High-resolution meteorological element reconstruction method, photovoltaic power prediction method and related devices

Abstract

The invention discloses a high-resolution meteorological element reconstruction method, a photovoltaic power prediction method and related devices, and belongs to the technical field of meteorological data reconstruction and photovoltaic prediction, and the method comprises the steps: employing ERA5 meteorological data as input, employing NSRDB meteorological data as a supervision target, and carrying out the iterative training of a quantum-classical hybrid model till the model converges; and inputting the ERA5 meteorological data of the target period into the trained model to obtain 10-minute-level meteorological data of the target period. The potential of quantum calculation in the aspects of feature enhancement, complex regression and global optimization is mined through three different paths of a quantum enhancement module, a classical learning module ensures that final regression output is stable and reliable, and aggregation consistency loss constraints are introduced into a training objective function of a model to meet the law of conservation of energy. And introducing shape regularization constraint to enable the output meteorological data to accord with the continuous smooth characteristic of the atmospheric physical process, and generating high-resolution meteorological data with physical consistency and good generalization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological data reconstruction and photovoltaic prediction, and in particular to a high-resolution meteorological element reconstruction method, a photovoltaic power prediction method and related devices. BACKGROUND

[0002] In the field of photovoltaic power prediction, in order to obtain high temporal resolution meteorological driving data, two databases are commonly used in engineering: ERA5 reanalysis data (hourly, 1h) and the United States National Solar Radiation Database NSRDB (10-minute level). NSRDB-10min provides 10-minute resolution ground meteorological true value data from 2016 to 2020, but lacks equivalent resolution public ground true value after 2020. Therefore, for the target period of 2025 and beyond, only hourly ERA5 data can be obtained, and corresponding 10-minute level meteorological true value cannot be obtained for direct supervised training. This brings challenges to down-sampling of high-resolution meteorological elements across decades (such as generalizing the model of 2016-2020 to 2025).

[0003] Although traditional statistical interpolation or deep learning down-sampling methods can improve the temporal resolution to some extent, when facing strong non-linear relationships and cross-decade data distribution drift, two prominent problems often arise: 1. Energy non-conservation: simple down-sampling may result in an inconsistency between the hourly cumulative value and the sum of the six 10-minute segments. This means that the hourly value does not remain consistent in terms of physical quantities after subdivision, violating the energy conservation and other physical constraints.

[0004] 2. Non-physical high-frequency oscillation: unconstrained high-resolution prediction may exhibit dramatic 10-minute level fluctuations (such as unreasonable peaks or sudden drops), i.e., high-frequency oscillation phenomena that do not conform to actual physical processes. This is particularly evident when there is a strong non-linear relationship or a decade shift in data distribution.

[0005] The above defects make it difficult for existing down-sampling solutions to reliably generate physically consistent 10-minute level meteorological data under the stringent condition of "training using 2016-2020 data and applying in 2025 with a lack of high-resolution true value after 2020". SUMMARY

[0006] The present application aims to overcome the problems of the prior art and provides a high-resolution meteorological element reconstruction method, a photovoltaic power prediction method and related devices.

[0007] The purpose of the present application is achieved by the following technical solution: a high-resolution meteorological element reconstruction method, comprising the following steps: The ERA5 hourly meteorological data and NSRDB 10-minute meteorological data of the historical period are obtained and preprocessed, and the preprocessing includes data alignment processing; The preprocessed ERA5 hourly meteorological data is used as an input sample, and the time-aligned NSRDB 10-minute meteorological data is used as a supervised target to iteratively train the quantum-classical hybrid model until the quantum-classical hybrid model converges; the training objective function of the quantum-classical hybrid model introduces an aggregated consistency loss constraint and / or a shape regularization constraint, the aggregated consistency loss is used to measure the deviation of the sum of 6 10-minute meteorological values within each hour from the corresponding ERA5 hourly meteorological value, and the shape regularization constraint is used to measure the deviation of the smoothness of the 10-minute sequence output by the model; the quantum-classical hybrid model includes a quantum enhancement module and a classical learning module; the quantum enhancement module performs feature enhancement processing on the input sample based on quantum feature mapping and quantum kernel estimation, or performs regression processing on the input sample based on a variational quantum circuit, or performs constraint optimization processing on the 10-minute preliminary meteorological prediction data output by the classical learning module based on a quadratic unconstrained binary optimization model, and correspondingly outputs enhanced features or prediction distribution parameters of 10-minute meteorological data or optimized 10-minute meteorological prediction data; the classical learning module generates 10-minute meteorological prediction data according to the enhanced features or the prediction distribution parameters, or the classical learning module outputs 10-minute preliminary meteorological prediction data according to the input sample; The ERA5 hourly meteorological data of the target period is input into the trained quantum-classical hybrid model to output 10-minute meteorological data of the target period.

[0008] In an example, the feature enhancement processing includes encoding the input sample into a quantum state using a parameterized quantum circuit, measuring the quantum state to obtain a quantum kernel matrix, and inputting the quantum kernel matrix as enhanced features into the classical learning module; The regression processing includes outputting prediction distribution parameters of 10-minute meteorological data according to the input sample; wherein the training objective function includes an aggregated consistency loss and a shape regularization loss, and the parameters of the variational quantum circuit and the classical learning module are optimized by a gradient descent method; The constraint optimization processing includes converting the 10-minute preliminary meteorological prediction data and the aggregated consistency constraint into a quadratic unconstrained binary optimization problem, and solving or approximately solving the quadratic unconstrained binary optimization problem using a quantum annealing algorithm or a variational quantum optimization algorithm, so as to optimize the 10-minute preliminary meteorological prediction data output by the classical learning module into 10-minute meteorological prediction data satisfying the aggregated consistency constraint.

[0009] In an example, the classical learning module is a multi-output regression model or a sequence-to-sequence neural network model.

[0010] In an example, the quantum-enhanced module is executed on a physical quantum processor, or by a quantum circuit simulator running on a classical computer, or by a quantum-inspired classical algorithm; the quantum circuit simulator simulates quantum feature mapping using a tensor network or a state vector, solves a quadratic unconstrained binary optimization model using a simulated annealing algorithm; when the quantum-enhanced module is executed by a quantum-inspired classical algorithm, a quantum-inspired classical kernel function is used to implement quantum feature mapping, and a simulated annealing algorithm is used to solve a quadratic unconstrained binary optimization model.

[0011] In an example, the method further comprises a quantile adjustment step: A quantile mapping function is constructed to convert the 10-minute meteorological data output by the quantum-classical hybrid model to obtain corrected 10-minute meteorological data.

[0012] It should be further pointed out that the technical features of the above-mentioned examples of the method can be combined or replaced to form new technical solutions.

[0013] The application also includes a photovoltaic power prediction method, which is implemented based on the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints formed by any one or more combinations, comprising the following steps: The 10-minute meteorological data is aligned with the historical photovoltaic power data, and the aligned historical photovoltaic power data is used to train the power prediction model, and a mask weighted supervision term and an hour window consistency regularization term are added in the training loss; the mask weighted supervision term is defined based on a solar elevation angle or an irradiance threshold, and is used to mark invalid data points; the hour window consistency regularization term is used to ensure that the predicted 10-minute power is consistent with the hour-level power reference value on an hourly scale; The 10-minute meteorological data of the target period is input into the trained power prediction model to obtain 10-minute photovoltaic power prediction values.

[0014] In an example, the photovoltaic power prediction method further comprises correcting the 10-minute photovoltaic power prediction values, comprising: According to the actual power observation data of the last K hours, the 10-minute photovoltaic power prediction values output by the power prediction model are online re-calibrated, and the parameters of the re-calibration are dynamically updated by a least squares regression with a forgetting factor to correct the prediction bias in real time; and / or, According to the performance decay rate of the photovoltaic module, the temperature difference between the historical period and the target period, a periodic adjustment loss function of the power prediction model is constructed, and the 10-minute photovoltaic power prediction values are corrected by using the periodic adjustment loss function.

[0015] It should be further explained that the technical features corresponding to each of the above methods can be combined or replaced to form new technical solutions.

[0016] The application also includes a computer program product comprising a computer program which, when executed by a processor, implements the steps of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the steps of the photovoltaic power prediction method formed by any one of the above examples or a combination of multiple examples.

[0017] The application also includes a storage medium having computer instructions stored thereon, which, when executed, perform the steps of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the steps of the photovoltaic power prediction method formed by any one of the above examples or a combination of multiple examples.

[0018] The application also includes a terminal comprising a memory and a processor, the memory having computer instructions stored thereon executable on the processor, and the processor executing the computer instructions to perform the steps of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the steps of the photovoltaic power prediction method formed by any one of the above examples or a combination of multiple examples.

[0019] Compared with the prior art, the application has the following advantages: 1. In an example, the method of the application introduces a quantum-classical hybrid model, and the quantum-enhanced module mines the potential of quantum computing in feature enhancement, complex regression, and global optimization through three different paths, and the classical learning module ensures that the final regression output is stable and reliable. At the same time, an aggregated consistency loss constraint is introduced to satisfy the law of conservation of energy, and a shape regularization constraint is introduced to make the output 10-minute meteorological data conform to the continuous smoothness of atmospheric physical processes, generating 10-minute meteorological data with physical consistency and good generalization performance, providing a high-quality data basis for downstream applications such as photovoltaic power prediction.

[0020] 2. In an example, high-dimensional nonlinear features are provided through a quantum kernel matrix, which can capture more complex hourly to minute-level mapping relationships, enhance the expression ability and learning efficiency of the classical model, and improve the generalization performance across different time domains; a variational quantum circuit is used to learn complex input-output mappings and is trained together with physical constraints to ensure that its output is both accurate and physically reasonable; the aggregated consistency constraint is converted into a strict mathematical optimization problem, which can ensure that the output results satisfy the aggregated consistency, and thus ensure that the output 10-minute meteorological data satisfies the physical laws.

[0021] 3. In an example, the quantum enhancement module supports physical quantum processors, classical simulators, and quantum-inspired classical algorithms, ensuring the compatibility and implementability of the method of the present application on existing classical circuits and future quantum computing hardware, greatly improving the practical value and landing prospects of the technology.

[0022] 4. In an example, the statistical distribution deviation of the model prediction value is corrected by the quantile mapping method, especially solving the distribution drift problem between historical training data and future application data caused by climate change or system error, improving the robustness of long-term prediction.

[0023] 5. In an example, mask weighting is introduced in the power model training, effectively focusing on the effective period of power generation, improving the training efficiency and accuracy; the hour window consistency regularization term is introduced to ensure that the cumulative amount of high-frequency power prediction also remains reasonable on a longer time scale.

[0024] 6. In an example, through online recalibration processing, the systematic deviation of the model can be corrected in real time and adaptively, and it can quickly respond to short-term changes in the state of the power station (such as dust shielding, sensor drift); through the age domain correction, considering the component attenuation and temperature difference, actively compensating for the long-term performance drift caused by equipment aging and environmental changes, ensuring the accuracy of cross-epoch prediction. BRIEF DESCRIPTION OF DRAWINGS

[0025] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings, which are provided to provide further understanding of the present application, and form a part of the present application. The same reference numerals are used to represent the same or similar parts in the accompanying drawings. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application.

[0026] Figure 1 A high-resolution meteorological element reconstruction method flowchart is provided for an example of the present application; Figure 2 A photovoltaic power prediction method flowchart is provided for an example of the present application; Figure 3 A photovoltaic power prediction method flowchart based on quantum-enhanced constraints is provided for an example of the present application. DETAILED DESCRIPTION

[0027] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0028] Furthermore, the technical features involved in the different embodiments of the application described below can be combined with each other as long as they do not conflict with each other.

[0029] In an example, as shown in FIG. 1, a high-resolution meteorological element reconstruction method based on quantum-enhanced constraints comprises the following steps: Figure 1 S10: Obtain and preprocess ERA5 hourly meteorological data and NSRDB 10-minute meteorological data for a historical period (e.g., 2016-2020), which includes data alignment processing. S10: Obtain and preprocess ERA5 hourly meteorological data and NSRDB 10-minute meteorological data for a historical period (e.g., 2016-2020), which includes data alignment processing.

[0030] Among them, the ERA5 hourly meteorological data comes from the ERA5 reanalysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF), with a time resolution of 1 hour, containing meteorological elements affecting photovoltaic power generation such as global horizontal irradiance (GHI), air temperature, humidity, etc. In this example, the ERA5 hourly data corresponding to the site or region from 2016 to 2020 is collected as the input feature sequence.

[0031] The NSRDB 10-minute meteorological data comes from ground observation data of the National Solar Radiation Database (NSRDB), with a time resolution of 10 minutes, currently providing high-frequency meteorological true values (especially irradiance-related data) from 2016 to 2020. In this example, the 10-minute data matching ERA5 during this period is used as the target output sequence for training.

[0032] In data preprocessing, ERA5 and NSRDB data need to be aligned: for example, by timestamp matching to associate ERA5 hourly records with corresponding 6 NSRDB 10-minute records, and eliminating missing values or outliers (which can be marked by setting a mask).

[0033] Preferably, during data preprocessing, derived features such as clearness index can be calculated for subsequent quantile mapping adaptation steps. The clearness index is usually defined as the ratio of actual irradiance to irradiance under theoretical clear sky conditions, which can represent the degree of cloud cover and is a commonly used standardized index for comparing irradiance distribution across regions or time periods.

[0034] S20: using the pre-processed ERA5 hourly meteorological data as input samples and the time-aligned NSRDB 10-minute meteorological data as supervised targets, iteratively training the quantum-classical hybrid model until the quantum-classical hybrid model converges; the training objective function of the quantum-classical hybrid model introduces an aggregated consistency loss constraint and / or a shape regularization constraint, the aggregated consistency loss is used to measure the deviation of the sum of the 6 10-minute meteorological values in each hour from the corresponding ERA5 hourly meteorological value, and the shape regularization constraint is used to measure the deviation of the smoothness of the 10-minute sequence output by the model; the quantum-classical hybrid model includes a quantum feature enhancement module (QFE) and a classical learning module; the quantum feature enhancement module performs feature enhancement processing on the input samples based on quantum feature mapping and quantum kernel estimation, or performs regression processing on the input samples based on a variational quantum circuit, or performs constraint optimization processing on the 10-minute preliminary meteorological prediction data output by the classical learning module based on a quadratic unconstrained binary optimization model, and correspondingly outputs enhanced features or prediction distribution parameters of 10-minute meteorological data or optimized 10-minute meteorological prediction data; the classical learning module generates 10-minute meteorological prediction data according to the enhanced features or the prediction distribution parameters, or the classical learning module outputs 10-minute preliminary meteorological prediction data according to the input samples.

[0035] Specifically, in the training phase, each corresponding ERA5 hourly sample and its 6 NSRDB 10-minute true value samples are input into the quantum-classical hybrid model, and the parameters of the quantum enhancement module and the classical learning module are iteratively optimized. The loss function is composed of three parts: prediction error loss (measuring the deviation of the 10-minute meteorological prediction data output by the model from the NSRDB true value, such as mean square error), aggregated consistency loss (measuring the difference between the sum of the 6 slices per hour and the ERA5 hourly value), and shape regularization loss (measuring the deviation of the smoothness of the output sequence). In this example, the total loss function of the quantum-classical hybrid model can be represented as: wherein is the prediction error loss (such as MSE); is the aggregated consistency loss; is the shape regularization loss, and are weight coefficients.

[0036] The aggregated consistency constraint is used to ensure that the sum of the 6 10-minute meteorological values output by the model is equal to the original 1-hour meteorological value, realizing the energy conservation between the hour ↔ 10-minute granularity. To this end, an aggregated consistency loss is introduced to punish the deviation of the output slice from the hourly value in training. The calculation expression of the aggregated consistency constraint is: where, denotes the aggregated consistency loss value; denotes the total number of samples in the training batch (Batch Size); denotes the sample index; denotes the time step index, denoting the -th 10-minute slice within an hour (value range ); denotes the meteorological prediction value (e.g., radiation intensity) of the -th sample output by the model at the -th 10-minute time point; denotes the original ERA5 hourly meteorological true value corresponding to the -th sample (as the sum benchmark).

[0037] Meanwhile, the present application converts this global consistency constraint into a quadratic unconstrained binary optimization (QUBO) problem, and solves or approximately solves it with the help of quantum annealing algorithms or variational quantum optimization algorithms (such as QAOA, VQE) to obtain the optimal or approximately optimal slice allocation that satisfies the constraint. This quantum optimization strategy can efficiently search the solution space that satisfies the global constraint, thereby globally optimizing the slice allocation scheme within each hour. Shape regularization is used to constrain the smoothness and physical reasonableness of the 10-minute output sequence (10-minute meteorological prediction data), avoiding sharp oscillations. Specifically, second-order difference total variation (i.e., minimizing the fluctuation amplitude of adjacent slice increments) or slope monotonicity / upper bound constraints can be used to regularize the smoothness and monotonicity of the output curve to limit the amplitude of high-frequency oscillations. Shape regularization constraints are used to ensure that short-time sequence changes conform to actual physical processes (e.g., cloud cover changes will not oscillate sharply within 10 minutes beyond a reasonable range). In this example, the shape regularization constraint expression is: denotes the shape regularization loss value; denotes the meteorological prediction value of the current time point (the i -th 10-minute) output by the model; denotes the meteorological prediction values of the next time point and the previous time point, respectively, output by the model; denotes the second-order difference of the sequence, which corresponds to the fluctuation amplitude of the rate of change in a physical sense (i.e., the acceleration term); denotes the L1 norm (absolute value), which is used to punish sharp oscillations in any direction and ensure curve smoothness.

[0038] Further, the quantum feature enhancement module QFE can complete this function through one of the three implementation forms (hereinafter referred to as QFE-1, QFE-2, QFE-3): QFE-1 is quantum feature mapping and quantum kernel estimation, including a parameterized quantum state preparation circuit and a quantum state measurement submodule connected in sequence. The parameterized quantum state preparation circuit encodes the input sample (such as several meteorological elements of the current hour, such as temperature, humidity, clear sky index, hourly GHI, and necessary context (such as previous time trend, cloud stability index, etc.)) into a quantum state vector, and then performs measurement sampling on the quantum state through the quantum state measurement submodule to estimate the quantum kernel matrix (10-minute scale weather change information). The quantum kernel essentially measures the similarity of different input samples in the quantum high-dimensional Hilbert space. During training, the calculated kernel matrix can be provided to the classical learner C to model the output. By selecting an appropriate quantum feature mapping (such as a feature encoding based on a parameterized rotation gate), QFE-1 can introduce more rich nonlinear features than traditional kernel methods, i.e.: the quantum feature mapping / kernel method uses quantum state high-dimensional embedding and quantum kernel calculation, and the model can capture more complex hour-to-minute mapping relationships, while using the expressive power of the quantum state space to improve the generalization ability of the model across the years. This implementation only needs quantum circuits to perform state space mapping and does not directly output prediction values, which is suitable for providing high-dimensional features before a classical regressor. When QFE-1 and the classical learning module form a quantum-classical hybrid model, the objective function of the classical learning module introduces an aggregated consistency loss and a shape regularization loss, and minimizes the prediction error of the classical learning module, and outputs a prediction sequence (10-minute meteorological prediction data) that satisfies the constraints as much as possible.

[0039] QFE-2 is a variational quantum circuit (VQC), including a parameterized quantum state preparation circuit, a parameterized quantum evolution circuit and a quantum state measurement submodule connected in sequence. By designing a variational quantum circuit, the distribution parameters (such as mean, standard deviation, etc.) of 6 10-minute slices corresponding to each hour are directly output. For example, the quantum circuit takes the current hourly meteorological data as input, encodes the quantum gate of the parameterized quantum state preparation circuit into the initial quantum state, evolves through the parameterized quantum gate of the parameterized quantum evolution circuit, and finally measures the statistical quantities (such as expected value, quantile, etc.) related to the 6 slices by the quantum state measurement submodule. These statistical quantities can be used as the prediction of the 6 slice values or the description of their distribution patterns after appropriate scaling. In the training process, the parameter shift method is used to efficiently calculate the gradient of the quantum circuit with respect to the loss function, and the prediction error, the aggregation consistency constraint and the shape regularization are jointly applied to the quantum circuit parameters, and the parameters of the variational quantum circuit and the classical learning module are optimized using a classical optimization algorithm (such as gradient descent). QFE-2 can fully utilize the ability of quantum circuit to express complex functions, directly fit the output, and integrate physical constraints in training to ensure physical consistency of the output.

[0040] QFE-3 is a constrained QUBO optimization, including a QUBO problem mapping submodule and a quantum annealer. The QUBO problem mapping submodule converts the 10-minute level preliminary meteorological prediction data and the aggregation consistency constraint into a quadratic unconstrained binary optimization problem, corresponding to an energy allocation problem of 6 slices per hour (quadratic unconstrained binary optimization QUBO problem), whose objective function is where, the transpose of the decision vector, denotes the QUBO matrix, and then the quantum annealing or variational quantum optimization algorithm is used to find the slice allocation vector that minimizes the objective function In particular, the total amount per hour (e.g. total irradiance energy) is discretized into several portions with certain resolution, and how to allocate these portions to the 6 sub-periods is formulated as a combinatorial optimization problem, with the constraints (hourly sum fixed) embedded in the objective function or penalty term of QUBO, using quantum annealing computers or variational quantum optimization algorithms to efficiently find (or approximately find) the optimal slicing allocation scheme satisfying the constraints in a large search space. QFE-3 does not directly deal with the nonlinear mapping of input features, but focuses on optimizing the allocation of slices at a global level to satisfy physical constraints and optimize certain objectives (e.g. to make the slice distribution as smooth as possible or conform to shape priors). During the model training process, QFE-3 works with the classical prediction output: the classical learning module first gives a preliminary slice prediction, which is then globally optimized by the QFE-3 module to output an improved solution that satisfies the aggregation consistency, and thus obtain the optimized 10-minute-level weather prediction data. When QFE-3 and the classical learning module form a quantum-classical hybrid model, the objective function of the classical learning module introduces an aggregation consistency loss and a shape regularization loss, and minimizes the prediction error of the classical learning module to output a preliminary prediction sequence that satisfies the constraints as much as possible. At the same time, QFE-3 converts the preliminary prediction sequence and the aggregation consistency constraint into a quadratic unconstrained binary optimization problem, and uses quantum annealing computers or variational quantum optimization algorithms to solve it, finding a final sequence that is closest to the preliminary prediction sequence and strictly satisfies the aggregation consistency constraint, i.e. outputting the optimized 10-minute-level weather prediction data.

[0041] Preferably, QFE-3 is integrated with a quantum random number generator for random sampling, data augmentation, and other steps in constraint solving. Utilizing quantum randomness can improve the diversity of global optimization search and the robustness of the model. Similarly, when no quantum hardware is available, a high-quality classical random source can be used instead of a quantum random number generator.

[0042] In each training iteration, if QFE-1 or QFE-2 is used, the implementation itself can obtain gradients through standard backpropagation or parameter shift method, directly updating the parameters of the quantum enhancement module and the classical learning module; if QFE-3 is used, for the output of the current classical learning module, the improved slice is obtained by solving QUBO through quantum annealing or simulated annealing, then the loss is calculated by comparing the improved result with the true value, and the classical learning module parameters are adjusted accordingly (QFE-3 itself has no learnable parameters, but it guides the classical learning module to optimize in the direction of satisfying the constraints). This cycle continues until the loss converges. After training is complete, the final determined model parameters are obtained, including the configuration of the QFE module (if QFE-1, the parameters of the quantum feature mapping are determined, QFE-2 is the quantum circuit parameters), and the parameters of the classical learning module C, while the model has embedded the mechanism to ensure physical consistency.

[0043] The classical learning module is used to receive the enhanced features or prediction distribution parameters from the QFE module output, and predict the corresponding 6 10-minute weather values per hour, or output 10-minute level preliminary weather prediction data according to the input ERA5 hourly weather data. The classical learning module can be a multi-output regression model, such as a gradient boosting decision tree (GBDT) model that simultaneously regresses 6 values, or a lightweight sequence-to-sequence neural network model that outputs a sequence of length 6. The sequence-to-sequence neural network model includes a sequentially connected encoder and a decoder. The encoder is used to compress the input sample or enhanced feature or prediction distribution parameter into a fixed-dimensional context vector, which contains all the semantic information of the input data. The decoder outputs 10-minute level weather prediction data according to the context vector. In this example, a gradient boosting decision tree model with monotonic constraint is preferred. The monotonic constraint means that certain input features and outputs have a physical monotonic relationship (for example, when the solar elevation angle increases, the irradiance does not decrease), which ensures physical rationality. Gradient boosting trees can efficiently handle nonlinear relationships, while monotonic constraints can prevent overfitting and predictions that violate common sense. This example can improve the accuracy of weather down-sampling mapping while maintaining realizability through the collaborative processing of the quantum enhancement module and the classical learning module.

[0044] S30: input ERA5 hourly weather data of the target period (such as 2025 and the future period) into the trained quantum-classical hybrid model, and make the quantum-classical hybrid model output 10-minute level weather data of the target period.

[0045] Specifically, step S30 is a target period downsampling inference step, which only needs to obtain ERA5 hourly weather forecasts or ERA5 hourly data, without any 10-minute true value as input. Specifically, the hourly-level features are input into the QFE module. If it is the QFE-1 mode, a quantum state is generated by quantum feature mapping and a quantum feature vector / kernel matrix is calculated, and then is passed to the classical learning module; if it is the QFE-2 mode, the prediction distribution parameters of the 6 slices of the hour are directly calculated by the quantum circuit; for the QFE-1 mode and the QFE-2 mode, the classical learning module receives the outputs of both, calculates the prediction of the 6 10-minute slices of the hour, and realizes the high-resolution reconstruction of the corresponding hour of meteorological elements. If it is the QFE-3 mode, the classical learning module first generates 6 preliminary prediction data, and then adjusts them to the final sequence output that meets the constraints through quantum optimization, that is, the 10-minute resolution meteorological reconstruction sequence is obtained. The above process is repeated to traverse each hour in the target period, and a complete 2025 10-minute meteorological data sequence is generated. These reconstructed 10-minute meteorological data can be used as input data for subsequent photovoltaic power prediction models, such as inputting 10-minute granularity irradiance and temperature into the existing photovoltaic power station power model to calculate the expected power generation curve. Of course, the 10-minute meteorological data can also be applied to downstream applications such as weather prediction. Due to the incorporation of physical constraints and adaptive regularization in model training, the model can still maintain the consistency of hourly cumulative values in 2025 and produce reasonable smooth fine-grained changes.

[0046] The present example provides a method for generating 10-minute meteorological data with physical consistency under the harsh condition of "2016-2020 dual library training, 2025 only ERA5 hourly data available", which can ensure the conservation of hourly cumulative amount and meet the physical laws such as energy balance.

[0047] In an example, after the model training is completed, the model is subjected to physical constraint fusion and verification: For aggregation consistency, it can be verified whether the hourly cumulative value of the model prediction output is strictly consistent with the ERA5 input. If there is a slight deviation, a post-processing step can be added to force the proportion adjustment of the 6 values per hour to match the ERA5 hourly value. In fact, since the aggregation consistency constraint has been incorporated in the model training, generally no large adjustment is needed.

[0048] For shape prior, the smoothness and monotonicity of the prediction sequence on the 10-minute scale are checked. If there are still slight oscillations locally, a low-pass filter can be applied for smoothing, or a maximum and minimum clipping method can be used to clip abnormal peaks. However, the above adjustment methods are usually not needed when the training is sufficient.

[0049] In an example, an age domain adaptive mechanism is introduced, i.e., a quantile mapping method is used to correct the 10-minute level meteorological values output by the quantum-classical hybrid model: A quantile mapping function is constructed to convert the 10-minute level meteorological data output by the quantum-classical hybrid model to obtain corrected 10-minute level meteorological data. For example, the model predicts an irradiance value, and according to the target period quantile mapping function, the predicted irradiance value is converted to the possible value in the target period to correct the meteorological data. Optionally, the quantile mapping function is designed as a trainable layer embedded at the end of the quantum-classical hybrid model as a correction module of the correction model.

[0050] Optionally, during the model training iteration process, the correction module of the model is used to call the quantile mapping function to correct the data every several iteration periods, so that the model gradually adapts to the future distribution.

[0051] Optionally, when the model is physically constrained and verified, the quantile mapping method is used to correct the 10-minute level meteorological values output by the quantum-classical hybrid model. Specifically, for age distribution alignment, the ERA5 data of several years before 2025 can be used to run the quantum-classical hybrid model to determine whether there is a difference between the statistical distribution (e.g., annual average, quantile) of the 10-minute data output by the model and the true value of the NSRDB from 2016-2020, and if so, the correction module parameters of the model (e.g., re-adjust the quantile mapping) can be adjusted to eliminate the deviation, so as to ensure that the model has been calibrated to the maximum extent for the expected 2025 meteorological distribution before application.

[0052] In an example, the quantum enhancement module is executed on a physical quantum processor, and in the absence of a quantum processor, the quantum enhancement module is executed on a quantum circuit simulator running on a classical computer or through a quantum-inspired classical algorithm. Specifically, when a classical computer is used to simulate a quantum circuit, the quantum circuit simulator is configured to simulate quantum feature mapping using a tensor network or a state vector, or to solve a quadratic unconstrained binary optimization model using a simulated annealing algorithm. When the quantum enhancement module is executed based on a quantum-inspired classical algorithm, a quantum-inspired classical kernel function is used to implement quantum feature mapping, a quantum kernel is simulated through classical random feature mapping, and a quadratic unconstrained binary optimization model can also be solved using a simulated annealing algorithm. This example provides an alternative implementation of quantum computing, i.e., a classical simulation implementation path, which can fully utilize quantum acceleration when quantum hardware is available, and can also be implemented through classical means when quantum hardware is lacking, ensuring the practicality and deployability of the method.

[0053] Combining the above examples, the preferred example of the high-resolution meteorological element reconstruction method based on quantum enhancement constraint of the present application is obtained, which includes the following steps: S10': Obtain and preprocess ERA5 hourly meteorological data and NSRDB 10-minute meteorological data of a historical period; S20': Use the preprocessed ERA5 hourly meteorological data as input samples and the time-aligned NSRDB 10-minute meteorological data as supervised targets to iteratively train the quantum-classical hybrid model until the quantum-classical hybrid model converges; the training objective function of the quantum-classical hybrid model introduces an aggregated consistency loss constraint and / or a shape regularization constraint; S30': Build a quantile mapping function to convert the 10-minute meteorological data output by the quantum-classical hybrid model to obtain corrected 10-minute meteorological data; S30': Fuse and verify the model with physical constraints; S40': Input the ERA5 hourly meteorological data of a target period into the trained quantum-classical hybrid model to output 10-minute meteorological data of the target period.

[0054] The application also includes a photovoltaic power prediction method based on the quantum-enhanced constraint-based high-resolution meteorological element reconstruction method formed by any one of the above examples or a combination of multiple examples, comprising the following steps: S100: Align the 10-minute meteorological data with the historical photovoltaic power data, and then use the aligned historical photovoltaic power data to train the power prediction model, adding a mask weighting supervision item and an hour window consistency regularization item to the training loss.

[0055] The mask weighting supervision item is defined based on the solar elevation angle or the irradiance threshold and is used to mark invalid data points. Specifically, the mask weighting supervision item calculation expression is: The mask weighting supervision loss value is represented as The total number of sample time points in the training batch is represented as The time point index is represented as The binary mask is represented as; when the solar zenith angle at this time is less than the set threshold (or the irradiance is greater than 0, i.e. during the daytime valid power generation period), Otherwise (night) This item is used to eliminate invalid data at night; The sample weight coefficient is represented as; it is used to weight the prediction error of different time periods; it is usually set to be proportional to the irradiance intensity or power amplitude at this time (for example ), which means the model's prediction bias during the peak generation period will be more severely punished, thus improving the accuracy of critical periods; denotes the photovoltaic power prediction value of the model output at the th time point; denotes the photovoltaic power ground truth at the th time point; denotes the total number of valid (daytime) samples within a batch, used for loss normalization to prevent loss value magnitude fluctuations due to day and night length changes.

[0056] Further, the hour window consistency regularization term is used to ensure that the predicted 10-minute-level power is consistent with the hourly-level power reference value at the hourly scale. The hour window consistency regularization term calculation expression is: wherein, denotes the hour window consistency regularization loss value; denotes the index of the hourly time window; denotes the total number of hours involved in the calculation (or the set of hourly windows within a batch) denotes the fine-grained time step index within the window (for 1 hour to 10 minutes, the value is 1 to 6); denotes the meteorological prediction value (such as irradiance) generated by the model for the th 10-minute slice within the th hour; denotes the complete high-resolution sequence set generated within the th hour; denotes the aggregation operator (Aggregation Operator), defined according to the nature of the physical quantity: for irradiance integration, it is a summation operator; for state quantities such as temperature or wind speed, it may be a mean operator (Mean), in this formula, it mainly refers to summation. denotes the coarse-grained ground truth corresponding to the th hour (i.e. ERA5 hourly data), serving as the physical constraint boundary within the window; denotes the squared L2 norm; for a scalar (single numerical value), this is equivalent to the square of the number.

[0057] Further, to obtain photovoltaic power prediction capability, a power prediction model is trained using historical data. A historical interval (e.g., records of a certain photovoltaic power station from 2016 to 2020) containing photovoltaic power generation measured power data is selected. The measured 5-minute power data is aggregated and averaged (or summed) every 6 points into 10-minute data to correspond to the time granularity of the aforementioned 10-minute level meteorological data. The 10-minute granularity meteorological features (which can come from NSRDB true values or simulated values based on ERA5 downsampling) are aligned with the corresponding 10-minute power data to build a supervised learning sample.

[0058] When training the photovoltaic power model, a mask weighted loss is introduced: for each 10-minute sample, different weights are given according to the pre-generated mask M(t). For example, at night (when the solar elevation is negative or the irradiance is close to 0), the power data is meaningless, thereby reducing the contribution of the sample to the loss; during the day when normal power generation occurs, the sample is given normal weight, thereby avoiding interference of the model by zero power or abnormal data at night. The loss function also includes a consistency regularization of the hour window aggregation, which is used to constrain the sum of the 6 10-minute power predictions of the power model in any hour to be close to the more coarse-grained hourly total power generation (if known), thereby corresponding to the aggregation consistency on the meteorological side. If the hourly total is unknown, a sliding window smoothing method can be used to keep the adjacent hour slice fluctuations within a reasonable range. Through the above training method, the robustness of the power model under the condition of data missing or imbalance is improved, and the power prediction is ensured not to deviate from the actual total at the hourly level, so that the power model can accurately fit the high-frequency changes and also ensure the reasonableness of the low-frequency total.

[0059] S200: input the 10-minute level meteorological data of the target period into the trained power prediction model to obtain 10-minute level photovoltaic power prediction values.

[0060] Optionally, the photovoltaic power model itself can adopt a structure similar to a classic learning module, such as a gradient boosting tree regression (considering the nonlinear relationship between power and meteorology), or a sequence-to-sequence neural network model or a small neural network. The model can input the meteorological prediction data at the current time and the cross-hour context information (such as the irradiance change trend, cloud cover change, etc. in the last hour or the previous hours) to represent the hysteresis effect of cloud state change, and then output 10-minute level photovoltaic power prediction values, thereby improving the prediction stability.

[0061] In an example, further including a correction process on the 10-minute level photovoltaic power prediction values, including an online re-calibration and / or power correction step.

[0062] Online recalibration includes: recalibrating the 10-minute photovoltaic power prediction values ​​output by the power prediction model based on the actual power observation data of the most recent K hours. The recalibrated parameters are dynamically updated using least squares regression with a forgetting factor to correct prediction bias in real time. Specifically, after the photovoltaic power prediction model is deployed online, it can be continuously calibrated using real-time power monitoring data. A sliding time window of the most recent K hours (e.g., the most recent 48 hours) is maintained, and the model prediction values ​​and actual monitoring values ​​are collected every 10 minutes within this window. A linear correction function is fitted using least squares regression with a forgetting factor to correct the original model prediction. Adjusted to corrected prediction The forgetting factor gives greater weight to newer data to quickly respond to recent systematic biases (such as changes in sensor calibration or sudden environmental events). Correction coefficients are updated periodically (e.g., hourly or daily). This data is then applied to subsequent prediction outputs to achieve online small-window recalibration. This process is equivalent to an adaptive post-processing module, further improving the reliability and accuracy of predictions during long-term runtime.

[0063] Power correction involves constructing a periodically adjusted loss function for the power prediction model based on the average annual performance degradation rate of photovoltaic modules and the temperature difference between historical and target periods. This periodically adjusted loss function is then used to derate the 10-minute photovoltaic power prediction values. Specifically, power correction is essentially a cross-decadal adaptive strategy. It adds a penalty term specific to the target period characteristics at regular intervals (e.g., every 1000 iterations) to the conventional loss function (e.g., prediction error). For example, the true power values ​​in the training data are uniformly multiplied by a correction factor of 0.97. Subsequent loss calculations then use the corrected power values ​​learned by the power prediction model. Through power correction, prediction inconsistencies caused by climate change or equipment aging can be mitigated, enhancing the generalization ability of the target period data output by the power prediction model.

[0064] It should be noted that the training and recalibration of the power prediction model do not require the support of 10-minute meteorological data for the target period, such as 2025. The power model training mainly relies on existing historical power data, while the application in 2025 relies on the 10-minute meteorological forecast data generated by this invention and a small amount of real-time power feedback. Therefore, the method of this invention can complete the task of power prediction from meteorological forecast to end-to-end even in the absence of high-resolution meteorological data.

[0065] Combining the examples of the two methods described above yields a preferred example of the present invention's method for predicting photovoltaic power from hourly meteorological data to ten-minute levels based on quantum-enhanced constraints, such as... Figure 3As shown, 2016–2020 is taken as the historical period, and 2025 as the target period. ERA5 and NSRDB meteorological data from 2016–2020 are used for training. The goal is to be able to predict 10-minute meteorological elements using only ERA5 hourly data by 2025, and further apply this to photovoltaic power forecasting. The specific steps include: S1: Acquire and preprocess ERA 5-hour meteorological data and NSRDB 10-minute meteorological data from 2016–2020; S2: Using preprocessed ERA 5-hour meteorological data as input samples and time-aligned NSRDB 10-minute meteorological data as supervision targets, the quantum-classical hybrid model is iteratively trained until the quantum-classical hybrid model converges; the training objective function of the quantum-classical hybrid model introduces aggregation consistency loss constraints and / or shape regularization constraints. A quantile mapping function is constructed to transform the 10-minute meteorological data output by the quantum-classical hybrid model, resulting in corrected 10-minute meteorological data. S3: Input the 2025 ERA 5-hour meteorological data into the trained quantum-classical hybrid model, so that the quantum-classical hybrid model outputs 10-minute meteorological data for the target period.

[0066] S4: Train the power prediction model using aligned historical photovoltaic power data, and add a mask weighted supervision term and an hourly window consistency regularization term to the training loss; based on the actual power observation data of the most recent K hours, recalibrate the 10-minute photovoltaic power prediction value output by the power prediction model online, and use the periodic adjustment loss function to correct the 10-minute photovoltaic power prediction value. S5: Input the 10-minute meteorological data for 2025 into the trained power prediction model to obtain the 10-minute photovoltaic power prediction value.

[0067] This method comprises two stages: data acquisition and model training during the historical monitoring period, and inference and recalibration during the target application period. During the historical monitoring period (2016–2020), ERA 5-hour meteorological data and NSRDB 10-minute meteorological data were acquired to train the quantum-classical hybrid model and the power prediction model. During the target application period (2025), only ERA 5-hour data was used as input. The trained quantum-classical hybrid model generated 10-minute meteorological data, and the power prediction model was used to predict photovoltaic power, with online calibration performed simultaneously. The entire process strictly does not rely on 10-minute ground truth labels after 2020, thus solving the problem of a lack of high-resolution ground truth data for the target period.

[0068] The application also provides a cross-era mapping system for hour-level meteorological data to ten-minute-level photovoltaic power prediction based on quantum-enhanced constraints, comprising: A data acquisition and preprocessing module acquires and preprocesses ERA5 hour-level meteorological data and NSRDB ten-minute-level meteorological data of a historical period.

[0069] A first training module uses the preprocessed ERA5 hour-level meteorological data as input samples and time-aligned NSRDB ten-minute-level meteorological data as a supervised target to iteratively train a quantum-classical hybrid model until the quantum-classical hybrid model converges; the training objective function of the quantum-classical hybrid model introduces an aggregated consistency loss constraint and / or a shape regularization constraint.

[0070] A quantum enhancement module performs feature enhancement processing on input samples according to quantum feature mapping and quantum kernel estimation, or performs regression processing on input samples based on a variational quantum circuit, or performs constraint optimization processing on ten-minute-level preliminary meteorological prediction data output by a classical learning module based on a quadratic unconstrained binary optimization model, and correspondingly outputs enhanced features or prediction distribution parameters of ten-minute-level meteorological data or optimized ten-minute-level meteorological prediction data.

[0071] A classical learning module generates ten-minute-level meteorological prediction data according to enhanced features or prediction distribution parameters, or outputs ten-minute-level preliminary meteorological prediction data according to input samples.

[0072] A second training module trains a power prediction model using aligned historical photovoltaic power data, and adds a mask weighting supervision item and an hour window consistency regularization item to the training loss.

[0073] An online re-calibration and power correction module performs online re-calibration on ten-minute-level photovoltaic power prediction values output by the power prediction model according to actual power observation data of the last K hours, and performs correction processing on the ten-minute-level photovoltaic power prediction values using a periodic adjustment loss function.

[0074] The above modules can be integrated on the same computing platform to work cooperatively. Through the above software and hardware implementation means, the scheme of the application can be conveniently deployed in an actual prediction system. For example, a software system containing the above modules is deployed in a power grid dispatching center, and the latest ERA5 or numerical weather prediction hour data can be down-sampled to ten-minute prediction power station data in real time every day using the method of the application, and the model is self-corrected according to the real-time power data returned by the power station, ensuring long-term accurate and reliable prediction.

[0075] The present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the photovoltaic power prediction method according to any one of the examples or combinations of multiple examples. The processor can be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present application.

[0076] The present application also provides a storage medium having the same inventive concept as the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the photovoltaic power prediction method according to any one of the examples or combinations of multiple examples, and having computer instructions stored thereon, which, when executed, perform the steps of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the photovoltaic power prediction method according to any one of the examples or combinations of multiple examples.

[0077] Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product stored in a storage medium, including instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0078] The present application also provides a terminal having the same inventive concept as any one of the examples or combinations of multiple examples of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the photovoltaic power prediction method, comprising a memory and a processor, the memory having computer instructions stored thereon that can be executed on the processor, and the processor executing the computer instructions to perform the steps of the high-resolution meteorological element reconstruction method based on quantum-enhanced constraints or the photovoltaic power prediction method. The processor can be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present application.

[0079] In an example, the terminal, i.e., the electronic device, is in the form of a general-purpose computing device, and the components of the electronic device can include but are not limited to the at least one processing unit (processor), the at least one storage unit, and a bus connecting different system components, including the storage unit and the processing unit.

[0080] The storage unit stores program codes which can be executed by the processing unit, so that the processing unit executes the steps described in the above "Exemplary Methods" section according to various exemplary embodiments of the present application. For example, the processing unit can execute the above-mentioned high-resolution meteorological element reconstruction method or photovoltaic power prediction method based on quantum-enhanced constraints.

[0081] The storage unit can include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) 3201 and / or a cache memory, and can further include a read-only memory (ROM).

[0082] The storage unit can also include a program / utility, having a set (at least one) of program modules that are executed by the processing unit 3202, and / or can include program / data, each or a combination thereof which can include implementation of a network environment.

[0083] The bus can be representative of one or more of several types of bus structures, including a storage unit bus or bus controller, a peripheral bus, a graphics acceleration port, a processing unit bus, or a local bus using any of a variety of bus architectures.

[0084] The electronic device can also communicate with one or more external devices such as a keyboard or a pointing device, through an I / O interface. Additionally, the electronic device can communicate with one or more devices that enable a user to interact with the electronic device, and / or one or more devices (e.g., a router, a modem, etc.) that enable the electronic device to communicate with one or more other computing devices. Such communication can occur via an I / O interface. Still yet, the electronic device can communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or the Internet) through a network adapter. It should be appreciated that the electronic device can be a part of a larger system, including but not limited to a network of electronic devices, a cloud computing system, or the like. The network adapter can communicate with the other system devices and / or systems through the bus. It will be appreciated that other hardware and / or software modules can be used in conjunction with the electronic device. Such modules can include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0085] Through the above description, those skilled in the art can easily understand that the example embodiments described herein can be implemented by software or by software in combination with necessary hardware. Therefore, the technical solutions according to the example embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash disk, a mobile hard disk, or the like) or on a network, and includes a number of instructions to make a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) execute the methods of the example embodiments of the present application.

[0086] The above detailed description is a detailed description of the present application, and cannot be considered as limiting the specific embodiments of the present application to these descriptions. For those skilled in the art, without departing from the concept of the present application, a number of simple deductions and substitutions can be made, which should be considered as falling within the scope of protection of the present application.

Claims

1. A high-resolution meteorological element reconstruction method, characterized by, The method comprises the following steps: obtaining and preprocessing ERA5 hourly meteorological data and NSRDB 10-minute meteorological data of a historical period, the preprocessing comprising data alignment processing; iteratively training a quantum-classical hybrid model by taking the preprocessed ERA5 hourly meteorological data as input samples and taking the time-aligned NSRDB 10-minute meteorological data as a supervised target until the quantum-classical hybrid model converges; the training objective function of the quantum-classical hybrid model introduces an aggregated consistency loss constraint and / or a shape regularization constraint, the aggregated consistency loss is used to measure the deviation of the sum of 6 10-minute meteorological values within each hour from the corresponding ERA5 hourly meteorological value, and the shape regularization constraint is used to measure the deviation of the smoothness of the 10-minute sequence output by the model; the quantum-classical hybrid model comprises a quantum enhancement module and a classical learning module; the quantum enhancement module performs feature enhancement processing on the input samples based on quantum feature mapping and quantum kernel estimation, or performs regression processing on the input samples based on a variational quantum circuit, or performs constraint optimization processing on 10-minute preliminary meteorological prediction data output by the classical learning module based on a quadratic unconstrained binary optimization model, and correspondingly outputs enhanced features or prediction distribution parameters of 10-minute meteorological data or optimized 10-minute meteorological prediction data; the classical learning module generates 10-minute meteorological prediction data according to the enhanced features or the prediction distribution parameters, or the classical learning module outputs 10-minute preliminary meteorological prediction data according to the input samples; input ERA5 hourly meteorological data of a target period into the trained quantum-classical hybrid model to make the quantum-classical hybrid model output 10-minute meteorological data of the target period.

2. The high-resolution meteorological element reconstruction method of claim 1, wherein, the feature enhancement processing comprises encoding the input samples into quantum states by using a parameterized quantum circuit, measuring the quantum states to obtain a quantum kernel matrix, and inputting the quantum kernel matrix as enhanced features into the classical learning module; the regression processing comprises outputting prediction distribution parameters of 10-minute meteorological data according to the input samples; wherein the training objective function comprises an aggregated consistency loss and a shape regularization loss, and the parameters of the variational quantum circuit and the classical learning module are optimized by a gradient descent method; the constraint optimization processing comprises converting the 10-minute preliminary meteorological prediction data and the aggregated consistency constraint into a quadratic unconstrained binary optimization problem, and solving or approximately solving the quadratic unconstrained binary optimization problem by using a quantum annealing algorithm or a variational quantum optimization algorithm, so as to optimize the 10-minute preliminary meteorological prediction data output by the classical learning module into 10-minute meteorological prediction data satisfying the aggregated consistency constraint.

3. The high-resolution meteorological element reconstruction method of claim 1, wherein, The classical learning module is a multi-output regression model or a sequence-to-sequence neural network model.

4. The high-resolution meteorological element reconstruction method of claim 1, wherein, The quantum enhancement module is executed on a physical quantum processor, or by a quantum circuit simulator running on a classical computer, or by a quantum-inspired classical algorithm; the quantum circuit simulator simulates quantum feature mapping using a tensor network or a state vector, or solves a quadratic unconstrained binary optimization model using a simulated annealing algorithm; when the quantum enhancement module is executed by a quantum-inspired classical algorithm, a quantum-inspired classical kernel function is used to implement quantum feature mapping, and a simulated annealing algorithm is used to solve a quadratic unconstrained binary optimization model.

5. The high-resolution meteorological element reconstruction method of claim 1, wherein, The method further comprises a quantile adjustment step: A quantile mapping function is constructed to convert the 10-minute meteorological data output by the quantum-classical hybrid model to obtain corrected 10-minute meteorological data.

6. A photovoltaic power prediction method, implemented on the basis of the high-resolution meteorological element reconstruction method according to any one of claims 1-5, characterized in that, The method comprises the following steps: The 10-minute meteorological data is aligned with historical photovoltaic power data, and the trained power prediction model is then trained using the aligned historical photovoltaic power data, with a mask weighted supervision term and an hour window consistency regularization term added to the training loss; the mask weighted supervision term is defined based on a solar elevation angle or irradiance threshold to mark invalid data points; the hour window consistency regularization term is used to ensure that the predicted 10-minute power is consistent with the hourly power reference value at the hourly scale; The 10-minute meteorological data of the target period is input into the trained power prediction model to obtain 10-minute photovoltaic power prediction values.

7. The photovoltaic power prediction method of claim 6, wherein, The method further comprises correcting the 10-minute photovoltaic power prediction values, including: According to the actual power observation data of the last K hours, the 10-minute photovoltaic power prediction values output by the power prediction model are online recalibrated, and the parameters of the recalibration are dynamically updated by least squares regression with a forgetting factor to correct the prediction bias in real time; and / or A periodic adjustment loss function of the power prediction model is constructed according to the performance degradation rate of the photovoltaic module and the temperature difference between the historical period and the target period, and the 10-minute photovoltaic power prediction values are corrected using the periodic adjustment loss function.

8. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the high-resolution meteorological element reconstruction method of any one of claims 1-5 or the steps of the photovoltaic power prediction method of claim 6 or 7.

9. A storage medium having stored thereon computer instructions, characterized in that, The computer program, when executed by a processor, implements the steps of the high-resolution meteorological element reconstruction method of any one of claims 1-5 or the steps of the photovoltaic power prediction method of claim 6 or 7.

10. A terminal comprising a memory and a processor, said memory having stored thereon computer instructions executable on said processor, characterized in that, The computer program, when executed by a processor, implements the steps of the high-resolution meteorological element reconstruction method of any one of claims 1-5 or the steps of the photovoltaic power prediction method of claim 6 or 7.

Images