Photovoltaic power uncertainty prediction method and system

Abstract

This invention relates to a method and system for predicting photovoltaic power uncertainty. The method includes acquiring weather forecast data for the time to be predicted; inputting the weather forecast data into a preset weather forecast data uncertainty measurement model to obtain corresponding probability representations; and sampling based on these probability representations to obtain weather forecast data samples. It also involves acquiring historical sequence operation data up to the time to be predicted within a preset time period, inputting the weather forecast data samples and the historical sequence operation data into a preset deep probability prediction model to obtain the probability distribution of each predicted power generation value; and performing a consistent fusion of the probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value. This invention improves the accuracy of photovoltaic power prediction.

Description

Technical Field

[0001] This invention belongs to the technical field of photovoltaic power generation prediction, and in particular relates to a method and system for predicting photovoltaic power uncertainty. Background Technology

[0002] Currently, the proportion of photovoltaic power generation in the power system continues to increase. Dispatch operation and risk assessment place higher demands on the reliability and availability of power generation prediction results. The prediction objects are affected by factors such as rapid fluctuations in meteorological conditions, data acquisition and time alignment constraints, and statistical biases of different data sources, resulting in obvious randomness and instability in prediction results under different scenarios.

[0003] Existing photovoltaic power generation prediction methods typically build prediction models based on weather forecast data and historical operating data and output prediction results. In practical applications, weather forecast input is often treated as a fixed quantity and a single input is used to obtain a single prediction result or only a coarse-grained range expression is given. As a result, when weather forecast errors and model output fluctuations coexist, problems such as insufficient consistency in the distribution of prediction results, increased bias in risk characterization, and inconvenience for downstream decision-making can easily occur. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting photovoltaic power uncertainty, in order to solve the technical problem that existing photovoltaic power generation prediction methods cause the probability characterization of the prediction results to deviate from reality due to inconsistent processing of random fluctuations in weather forecast input and prediction output.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a method for predicting photovoltaic power uncertainty, the method comprising:

[0007] Acquire weather forecast data for each power station within a preset range, and perform spatial correlation optimization processing on the weather forecast data to obtain the weather forecast data for the target photovoltaic power station at the time to be predicted. Input the weather forecast data at the time to be predicted into a preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation, and sample the weather forecast data based on the probability representation.

[0008] Historical sequence operation data up to the time to be predicted is obtained according to a preset time period. The weather forecast data sample and the historical sequence operation data are input into a preset deep probability prediction model to obtain the probability distribution of each predicted power generation value.

[0009] At the distribution level, the probability distributions are uniformly fused to obtain the final probability distribution of the predicted power generation value. The deep probabilistic prediction model is trained as follows:

[0010] Acquire historical operational data and corresponding weather forecast data to generate model training samples;

[0011] Time series features are extracted from the model training samples to obtain corresponding output features. Based on an approximate Gaussian process structure, feature randomization is performed on the output features to obtain probability distribution parameters. Then, based on the probability distribution parameters, the probability distribution of the predicted power generation value is determined. The corresponding loss function is calculated according to the probability distribution and the model training samples, and the model parameters are optimized and updated through the loss function to obtain the deep probability prediction model.

[0012] By employing the above technical solution, weather forecast data at the time to be predicted is acquired and input into a weather forecast data uncertainty measurement model to obtain a probabilistic representation. Based on this, weather forecast data samples are obtained. This allows for the explicit quantification of the mapping uncertainty from weather forecast to actual weather in probabilistic form and its transformation into multi-meteorological scenario inputs, thereby reducing the sensitivity of the prediction results to single weather forecast errors. Furthermore, by acquiring historical sequence operational data up to the time to be predicted and inputting the weather forecast data samples and historical sequence operational data into a deep probabilistic prediction model, the probability distribution of each predicted power generation value is obtained. This enables the output of corresponding power probability distribution results under multiple scenario inputs, thereby improving the handling of fluctuations and randomness. This enhances the model's ability to characterize data and improves the interpretability of results. By unifying and fusing various probability distributions at the distribution level to obtain the final probability distribution, it is possible to unify and fuse input uncertainty propagation and output uncertainty at the distribution level. This avoids distribution distortion caused by simply weighting multiple point predictions and facilitates direct reading of quantiles for downstream applications. By using model training methods to extract features from model training samples and perform random feature calculations on output features to obtain probability distribution parameters, and then calculating the loss function to update model parameters, the probability prediction model can learn temporal feature representations and probability distribution expressions simultaneously during the training phase. This enables end-to-end training and improves prediction accuracy and generalization ability.

[0013] In one example, the present invention can be further configured as follows: inputting the weather forecast data into a preset weather forecast data uncertainty measurement model to obtain a corresponding probability representation, and sampling according to the probability representation to obtain a weather forecast data sample, including:

[0014] Acquire weather forecast data for the target photovoltaic power station and surrounding stations within a preset radius of the target photovoltaic power station, generate node features corresponding to each station, and generate spatial correlation features based on the geospatial information between each station;

[0015] A multi-source weather forecast spatial map model is constructed based on the node features and the spatial correlation features. The multi-source weather forecast spatial map model is then input into a preset S-GTN network to generate weather forecast data for the time to be predicted.

[0016] The weather forecast data at the time to be predicted is input into the weather forecast data uncertainty measurement model to obtain several quantiles used to characterize the uncertainty of the weather forecast data;

[0017] Based on several quantiles, the inverse function of the approximate cumulative distribution function corresponding to the probability representation is constructed using linear interpolation.

[0018] Random numbers are drawn from the standard uniform distribution based on the inverse function of the approximate cumulative distribution function, and the random numbers are mapped to obtain the weather forecast data sample.

[0019] By adopting the above technical solution, weather forecast data of the target photovoltaic power station and surrounding stations within a preset radius of the target photovoltaic power station are acquired. Based on the geospatial information between the stations, spatial correlation features are generated, a multi-source weather forecast spatial map model is constructed, and the model is input into a preset S-GTN network to generate weather forecast data for the time to be predicted. Spatial correlation can be used to collaboratively correct the weather forecast information of the target station, thereby reducing the impact of single-station forecast bias on subsequent uncertainty modeling and improving the reliability of the input data. By inputting the weather forecast data into the uncertainty measurement model to obtain several quantiles, and constructing the inverse function of the approximate cumulative distribution function based on the linear interpolation method, random numbers are extracted from the standard uniform distribution according to the inverse function of the approximate cumulative distribution function. The random numbers are then used as independent variables to be input into the inverse function of the approximate cumulative distribution function for inverse mapping to obtain weather forecast data samples. This can convert discrete quantile information into sampleable probability representations and generate diverse scene samples, thereby improving the continuity and stability of the sampling results and enhancing the reliability of subsequent probability prediction.

[0020] In one example, the present invention can be further configured as follows: The step of drawing random numbers from a standard uniform distribution based on the inverse function of the approximate cumulative distribution function, and mapping the random numbers to obtain the weather forecast data sample, includes:

[0021] The quantiles are sorted according to the magnitude of the quantile level values ​​to obtain an ordered quantile sequence;

[0022] Random numbers are drawn from a standard uniform distribution. The random numbers are compared with the magnitudes of each quantile level to determine the adjacent quantile levels. The corresponding adjacent quantiles are then obtained based on the adjacent quantile levels.

[0023] The weather forecast data sample is obtained by calculating using the inverse function of the approximate cumulative distribution function based on the adjacent quantile level, the adjacent quantile, and the random number.

[0024] By adopting the above technical solution, an ordered quantile sequence is obtained by sorting several quantiles and adjacent quantile level intervals and quantile intervals are determined based on random numbers. Then, weather forecast data samples are calculated based on adjacent quantile level intervals, quantile intervals and random numbers. This can ensure that the sampling mapping process is monotonically consistent within the quantile intervals and avoid abnormal sample landing points, thereby improving the controllability of the sampling process and enhancing the distribution consistency of the generated weather forecast data samples.

[0025] In one example, the present invention can be further configured as follows: the construction of the weather forecast data uncertainty measurement model includes:

[0026] Acquire historical weather forecast data and corresponding actual weather data within a preset radius of the photovoltaic site, and generate uncertainty measurement training samples;

[0027] A preset number of quantile regression models are trained based on the aforementioned uncertainty measure training samples;

[0028] The multiple quantile regression models are combined to generate the weather forecast data uncertainty measurement model.

[0029] By adopting the above technical solution, training samples are generated by acquiring historical weather forecast data and corresponding actual weather data within a preset radius of the photovoltaic site, and then training multiple quantile regression models and combining them to generate a weather forecast data uncertainty measurement model. This enables the uncertainty measurement model to learn the mapping rules at different quantile levels based on historical statistical relationships, thereby improving the ability to characterize different risk levels and providing a more robust probability basis for subsequent sampling.

[0030] In one example, the present invention can be further configured as follows: The step of obtaining historical sequence operational data up to the time to be predicted based on a preset time period, and inputting the weather forecast data sample and the historical sequence operational data into a preset deep probability prediction model to obtain the probability distribution of each predicted power generation value includes:

[0031] Based on the time to be predicted, determine the historical start time and historical end time corresponding to the preset time period to obtain the corresponding time range;

[0032] Within the stated time range, power generation data and actual irradiance data are acquired and time-aligned according to a unified time reference to obtain the historical sequence operation data.

[0033] The historical sequence data is correlated with the weather forecast data sample to obtain model input data. This model input data is then fed into the deep probabilistic prediction model to generate the output mean and output variance as probability distribution parameters. Based on these parameters, a probability distribution for the predicted power generation value is generated.

[0034] By adopting the above technical solution, a time range is obtained by determining a preset time period based on the time to be predicted. Within this time range, power generation data and actual irradiance data are acquired and time-aligned to form historical sequence operation data. This ensures that the input data is continuous and consistent under a unified time reference and reduces errors caused by time misalignment, thereby improving the quality of model input and enhancing prediction stability. By associating historical sequence operation data with weather forecast data samples to obtain model input data and inputting it into a deep probabilistic prediction model to generate output features, a learnable joint representation of historical operation status and meteorological inputs from multiple scenarios can be formed, thereby enhancing the model's comprehensive ability to express multi-source influencing factors. By determining the output mean and output variance as probability distribution parameters based on the output features and generating a probability distribution of predicted power generation values, the prediction results can be output in the form of distribution parameters while retaining uncertainty information. This facilitates the direct acquisition of quantile results and supports risk assessment and scheduling decisions.

[0035] In one example, the present invention can be further configured as follows: the step of uniformly fusing the probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value includes:

[0036] Extract the corresponding probability distribution parameters from each of the probability distributions to obtain multiple sets of output mean and output variance;

[0037] The fusion mean and fusion variance are calculated based on the multiple sets of output means and output variances;

[0038] The final probability distribution of the predicted power generation value is generated based on the fusion mean and the fusion variance.

[0039] By adopting the above technical solution, multiple sets of output mean and output variance are obtained by extracting probability distribution parameters from each probability distribution and calculating the fusion mean and fusion variance. Then, the final probability distribution is generated based on the fusion mean and fusion variance. This can unify and aggregate the distribution information corresponding to multiple scenarios at the parameter level and maintain the consistency of distribution expression, thereby improving the robustness of the final probability distribution and reducing the impact of single sample distribution bias on the output.

[0040] In one example, the present invention can be further configured as follows: extracting time-series features from the model training samples to obtain corresponding output features; performing random feature calculation on the output features to obtain probability distribution parameters; then determining the probability distribution of the predicted power generation value based on the probability distribution parameters; and calculating the corresponding loss function based on the probability distribution and the model training samples, including:

[0041] Based on the Informer-De model, time series features are extracted from the training samples of the model to obtain the corresponding output features;

[0042] Random features are obtained by performing random feature calculation on the output features based on an approximate Gaussian process.

[0043] Based on the random features, the corresponding output mean and output variance are calculated respectively, and used as the probability distribution parameters;

[0044] Based on the probability distribution parameters, a probability distribution for the predicted power generation value is constructed. The difference between the power generation data in the model training samples and the output mean is calculated. The difference is then scaled by combining the output variance to obtain the loss function.

[0045] By employing the above technical solution, random features are obtained by performing random feature calculation on the output features based on an approximate Gaussian process. This allows deep output features to be mapped to a random feature space for probabilistic modeling while maintaining computational scalability, thereby improving the efficiency and stability of probability parameter calculation. By calculating the output mean and output variance as probability distribution parameters based on the random features, both central tendency and dispersion can be represented simultaneously in the same computational chain, thus improving the completeness of uncertainty characterization and facilitating the output distribution results. By constructing a probability distribution based on the probability distribution parameters and calculating the difference between the power generation data of the training samples and the output mean, and then combining the output variance for scaling adjustment to obtain the loss function, a consistent measurement of errors at different uncertainty levels can be performed during the training phase, thereby improving the targeting of model optimization and enhancing the credibility of prediction results.

[0046] In one example, the present invention can be further configured as follows: the step of optimizing and updating the model parameters through the loss function to obtain the depth probability prediction model includes:

[0047] The update amount of the model parameters is determined based on the loss function, and the model parameters are updated to obtain the updated model parameters.

[0048] Obtain the L2 norm of the matrix corresponding to the updated model parameters;

[0049] Based on the matrix L2 norm, spectral normalization is performed on the updated model parameters to obtain spectrally normalized model parameters, which are then used as model parameters for subsequent iterations.

[0050] By employing the above technical solution, and determining the update amount of model parameters based on the loss function and performing parameter updates on the model parameters, the model parameters can be iteratively converged along the direction of reducing prediction error, thereby improving the fitting ability of the deep probabilistic prediction model to the probability distribution. By obtaining the L2 norm of the matrix corresponding to the updated model parameters and performing spectral normalization on the updated model parameters, the parameter scale can be constrained and drastic fluctuations in parameter updates can be suppressed, thereby improving the stability of the training process and the robustness of the model to perturbations and reducing the risk of overfitting. By using the spectrally normalized model parameters as the model parameters for subsequent iterations, the consistency of parameter constraints can be maintained in continuous iterations, thereby improving the controllability of model training and the stability of the final prediction results.

[0051] In one example, the present invention can be further configured such that the photovoltaic power uncertainty prediction method further includes:

[0052] During the inference test of the deep probabilistic prediction model, the GCRPS index is calculated based on the probability distribution of the predicted power generation value and the corresponding actual power generation value.

[0053] If the GCRPS index is greater than a preset threshold, the input data of the deep probability prediction model corresponding to the GCRPS index and the corresponding actual power generation data are saved as retraining samples.

[0054] If the retrained samples meet the preset triggering conditions, the deep probability prediction model is retrained and optimized based on the retrained samples.

[0055] By employing the above technical solution, the GCRPS index is calculated based on the probability distribution of the predicted power generation value and the corresponding actual power generation value during the inference test of the deep probabilistic prediction model. This allows for a quantitative assessment of the consistency between the predicted and actual results at the distribution level, thereby promptly identifying prediction biases and distortions in the characterization of uncertainties. Furthermore, by saving the corresponding model input data and actual power generation data as retraining samples when the GCRPS index exceeds a preset threshold, biased samples can be transformed into usable improvement data, providing targeted sample support for subsequent model updates. Finally, by retraining and optimizing the deep probabilistic prediction model based on the retraining samples when preset trigger conditions are met, the model parameters can be adapted to new operating conditions or changes in data distribution, thereby improving the prediction accuracy and long-term stability of the model during continuous operation.

[0056] In a second aspect, the present invention provides a photovoltaic power uncertainty prediction system, the system comprising:

[0057] The uncertainty measurement module is used to acquire weather forecast data of each power station within a preset range, and perform spatial correlation optimization processing on the weather forecast data to obtain the weather forecast data of the target photovoltaic power station at the time to be predicted. The weather forecast data at the time to be predicted is input into a preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation, and the weather forecast data sample is obtained by sampling according to the probability representation.

[0058] The probability prediction module is used to obtain historical sequence operation data up to the time to be predicted according to a preset time period, input the weather forecast data sample and the historical sequence operation data into a preset deep probability prediction model, and obtain the probability distribution of each predicted power generation value.

[0059] The distribution fusion module is used to perform consistent fusion of the various probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value. The deep probability prediction model is trained as follows:

[0060] The sample construction module is used to acquire historical operational data and corresponding weather forecast data to generate model training samples;

[0061] The random feature modeling module is used to extract time series features from the model training samples to obtain corresponding output features, perform random feature calculations on the output features to obtain probability distribution parameters, determine the probability distribution of the predicted power generation value based on the probability distribution parameters, calculate the corresponding loss function according to the probability distribution and the model training samples, and optimize and update the model parameters through the loss function to obtain the deep probability prediction model.

[0062] By employing the above technical solution, weather forecast data at the time to be predicted is acquired and input into a weather forecast data uncertainty measurement model to obtain a probabilistic representation. Based on this, weather forecast data samples are obtained. This allows for the explicit quantification of the mapping uncertainty from weather forecast to actual weather in probabilistic form and its transformation into multi-meteorological scenario inputs, thereby reducing the sensitivity of the prediction results to single weather forecast errors. Furthermore, by acquiring historical sequence operational data up to the time to be predicted and inputting the weather forecast data samples and historical sequence operational data into a deep probabilistic prediction model, the probability distribution of each predicted power generation value is obtained. This enables the output of corresponding power probability distribution results under multiple scenario inputs, thereby improving the handling of fluctuations and randomness. This enhances the model's ability to characterize data and improves the interpretability of results. By unifying and fusing various probability distributions at the distribution level to obtain the final probability distribution, it is possible to unify and fuse input uncertainty propagation and output uncertainty at the distribution level. This avoids distribution distortion caused by simply weighting multiple point predictions and facilitates direct reading of quantiles for downstream applications. By using model training methods to extract features from model training samples and perform random feature calculations on output features to obtain probability distribution parameters, and then calculating the loss function to update model parameters, the probability prediction model can learn temporal feature representations and probability distribution expressions simultaneously during the training phase. This enables end-to-end training and improves prediction accuracy and generalization ability. Attached Figure Description

[0063] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0064] Figure 1 This is a flowchart of a photovoltaic power uncertainty prediction method in an embodiment of the present invention;

[0065] Figure 2 This is a structural block diagram of the photovoltaic power uncertainty prediction system according to an embodiment of the present invention. Detailed Implementation

[0066] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0067] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0068] Example 1

[0069] like Figure 1 As shown, this invention discloses a method for predicting photovoltaic power uncertainty, which specifically includes the following steps:

[0070] S10: Obtain weather forecast data for each power station within a preset range, and perform spatial correlation optimization processing on the weather forecast data to obtain the weather forecast data for the target photovoltaic power station at the time to be predicted. Input the weather forecast data at the time to be predicted into the preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation, and sample the weather forecast data based on the probability representation.

[0071] Specifically, weather forecast data corresponding to the time to be predicted is obtained from the meteorological department as input variables, and time alignment is completed under the time benchmark consistent with the historical actual weather data. The aligned weather forecast data is sent into the preset weather forecast data uncertainty measurement model to output a probability representation used to characterize the uncertainty of the mapping from weather forecast to actual weather. Then, based on the probability representation, multiple weather forecast data samples with different values ​​are generated at the same time to be predicted for subsequent probability prediction calculations.

[0072] S20: Obtain historical sequence operation data up to the time to be predicted based on the preset time period, input the weather forecast data sample and the historical sequence operation data into the preset deep probability prediction model, and obtain the probability distribution of each predicted power generation value.

[0073] Specifically, historical power generation data of the photovoltaic power station up to the time to be predicted is obtained through a power meter, and historical meteorological-related operational data under the same time benchmark as the historical power generation data are obtained to form historical sequence operational data. Each weather forecast data sample and the historical sequence operational data are combined into the model input according to a preset data organization method, so that the deep probability prediction model outputs the corresponding probability distribution set of predicted power generation value for each weather forecast data sample for subsequent fusion.

[0074] S30: At the distribution level, the probability distributions are fused to obtain the final probability distribution of the predicted power generation value.

[0075] Specifically, the probability distributions of multiple predicted power generation values ​​obtained from different weather forecast data samples are unified and fused in the same distribution expression space, so that the fusion result is directly output as the final probability distribution that can be used for quantile reading and risk assessment, and the final probability distribution is output as the prediction result for the time to be predicted.

[0076] S201: Obtain historical operational data and corresponding weather forecast data to generate model training samples.

[0077] Specifically, historical power generation data of photovoltaic power plants are obtained through power meters. Historical weather forecast data and historical actual weather data corresponding to the historical power generation data are obtained from meteorological departments and time-aligned. The samples are divided into training set and test set at a ratio of 8:2. The power generation data, weather forecast data and actual weather data in the training set are standardized using Z-Score and satisfy Z=(x-μ) / σ, where μ and σ represent the mean and standard deviation of the corresponding data in the training set, respectively. The month, date, hour and minute in the time data are standardized using maximum value and satisfy Z=x / max, where max=12 for month data.

[0078] S202: Extract time series features from the model training samples to obtain the corresponding output features. Perform random feature calculation on the output features based on the approximate Gaussian process structure to obtain the probability distribution parameters. Then, determine the probability distribution of the predicted power generation value based on the probability distribution parameters. Calculate the corresponding loss function based on the probability distribution and the model training samples, and optimize and update the model parameters through the loss function to obtain the deep probabilistic prediction model.

[0079] Specifically, during the training process, historical operational data and corresponding weather forecast data are organized into model training samples and input into a time series feature extraction structure to obtain output features for probabilistic modeling. The output features are then input into a feature randomization structure to obtain probability distribution parameters. Based on the probability distribution parameters, a probability distribution for the predicted power generation value is constructed. The probability distribution and the actual power generation data in the model training samples are used together to calculate the loss function. The loss function is then used to back-optimize and update the parameters of the deep probabilistic prediction model to complete end-to-end training and obtain a deep probabilistic prediction model that can be used in the prediction stage.

[0080] In one embodiment, in step S10, spatial correlation optimization processing is performed on the weather forecast data to obtain the weather forecast data for the target photovoltaic power station at the predicted time. The weather forecast data for the predicted time is input into a preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation. Based on the probability representation, weather forecast data samples are obtained, including:

[0081] S11: Obtain weather forecast data for the target photovoltaic power station and surrounding stations within a preset radius of the target photovoltaic power station, generate node features corresponding to each station, and generate spatial correlation features based on the geospatial information between each station.

[0082] Specifically, by determining the site location identifier of the target photovoltaic power station and retrieving the set of surrounding sites within its preset radius, weather forecast data of the target photovoltaic power station and each surrounding site are obtained at a unified time to be predicted. The weather forecast data of each site at the time to be predicted is organized into a node feature vector corresponding to that site and a node feature set is formed according to the site index. At the same time, the latitude, longitude and altitude information of each site are read and the geospatial information between each pair of sites is calculated accordingly. The geospatial information includes the geographical distance between sites, relative orientation and altitude difference. The geospatial information is converted into edge feature vectors and a spatial association feature set is constructed according to the site pairs, so that the node feature set and the spatial association feature set correspond under the same site index system, so as to be used for subsequent construction of a multi-source weather forecast spatial map model.

[0083] S12: Construct a multi-source weather forecast spatial map model based on node characteristics and spatial correlation characteristics, input the multi-source weather forecast spatial map model into the preset S-GTN network, and generate weather forecast data for the time to be predicted.

[0084] Specifically, a multi-source weather forecast spatial graph model is constructed based on the target photovoltaic power station and surrounding site sets, including node sets and edge sets. The node feature vectors corresponding to each site are used as node attributes of the graph model and denoted as node features h. a And the node feature vector corresponding to node b, which is adjacent to node a, is denoted as node feature h. b The spatial association feature vectors between any two sites are used as edge attributes in the graph model and denoted as edge features f. ab Adjacency relationships are established to obtain a graph structure input consistent with the spatial topology of the sites. Subsequently, the multi-source weather forecast spatial graph model is input into the pre-defined S-GTN network, and the node features are first weighted based on edge features. Specifically, for each edge, the features of the nodes at both ends and the edge features are input together into a multilayer perceptron to obtain the neighborhood weight information q. ab ,satisfy And for each node, a set of neighborhood weight information adjacent to it is aggregated to form f. a , will f a The input to the multilayer perceptron yields the fused neighborhood weights, which are then multiplied by the node features to obtain the weighted node features. , , will each node Stacking the matrix H yields a matrix that retains only node features, while preserving site connectivity results in an adjacency mask matrix L. H and L are then input into the graph self-attention layer to obtain weighted node graph features. ,according to The attention aggregation result is calculated, where, , For querying the matrix, The key matrix, For value matrices, for Dimensions The matrix dot product is represented, and the node graph feature F is obtained by connecting it to the residual through a multilayer perceptron. H , Finally, F H The input multilayer perceptron performs output mapping to generate weather forecast data for the time to be predicted, so that the generated weather forecast data for the time to be predicted corresponds to the target photovoltaic power station and serves as input data for the subsequent weather forecast data uncertainty measurement model.

[0085] S13: Input the weather forecast data for the time to be predicted into the weather forecast data uncertainty measurement model to obtain several quantiles used to characterize the uncertainty of the weather forecast data.

[0086] Specifically, the weather forecast data for the time to be predicted is used as the independent variable x. i Input weather forecast data uncertainty measurement model, and use the conditional quantiles output by the uncertainty measurement model at different quantile levels τ as discrete representations of probability, such that each quantile satisfies , where y i For actual weather data, x i For weather forecast data, beta τ To optimize the objective parameters, τ is the quantile level and Q τ (y i |x i ) means that given x i time y i The conditional quantiles are obtained, thus yielding a set of quantiles used to characterize uncertainty.

[0087] S14: Based on several quantiles, construct the inverse function of the approximate cumulative distribution function corresponding to the probability representation using linear interpolation.

[0088] Specifically, several quantiles and their corresponding quantile levels τ are considered as discrete sampling points on the cumulative distribution function, such that each quantile corresponds to a cumulative probability τ, forming a quantile sequence that increases in τ. A linear interpolation relationship is introduced between adjacent quantiles to obtain an approximate inverse function of the cumulative distribution function that can be queried for any cumulative probability value. ,satisfy Where x is a function variable, and They represent the first And the lth quantile level.

[0089] S15: Based on the inverse function of the approximate cumulative distribution function, random numbers are drawn from the standard uniform distribution and mapped by the inverse function to obtain weather forecast data samples.

[0090] Specifically, n random numbers u are drawn from a standard uniform distribution. j As a sampling driving quantity, it locates each u j Adjacent quantile levels and adjacent quantiles, and each u j Substituting the corresponding adjacent quantile levels and adjacent quantiles into the inverse function of the approximate cumulative distribution function, a set of weather forecast data samples corresponding to the weather forecast data at the time to be predicted is obtained for subsequent power probability prediction.

[0091] In one embodiment, step S13, namely, drawing random numbers from a standard uniform distribution based on the inverse function of the approximate cumulative distribution function and mapping the random numbers using the inverse function to obtain weather forecast data samples, includes:

[0092] S131: Sort several quantiles according to the size of the quantile level values ​​corresponding to the quantiles to obtain an ordered quantile sequence.

[0093] Specifically, the quantiles output by the uncertainty measurement model are sorted in ascending order according to their quantile level values, while maintaining the correspondence with the quantile level τ, to obtain an ordered quantile sequence. This ordered quantile sequence covers the boundary of the sampling interval and provides endpoint values ​​of adjacent intervals for random number positioning.

[0094] S132: Draw random numbers from a standard uniform distribution, compare the random numbers with the magnitudes of each quantile level to determine adjacent quantile levels, and obtain the corresponding adjacent quantiles based on the adjacent quantile levels.

[0095] Specifically, random numbers u are generated from a standard uniform distribution. j and u j Consider the cumulative probability coordinates as being located on the quantile axis, and search for those that satisfy... Adjacent quantile levels to determine u j The boundary of the quantile interval falls within the range, and the data is read from τ. m and τ m+1 The corresponding adjacent quantile Q τm With Q τm+1 .

[0096] S133: Based on adjacent quantile levels, adjacent quantiles, and random numbers, the inverse function of the approximate cumulative distribution function is used to calculate and obtain weather forecast data samples.

[0097] Specifically, based on u j Linear interpolation is performed at the relative positions within adjacent quantile intervals to obtain the sampling results. and satisfy This allows us to obtain a sample of weather forecast data corresponding to the weather forecast data at the time to be predicted.

[0098] In one embodiment, step S10, namely the construction of the weather forecast data uncertainty measurement model, includes:

[0099] S101: Obtain historical weather forecast data and corresponding actual weather data within a preset radius of the photovoltaic site, and generate uncertainty measurement training samples.

[0100] Specifically, the location identifiers of photovoltaic sites are determined and a set of sites within a preset radius is retrieved. Historical weather forecast data is obtained from the meteorological department, and actual weather data at the same time is simultaneously acquired. The historical weather forecast data is used as the input variable, and the actual weather data at the corresponding time is used as the output variable. After alignment according to a unified time base, (x i ,y i Pair sample sets to construct training samples for uncertainty measurement, and perform normalization on the training samples consistent with the training set when necessary to ensure the numerical stability of model training.

[0101] S102: Train a quantile regression model with a preset number of training samples based on uncertainty measures.

[0102] Specifically, quantile levels τ were set to 0.05, 0.25, 0.5, 0.75, and 0.95, respectively, and corresponding quantile regression models were constructed for each τ, so that each model... The output form is then optimized using stochastic gradient descent. τ Optimization target satisfies This allows us to obtain multiple quantile regression model parameters for different τ values.

[0103] S103: Combine multiple quantile regression models to generate a weather forecast data uncertainty measurement model.

[0104] Specifically, multiple quantile regression models trained at different quantile levels τ are combined using a unified input interface, enabling them to function correctly when receiving any weather forecast data x. i It can output multiple conditional quantiles in parallel. Multiple conditional quantiles are used as components of the probability representation to form a weather forecast data uncertainty measurement model for subsequent sampling.

[0105] In one embodiment, step S20 involves obtaining historical sequence operation data up to the time to be predicted based on a preset time period, inputting weather forecast data samples and historical sequence operation data into a preset deep probability prediction model to obtain the probability distribution of each predicted power generation value, including:

[0106] S21: Based on the time to be predicted, determine the historical start time and historical end time corresponding to the preset time period to obtain the corresponding time range.

[0107] Specifically, the time to be predicted is used as the historical cutoff time, and the historical start time is determined by tracing back based on the preset time period length. The historical start time and the historical cutoff time together limit the continuous time range used to construct the historical sequence running data. At the same time, the time range is matched with the data sampling interval to ensure that subsequent time alignment can cover the complete historical sequence segment.

[0108] S22: Acquire power generation data and actual irradiance data within a time range, and align them according to a unified time base to obtain historical sequence operation data.

[0109] Specifically, power generation data and actual irradiance data are acquired at sampling intervals within a time range, and time alignment of the two types of data is performed based on a unified timestamp to form a synchronization sequence. The aligned historical power generation data and historical actual irradiance data are combined into historical sequence running data, and when necessary, the historical sequence running data is numerically normalized using the Z-Score normalization relationship Z=(x-μ) / σ consistent with the training set to maintain consistency with the input distribution during the model training phase.

[0110] S23: Correlate historical sequence operation data with weather forecast data samples to obtain model input data, and then input the model input data into the deep probabilistic prediction model to generate output mean and output variance as probability distribution parameters. Based on the probability distribution parameters, generate the probability distribution of predicted power generation value.

[0111] Specifically, the historical power generation data s from the historical sequence operation data i Compared with historical weather data y i and the corresponding historical time data t i Input feature embedding layer and satisfy , , , will F e Input feature encoding layer and satisfy and ,in And W q For query matrix, W k For the key matrix, W v Given a value matrix, the first feature encoding layer outputs the feature F. s Input to the second feature encoding layer to obtain the output feature F of the second feature encoding layer g At the same time, weather forecast data samples With corresponding future time data Input decoding feature embedding layer and satisfy , , , will F d The input feature decoding layer produces the output feature F. h Then F g With F h Input feature cross-fusion layer and satisfy and ,in and For querying the matrix, The key matrix, The value matrix is ​​used to obtain the output feature F, which serves as the output of the depth probability prediction model. o The output feature F o The input is an approximate Gaussian process structure, and the output layer is approximated using random features, such that the random features satisfy... L o For output feature F o Feature dimensions, W L For a fixed weight parameter whose values ​​are sampled from a standard normal distribution, b L For fixed weight parameters and their values ​​from a uniform distribution After sampling and obtaining the random feature R, the mean and variance of the output distribution are calculated and satisfied. and A and G t As trainable model parameters, M is used as the output mean and V as the output variance to form probability distribution parameters, and a probability distribution N(M,V) for predicting power generation values ​​is generated accordingly.

[0112] In one embodiment, step S30, namely, performing uniform fusion of the probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value, includes:

[0113] S31: Extract the corresponding probability distribution parameters from each probability distribution to obtain multiple sets of output mean and output variance.

[0114] Specifically, the multiple prediction probability distributions obtained from different weather forecast data samples are uniformly represented as... The form is used to read the output mean from each probability distribution. With output variance This is to form multiple sets of output mean and output variance for subsequent fusion calculations.

[0115] S32: Calculate the fusion mean and fusion variance based on the multiple output means and output variances.

[0116] Specifically, the averages of multiple outputs are summed to obtain a fused mean that satisfies... Simultaneously, the quadratic terms of the variances and corresponding means of multiple outputs are summarized to obtain the fusion variance and satisfy the following conditions: This ensures that the fusion results maintain a consistent representation of mean and variance at the distribution level.

[0117] S33: Generate the final probability distribution of the predicted power generation value based on the fusion mean and fusion variance.

[0118] Specifically, the fused mean M o With fusion variance V o The final probability distribution parameter is used to construct the final probability distribution N(M) of the predicted power generation value. o V o The final probability distribution is then used as the fusion output for subsequent quantile readings and risk assessments.

[0119] In one embodiment, in step S202, time-series features are extracted from the model training samples to obtain corresponding output features. Random feature calculation is performed on the output features to obtain probability distribution parameters. Then, the probability distribution of the predicted power generation value is determined based on the probability distribution parameters. The corresponding loss function is calculated based on the probability distribution and the model training samples, including:

[0120] S2021: Based on the Informer-De model, time series features are extracted from the model training samples to obtain the corresponding output features.

[0121] Specifically, the historical power generation data s from the historical sequence operation data i Compared with historical weather data y i and the corresponding historical time data t i Input feature embedding layer and satisfy , , , will F e Input feature encoding layer and satisfy and ,in And W q For query matrix, W k For the key matrix, W v Given a value matrix, the first feature encoding layer outputs the feature F. s Input to the second feature encoding layer to obtain the output feature F of the second feature encoding layer g At the same time, weather forecast data samples With corresponding future time data Input decoding feature embedding layer and satisfy , , , will F d The input feature decoding layer produces the output feature F.h Then F g With F h Input feature cross-fusion layer and satisfy and ,in and For querying the matrix, The key matrix, The value matrix is ​​used to obtain the output feature F, which serves as the output of the depth probability prediction model. o .

[0122] S2022: Random features are obtained by performing random feature calculation on the output features based on an approximate Gaussian process.

[0123] Specifically, the output feature F obtained by extracting time series features o The input is an approximate Gaussian process structure, and the output layer is approximated using a random feature mapping method, such that the random features satisfy... L o For output feature F o Feature dimensions, W L For a fixed weight parameter whose values ​​are sampled from a standard normal distribution, b L With fixed weight parameters and their values ​​sampled from a uniform distribution U(0,2π), this formula is used to obtain the random feature R used for mean and variance calculation.

[0124] S2023: Calculate the corresponding output mean and output variance based on the random features, as probability distribution parameters.

[0125] Specifically, the mean and variance of the output distribution are calculated based on the random feature R and satisfy the following conditions: and A and G t As trainable model parameters, M is used as the output mean and V as the output variance to obtain a set of probability distribution parameters used to construct the probability distribution of predicted power generation values.

[0126] S2024: Construct a probability distribution for the predicted power generation value based on probability distribution parameters, calculate the difference between the power generation data in the model training samples and the output mean, and scale the difference using the output variance to obtain the loss function. Specifically, construct a probability distribution N(M,V) for the predicted power generation value based on the output mean M and the output variance V. Calculate the difference between the actual power generation data in the model training samples and the predicted power generation mean corresponding to the output mean, and introduce the output variance to scale the difference. The loss function satisfies... Where n is the total number of predicted values ​​in a batch, m j Let v be the predicted mean of the j-th photovoltaic output power in the probability distribution of predicted power generation values.j Let s be the variance of the predicted photovoltaic output power for the j-th photovoltaic power generation value in the probability distribution of predicted power generation values. j This corresponds to the actual output power value. Based on the loss function, the model parameters of the Informer-De model and the trainable model parameters A are updated using the backpropagation algorithm. Simultaneously, during training, the trainable model parameters Gt are updated based on the random features R, satisfying... .

[0127] In one embodiment, step S202, namely, optimizing and updating the model parameters through a loss function to obtain a depth probability prediction model, includes:

[0128] S2025: Determine the update amount of the model parameters based on the loss function, and perform parameter updates on the model parameters to obtain the updated model parameters.

[0129] Specifically, gradient information is calculated for the model parameters of the current iteration based on the loss function. The update amount of the model parameters is determined based on the gradient information, and iterative update processing is performed on the model parameters according to the preset parameter update rules to obtain the updated model parameters. The model parameters include trainable parameters used to generate probability distribution parameters in the deep probabilistic prediction model and trainable parameters used to generate output features. The updated model parameters are used to generate the probability distribution of the corresponding predicted power generation value in the next iteration and participate in the subsequent loss function calculation.

[0130] S2026: Obtain the L2 norm of the matrix corresponding to the updated model parameters.

[0131] Specifically, the updated model parameters are converted into matrix form according to the preset expansion rules of the parameter hierarchy or parameter tensors, and the corresponding matrix 2 norm is calculated based on the matrix form to characterize the spectral scale of the updated model parameters. The matrix 2 norm is denoted as... In the case of multiple parameter matrices in the model, the matrix 2 norm corresponding to each parameter matrix is ​​calculated and summarized into a set of norms required for normalization, which is then used for subsequent spectral normalization of the model parameters.

[0132] S2027: Perform spectral normalization on the updated model parameters based on the matrix L2 norm to obtain spectrally normalized model parameters, and use the spectrally normalized model parameters as the model parameters for subsequent iterations.

[0133] Specifically, for each updated parameter matrix W l Based on its corresponding matrix L2 norm Normalization scaling is performed to ensure the normalized parameter matrix meets the preset spectral scale constraints. The model parameter update calculation method is as follows: This yields the spectrally normalized model parameter set, which is then used to replace the updated model parameters in subsequent iterations. This ensures that the deep probability prediction model maintains parameter scaling stability when generating the output mean and output variance in subsequent iterations, and is used for the next round of loss function calculation and parameter update.

[0134] In one embodiment, the photovoltaic power uncertainty prediction method further includes:

[0135] S40: During the inference test of the deep probabilistic prediction model, the GCRPS index is calculated based on the probability distribution of the predicted power generation value and the corresponding actual power generation value.

[0136] Specifically, during the inference testing phase, the probability distribution of the predicted power generation value is obtained based on weather forecast data samples and historical sequence operation data at the time to be predicted. Simultaneously, the actual power generation value corresponding to the time to be predicted is acquired. The actual power generation value and the probability distribution are used as evaluation inputs to calculate the GCRPS index, which must meet the following requirements. ,in M and V are the mean and variance of the output distribution of the photovoltaic power generation probability prediction module, and S is the corresponding actual power generation value. and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively, so that the GCRPS index can be used to reflect the reasonableness of the probability distribution's coverage of the actual power generation value and the deviation in the distribution characterization, and serve as a criterion for subsequent sample selection and retraining triggering.

[0137] S50: When the GCRPS index is greater than the preset threshold, the input data of the deep probabilistic prediction model corresponding to the GCRPS index and the corresponding actual power generation data are saved as retraining samples.

[0138] Specifically, the GCRPS index calculated during the inference test is compared with a preset threshold. When the GCRPS index is greater than the preset threshold, the corresponding time to be predicted is determined as the deviation time. The input data of the deep probabilistic prediction model corresponding to the deviation time is extracted as input sample data. The input data includes weather forecast data samples and related segments in historical sequence operation data corresponding to the deviation time. At the same time, the actual power generation data corresponding to the deviation time is extracted as supervision sample data. The input sample data and supervision sample data are paired and stored to form a retraining sample set and the corresponding time identifier and threshold trigger identifier are recorded for subsequent retraining calls.

[0139] S60: If the retraining samples meet the preset triggering conditions, the deep probability prediction model is retrained and optimized based on the retraining samples.

[0140] Specifically, the number of samples, sample coverage period, or threshold trigger number of the retrained sample set are compared with the preset trigger conditions. When the preset trigger conditions are met, the retrained sample set is invoked to perform parameter re-optimization on the deep probabilistic prediction model. The loss function is recalculated based on the retrained sample set, and the model parameters are iteratively updated to obtain the retrained model parameters. After the retraining is completed, the retrained deep probabilistic prediction model is used in the subsequent inference test stage to output the probability distribution of the predicted power generation value to replace the model parameter configuration before retraining.

[0141] Example 2

[0142] like Figure 2 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a photovoltaic power uncertainty prediction system, comprising:

[0143] The uncertainty measurement module is used to acquire weather forecast data of each power station within a preset range, and perform spatial correlation optimization processing on the weather forecast data to obtain the weather forecast data of the target photovoltaic power station at the time to be predicted. The weather forecast data at the time to be predicted is input into the preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation, and the weather forecast data sample is obtained by sampling according to the probability representation.

[0144] The probability prediction module is used to obtain historical sequence operation data up to the time to be predicted based on a preset time period, input the weather forecast data sample and the historical sequence operation data into the preset deep probability prediction model, and obtain the probability distribution of each predicted power generation value.

[0145] The distribution fusion module is used to unify and fuse various probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value. The deep probabilistic prediction model is trained as follows:

[0146] The sample construction module is used to acquire historical operational data and corresponding weather forecast data to generate model training samples;

[0147] The stochastic modeling module is used to extract time-series features from the model training samples to obtain the corresponding output features. The output features are then randomly calculated to obtain probability distribution parameters. Based on the probability distribution parameters, the probability distribution of the predicted power generation value is determined. The corresponding loss function is calculated according to the probability distribution and the model training samples. The model parameters are then optimized and updated using the loss function to obtain a deep probabilistic prediction model.

[0148] Optionally, the uncertainty measurement module includes:

[0149] The site acquisition submodule is used to acquire weather forecast data of the target photovoltaic power station and surrounding stations within a preset radius of the target photovoltaic power station, generate node features corresponding to each station, and generate spatial correlation features based on the geospatial information between each station.

[0150] The spatial graph inference submodule is used to construct a multi-source weather forecast spatial graph model based on the node features and the spatial correlation features, input the multi-source weather forecast spatial graph model into a preset S-GTN network, and generate weather forecast data for the time to be predicted.

[0151] The quantile output submodule is used to input the weather forecast data at the time to be predicted into the weather forecast data uncertainty measurement model to obtain several quantiles used to characterize the uncertainty of the weather forecast data.

[0152] The distribution construction submodule is used to construct the inverse function of the approximate cumulative distribution function corresponding to the probability representation based on several quantiles using linear interpolation.

[0153] The inversion sampling submodule is used to extract random numbers from a standard uniform distribution based on the inverse function of the approximate cumulative distribution function, and then map these random numbers to obtain weather forecast data samples. Optionally, the inversion sampling submodule includes:

[0154] The quantile sorting unit is used to sort a number of quantiles according to the size of the quantile level value corresponding to the quantile, so as to obtain an ordered quantile sequence.

[0155] Interval positioning units are used to generate random numbers from a standard uniform distribution and determine adjacent quantile levels and adjacent quantiles based on the position of the random numbers in the quantile level.

[0156] The interpolation generation unit is used to calculate the weather forecast data sample based on the adjacent quantile levels, adjacent quantiles, and random numbers using the inverse function of the approximate cumulative distribution function.

[0157] Optional, the construction of a model for measuring uncertainty in weather forecast data includes:

[0158] The sample acquisition submodule is used to acquire historical weather forecast data and corresponding actual weather data within a preset radius of the photovoltaic site, and generate uncertainty measurement training samples.

[0159] The quantile training submodule is used to train a preset number of quantile regression models based on training samples with uncertainty measures.

[0160] The model combination submodule is used to combine multiple quantile regression models to generate a weather forecast data uncertainty measurement model.

[0161] Optionally, the probability prediction module includes:

[0162] The time period determination submodule is used to determine the historical start time and historical end time corresponding to the preset time period based on the time to be predicted, so as to obtain the corresponding time range.

[0163] The sequence acquisition submodule is used to acquire power generation data and actual irradiance data within a time range, and to align the data according to a unified time base to obtain historical sequence operation data.

[0164] The input association submodule is used to associate historical sequence data with weather forecast data samples to obtain model input data;

[0165] The parameter generation submodule is used to input the model input data into the deep probabilistic prediction model, generate the output mean and output variance as probability distribution parameters, and generate the probability distribution of the predicted power generation value based on the probability distribution parameters.

[0166] Optionally, the distributed fusion module includes:

[0167] The parameter extraction submodule is used to extract the corresponding probability distribution parameters from each probability distribution to obtain multiple sets of output mean and output variance.

[0168] The parameter fusion submodule is used to calculate the fusion mean and fusion variance based on multiple sets of output means and output variances;

[0169] The distribution generation submodule is used to generate the final probability distribution of the predicted power generation value based on the fusion mean and fusion variance.

[0170] Optional, the stochastic modeling module includes:

[0171] The time series feature extraction module is used to extract time series features from the training samples of the Informer-De model to obtain the corresponding output features.

[0172] The random mapping module is used to perform random feature calculation on the output features based on an approximate Gaussian process to obtain random features;

[0173] The statistical calculation module is used to calculate the corresponding output mean and output variance based on random features, as probability distribution parameters.

[0174] The loss calculation module is used to construct the probability distribution of the predicted power generation value based on the probability distribution parameters, calculate the difference between the power generation data in the model training samples and the output mean, and combine the output variance to scale the difference to obtain the loss function.

[0175] Optionally, the stochastic modeling module also includes:

[0176] The parameter update submodule is used to determine the update amount of the model parameters based on the loss function, and to perform parameter updates on the model parameters to obtain the updated model parameters;

[0177] The norm calculation submodule is used to obtain the 2-norm of the matrix corresponding to the updated model parameters;

[0178] The spectral normalization submodule is used to perform spectral normalization on the updated model parameters based on the matrix L2 norm, obtain the spectral normalized model parameters, and use the spectral normalized model parameters as the model parameters for subsequent iterations.

[0179] Optionally, photovoltaic power uncertainty prediction methods also include:

[0180] The indicator evaluation module is used to calculate the GCRPS indicator based on the probability distribution of the predicted power generation value and the corresponding actual power generation value during the inference test of the deep probabilistic prediction model.

[0181] The sample accumulation module is used to save the input data of the deep probabilistic prediction model corresponding to the GCRPS index and the corresponding actual power generation data as retraining samples when the GCRPS index is greater than a preset threshold.

[0182] The trigger retraining module is used to retrain and optimize the deep probabilistic prediction model based on the retraining samples when the retraining samples meet the preset triggering conditions.

[0183] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," 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 invention. In this specification, 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.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for predicting photovoltaic power uncertainty, characterized in that, The method includes: Acquire weather forecast data for each power station within a preset range, and perform spatial correlation optimization processing on the weather forecast data to obtain the weather forecast data for the target photovoltaic power station at the time to be predicted. Input the weather forecast data at the time to be predicted into a preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation, and sample the weather forecast data based on the probability representation. Historical sequence operation data up to the time to be predicted is obtained according to a preset time period. The weather forecast data sample and the historical sequence operation data are input into a preset deep probability prediction model to obtain the probability distribution of each predicted power generation value. At the distribution level, the probability distributions are uniformly fused to obtain the final probability distribution of the predicted power generation value. The deep probabilistic prediction model is trained as follows: Acquire historical operational data and corresponding weather forecast data to generate model training samples; Time series features are extracted from the model training samples to obtain corresponding output features. Based on an approximate Gaussian process structure, feature randomization is performed on the output features to obtain probability distribution parameters. Then, based on the probability distribution parameters, the probability distribution of the predicted power generation value is determined. The corresponding loss function is calculated according to the probability distribution and the model training samples, and the model parameters are optimized and updated through the loss function to obtain the deep probability prediction model.

2. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The process involves performing spatial correlation optimization on the weather forecast data to obtain the weather forecast data for the target photovoltaic power station at the desired time. This data is then input into a preset weather forecast data uncertainty measurement model to obtain a corresponding probability representation. Finally, weather forecast data samples are obtained by sampling based on the probability representation, including: Acquire weather forecast data for the target photovoltaic power station and surrounding stations within a preset radius of the target photovoltaic power station, generate node features corresponding to each station, and generate spatial correlation features based on the geospatial information between each station; A multi-source weather forecast spatial map model is constructed based on the node features and the spatial correlation features. The multi-source weather forecast spatial map model is then input into a preset S-GTN network to generate weather forecast data for the time to be predicted. The weather forecast data at the time to be predicted is input into the weather forecast data uncertainty measurement model to obtain several quantiles used to characterize the uncertainty of the weather forecast data; Based on several quantiles, the inverse function of the approximate cumulative distribution function corresponding to the probability representation is constructed using linear interpolation. Random numbers are drawn from the standard uniform distribution based on the inverse function of the approximate cumulative distribution function, and the random numbers are mapped to obtain the weather forecast data sample.

3. The photovoltaic power uncertainty prediction method according to claim 2, characterized in that, The step of drawing random numbers from a standard uniform distribution based on the inverse function of the approximate cumulative distribution function, mapping the random numbers to obtain the weather forecast data sample includes: The quantiles are sorted according to the magnitude of the quantile level values ​​to obtain an ordered quantile sequence; Random numbers are drawn from a standard uniform distribution. The random numbers are compared with the magnitudes of each quantile level to determine the adjacent quantile levels. The corresponding adjacent quantiles are then obtained based on the adjacent quantile levels. The weather forecast data sample is obtained by calculating using the inverse function of the approximate cumulative distribution function based on the adjacent quantile level, the adjacent quantile, and the random number.

4. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The construction of the uncertainty measurement model for the weather forecast data includes: Acquire historical weather forecast data and corresponding actual weather data within a preset radius of the photovoltaic site, and generate uncertainty measurement training samples; A preset number of quantile regression models are trained based on the aforementioned uncertainty measure training samples; The multiple quantile regression models are combined to generate the weather forecast data uncertainty measurement model.

5. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The process involves acquiring historical sequence operational data up to the time to be predicted based on a preset time period, inputting the weather forecast data sample and the historical sequence operational data into a preset deep probability prediction model to obtain the probability distribution of each predicted power generation value, including: Based on the time to be predicted, determine the historical start time and historical end time corresponding to the preset time period to obtain the corresponding time range; Within the stated time range, power generation data and actual irradiance data are acquired and time-aligned according to a unified time reference to obtain the historical sequence operation data. The historical sequence operation data is correlated with the weather forecast data sample to obtain model input data. The model input data is then input into the deep probability prediction model to generate the output mean and output variance as probability distribution parameters. Based on the probability distribution parameters, the probability distribution of the predicted power generation value is generated.

6. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The step of unifying and fusing the probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value includes: Extract the corresponding probability distribution parameters from each of the probability distributions to obtain multiple sets of output mean and output variance; The fusion mean and fusion variance are calculated based on the multiple sets of output means and output variances; The final probability distribution of the predicted power generation value is generated based on the fusion mean and the fusion variance.

7. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The process involves extracting time-series features from the model training samples to obtain corresponding output features, performing random feature calculations on the output features based on an approximate Gaussian process structure to obtain probability distribution parameters, determining the probability distribution of the predicted power generation value based on the probability distribution parameters, and calculating the corresponding loss function based on the probability distribution and the model training samples, including: Based on the Informer-De model, time series features are extracted from the training samples of the model to obtain the corresponding output features; Random features are obtained by performing random feature calculation on the output features based on an approximate Gaussian process. Based on the random features, the corresponding output mean and output variance are calculated respectively, and used as the probability distribution parameters; Based on the probability distribution parameters, a probability distribution of the predicted power generation value is constructed. The difference between the power generation data in the model training samples and the output mean is calculated. The difference is then scaled by combining the output variance to obtain the loss function.

8. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The step of optimizing and updating the model parameters through the loss function to obtain the depth probability prediction model includes: The update amount of the model parameters is determined based on the loss function, and the model parameters are updated to obtain the updated model parameters. Obtain the L2 norm of the matrix corresponding to the updated model parameters; Based on the matrix L2 norm, spectral normalization is performed on the updated model parameters to obtain spectrally normalized model parameters, which are then used as model parameters for subsequent iterations.

9. The photovoltaic power uncertainty prediction method according to claim 1, characterized in that, The method further includes: During the inference test of the deep probabilistic prediction model, the GCRPS index is calculated based on the probability distribution of the predicted power generation value and the corresponding actual power generation value. If the GCRPS index is greater than a preset threshold, the input data of the deep probability prediction model corresponding to the GCRPS index and the corresponding actual power generation data are saved as retraining samples. If the retrained samples meet the preset triggering conditions, the deep probability prediction model is retrained and optimized based on the retrained samples.

10. A photovoltaic power uncertainty prediction system, characterized in that, The system includes: The uncertainty measurement module is used to acquire weather forecast data of each power station within a preset range, and perform spatial correlation optimization processing on the weather forecast data to obtain the weather forecast data of the target photovoltaic power station at the time to be predicted. The weather forecast data at the time to be predicted is input into a preset weather forecast data uncertainty measurement model to obtain the corresponding probability representation, and the weather forecast data sample is obtained by sampling according to the probability representation. The probability prediction module is used to obtain historical sequence operation data up to the time to be predicted according to a preset time period, input the weather forecast data sample and the historical sequence operation data into a preset deep probability prediction model, and obtain the probability distribution of each predicted power generation value. The distribution fusion module is used to perform consistent fusion of the various probability distributions at the distribution level to obtain the final probability distribution of the predicted power generation value. The deep probability prediction model is trained as follows: The sample construction module is used to acquire historical operational data and corresponding weather forecast data to generate model training samples; The random feature modeling module is used to extract time series features from the model training samples to obtain corresponding output features, perform random feature calculations on the output features to obtain probability distribution parameters, determine the probability distribution of the predicted power generation value based on the probability distribution parameters, calculate the corresponding loss function according to the probability distribution and the model training samples, and optimize and update the model parameters through the loss function to obtain the deep probability prediction model.

Images