A day-ahead power probability forecasting method and device under meteorological uncertainty

Abstract

This application provides a method and apparatus for predicting day-ahead power probability under meteorological uncertainty, relating to the field of day-ahead power probability prediction technology, and solving the technical problem of poor coordination between power range and density prediction results in existing technologies when predicting day-ahead power probability under meteorological uncertainty. The method specifically includes: acquiring meteorological station data; preprocessing the meteorological station data; quantifying the meteorological uncertainty in the meteorological station data by using probability distribution fitting and combining it with time series deviation correction, obtaining uncertainty quantification features; inputting the preprocessed meteorological station data and uncertainty quantification features into a two-branch collaborative model, outputting intermediate results for power range prediction and power density prediction; iteratively optimizing the two-branch collaborative model using an alternating optimization strategy, introducing collaborative constraints and adaptive weight adjustment, until the two-branch collaborative model converges, and outputting the power probability prediction result. This application is used for day-ahead power probability prediction.

Description

Technical Field

[0001] This application relates to the field of day-ahead power probability prediction technology, and in particular to a method and apparatus for day-ahead power probability prediction under meteorological uncertainty. Background Technology

[0002] Probabilistic prediction of day-ahead power of new energy sources under meteorological uncertainties is a key technical support for power system dispatch and trading risk control.

[0003] Existing technologies typically begin by acquiring NWP meteorological data and station power data and performing basic preprocessing. Then, meteorological uncertainties are quantified holistically through a single probability distribution fitting. Subsequently, a single-branch model is used to simultaneously predict power ranges and density. Some technologies perform simple overall iterative optimization of the model without setting collaborative constraints between branches. These technologies fail to hierarchically decompose and accurately quantify the primary uncertainties of meteorological factor fluctuations and the secondary uncertainties of meteorological prediction errors. Meteorological characteristics quantified using only a single method cannot truly reflect actual meteorological fluctuation patterns. Furthermore, single-branch models struggle to balance the reasonableness of range prediction boundaries with the probabilistic accuracy of density predictions. The lack of collaborative constraints in overall optimization can easily lead to deviations or even conflicts between the two prediction results, resulting in a low degree of matching between the output power probability prediction and actual operating conditions, failing to accurately meet the day-ahead power prediction requirements under meteorological uncertainties.

[0004] Therefore, existing technologies suffer from poor coordination between power range and density prediction results when predicting day-ahead power probability under meteorological uncertainties. Summary of the Invention

[0005] This application provides a method and apparatus for predicting day-ahead power probability under meteorological uncertainty, which solves the technical problem of poor coordination between power range and density prediction results in the prior art when predicting day-ahead power probability under meteorological uncertainty.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for predicting day-ahead power probability under meteorological uncertainty is provided, comprising: acquiring meteorological station data; the meteorological station data includes NWP data and target station power data; preprocessing the meteorological station data; quantifying the meteorological uncertainty in the meteorological station data by using probability distribution fitting and combining it with time series deviation correction, and obtaining uncertainty quantification features; inputting the preprocessed meteorological station data and uncertainty quantification features into a two-branch collaborative model, and outputting intermediate results of power interval prediction and power density prediction; the two-branch collaborative model includes a feature fusion layer, a first branch, and a second branch; the feature fusion layer is used to fuse the preprocessed meteorological station data and uncertainty quantification features to obtain fused features; the first branch is used to perform power interval prediction based on upper and lower boundary constraints; the second branch is used to determine the power probability distribution law to perform power density prediction; iteratively optimizing the two-branch collaborative model using an alternating optimization strategy, introducing collaborative constraints and adaptive weight adjustment, until the two-branch collaborative model converges, and outputting power probability prediction results; the power probability prediction results include interval prediction results and density prediction results.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the meteorological station data is preprocessed, including: performing time-series alignment on the meteorological station data, aligning the NWP data based on the timestamp of the target station power data, and removing data with time-series misalignment; using the KNN interpolation method to fill in missing values ​​in the time-series aligned meteorological station data; and performing normalization processing on the meteorological station data after missing value filling.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves using probability distribution fitting and combining it with time-series deviation correction to quantify meteorological uncertainty in meteorological station data, thereby obtaining uncertainty quantification features. This includes: decomposing meteorological uncertainty in meteorological station data into primary uncertainty retaining the fluctuations of the meteorological factors themselves and secondary uncertainty retaining the meteorological forecast error; fitting the core meteorological factors in the meteorological station data with probability distributions to quantify primary uncertainty; the core meteorological factors are wind speed and light intensity in the NWP data; calculating the meteorological forecast deviation using a sliding window, combining the meteorological forecast deviation to correct secondary uncertainty, and integrating it with the quantified primary uncertainty to obtain uncertainty quantification features.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves power range prediction based on upper and lower boundary constraints, including: using a Transformer encoder to extract temporal correlation features and uncertainty correlation features from the fused features; outputting the upper and lower boundaries of the power prediction range through a dual-output fully connected layer; and constraining the upper and lower boundaries to ensure that the lower boundary is not less than 0 and the upper boundary does not exceed the rated power of the target power station, thereby obtaining intermediate results for power range prediction.

[0010] In conjunction with the first aspect mentioned above, one possible implementation involves determining the power probability distribution pattern for power density prediction, including: using a CNN-LSTM structure to collaboratively extract local correlation features and temporal correlation features of the fused features, wherein the CNN layer captures the local correlation features of meteorological factors and uncertainties in the fused features, and the LSTM layer captures the temporal correlation features of the fused features; the distribution parameter prediction layer outputs Gaussian mixture distribution parameters corresponding to the power probability; the Gaussian mixture distribution parameters include mean, variance, and mixture weights; and the power probability distribution pattern is determined based on the Gaussian mixture distribution parameters to obtain intermediate results for power density prediction.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves iteratively optimizing the two-branch collaborative model using an alternating optimization strategy. This includes introducing collaborative constraints and adaptive weight adjustment until the two-branch collaborative model converges and outputs a power probability prediction result. The steps include: fixing the initial loss weights and pre-training the two-branch collaborative model using the AdamW optimizer to quickly converge to a local optimum; fixing the second branch parameters and optimizing the first branch using collaborative constraints; fixing the first branch parameters and optimizing the second branch using collaborative constraints; adaptively adjusting the loss weights based on the two-branch loss reduction rate after each alternating iteration; determining whether the two-branch loss reduction rate is less than a preset threshold; if it is less, convergence occurs, and a power probability prediction result is output; if it is not less, iterative optimization continues.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the core meteorological factors in the meteorological station data are fitted using probability distributions, including: fitting the wind speed in the NWP data using a Weibull distribution, and fitting the light intensity in the NWP data using a Beta distribution.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the power probability distribution law is determined based on the Gaussian mixture distribution parameters to obtain intermediate results for power density prediction. This includes: calculating the probability density corresponding to different day-ahead power values ​​based on the mean, variance, and mixture weight of the Gaussian mixture distribution, plotting the power probability density curve, determining the distribution probability of power in different intervals, and obtaining intermediate results for power density prediction.

[0014] In conjunction with the first aspect above, in one possible implementation, the cooperative constraint is that the absolute value of the deviation between the intermediate results of the power range prediction and the intermediate results of the power density prediction does not exceed 5% of the rated power of the target power station.

[0015] Secondly, a device for predicting day-ahead power probability under meteorological uncertainty is provided, comprising: a communication unit and a processing unit; the communication unit is used to acquire meteorological station data; the meteorological station data includes NWP data and target station power data; the processing unit is used to preprocess the meteorological station data; by using probability distribution fitting and combining time series deviation correction, the meteorological uncertainty in the meteorological station data is quantified to obtain uncertainty quantification features; the preprocessed meteorological station data and uncertainty quantification features are input into a two-branch collaborative model, and intermediate results of power interval prediction and power density prediction are output; the two-branch collaborative model includes a feature fusion layer, a first branch and a second branch; the feature fusion layer is used to fuse the preprocessed meteorological station data and uncertainty quantification features to obtain fused features; the first branch is used to perform power interval prediction based on upper and lower boundary constraints; the second branch is used to determine the power probability distribution law to perform power density prediction; the two-branch collaborative model is iteratively optimized using an alternating optimization strategy, introducing collaborative constraints and adaptive weight adjustment, until the two-branch collaborative model converges, and outputs power probability prediction results; the power probability prediction results include interval prediction results and density prediction results.

[0016] This application provides a method and apparatus for predicting day-ahead power probability under meteorological uncertainty, which can accurately quantify meteorological uncertainty in layers and achieve synergistic optimization of power range and density prediction, effectively improving the accuracy and reliability of day-ahead power probability prediction results. The existing technology has the technical problem of poor synergy between power range and density prediction results when predicting day-ahead power probability under meteorological uncertainty.

[0017] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a day-ahead power probability prediction method under meteorological uncertainty provided in an embodiment of this application; Figure 2A flowchart illustrating another day-ahead power probability prediction method under meteorological uncertainty provided in this application embodiment; Figure 3 A flowchart illustrating another day-ahead power probability prediction method under meteorological uncertainty provided in this application embodiment; Figure 4 A flowchart illustrating another day-ahead power probability prediction method under meteorological uncertainty provided in this application embodiment; Figure 5 A flowchart illustrating another day-ahead power probability prediction method under meteorological uncertainty provided in this application embodiment; Figure 6 A flowchart illustrating another day-ahead power probability prediction method under meteorological uncertainty provided in this application embodiment; Figure 7 This is a schematic diagram of a day-ahead power probability prediction device under meteorological uncertainty, provided as an embodiment of this application. Detailed Implementation

[0019] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0020] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0021] To address the technical problem of poor coordination between power range and density prediction results in day-ahead power probability prediction under meteorological uncertainty in existing technologies, this application provides a method for day-ahead power probability prediction under meteorological uncertainty. The method includes: first, acquiring and preprocessing meteorological station data containing NWP data and target station power data; then, obtaining uncertainty quantification features by hierarchically quantifying meteorological uncertainty through probability distribution fitting combined with time-series deviation correction; subsequently, inputting the preprocessed data and uncertainty quantification features into a two-branch collaborative model containing a feature fusion layer, a first branch for power range prediction, and a second branch for power density prediction; outputting two types of intermediate prediction results; and finally, employing an alternating optimization strategy, introducing collaborative constraints and adaptive weight adjustment to iteratively optimize the model until convergence, outputting a power probability prediction result containing both range and density prediction results. Based on this, meteorological uncertainty can be accurately quantified hierarchically, achieving collaborative optimization of power range and density prediction, and effectively improving the coordination and accuracy of day-ahead power probability prediction results under meteorological uncertainty.

[0022] like Figure 1 As shown in the embodiments of this application, the day-ahead power probability prediction method under meteorological uncertainty includes: S101. Obtain meteorological station data.

[0023] Among them, meteorological station data is the basic data used for power probability prediction, including NWP data and target power data. NWP data is numerical weather prediction data, and target power data is actual power monitoring data of new energy power plants.

[0024] In this embodiment, data is acquired from the sensing devices of meteorological monitoring stations and target new energy power plants through a communication unit. The acquired dataset becomes meteorological station data containing NWP data and target power plant power data. The time granularity and period of data acquisition can be adjusted according to actual forecasting needs.

[0025] It should be noted that the meteorological station data obtained must ensure the validity and completeness of the data source to avoid invalid data interfering with the subsequent forecasting process.

[0026] As an example, power monitoring data is acquired from wind / solar power plants at an hourly granularity, while NWP data such as wind speed and solar irradiance for the corresponding time period are acquired simultaneously.

[0027] Based on the above steps, complete meteorological station basic data covering the forecast period and matching the target stations can be obtained.

[0028] S102. Preprocess the meteorological station data.

[0029] Preprocessing refers to the normalization of raw meteorological station data to eliminate invalid interference factors and improve data quality.

[0030] In this embodiment of the application, the collected meteorological station data undergoes multi-stage normalization processing. By aligning the data benchmark in the time dimension, missing key data is supplemented and filled in. Then, the data is processed to unify the dimensions, so that the processed data can be adapted to the input requirements of subsequent models.

[0031] It should be noted that each step of the preprocessing process must be performed in sequence, and the processing method can be flexibly adjusted according to the actual missing data and distribution characteristics.

[0032] As an example, time stamps are aligned between NWP data and power data that are out of sync, interpolation is used to fill in the missing wind speed data with a low missing rate, and then all data are mapped to the same numerical range.

[0033] Based on the above steps, we can obtain regularized and standardized meteorological station data, which provides a high-quality data foundation for subsequent quantification of meteorological uncertainties.

[0034] S103. By using probability distribution fitting and combining it with time series deviation correction, the meteorological uncertainty in meteorological station data is quantified to obtain the uncertainty quantification characteristics.

[0035] Among them, uncertainty quantification features refer to the set of features that can characterize various uncertainties in meteorological data, and can accurately reflect the degree of influence of meteorological factors on power prediction.

[0036] In this embodiment of the application, the meteorological uncertainty in the meteorological station data is first decomposed into types, and then the appropriate quantification method is used to extract features for different types of uncertainty. Finally, the various quantified uncertainty features are integrated to form a unified uncertainty quantification feature.

[0037] It should be noted that the fitting type of the probability distribution must match the distribution characteristics of the meteorological factors, and the window size for time series deviation correction can be flexibly set according to the forecast duration.

[0038] As an example, after decomposing meteorological uncertainties into two categories, probability distributions are fitted to wind speed and light intensity respectively. Then, the prediction bias is calculated through a sliding window to correct the other type of uncertainty, and finally, a feature vector that integrates the two types of uncertainties is obtained.

[0039] Based on the above steps, precise hierarchical quantification of meteorological uncertainty can be achieved, resulting in uncertainty quantification characteristics that can truly reflect the patterns of meteorological fluctuations.

[0040] S104. Input the preprocessed meteorological station data and uncertainty quantification features into the dual-branch collaborative model, and output the intermediate results of power interval prediction and power density prediction.

[0041] Among them, the dual-branch collaborative model refers to a prediction model that includes a feature fusion layer and two functional branches. The intermediate results of power interval prediction are the boundary data of the power confidence interval, and the intermediate results of power density prediction are the relevant data of the power probability distribution.

[0042] In this embodiment, the two types of input features are fused by the feature fusion layer of the dual-branch collaborative model. After obtaining the fused features, they are input into the first branch and the second branch respectively. The first branch is used to predict the power interval based on the upper and lower boundary constraints, and the second branch is used to determine the power probability distribution law to predict the power density. Each branch completes feature extraction and result output according to its own prediction logic, and finally obtains the intermediate results of the two types of power prediction simultaneously.

[0043] It should be noted that the prediction processes of the two functional branches are independent of each other, and the intermediate output results must match the power rating range of the target power station.

[0044] As an example, the preprocessed meteorological data and uncertainty quantification features are input into the model. After the fusion layer outputs the fusion features, the first branch outputs the intermediate results of power interval prediction, and the second branch outputs the intermediate results of power density prediction, forming two types of intermediate results.

[0045] Based on the above steps, intermediate prediction results for power range and density can be obtained simultaneously, achieving parallel output of the two types of predictions, which is different from the single-branch single prediction method of existing technologies.

[0046] S105. The alternating optimization strategy is used to iteratively optimize the dual-branch collaborative model, introducing collaborative constraints and adaptive weight adjustment until the dual-branch collaborative model converges, and the power probability prediction result is output.

[0047] Among them, the power probability prediction result refers to the final power prediction data that integrates interval prediction and density prediction, including the confidence interval of power (interval prediction result) and the probability distribution law (density prediction result).

[0048] In this embodiment, the dual-branch collaborative model is first pre-trained, and then the two functional branches are optimized separately through alternating iterations. During the optimization process, collaborative constraints are introduced to ensure the consistency of results between branches. At the same time, the loss weight is adaptively adjusted according to the model training state until the model meets the convergence condition. Finally, the two types of intermediate results are integrated to output the final prediction result.

[0049] It should be noted that the threshold for judging model convergence can be set according to the actual prediction accuracy requirements, and the number of iterations for alternating optimization should take into account both training efficiency and model accuracy.

[0050] Based on the above steps, the device can achieve accurate optimization of the dual-branch collaborative model, ensure the collaborative consistency of the two types of prediction results, and output high-quality power probability prediction results.

[0051] Based on the above technical solution, by stratifying meteorological uncertainty, constructing a dual-branch collaborative model, and introducing alternating optimization and collaborative constraint strategies, parallel development and collaborative optimization of power range and density prediction are achieved. This effectively solves the problem of poor synergy between the two types of prediction results in the existing technology, improves the accuracy and reliability of day-ahead power probability prediction under meteorological uncertainty, and the output prediction results are more in line with the actual operating conditions of new energy power plants, which can provide a more scientific decision-making basis for power system dispatch and trading risk control.

[0052] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above S102 can be specifically implemented through the following S201, S202 and S203, which are explained in detail below: S201. Perform time-series alignment on meteorological station data. Using the timestamp of the target power station data as a reference, align the NWP data and remove data with time-series misalignment.

[0053] Time alignment refers to unifying the timestamps of data from different sources to ensure data matching in the time dimension.

[0054] In this embodiment of the application, the time dimension of the NWP data is calibrated and matched based on the timestamp of the target power data, and misaligned data with different timestamps are identified and removed.

[0055] It should be noted that time alignment must ensure consistency in time granularity to avoid mixing data of different granularities.

[0056] As an example, the power data timestamps at the 15-minute granularity of the power station are used as a benchmark to align with the NWP wind speed and illumination data at the same granularity, and NWP data with timestamp deviations exceeding 5 minutes are removed.

[0057] Based on the above steps, meteorological station data that perfectly matches the time dimension can be obtained.

[0058] S202. Use KNN interpolation to fill in missing values ​​in the time-series aligned meteorological station data.

[0059] Among them, KNN interpolation is a numerical interpolation method based on the k-nearest neighbor algorithm, which completes the missing values ​​by using the features of neighboring data.

[0060] In this embodiment of the application, missing value detection is performed on the time-aligned meteorological station data. For the identified missing data points, the KNN interpolation method is used to calculate the feature values ​​of the adjacent valid data to complete the accurate filling of missing values.

[0061] It should be noted that the number of neighboring points in the KNN interpolation method can be adjusted according to the data distribution characteristics.

[0062] As an example, when scattered missing points were detected in the wind speed data, the 5 nearest neighbor interpolation method was used to calculate and fill in the missing values ​​using the surrounding valid wind speed data.

[0063] Based on the above steps, complete meteorological station data without missing values ​​can be obtained.

[0064] S203. Normalize the meteorological station data after filling in the missing values.

[0065] Normalization refers to mapping data with different dimensions and numerical ranges to a unified interval to eliminate dimensional interference.

[0066] In this embodiment of the application, the meteorological station data with missing values ​​filled are subjected to Min-Max standardization processing, which maps the data of each dimension to a preset numerical range to achieve the unification of data units.

[0067] Based on the above steps, standardized meteorological station data with unified dimensions and adapted to model input can be obtained.

[0068] Based on the above technical solution, through continuous preprocessing operations such as time alignment, KNN interpolation missing value filling and normalization, multiple optimizations of meteorological station data in terms of time dimension, data integrity and numerical standardization are achieved. This effectively eliminates problems such as time misalignment, data missingness and dimensional interference in the original data, resulting in high-quality standardized data and improving data adaptability and effectiveness.

[0069] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above S103 can be implemented by the following S301, S302 and S303, which are explained in detail below: S301. Decompose the meteorological uncertainty in meteorological station data into the primary uncertainty of retaining the fluctuation of the meteorological factors themselves and the secondary uncertainty of meteorological forecast error.

[0070] Primary uncertainty is the uncertainty caused by the natural fluctuations of meteorological factors themselves, while secondary uncertainty is the uncertainty of prediction error caused by the deviation between NWP data and measured data.

[0071] In this embodiment of the application, the causes of meteorological uncertainty in meteorological station data are analyzed, and the uncertainty is decomposed into two independent types of uncertainty: primary uncertainty and secondary uncertainty, according to the source of fluctuation, so as to achieve a hierarchical definition of meteorological uncertainty.

[0072] It should be noted that the decomposition process must be based on the fluctuation characteristics and sources of meteorological data to ensure that the two types of uncertainty do not overlap and are fully covered.

[0073] As an example, the random fluctuations of wind speed and light intensity are classified as primary uncertainties, while the deviation between the wind speed predicted by the NWP and the wind speed measured at the site is classified as secondary uncertainties.

[0074] Based on the above steps, a precise hierarchical decomposition of meteorological uncertainties can be completed, clarifying the sources and characteristics of different types of uncertainties.

[0075] S302. The core meteorological factors in the meteorological station data are fitted with probability distributions to quantify the primary uncertainty.

[0076] Among them, the core meteorological factors are wind speed and light intensity, which play a key role in power prediction in NWP data, and probability distribution fitting is used to characterize the fluctuation pattern of meteorological factors through a suitable distribution model.

[0077] In this embodiment of the application, wind speed and light intensity data are extracted from meteorological station data. Based on the distribution characteristics of the two core meteorological factors, appropriate probability distribution models are used for fitting. The fluctuation degree is characterized by the distribution parameters, thereby realizing the quantification of primary uncertainty.

[0078] Optionally, the probability distribution of wind speed is fitted to satisfy the following formula: in, Let be the wind speed probability density function. This is the actual wind speed value. For shape parameters, For scale parameters; The probability distribution of light intensity fits the following formula: in, Let be the probability density function of light intensity. This is the normalized light intensity value. , For the distribution shape parameter, (·) is the gamma function.

[0079] As an example, the shape and scale parameters are obtained by fitting wind speed data using the Weibull distribution, and the light intensity parameters are obtained by fitting light intensity data using the Beta distribution. , The parameters quantify the volatility uncertainty of the two types of factors.

[0080] Based on the above steps, a quantitative characterization of primary uncertainty can be achieved, and quantitative parameters of primary uncertainty corresponding to each core meteorological factor can be obtained.

[0081] S303. The meteorological forecast deviation is calculated using a sliding window. The secondary uncertainty is corrected by combining the meteorological forecast deviation with the quantified primary uncertainty to obtain the uncertainty quantification characteristics.

[0082] Among them, the sliding window refers to a time window that slides according to a preset duration, the meteorological forecast deviation is the difference between the NWP forecast value and the measured value, and the uncertainty quantification feature is a unified feature set that integrates the two types of uncertainty.

[0083] In this embodiment of the application, a sliding window of preset duration is set, and the meteorological forecast deviation within each window is calculated by sliding the time series of meteorological station data. The secondary uncertainty is corrected and quantified based on the deviation. Then, the corrected secondary uncertainty quantification result is fused with the primary uncertainty quantification parameter obtained in S302 to form a unified uncertainty quantification feature.

[0084] Optionally, the calculation of weather forecast deviation within the sliding window satisfies the following formula: in, Let be the weather forecast deviation at time t, and w be the sliding window length. Let i be the predicted NWP value at time i. Let be the measured value at time i.

[0085] It should be noted that the length of the sliding window can be flexibly adjusted according to the prediction duration and data time granularity, and feature integration must ensure that the dimensions of the two types of uncertain features are matched.

[0086] As an example, a 24-hour sliding window is set to calculate the prediction bias of wind speed and light intensity and correct the secondary uncertainty. The bias is then concatenated with the probability distribution fitting parameters of wind speed and light intensity to obtain a multidimensional uncertainty quantification feature vector.

[0087] Based on the above steps, the secondary uncertainty can be corrected and quantified, and the two types of uncertainty can be fused to obtain quantitative characteristics that can comprehensively characterize meteorological uncertainty.

[0088] Based on the above technical solution, by decomposing meteorological uncertainties in layers, quantifying them in a targeted manner, and fusing their features, we can achieve accurate and comprehensive quantification of meteorological uncertainties. Compared with the existing technology's overall single quantification method for meteorological uncertainties, this method can accurately characterize meteorological uncertainties from different sources and form a unified feature that adapts to the model input through feature integration. This provides reliable uncertainty feature support for the accurate prediction of the subsequent dual-branch collaborative model and effectively improves the model's adaptability to meteorological fluctuations.

[0089] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 4 As shown, the power range prediction based on upper and lower boundary constraints in S104 can be specifically implemented through the following S401, S402, and S403, which are explained in detail below: S401. Use the Transformer encoder to extract temporal correlation features and uncertainty correlation features from the fused features.

[0090] Among them, the Transformer encoder is a feature extraction module based on the self-attention mechanism. The temporal correlation feature is the correlation feature of the fused feature over time, and the uncertainty correlation feature is the correlation feature between meteorological uncertainty and power in the fused feature.

[0091] In this embodiment, the fused features output from the dual-branch collaborative model feature fusion layer are input to the Transformer encoder. The encoder's self-attention mechanism captures the temporal correlations of different time dimensions in the fused features, while simultaneously mining the correlation features between meteorological uncertainty and power change, thus completing the extraction and fusion of the two types of features.

[0092] Optionally, the calculation of the self-attention mechanism satisfies the following formula: in, For querying the matrix, The key matrix, For value matrices, Let be the dimension of the key matrix. It is a normalized exponential function.

[0093] It should be noted that the number of layers and attention heads in the Transformer encoder can be flexibly adjusted according to the feature dimension and prediction requirements.

[0094] As an example, a Transformer structure with 3 layers of encoders and 8 attention heads is used to process the fused features and extract the temporal correlation features of power changes with wind speed and light intensity, as well as the impact features of uncertainty on power.

[0095] Based on the above steps, the core related features in the fusion features can be accurately extracted, providing feature support for subsequent power range boundary prediction.

[0096] S402, The upper and lower boundaries of the output power prediction range through the dual-output fully connected layer.

[0097] Among them, the dual-output fully connected layer refers to a fully connected neural network layer with two output nodes, which correspond to the upper and lower boundaries of the power prediction interval, respectively.

[0098] In this embodiment, the associated features extracted by the Transformer encoder are input into a dual-output fully connected layer. Through linear transformation and activation operations of the fully connected layer, the upper and lower boundary values ​​of the power prediction at the corresponding future time are output respectively.

[0099] Optionally, the forward pass of the dual-output fully connected layer satisfies the following formula: in, This is the upper boundary value of the power. This is the lower boundary value of the power. , This is the weight matrix of the fully connected layer. , Here, x represents the bias term, and x represents the associated features extracted by the encoder.

[0100] It should be noted that the number of hidden layer neurons in a dual-output fully connected layer needs to match the dimension of the input features.

[0101] As an example, the extracted high-dimensional correlation features are input into a dual-output fully connected layer with 64 hidden layer neurons, which outputs the initial values ​​of the upper and lower boundaries of the power prediction for a certain period of time a certain day.

[0102] Based on the above steps, the initial upper and lower boundary values ​​of the power prediction range can be obtained, thus achieving preliminary prediction of the power range.

[0103] S403. Constrain the upper and lower boundaries to ensure that the lower boundary is not less than 0 and the upper boundary does not exceed the rated power of the target power station, and obtain the intermediate results of the power range prediction.

[0104] Among them, the target power station rated power is the maximum output power designed for the new energy power station, and the intermediate results of the power range prediction are the effective power range data after boundary constraints.

[0105] In this embodiment, the initial upper and lower boundary values ​​of the dual-output fully connected layer are numerically constrained. The lower boundary value less than 0 is corrected to 0, and the upper boundary value exceeding the rated power of the target power station is corrected to the rated power value, so as to obtain power range data that conforms to the actual operating conditions.

[0106] It should be noted that boundary constraints should be implemented immediately after model prediction to prevent invalid values ​​from entering the subsequent model optimization stage.

[0107] As an example, for a photovoltaic power station with a rated power of 50MW, the output lower boundary value of -2MW is corrected to 0, and the upper boundary value of 55MW is corrected to 50MW to obtain the effective power range.

[0108] Based on the above steps, intermediate power range prediction results that conform to actual operating conditions can be obtained, ensuring the physical validity of the prediction results.

[0109] Based on the above technical solution, a complete power range prediction process is formed by extracting core correlation features from the fused features using a Transformer encoder, achieving preliminary prediction of power range boundaries with a dual-output fully connected layer, and ensuring the validity of the prediction results through physical boundary constraints. Compared with existing range prediction methods, this process can more accurately capture the temporal and uncertain correlation between meteorological conditions and power, and avoids invalid prediction results through boundary constraints, effectively improving the accuracy and validity of power range prediction and providing reliable intermediate range results for subsequent power probability prediction.

[0110] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 5 As shown, the determination of the power probability distribution law and the prediction of power density in S104 above can be specifically achieved through the following S501, S502 and S503, which are explained in detail below: S501. Employ a CNN-LSTM structure to collaboratively extract local correlation features and temporal correlation features from the fused features.

[0111] Among them, the CNN-LSTM structure is a feature extraction structure that integrates convolutional neural networks and long short-term memory networks. The local correlation features are the local correlation features between meteorological factors and uncertainty in the fused features, and the temporal correlation features are the temporal correlation features of the fused features that change over time.

[0112] In this embodiment, the fused features output by the feature fusion layer are input into a CNN-LSTM structure. First, the local correlation features in the fused features are captured by the convolution operation of the CNN layer. Then, the output of the CNN layer is input into the LSTM layer. The temporal correlation features of the fused features are captured by the gating mechanism, thereby achieving the collaborative extraction of the two types of features.

[0113] It should be noted that the kernel size of the CNN layer and the hidden layer dimension of the LSTM layer can be flexibly adjusted according to the dimension of the fused features.

[0114] As an example, two CNN layers with a kernel size of 3 and one LSTM layer with a hidden layer dimension of 64 are used to process the fused features and extract the local correlation between meteorological factors and uncertainty, as well as the temporal variation features of power.

[0115] Based on the above steps, local and temporal correlation features in the fusion features can be accurately extracted collaboratively, providing core feature support for subsequent power probability distribution parameter prediction.

[0116] S502, Gaussian mixture distribution parameters corresponding to the probability of output power of the predicted layer are obtained by using distribution parameters.

[0117] Among them, the distribution parameter prediction layer is a neural network layer used to output probability distribution parameters. The Gaussian mixture distribution parameters include mean, variance and mixture weights, which are the core parameters characterizing the power probability distribution law.

[0118] In this embodiment, the collaborative features extracted by the CNN-LSTM structure are input into the distribution parameter prediction layer. Through linear transformation and activation operations within the layer, Gaussian mixture distribution parameters corresponding to the power probability at future time are output. Each parameter represents the degree of concentration, dispersion and proportion of each mixture component of the power distribution.

[0119] Optionally, the probability density function of a Gaussian mixture distribution can be calculated as follows: in, Let be the probability density function of a Gaussian mixture distribution. The mixing weight of the k-th mixing component. Let be the mean of the k-th mixture component. Let V be the variance of the k-th mixture component. Let K be the probability density function of a normal distribution, and K be the number of mixture components.

[0120] It should be noted that the output dimension of the distribution parameter prediction layer must match the number of components of the Gaussian mixture distribution, and the sum of the mixture weights must satisfy the normalization constraint.

[0121] As an example, the collaborative features are input into the distribution parameter prediction layer, which outputs the mean, variance, and mixing weights of the two mixture components, for a total of six Gaussian mixture distribution parameters.

[0122] Based on the above steps, Gaussian mixture distribution parameters that can characterize the power probability distribution can be obtained, laying the foundation for determining the power probability distribution law.

[0123] S503. Determine the power probability distribution law based on the Gaussian mixture distribution parameters to obtain intermediate results for power density prediction.

[0124] Among them, the intermediate results of power density prediction include predicted data containing power probability density curves and probability distributions of different power intervals, which can intuitively reflect the probability distribution characteristics of power at future times.

[0125] In this embodiment, based on the mean, variance, and mixed weights of the distribution parameter prediction layer output, the probability density corresponding to different day-ahead power values ​​is calculated, the power probability density curve is plotted, and the distribution probability of power in different intervals is statistically analyzed to obtain intermediate power density prediction results.

[0126] Optionally, the probability density calculation for different power values ​​satisfies the following formula: in, Let P be the probability density function corresponding to the power value P.

[0127] As an example, based on Gaussian mixture distribution parameters, the probability density corresponding to each 1MW power value in the range of 0-50MW is calculated and a curve is plotted. At the same time, the distribution probability of power in the ranges of 0-10MW, 10-30MW, and 30-50MW is determined to obtain intermediate results of power density prediction.

[0128] Based on the above steps, intermediate power density prediction results that accurately reflect the power probability distribution can be obtained, thus achieving a quantitative representation of power probability information.

[0129] Based on the above technical solution, a CNN-LSTM structure is used to accurately and collaboratively extract local and temporal correlation features of the fused features. The core Gaussian mixture distribution parameters are output by the distribution parameter prediction layer. Finally, the probability density is calculated by combining the parameters, curves are plotted, and interval distribution probabilities are determined, forming a complete power density prediction process. Compared with existing density prediction methods, this approach can more comprehensively capture the multidimensional correlations of the fused features. The Gaussian mixture distribution more closely matches the actual probability distribution characteristics of power, effectively improving the accuracy and intuitiveness of power density prediction and providing reliable intermediate density results for the subsequent integration of power probability prediction results.

[0130] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 6 As shown, the above S105 can be specifically implemented through the following S601, S602, S603 and S604, which are explained in detail below: S601. With fixed initial loss weights, the AdamW optimizer is used to pre-train the dual-branch collaborative model, which quickly converges to a local optimum.

[0131] The initial loss weights are the loss ratios preset for the first and second branches before model pre-training. The AdamW optimizer is an adaptive learning rate optimizer with weight decay. The local optimum is the parameter state where the model loss first drops to a minimum value.

[0132] In this embodiment, initial loss weights are configured for the first branch and the second branch respectively. The overall loss of the dual-branch collaborative model is used as the optimization objective. The AdamW optimizer is used to iteratively update all parameters of the model, thereby driving the model to converge quickly to a local optimum.

[0133] Optionally, the overall pre-training loss of the model satisfies the following formula: in, For the overall pre-training loss of the model, The initial loss weights for the first branch, The initial loss weights for the second branch, For the first branch loss, This is the loss for the second branch.

[0134] As an example, and All values ​​are set to 1, and the model is pre-trained using the AdamW optimizer with a learning rate of 0.001 until the overall loss of the model tends to stabilize.

[0135] Based on the above steps, the model can quickly acquire basic predictive capabilities, laying the parameter foundation for subsequent alternating optimization.

[0136] S602. Fix the parameters of the second branch and optimize the first branch by combining the cooperative constraints; fix the parameters of the first branch and optimize the second branch by combining the cooperative constraints.

[0137] Among them, the cooperative constraint is that the absolute value of the deviation between the intermediate results of the power range prediction and the intermediate results of the power density prediction shall not exceed 5% of the rated power of the target power station.

[0138] In this embodiment, all parameters of the second branch are first fixed, and the collaborative constraints are incorporated into the loss function of the first branch to construct a constrained loss function. ,by Optimize the parameters of the first branch to achieve the objective; then fix all parameters of the first branch and incorporate the same collaborative constraints into the loss function of the second branch to construct a constrained loss function. ,by Optimize the parameters of the second branch for the objective.

[0139] Alternatively, the bi-branch loss with collaborative constraints satisfies the following formula: in, The first branch is used to predict the loss. For the second branch, predict the loss. To constrain the penalty coefficient, For intermediate results of interval prediction, For intermediate results of density prediction, The rated power of the target power station.

[0140] It should be noted that the collaborative constraint judgment conditions of the two branches are completely consistent, and the penalty coefficients must be kept the same to ensure that the constraint strength is uniform. , Consistent with the basic loss implications of the first and second branches in S601, in this step... , The overall branch loss after incorporating the collaborative constraint penalty term is an optimization of the basic loss.

[0141] Based on the above steps, alternating optimization of the two branches can be achieved, and the losses of both branches can be incorporated into the collaborative constraints to ensure the consistency of the prediction results during the optimization process.

[0142] S603. After each alternation iteration, the loss weight is adaptively adjusted based on the rate of decrease of the bi-branch loss.

[0143] The bi-branch loss reduction rate is based on the condition with collaborative constraints. and The calculated loss change ratios are used to adaptively adjust the loss weights, which are then dynamically updated based on the convergence rates of the two losses. and .

[0144] In this embodiment of the application, after completing one alternating iteration of "optimizing the first branch + optimizing the second branch", the following calculations are performed respectively. and The loss decrease rate in a single iteration is used to dynamically adjust the loss weight based on the difference between the two decrease rates, and the corresponding weight is increased for the branch that converges more slowly.

[0145] Optionally, the rate of decrease of the bi-branch loss and the weight adjustment satisfy the following formula: in, for Loss reduction rate for Loss reduction rate / For the bi-branch loss before iteration, / The loss is the two-branch loss after iteration. This is the weighting adjustment factor. The preset baseline descent rate, / The adjusted weights for the bi-branch loss.

[0146] As an example, suppose =5% =0.3, if =2% =6%, then increase ,reduce This allows the training to focus more on optimizing the first branch.

[0147] Based on the above steps, dynamic adaptation of loss weights can be achieved, allowing model training resources to be tilted towards branches with slower convergence, thereby improving the overall convergence efficiency of the two branches.

[0148] S604. Determine whether the rate of decrease of the bi-branch loss is less than the preset threshold. If it is less, converge and output the power probability prediction result; if it is not less, continue iterative optimization.

[0149] Among them, the two-branch loss reduction rate is based on Calculated and based on Calculated The preset threshold is the critical value for determining model convergence, and the power probability prediction result is the day-ahead power data that integrates the final prediction results of the two branches.

[0150] In this embodiment of the application, after completing a single alternation iteration and weight adjustment, the following are determined: and If all values ​​are less than the preset threshold, the model is considered converged. The power interval prediction results of the first branch and the power density prediction results of the second branch are integrated, and the final power probability prediction result is output. If any decrease rate fails to meet the target, the process returns to S602 to continue the alternating optimization.

[0151] It should be noted that model convergence determination requires both branches to meet the threshold requirement of the rate of decrease in loss; convergence cannot be determined by a single branch.

[0152] As an example, the preset threshold is set to 0.001. If after iteration... =0.0008、 =0.0007, if all values ​​are less than the threshold, the model converges, and the final prediction result is output.

[0153] Based on the above steps, the overall convergence state of the model can be accurately determined, ensuring that the output power probability prediction results meet the accuracy requirements of the two-branch model.

[0154] Based on the above technical solution, the model is pre-trained to quickly converge to a local optimum. Then, collaborative constraints are simultaneously incorporated into the bi-branch loss to complete alternating optimization. The loss weights are dynamically adjusted based on the actual decrease rate of the bi-branch loss. Finally, the model convergence is jointly determined based on the decrease rate of the bi-branch loss. This forms a model optimization process that takes into account both convergence efficiency and result consistency. The bi-branch optimization process is controlled by collaborative constraints, and the loss weights are matched with the convergence state in real time. This effectively ensures the synergy between the power range and density prediction results, while improving the targeting and efficiency of model training. The final output power probability prediction results are more accurate and more consistent.

[0155] The foregoing mainly describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as a day-ahead power probability prediction device under meteorological uncertainty, includes at least one of the hardware structures and software modules corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0156] This application embodiment can divide the day-ahead power probability prediction device under meteorological uncertainty into functional units based on the above method example. For example, each function can be divided into its own functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0157] When using integrated units, Figure 7 A possible structural schematic diagram of the day-ahead power probability prediction device under meteorological uncertainty (hereinafter referred to as day-ahead power probability prediction device 70 under meteorological uncertainty) involved in the above embodiments is shown. The day-ahead power probability prediction device 70 under meteorological uncertainty includes a processing unit 701 and a communication unit 702, and may also include a storage unit 703. Figure 7 The schematic diagram shown can be used to illustrate the structure of the day-ahead power probability prediction device under meteorological uncertainty involved in the above embodiments.

[0158] when Figure 7The schematic diagram shown illustrates the structure of the day-ahead power probability prediction device under meteorological uncertainty involved in the above embodiments. The processing unit 701 is used to control and manage the operation of the day-ahead power probability prediction device under meteorological uncertainty, the communication unit 702 is used for the day-ahead power probability prediction device under meteorological uncertainty to communicate with other devices, and the storage unit 703 is used to store the program code and data of the day-ahead power probability prediction device under meteorological uncertainty.

[0159] For example, communication unit 702 is used to acquire meteorological station data; the meteorological station data includes NWP data and target station power data; Processing unit 701 is used to preprocess meteorological station data; it uses probability distribution fitting and combines time series deviation correction to quantify the meteorological uncertainty in the meteorological station data to obtain uncertainty quantification features; it inputs the preprocessed meteorological station data and uncertainty quantification features into a two-branch collaborative model, and outputs intermediate results of power interval prediction and power density prediction; the two-branch collaborative model includes a feature fusion layer, a first branch, and a second branch; the feature fusion layer is used to fuse the preprocessed meteorological station data and uncertainty quantification features to obtain fused features; the first branch is used to perform power interval prediction based on upper and lower boundary constraints; the second branch is used to determine the power probability distribution law to perform power density prediction; an alternating optimization strategy is used to iteratively optimize the two-branch collaborative model, introducing collaborative constraints and adaptive weight adjustment, until the two-branch collaborative model converges, and outputs power probability prediction results; the power probability prediction results include interval prediction results and density prediction results.

[0160] In one possible implementation, the processing unit 701 is further configured to perform time-series alignment on the meteorological station data, align the NWP data based on the timestamp of the target station power data, and remove data with time-series misalignment; use KNN interpolation to fill in missing values ​​in the time-series aligned meteorological station data; and perform normalization processing on the meteorological station data after missing value filling.

[0161] In one possible implementation, the processing unit 701 is further configured to decompose the meteorological uncertainty in the meteorological station data into primary uncertainty that retains the fluctuation of the meteorological factors themselves and secondary uncertainty that contains the meteorological forecast error; to fit the core meteorological factors in the meteorological station data with probability distributions to quantify the primary uncertainty; the core meteorological factors are wind speed and light intensity in the NWP data; to calculate the meteorological forecast deviation using a sliding window, to correct the secondary uncertainty by combining the meteorological forecast deviation, and to integrate the quantified primary uncertainty with the quantified uncertainty to obtain the uncertainty quantification feature.

[0162] In one possible implementation, the processing unit 701 is further configured to extract the temporal correlation features and uncertainty correlation features from the fused features using a Transformer encoder; output the upper and lower boundaries of the power prediction interval through a dual-output fully connected layer; constrain the upper and lower boundaries to ensure that the lower boundary is not less than 0 and the upper boundary does not exceed the rated power of the target power station, thereby obtaining intermediate results of the power interval prediction.

[0163] In one possible implementation, the processing unit 701 is further configured to use a CNN-LSTM structure to collaboratively extract local correlation features and temporal correlation features of the fused features. Specifically, the CNN layer captures the local correlation features of meteorological factors and uncertainties in the fused features, and the LSTM layer captures the temporal correlation features of the fused features. The distribution parameter prediction layer outputs Gaussian mixture distribution parameters corresponding to the power probability. The Gaussian mixture distribution parameters include the mean, variance, and mixture weights. The power probability distribution law is determined based on the Gaussian mixture distribution parameters to obtain intermediate results of power density prediction.

[0164] In one possible implementation, the processing unit 701 is further configured to: fix the initial loss weights; pre-train the dual-branch collaborative model using the AdamW optimizer to quickly converge to a local optimum; fix the parameters of the second branch and optimize the first branch using collaborative constraints; fix the parameters of the first branch and optimize the second branch using collaborative constraints; after each alternating iteration, adaptively adjust the loss weights based on the dual-branch loss reduction rate; determine whether the dual-branch loss reduction rate is less than a preset threshold; if it is less, convergence is achieved, and the power probability prediction result is output; if it is not less, it continues iterative optimization.

[0165] In one possible implementation, the processing unit 701 is further configured to fit the wind speed in the NWP data using a Weibull distribution and the light intensity in the NWP data using a Beta distribution.

[0166] In one possible implementation, the processing unit 701 is further configured to calculate the probability density corresponding to different day-ahead power values ​​based on the mean, variance, and mixing weight of the Gaussian mixture distribution, plot the power probability density curve, determine the distribution probability of power in different intervals, and obtain intermediate results of power density prediction.

[0167] In one possible implementation, the cooperative constraint is that the absolute value of the deviation between the intermediate results of the power range prediction and the intermediate results of the power density prediction does not exceed 5% of the rated power of the target power station.

[0168] The processing unit 701 can be a processor or a controller, and the communication unit 702 can be a communication interface, transceiver, transceiver circuit, transceiver device, etc. The term "communication interface" is a general term and may include one or more interfaces. The storage unit 703 can be a memory. When the day-ahead power probability prediction device 70 under meteorological uncertainty is a chip, the processing unit 701 can be a processor or a controller, and the communication unit 702 can be an input interface and / or an output interface, pins, or circuits, etc. The storage unit 703 can be a storage unit within the chip (e.g., a register, cache, etc.) or a storage unit located outside the chip (e.g., read-only memory (ROM), random access memory (RAM, etc.).

[0169] The communication unit can also be referred to as a transceiver unit. The antenna and control circuit with transceiver functions in the day-ahead power probability prediction device 70 under meteorological uncertainty can be considered as the communication unit 702 of the day-ahead power probability prediction device 70 under meteorological uncertainty, and the processor with processing functions can be considered as the processing unit 701 of the day-ahead power probability prediction device 70 under meteorological uncertainty. Optionally, the device in the communication unit 702 used to implement the receiving function can be considered as the communication unit, which is used to execute the receiving steps in the embodiments of this application. The communication unit can be a receiver, a receiver circuit, etc. The device in the communication unit 702 used to implement the transmitting function can be considered as the transmitting unit, which is used to execute the transmitting steps in the embodiments of this application. The transmitting unit can be a transmitter, a transmitter, a transmitting circuit, etc.

[0170] Figure 7 If the integrated units in the process are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. Storage media for storing computer software products include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0171] Figure 7 The units in the process can also be called modules; for example, a processing unit can be called a processing module.

[0172] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0173] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0174] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for predicting day-ahead power probability under meteorological uncertainty, characterized in that, include: Obtain meteorological station data; The meteorological station data includes NWP data and target station power data; The meteorological station data is preprocessed; By using probability distribution fitting and combining it with time series deviation correction, the meteorological uncertainty in the meteorological station data is quantified to obtain uncertainty quantification characteristics; The preprocessed meteorological station data and the uncertainty quantification features are input into a two-branch collaborative model, which outputs intermediate results for power interval prediction and power density prediction. The two-branch collaborative model includes a feature fusion layer, a first branch, and a second branch. The feature fusion layer is used to fuse the preprocessed meteorological station data and the uncertainty quantification features to obtain fused features. The first branch is used to predict the power range based on the upper and lower boundary constraints; the second branch is used to determine the power probability distribution law and predict the power density. The dual-branch collaborative model is iteratively optimized using an alternating optimization strategy, introducing collaborative constraints and adaptive weight adjustment until the dual-branch collaborative model converges, and outputs power probability prediction results; the power probability prediction results include interval prediction results and density prediction results.

2. The method according to claim 1, characterized in that, The preprocessing of the meteorological station data includes: The meteorological station data is time-series aligned by using the timestamp of the target power station data as a reference to align the NWP data and remove data with time-series misalignment. KNN interpolation was used to fill in missing values ​​in the time-series aligned meteorological station data. The meteorological station data after missing values ​​were filled were normalized.

3. The method according to claim 1, characterized in that, The method employs probability distribution fitting and combines it with time-series deviation correction to quantify the meteorological uncertainty in the meteorological station data, thereby obtaining uncertainty quantification features, including: Meteorological uncertainty in meteorological station data is decomposed into primary uncertainty, which retains the fluctuations of the meteorological factors themselves, and secondary uncertainty, which retains the errors in meteorological forecasts. The core meteorological factors in the meteorological station data were fitted with probability distributions to quantify the primary uncertainty; the core meteorological factors were wind speed and light intensity in the NWP data. A sliding window is used to calculate the meteorological forecast deviation, and the secondary uncertainty is corrected by combining the meteorological forecast deviation with the quantified primary uncertainty to obtain the uncertainty quantification feature.

4. The method according to claim 1, characterized in that, The power range prediction based on upper and lower boundary constraints includes: The Transformer encoder is used to extract temporal correlation features and uncertainty correlation features from the fused features; The upper and lower boundaries of the output power prediction range are determined by the dual-output fully connected layer. Constraints are imposed on the upper and lower boundaries to ensure that the lower boundary is not less than 0 and the upper boundary does not exceed the rated power of the target power station, thus obtaining intermediate results for power range prediction.

5. The method according to claim 1, characterized in that, The process of determining the power probability distribution and predicting power density includes: A CNN-LSTM structure is used to collaboratively extract local correlation features and temporal correlation features of the fused features. Specifically, the CNN layer captures the local correlation features of meteorological factors and uncertainties in the fused features, and the LSTM layer captures the temporal correlation features of the fused features. The Gaussian mixture distribution parameters corresponding to the probability of the output power of the prediction layer are obtained by using the distribution parameters; the Gaussian mixture distribution parameters include the mean, variance, and mixing weights. Based on the Gaussian mixture distribution parameters, the power probability distribution law is determined, and intermediate results for power density prediction are obtained.

6. The method according to claim 1, characterized in that, The alternating optimization strategy is used to iteratively optimize the dual-branch collaborative model, introducing collaborative constraints and adaptive weight adjustment, until the dual-branch collaborative model converges, outputting power probability prediction results, including: With fixed initial loss weights, the AdamW optimizer is used to pre-train the dual-branch collaborative model, which quickly converges to a local optimum. Fix the parameters of the second branch and optimize the first branch by combining the cooperative constraints; fix the parameters of the first branch and optimize the second branch by combining the cooperative constraints. After each alternation iteration, the loss weight is adaptively adjusted based on the rate of decrease of the bi-branch loss. Determine whether the rate of decrease of the bi-branch loss is less than a preset threshold. If it is less, convergence is achieved and the power probability prediction result is output. If it is not less, iterative optimization continues.

7. The method according to claim 3, characterized in that, The core meteorological factors in the meteorological station data are fitted with probability distributions, including: wind speed in the NWP data is fitted with a Weibull distribution, and light intensity in the NWP data is fitted with a Beta distribution.

8. The method according to claim 5, characterized in that, The step of determining the power probability distribution law based on the Gaussian mixture distribution parameters and obtaining intermediate power density prediction results includes: calculating the probability density corresponding to different daily power values ​​based on the mean, variance and mixture weight of the Gaussian mixture distribution, plotting the power probability density curve, determining the distribution probability of power in different intervals, and obtaining intermediate power density prediction results.

9. The method according to claim 6, characterized in that, The collaborative constraint is that the absolute value of the deviation between the intermediate results of the power range prediction and the intermediate results of the power density prediction does not exceed 5% of the rated power of the target power station.

10. A day-ahead power probability prediction device under meteorological uncertainty, characterized in that, The device includes: a communication unit and a processing unit; The communication unit is used to acquire meteorological station data; the meteorological station data includes NWP data and target station power data; The processing unit is used to preprocess the meteorological station data; quantify the meteorological uncertainty in the meteorological station data by using probability distribution fitting and combining it with time series deviation correction, and obtain uncertainty quantification features; input the preprocessed meteorological station data and the uncertainty quantification features into a two-branch collaborative model, and output intermediate results of power interval prediction and power density prediction; the two-branch collaborative model includes a feature fusion layer, a first branch, and a second branch; the feature fusion layer is used to fuse the preprocessed meteorological station data and the uncertainty quantification features to obtain fused features; the first branch is used to perform power interval prediction based on upper and lower boundary constraints; the second branch is used to determine the power probability distribution law to perform power density prediction; the two-branch collaborative model is iteratively optimized using an alternating optimization strategy, introducing collaborative constraints and adaptive weight adjustment, until the two-branch collaborative model converges, and outputs power probability prediction results; the power probability prediction results include interval prediction results and density prediction results.

Images