Multi-mode photovoltaic power generation power prediction system based on logistics unmanned aerial vehicle cluster

Abstract

The invention relates to the field of power prediction, and discloses a multi-modal photovoltaic power generation power prediction system based on a logistics unmanned aerial vehicle cluster, and the system comprises the steps: obtaining multi-modal data through employing the unmanned aerial vehicle cluster; processing and evaluating the multi-modal data, and evaluating the accuracy, integrity and consistency of each modal data; constructing a respective prediction model for each model, and performing iterative training by using the obtained data to obtain a feature extraction value of each modal; and carrying out multi-modal fusion on the feature extraction values obtained by each modal based on Transform, inputting the feature extraction values into an integrated prediction model, and predicting to obtain the power generation power of each optical storage and charging cluster. Through multi-modal data complementarity enhancement, the unmanned aerial vehicle collects local cloud layer images, high-resolution meteorological sensor data and historical power generation time sequence data, and the space-time blind area of traditional single meteorological station data is effectively made up. A complementary prediction mechanism is formed by adopting the dynamic feature extraction of the CNN on the cloud layer image and the time sequence modeling of the LSTM on the historical power generation data in combination with the physical rule regression analysis of the meteorological data.

Description

Technical Field

[0001] This invention relates to the field of power prediction technology, and more specifically, to a multimodal photovoltaic power generation prediction system based on a logistics drone swarm. Background Technology

[0002] As the global energy structure shifts towards renewable energy, photovoltaic power generation has attracted widespread attention due to its clean and efficient characteristics. However, the power output of photovoltaic power generation is subject to considerable uncertainty due to the complex influence of factors such as weather and geographical environment.

[0003] Traditional photovoltaic (PV) forecasting has limitations: 1. The meteorological data collected for PV forecasting is narrow in scope. Traditional methods rely on fixed weather stations or single-point monitoring, which only have limited sensing capabilities and are difficult to capture microclimate changes (such as local cloud movement and sudden weather events); 2. The application of drones is singular and inefficient: Currently, drones in the energy sector are mostly used for equipment inspection or single-point data collection. To achieve full coverage monitoring of dispersed integrated PV-storage-charging clusters, a large number of drones need to be deployed, which is costly and results in poor data coordination. Currently, there are cases where drones are used to collect environmental, weather, and cloud data at individual PV solar panel deployment points, but for integrated PV-storage-charging clusters, this singular application method is too costly. This technical solution combines a logistics drone cluster with a PV power station cluster. By utilizing the drone cluster to perform logistics tasks, it simultaneously collects short-, medium-, and long-distance weather data that affects PV power generation efficiency, thereby assisting each PV power station in achieving accurate PV power generation efficiency forecasting.

[0004] To address the limitations of existing photovoltaic power generation forecasting methods, such as the limited range of meteorological data collection and the single application of drones serving only a single photovoltaic point, this patent proposes a multimodal photovoltaic power generation forecasting system based on the integration of photovoltaic, energy storage, and charging clusters with wide-area collaboration of drones. Summary of the Invention

[0005] To address at least one of the aforementioned problems, the present invention first provides a multimodal photovoltaic power generation prediction system based on a logistics drone swarm, comprising the following steps:

[0006] Step S1: Acquire multimodal data;

[0007] Step S2: Process and evaluate the multimodal data, assessing the accuracy, completeness, and consistency of each modality.

[0008] Step S3: Build a prediction model for each model and use the obtained data for iterative training to obtain the feature extraction values ​​for each modality;

[0009] Step S4: The feature extraction values ​​obtained from each mode are fused using Transformer, and then input into the integrated prediction model to predict the power generation of each photovoltaic-storage-charging cluster.

[0010] Optionally, the multimodal data obtained in step S1 mainly includes: historical photovoltaic power generation data; cloud images and environmental images at short, medium and long distances above each photovoltaic-storage-charging cluster point obtained by using a drone swarm; local weather forecast information obtained by web crawling; and real-time temperature and humidity information of each location obtained by temperature and humidity sensors.

[0011] Optionally, the specific data evaluation operations in step S2 include:

[0012] Step S21: Evaluate the authenticity of the data by comparing the corresponding values ​​in the multimodal data with the actual measured values ​​and calculating the error rate;

[0013] Step S22: Assess the completeness of the data, i.e., check whether the data contains all the necessary information, whether any values ​​are missing or important content is omitted.

[0014] Step S23: Evaluate the consistency of the data, that is, the logical consistency of multimodal data between different modalities and within the same modality.

[0015] Optionally, the specific operations in step S3 include:

[0016] Step S31: Perform feature engineering on historical power generation data and photovoltaic-storage-charging cluster tag data, extract features and input them into the first time series model (Prophet+LSTM) to obtain the periodicity and trend of the time series data, and extract the first time series dynamic feature value;

[0017] Step S32: Input the cloud image data and surrounding environment image data obtained by the drone into the second temporal model (CNN) to obtain the image edge, texture and coverage data, and extract the second spatial feature extraction value.

[0018] Step S33: Input the temperature, humidity, wind force and weather data into the third time series model (LSTM+Attention) to obtain the third feature extraction value of the time series meteorological-power generation correlation.

[0019] Optionally, the specific operations in step S31 include: taking historical power generation probabilities as input, labeling time-related features, inputting them into the Prophet model, and decomposing the power generation data into a trend term g(t), a seasonal term s(t), and a holiday term h(t):

[0020]

[0021] Then the Prophet residual The original power value is input into an LSTM for enhancement, and the corrected power timing characteristics are output, thereby compensating for the LSTM's neglect of macroscopic periodicity.

[0022] Optionally, the specific operations in step S32 include: taking the RGB cloud image captured by the drone as input, inputting a lightweight ResNet18 (pre-trained weight initialization), and modifying the last layer of the model by changing global average pooling to a fully connected layer, so that it can better extract the corresponding feature vectors and provide micro-environment information that weather stations cannot cover.

[0023] Optionally, the specific operations in step S33 include: taking meteorological time series data as input, using an Encoder-Decoder structure, extracting meteorological time series features through LSTM, then dynamically focusing on key time steps through an Attention mechanism, and finally outputting the correlation feature vector between meteorology and power generation; setting a two-level attention mechanism in the attention mechanism, namely a time attention mechanism and a variable attention mechanism.

[0024] Optionally, the specific operations in step S4 include:

[0025] Step S41: Embed the feature extraction values ​​of different modalities as input into the Transformer model, and use the self-attention mechanism of the Transformer to dynamically capture the correlation and importance weights between each modality;

[0026] Step S42: Extract features from multiple subspaces through a multi-head attention mechanism, and further fuse these features through a feedforward network to generate a comprehensive feature representation;

[0027] Step S43: Input the fused features into the integrated prediction model to predict the power generation of each photovoltaic-storage-charging cluster.

[0028] Step S44: Obtain the photovoltaic power prediction trajectory based on the time and the corresponding photovoltaic cluster point.

[0029] Optionally, the specific operations in steps S41 and S42 include:

[0030] The feature vectors extracted from the three modalities are as follows:

[0031] Meteorological characteristics: Image features: Historical power generation characteristics:

[0032] When inputting various features into the Transformer, the dimensions need to be unified through linear projection:

[0033] ,in The weight matrix is ​​a learnable weight matrix;

[0034] The multi-head attention mechanism in the Transformer coding layer can be represented as:

[0035] Each head is:

[0036]

[0037] Where Q, K, and V are obtained by linear transformation of the input feature Z;

[0038] After N layers of Transformer encoding, the [CLS] flag bit is taken as the fused feature:

[0039]

[0040] The final prediction is:

[0041]

[0042] That is, the power generation capacity of each photovoltaic-storage-charging cluster.

[0043] Optionally, step S1 includes collecting data through a drone swarm, which may include a logistics drone swarm and a photovoltaic power station swarm.

[0044] Compared with existing technologies, the multimodal photovoltaic power generation prediction system based on logistics drone swarms in this invention has the following outstanding effects and advantages:

[0045] 1. Significantly improves photovoltaic power generation forecast accuracy: Through multimodal data complementarity enhancement, it effectively compensates for the spatiotemporal blind spots of traditional single meteorological station data by fusing local cloud images (spatial features) collected by UAVs, high-resolution meteorological sensor data (temperature, humidity, wind speed), and historical power generation time series data (temporal features). Specifically, it employs CNN (Convolutional Neural Network) for dynamic feature extraction of cloud images (such as cloud movement speed and coverage) and LSTM for temporal modeling of historical power generation data, combined with regression analysis of the physical laws of meteorological data, to form a complementary prediction mechanism.

[0046] 2. Reduce the cost of drone applications and improve operational efficiency: This invention uses the same drone swarm to perform meteorological detection (cloud data collection) in a time-sharing manner, collecting weather data for each photovoltaic power station within a short, medium and long range, thus expanding the scope of data collection to include more photovoltaic power stations.

[0047] 3. Economic advantages: Hardware investment costs are reduced through the reuse of drone functions and optimization of cluster scheduling.

[0048] 4. Environmental benefits: Accurate forecasting reduces carbon emissions from standby thermal power units.

[0049] 5. Technical versatility: The proposed multimodal fusion framework can be extended to other new energy fields such as wind power and hydrogen energy, requiring only adjustments to the data input type and model parameters. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of a multimodal photovoltaic power generation prediction system based on a logistics drone swarm, according to an embodiment of the present invention.

[0051] Figure 2 This is a flowchart of a multimodal photovoltaic power generation prediction system based on a logistics drone swarm, according to an embodiment of the present invention.

[0052] Figure 3 This is a schematic diagram illustrating the relationships between the models in step S3 of this embodiment of the invention;

[0053] Figure 4 This is a flowchart of the multimodal fusion process in step S4 of an embodiment of the present invention;

[0054] Figure 5 This is a comparison chart of the actual power generation and the predicted power generation on a certain day. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] This invention provides a multimodal photovoltaic power generation prediction system based on a logistics drone swarm, combined with... Figures 1 to 5 As shown, the main technical intersection areas of this invention include:

[0057] 1. Accurate prediction of photovoltaic power generation;

[0058] 2. Coordinated control of UAV swarms and collection of short, medium and long-range weather data;

[0059] 3. Multimodal data collection and construction, training, and prediction of large multimodal models.

[0060] This technical solution includes:

[0061] like Figure 1 The four components shown are: 1) an integrated photovoltaic, energy storage and charging cluster, 2) a wide-area unmanned aerial vehicle (UAV) operation system, 3) a data center, and 4) a multimodal prediction center.

[0062] A photovoltaic-storage-charging integrated cluster is a comprehensive energy system that integrates photovoltaic power generation, energy storage systems, and charging facilities. The integrated system consists of photovoltaic power generation devices, energy storage equipment, and charging piles, achieving energy conversion and storage through power electronics technology. A drone wide-area operation system can collect data near photovoltaic power plants while simultaneously charging the drones using the integrated cluster's charging facilities. This system includes logistics drone clusters and photovoltaic power plant clusters. The acquired data is transmitted to a data center, which integrates and processes the data. A multimodal prediction center then models and deeply fuses the data to form a final prediction model, forecasting the power generation value of each photovoltaic-storage-charging cluster.

[0063] like Figure 2 As shown, this application includes the following steps:

[0064] Step S1: Acquire multimodal data using a drone swarm; collect historical power generation data of photovoltaics through a drone swarm, which includes a logistics drone swarm and a photovoltaic power station swarm; acquire short, medium and long-range cloud and environmental images over each photovoltaic-storage-charging cluster point using the drone swarm; obtain local weather forecast information through web crawling and other means, and obtain real-time temperature and humidity information for each point through temperature and humidity sensors.

[0065] Current drone applications are limited and inefficient: In the energy sector, drones are primarily used for equipment inspection or single-point data collection. To achieve full coverage monitoring of dispersed integrated photovoltaic-storage-charging clusters, a large number of drones need to be deployed, resulting in high costs and poor data coordination. While some drones are used to collect environmental, weather, and cloud data from individual photovoltaic solar panel deployment sites, this single-application approach is prohibitively expensive for integrated photovoltaic-storage-charging clusters. Typically, multiple drones are needed for a single site, and these drones are specifically designed for photovoltaic power generation forecasting. Because they serve a single site, their operational range is limited, making it difficult to capture microclimate changes (such as local cloud movement). Furthermore, equipping a single site with numerous drones is costly and results in low drone utilization.

[0066] This technical solution combines distributed photovoltaic power stations with an existing logistics drone swarm. This results in a larger number of drones collecting weather data for the photovoltaic power stations, with a wider operating range and more accurate data, leading to more accurate photovoltaic power generation forecasts. Simultaneously, the drone swarm, tasked with logistics, naturally operates and collects weather data over a wider area while completing its logistics missions. Since the drone swarm serves multiple photovoltaic sites simultaneously, it can be reused across sites, reducing costs. The drone swarm in this application moves between sites during logistics operations, requiring fewer drones to achieve the same effect, thus reducing hardware deployment costs. Because the drones in our solution simultaneously perform photovoltaic forecasting data collection and logistics tasks, drone utilization is also higher. The combination of the logistics drone swarm and the photovoltaic power station swarm utilizes the drone swarm's logistics operations to simultaneously collect short-, medium-, and long-range weather data affecting photovoltaic power generation efficiency, thereby assisting each photovoltaic power station in achieving accurate photovoltaic power generation efficiency forecasts.

[0067] For drones to perform logistics and weather monitoring in a time-sharing or simultaneous manner, drone resources can be dynamically allocated, prioritizing the collection of data from areas with the greatest impact on forecasts. Specifically, the following operations can be performed:

[0068] Dynamic scheduling algorithm for UAV swarms based on predictive sensitivity: Establish a sensitivity assessment model and calculate the Weather Sensitivity Index (WSI) of each photovoltaic-storage-charging cluster in real time.

[0069] WSI=α*( Prediction error / Cloud coverage) + β*( Prediction error / Temperature gradient (α, β are weights, learned through regression from historical data)

[0070] Develop drone scheduling strategies: prioritize drone monitoring in high WSI areas; automatically overlay meteorological data collection points when planning logistics routes (e.g., detour by 5% to cover highly sensitive areas).

[0071] Increased hardware support: The drone is equipped with a retractable weather probe that automatically deploys to collect data during flight (avoiding energy consumption while hovering).

[0072] Step S2: Process and evaluate the multimodal data. After receiving the multimodal data obtained in Step S1, it is necessary to ensure the accuracy, completeness, and consistency of the data. This ensures that no errors occur during Step S3. Specific data evaluation operations include:

[0073] For the collected multimodal data (meteorological, image, time-series power generation data, etc.), based on technical principles and practical application requirements, suitable data processing methods can be summarized into the following four categories:

[0074] The first category is time-series data processing methods, which are applicable to time-related sequences such as historical power generation, temperature / humidity sensor sequences, and wind speed.

[0075] ① Using statistical modeling methods:

[0076] The Prophet model decomposes trend, seasonal, and holiday components, making it suitable for photovoltaic power generation data with obvious periodicity.

[0077] ARIMA / SARIMA models are suitable for stationary time series or time series data with approximately linear characteristics.

[0078] ② Using deep learning models:

[0079] LSTM / GRU model: solves long-term dependency problems (e.g., using LSTM to enhance Prophet residuals) and captures the temporal patterns of power generation data.

[0080] Transformer model: Captures long-distance dependencies through self-attention mechanism, suitable for ultra-short-term prediction of multimodal fusion.

[0081] The second category is spatial data processing methods, which are applicable to space-related sequences such as cloud images, environmental RGB images, and infrared thermal imaging captured by UAVs.

[0082] ① Using traditional image processing methods:

[0083] Threshold segmentation + morphological operations: Quickly extract cloud-covered areas.

[0084] Optical flow: Estimating cloud movement speed using consecutive frame images.

[0085] ② Using deep learning models:

[0086] CNN (Convolutional Neural Network): The ResNet18 and MobileNet (lightweight deployment) models are built as the basic architecture to extract spatial features such as cloud texture, edges, and coverage, and generate feature vectors.

[0087] U-Net: Semantic segmentation of cloud regions, quantification of cloud occlusion ratio.

[0088] The third category is multimodal data fusion methods, which are suitable for integrating meteorological data, image features, and historical power generation time series.

[0089] ① Utilizing traditional fusion methods:

[0090] Weighted average / feature concatenation: simple but ignores the correlation between modalities (e.g., directly concatenating LSTM output and CNN features).

[0091] ②Utilizing a deep learning fusion architecture:

[0092] Late Fusion: Each modality is trained independently, and the output results are fused (e.g., features are extracted from each modality first).

[0093] Utilizing attention-based fusion:

[0094] Transformer: Dynamically weights the importance of each modality (e.g., by assigning weights through multi-head attention).

[0095] Cross-attention: Establishing intermodal relationships (e.g., the interaction between cloud image features and meteorological data).

[0096] Multi-task learning (MTL): Sharing underlying features to simultaneously optimize power generation forecasting and cloud segmentation tasks.

[0097] The fourth category is lightweight and edge computing methods, which are suitable for real-time processing of drone data and reduce transmission bandwidth requirements.

[0098] ① Utilizing model compression technology:

[0099] Knowledge distillation: Large models guide the training of small models (e.g., using ResNet50 to guide ResNet18).

[0100] Quantization: FP32 → INT8 reduces computational cost.

[0101] ②Utilize the edge-friendly model:

[0102] Lightweight CNNs: MobileNetV3, EfficientNet-Lite (suitable for cloud recognition deployment on drones).

[0103] Simplified time series models: TinyLSTM, Temporal Convolutional Networks (TCN).

[0104] The above data processing methods are all existing technologies. The specific operations of data processing will not be described in detail here. Appropriate methods should be selected to process different types of data.

[0105] Step S21: Evaluate the authenticity of the data. For example, for data where true values ​​can be obtained, such as temperature and humidity in meteorological data, compare the corresponding values ​​in the multimodal data with the actual measured values ​​and calculate the error rate. For example, if the actual temperature is 25℃, but the data records it as 24℃, then the error is 1℃.

[0106] Step S22: Assess the completeness of the data, i.e., check whether the data contains all the necessary information, whether any values ​​are missing or important content is omitted.

[0107] Step S23: Evaluate the consistency of the data, that is, the logical consistency of multimodal data between different modalities and within the same modality.

[0108] The processed data is divided into multiple datasets, trained and tested using an LSTM network to obtain prediction errors. Then, the prediction model is corrected based on the prediction errors to obtain an optimized prediction model, thereby enhancing the reliability of the prediction model.

[0109] like Figure 3 As shown, step S3 involves constructing a prediction model for each model and iteratively training it using the obtained data to obtain the feature extraction values ​​for each modality. Step S3 uses the data evaluated in step S2 as input. The model is divided into three parts. Because the collected modal data are of different types, a late fusion method is used. Therefore, the training model is divided into three parts, specifically including:

[0110] Step S31: Perform feature engineering on historical power generation data and photovoltaic-storage-charging cluster tag data, extract features and input them into the first time series model (Prophet+LSTM). Based on the Prophet model and LSTM (Long Short-Term Memory model), extract the first time series dynamic feature values ​​such as periodicity and trend of the time series data.

[0111] The specific operations in step S31 include: taking historical power generation probabilities as input, labeling time-related features, inputting them into the Prophet model, and decomposing the power generation data into a trend term g(t), a seasonal term s(t), and a holiday term h(t):

[0112]

[0113] Then the Prophet residual The original power value is input into the LSTM for enhancement, and the corrected power timing characteristics are output, thereby effectively compensating for the LSTM's neglect of macroscopic periodicity.

[0114] Step S32: Input the cloud image data and surrounding environment images obtained by the drone into the second temporal model (CNN). Use the CNN convolutional neural network to extract features from the cloud image data to obtain spatial second feature extraction values ​​such as image edges, textures, and coverage.

[0115] The specific operations in step S32 include: taking the historical power generation probability as input, taking the RGB cloud image taken by the drone as input, inputting a lightweight ResNet18 (pre-trained weight initialization), and modifying the last layer of the model by changing the global average pooling to a fully connected layer, so that it can better extract the corresponding feature vectors, which can provide micro-environment information that weather stations cannot cover.

[0116] Step S33: Input meteorological data such as temperature, humidity, wind force, and weather into the third time series model (LSTM+Attention), and use LSTM+Attention to construct a time series meteorological-power generation correlation feature extraction model from meteorological data (X: temperature, humidity, wind force, etc.) to power generation data (Y);

[0117] The specific operations in step S33 include: taking meteorological time-series data as input, this module adopts an Encoder-Decoder structure, the core of which is to extract meteorological time-series features through LSTM, then dynamically focus on key time steps through the Attention mechanism, and finally output the correlation feature vector between meteorology and power generation; in the attention mechanism, a two-level attention mechanism is set, the interpretability of which is time attention, which can extract which historical moments are the most critical; and variable attention mechanism, which can extract which meteorological factors are dominant.

[0118] Furthermore, when processing image / time series / meteorological data, implicit correlations between modalities can be utilized: for example, the spatial mapping between cloud images and ground humidity, and specific cloud textures corresponding to historical power generation troughs.

[0119] This approach leverages physics-guided data augmentation for correlation mining, using graph neural networks (GNNs) to construct modal relationship graphs and establish node and edge weights. For example:

[0120] Node = {cloud pixel block, weather station, photovoltaic module};

[0121] Edge weight = correlation between cloud movement direction and wind speed vector;

[0122] When meteorological data for a certain area is missing, virtual meteorological data can be generated through GNN relationship graphs. For example, a correlation model of virtual humidity = f(cloud texture features, measured humidity of adjacent stations, wind speed) can be constructed to establish virtual humidity and make up for the lack of humidity data by using the associated cloud texture features, measured humidity of adjacent stations and wind speed.

[0123] This further enhances the data, supplementing data gaps caused by drone communication interruptions. By leveraging physical principles (such as cloud-wind-humidity propagation models), it improves the robustness of predictions with small sample sizes, overcoming the limitations of traditional data augmentation which is limited to a single modality and strengthening the correlation between different modalities.

[0124] like Figure 4 As shown, in step S4, the feature extraction values ​​obtained from each mode are fused using Transformer and input into the integrated prediction model to predict the power generation of each photovoltaic-storage-charging cluster.

[0125] Step S41: First, embed the feature extraction values ​​of different modalities as input into the Transformer model, and use the self-attention mechanism of the Transformer to dynamically capture the correlation and importance weights between each modality.

[0126] Step S42: Through the multi-head attention mechanism, the model can extract features from multiple subspaces and further fuse these features through a feedforward network to generate a comprehensive feature representation.

[0127] The specific operations in steps S41 and S42 include:

[0128] The feature vectors extracted from the three modalities are as follows:

[0129] Meteorological characteristics: Image features: Historical power generation characteristics:

[0130] When inputting various features into the Transformer, the dimensions need to be unified through linear projection:

[0131] ,in The weight matrix is ​​a learnable weight matrix;

[0132] The multi-head attention mechanism in the Transformer coding layer can be represented as:

[0133] Each head is:

[0134]

[0135] Where Q, K, and V are obtained by linear transformation of the input feature Z;

[0136] After N layers of Transformer encoding, the [CLS] flag bit is taken as the fused feature:

[0137]

[0138] The final prediction is:

[0139]

[0140] Specifically, in step S3, each model can be used to construct its own prediction model. For historical photovoltaic power generation data, the periodicity and trend characteristics of the time series data can be extracted based on the Prophet model and LSTM (Long Short-Term Memory model). For image data such as clouds, spatial features such as edges, textures, and coverage can be extracted using a CNN convolutional neural network. For meteorological data such as temperature, humidity, wind, and weather, an LSTM+Attention model can be used to construct a time series meteorological-power generation correlation feature extraction model from meteorological data (X: temperature, humidity, wind, etc.) to power generation data (Y). The obtained data is then used for iterative training to obtain the feature extraction values ​​of each modality. In step S4, the feature extraction values ​​obtained from each modality are fused using a Transformer and input into the ensemble prediction model to predict the power generation of each photovoltaic-storage-charging cluster.

[0141] S43. Finally, the fused features are input into the integrated prediction model to predict the power generation of each photovoltaic-storage-charging cluster.

[0142] This application combines temporal models (LSTM), spatial models (CNN), and dynamic association models (Attention), and achieves deep fusion of heterogeneous data through dynamic weighting using Transformer, thus overcoming the shortcomings of existing technologies that rely on single data sources and simple data fusion.

[0143] like Figure 5 As shown in Figure S44, the photovoltaic power prediction trajectory is obtained based on the time and the corresponding photovoltaic cluster point.

[0144] Compared with existing technologies, the multimodal photovoltaic power generation prediction system based on logistics drone swarms in this invention has the following outstanding effects and advantages:

[0145] 1. Significantly improves photovoltaic power generation forecast accuracy: Through multimodal data complementarity enhancement, it effectively compensates for the spatiotemporal blind spots of traditional single meteorological station data by fusing local cloud images (spatial features) collected by UAVs, high-resolution meteorological sensor data (temperature, humidity, wind speed), and historical power generation time series data (temporal features). Specifically, it employs CNN (Convolutional Neural Network) for dynamic feature extraction of cloud images (such as cloud movement speed and coverage) and LSTM for temporal modeling of historical power generation data, combined with regression analysis of the physical laws of meteorological data, to form a complementary prediction mechanism.

[0146] 2. Reduce the cost of drone applications and improve operational efficiency: This invention uses the same drone swarm to perform meteorological detection (cloud data collection) in a time-sharing manner, collecting weather data for each photovoltaic power station within a short, medium and long range, thus expanding the scope of data collection to include more photovoltaic power stations.

[0147] 3. Economic advantages: Hardware investment costs are reduced through the reuse of drone functions and optimization of cluster scheduling.

[0148] 4. Environmental benefits: Accurate forecasting reduces carbon emissions from standby thermal power units.

[0149] 5. Technical versatility: The proposed multimodal fusion framework can be extended to other new energy fields such as wind power and hydrogen energy, requiring only adjustments to the data input type and model parameters.

[0150] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0151] The above embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of this disclosure, and these modifications and improvements all fall within the protection scope of this disclosure.

Claims

1. A multimodal photovoltaic power generation prediction system based on a logistics drone swarm, characterized in that, Includes the following steps: Step S1: Acquire multimodal data; Step S2: Process and evaluate the multimodal data, assessing the accuracy, completeness, and consistency of each modality. Step S3: Build a prediction model for each model and use the obtained data for iterative training to obtain the feature extraction values ​​for each modality; Step S4: The feature extraction values ​​obtained from each mode are fused using Transformer, and then input into the integrated prediction model to predict the power generation of each photovoltaic-storage-charging cluster.

2. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 1, characterized in that, The multimodal data obtained in step S1 mainly includes: historical photovoltaic power generation data; cloud images and environmental images at short, medium and long distances above each photovoltaic-storage-charging cluster point obtained by using a drone swarm; local weather forecast information obtained by web crawling; and real-time temperature and humidity information of each location obtained by temperature and humidity sensors.

3. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 1, characterized in that, The specific data evaluation operations in step S2 include: Step S21: Evaluate the authenticity of the data by comparing the corresponding values ​​in the multimodal data with the actual measured values ​​and calculating the error rate; Step S22: Assess the completeness of the data, i.e., check whether the data contains all the necessary information, whether any values ​​are missing or important content is omitted. Step S23: Evaluate the consistency of the data, that is, the logical consistency of multimodal data between different modalities and within the same modality.

4. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 1, characterized in that, The specific operations in step S3 include: Step S31: Perform feature engineering on historical power generation data and photovoltaic-storage-charging cluster tag data, extract features and input them into the first time series model (Prophet+LSTM) to obtain the periodicity and trend of the time series data, and extract the first time series dynamic feature value; Step S32: Input the cloud image data and surrounding environment image data obtained by the drone into the second temporal model (CNN) to obtain the image edge, texture and coverage data, and extract the second spatial feature extraction value. Step S33: Input the temperature, humidity, wind force and weather data into the third time series model (LSTM+Attention) to obtain the third feature extraction value of the time series meteorological-power generation correlation.

5. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 4, characterized in that, The specific operations in step S31 include: taking historical power generation probabilities as input, labeling time-related features, inputting them into the Prophet model, and decomposing the power generation data into a trend term g(t), a seasonal term s(t), and a holiday term h(t): Then the Prophet residual The original power value is input into an LSTM for enhancement, and the corrected power timing characteristics are output, thereby compensating for the LSTM's neglect of macroscopic periodicity.

6. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 4, characterized in that, The specific operations in step S32 include: taking the RGB cloud image captured by the drone as input, inputting a lightweight ResNet18 (pre-trained weight initialization), and modifying the last layer of the model by changing global average pooling to a fully connected layer, so that it can better extract the corresponding feature vectors and provide micro-environment information that weather stations cannot cover.

7. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 4, characterized in that, The specific operations in step S33 include: taking meteorological time series data as input, using an Encoder-Decoder structure, extracting meteorological time series features through LSTM, then dynamically focusing on key time steps through an Attention mechanism, and finally outputting the correlation feature vector between meteorology and power generation; setting a two-level attention mechanism in the attention mechanism, namely a time attention mechanism and a variable attention mechanism.

8. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to claim 1, characterized in that, The specific operations in step S4 include: Step S41: Embed the feature extraction values ​​of different modalities as input into the Transformer model, and use the self-attention mechanism of the Transformer to dynamically capture the correlation and importance weights between each modality; Step S42: Extract features from multiple subspaces through a multi-head attention mechanism, and further fuse these features through a feedforward network to generate a comprehensive feature representation; Step S43: Input the fused features into the integrated prediction model to predict the power generation of each photovoltaic-storage-charging cluster. Step S44: Obtain the photovoltaic power prediction trajectory based on the time and the corresponding photovoltaic cluster point.

9. The multimodal photovoltaic power generation prediction system based on a logistics drone swarm according to claim 8, characterized in that, The specific operations in steps S41 and S42 include: The feature vectors extracted from the three modalities are as follows: Meteorological characteristics: Image features: Historical power generation characteristics: When inputting various features into the Transformer, the dimensions need to be unified through linear projection: ,in The weight matrix is ​​a learnable weight matrix; The multi-head attention mechanism in the Transformer coding layer can be represented as: Each head is: Where Q, K, and V are obtained by linear transformation of the input feature Z; After N layers of Transformer encoding, the [CLS] flag bit is taken as the fused feature: The final prediction is: That is, the power generation capacity of each photovoltaic-storage-charging cluster.

10. The multimodal photovoltaic power generation prediction system based on logistics drone swarms according to any one of claims 1-9, characterized in that, Step S1 includes collecting data through a drone swarm, which includes a logistics drone swarm and a photovoltaic power station swarm.

Images