Method and system for improving photovoltaic power prediction precision based on meteorological cloud picture recognition
By collecting meteorological data from multimodal air and ground stations and constructing a dual-channel meteorological cloud image recognition system, and using generative adversarial networks for feature mapping modeling, the problems of insufficient utilization of multi-source meteorological data and low cloud image recognition accuracy in photovoltaic power prediction were solved, thereby improving the accuracy and stability of photovoltaic power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-15
- Publication Date
- 2026-03-13
AI Technical Summary
Existing photovoltaic power prediction methods suffer from insufficient utilization of multi-source meteorological data, low accuracy of cloud image recognition, and large prediction errors, making it difficult to meet the requirements for high accuracy and real-time performance.
Multimodal air-to-ground meteorological data were collected, and a dual-channel meteorological cloud image recognition system was established, including a cloud image feature extraction channel and a feature fusion modeling channel. Feature mapping modeling was performed using generative adversarial networks to construct a cloud cluster irradiance mapping model. Combined with historical power generation data from photovoltaic power plants, prediction training and fitting were performed to construct a photovoltaic power optimization predictor.
It improves the accuracy and stability of photovoltaic power prediction, and realizes high-precision and real-time power prediction and control of photovoltaic power plants.
Smart Images
Figure CN121661518A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power prediction technology, and specifically to a method and system for improving the accuracy of photovoltaic power prediction based on meteorological cloud map recognition. Background Technology
[0002] In photovoltaic power forecasting, complex and variable meteorological conditions significantly affect solar irradiance due to factors such as cloud distribution, thickness, and movement speed. Traditional forecasting methods often rely on a single data source, such as ground meteorological observations or historical power data, which fails to comprehensively reflect the actual characteristics of irradiance variations. Furthermore, meteorological cloud images have limited recognition capabilities and are not precise enough in depicting cloud morphology and dynamic changes, resulting in incomplete model input features and significant biases in the forecast results. This makes it difficult to meet the high-precision, real-time power forecasting requirements of photovoltaic power plants. Summary of the Invention
[0003] This application provides a method and system for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition, which is used to address the technical problems of insufficient utilization of multi-source meteorological data, low cloud image recognition accuracy and large prediction error in existing photovoltaic power prediction technologies.
[0004] In view of the above problems, this application provides a method and system for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition.
[0005] The first aspect of this application provides a method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition, the method comprising: Multimodal air-to-ground meteorological data of the target photovoltaic power station are collected, including all-sky image data of the photovoltaic power station, satellite cloud image data, and ground meteorological station observation data. A dual-channel meteorological cloud image recognition system is established, comprising a cloud image feature extraction channel and a feature fusion modeling channel. The dual-channel system is used to extract features and fuse and model the multimodal air-to-ground meteorological data to generate a cloud cluster irradiance mapping model. Historical power generation datasets of the photovoltaic power station are obtained. Based on the cloud cluster irradiance mapping model and the historical power generation datasets of the photovoltaic power station, prediction training, fitting, and accuracy improvement optimization are performed to construct a photovoltaic power optimization predictor. Power prediction control is then performed using the photovoltaic power optimization predictor.
[0006] A second aspect of this application provides a photovoltaic power prediction accuracy improvement system based on meteorological cloud image recognition, the system comprising: The system includes a meteorological data acquisition module for collecting multimodal air-to-ground meteorological data from the target photovoltaic power station. This multimodal air-to-ground meteorological data includes all-sky image data of the photovoltaic power station, satellite cloud image data, and ground meteorological station observation data. A channel construction module is used to build a dual-channel meteorological cloud image recognition system, comprising a cloud image feature extraction channel and a feature fusion modeling channel. A mapping model generation module is used to extract features and perform fusion modeling on the multimodal air-to-ground meteorological data using the dual-channel meteorological cloud image recognition system, generating a cloud irradiance mapping model. A prediction module is used to acquire historical power generation datasets from the photovoltaic power station, perform prediction training, fitting, and accuracy improvement optimization based on the cloud irradiance mapping model and the historical power generation datasets, construct a photovoltaic power optimization predictor, and perform power prediction control through the photovoltaic power optimization predictor.
[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application collects multimodal air-to-ground meteorological data from a target photovoltaic (PV) power plant. This multimodal air-to-ground meteorological data includes PV all-sky image data, satellite cloud image data, and ground meteorological station observation data. A dual-channel meteorological cloud image recognition system is constructed, comprising a cloud image feature extraction channel and a feature fusion modeling channel. The system uses this dual-channel system to extract features and fuse modeling from the multimodal air-to-ground meteorological data, generating a cloud irradiance mapping model. Historical power generation data from the PV power plant is acquired. Based on the cloud irradiance mapping model and the historical power generation data, prediction training, fitting, and accuracy improvement optimization are performed to construct a PV power optimization predictor. Power prediction control is then performed using this PV power optimization predictor. This invention addresses the technical problems of insufficient utilization of multi-source meteorological data, low cloud image recognition accuracy, and large prediction errors in existing PV power prediction technologies. By constructing a dual-channel meteorological cloud image recognition system and fusing multimodal meteorological data for modeling and optimization, the technical effect of improving the accuracy and stability of PV power prediction is achieved. Attached Figure Description
[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0009] Figure 1 A schematic diagram of the method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition provided in the embodiments of this application; Figure 2A schematic diagram of the photovoltaic power prediction accuracy improvement system based on meteorological cloud map recognition provided in this application embodiment.
[0010] Figure labeling: Meteorological data acquisition module 11, channel construction module 12, mapping model generation module 13, forecasting module 14. Detailed Implementation
[0011] This application provides a method and system for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition. It addresses the technical problems of insufficient utilization of multi-source meteorological data, low accuracy of cloud image recognition, and large prediction errors in existing photovoltaic power prediction technologies. By constructing a dual-channel meteorological cloud image recognition system and fusing multimodal meteorological data for modeling and optimization, the technical effect of improving the accuracy and stability of photovoltaic power prediction is achieved.
[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.
[0014] Example 1, as Figure 1 As shown, this application provides a method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition, the method comprising: Step S100: Collect multimodal air-to-ground meteorological data of the target photovoltaic power station. The multimodal air-to-ground meteorological data includes photovoltaic all-sky image data, satellite cloud image data and ground meteorological station observation data.
[0015] In this embodiment, multi-source meteorological monitoring equipment is installed at the target photovoltaic power station to simultaneously collect and record different types of meteorological information, thereby obtaining multi-modal air-to-ground meteorological data for the target photovoltaic power station. In this process, firstly, a photovoltaic all-sky imager is used to acquire photovoltaic all-sky image data to reflect the distribution, thickness, and movement direction of clouds in the sky; secondly, meteorological satellite observation data is retrieved to obtain satellite cloud image data, presenting the evolution characteristics of large-scale cloud clusters from a macroscopic perspective; finally, ground meteorological station observation devices record ground meteorological parameters such as temperature, humidity, wind speed, air pressure, and solar irradiance in real time. Through the comprehensive collection and time-synchronized processing of the above multi-source information, multi-modal air-to-ground meteorological data is formed.
[0016] Step S200: Establish a dual-channel meteorological cloud image recognition system, which includes a cloud image feature extraction channel and a feature fusion modeling channel.
[0017] Furthermore, the method provided in the application embodiments, which establishes a dual-channel system for meteorological cloud image recognition, also includes: Based on the modal characteristics of the multimodal air-to-ground meteorological data, the data preprocessing steps are analyzed to obtain the multimodal data preprocessing step channel; according to the meteorological cloud image feature extraction target, the output data of the multimodal data preprocessing step channel is correlated by a feature algorithm to determine the multimodal data feature extraction channel; the multimodal data preprocessing step channel and the multimodal data feature extraction channel are connected in series and integrated to obtain the cloud image feature extraction channel; a feature fusion modeling channel is constructed, and the feature fusion modeling channel is linked to the cloud image feature extraction channel to build the dual-channel meteorological cloud image recognition system.
[0018] In this embodiment, to establish a dual-channel meteorological cloud image recognition system, a data preprocessing step is first analyzed based on the modal characteristics of multimodal air-to-ground meteorological data. During this process, format unification, time alignment, and spatial registration are performed on photovoltaic all-sky image data, satellite cloud image data, and ground meteorological station observation data, respectively. Photovoltaic all-sky image data undergoes grayscale normalization and noise suppression to improve image clarity and brightness consistency; satellite cloud image data achieves spatial resolution matching through resampling and geographic coordinate correction; and ground meteorological station observation data undergoes interpolation completion and normalization to ensure temporal series continuity and numerical scale consistency. Through these steps, the multimodal data preprocessing step channel is obtained.
[0019] Next, feature algorithms are used to correlate the output data of the multimodal data preprocessing steps according to the meteorological cloud image feature extraction objectives. In this process, convolutional neural networks are used to extract cloud morphology and texture features from photovoltaic all-sky image data and satellite cloud image data to identify cloud shape, thickness, and coverage. Ground meteorological station observation data are analyzed using time series analysis to extract the changing trends of meteorological parameters such as temperature, humidity, wind speed, and irradiance. To determine the correlation between image features and meteorological features, the Pearson correlation coefficient method is used to calculate the linear correlation between the two sets of features over time, and mutual information analysis is used to measure their nonlinear dependence. Then, low-correlation or redundant features are eliminated through feature significance testing, retaining the feature set that accurately characterizes cloud evolution and irradiance changes, thus obtaining the multimodal data feature extraction channels.
[0020] Next, the multimodal data preprocessing channel and the multimodal data feature extraction channel are integrated in series. A continuous path from raw meteorological data input to feature output is established through feature stitching. In this process, a batch normalization method is used to maintain the stability of feature distribution, so that spatial features and temporal features are expressed consistently at the same scale, thereby obtaining the cloud image feature extraction channel.
[0021] Next, a feature fusion modeling channel was constructed. Based on the output features of the cloud image feature extraction channel, features from photovoltaic all-sky images, satellite cloud images, and ground meteorological features were extracted. A multimodal feature fusion function was then established, combining these features with corresponding photovoltaic irradiance features to fuse features from different sources, resulting in a meteorological cloud image fusion feature set. A generative adversarial network (GAN) was introduced to dynamically map the meteorological cloud image fusion feature set to photovoltaic irradiance features. Adversarial training between the generator and discriminator optimized the mapping accuracy and model convergence performance, forming a feature fusion modeling channel with high recognition capabilities. Finally, the feature fusion modeling channel was linked to the cloud image feature extraction channel, achieving the integration and fusion of feature information, thus completing the construction of a dual-channel meteorological cloud image recognition system.
[0022] Furthermore, in the method provided in the application embodiments, constructing the feature fusion modeling channel further includes: Based on the output features of the cloud image feature extraction channel, photovoltaic all-sky image features, satellite cloud image features, ground meteorological features, and corresponding photovoltaic irradiance features are obtained; a multimodal feature fusion function is constructed, and the photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features are weighted and fused based on the multimodal feature fusion function to obtain a meteorological cloud image fusion feature set; a generative adversarial network is introduced to dynamically map and model the meteorological cloud image fusion feature set and the corresponding photovoltaic irradiance features, and a feature fusion modeling channel is constructed.
[0023] In this embodiment, based on the output features of the cloud image feature extraction channel, photovoltaic all-sky image features, satellite cloud image features, ground meteorological features, and corresponding photovoltaic irradiance features are first acquired. The photovoltaic all-sky image features are obtained using a convolutional neural network feature extraction algorithm to extract cloud morphology features, edge texture features, and cloud coverage ratio features to reflect the cloud cover distribution and changing trends in local sky areas. Satellite cloud image features utilize multi-scale convolutional analysis to extract cloud movement direction, cloud band thickness, and regional structural change features, reflecting the dynamic processes of the atmospheric system at a macroscopic scale. Ground meteorological features are extracted using time series analysis methods, modeling time-series data such as temperature, humidity, wind speed, and air pressure based on a long short-term memory network to capture the impact of meteorological changes on irradiance intensity. Photovoltaic irradiance features are obtained by combining real-time data collected by ground irradiance sensors with historical power data. After sliding window filtering and polynomial smoothing, irradiance change rate, average irradiance value, and irradiance stability index are extracted to characterize the dynamic fluctuation characteristics of irradiance features. A timestamp alignment method is used to achieve time-series synchronization of multi-source data, resulting in unified feature data.
[0024] After acquiring the feature data, a multimodal feature fusion function is constructed to fuse features from photovoltaic all-sky images, satellite cloud images, and ground meteorological features. In this process, based on multimodal air-to-ground meteorological data, meteorological cloud cover monitoring data is first acquired, and the impact of multimodal data is assessed based on cloud cover change trends to determine the precise impact of the multimodal data. Then, based on this impact, dynamic decision weights for the multimodal data are determined. By fitting the contribution relationships of different data sources to irradiance changes, a weighted fitting of multimodal features is performed to construct the multimodal feature fusion function. Finally, the constructed multimodal feature fusion function is used to weight and fuse features from photovoltaic all-sky images, satellite cloud images, and ground meteorological features to obtain a meteorological cloud image fusion feature set.
[0025] Subsequently, a generative adversarial network (GAN) is introduced to dynamically map and model the fused feature set of meteorological cloud images and the corresponding photovoltaic irradiance features. In this process, an initial generator and an initial discriminator are first established based on the GAN framework. The generator is used to generate irradiance prediction results based on the fused features, and the discriminator is used to distinguish the prediction results from the actual irradiance features. During the model training phase, an empirical model of cloud image irradiance is injected into the generator structure, enabling it to perform irradiance prediction training based on the fused feature set of meteorological cloud images, thus obtaining an irradiance generator with irradiance generation capabilities. Then, using actual photovoltaic irradiance features as the training benchmark, the output of the irradiance generator is trained to generate an irradiance discriminator, enhancing the model's ability to discriminate irradiance prediction errors. Through dynamic adversarial training between the irradiance generator and the irradiance discriminator, the irradiance prediction performance of the generator and the recognition accuracy of the discriminator are gradually optimized, achieving a high-precision dynamic mapping from meteorological features to irradiance features. Finally, based on the converged irradiance generator and irradiance discriminator, a feature fusion modeling channel is constructed.
[0026] Furthermore, in the method provided in the application embodiments, constructing the multimodal feature fusion function further includes: Based on the multimodal air-to-ground meteorological data, meteorological cloud cover monitoring data is obtained; the impact of the multimodal air-to-ground meteorological data is assessed according to the meteorological cloud cover monitoring data to determine the precise impact of the multimodal data; based on the precise impact of the multimodal data, the dynamic decision weights of the multimodal data are determined; based on the dynamic decision weights of the multimodal data, a weighted fitting of multimodal features is performed to construct the multimodal feature fusion function.
[0027] In this embodiment, meteorological cloud cover monitoring data is first obtained based on multimodal air-to-ground meteorological data. After brightness correction and spatial registration of photovoltaic all-sky image data and satellite cloud image data, threshold segmentation is used to distinguish between clouds and clear sky. Then, the cloud coverage rate at each moment is calculated according to the regional pixel ratio, and the data are concatenated into a continuous cloud cover change sequence with a fixed time step, thereby obtaining meteorological cloud cover monitoring data.
[0028] Next, the impact of multimodal air-to-ground meteorological data is assessed based on cloud cover monitoring data. In this process, cloud cover monitoring sequences are paired with time series of temperature, humidity, wind speed, air pressure, and surface irradiance. Each sequence is normalized by removing the mean and standard deviation, converting it to a form with a mean of zero and a variance of one. Then, the Pearson correlation coefficient is calculated to measure the degree of synchronization between the two sequences. Specifically, the mean of each sequence is first calculated, then the product of the deviations and covariance at the corresponding time points is calculated, and finally, the covariance is divided by the product of the two standard deviations to obtain the correlation coefficient. This calculation is repeated within a small time lag, and the highest correlation coefficient is taken as the impact of that factor relative to cloud cover changes, ultimately forming a precise impact set of multimodal data.
[0029] Then, based on the precise impact degree of the multimodal data, the dynamic decision weights of the multimodal data are determined. In this process, the absolute value of each impact degree is compared with the sum of all impact degrees to obtain its proportion in the total impact. These proportions are then normalized so that the sum of all weights equals one. When the meteorological environment changes, the proportions are recalculated based on the new impact degrees to obtain the dynamic decision weights of the multimodal data reflecting the current meteorological scenario.
[0030] Finally, multimodal feature weighting is performed based on dynamic decision weights for multimodal data. In this process, photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features are aligned along the time dimension to form a feature vector set. Features from different sources are weighted and combined according to dynamic decision weights, and the least squares method is used to determine the weighting coefficients, minimizing the error between the weighted result and the actual observed irradiance. Through multiple iterations to optimize the fitting parameters, the fusion result stably reflects the correlation between multimodal features. Ultimately, through the above weighted fitting process, a multimodal feature fusion function is established that can uniformly describe the correlation between photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features.
[0031] Furthermore, the method provided in the application embodiment introduces a generative adversarial network to dynamically map and model the meteorological cloud image fusion feature set and the corresponding photovoltaic irradiance features, and constructs a feature fusion modeling channel, further including: Based on the generative adversarial network, an initial generator and an initial discriminator are obtained; the injected cloud image irradiance empirical model is used to train the initial generator for irradiance prediction based on the meteorological cloud image fusion feature set to obtain an irradiance generator; the irradiance generator is trained for irradiance discrimination according to the corresponding photovoltaic irradiance features to generate an irradiance discriminator; dynamic mapping modeling is performed based on the irradiance generator and the irradiance discriminator to construct a feature fusion modeling channel.
[0032] In this embodiment, an initial generator and an initial discriminator are first obtained based on a generative adversarial network (GAN). The generator employs a convolutional neural network (CNN) method, extracting spatial structure information from the fused features of meteorological cloud images through multi-layer convolutional operations, and generating an irradiance prediction map using a deconvolutional structure. The discriminator employs a convolutional classification network method, extracting local and global features from the input irradiance map through convolutional layers, and outputting a true / false judgment result using the sigmoid function. The parameters of both are set using a random initialization method to ensure stable convergence of the model in the early stages of training, thereby obtaining an initial generator and an initial discriminator with basic mapping capabilities.
[0033] Subsequently, the initial generator is trained using a cloud image irradiance empirical model based on a meteorological cloud image fusion feature set for irradiance prediction. This cloud image irradiance empirical model is a model established based on meteorological observations and physical radiation laws, used to describe the relationship between cloud distribution and solar irradiance. This model analyzes historical meteorological data, combining parameters such as cloud cover, cloud top height, and solar incidence angle, and uses linear regression to establish the correspondence between irradiance changes and cloud optical thickness, forming a reference model that can be used to calculate theoretical irradiance distribution. During training, the initial generator uses the meteorological cloud image fusion feature set as input to generate an irradiance prediction map; the empirical model provides the target irradiance value as a supervision signal. The initial generator calculates the difference between the predicted value and the empirical model output, using the mean squared error as the loss function, and continuously adjusts the network parameters through a backpropagation algorithm, ultimately obtaining an irradiance generator capable of accurately predicting irradiance distribution based on meteorological characteristics.
[0034] After obtaining the irradiance generator, it is trained using photovoltaic irradiance features to generate an irradiance discriminator. In this process, a binary classification method is employed. The discriminator is input into both the actual irradiance observation map and the predicted irradiance map output by the irradiance generator, and the error between the classification result and the true label is calculated. The classification error is quantified using the cross-entropy loss function, and the initial discriminator parameters are updated using the gradient descent algorithm, enabling the initial discriminator to distinguish between predicted and actual irradiance, thus obtaining a converged irradiance discriminator.
[0035] Finally, dynamic mapping modeling is performed based on the irradiance generator and irradiance discriminator. In this process, the output results of the irradiance generator and irradiance discriminator are first evaluated for loss, and the performance loss parameters of the generator and discriminator are calculated respectively. Then, based on the loss parameters, the generator and discriminator are iteratively trained adversarially, allowing them to continuously optimize during the generation and discrimination processes, thus forming a stable irradiance generation and prediction network. Finally, this irradiance generation and prediction network is used to perform dynamic mapping modeling on the meteorological cloud image fusion feature set, establishing a nonlinear mapping relationship between meteorological features and irradiance features, and constructing a feature fusion modeling channel.
[0036] Furthermore, in the method provided in the application embodiments, the dynamic mapping modeling based on the irradiation generator and the irradiation discriminator, and the construction of a feature fusion modeling channel, further includes: Loss assessment is performed on the irradiation generator and the irradiation discriminator to obtain generator performance loss parameters and discriminator performance loss parameters; based on the generator performance loss parameters and discriminator performance loss parameters, iterative adversarial training is performed on the irradiation generator and the irradiation discriminator to construct an irradiation generation prediction network; based on the irradiation generation prediction network, dynamic mapping modeling is performed on the meteorological cloud image fusion feature set to construct the feature fusion modeling channel.
[0037] In this embodiment, when evaluating the loss of the irradiance generator and the irradiance discriminator, the generator performance loss parameter and the discriminator performance loss parameter are calculated separately to measure the difference between the predicted results and the actual data. The generator performance loss parameter is calculated using the mean squared error method, reflecting the accuracy of the generator's irradiance distribution prediction by comparing the pixel-level differences between the generated irradiance map and the actual irradiance observation map. The discriminator performance loss parameter is calculated using the binary cross-entropy method, measuring the discriminator's classification ability in distinguishing between real and predicted samples by comparing the deviation between the discrimination result and the true label.
[0038] Next, iterative adversarial training is performed on the radiation generator and radiation discriminator based on the generator's performance loss parameters and the discriminator's performance loss parameters. In this process, an alternating optimization method is used. In each training round, the parameters of the radiation discriminator are fixed, while the parameters of the radiation generator are updated. This allows the radiation generator to minimize its own performance loss function through backpropagation, thereby generating radiation prediction results that are closer to the actual radiation distribution. Subsequently, the parameters of the radiation generator are fixed again, and the radiation discriminator is trained to maximize its discrimination accuracy, thus improving its ability to identify the difference between generated and actual radiation. Through multiple rounds of iterative training, the radiation generator and radiation discriminator continuously optimize their parameters in a game-like process, ultimately forming an radiation generation and prediction network that converges in both prediction and discrimination tasks.
[0039] Finally, based on the irradiance generation and prediction network, a dynamic mapping model is performed on the meteorological cloud image fusion feature set. In this process, the meteorological cloud image fusion feature set is input into the irradiance generation and prediction network. The network, through a multi-layer nonlinear mapping structure, converts photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features into irradiance distribution features. During the mapping process, the irradiance generation and prediction network automatically establishes a dynamic spatial and temporal response model based on the coupling relationship between meteorological features, ensuring that the output irradiance prediction results accurately reflect the impact of changes in meteorological conditions on irradiance distribution. Through this modeling process, a dynamic mapping relationship is formed between the meteorological cloud image fusion feature set and photovoltaic irradiance features, thus completing the feature fusion modeling channel.
[0040] Step S300: The meteorological cloud image recognition dual channel is used to extract features and fuse model the meteorological data of the multimodal air-to-ground station to generate a cloud irradiance mapping model.
[0041] In this embodiment, when using a dual-channel meteorological cloud image recognition system to extract features and fuse model multimodal air-to-ground station meteorological data, the cloud image feature extraction channel is first used to aggregate features from the multimodal air-to-ground station meteorological data. This process integrates multi-source meteorological information into a structured feature representation by unifying the encoding and feature mapping of photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features. This makes different modal data comparable and correlated in the same feature space, thereby forming a comprehensive feature set that reflects the spatiotemporal changes of clouds and the dynamic characteristics of meteorological conditions.
[0042] After obtaining the comprehensive feature set, the features are fused and modeled through a feature fusion modeling channel. This channel models the nonlinear correlation between features based on a multimodal feature fusion function and an irradiance generation and prediction network. The multimodal feature fusion function weights and combines features from different modalities, strengthening feature signals that contribute more significantly to irradiance changes. The irradiance generation and prediction network learns the complex mapping relationships between features through an adversarial training process between the irradiance generator and the irradiance discriminator, enabling the fused features to accurately reflect the dynamic relationship between meteorological state changes and irradiance response. Through multiple rounds of iterative optimization, the feature fusion modeling channel achieves a stable mapping from fused features to irradiance changes.
[0043] Finally, through the synergistic effect of the cloud image feature extraction channel and the feature fusion modeling channel, a cloud cluster irradiance mapping model is formed. This cloud cluster irradiance mapping model comprehensively correlates the cloud distribution characteristics reflected in the meteorological cloud image with the ground meteorological conditions, establishing a dynamic mapping relationship between meteorological characteristics and changes in photovoltaic irradiance.
[0044] Step S400: Obtain the historical power generation dataset of the photovoltaic power station, perform prediction training fitting and accuracy improvement optimization based on the cloud irradiance mapping model and the historical power generation dataset of the photovoltaic power station, construct a photovoltaic power optimization predictor, and perform power prediction control through the photovoltaic power optimization predictor.
[0045] In this embodiment, the historical power generation dataset of a photovoltaic power plant is first obtained from a preset database. After obtaining the historical power generation dataset, the power generation, component temperature, inverter output voltage and current, and corresponding irradiance information in the dataset are first cleaned and time-aligned to ensure correspondence with meteorological data in the time dimension. Subsequently, a multi-task photovoltaic prediction framework is constructed, which includes a photovoltaic multi-time period prediction task and an irradiance-power conversion task. By training and fitting the historical power generation dataset of the photovoltaic power plant, a basic photovoltaic power predictor and an irradiance-power converter are obtained.
[0046] After obtaining the basic photovoltaic power predictor and irradiance-to-power converter, they are jointly trained using a cloud irradiance mapping model. By utilizing the irradiance distribution characteristics provided by the cloud irradiance mapping model, multiple rounds of training and parameter correction are performed on the basic photovoltaic power predictor and irradiance-to-power converter to ensure that the predicted output is consistent with the actual power changes. After multiple optimizations, an optimized photovoltaic power predictor is constructed.
[0047] Next, when performing power prediction control through the photovoltaic power optimization predictor, real-time meteorological cloud image data, satellite cloud image data, and ground meteorological station observation data are input into the photovoltaic power optimization predictor. The photovoltaic power optimization predictor uses the future irradiance distribution information generated by the cloud irradiance mapping model to calculate the power prediction value for different time periods and obtain the power prediction result.
[0048] After obtaining the power prediction results, the photovoltaic power optimization predictor compares and analyzes these results with the real-time operating data of the photovoltaic power plant. Based on the deviation between the predicted and actual power, it generates a power adjustment command. This command guides the operation and scheduling of the inverters, combiner units, and energy storage devices in the photovoltaic power plant, ensuring stable power output within the predicted trend. When the prediction indicates a decrease in irradiance in the short term, the photovoltaic power optimization predictor will reduce output power or adjust the charging and discharging strategies of the energy storage devices in advance to balance power fluctuations. Conversely, when the prediction indicates an impending increase in irradiance, it will increase power output capacity in advance to maximize power generation.
[0049] Furthermore, the method provided in the application embodiments for constructing a photovoltaic power optimization predictor further includes: A multi-task framework for photovoltaic (PV) prediction is constructed, comprising a PV multi-time period prediction task and an irradiance-power conversion task. The historical power generation dataset of the PV power plant is trained and fitted according to the PV prediction multi-task framework to obtain a basic PV power predictor and an irradiance-power converter. The basic PV power predictor and the irradiance-power converter are then trained, fitted, and their accuracy improved and optimized using the cloud irradiance mapping model to construct an optimized PV power predictor.
[0050] In this embodiment, when constructing the photovoltaic (PV) prediction multi-task framework, the PV multi-time period prediction task and the irradiance-power conversion task are placed in the same training structure, allowing the two tasks to learn time-series patterns and power response patterns respectively while sharing input. Specifically, the PV multi-time period prediction task uses a long short-term memory network for sliding window modeling. The input consists of timestamp-aligned power generation, module temperature, inverter output voltage and current, and a sequence of meteorological elements; the output is a power generation sequence with multiple prediction time intervals. The irradiance-power conversion task uses a feedforward neural network to map and learn irradiance, module temperature, inverter output voltage and current, etc., at the same moment, outputting an estimated power generation value for that moment. This structure yields a PV prediction multi-task framework that simultaneously possesses sequence prediction capabilities and irradiance response characterization capabilities.
[0051] Next, following the multi-task framework for photovoltaic (PV) prediction, a supervised learning process is employed to train and fit the historical power generation dataset of PV power plants, using mean squared error as the loss function. The PV multi-time-period prediction task is trained using historical multi-step power sequences as input and future multi-step power as the target, enabling the model to capture the temporal dependence of intraday and short-term fluctuations. The irradiance-power conversion task is trained using irradiance and equipment operation measurements at the same moment as input and the measured power at that moment as the target, enabling the model to learn the instantaneous impact of irradiance changes on power output. During training, standardization is used to ensure that inputs of different dimensions are within a comparable range, and early stopping strategies and cross-validation are employed to suppress overfitting. After training, a basic PV power predictor for multi-time-period sequence prediction and an irradiance-power converter for instantaneous mapping are obtained.
[0052] Finally, the cloud irradiance mapping model was combined with the basic photovoltaic power predictor and the irradiance-power converter for prediction training, fitting, and accuracy improvement optimization. In this process, the cloud irradiance mapping model was first coupled with the irradiance-power converter for power prediction. By inputting the irradiance distribution characteristics output by the cloud irradiance mapping model into the irradiance-power converter, it was able to generate power prediction results corresponding to irradiance under different cloud movement and irradiance variation conditions, thus obtaining the cloud power predictor. Next, the cloud power predictor and the basic photovoltaic power predictor were jointly trained. During training, the dynamic irradiance information provided by the cloud power predictor was used to correct the time series prediction results of the basic photovoltaic power predictor. The parameters of both were continuously optimized through gradient backpropagation and error minimization methods, making the predicted power more consistent with the actual changing trend. After joint training and multiple rounds of iterative optimization, an optimized photovoltaic power predictor was constructed.
[0053] Furthermore, in the method provided in the application embodiments, the basic photovoltaic power predictor and the irradiance-power converter are subjected to prediction training fitting and accuracy improvement optimization in combination with the cloud irradiance mapping model to construct a photovoltaic power optimization predictor, which further includes: The cloud irradiance mapping model and the irradiance-power converter are coupled for power prediction to obtain a cloud power predictor; the cloud power predictor is then combined with the basic photovoltaic power predictor for prediction training, fitting, and accuracy improvement optimization to construct a photovoltaic power optimization predictor.
[0054] In this embodiment, when coupling the cloud irradiance mapping model and the irradiance-power converter for power prediction, the cloud irradiance mapping model is first used to generate an irradiance feature sequence containing cloud distribution characteristics, irradiance intensity variation trends, and spatial irradiance distribution information. This irradiance feature sequence is then injected into the irradiance-power converter. The converter uses a nonlinear mapping method to correlate the irradiance variation characteristics with the photovoltaic module power output characteristics, learning the dynamic response relationship between irradiance changes and power output. Through repeated training and error correction, the irradiance-power converter can accurately predict photovoltaic power output based on cloud movement characteristics under different meteorological conditions, thus forming a cloud power predictor.
[0055] Subsequently, the cloud power predictor and the basic photovoltaic power predictor are combined for prediction training, fitting, and accuracy improvement optimization. In this step, feature fusion and parameter co-adjustment are first performed on the cloud power predictor and the basic photovoltaic power predictor through joint prediction training, so that time series features and irradiance variation features are coupled in the same prediction framework to obtain a joint photovoltaic power predictor. Then, an online learning mechanism is introduced on the basis of the joint photovoltaic power predictor. By updating the model parameters and error correction in real time, adaptive optimization to dynamic meteorological changes is achieved, thereby constructing an optimized photovoltaic power predictor.
[0056] Furthermore, in the method provided in the application embodiments, the method of combining the cloud power predictor with the basic photovoltaic power predictor for prediction training fitting and accuracy improvement optimization to construct a photovoltaic power optimized predictor further includes: The cloud power predictor and the basic photovoltaic power predictor are jointly trained and fitted to obtain a photovoltaic power joint predictor; an online learning mechanism is introduced to improve and optimize the accuracy of the photovoltaic power joint predictor, and the photovoltaic power optimized predictor is constructed.
[0057] In this embodiment, the power prediction sequence output by the cloud power predictor and the power prediction result output by the basic photovoltaic power predictor are first synchronized in time and normalized in terms of features, making the two prediction results comparable within the same time scale and numerical range. Then, a multi-input regression fusion algorithm is used to construct a joint prediction framework, using the irradiance variation characteristics of the cloud power predictor and the time series characteristics of the basic photovoltaic power predictor as input variables, and the measured power in the historical power generation dataset of photovoltaic power plants as the supervision target for joint training. During joint training, mean squared error is used as the loss function, and the parameters of the cloud power predictor and the basic photovoltaic power predictor are iteratively updated through backpropagation, enabling the model to simultaneously optimize prediction accuracy and timeliness based on the fusion of meteorological irradiance characteristics and time series characteristics. After multiple rounds of iterative training, a photovoltaic power joint predictor that can simultaneously reflect irradiance variation and temporal patterns is obtained.
[0058] After obtaining the photovoltaic power joint predictor, an online learning mechanism is introduced to improve its accuracy. This mechanism uses real-time input meteorological cloud imagery, satellite cloud imagery, and ground meteorological station observations as incremental samples, employing a sliding window update strategy and incremental gradient descent algorithm to update model parameters. Specifically, when new meteorological samples and measured power data arrive, the error between the current predicted output and the measured power is calculated. This error is used as the basis for parameter correction, locally adjusting the weight matrix of the photovoltaic power joint predictor to make the predicted output closer to the actual power changes under current meteorological conditions. This process achieves adaptive optimization of the model, enabling the photovoltaic power joint predictor to maintain prediction stability and high accuracy under different seasons, different cloud movement patterns, and sudden meteorological changes. Through continuous optimization via the online learning mechanism, an optimized photovoltaic power predictor is ultimately constructed.
[0059] In summary, the embodiments of this application have at least the following technical effects: This application collects multimodal air-to-ground meteorological data from a target photovoltaic (PV) power plant. This multimodal air-to-ground meteorological data includes PV all-sky image data, satellite cloud image data, and ground meteorological station observation data. A dual-channel meteorological cloud image recognition system is constructed, comprising a cloud image feature extraction channel and a feature fusion modeling channel. The system uses this dual-channel system to extract features and fuse modeling from the multimodal air-to-ground meteorological data, generating a cloud irradiance mapping model. Historical power generation data from the PV power plant is acquired. Based on the cloud irradiance mapping model and the historical power generation data, prediction training, fitting, and accuracy improvement optimization are performed to construct a PV power optimization predictor. Power prediction control is then performed using this PV power optimization predictor. This invention addresses the technical problems of insufficient utilization of multi-source meteorological data, low cloud image recognition accuracy, and large prediction errors in existing PV power prediction technologies. By constructing a dual-channel meteorological cloud image recognition system and fusing multimodal meteorological data for modeling and optimization, the technical effect of improving the accuracy and stability of PV power prediction is achieved.
[0060] Example 2 is based on the same inventive concept as the photovoltaic power prediction accuracy improvement method based on meteorological cloud map recognition in the previous examples, such as... Figure 2 As shown, this application provides a photovoltaic power prediction accuracy improvement system based on meteorological cloud image recognition. The system and method embodiments in this application are based on the same inventive concept. The system includes: The meteorological data acquisition module 11 is used to acquire multimodal air-to-ground meteorological data of the target photovoltaic power station. The multimodal air-to-ground meteorological data includes photovoltaic all-sky image data, satellite cloud image data, and ground meteorological station observation data. The channel construction module 12 is used to build a dual channel for meteorological cloud image recognition. The dual channel for meteorological cloud image recognition includes a cloud image feature extraction channel and a feature fusion modeling channel. The mapping model generation module 13 is used to perform feature extraction and fusion modeling on the multimodal air-to-ground meteorological data using the dual channel for meteorological cloud image recognition to generate a cloud irradiance mapping model. The prediction module 14 is used to obtain the historical power generation dataset of the photovoltaic power station, perform prediction training fitting and accuracy improvement optimization based on the cloud irradiance mapping model and the historical power generation dataset of the photovoltaic power station, construct a photovoltaic power optimization predictor, and perform power prediction control through the photovoltaic power optimization predictor.
[0061] Furthermore, the system is also used to implement the following functions: Based on the modal characteristics of the multimodal air-to-ground meteorological data, the data preprocessing steps are analyzed to obtain the multimodal data preprocessing step channel; according to the meteorological cloud image feature extraction target, the output data of the multimodal data preprocessing step channel is correlated by a feature algorithm to determine the multimodal data feature extraction channel; the multimodal data preprocessing step channel and the multimodal data feature extraction channel are connected in series and integrated to obtain the cloud image feature extraction channel; a feature fusion modeling channel is constructed, and the feature fusion modeling channel is linked to the cloud image feature extraction channel to build the dual-channel meteorological cloud image recognition system.
[0062] Furthermore, the system is also used to implement the following functions: Based on the output features of the cloud image feature extraction channel, photovoltaic all-sky image features, satellite cloud image features, ground meteorological features, and corresponding photovoltaic irradiance features are obtained; a multimodal feature fusion function is constructed, and the photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features are weighted and fused based on the multimodal feature fusion function to obtain a meteorological cloud image fusion feature set; a generative adversarial network is introduced to dynamically map and model the meteorological cloud image fusion feature set and the corresponding photovoltaic irradiance features, and a feature fusion modeling channel is constructed.
[0063] Furthermore, the system is also used to implement the following functions: Based on the multimodal air-to-ground meteorological data, meteorological cloud cover monitoring data is obtained; the impact of the multimodal air-to-ground meteorological data is assessed according to the meteorological cloud cover monitoring data to determine the precise impact of the multimodal data; based on the precise impact of the multimodal data, the dynamic decision weights of the multimodal data are determined; based on the dynamic decision weights of the multimodal data, a weighted fitting of multimodal features is performed to construct the multimodal feature fusion function.
[0064] Furthermore, the system is also used to implement the following functions: Based on the generative adversarial network, an initial generator and an initial discriminator are obtained; the injected cloud image irradiance empirical model is used to train the initial generator for irradiance prediction based on the meteorological cloud image fusion feature set to obtain an irradiance generator; the irradiance generator is trained for irradiance discrimination according to the corresponding photovoltaic irradiance features to generate an irradiance discriminator; dynamic mapping modeling is performed based on the irradiance generator and the irradiance discriminator to construct a feature fusion modeling channel.
[0065] Furthermore, the system is also used to implement the following functions: Loss assessment is performed on the irradiation generator and the irradiation discriminator to obtain generator performance loss parameters and discriminator performance loss parameters; based on the generator performance loss parameters and discriminator performance loss parameters, iterative adversarial training is performed on the irradiation generator and the irradiation discriminator to construct an irradiation generation prediction network; based on the irradiation generation prediction network, dynamic mapping modeling is performed on the meteorological cloud image fusion feature set to construct the feature fusion modeling channel.
[0066] Furthermore, the system is also used to implement the following functions: A multi-task framework for photovoltaic (PV) prediction is constructed, comprising a PV multi-time period prediction task and an irradiance-power conversion task. The historical power generation dataset of the PV power plant is trained and fitted according to the PV prediction multi-task framework to obtain a basic PV power predictor and an irradiance-power converter. The basic PV power predictor and the irradiance-power converter are then trained, fitted, and their accuracy improved and optimized using the cloud irradiance mapping model to construct an optimized PV power predictor.
[0067] Furthermore, the system is also used to implement the following functions: The cloud irradiance mapping model and the irradiance-power converter are coupled for power prediction to obtain a cloud power predictor; the cloud power predictor is then combined with the basic photovoltaic power predictor for prediction training, fitting, and accuracy improvement optimization to construct a photovoltaic power optimization predictor.
[0068] Furthermore, the system is also used to implement the following functions: The cloud power predictor and the basic photovoltaic power predictor are jointly trained and fitted to obtain a photovoltaic power joint predictor; an online learning mechanism is introduced to improve and optimize the accuracy of the photovoltaic power joint predictor, and the photovoltaic power optimized predictor is constructed.
[0069] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition, characterized in that, The method includes: Collect multimodal air-to-ground meteorological data of the target photovoltaic power station, including photovoltaic all-sky image data, satellite cloud image data and ground meteorological station observation data; A dual-channel meteorological cloud image recognition system is established, comprising a cloud image feature extraction channel and a feature fusion modeling channel. The meteorological cloud image recognition dual-channel is used to extract features and fuse model the meteorological data from the multimodal air-to-ground station, generating a cloud irradiance mapping model; Historical power generation datasets of photovoltaic power plants are obtained. Based on the cloud irradiance mapping model and the historical power generation datasets of photovoltaic power plants, prediction training and fitting are performed and accuracy is improved and optimized to construct a photovoltaic power optimization predictor. Power prediction control is then performed through the photovoltaic power optimization predictor.
2. The method for improving photovoltaic power prediction accuracy based on meteorological cloud image recognition as described in claim 1, characterized in that, Establish a dual-channel system for meteorological cloud image recognition, including: Based on the modal characteristics of the multimodal air-to-ground meteorological data, the data preprocessing steps are analyzed to obtain the multimodal data preprocessing step channel; Based on the meteorological cloud image feature extraction target, the output data of the multimodal data preprocessing step channel are correlated using a feature algorithm to determine the multimodal data feature extraction channel; The multimodal data preprocessing step channel and the multimodal data feature extraction channel are connected in series and integrated to obtain the cloud map feature extraction channel; A feature fusion modeling channel is constructed, and the feature fusion modeling channel is linked to the cloud image feature extraction channel to build a dual channel for meteorological cloud image recognition.
3. The method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition as described in claim 2, characterized in that, Constructing a feature fusion modeling channel includes: Based on the output features of the cloud image feature extraction channel, photovoltaic all-sky image features, satellite cloud image features, ground meteorological features, and corresponding photovoltaic irradiance features are obtained. A multimodal feature fusion function is constructed, and the photovoltaic all-sky image features, satellite cloud image features, and ground meteorological features are weighted and fused based on the multimodal feature fusion function to obtain a meteorological cloud image fusion feature set; Generative adversarial networks are introduced to dynamically map and model the meteorological cloud image fusion feature set and the corresponding photovoltaic irradiance features, thereby constructing a feature fusion modeling channel.
4. The method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition as described in claim 3, characterized in that, Constructing a multimodal feature fusion function includes: Meteorological cloud cover monitoring data is obtained based on the multimodal air-to-ground meteorological data. The impact of the multimodal air-to-ground station meteorological data is assessed based on the meteorological cloud cover monitoring data to determine the precise impact of the multimodal data. Based on the precise impact of the multimodal data, determine the dynamic decision weights for the multimodal data; Based on the dynamic decision weights of the multimodal data, a weighted fitting of multimodal features is performed to construct the multimodal feature fusion function.
5. The method for improving photovoltaic power prediction accuracy based on meteorological cloud image recognition as described in claim 3, characterized in that, Generative adversarial networks are introduced to dynamically map and model the meteorological cloud image fusion feature set and the corresponding photovoltaic irradiance features, constructing a feature fusion modeling channel, including: Based on the generative adversarial network, an initial generator and an initial discriminator are obtained; The injected cloud image irradiance empirical model is used to train the initial generator for irradiance prediction based on the meteorological cloud image fusion feature set to obtain the irradiance generator. The irradiation generator is trained to discriminate based on the corresponding photovoltaic irradiation characteristics to generate an irradiation discriminator. Dynamic mapping modeling is performed based on the irradiation generator and the irradiation discriminator to construct a feature fusion modeling channel.
6. The method for improving photovoltaic power prediction accuracy based on meteorological cloud image recognition as described in claim 5, characterized in that, Dynamic mapping modeling is performed based on the irradiation generator and the irradiation discriminator, and a feature fusion modeling channel is constructed, including: A loss assessment is performed on the irradiation generator and the irradiation discriminator to obtain generator performance loss parameters and discriminator performance loss parameters. Based on the generator performance loss parameters and the discriminator performance loss parameters, iterative adversarial training is performed on the irradiation generator and the irradiation discriminator to construct an irradiation generation prediction network. Based on the irradiation generation and prediction network, the meteorological cloud image fusion feature set is dynamically mapped and modeled to construct the feature fusion modeling channel.
7. The method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition as described in claim 1, characterized in that, Constructing a photovoltaic power optimization predictor includes: A photovoltaic prediction multi-task framework is constructed, which includes photovoltaic multi-time period prediction tasks and irradiance-power conversion tasks. The photovoltaic power prediction multi-task framework is used to predict, train and fit the historical power generation dataset of the photovoltaic power plant to obtain a basic photovoltaic power predictor and an irradiance-power converter. By combining the cloud irradiance mapping model, the basic photovoltaic power predictor and irradiance-power converter are subjected to prediction training fitting and accuracy improvement optimization to construct a photovoltaic power optimized predictor.
8. The method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition as described in claim 7, characterized in that, By combining the cloud irradiance mapping model with the basic photovoltaic power predictor and irradiance-power converter, prediction training, fitting, and accuracy improvement optimization are performed to construct an optimized photovoltaic power predictor, including: The cloud irradiance mapping model and the irradiance-power converter are coupled for power prediction to obtain a cloud power predictor. By combining the cloud power predictor with the basic photovoltaic power predictor, a photovoltaic power optimization predictor is constructed through prediction training, fitting, and accuracy improvement optimization.
9. The method for improving the accuracy of photovoltaic power prediction based on meteorological cloud image recognition as described in claim 8, characterized in that, By combining the cloud power predictor with the basic photovoltaic power predictor for prediction training, fitting, and accuracy improvement optimization, a photovoltaic power optimization predictor is constructed, including: A joint prediction training and fitting process is performed on the cloud power predictor and the basic photovoltaic power predictor to obtain a joint photovoltaic power predictor. An online learning mechanism is introduced to improve and optimize the accuracy of the photovoltaic power joint predictor, and the photovoltaic power optimized predictor is constructed.
10. A photovoltaic power prediction accuracy improvement system based on meteorological cloud image recognition, characterized in that, The system is used to execute the photovoltaic power prediction accuracy improvement method based on meteorological cloud image recognition as described in any one of claims 1-9, and the system includes: The meteorological data acquisition module is used to collect multimodal air-to-ground meteorological data of the target photovoltaic power station. The multimodal air-to-ground meteorological data includes photovoltaic all-sky image data, satellite cloud image data and ground meteorological station observation data. The channel construction module is used to build a dual channel for meteorological cloud image recognition, which includes a cloud image feature extraction channel and a feature fusion modeling channel. The mapping model generation module is used to extract features and fuse model the multimodal air-to-ground station meteorological data using the dual-channel meteorological cloud image recognition to generate a cloud irradiance mapping model. The prediction module is used to acquire historical power generation datasets of photovoltaic power plants, perform prediction training and fitting and accuracy improvement optimization based on the cloud irradiance mapping model and the historical power generation datasets of photovoltaic power plants, construct a photovoltaic power optimization predictor, and perform power prediction control through the photovoltaic power optimization predictor.