Photovoltaic power probability prediction method and system for extreme weather scenarios

Abstract

This invention relates to a method and system for probabilistic prediction of photovoltaic power output under extreme weather scenarios, belonging to the field of photovoltaic power generation prediction technology. The method collects meteorological, unit operation, and grid connection data from photovoltaic power plants. Through extreme weather sample screening, improved IQR-3σ double-layer anomaly detection, meteorological neighborhood spatiotemporal weighted interpolation, variational mode decomposition, and time-frequency feature enhancement, a standardized time-frequency conditional feature tensor is obtained. A conditional generative adversarial network containing a quantum heuristic condition generator and a multi-scale extreme perception condition discriminator is constructed. Parameters are optimized through a three-stage training process: basic adversarial training, extreme sample weighting, and Focal-CVaR tail fine-tuning. An initial probability interval is constructed based on multi-scenario prediction sequences. After confidence calibration, prediction intervals at different confidence levels are output. The median of the sequence is used as the point prediction result, providing support for photovoltaic power plant scheduling and grid risk management.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation prediction technology, specifically relating to a method and system for predicting photovoltaic power probability in extreme weather scenarios. Background Technology

[0002] With the continuous expansion of new energy power generation, the proportion of photovoltaic (PV) power generation integrated into the power system is constantly increasing. PV output power is influenced by various meteorological factors such as solar irradiance, cloud cover, ambient temperature, wind speed, and humidity, exhibiting strong randomness, volatility, and intermittency. Under extreme weather conditions such as heavy rain, severe convection, dust storms, and rapid cloud movement, irradiance can change drastically in a short period, causing a rapid rise or fall in PV output power, increasing the difficulty of peak shaving and frequency regulation, reserve capacity allocation, and PV power plant operation and dispatch. Therefore, accurately predicting PV power under extreme weather conditions and quantifying its uncertainties is of great significance for PV power plant operation management and grid safety dispatch.

[0003] Existing photovoltaic power prediction methods mainly include deterministic prediction and probabilistic prediction. Deterministic prediction typically outputs only a single power prediction value, making it difficult to reflect the prediction error range and the probability of different outcomes. Probabilistic prediction can describe power uncertainty through probability distributions, quantiles, or prediction intervals, but existing methods mostly employ models such as quantile regression, Gaussian processes, Bayesian estimation, and deep neural networks. When the number of extreme weather samples is small, there are complex nonlinear couplings between multiple meteorological variables, and the power distribution exhibits skewed or heavy-tailed characteristics, problems such as insufficient tail distribution characterization, low prediction interval coverage, or excessively wide intervals are prone to occur, making it difficult to balance the reliability and sharpness of the prediction interval.

[0004] Generative Adversarial Networks (GANs) can learn complex data distributions through adversarial training between generators and discriminators, and can be used to generate multiple photovoltaic power prediction scenarios. However, classic GANs tend to favor distributions under normal operating conditions when the proportion of extreme samples is low, and may suffer from problems such as training instability, pattern collapse, and insufficient diversity of generated samples. If the generator does not fully integrate information on weather, operating status, and historical power, the generated results may not accurately reflect the power distribution under specific weather and operating conditions.

[0005] Quantum-inspired machine learning utilizes concepts such as quantum state superposition, parameterized rotation, and entanglement to construct high-dimensional feature maps, enhancing the model's ability to express complex nonlinear relationships. However, existing technologies lack a unified design for extreme sample screening, outlier processing, multi-scale time-frequency feature extraction, tail risk constraints, and prediction interval calibration. Furthermore, conventional outlier removal methods may inadvertently delete real extreme power fluctuation samples, and uncalibrated prediction intervals may exhibit discrepancies between actual coverage and nominal confidence levels.

[0006] Therefore, it is necessary to propose a photovoltaic power probability prediction method and system suitable for extreme weather scenarios to improve the conditional distribution fitting ability, tail prediction accuracy and probability interval reliability under extreme conditions. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for predicting photovoltaic power probability in extreme weather scenarios, thereby improving the accuracy and reliability of photovoltaic power prediction within probability intervals under extreme weather conditions and providing support for grid safety dispatch and power plant risk management.

[0008] To achieve the above objectives, the technical solution of the present invention is: a method for probabilistic prediction of photovoltaic power in extreme weather scenarios, comprising:

[0009] Step 1: Collect meteorological data, unit operation data and grid connection data of photovoltaic power station, screen extreme weather samples, and perform abnormal data detection, missing data interpolation, variational mode decomposition, feature normalization and sliding time window construction on the screened samples in sequence to obtain standardized time-frequency conditional feature tensors;

[0010] Step 2: Construct a conditional generative adversarial network including a quantum heuristic conditional generator and a multi-scale extreme perception conditional discriminator. Input the standardized time-frequency conditional feature tensor and random noise into the quantum heuristic conditional generator to output a photovoltaic power prediction sequence. The multi-scale extreme perception conditional discriminator then distinguishes between the real power sequence and the photovoltaic power prediction sequence.

[0011] Step 3: Optimize the parameters of the quantum heuristic condition generator and the multi-scale extreme perception condition discriminator through three stages in sequence: basic condition adversarial training, extreme sample dynamic weighted training, and Focal-CVaR joint tail fine-tuning.

[0012] Step 4: Generate multiple photovoltaic power prediction sequences based on the same standardized time-frequency conditional feature tensor, construct an initial probability prediction interval by kernel density smoothing quantile estimation, and then perform confidence calibration on the initial probability prediction interval according to the interval coverage deviation of the calibration sample to obtain photovoltaic power probability prediction intervals at different confidence levels.

[0013] Step 5: Take the median of multiple photovoltaic power prediction sequences as the point prediction result, output the point prediction result and the photovoltaic power probability prediction interval under different confidence levels, and evaluate the prediction effect through evaluation indicators.

[0014] Furthermore, step 1 includes the following steps:

[0015] Step 1.1: Multi-source data acquisition: Collect meteorological data (irradiance, ambient temperature, wind speed, humidity, cloud cover), generator operation data (PV module output power, module temperature, inverter efficiency), and grid connection data (grid connection point voltage, active power, reactive power); all data are aligned with a unified timestamp, and the sampling resolution is uniformly set to 15 minutes to form the original dataset:

[0016]

[0017] in, for Multidimensional feature vector at time step, This represents the total number of sampling points.

[0018] Step 1.2: Targeted screening of extreme weather samples: Calculate the first... Power fluctuation at each sampling time and irradiance mutation rate Construct a comprehensive extreme index:

[0019]

[0020] in, for The comprehensive extreme index at each sampling time; , These are the median of the total sequence power volatility and the irradiance mutation rate, respectively. , The weighting coefficients of the two indicators satisfy the following conditions: The comprehensive extreme index is greater than the preset quantile threshold. The samples were identified as candidate samples for extreme operating conditions. By combining changes in meteorological elements, inverter operating status, grid connection status, and irradiance-power response relationship, non-meteorological power surges caused by equipment failure, power rationing, communication anomalies, or shutdowns were excluded, resulting in an extreme weather sample set.

[0021] Step 1.3: Improved IQR-3σ Two-Layer Anomaly Detection: Calculate the global IQR deviation and the local sliding window 3σ deviation, where the global IQR deviation is:

[0022]

[0023] The local sliding window 3σ deviation is:

[0024]

[0025] The overall anomaly confidence level is:

[0026]

[0027] in, for The two-layer anomaly confidence score for time-based data shows that the higher the value, the higher the probability of an anomaly. This represents the global IQR deviation. The median of the sequence. Interquartile range; and These represent the first and third quartiles of the sequence, respectively; The 3σ deviation of the local sliding window. , The mean and standard deviation of the series within the window; The global deviation weight is used. When the comprehensive anomaly confidence level is greater than the preset anomaly threshold, the corresponding sample is marked as a statistically abnormal sample. For the extreme weather samples obtained in step 1.2, physical consistency is further checked based on the photovoltaic power rated boundary, irradiance-power consistency, continuity of adjacent time periods, and equipment operating status. Only when they simultaneously meet the statistical anomaly conditions and physical anomaly conditions are they judged as invalid abnormal data; otherwise, they are retained as valid extreme weather samples.

[0028] Step 1.4: Meteorological Neighborhood Spatiotemporal Weighted Interpolation: For the missing data resulting from outlier removal, interpolation is performed using spatial distance weighting, temporal proximity weighting, and meteorological correlation weighting.

[0029]

[0030] The weights for the three categories are as follows:

[0031]

[0032]

[0033]

[0034] in, For the interpolation results of missing values, This represents the total number of valid neighboring samples selected. The spatial distance weight is denoted as , where is the spatial distance. The spatial smoothing coefficient is... The closer the distance, the greater the weight; For time proximity weights, For the time difference, This is a time smoothing coefficient; the closer the time, the greater the weight. For meteorological correlation weights, The Pearson correlation coefficient is the meteorological element corresponding to the neighboring sample and the target sample. The stronger the correlation, the greater the weight.

[0035] Step 1.5: Variational Mode Decomposition Time-Frequency Feature Enhancement: Variational mode decomposition is performed on the irradiance and photovoltaic output power time series to obtain multiple mode components; a mode effectiveness index is constructed based on the energy proportion and kurtosis of each mode component.

[0036]

[0037] in, For the first The effectiveness index of a modal component indicates that the mode contains more effective information and less noise. For the first The energy percentage of each mode reflects the degree to which that component contributes to the original signal; For the first The kurtosis value of each mode reflects the characteristics of the signal pulse. This is a reference kurtosis value obtained based on historical valid samples. This represents the maximum deviation of all modal kurtosis from the reference kurtosis. , Let be the weighting coefficient for energy percentage and kurtosis, and satisfy: The modality retention index is not lower than a preset threshold. The modal components are obtained, and their time-domain statistical features and frequency-domain energy features are extracted and spliced ​​with the original meteorological and operational features to form a time-frequency fusion feature set.

[0038] Step 1.6: Normalization and Conditional Temporal Feature Construction: Max-min normalization is used to map all features to the [0,1] interval to eliminate dimensional differences; a sliding time window is used to construct the temporal input samples, with a window length of L and a step size of 1, and the final output is a standardized time-frequency conditional feature matrix. Where N is the number of samples and D is the enhanced feature dimension, the standardized time-frequency conditional feature tensor serves as the common conditional input for the quantum heuristic conditional generator and the multi-scale extreme perception conditional discriminator.

[0039] Furthermore, step 2 includes the following steps:

[0040] Step 2.1: Quantum bit ground state encoding: normalizing the time-frequency conditional feature tensor Input the conditional feature encoding layer to obtain the conditional feature vector:

[0041]

[0042] The conditional feature vector With standard normal random noise vector Perform splicing, and calculate the number of splicing results. Initial rotation angle of the quantum heuristic bit state:

[0043]

[0044] Constructing the initial quantum heuristic state:

[0045]

[0046] in, For the initial quantum state, This represents the concatenation result of a random noise vector and a conditional eigenvector. For tensor product operations, The activation function for Sigmoid. For the first The initial rotation angle of each qubit; , These are trainable weights and biases.

[0047] Step 2.2: Conditional Weighted Superposition State Mapping: Calculate the superposition state weighting coefficients based on the conditional eigenvectors.

[0048]

[0049] Constructing a conditionally weighted superposition state:

[0050]

[0051] in, It is a weighted superposition state; To calculate the ground state, there are a total of 2^n states, achieving an exponential dimensionality increase in the feature space; This represents the original probability amplitude of the ground state. These are the ground-state weighting coefficients adaptively calculated based on the conditional feature encoding results.

[0052] Step 2.3: Rotation-Entanglement Co-evolution Layer: Construction The layered parameterized quantum heuristic transformation unit comprises a single-bit Y-axis rotation gate and a mutual information adaptive parameterized controlled rotation entanglement gate in each layer; the single-bit Y-axis rotation gate is as follows:

[0053]

[0054] Mutual information adaptive controlled entanglement gate:

[0055]

[0056] in,

[0057] In the formula, For the first Layer A bit-sized Y-axis rotary gate To enable trainable rotation parameters, allowing for continuous transformation of single-qubit states; Indicates the action on the first The unit operator of the nth qubit, whose function is to maintain the nth qubit The state of each qubit remains unchanged. and They represent the first Each qubit is used to calculate the ground state. and The projection operator on the control bit is used to determine which ground state component the control bit is in. Represents a single-qubit gate that rotates about the Y-axis of a Bloch sphere. Indicates the first The conditional feature and the first Mutual information between conditional features; Indicates the first Layer entanglement strength coefficient.

[0058] go through After layer rotation-entanglement co-evolution, the final quantum heuristic state is obtained:

[0059]

[0060] in, Indicates the first Evolution matrix corresponding to layer parameterized transformation unit

[0061] Step 2.4: Quantum Heuristic State Measurement Decoding: Measure the preset observable of the final quantum heuristic state to obtain the measurement characteristics:

[0062]

[0063] Measuring feature vectors Input to a fully connected decoding layer, obtain a length of Photovoltaic power prediction sequence:

[0064]

[0065] in, Indicates the action on the first A measurement operator for the state of a quantum heuristic bit; Indicates the rated power of the photovoltaic power station; and This represents the trainable parameters of the fully connected decoding layer.

[0066] Step 2.5: Multi-scale extreme sensing condition discrimination: The actual photovoltaic power sequence... Or generate photovoltaic power prediction sequences With conditional eigenvectors The input is fused to form a conditional discriminant input; one-dimensional dilated convolution branches with different dilation rates are used to extract features at multiple time scales.

[0067]

[0068] or:

[0069]

[0070] in, Indicates the first The dilation rate of a one-dimensional dilated convolution branch.

[0071] The fusion weights for each scale branch are calculated using a trainable attention mechanism:

[0072]

[0073] And obtain multi-scale fusion features:

[0074]

[0075] The multi-scale fused features are input into the discriminant output layer to obtain the sample authenticity probability:

[0076]

[0077] in, It is a multi-scale fusion feature; For the first Convolutional features of each dilation rate branch; For the first The information entropy of each branch feature reflects the amount of information contained in the feature; For adaptive fusion weights; Indicates the number of branches in the dilated convolution; , , , and All of these represent trainable parameters.

[0078] Furthermore, step 3 includes the following steps: Step 3 is used to progressively train the quantum-inspired conditional generator and the multi-scale extreme-sensing conditional discriminator constructed in step 2. The three training phases are performed sequentially, with each subsequent phase optimizing the model parameters obtained in the previous phase without re-initializing them. During training, the gradients of the model parameters are calculated based on the loss functions corresponding to each phase, and the trainable parameters of the generator and discriminator are iteratively updated through backpropagation and the Adam optimizer. Specifically, the basic adversarial training phase is used to learn the overall distribution of photovoltaic power under given conditions; the extreme sample dynamic weighted training phase is used to improve the impact of extreme weather samples on model parameter updates; and the Focal-CVaR joint tail fine-tuning phase is used to further enhance the learning of high-error, difficult samples and tail-risk samples.

[0079] Step 3.1: Basic Adversarial Training Phase: First, the quantum heuristic conditional generator and the multi-scale extreme sensing conditional discriminator are trained alternately and iteratively using a conditional generation adversarial objective function, enabling the generator to initially learn the mapping relationship between conditional features and photovoltaic power, as well as the overall probability distribution.

[0080]

[0081] in, For the value function of adversarial games; Represents the joint distribution of real data Calculate the mathematical expectation of the sample. This represents the joint distribution of the true photovoltaic power sequence and conditional characteristics. This represents the log-likelihood term of the discriminator with respect to the real samples. Represents the prior distributions of random noise, respectively. and conditional feature distribution Sampling is performed on the generated samples, and the expected value of the discrimination results is calculated. This represents the log-likelihood term of the discriminator for the generated samples. This indicates that the discriminator is based on conditional features. The following will generate a photovoltaic power sequence. The probability of being identified as a real sample. Represents the prior distribution of random noise. This represents the conditional characteristic distribution.

[0082] Step 3.2: Dynamically Weighted Training of Extreme Samples: Training continues based on the model parameters obtained in Step 3.1. Considering the relatively small number of extreme weather samples in the training set, to enhance the impact of extreme samples on model parameter updates, an S-curve is used to construct the first... Dynamic weights for extreme samples across training rounds:

[0083]

[0084] in, and These represent the minimum and maximum values ​​of the extreme sample weights, respectively. Indicates the slope of the S-curve. This indicates the central training round where the weights of extreme samples increase rapidly.

[0085] No. The sample weights corresponding to each training sample are:

[0086]

[0087] Weighted loss:

[0088]

[0089] Where B represents the training batch size. This represents the log-likelihood of the discriminator with respect to the real samples. For the first A real photovoltaic power sequence, To and Corresponding conditional features This indicates that the discriminator is based on conditional features. Below, the actual photovoltaic power sequence The probability of being identified as a real sample. This term represents the log-likelihood of the discriminator with respect to the generated samples. This indicates that the discriminator is based on conditional features. The probability of determining the generated power sequence as a real sample is then calculated. For the generator based on random noise and conditional features The generated photovoltaic power sequence.

[0090] Step 3.3: Focal-CVaR Joint Tail Fine-tuning: Addressing the issues of large prediction errors for some samples and high tail risk in the error distribution under extreme conditions, the sample prediction error is calculated based on the prediction results of multiple scenarios under the same conditions. Focal hard sample loss and CVaR tail risk loss are constructed separately. The Focal loss is used to mitigate the impact of high-error hard samples on parameter updates, while the CVaR loss is used to focus on constraining the tail samples with large prediction errors. The two are combined with dynamic weights to form a joint tail loss, which is then used to update the generator parameters through backpropagation and the Adam optimizer. The first step is calculated based on the median of the prediction results from multiple scenarios under the same conditions. Prediction error per sample:

[0091]

[0092] in, For the first A real photovoltaic power sequence, Indicates features under the same conditions Next, input to the generator Groups of independent random noise ,get The photovoltaic power generation results are then used to take the median of these results as the nth photovoltaic power generation result. Predicted point values ​​for each sample;

[0093] Constructing Focal hard sample loss:

[0094]

[0095] Construction of Conditional Risk Value Loss:

[0096]

[0097] in, Represents the prediction error sequence Quantiles This indicates the confidence level or tail percentile level of CVaR. , Represents the Focal focusing factor. This represents the error scale parameter.

[0098] Constructing a joint tail loss:

[0099]

[0100] Among them, the dynamic balancing weights are:

[0101]

[0102] Total generator loss:

[0103]

[0104] in, This indicates that the generator is subject to conditional resistance to loss. This represents the tail loss weights that increase with the number of training epochs. Indicates the total number of training rounds. This represents the maximum value of the dynamic weight. This represents the minimum value of the dynamic weight.

[0105] Through continuous training in the above three stages, the model parameters gradually transition from overall distribution learning to extreme sample reinforcement and tail risk optimization. While maintaining basic prediction capabilities, the model's ability to learn from photovoltaic power mutations and high-error samples under extreme weather conditions is improved.

[0106] Furthermore, step 4 includes the following steps:

[0107] Step 4.1: Generation of multi-scenario photovoltaic power sequences: For the same standardized time-frequency conditional feature tensor corresponding to the time to be predicted Input to the trained quantum heuristic conditional generator Generate a set of mutually independent standard normal random noise. Photovoltaic power prediction sequence:

[0108]

[0109] Construct a sample set of photovoltaic power prediction distribution:

[0110]

[0111] Step 4.2: Constructing the initial interval for kernel density smoothing quantiles: For the h-th prediction time in the prediction time domain, the power probability density function is calculated using the kernel density estimation method:

[0112]

[0113] The corresponding cumulative distribution function is:

[0114]

[0115] in, Represents the kernel function. Let the value of the photovoltaic power random variable be denoted as . Indicates the first The kernel density estimation bandwidth corresponding to each prediction time point Indicates the first The predicted sequence in the 1st Prediction power at each prediction time, This represents the integral variable.

[0116] For nominal confidence level The lower and upper bounds of the initial probability prediction interval are as follows:

[0117]

[0118]

[0119] Indicates the first Cumulative distribution function at each prediction time The generalized inverse function of is also called the quantile function;

[0120] Step 4.3: Quantum superposition state confidence calibration: Establishing the target confidence level With quantum superposition state probability amplitude Mapping relationship between them:

[0121]

[0122] The interval calibration coefficient is calculated based on the interval coverage deviation on an independent calibration sample set that was not involved in model parameter training:

[0123]

[0124] And the range calibration coefficients are subjected to amplitude limiting processing:

[0125]

[0126] in, This represents the inverse cumulative distribution function of the standard normal distribution; Indicates the interval coverage on an independent calibration sample set or a rolling calibration window; This represents the coverage deviation feedback coefficient; and These represent the lower and upper limits of the interval calibration coefficient, respectively.

[0127] Calculate the first Median of scene predictions at each prediction time:

[0128]

[0129] The lower bound of the calibrated final probability prediction interval is:

[0130]

[0131] The upper bound of the calibrated final probability prediction interval is:

[0132]

[0133] Make the final probability prediction interval satisfy:

[0134] .

[0135] Furthermore, the evaluation indicators in step 5 include interval coverage, average interval width, average absolute error, and root mean square error, which comprehensively evaluate the probability prediction effect from three dimensions: reliability, accuracy, and precision.

[0136] (1) Interval Coverage Rate (PICP)

[0137]

[0138] in, Indicates the number of evaluation samples. Indicates the first The actual photovoltaic power at each evaluation moment; This is an indicator function, taking the value 1 when the condition is met and 0 otherwise; it measures the proportion of the actual power falling within the prediction interval, reflecting the reliability of the probabilistic prediction.

[0139] (2) Interval Average Width (MPIW)

[0140]

[0141] This measures the average width of the prediction interval; a smaller value indicates a more accurate prediction.

[0142] (3) Mean Absolute Error (MAE)

[0143]

[0144] The average absolute deviation between the predicted and actual values ​​of a point is measured, reflecting the accuracy of the prediction.

[0145] (4) Root Mean Square Error (RMSE)

[0146]

[0147] in, Indicates the first The evaluation results are calculated using the median of the photovoltaic power prediction results for multiple scenarios at each evaluation time. The interval coverage rate is used to evaluate the reliability of the probability prediction interval, the average interval width is used to evaluate the sharpness of the probability prediction interval, and the mean absolute error and root mean square error are used to evaluate the accuracy of the point prediction results. Finally, probability prediction intervals and point prediction results at different confidence levels are output, providing data support for photovoltaic power scheduling.

[0148] The present invention also provides a photovoltaic power probability prediction system for extreme weather scenarios, which performs the method described above, including:

[0149] The multi-scale time-frequency coupled data preprocessing module is used to collect meteorological data, unit operation data and grid connection data of photovoltaic power plants and align them with a unified timestamp. It filters extreme weather samples and sequentially performs improved IQR-3σ double-layer anomaly detection, meteorological neighborhood spatiotemporal weighted interpolation, variational mode decomposition, feature normalization and sliding time window construction, and outputs a standardized time-frequency conditional feature tensor.

[0150] The quantum superposition-entanglement cooperative conditional generative adversarial network module includes a quantum heuristic conditional generator and a multi-scale extreme perception conditional discriminator. The quantum heuristic conditional generator is used to take random noise and the standardized time-frequency conditional feature tensor as input, and generate a photovoltaic power prediction sequence through conditional feature encoding, weighted superposition state mapping, rotation-entanglement cooperative evolution and measurement decoding. The multi-scale extreme perception conditional discriminator is used to fuse the conditional features with the real power sequence and the photovoltaic power prediction sequence respectively, and output the sample authenticity probability through multi-scale dilated convolution and attention fusion.

[0151] The dynamic dual-weighted tail-aware training module is used to sequentially perform basic conditional adversarial training, extreme sample dynamic weighted training, and Focal-CVaR joint tail fine-tuning to adjust the parameters of the quantum heuristic condition generator and the multi-scale extreme-aware conditional discriminator.

[0152] The probability prediction interval generation and confidence calibration module is used to generate multiple photovoltaic power prediction sequences based on the same standardized time-frequency conditional feature tensor. It constructs an initial probability prediction interval by kernel density smoothing quantile estimation, calculates the interval calibration coefficient based on the interval coverage deviation of the calibration sample, calibrates the initial probability prediction interval, and outputs photovoltaic power probability prediction intervals at different confidence levels.

[0153] The prediction result output module is used to take the median of the multiple photovoltaic power prediction sequences as the point prediction result, output the point prediction result and the photovoltaic power probability prediction interval under different confidence levels, and calculate the interval coverage, interval average width, average absolute error and root mean square error after obtaining the actual power, and evaluate the prediction result from three aspects: interval reliability, interval sharpness and point prediction accuracy.

[0154] Compared with the prior art, the present invention has the following beneficial effects:

[0155] 1. Improve the quality of extreme weather sample identification and data preprocessing.

[0156] This invention constructs a comprehensive extreme index by fusing power fluctuation rate and irradiance mutation rate, and combines meteorological changes, equipment status, and power response characteristics to screen effective extreme weather samples, reducing the interference of non-meteorological factors on sample identification. Simultaneously, it employs improved IQR-3σ dual-layer anomaly detection, meteorological neighborhood spatiotemporal weighted interpolation, and variational mode decomposition to achieve anomaly data identification, missing data recovery, and time-frequency feature enhancement, thereby improving the reliability of input data.

[0157] 2. Improve the accuracy of photovoltaic power point prediction under extreme weather conditions.

[0158] This invention combines time-frequency condition features with random noise by constructing a quantum-heuristic condition generator and a multi-scale extreme sensing condition discriminator, thereby enhancing the model's ability to learn from short-term power mutations, multi-factor coupling, and nonlinear change patterns. This effectively improves the problem of large prediction errors in traditional models under extreme weather scenarios.

[0159] 3. Improve the reliability and accuracy of probability prediction intervals.

[0160] This invention constructs a prediction distribution by generating multiple sets of power prediction sequences, and generates a probability prediction interval by combining kernel density smoothing quantile estimation and confidence calibration mechanism; at the same time, it uses Focal-CVaR combined with tail loss to constrain high-error samples, improves the coverage of the prediction interval, reduces the problem of excessively wide intervals, and achieves a balance between reliability and sharpness.

[0161] 4. Enhanced tail distribution characterization capability for extreme power applications.

[0162] This invention improves the model's attention to low-probability, high-amplitude power fluctuation samples by using dynamic weighting of extreme samples, multi-scale feature extraction, and Focal-CVaR tail risk constraints, thereby enhancing the predictive ability of power abrupt changes under extreme weather conditions such as rainstorms, strong cloud cover, and dust storms.

[0163] 5. Improve the model's generalization ability under extremely small sample conditions.

[0164] This invention utilizes the superposition state mapping and rotation-entanglement co-evolution mechanism in quantum heuristic generative structures to enhance the expressive power of high-dimensional feature spaces, improve the diversity of generated samples, and reduce the problems of pattern collapse and insufficient distribution coverage that traditional generative adversarial networks are prone to under extreme weather and small sample conditions. Attached Figure Description

[0165] Figure 1 This is a schematic diagram of the multi-scale time-frequency coupled data acquisition and preprocessing module provided by the present invention.

[0166] Figure 2 A schematic diagram of the quantum superposition-entanglement cooperative generator structure provided by the present invention.

[0167] Figure 3 A schematic diagram of the dynamic dual-weighted tail perception adversarial training process provided by the present invention.

[0168] Figure 4 A schematic diagram illustrating the probability interval calibration principle of the quantum superposition state confidence calibration of the present invention.

[0169] Figure 5 This is a schematic diagram of the overall system architecture provided by the present invention.

[0170] Figure 6This is a comparison chart of the actual power and the predicted power of the model.

[0171] Figure 7 This model's 90% probability interval prediction chart.

[0172] Figure 8 Scatter plot of actual and predicted values ​​from this model.

[0173] Figure 9 Extreme operating condition prediction results diagram. Detailed Implementation

[0174] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0175] This invention provides a method for probabilistic prediction of photovoltaic power in extreme weather scenarios, comprising:

[0176] Step 1: Collect meteorological data, unit operation data and grid connection data of photovoltaic power station, screen extreme weather samples, and perform abnormal data detection, missing data interpolation, variational mode decomposition, feature normalization and sliding time window construction on the screened samples in sequence to obtain standardized time-frequency conditional feature tensors;

[0177] Step 2: Construct a conditional generative adversarial network including a quantum heuristic conditional generator and a multi-scale extreme perception conditional discriminator. Input the standardized time-frequency conditional feature tensor and random noise into the quantum heuristic conditional generator to output a photovoltaic power prediction sequence. The multi-scale extreme perception conditional discriminator then distinguishes between the real power sequence and the photovoltaic power prediction sequence.

[0178] Step 3: Optimize the parameters of the quantum heuristic condition generator and the multi-scale extreme perception condition discriminator through three stages in sequence: basic condition adversarial training, extreme sample dynamic weighted training, and Focal-CVaR joint tail fine-tuning.

[0179] Step 4: Generate multiple photovoltaic power prediction sequences based on the same standardized time-frequency conditional feature tensor, construct an initial probability prediction interval by kernel density smoothing quantile estimation, and then perform confidence calibration on the initial probability prediction interval according to the interval coverage deviation of the calibration sample to obtain photovoltaic power probability prediction intervals at different confidence levels.

[0180] Step 5: Take the median of multiple photovoltaic power prediction sequences as the point prediction result, output the point prediction result and the photovoltaic power probability prediction interval under different confidence levels, and evaluate the prediction effect through evaluation indicators.

[0181] The present invention also provides a photovoltaic power probability prediction system for extreme weather scenarios, which performs the method described above, including:

[0182] The multi-scale time-frequency coupled data preprocessing module is used to collect meteorological data, unit operation data and grid connection data of photovoltaic power plants and align them with a unified timestamp. It filters extreme weather samples and sequentially performs improved IQR-3σ double-layer anomaly detection, meteorological neighborhood spatiotemporal weighted interpolation, variational mode decomposition, feature normalization and sliding time window construction, and outputs a standardized time-frequency conditional feature tensor.

[0183] The quantum superposition-entanglement cooperative conditional generative adversarial network module includes a quantum heuristic conditional generator and a multi-scale extreme perception conditional discriminator. The quantum heuristic conditional generator is used to take random noise and the standardized time-frequency conditional feature tensor as input, and generate a photovoltaic power prediction sequence through conditional feature encoding, weighted superposition state mapping, rotation-entanglement cooperative evolution and measurement decoding. The multi-scale extreme perception conditional discriminator is used to fuse the conditional features with the real power sequence and the photovoltaic power prediction sequence respectively, and output the sample authenticity probability through multi-scale dilated convolution and attention fusion.

[0184] The dynamic dual-weighted tail-aware training module is used to sequentially perform basic conditional adversarial training, extreme sample dynamic weighted training, and Focal-CVaR joint tail fine-tuning to adjust the parameters of the quantum heuristic condition generator and the multi-scale extreme-aware conditional discriminator.

[0185] The probability prediction interval generation and confidence calibration module is used to generate multiple photovoltaic power prediction sequences based on the same standardized time-frequency conditional feature tensor. It constructs an initial probability prediction interval by kernel density smoothing quantile estimation, calculates the interval calibration coefficient based on the interval coverage deviation of the calibration sample, calibrates the initial probability prediction interval, and outputs photovoltaic power probability prediction intervals at different confidence levels.

[0186] The prediction result output module is used to take the median of the multiple photovoltaic power prediction sequences as the point prediction result, output the point prediction result and the photovoltaic power probability prediction interval under different confidence levels, and calculate the interval coverage, interval average width, average absolute error and root mean square error after obtaining the actual power, and evaluate the prediction result from three aspects: interval reliability, interval sharpness and point prediction accuracy.

[0187] The following is a detailed implementation process of the present invention.

[0188] like Figure 5 As shown, this invention provides a method for probabilistic prediction of photovoltaic power in extreme weather scenarios, and the specific implementation is as follows:

[0189] Step 1 as follows Figure 1 As shown:

[0190] Step 1.1: Multi-source data acquisition: Collect meteorological data (irradiance, ambient temperature, wind speed, humidity, cloud cover), generator operation data (PV module output power, module temperature, inverter efficiency), and grid connection data (grid connection point voltage, active power, reactive power); all data are aligned with a unified timestamp, and the sampling resolution is uniformly set to 15 minutes to form the original dataset:

[0191]

[0192] in, for Multidimensional feature vector at time step, This represents the total number of sampling points.

[0193] Step 1.2: Targeted screening of extreme weather samples: Calculate the first... Power fluctuation at each sampling time and irradiance mutation rate Construct a comprehensive extreme index:

[0194]

[0195] in, for The comprehensive extreme index at each sampling time; , These are the median of the total sequence power volatility and the irradiance mutation rate, respectively. , The weighting coefficients of the two indicators satisfy the following conditions: The comprehensive extreme index is greater than the preset quantile threshold. The samples were identified as candidate samples for extreme operating conditions. By combining changes in meteorological elements, inverter operating status, grid connection status, and irradiance-power response relationship, non-meteorological power surges caused by equipment failure, power rationing, communication anomalies, or shutdowns were excluded, resulting in an extreme weather sample set.

[0196] Step 1.3: Improved IQR-3σ Two-Layer Anomaly Detection: Calculate the global IQR deviation and the local sliding window 3σ deviation, where the global IQR deviation is:

[0197]

[0198] The local sliding window 3σ deviation is:

[0199]

[0200] The overall anomaly confidence level is:

[0201]

[0202] in, for The two-layer anomaly confidence score for time-based data shows that the higher the value, the higher the probability of an anomaly. This represents the global IQR deviation. The median of the sequence. Interquartile range; and These represent the first and third quartiles of the sequence, respectively; The 3σ deviation of the local sliding window. , The mean and standard deviation of the series within the window; The global deviation weight is used. When the comprehensive anomaly confidence level is greater than the preset anomaly threshold, the corresponding sample is marked as a statistically abnormal sample. For the extreme weather samples obtained in step 1.2, physical consistency is further checked based on the photovoltaic power rated boundary, irradiance-power consistency, continuity of adjacent time periods, and equipment operating status. Only when they simultaneously meet the statistical anomaly conditions and physical anomaly conditions are they judged as invalid abnormal data; otherwise, they are retained as valid extreme weather samples.

[0203] Step 1.4: Meteorological Neighborhood Spatiotemporal Weighted Interpolation: For the missing data resulting from outlier removal, interpolation is performed using spatial distance weighting, temporal proximity weighting, and meteorological correlation weighting.

[0204]

[0205] The weights for the three categories are as follows:

[0206]

[0207]

[0208]

[0209] in, For the interpolation results of missing values, This represents the total number of valid neighboring samples selected. Here, represents the spatial weight, and represents the spatial distance. The spatial smoothing coefficient is... The closer the distance, the greater the weight; As time weight, For the time difference, This is a time smoothing coefficient; the closer the time, the greater the weight. For meteorological correlation weights, The Pearson correlation coefficient is the meteorological element corresponding to the neighboring sample and the target sample. The stronger the correlation, the greater the weight.

[0210] Step 1.5: Variational Mode Decomposition Time-Frequency Feature Enhancement: Variational mode decomposition is performed on the irradiance and photovoltaic output power time series to obtain multiple mode components; a mode effectiveness index is constructed based on the energy proportion and kurtosis of each mode component.

[0211]

[0212] in, For the first The effectiveness index of a modal component indicates that the mode contains more effective information and less noise. For the first The energy percentage of each mode reflects the degree to which that component contributes to the original signal; For the first The kurtosis value of each mode reflects the characteristics of the signal pulse. This is a reference kurtosis value obtained from statistical analysis of historical valid samples; , Let be the weighting coefficient for energy percentage and kurtosis, and satisfy: The modality retention index is not lower than a preset threshold. The modal components are obtained, and their time-domain statistical features and frequency-domain energy features are extracted and spliced ​​with the original meteorological and operational features to form a time-frequency fusion feature set.

[0213] Step 1.6: Normalization and Conditional Temporal Feature Construction: Max-min normalization is used to map all features to the [0,1] interval to eliminate dimensional differences; a sliding time window is used to construct the temporal input samples, with a window length of L and a step size of 1, and the final output is a standardized time-frequency conditional feature matrix. Where N is the number of samples and D is the enhanced feature dimension, the standardized time-frequency conditional feature tensor serves as the common conditional input for the quantum heuristic conditional generator and the multi-scale extreme perception conditional discriminator.

[0214] Step 2 as follows Figure 2 As shown:

[0215] Step 2.1: Quantum bit ground state encoding: normalizing the time-frequency conditional feature tensor Input the conditional feature encoding layer to obtain the conditional feature vector:

[0216]

[0217] The conditional feature vector With standard normal random noise vector Perform splicing, and calculate the number of splicing results. Initial rotation angle of the quantum heuristic bit state:

[0218]

[0219] Constructing the initial quantum heuristic state:

[0220]

[0221] in, For the initial quantum state, This represents the concatenation result of a random noise vector and a conditional eigenvector. For tensor product operations, For the first The initial rotation angle of each qubit; , These are trainable weights and biases.

[0222] Step 2.2: Conditional Weighted Superposition State Mapping: Calculate the superposition state weighting coefficients based on the conditional eigenvectors.

[0223]

[0224] Constructing a conditionally weighted superposition state:

[0225]

[0226] in, It is a weighted superposition state; To calculate the ground state, there are a total of 2^n states, achieving an exponential dimensionality increase in the feature space; This represents the original probability amplitude of the ground state. These are the ground-state weighting coefficients adaptively calculated based on the conditional feature encoding results.

[0227] Step 2.3: Rotation-Entanglement Co-evolution Layer: Construction The layered parameterized quantum heuristic transformation unit comprises a single-bit Y-axis rotation gate and a mutual information adaptive parameterized controlled rotation entanglement gate in each layer; the single-bit Y-axis rotation gate is as follows:

[0228]

[0229] Mutual information adaptive controlled entanglement gate:

[0230]

[0231] in,

[0232] In the formula, For the first Layer A bit-sized Y-axis rotary gate To enable trainable rotation parameters, allowing for continuous transformation of single-qubit states; Indicates the first The conditional feature and the first Mutual information between conditional features; Indicates the first Layer entanglement strength coefficient.

[0233] go through After layer rotation-entanglement co-evolution, the final quantum heuristic state is obtained:

[0234]

[0235] in, Indicates the first Evolution matrix corresponding to layer parameterized transformation unit

[0236] Step 2.4: Quantum Heuristic State Measurement Decoding: Measure the preset observable of the final quantum heuristic state to obtain the measurement characteristics:

[0237]

[0238] Measuring feature vectors Input to a fully connected decoding layer, obtain a length of Photovoltaic power prediction sequence:

[0239]

[0240] in, Indicates the action on the first A measurement operator for the state of a quantum heuristic bit; Indicates the rated power of the photovoltaic power station; and This represents the trainable parameters of the fully connected decoding layer.

[0241] Step 2.5: Multi-scale extreme sensing condition discrimination: The actual photovoltaic power sequence... Or generate photovoltaic power prediction sequences With conditional eigenvectors The input is fused to form a conditional discriminant input; one-dimensional dilated convolution branches with different dilation rates are used to extract features at multiple time scales.

[0242]

[0243] or:

[0244]

[0245] in, Indicates the first The dilation rate of a one-dimensional dilated convolution branch.

[0246] The fusion weights for each scale branch are calculated using a trainable attention mechanism:

[0247]

[0248] And obtain multi-scale fusion features:

[0249]

[0250] The multi-scale fused features are input into the discriminant output layer to obtain the sample authenticity probability:

[0251]

[0252] in, It is a multi-scale fusion feature; For the first Convolutional features of each dilation rate branch; For the first The information entropy of each branch feature reflects the amount of information contained in the feature; For adaptive fusion weights; Indicates the number of branches in the dilated convolution; , , , and All of these represent trainable parameters.

[0253] Step 3 as follows Figure 3 As shown:

[0254] Step 3.1: Basic Adversarial Training Phase: The quantum heuristic conditional generator and the multi-scale extreme perception conditional discriminator are trained iteratively using a conditional generation adversarial objective function.

[0255]

[0256] in, For the value function of adversarial games; This represents the joint distribution of the true photovoltaic power sequence and conditional characteristics. Represents the prior distribution of random noise. This represents the conditional characteristic distribution.

[0257] Step 3.2: Dynamically Weighted Training of Extreme Samples: Constructing the first step using an S-curve Dynamic weights for extreme samples across training rounds:

[0258]

[0259] in, and These represent the minimum and maximum values ​​of the extreme sample weights, respectively. Indicates the slope of the S-curve. This indicates the central training round where the weights of extreme samples increase rapidly.

[0260] No. The sample weights corresponding to each training sample are:

[0261]

[0262] Weighted loss:

[0263]

[0264] Where B represents the training batch size.

[0265] Step 3.3: Focal-CVaR Joint Tail Fine-tuning: Calculate the first tail based on the median of prediction results from multiple scenarios under the same conditions. Prediction error per sample:

[0266]

[0267] Constructing Focal hard sample loss:

[0268]

[0269] Construction of Conditional Risk Value Loss:

[0270]

[0271] in, Represents the prediction error sequence Quantiles , Represents the Focal focusing factor. This represents the error scale parameter.

[0272] Constructing a joint tail loss:

[0273]

[0274] Among them, the dynamic balancing weights are:

[0275]

[0276] Total generator loss:

[0277]

[0278] in, This indicates that the generator is subject to conditional resistance to loss. This represents the tail loss weights that increase with the number of training epochs. This indicates the total number of training rounds.

[0279] Step 4 as follows Figure 4 As shown:

[0280] Step 4.1: Generation of multi-scenario photovoltaic power sequences: For the same standardized time-frequency conditional feature tensor corresponding to the time to be predicted Input to the trained quantum heuristic conditional generator Generate a set of mutually independent standard normal random noise. Photovoltaic power prediction sequence:

[0281]

[0282] Construct a sample set of photovoltaic power prediction distribution:

[0283]

[0284] Step 4.2: Constructing the initial interval for kernel density smoothing quantiles: For the h-th prediction time in the prediction time domain, the power probability density function is calculated using the kernel density estimation method:

[0285]

[0286] The corresponding cumulative distribution function is:

[0287]

[0288] in, Represents the kernel function. Indicates the first The kernel density estimation bandwidth corresponding to each prediction time point Indicates the first The predicted sequence in the 1st Prediction power at each prediction time.

[0289] For nominal confidence level The lower and upper bounds of the initial probability prediction interval are as follows:

[0290]

[0291]

[0292] Step 4.3: Quantum superposition state confidence calibration: Establishing the target confidence level With quantum superposition state probability amplitude Mapping relationship between them:

[0293]

[0294] The interval calibration coefficient is calculated based on the interval coverage deviation on an independent calibration sample set that was not involved in model parameter training:

[0295]

[0296] And the range calibration coefficients are subjected to amplitude limiting processing:

[0297]

[0298] in, This represents the inverse cumulative distribution function of the standard normal distribution; Indicates the interval coverage on an independent calibration sample set or a rolling calibration window; This represents the coverage deviation feedback coefficient; and These represent the lower and upper limits of the interval calibration coefficient, respectively.

[0299] Calculate the first Median of scene predictions at each prediction time:

[0300]

[0301] The lower bound of the calibrated final probability prediction interval is:

[0302]

[0303] The upper bound of the calibrated final probability prediction interval is:

[0304]

[0305] Make the final probability prediction interval satisfy:

[0306]

[0307] The evaluation metrics in step 5 include interval coverage, average interval width, average absolute error, and root mean square error, which comprehensively assess the probabilistic prediction performance from three dimensions: reliability, accuracy, and precision.

[0308] (1) Interval Coverage Rate (PICP)

[0309]

[0310] in, Indicates the number of evaluation samples. Indicates the first The actual photovoltaic power at each evaluation moment; This is an indicator function, taking the value 1 when the condition is met and 0 otherwise; it measures the proportion of the actual power falling within the prediction interval, reflecting the reliability of the probabilistic prediction.

[0311] (2) Interval Average Width (MPIW)

[0312]

[0313] This measures the average width of the prediction interval; a smaller value indicates a more accurate prediction.

[0314] (3) Mean Absolute Error (MAE)

[0315]

[0316] The average absolute deviation between the predicted and actual values ​​of a point is measured, reflecting the accuracy of the prediction.

[0317] (4) Root Mean Square Error (RMSE)

[0318]

[0319] in, Indicates the first The evaluation results are calculated using the median of the photovoltaic power prediction results for multiple scenarios at each evaluation time. The interval coverage rate is used to evaluate the reliability of the probability prediction interval, the average interval width is used to evaluate the sharpness of the probability prediction interval, and the mean absolute error and root mean square error are used to evaluate the accuracy of the point prediction results. Finally, probability prediction intervals and point prediction results at different confidence levels are output, providing data support for photovoltaic power scheduling.

[0320] Experimental run and result verification

[0321] To verify the predictive performance of the model in this invention, ExtraTreesRegressor, Ridge, LGBMRegressor, and XGBRegressor were selected as comparative models, and photovoltaic power prediction experiments were conducted using the same data partitioning method, input features, and prediction time domain. For the extreme operating condition samples selected by the comprehensive extremeity index, the deterministic prediction performance and probabilistic interval prediction performance of each model were evaluated. The root mean square error (RMSE) and mean absolute error (MAE) were used to measure the deterministic prediction error, and the experimental results are shown in Table 1. At a 90% nominal confidence level, the reliability and sharpness of the probabilistic prediction interval were evaluated using prediction interval coverage (PICP) and prediction interval mean width (MPIW), and the experimental results are shown in Table 2. All models were tested under uniform experimental conditions and evaluation criteria to ensure the consistency and comparability of the comparison results.

[0322] Table 1 Comparison of deterministic prediction performance of different model schemes

[0323] Model Scheme RMSE (kW) MAE (kW) ExtraTreesRegressor 45.88 18.37 Ridge Returns 44.54 19.44 LGBMRegressor 44.37 17.13 XGBRegressor 43.10 16.69 This model 41.18 16.56

[0324] Table 2 Comparison of Probability Interval Prediction Performance of Different Model Schemes

[0325] Model Scheme 90% Predicted Interval Coverage Rate (PICP) 90% prediction interval average width MPIW (kW) LGBMRegressor 75.27 72.19 Ridge Returns 84.52 68.23 ExtraTreesRegressor 88.51 72.14 XGBRegressor 90.14 69.85 This model 90.46 67.22

[0326] As shown in Table 1, our model achieves the lowest RMSE and MAE in deterministic prediction, at 41.18 kW and 16.56 kW respectively, both outperforming ExtraTreesRegressor, Ridge, LGBMRegressor, and XGBRegressor. Among these, XGBRegressor, with RMSE and MAE of 43.10 kW and 16.69 kW respectively, is a better performing model compared to the others, but still lower than the prediction accuracy of our model. The results indicate that our model can more effectively learn the nonlinear variation of photovoltaic power and reduce point prediction errors under extreme conditions.

[0327] As shown in Table 2, the PICP (Predicted Interval Coverage Probability) of this model at a 90% nominal confidence level is 90.46%, achieving the preset coverage level. Simultaneously, its average predicted interval width (MPIW) is 67.22 kW, the smallest among all models. In contrast, while XGBRegressor achieves a coverage rate close to 90%, its average interval width is 69.85 kW. Overall, this model effectively reduces interval width and redundancy while maintaining high coverage reliability, achieving a good balance between probabilistic prediction reliability and interval sharpness.

[0328] Furthermore, to intuitively verify the photovoltaic power prediction effect of the model of this invention, a visual analysis was performed on the point prediction results, probability prediction intervals, the correlation between predicted and actual values, and the prediction results of extreme operating condition samples, such as... Figures 6-9 As shown.

[0329] like Figure 6 As shown, the predicted power of this model is basically consistent with the actual power change trend, and it can track the rise, fall and local fluctuation of photovoltaic power well, indicating that the model has good point prediction capability.

[0330] like Figure 7 As shown, the 90% probability prediction interval constructed by this model can cover most of the actual power observations, and the interval width can be adjusted according to the degree of power fluctuation, demonstrating good interval reliability and sharpness.

[0331] like Figure 8 As shown, the scatter plots of the predicted and actual values ​​are mainly distributed near the ideal fitting line, indicating that the prediction results of this model have good consistency with the actual power.

[0332] like Figure 9As shown, under the extreme operating condition samples obtained by screening with the comprehensive extreme degree index, the model can still track the actual power change well, indicating that the proposed method has a certain ability to predict and characterize extreme power fluctuations.

[0333] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for probabilistic prediction of photovoltaic power in extreme weather scenarios, characterized in that, include: Step 1: Collect meteorological data, unit operation data and grid connection data of photovoltaic power station, screen extreme weather samples, and perform abnormal data detection, missing data interpolation, variational mode decomposition, feature normalization and sliding time window construction on the screened samples in sequence to obtain standardized time-frequency conditional feature tensors; Step 2: Construct a conditional generative adversarial network including a quantum heuristic conditional generator and a multi-scale extreme perception conditional discriminator. Input the standardized time-frequency conditional feature tensor and random noise into the quantum heuristic conditional generator to output a photovoltaic power prediction sequence. The multi-scale extreme perception conditional discriminator then distinguishes between the real power sequence and the photovoltaic power prediction sequence. Step 3: Optimize the parameters of the quantum heuristic condition generator and the multi-scale extreme perception condition discriminator through three stages in sequence: basic condition adversarial training, extreme sample dynamic weighted training, and Focal-CVaR joint tail fine-tuning. Step 4: Generate multiple photovoltaic power prediction sequences based on the same standardized time-frequency conditional feature tensor, construct an initial probability prediction interval by kernel density smoothing quantile estimation, and then perform confidence calibration on the initial probability prediction interval according to the interval coverage deviation of the calibration sample to obtain photovoltaic power probability prediction intervals at different confidence levels. Step 5: Take the median of multiple photovoltaic power prediction sequences as the point prediction result, output the point prediction result and the photovoltaic power probability prediction interval under different confidence levels, and evaluate the prediction effect through evaluation indicators.

2. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Collect meteorological data, generator operation data, and grid connection data with a sampling resolution of 15 minutes. After aligning them with a unified timestamp, the raw dataset is obtained. The meteorological data includes irradiance, ambient temperature, wind speed, humidity, and cloud cover. The generator operation data includes photovoltaic module output power, module temperature, and inverter efficiency. The grid connection data includes grid connection point voltage, active power, and reactive power. Step 1.2: Calculate the power fluctuation rate at each sampling time. and irradiance mutation rate Construct a comprehensive extreme index ,in, for The comprehensive extreme index at each sampling time, , These are the median of the total sequence power volatility and the irradiance mutation rate, respectively. , The weighting coefficients of the two indicators satisfy the following conditions: ,Will Greater than the preset quantile threshold The samples were used as candidate samples for extreme operating conditions. After excluding power mutations caused by non-meteorological factors by combining changes in meteorological elements, inverter operating status, grid connection status and irradiance-power response relationship, an extreme weather sample set was obtained. Step 1.3: Perform improved IQR-3σ two-layer anomaly detection on the extreme weather sample set: calculate the global IQR deviation. Deviation from local sliding window 3σ The overall anomaly confidence level is obtained by weighted summation. , As the global deviation weight, Samples exceeding the preset anomaly threshold are marked as statistically abnormal samples. Then, the extreme weather samples undergo physical consistency verification. Only when a sample meets both the statistical anomaly and physical anomaly conditions is it determined to be invalid abnormal data and removed. The remaining samples are retained as valid extreme weather samples. Step 1.4: For the missing data generated after removing invalid outlier data, interpolation is performed to complete the missing data using spatial distance weight, temporal proximity weight, and meteorological correlation weight. Step 1.5: Perform variational mode decomposition on the irradiance time series and photovoltaic output power time series to obtain multiple mode components. Calculate the mode effectiveness index based on the energy proportion and kurtosis of each mode component. ,in, For the first The effectiveness index of each modal component , Let be the weighting coefficient for energy percentage and kurtosis, and satisfy: , The energy percentage of the k-th modal component. Let be the kurtosis value of the k-th modal component. For reference kurtosis value, The maximum deviation of all modal kurtosis from the reference kurtosis is retained. Not lower than the preset threshold The modal components are extracted, and the time-domain statistical features and frequency-domain energy features of the modal components are retained. These features are then concatenated with the original meteorological and operational features to obtain a time-frequency fusion feature set. Step 1.6: After performing max-min normalization on the time-frequency fusion feature set, construct time-series input samples through a sliding time window to obtain the standardized time-frequency conditional feature tensor. , where N is the number of samples, L is the sliding window length, and D is the dimension of the enhanced features.

3. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Convert the standardized time-frequency conditional feature tensor... The input conditional feature encoding layer yields the conditional feature vector. ,Will After concatenating with the standard normal random noise vector z, the first... Initial rotation angle of a quantum heuristic bit Constructing the initial quantum state ,in, For the initial quantum state, The activation function for Sigmoid. This represents the concatenation result of a random noise vector and a conditional eigenvector. For tensor product operations, For the first The initial rotation angle of each qubit; , For trainable weights and biases; Step 2.2, based on Calculate the weighting coefficients of the superposition state Constructing a conditionally weighted superposition state ,in, It is a weighted superposition state; To calculate the ground state, a total of 2 n indivual; This represents the original probability amplitude of the ground state. These are the ground-state weighting coefficients adaptively calculated based on the conditional feature encoding results; Step 2.3, Construction Layer-parameterized quantum heuristic transformation units, each layer including a single-bit Y-axis rotation gate. Mutual information adaptive parameterized controlled rotational entanglement gate ,in, , For trainable rotation parameters, Indicates the action on the first The unit operator of the nth qubit, whose function is to maintain the nth qubit The state of each qubit remains unchanged. and They represent the first Each qubit is used to calculate the ground state. and The projection operator on the control bit is used to determine which ground state component the control bit is in. Represents a single-qubit gate that rotates about the Y-axis of a Bloch sphere. Indicates the first The conditional feature and the first Mutual information between conditional features Indicates the first Layer entanglement strength coefficient, κ is the scaling factor, after... The final quantum heuristic state is obtained after the layer transformation. , Indicates the first The evolution matrix corresponding to the layer parameterized transformation unit; Step 2.4: Measure the pre-defined observable of the final quantum heuristic state to obtain the measurement feature vector. , , Indicates the action on the first A measurement operator for each quantum heuristic bit state is input into a fully connected decoding layer to obtain a photovoltaic power prediction sequence of length H. ,in Indicates the rated power of the photovoltaic power station; and This represents the trainable parameters of the fully connected decoding layer; Step 2.5, convert the actual power sequence P or the photovoltaic power prediction sequence. With the conditional feature vector The input for discrimination is obtained by concatenation. Multiple one-dimensional dilated convolutional branches with different dilation rates are used to extract multi-timescale features. After calculating the fusion weights of the features from each branch through a trainable attention mechanism, the multi-scale features are fused. The fused features are then input into the discriminative output layer to obtain the probability of sample authenticity. ,in For multi-scale fusion features, , These are trainable parameters.

4. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 3, characterized in that, Step 3 specifically includes: Step 3.1, Basic Conditional Adversarial Training Phase: Using Conditional Generative Adversarial Objective Functions The generator and discriminator parameters are optimized alternately and iteratively, where To counter the game value function, Represents the joint distribution of real data Calculate the mathematical expectation of the sample. This represents the joint distribution of the true photovoltaic power sequence and conditional characteristics. This represents the log-likelihood term of the discriminator with respect to the real samples. Represents the prior distributions of random noise, respectively. and conditional feature distribution Sampling is performed on the generated samples, and the expected value of the discrimination results is calculated. This represents the log-likelihood term of the discriminator for the generated samples. This indicates that the discriminator is based on conditional features. The following will generate a photovoltaic power sequence. The probability of being identified as a real sample. Represents the prior distribution of random noise. Represents the conditional characteristic distribution; Step 3.2, Extreme Sample Dynamic Weighting Training Phase: The dynamic weights of the extreme samples in the r-th training round are calculated using an S-curve. ,in and These represent the minimum and maximum values ​​of the extreme sample weights, respectively. Indicates the slope of the S-curve. The central training round represents the period where the weights of extreme samples increase rapidly, and the i-th training sample is assigned a sample weight. : ,based on Calculate the weighted discriminant loss B is the training batch size. This represents the log-likelihood of the discriminator with respect to the real samples. For the first A real photovoltaic power sequence, To and Corresponding conditional features, This indicates that the discriminator is based on conditional features. Below, the actual photovoltaic power sequence The probability of being identified as a real sample. This term represents the log-likelihood of the discriminator with respect to the generated samples. This indicates that the discriminator is based on conditional features. The probability of determining the generated power sequence as a real sample is then calculated. For the generator based on random noise and conditional features The generated photovoltaic power sequence; Step 3.3, Focal-CVaR Joint Tail Fine-tuning Stage: Calculate the... Prediction error per sample ,in, For the first A real photovoltaic power sequence, Indicates features under the same conditions Next, input to the generator Groups of independent random noise ,get The photovoltaic power generation results are then used to take the median of these results as the nth photovoltaic power generation result. The point prediction values ​​for each sample are used to construct the Focal hard sample loss. ,in Represents the Focal focusing factor. Represents the error scale parameter; constructs the conditional value at risk loss. ,in Represents the prediction error sequence Quantiles This indicates the confidence level or tail percentile level of CVaR. ; Construct a joint tail loss The dynamic balance weight , Indicates the total number of training rounds. This represents the maximum value of the dynamic weight. The minimum value of the dynamic weights; the total generator loss is ,in This indicates that the generator is subject to conditional resistance to loss. This represents the tail loss weights that increase with training epochs, based on Optimize generator parameters.

5. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Standardize the time-frequency conditional feature tensor for the time to be predicted. Input S sets of mutually independent standard normal random noise into the trained quantum heuristic conditional generator to generate S photovoltaic power prediction sequences. The photovoltaic power prediction distribution sample set was obtained. ; Step 4.2: For the first [item] in the prediction time domain... At each prediction time, the power probability density function is calculated using kernel density estimation. ,in Represents the kernel function. Let the value of the photovoltaic power random variable be denoted as . Indicates the first The kernel density estimation bandwidth corresponding to each prediction time point Indicates the first The predicted sequence in the 1st The predicted power at each prediction time point corresponds to the cumulative distribution function as follows: Regarding nominal confidence levels , Let represent the integral variable, and the lower bound of the initial probability prediction interval is . The upper boundary is , Indicates the first Cumulative distribution function at each prediction time The generalized inverse function of is also called the quantile function; Step 4.3: Calculate the interval calibration coefficient based on the interval coverage deviation of the independent calibration sample set. ,in This represents the inverse cumulative distribution function of the standard normal distribution. For the target confidence level, Indicates the interval coverage on an independent calibration sample set or a rolling calibration window; This represents the coverage deviation feedback coefficient, for After applying amplitude limiting, calculate the median of the scene prediction at the h-th prediction time. The lower bound of the calibrated final probability prediction interval is The upper boundary is ,satisfy .

6. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 2, characterized in that, In step 1.2, the power fluctuations caused by non-meteorological factors excluded include non-meteorological power fluctuations caused by equipment failure, power rationing, communication anomalies, and shutdowns.

7. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 2, characterized in that, In step 1.3, the criteria for physical consistency verification include photovoltaic power rated boundary, irradiance-power consistency, power continuity between adjacent time periods, and equipment operating status.

8. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 5, characterized in that, In step 4.3, the interval calibration coefficient is... The expression for limiting is: ,in and These represent the lower and upper limits of the interval calibration coefficient, respectively. This represents the truncation / limiting function.

9. The photovoltaic power probability prediction method for extreme weather scenarios according to claim 1, characterized in that, In step 5, the evaluation indicators include interval coverage, average interval width, average absolute error, and root mean square error.

10. A photovoltaic power probability prediction system for extreme weather scenarios, executing the method as described in any one of claims 1-9, characterized in that, include: The multi-scale time-frequency coupled data preprocessing module is used to collect meteorological data, unit operation data and grid connection data of photovoltaic power plants and align them with a unified timestamp. It filters extreme weather samples and sequentially performs improved IQR-3σ double-layer anomaly detection, meteorological neighborhood spatiotemporal weighted interpolation, variational mode decomposition, feature normalization and sliding time window construction, and outputs a standardized time-frequency conditional feature tensor. The quantum superposition-entanglement cooperative conditional generative adversarial network module includes a quantum heuristic conditional generator and a multi-scale extreme perception conditional discriminator. The quantum heuristic conditional generator is used to take random noise and the standardized time-frequency conditional feature tensor as input, and generate a photovoltaic power prediction sequence through conditional feature encoding, weighted superposition state mapping, rotation-entanglement cooperative evolution and measurement decoding. The multi-scale extreme perception conditional discriminator is used to fuse the conditional features with the real power sequence and the photovoltaic power prediction sequence respectively, and output the sample authenticity probability through multi-scale dilated convolution and attention fusion. The dynamic dual-weighted tail-aware training module is used to sequentially perform basic conditional adversarial training, extreme sample dynamic weighted training, and Focal-CVaR joint tail fine-tuning to adjust the parameters of the quantum heuristic condition generator and the multi-scale extreme-aware conditional discriminator. The probability prediction interval generation and confidence calibration module is used to generate multiple photovoltaic power prediction sequences based on the same standardized time-frequency conditional feature tensor. It constructs an initial probability prediction interval by kernel density smoothing quantile estimation, calculates the interval calibration coefficient based on the interval coverage deviation of the calibration sample, calibrates the initial probability prediction interval, and outputs photovoltaic power probability prediction intervals at different confidence levels. The prediction result output module is used to take the median of the multiple photovoltaic power prediction sequences as the point prediction result, and output the point prediction result and the photovoltaic power probability prediction interval under different confidence levels.

Images