deviation of a single station or the average operating status of the region. This embodiment of the application completes the optimization and correction of the basic probability of a single station through a dedicated correction formula. The specific joint posterior probability correction formula is as follows:
[0053] in, Let be the final posterior probability of power rationing after correction at time t, with its numerical range strictly constrained to be between 0 and 1; Let t be the posterior probability of the initial power outage at the power station itself, which is the basic probability value calculated based solely on data from a single station. This is a preset correction sensitivity coefficient used to adjust the strength of the regional correction factor's correction to the base probability. This is the regional resource coordination correction factor corresponding to time t.
[0054] The specific correction rules are as follows: When the local meteorological conditions of the target power station are favorable, its theoretical power generation capacity is sufficient, and the wind and solar resources of similar power stations and adjacent new energy power stations in the region are also abundant, and multiple dimensions verify that the current power generation conditions are excellent, but the actual power generation capacity of the target power station is still significantly low. A larger value corresponds to a higher posterior probability of the final power curtailment after correction; conversely, if the low output of the target power station can be reasonably explained by objective factors such as weak wind and solar resources, low irradiance, equipment failure, or abnormal data collection, the weight of the power curtailment judgment is reduced, and the posterior probability of the final power curtailment is lowered. This single-station-region joint correction method effectively avoids misjudgment problems caused by fluctuations in single-station data, significantly improving the stability and accuracy of identifying large-scale regional power curtailment scenarios.
[0055] After completing probability correction and obtaining the final power curtailment probability, this embodiment of the application performs refined classification and differentiated weight assignment on all training samples to construct power curtailment labels and training weighting factors adapted to model training. Specifically, the final power curtailment probability is divided into four levels: normal level, suspected power curtailment level, relatively high power curtailment level, and high power curtailment level, achieving accurate classification of sample operating conditions. At the sample training weight configuration level, a differentiated assignment strategy is adopted: for samples with high data integrity and strong credibility, and a final power curtailment probability lower than the first preset threshold, they are judged as normal power generation samples and given higher training weights, allowing them to dominate the model's learning of normal output patterns; for samples with a final power curtailment probability higher than the second preset threshold, clearly existing power curtailment conditions, or samples with insufficient data integrity and low credibility, their model training weights are significantly reduced, retaining only their data trend constraint information, effectively preventing the model from mistakenly learning restricted output conditions as normal power generation capacity of the station.
[0056] In the specific training process of the model, this application adds a posterior consistency constraint term to the conventional power prediction error loss constraint, constructing a dual-constraint training mechanism to further improve the robustness of the model. The embodiments of this application implement sample difference constraint training through a weighted loss function, the specific loss function formula of which is as follows:
[0057] in, This represents the total loss value for a single iteration of model training. The confidence training weights are the corresponding to the i-th training sample. The predicted power generation of the i-th sample output by the model; The actual power generation of the station corresponding to the i-th sample; These are the consistency constraint weighting coefficients, used to adjust the influence strength of the posterior constraint terms; This is a posterior consistency constraint term, used to ensure that the model's predicted output is highly consistent with the power rationing identification tag and the final power rationing probability determination result.
[0058] By employing the aforementioned dual-constraint training method, the problem of the model mislearning the output characteristics of power curtailment samples can be effectively avoided, significantly improving the model's robustness and generalization ability in scenarios where power curtailment samples are mixed into the training set. Simultaneously, this application's embodiments adopt a rolling update strategy combining sliding window and online learning. Relying on real-time updated sample weights and power curtailment labels, the prediction model parameters are continuously iteratively optimized, enabling the model to adapt to temporal changes in wind and solar resources, evolving operating conditions of power plant equipment, and dynamic changes in power curtailment scenarios, achieving long-term stable operation of power prediction and power curtailment status identification for new energy power plants.
[0059] Example 2 This application also relates to a power curtailment sample identification and training weighting system for new energy power prediction. The entire system adopts a layered architecture to complete the entire process of data flow, power curtailment identification, model training, and display. The overall hierarchy of the layered architecture, from top to bottom, is as follows: data access layer, identification service layer, training service layer, model management layer, site access layer, and result display layer. The overall structure of the layered architecture is as follows: Figure 2 As shown.
[0060] like Figure 2 As shown, the data access layer serves as the data input terminal for the entire system, undertaking the acquisition and preprocessing of raw data from multiple sources across the entire domain. This corresponds to the data access module within the system. This module is responsible for collecting various heterogeneous data, including the historical actual power generation of the target power station, station meteorological data, power and meteorological data of regional wind and solar power stations, equipment status, and time-series characteristics. It uniformly performs time alignment, missing data filling, anomaly filtering, and standardization processing, synchronously calculates the theoretical power generation capacity and power deficit of the target power station at each time point, and outputs standardized characteristic data that has undergone quality control and basic calculations, which is then transmitted to the downstream identification service layer.
[0061] The identification service layer integrates three core computing units: a station-specific posterior probability calculation module, a regional resource collaborative correction factor generation module, and a joint posterior correction module. It receives standardized time-series features from the data access layer to perform power curtailment probability projections. The station-specific posterior probability calculation module extracts multi-dimensional features from single-station meteorology, output deviation, power change rate, and equipment status, and outputs the basic posterior probability of power curtailment based solely on single-station data using a gradient boosting tree model. The regional resource collaborative correction factor generation module calculates wind and solar resource indices based on regional wind and solar power station data, merges them to obtain the overall sufficiency of regional resources, and calculates the regional resource collaborative correction factor using a weighted formula. The joint posterior correction module introduces correction factors to jointly correct the basic probability of single-station power curtailment, outputting the final posterior probability of power curtailment incorporating regional resource information. The identification service layer then pushes this final probability data down to the training service layer.
[0062] The training service layer includes a built-in power rationing label generation module, a training weighted generation module, and a prediction model training module. It is responsible for classifying samples and training them under constraints based on power rationing probabilities. The power rationing label generation module classifies power rationing levels into four categories: normal, suspected, relatively high, and high, based on the final posterior probability of power rationing. The training weighted generation module allocates training weights differently based on sample credibility, reducing the loss contribution weight of samples with high power rationing risk. The prediction model training module uses a weighted loss function to iterate the model and adds a posterior consistency constraint term to ensure that the predicted output matches the power rationing identification result. The trained model parameters are then delivered to the model management layer for unified management.
[0063] The model management layer is responsible for storing all prediction models, archiving versions, and caching parameters, while also carrying the operational logic of the online update module. The online update module relies on a sliding window online learning mechanism to continuously receive new time-series samples, power rationing labels, and weighting factors from the power plant, iteratively optimizes the model's internal parameters, and distributes the updated, usable model to the power plant's calling layer.
[0064] As a middleware layer for business interaction, the power station call layer provides standardized model call interfaces for each new energy power station. The power station operation and maintenance terminal can initiate real-time power output prediction and power curtailment status identification requests as needed. The power station call layer retrieves the latest iterative model stored in the model management layer to complete online calculations and transmits the power prediction results and power curtailment level determination results to the result display layer in a unified manner.
[0065] The results display layer serves as the visualization output of the entire system, providing operation and maintenance personnel with an intuitive view of all business data, including the target power station's theoretical generating power, real-time generating power, power deficit, basic probability of single-station power curtailment, regional resource correction factor, final power curtailment posterior probability, power curtailment classification label, and power station output prediction curve. This allows staff to intuitively understand whether power stations are subject to grid connection curtailment and the level of power curtailment risk at different times, supporting the grid's new energy dispatch and power station operation and maintenance decision-making.
[0066] The entire system transmits data and calculation results from top to bottom at each level. Each internal functional module corresponds to the business functions of each level in the hierarchical architecture, fully implementing all the power curtailment identification and weighted training logic of S1 to S4 mentioned above in this application. A closed loop is formed from the data source to the visualization output. It can not only complete the power curtailment labeling and model training of historical power plant samples offline in batches, but also identify the power curtailment status of power plants online in real time and adaptively update the prediction model, effectively reducing the probability of misjudgment of power curtailment at a single station and improving the robustness of the new energy power prediction model in mixed power curtailment sample scenarios.
[0067] Example 3 This invention provides a power curtailment sample identification and training weighting device for predicting renewable energy power output. Figure 3 This is a schematic diagram of a power curtailment sample identification and training weighting device for predicting renewable energy power, provided in an embodiment of the present invention. Figure 3 As shown, the device includes: The data foundation modeling unit 100 is used to collect multi-source heterogeneous data of the target site and wind farms and photovoltaic power stations in the area. After time alignment and quality control, it generates theoretical power reference values and the deviation between actual power by analyzing the physical output capacity of the target site under the current resource endowment and operating conditions. The single-station probability inference unit 200 is used to combine the theoretical power reference value, the deviation, and the local meteorological elements, power time-series dynamic characteristics, and equipment status of the target station to form a feature vector, which is then mapped by a gradient boosting tree model into a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station. Regional collaborative construction unit 300 is used to quantify the overall abundance of regional wind and solar resources by utilizing meteorological and theoretical power data from similar stations within the region, and to construct a collaborative correction factor independent of single-station judgment. The joint correction training unit 400 is used to correct the conditional probability estimate using the collaborative correction factor, generate sample confidence weights and power rationing level labels based on the corrected final probability, constrain the training of the prediction model, make the model output converge with the recognition result, and realize the final recognition of the power rationing state and the generation of training constraints.
[0068] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0069] Example 4 To implement the methods of the above embodiments, the present invention also provides an electronic device, which includes a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the various steps of the methods described above.
[0070] Example 5 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.
[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0072] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0073] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A method for identifying and weighting training curtailment samples for new energy power prediction, characterized in that, include: Collect multi-source heterogeneous data of the target site and wind farms and photovoltaic power stations in the area. After time alignment and quality control, analyze the physical output capacity of the target site under the current resource endowment and operating conditions to generate theoretical power reference values and the deviation between them and actual power. The theoretical power reference value, the deviation, the local meteorological elements of the target station, the power time-series dynamic characteristics, and the equipment status are combined to form a feature vector, which is then mapped by a gradient boosting tree model to a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station. By utilizing meteorological and theoretical power data from similar stations within the region, the overall abundance of wind and solar resources in the region is quantified, and a collaborative correction factor independent of individual station judgments is constructed. The conditional probability estimate is corrected using the collaborative correction factor. Sample confidence weights and power rationing level labels are generated based on the corrected final probability to constrain the training of the prediction model, so that the model output converges with the recognition result, thereby realizing the final recognition of the power rationing state and the generation of training constraints.
2. The method according to claim 1, characterized in that, Multi-source heterogeneous data from the target site and surrounding wind farms and photovoltaic power stations are collected. After time alignment and quality control, the physical output capacity of the target site under current resource endowment and operating conditions is analyzed to generate theoretical power reference values and their deviations from actual power, including: Collect historical actual power generation, meteorological data, and equipment status data of the target wind farm, as well as power data and meteorological data of other wind farms and photovoltaic power stations in the region, and simultaneously acquire time characteristics; Perform unified time alignment on the collected multi-source heterogeneous data to ensure that all data sequences are consistent in the time dimension; Missing data can be filled using linear interpolation or nearest-neighbor filling. Perform initial screening of abnormal data to identify and mark power jumps and meteorological extremes that exceed reasonable ranges; Based on the meteorological data of the target power station, the current wind and solar resource conditions are analyzed, the equipment availability is determined by combining the equipment status data, and the historical output characteristics are referenced to estimate the theoretical power generation capacity that the power station should have under unrestricted conditions. The theoretical power generation capacity is then compared with the actual grid-connected power collected to obtain the deviation.
3. The method according to claim 2, characterized in that, The theoretical power reference value, the deviation, and the local meteorological elements, power time-series dynamic characteristics, and equipment status of the target station are collectively used to form a feature vector. This vector is then mapped through a gradient boosting tree model to a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station. This estimate includes: The theoretical power reference value and the deviation amount are used as target feature inputs to characterize the degree of power output deviation. The local meteorological elements of the target station, the power time-series change rate, and the equipment operating status are used as auxiliary interpretation feature inputs. The power time-series change rate is used to capture the dynamic characteristics of power output fluctuations, and the equipment operating status is used to distinguish the equipment performance under power rationing and non-power rationing conditions. All features are merged into a multidimensional feature vector; A gradient boosting tree model is used as a probabilistic inference framework. The feature vector is used as input and the power curtailment probability is used as output. The gradient boosting tree model is trained to perform probabilistic mapping of the power curtailment state of a single station and output the conditional probability estimate.
4. The method according to claim 3, characterized in that, By utilizing meteorological and theoretical power data from similar stations within the region, the overall abundance of wind and solar resources in the region is quantified, and a collaborative correction factor independent of individual station judgments is constructed, including: Based on meteorological and theoretical power data of all or sampled stations in the region, the regional wind resource index and the regional solar resource index are calculated respectively. The wind resource index is determined by the regional wind power meteorological characteristics and the regional wind power theoretical power generation, while the solar resource index is determined by the regional photovoltaic meteorological characteristics and the regional photovoltaic theoretical power generation. The regional wind resource index and the regional light resource index are integrated into the overall abundance of regional resources, which represents the comprehensive abundance level of wind and light resources at the regional scale. Using the overall abundance of regional resources as the basic component, the regional wind resource index and the regional light resource index are weighted and combined to construct the collaborative correction factor.
5. The method according to claim 4, characterized in that, Correcting the conditional probability estimate using the collaborative correction factor includes: The collaborative correction factor is used as a prior weight and applied to the conditional probability estimate. When the local meteorological conditions at the target station are favorable, and the overall abundance of regional resources indicates that the regional scenic resources are abundant, and the deviation on which the conditional probability estimate is based is still significantly positive, the final probability value after correction is increased. The inference results of a single station are jointly corrected by a consistency test based on the sufficiency of regional resources, and the final probability is output.
6. The method according to claim 5, characterized in that, Based on the corrected final probability, sample credibility weights and power rationing level labels are generated, including: Based on the numerical range of the final probability, the corresponding samples are divided into four categories of power rationing levels: normal, suspected, relatively high, and high. For samples whose final probability value is lower than the first threshold, they are identified as normal samples and assigned a high confidence weight; For samples whose final probability value is higher than the second threshold, they are identified as power-limited samples and their training weights are reduced. For samples falling between these two values, their credibility is assessed based on a combination of their final probability value and data integrity, and they are assigned a moderate weight. The power rationing level label and the sample confidence weight are output for differential control of sample contribution in subsequent prediction model training.
7. The method according to claim 6, characterized in that, The training process of a constrained prediction model includes: During the training of the prediction model, the contribution of each sample to the loss function is weighted by the sample confidence weight, so that normal samples dominate the model learning process and the impact of power-limited samples on the model parameter update is suppressed. In addition to the conventional prediction error constraint, a posterior consistency constraint term is added to ensure that the predicted power output of the model is consistent with the power curtailment status indicated by the power curtailment level label, thereby preventing the model from mislearning the limited power output sample as a normal power generation pattern. Based on the corrected final probability and the sample confidence weights, the parameters of the prediction model are updated in a rolling manner using a sliding window or online learning method, so that the recognition and prediction results can maintain adaptive adjustment capabilities over time.
8. A power curtailment sample identification and training weighting device for predicting renewable energy power, characterized in that, include: The data foundation modeling unit is used to collect multi-source heterogeneous data of the target site and wind farms and photovoltaic power stations in the region. After time alignment and quality control, it generates theoretical power reference values and the deviation between actual power by analyzing the physical output capacity of the target site under the current resource endowment and operating conditions. The single-station probability inference unit is used to combine the theoretical power reference value, the deviation, and the local meteorological elements, power time-series dynamic characteristics, and equipment status of the target station to form a feature vector, which is then mapped by a gradient boosting tree model into a conditional probability estimate representing the possibility of power output limitation when relying solely on information from this station. Regional collaborative construction unit is used to quantify the overall abundance of regional wind and solar resources by utilizing meteorological and theoretical power data from similar stations within the region, and to construct a collaborative correction factor independent of single-station judgment. The joint correction training unit is used to correct the conditional probability estimate using the collaborative correction factor, generate sample confidence weights and power rationing level labels based on the corrected final probability, constrain the training of the prediction model, make the model output converge with the recognition result, and realize the final recognition of the power rationing state and the generation of training constraints.
9. An electronic device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.
