Rainfall prediction system and method based on cloud system parameter monitoring

Abstract

The invention belongs to the technical field of meteorological monitoring and prediction, and discloses a rainfall prediction system and method based on cloud system parameter monitoring. According to the system, cloud system and ground meteorological parameters are acquired through multi-source heterogeneous sensors such as high-resolution satellite remote sensing, ground radar, a low-altitude unmanned aerial vehicle and a marine meteorological buoy, data are transmitted through 5G, satellite communication and various wireless technologies, and preprocessing, feature extraction and model prediction are carried out through a data processing center by applying deep learning and other algorithms. The method comprises the steps of data acquisition, transmission, preprocessing, feature extraction, model prediction and result output. The rainfall prediction method can comprehensively and accurately collect data, efficiently process and analyze the data and output a visual rainfall prediction result, solves the problems that a traditional method is insufficient in data and low in precision, improves the accuracy and reliability of rainfall prediction, and has important application value in the fields of meteorological disaster early warning, agricultural production, urban planning and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological observation based on remote sensing satellites, and specifically to a precipitation prediction system and method based on cloud system parameter monitoring. Background Art

[0002] Precipitation, as one of the core elements of climate change, the study of its variation law plays a decisive role in the rational allocation of water resources, the scientific planning of agricultural production, and the effective implementation of disaster warning. In the process of studying precipitation, in the prior art, a precipitation prediction method based on deep learning with the publication number of CN114879281A obtains a precipitation prediction training data set through the official website of the China Statistical Yearbook. For.jpg format data, it uses the intelligent character recognition OCR technology of the Baidu Smart Cloud system and the split function to extract text data information, and uses the pandas library and the head function to load and check the data, and further completes data format conversion, normalization, and tensorization operations. On this basis, a DNN model based on TensorFlow is built, the shuffle and batch values are carefully set, and a suitable optimizer and loss function are assembled to predict feature data. The performance of the model is evaluated by the RMSD value and optimized, and finally the model is saved for subsequent precipitation prediction. Processing.jpg format data requires specific technical means; and the data quality is unstable, with noise and missing values;

[0003] In the prior art, the NRIET quantitative precipitation estimation method based on cloud classification and machine learning, disclosed in CN110346844A, first preprocesses radar data and rain gauge data, matches radar reflectivity and rain gauge precipitation data based on stations, identifies different cloud systems such as stratiform clouds and convective clouds according to the radar reflectivity intensity, and uses machine learning regression algorithms to perform real-time fitting training to obtain a relationship model between cumulative precipitation and radar combined reflectivity. Finally, this model is applied to radar combined reflectivity grid field data to achieve quasi-real-time quantitative precipitation estimation. This method lacks a unified and effective parameter selection standard, making the model construction process complex and time-consuming. Summary of the Invention

[0004] The purpose of the present invention is to provide a precipitation prediction system and method based on cloud system parameter monitoring to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A precipitation prediction system based on cloud system parameter monitoring, including: a data acquisition module, a data transmission module, a data processing center, and a storage module;

[0006] The data acquisition module uses multi-source heterogeneous sensors, innovatively integrating high-resolution satellite remote sensing, ground radar, and low-altitude UAV monitoring technologies to comprehensively collect cloud system parameter data, including cloud top height, cloud bottom height, cloud thickness, cloud movement speed, cloud water content, cloud temperature, cloud reflectivity, and cloud texture features. At the same time, a high-precision ground meteorological monitoring station is used to obtain ground meteorological parameter data such as air temperature, air pressure, humidity, wind speed, and wind direction to ensure the comprehensiveness and accuracy of the data. The data transmission module uses a high-speed transmission technology that combines 5G and satellite communication to transmit the collected cloud system parameter data and ground meteorological parameter data to the data processing center in a manner of ultra-low latency and high reliability to meet the real-time data processing requirements.

[0007] The data processing center includes:

[0008] Preprocessing unit: Based on the adaptive noise recognition algorithm of deep learning, the collected data is deeply cleaned to accurately remove noise and outliers in the data. A new missing value filling algorithm based on spatio-temporal correlation is used to effectively supplement the missing data by combining the data correlation of adjacent times and spatial positions. A standardization method based on the data distribution characteristics is used to make different types of data have a unified scale and improve the data availability.

[0009] Feature extraction unit: With the innovative algorithm that combines the generative adversarial network GAN and the convolutional neural network CNN, features that have an important impact on precipitation prediction are extracted from the preprocessed data. Especially, the morphological and structural features of clouds are deeply mined from the cloud system image data to enhance the representativeness of the features.

[0010] Model prediction unit: Using a multi-modal deep learning neural network model that integrates the attention mechanism, the extracted feature data is used as input for precipitation prediction. This model integrates multi-modal data such as cloud system parameters, ground meteorological parameters, and geographical information, and is trained with a large amount of historical cloud system parameter data and corresponding precipitation data to significantly improve the prediction accuracy.

[0011] Result output unit: Outputs the precipitation prediction results, including the predicted precipitation intensity, precipitation time range, and precipitation area information, and uses augmented reality AR and virtual reality VR technologies to generate a visual report to intuitively and immersively display the precipitation prediction situation.

[0012] The storage module uses distributed blockchain storage technology to securely and reliably store the collected data, preprocessed data, feature data, trained models, and prediction result information to ensure the immutability and high availability of the data.

[0013] 3. Further, the high-resolution satellite remote sensing sensor in the data acquisition module has multi-spectral imaging capabilities, can obtain cloud system information at different wavelengths, and through the analysis of multi-spectral data, can further accurately identify the material composition and microscopic structural characteristics of the cloud system, providing a richer data dimension for precipitation prediction; at the same time, the low-altitude unmanned aerial vehicle is equipped with a high-definition camera with adjustable focal length and a micro meteorological sensor, which can automatically adjust the flight path and monitoring parameters according to the height and shape of the cloud system, realizing refined data acquisition of specific cloud system areas.

[0014] 4. Further, on the basis of the integration of 5G and satellite communication, the data transmission module adopts data encryption and compression technologies; through advanced encryption algorithms, it ensures the security of data during transmission, preventing data from being stolen or tampered with; using efficient data compression algorithms, it compresses a large amount of collected data, reduces the transmission bandwidth requirements, improves the transmission efficiency, and can ensure the stable and fast transmission of data even in complex communication environments.

[0015] 5. Further, in the preprocessing unit of the data processing center, the adaptive noise recognition algorithm based on deep learning adopts transfer learning technology, which can quickly adapt to the data noise characteristics under different monitoring devices and environments, and can achieve efficient noise removal of new data without a large number of samples for retraining; at the same time, the new missing value filling algorithm based on spatio-temporal correlation combines Bayesian inference methods, which can more accurately estimate missing values according to the uncertainty of existing data, improving the integrity and quality of data.

[0016] Further, the innovative algorithm that combines the generative adversarial network GAN and the convolutional neural network CNN in the feature extraction unit introduces an attention mechanism, which can automatically focus on the regions and features that are most critical for precipitation prediction in cloud system images, such as the edges of clouds and the convective structures inside, further enhancing the accuracy and pertinence of feature extraction; moreover, this algorithm can adaptively adjust the feature extraction strategy according to different precipitation types (heavy rain, light rain, snowfall), improving the effectiveness of features.

[0017] Further, the multi-modal deep learning neural network model that integrates the attention mechanism in the model prediction unit adopts incremental learning technology, which can automatically update the model parameters during the continuous accumulation of new historical cloud system parameter data and precipitation data, continuously improving the prediction accuracy; at the same time, this model also combines geographic information system GIS technology, which can incorporate topographic and geomorphic geographical factors into the precipitation prediction model, considering the influence of terrain on cloud movement and precipitation formation, improving the accuracy of precipitation prediction in different geographical regions.

[0018] A precipitation prediction method based on cloud system parameter monitoring includes the following steps:

[0019] Step 1. Data collection: Using the multi-source heterogeneous sensor fusion technology, high-resolution satellite remote sensing, ground radar, and low-altitude drones are utilized to comprehensively collect cloud system parameter data, including cloud top height, cloud bottom height, cloud thickness, cloud movement speed, cloud water content, cloud temperature, cloud reflectivity, and cloud texture features. At the same time, through high-precision ground meteorological monitoring stations, ground meteorological parameter data such as air temperature, air pressure, humidity, wind speed, and wind direction are collected.

[0020] Step 2. Data transmission step: Through the high-speed transmission link integrating 5G and satellite communication, the collected cloud system parameter data and ground meteorological parameter data are transmitted to the data processing center in a ultra-low latency and highly reliable manner.

[0021] Step 3. Data preprocessing step: In the data processing center, data cleaning is performed based on the adaptive noise recognition algorithm of deep learning to remove noise and outliers in the data. A new missing value filling algorithm based on spatio-temporal correlation is used to supplement the missing data. The standardization method based on the data distribution characteristics is applied to make different types of data have a unified scale.

[0022] Step 4. Feature extraction step: With the innovative algorithm combining the generative adversarial network GAN and the convolutional neural network CNN, features that have an important impact on precipitation prediction are extracted from the preprocessed data, especially deeply mining the morphological features and structural features of clouds from the cloud system image data.

[0023] Step 5. Model prediction step: Using the multi-modal deep learning neural network model integrating the attention mechanism, the extracted feature data is used as input for precipitation prediction. This model integrates multi-modal data of cloud system parameters, ground meteorological parameters, and geographical information, and is trained with a large amount of historical cloud system parameter data and corresponding precipitation data.

[0024] Step 6. Result output step: The precipitation prediction results are output, including the predicted precipitation intensity, precipitation time range, and precipitation area information, and a visualization report is generated using augmented reality AR and virtual reality VR technologies to intuitively and immersively display the precipitation prediction situation.

[0025] Furthermore, in Step 1, for different application scenarios, urban meteorological disaster warning, agricultural irrigation water scheduling, and flood control of water conservancy projects, personalized data collection strategies are formulated. In the urban meteorological disaster warning scenario, the monitoring density of cloud systems and ground meteorological parameters in the surrounding areas of the city is increased. In the agricultural irrigation water scheduling scenario, according to the distribution area and growth cycle of crops, cloud systems and meteorological data over farmland are collected in a targeted manner to meet the special requirements for precipitation prediction in different scenarios.

[0026] Further, data fusion is added after step three to deeply fuse data from different sources and of different types, such as satellite remote sensing data, ground radar data, and UAV monitoring data. By making full use of the advantages of each data source, the inconsistency and redundancy between data are eliminated, forming a more comprehensive and accurate comprehensive dataset, providing a better data foundation for subsequent feature extraction and model prediction.

[0027] Further, in step six, the visualization report not only adopts augmented reality (AR) and virtual reality (VR) technologies, but also combines artificial intelligence voice interaction technology. Users can obtain detailed precipitation prediction information for a specific area and specific time through voice commands, realizing natural interaction with the visualization report. At the same time, the visualization report can display the precipitation prediction results in different forms, such as dynamic charts and 3D models, facilitating users to intuitively understand and analyze the precipitation prediction information.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] The present invention adopts a method based on cloud system parameter monitoring, bringing significant effects in many aspects. In data acquisition, the multi-source heterogeneous sensor fusion technology is used, combined with high-resolution satellite remote sensing, ground radar, low-altitude UAVs, and high-precision ground meteorological monitoring stations to comprehensively collect cloud system and ground meteorological parameters, covering a variety of key feature data, providing a rich and accurate data foundation for precipitation prediction, effectively solving the problems of insufficient coverage and low accuracy in traditional data acquisition. The 5G and satellite communication fusion technology is adopted, combined with data encryption and compression technologies. The high speed of 5G and the wide coverage of satellite communication ensure that data can be transmitted to the data processing center with ultra-low latency and high reliability, meeting the real-time requirements; the encryption technology ensures data security, preventing theft and tampering. The feature extraction unit uses an innovative algorithm that combines a generative adversarial network (GAN) and a convolutional neural network (CNN) and introduces an attention mechanism, which can accurately extract features crucial for precipitation prediction from the preprocessed data, especially deeply mining the morphological and structural features from cloud system images, enhancing the representativeness of features, improving the accuracy and pertinence of feature extraction, and better capturing information related to precipitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the system flow of the present invention;

[0031] Figure 2 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] See also Figure 1 —2. The present invention provides a technical solution: an implementation method of a precipitation prediction system and method based on cloud parameter monitoring:

[0034] When building a precipitation prediction system based on cloud parameter monitoring, the data acquisition module is the foundation of the entire system, and its accuracy and comprehensiveness directly determine the accuracy of subsequent predictions. The selection and deployment of high-resolution satellite remote sensing sensors is a complex and critical task. The multispectral imaging sensor carried by the satellite has excellent performance and can clearly capture the outline and texture of the cloud system in the visible light band, providing an intuitive basis for cloud morphology analysis; in the near-infrared band, the water vapor content in the cloud system can be effectively inverted by monitoring the reflectivity of the cloud system, which is crucial for judging the possibility and intensity of precipitation; the thermal infrared band can accurately detect the temperature distribution of the cloud system and reveal the energy changes and convection activities inside the cloud system. In order to ensure high-frequency coverage of the target area, the satellite orbit design comprehensively considers factors such as the rotation and revolution of the earth and the geographical location of the target area. By optimizing parameters such as orbital inclination, altitude and period, the satellite passes over the target area many times a day, thereby obtaining the latest and continuous cloud data, providing strong support for real-time monitoring of cloud evolution.

[0035] The deployment of ground radars needs to take into full consideration the diversity of geographical environments and monitoring needs. In flat and open plains, the S-band Doppler weather radar, with its detection range of up to 460 kilometers, is an ideal choice for monitoring the movement speed and radial velocity of cloud systems. The radar uses the Doppler effect to accurately measure the movement speed of particles in the cloud system by transmitting and receiving electromagnetic waves, and then analyzes the movement direction and intensity changes of the cloud system. In areas with complex terrain such as mountainous areas, the addition of X-band radars becomes a necessary measure because the terrain undulations have a great impact on the propagation of radar waves. Although the detection range of X-band radars is relatively short, their wavelength is shorter and they are more adaptable to terrain. They can effectively capture subtle changes in mountainous cloud systems caused by terrain blocking and lifting, such as the formation of terrain clouds, the gathering and dissipation of clouds in valleys, etc. Different types of radars are interconnected through high-speed data links, and distributed data management and collaborative processing technologies are used to achieve real-time sharing and fusion analysis of data, thereby improving the monitoring capabilities of complex cloud systems.

[0036] The application of low-altitude drones has opened up a new dimension for cloud system monitoring and provided refined data support. Drones with long endurance and high maneuverability are selected. The adjustable focal length high-definition cameras they carry are built-in with advanced image recognition and autofocus algorithms. During flight, the drones receive satellite and radar monitoring data through real-time communication with the ground control station, and automatically adjust the focal length according to the cloud system height to ensure clear cloud images are obtained. At the same time, the equipped micro meteorological sensors integrate high-precision temperature, humidity, air pressure, and wind speed sensors, which can collect meteorological parameters inside the cloud system in real time, providing first-hand data for studying the microphysical processes of cloud systems. The flight path of the drone is pre-planned by the ground control station based on satellite and radar monitoring data. Using Geographic Information System (GIS) technology, considering factors such as terrain and airspace restrictions, a safe and efficient flight route is formulated. When abnormal changes are found in a specific cloud system area, such as enhanced convection and sudden changes in cloud top height, the ground control station can dynamically adjust the drone flight path through the real-time communication link to achieve refined monitoring of key areas.

[0037] The construction of the data transmission module focuses on the 5G and satellite communication integration technology, aiming to achieve high-speed, stable, and secure data transmission. At the ground monitoring sites, 5G communication base stations are fully deployed. Utilizing the high bandwidth and low latency characteristics of the 5G network, it ensures that data can be quickly transmitted to the edge nodes of the data processing center. At the same time, each monitoring device is equipped with a satellite communication terminal as a backup transmission link to cope with emergencies such as insufficient 5G signal coverage or communication interruption. In terms of data encryption, advanced [specific encryption algorithm] is adopted, such as the encryption technology based on quantum key distribution. Using the principles of quantum mechanics to ensure the absolute security of the key, the transmitted data is encrypted to prevent data from being stolen or tampered with during transmission. In the data compression link, [specific compression algorithm] is used, such as the image compression algorithm based on deep learning and the lossless compression algorithm for the characteristics of meteorological data. For large-volume image data such as satellite remote sensing images, the deep learning compression algorithm is adopted to minimize the data volume to the greatest extent while ensuring that the key information of the image is not lost; for structured data such as ground meteorological parameters, the lossless compression algorithm is adopted to ensure data integrity while achieving efficient compression, reducing data transmission costs and bandwidth requirements.

[0038] As the core hub of the entire system, the data processing center undertakes the key tasks of data processing and analysis. In the preprocessing unit, a model is built based on the deep learning-based adaptive noise recognition algorithm, such as TensorFlow or PyTorch. By collecting a large amount of historical data containing various types of noise, including sensor noise, electromagnetic interference noise, and data transmission noise, etc., the model is trained in a supervised manner. The model learns the characteristic patterns of different noises during the training process, so as to accurately identify and remove the noises in the data. In terms of missing value imputation, a new missing value imputation algorithm based on spatio-temporal correlation is used, combining time series analysis and spatial interpolation methods. For time series data, time series analysis methods such as the autoregressive integrated moving average model ARIMA are adopted to predict the missing values according to the changing trends of historical data; for spatially distributed data, spatial interpolation methods such as Kriging interpolation are used, combining the data of adjacent monitoring points and the spatial position relationship to effectively supplement the missing data. Data standardization processing adopts, such as Z-Score standardization, Min-Max standardization, etc., according to the distribution characteristics of different data types, so that different types of data have a unified scale, eliminate the differences in data dimensions and orders of magnitude, and facilitate subsequent feature extraction and model training.

[0039] In the data processing center, efficient and stable data transmission is the key bridge to ensure the coordinated operation of each unit. During the process of data transmission from the acquisition end to the processing center, a variety of advanced technologies are adopted to ensure the integrity and timeliness of the data.

[0040] When the data is transmitted to the edge node of the data processing center through the 5G and satellite communication fusion link, it will face the problems of large-scale data aggregation and distribution. At this time, software-defined network (SDN) technology is adopted to intelligently manage the data traffic. The SDN controller can monitor the network traffic status in real time and dynamically adjust the data transmission path according to the priorities and real-time requirements of different data. For example, for the real-time cloud system monitoring data with extremely high timeliness requirements, the SDN controller will preferentially allocate high-bandwidth links to ensure the rapid transmission of data to the preprocessing unit; while for the update of historical data or the transmission of backup data, it can be arranged during the period with low network load, effectively improving the utilization rate of network resources.

[0041] The SDN controller plays a core role in traffic scheduling in the system. It establishes connections with network devices such as switches or routers using standard protocols like OpenFlow to collect network traffic information in real time. For example, within each second, the SDN controller can obtain detailed information such as the bandwidth utilization rate of each link, the packet transmission delay, and the source and destination addresses of the traffic. When a large amount of cloud system monitoring data floods into the data processing center from the acquisition end, the SDN controller performs traffic scheduling according to pre-set policies. For real-time cloud system monitoring data marked as high priority, such as the radar reflectivity data of a developing severe convective cloud system, the controller will look for available high-bandwidth links in its internal flow table. If it finds that a 10Gbps fiber optic link currently has a low utilization rate and meets the data transmission delay requirements, it will direct the traffic of this data to this link by issuing flow table entries, ensuring that the data can be transmitted to the preprocessing unit at the fastest speed, and its transmission delay can be controlled within 10 milliseconds.

[0042] For the update of historical data or the transmission of backup data, the SDN controller will handle it according to the time scheduling strategy. During the early morning hours when the network load is low, such as from 2 am to 4 am, the controller will centrally arrange the transmission tasks for this data. It will reasonably allocate the storage locations of the data according to the load conditions and storage space of each storage node, and select relatively idle network links for transmission. For example, for some historical cloud system image data that needs to be stored for a long time, the controller will transmit it to a storage node with a large storage capacity and a stable network connection. At the same time, using link aggregation technology, multiple 1Gbps links are bundled into a logical link to increase the transmission bandwidth and ensure the integrity and transmission efficiency of the data.

[0043] To further enhance the security and stability of data transmission, in the data interaction link within the data processing center, the virtual private network (VPN) technology is used. The data transmission between different functional units is carried out through encrypted VPN channels to prevent data from being stolen or tampered with during the internal network transmission process. For example, the data transmitted from the feature extraction unit to the model prediction unit will be encrypted by VPN and can only be decrypted and read after passing through specific key verification at the target unit, ensuring the security of the data flow within the processing center.

[0044] The virtual private network (VPN) technology: When data is transmitted between different functional units within the data processing center, such as from the feature extraction unit to the model prediction unit, the VPN technology based on the IPsec (Internet Protocol Security) protocol is adopted. On the data sending side, such as the server of the feature extraction unit, the IPsec VPN client software is installed. When there is data to be transmitted, the client software encrypts the data according to the pre-configured security policy. It first performs an integrity check on the data by calculating the hash value of the data (such as using the SHA-256 algorithm) to ensure that the data has not been tampered with during transmission. Then, the data is encrypted using an encryption algorithm (such as the AES-256 encryption algorithm), converting the original data into ciphertext form. At the same time, VPN-related identification information is added to the data packet header, such as the VPN internal address conversion information of the source address and destination address, the security protocol version number, etc.; On the data receiving side, that is, on the server of the model prediction unit, the corresponding IPsec VPN server software is installed. When the encrypted data is received, the server software first verifies the VPN identification information in the data packet header to confirm the legitimacy of the data source. Then, according to the pre-shared key information, the data is decrypted to restore the original data. Throughout the process, the VPN technology ensures the confidentiality and integrity of the data during the transmission within the internal network of the data processing center by establishing a secure tunnel, preventing the data from being illegally accessed or maliciously tampered with internally.

[0045] In terms of ensuring the reliability of data transmission, a combination of distributed storage and redundant transmission technologies is adopted. When the data arrives at the data processing center, it is distributed and stored on multiple storage nodes, and each node stores different parts or copies of the data. At the same time, during the data transmission process, key data is transmitted redundantly over multiple links. For example, when transmitting the preprocessed data to the feature extraction unit, in addition to the main transmission link, the data is also transmitted simultaneously through the backup link. Once the main link fails, the backup link can immediately take over the data transmission task to ensure the continuity of the data processing process and avoid system failures caused by data transmission interruptions.

[0046] Distributed Storage and Redundant Transmission Technology: When the data arrives at the data processing center, a distributed storage system such as Ceph is used for storage. The Ceph storage system divides the data into multiple objects of a fixed size (such as objects of 4MB size), and then evenly distributes these objects to multiple storage nodes through the consistent hashing algorithm. Each storage node will regularly send heartbeat messages to other nodes in the cluster to report its own storage status and health condition. When a storage node fails, such as a hard disk failure or a network connection interruption, other nodes will detect the abnormality of the node within a short time (usually within 1 minute) and automatically start the data recovery process; during the data transmission process, for the redundant transmission of critical data, taking the transmission of preprocessed data to the feature extraction unit as an example. In addition to the main transmission link, a backup link is also used for simultaneous transmission. The main link may use a high-speed Ethernet link, such as a 10Gbps fiber optic link, while the backup link can be a low-latency wireless link, such as a 5G network link. At the data sending end, data will be sent to both the main link and the backup link simultaneously, and information such as sequence numbers and timestamps will be added to the data packet header. At the receiving end, the server of the feature extraction unit will listen to the data reception conditions of both the main link and the backup link simultaneously. If the main link fails, such as packet loss or transmission interruption, the receiving end will immediately switch to the backup link to continue receiving data, and reorder and verify the data according to the sequence numbers and timestamps in the data packet header to ensure the integrity and continuity of the data and avoid system failures caused by data transmission interruptions.

[0047] In the data processing center, the feature extraction unit is a key link for mining the deep value of data. The innovative algorithm that combines the Generative Adversarial Network (GAN) and the Convolutional Neural Network (CNN) plays a core role here. When building a model with the TensorFlow deep learning framework, the network structures of the generator and the discriminator are carefully designed. The generator adopts multiple transposed convolutional layers. By learning the distribution law of a large amount of cloud system image data, it generates realistic cloud system samples from random noise, expands the data set, and enhances the generalization ability of the model. For example, when there are few samples of a certain rare cloud system in the training data, the generator can generate more similar samples, enabling the model to learn more comprehensive cloud system features. The discriminator consists of a series of convolutional layers and fully connected layers, and its task is to accurately distinguish between real data and the fake data generated by the generator. During the training process, the generator and the discriminator engage in a fierce confrontation. The generator continuously optimizes its generation strategy to generate more realistic samples, while the discriminator continuously improves its discrimination ability. Through this dynamic game, the model parameters are continuously adjusted to make the feature extraction more accurate.

[0048] In the CNN part, a network architecture consisting of multiple convolutional layers and pooling layers is constructed. In the first few convolutional layers, smaller convolutional kernels such as 3×3 or 5×5 are used to focus on extracting low-level features of cloud systems, such as edges, textures, and local shapes. These basic features provide support for subsequent high-level feature extraction. In the later convolutional layers, larger convolutional kernels are used, and techniques such as dilated convolution are combined to expand the receptive field and extract high-level semantic features of cloud systems, such as the overall morphology, internal structure of cloud systems, and the relationships between different cloud systems. The attention mechanism weights the features in different regions by introducing an attention weight matrix. For example, when analyzing typhoon cloud systems, the model will automatically focus on the eyewall region and spiral rainbands of the typhoon because these regions are closely related to heavy precipitation and strong wind weather. By assigning higher weights to these key regions, features that are important for precipitation prediction can be highlighted, improving the accuracy and pertinence of feature extraction.

[0049] The model prediction unit uses a multi-modal deep learning neural network model integrated with an attention mechanism to achieve accurate precipitation prediction. This model integrates multi-modal data such as cloud system parameters, surface meteorological parameters, and geographical information. Geographical information is obtained through the Geographic Information System (GIS). The terrain and landform data are digitized, converted into grid form or vector data, and mapped to the same spatial coordinate system as other data so that the model can comprehensively analyze. In the model training stage, a large amount of historical cloud system parameter data and corresponding precipitation data are used. These data are divided according to the ratio of 70% training set, 15% validation set, and 15% test set. The Adam optimization algorithm is used to train the model. This algorithm combines the advantages of Adagrad and Adadelta algorithms, can adaptively adjust the learning rate, converge quickly in the initial stage of training, and finely adjust the model parameters in the later stage. At the same time, incremental learning technology is used. Whenever new historical data accumulates, the model can automatically update the parameters. For example, when new annual precipitation data is added to the training set, the model can learn new precipitation patterns and rules based on the existing knowledge, further improving the prediction accuracy. In addition, to prevent model overfitting, the Dropout technique is adopted, randomly discarding some neuron connections during training to reduce the co-adaptation between neurons and enhance the generalization ability of the model.

[0050] The result output unit is a crucial link for presenting the model prediction results to users. By leveraging augmented reality (AR) and virtual reality (VR) technologies, visual reports are generated to provide users with an immersive experience. Using Unity3D as the development engine and combining AR development tools such as ARCore and ARKit, AR applications are developed. Through mobile phones or AR glasses, users can intuitively see the results of precipitation prediction in real-world scenarios. For example, in an urban planning scenario, users can see the impact of future precipitation on the urban drainage system. The water accumulation depths in different areas are displayed as overlays of different colors on the city map, and the precipitation time range is presented in the form of a dynamic timeline. Users can view precipitation conditions at different time periods by touching the screen or issuing voice commands. In terms of VR, users wear VR headsets and enter a virtual meteorological scenario. By operating the controllers, they can view precipitation prediction information omnidirectionally and from multiple angles, such as overlooking the distribution of precipitation areas from high altitudes and observing the relationship between cloud system structures and precipitation by delving into the interior of clouds. Combining artificial intelligence voice interaction technology based on natural language processing frameworks such as Baidu UNIT and iFlytek, users can obtain detailed precipitation prediction information for specific regions and specific times through voice commands. For example, when a user says, "Query the precipitation intensity in Haidian District, Beijing tomorrow," the system can quickly respond and highlight the relevant information in the visual report, greatly improving the efficiency and convenience for users to obtain information. At the same time, to meet the needs of different users, the result output also supports multiple formats, such as PDF, CSV, etc., facilitating data archiving and secondary analysis for professional users.

[0051] In the implementation process of the precipitation prediction method based on cloud system parameter monitoring, data collection, as the starting point of the entire process, plays a crucial role. The deployment of multi-source heterogeneous sensors needs to comprehensively consider the climate characteristics and geographical environment differences in different regions.

[0052] In coastal areas, with a significant maritime climate, the formation and evolution of cloud systems are deeply influenced by factors such as ocean water vapor and sea surface temperature. Taking the southeastern coastal area of China as an example, every summer, the warm and humid air currents from the ocean meet the cold air from the continent, making it extremely easy to form precipitation. To more accurately capture the water vapor conditions for cloud formation, in addition to conventional high-resolution satellite remote sensing, ground radar, and low-altitude drone monitoring equipment, it is necessary to add ocean meteorological buoys. These buoys are equipped with high-precision temperature sensors that can accurately measure the sea surface temperature with an accuracy of up to ±0.1°C. By continuously monitoring the sea surface temperature, the heat and water vapor transfer from the ocean to the atmosphere can be effectively judged. At the same time, the humidity sensors on the buoys use capacitive or resistive principles to accurately measure the sea surface humidity, with the error controlled within ±2%RH, providing reliable data for analyzing water vapor conditions. The wind speed sensors use ultrasonic or propeller technologies to accurately measure wind speed and direction, providing key parameters for studying the energy exchange between the ocean and the atmosphere.

[0053] For the acquisition of satellite remote sensing data, the changes in seasonal and meteorological conditions are important bases for adjusting acquisition parameters. In summer, convective activities are active and severe convective weather such as heavy rain occurs frequently. Taking the southern region of China as an example, at this time, the observation frequency of the satellite for a specific area should be increased from the conventional 2 - 3 times a day to once every 3 - 4 hours to ensure that the rapid changes in cloud systems can be captured in a timely manner. In terms of sensor parameter settings, adjust the band combination and exposure time of the satellite multi - spectral imager. For example, for severe convective cloud systems, increase the sensitivity to the near - infrared and thermal - infrared bands to more clearly observe the water vapor distribution and energy changes inside the cloud systems. At the same time, by optimizing the satellite orbit control algorithm, ensure that when the satellite passes over the target area, high - quality data can be stably obtained, avoiding data loss or distortion caused by orbit deviation.

[0054] The calibration and maintenance of ground radars are the keys to ensuring data accuracy and stability. Taking the S - band Doppler weather radar as an example, a comprehensive calibration should be carried out at least once a month. During the calibration process, use a standard reflector to accurately measure and adjust parameters such as the radar's transmit power, receive sensitivity, and pulse width, ensuring that the measurement error of radar reflectivity is controlled within ±1 dBZ and the measurement error of radial velocity is controlled within ±0.5 m / s. At the same time, regularly check and maintain the hardware devices such as the radar antenna and feeder, and promptly replace aging or damaged components to ensure the normal operation of the radar.

[0055] Before low - altitude unmanned aerial vehicles (UAVs) perform tasks, the planning of flight routes needs to consider multiple factors comprehensively. Use numerical weather prediction models and real - time meteorological data to predict the airflow conditions in the flight area in advance. For example, by analyzing atmospheric wind field data, avoid severe convective areas and upper - air jet streams to ensure flight safety. At the same time, combined with terrain data, use geographic information system (GIS) technology to plan the optimal flight path. When flying in mountainous areas, adjust the flight height and speed of the UAV according to the height and slope of the mountain to avoid collisions with the mountain. In addition, use machine - learning algorithms to analyze historical flight data and meteorological conditions to establish a flight risk assessment model. Through learning a large amount of flight data, the model can identify potential flight risk factors under different meteorological conditions, such as strong winds and low visibility, and issue early warnings in advance, providing a scientific basis for optimizing flight routes.

[0056] The Geographic Information System (GIS) not only digitizes terrain and landform data and performs coordinate mapping, but also further analyzes the correlation between different geographical elements and precipitation. By analyzing the relationship between the distribution of water areas such as rivers and lakes and precipitation, it is found that areas near large water bodies have unique patterns in terms of water vapor evaporation and transportation. During the model training process, the characteristics of these geographical elements are quantified and added to the model as additional input features; for areas near water bodies, a water vapor influence factor is calculated based on factors such as water body area and distance from the water body, and it is input into the model together with cloud system parameters and meteorological data. The model learns the interaction relationships between these features during the training process, so as to more accurately predict precipitation. When predicting precipitation in an area near a lake, the model comprehensively considers the water vapor contribution of the lake and the movement and development of the cloud system, improving the accuracy of precipitation prediction, especially showing better performance in predicting the spatial distribution and intensity change of precipitation.

[0057] The data transmission step is the core link to ensure data timeliness. In terms of wired communication, fiber optic networks are used to connect the main data collection sites and data processing centers. Fiber optic networks have the characteristics of high bandwidth and low latency, and can meet the real-time transmission requirements of large amounts of data. For example, in data collection sites in urban areas, by laying gigabit fiber optic networks, satellite remote sensing data, ground radar data, etc. are transmitted to the data processing center at high speed, and the data transmission rate can reach more than 1000 Mbps, ensuring that the data can reach the processing center for analysis in a short time.

[0058] For some remote areas or temporary monitoring points, wireless communication plays an important role. In addition to 5G and satellite communication, low-power wide-area network technologies such as LoRa and NB-IoT are combined to achieve the transmission of low-rate and small-data-volume sensor data. In small meteorological stations in mountainous areas, due to the complex terrain and difficult laying of wired networks, LoRa technology is used to transmit the collected meteorological parameters to nearby gateways. LoRa technology has the characteristics of long distance and low power consumption, and the transmission distance can reach several kilometers, which can meet the data transmission requirements of mountain meteorological stations. After the gateway aggregates and preliminarily processes the received data, it is transmitted to the data processing center through 5G or satellite links. To further improve the reliability of data transmission, data redundancy transmission technology is adopted. For key data, such as key frames of satellite remote sensing images and real-time monitoring data of ground radars, multi-link backup transmission is carried out. For example, data is transmitted simultaneously through 5G and satellite communication links to ensure that when a certain link fails, the data can be successfully transmitted through other links without affecting subsequent data processing and analysis.

[0059] In the data preprocessing step, different processing strategies are adopted for different types of data. For satellite remote sensing image data, in addition to removing noise and filling in missing values, geometric correction and radiometric correction are crucial. Geometric correction corrects the geometric distortion of the image caused by factors such as the Earth's curvature and satellite attitude through ground control points and satellite orbit parameters. The precise coordinates of the ground control points are obtained using a high-precision global positioning system (GPS) and ground control point measurement equipment. By establishing a geometric correction model, such as a polynomial model or a rational function model, the satellite remote sensing image is geometrically corrected to achieve sub-pixel level positioning accuracy. Radiometric correction corrects the radiance of the image according to the characteristics of the satellite sensor and the atmospheric transmission model. Using atmospheric radiation transfer software, such as MODTRAN, combined with the calibration parameters of the satellite sensor and real-time atmospheric parameters, the image is radiometrically corrected to ensure the comparability of data obtained at different times and by different satellites.

[0060] For ground radar data, quality control is carried out on data such as reflectivity and radial velocity to remove incorrect data caused by ground clutter, anomalous echoes, etc. Deep learning algorithms are used to automatically identify and classify radar echo images. Based on the convolutional neural network (CNN), a multi-scale feature extraction network is constructed to accurately identify ground clutter, precipitation echoes, anomalous echoes, etc. in the radar echo images. Through training with a large amount of labeled data, the recognition accuracy of the model reaches over 90%, quickly and accurately detecting anomalous echoes and improving the efficiency of data quality control.

[0061] In the feature extraction and model prediction steps, the methods are closely combined with the implementation in the system. When extracting features from the preprocessed data, an algorithm that combines the same generative adversarial network (GAN) and convolutional neural network (CNN) as in the system is used to ensure the consistency and accuracy of the features. The generative adversarial network can generate more realistic cloud system feature samples through the adversarial training of the generator and discriminator, expand the dataset, and enhance the generalization ability of the model. In the model prediction stage, according to different application scenarios and requirements, a suitable multi-modal deep learning neural network model is selected. For short-term precipitation prediction, lightweight models with higher computational efficiency, such as the long short-term memory network (LSTM) or gated recurrent unit (GRU) based on the recurrent neural network (RNN), can be selected. These models can quickly process time series data and give accurate prediction results in a short time. For long-term precipitation trend analysis, a more complex and more generalizable model, such as the Transformer model based on the attention mechanism, is used. It can fully mine the long-term dependencies and complex features in the data and improve the accuracy of long-term precipitation trend analysis.

[0062] The transfer learning technique is introduced. The model parameters trained in other regions or under similar meteorological conditions are used as the initial parameters for fine-tuning training in the target area. For example, a model for precipitation prediction in a monsoon climate region has been trained in a certain area. When precipitation prediction is needed in another area with similar monsoon climate characteristics, the parameters of this model are transferred to the new area, and a small amount of data from the new area is used for fine-tuning training. Through transfer learning, the training time and data requirements can be reduced, the adaptability of the model can be improved, and the model can quickly and accurately predict precipitation in the new area.

[0063] When the transfer learning technique is adopted, it is first necessary to determine the appropriate source model and target task. In the field of precipitation prediction, if a precipitation prediction model based on cloud system parameters and meteorological data has been established in region A with similar climate and certain commonalities in geographical environment. This source model has been trained with a large amount of historical data in region A and has good learning and fitting ability for the local precipitation pattern; when applying it to the target region B, the parameters of the source model will be initialized. During the initialization process, some key parameters will be adjusted according to the characteristics of the target region B. For example, if the terrain of region B is relatively complex, the relevant parameters in geographical information will be recalibrated. For the parameters representing the influence of terrain undulation on cloud movement, they will be adjusted according to the actual terrain data of region B (such as mountain height, slope and other information) to make it more in line with the actual situation of the target region.

[0064] In the fine-tuning training stage, a small amount of representative data from the target region B will be collected. For example, the cloud system parameters and corresponding precipitation data in different seasons and different weather types in the past year in the target region B are selected to form a small-scale training set. This training set is used to train the initialized model. During the training process, a relatively small learning rate, such as 0.001, is adopted to avoid over-adjusting the parameters of the source model and causing overfitting. At the same time, key attention is paid to the prediction performance indicators of the model in the target region B, such as root mean square error (RMSE) and coefficient of determination (R 2 ) etc. Through continuous iterative training, the model can better adapt to the precipitation pattern in the target region B; after 50 training cycles of fine-tuning, the RMSE index of the model in the target region B gradually decreases from the initial high value to an acceptable range, indicating that the prediction ability of the model in the target region B has been effectively improved through transfer learning, and it can quickly adapt to the precipitation prediction task in region B by using the knowledge learned in region A, reducing the large amount of data and computing resources required to retrain a completely new model in the target region.

[0065] The result output steps correspond to the result output unit of the system. Visual reports are customized and generated according to the needs of different user groups. For professional meteorological personnel, detailed data such as cloud system parameters and model prediction indicators are provided. For example, three-dimensional structure parameters of the cloud system, size distribution of cloud particles, and prediction error analysis of the model are provided to help professionals deeply analyze the precipitation process and model performance. For the general public, key information such as precipitation intensity, time, and area is presented in a concise and easy-to-understand manner. Through meteorological APPs or social media platforms, precipitation prediction results are presented in intuitive forms such as maps and charts. For example, different precipitation intensities are represented by areas of different colors, and the time range of precipitation is represented by a time axis, facilitating the public to understand weather changes and take preventive measures in advance.

Claims

1. A precipitation prediction system based on cloud parameter monitoring, characterized in that: include: Data acquisition module, data transmission module, data processing center and storage module; The data acquisition module adopts multi-source heterogeneous sensors, innovatively integrates high-resolution satellite remote sensing, ground radar and low-altitude drone monitoring technology, and comprehensively collects cloud parameter data, covering cloud top height, cloud bottom height, cloud thickness, cloud movement speed, cloud water content, cloud temperature, cloud reflectivity, and cloud texture characteristics; at the same time, high-precision ground meteorological monitoring stations are used to obtain ground meteorological parameter data of temperature, air pressure, humidity, wind speed, and wind direction to ensure the comprehensiveness and accuracy of the data; the data transmission module uses high-speed transmission technology that integrates 5G and satellite communications to transmit the collected cloud parameter data and ground meteorological parameter data to the data processing center in an ultra-low latency and high reliability manner to meet real-time data processing requirements; The data processing center includes: Preprocessing unit: Based on deep learning, the adaptive noise recognition algorithm deeply cleans the collected data and accurately removes noise and outliers in the data; A new missing value filling algorithm based on spatiotemporal correlation is used to effectively supplement missing data by combining the data correlation between adjacent moments and spatial locations. A standardization method based on data distribution characteristics is used to make different types of data have a unified scale and improve data availability. Feature extraction unit: With the help of an innovative algorithm combining generative adversarial networks (GAN) and convolutional neural networks (CNN), features that have an important impact on precipitation prediction are extracted from preprocessed data, especially the morphological and structural features of clouds are deeply mined from cloud image data to enhance the representativeness of features; Model prediction unit: A multimodal deep learning neural network model with an integrated attention mechanism is used to take the extracted feature data as input for precipitation prediction. The model integrates multimodal data of cloud parameters, ground meteorological parameters, and geographic information, and is trained with a large amount of historical cloud parameter data and corresponding precipitation data, significantly improving the prediction accuracy. Result output unit: outputs precipitation forecast results, including predicted precipitation intensity, precipitation time range, precipitation area information, and uses augmented reality AR and virtual reality VR technology to generate visual reports to intuitively and immersively display precipitation forecasts; The storage module adopts distributed blockchain storage technology to safely and reliably store the collected data, preprocessed data, feature data, trained models and prediction result information, ensuring that the data cannot be tampered with and has high availability.

2. A precipitation prediction method based on cloud parameter monitoring, applied to the precipitation prediction system based on cloud parameter monitoring of claim 1, characterized in that: The following steps are involved: Step 1: Data collection: Use multi-source heterogeneous sensor fusion technology, high-resolution satellite remote sensing, ground radar and low-altitude drones to comprehensively collect cloud parameter data, including cloud top height, cloud bottom height, cloud thickness, cloud movement speed, cloud water content, cloud temperature, cloud reflectivity, and cloud texture characteristics; at the same time, collect ground meteorological parameter data such as temperature, air pressure, humidity, wind speed and wind direction through high-precision ground meteorological monitoring stations; Step 2: Data transmission step: Through the high-speed transmission link integrated with 5G and satellite communications, the collected cloud parameter data and ground meteorological parameter data are transmitted to the data processing center in an ultra-low latency and high reliability manner; Step 3: Data preprocessing: In the data processing center, the deep learning-based adaptive noise recognition algorithm is used to clean the data to remove noise and outliers in the data; a new missing value filling algorithm based on spatiotemporal correlation is used to supplement the missing data; a standardization method based on data distribution characteristics is used to make different types of data have a unified scale; Step 4: Feature extraction: With the help of an innovative algorithm combining generative adversarial network (GAN) and convolutional neural network (CNN), features that have an important impact on precipitation prediction are extracted from the preprocessed data, especially the morphological and structural features of clouds are deeply mined from cloud image data; Step 5: Model prediction step: Utilize a multimodal deep learning neural network model with an integrated attention mechanism to take the extracted feature data as input for precipitation prediction; the model integrates multimodal data of cloud parameters, ground meteorological parameters, and geographic information, and is trained using a large amount of historical cloud parameter data and corresponding precipitation data; Step 6. Result output step: Output precipitation forecast results, including predicted precipitation intensity, precipitation time range, precipitation area information, and use augmented reality AR and virtual reality VR technology to generate a visual report to display the precipitation forecast situation intuitively and immersively.

3. The precipitation prediction system based on cloud parameter monitoring according to claim 1, characterized in that: The high-resolution satellite remote sensing sensor in the data acquisition module has multispectral imaging capabilities and can obtain cloud information at different wavelengths. By analyzing multispectral data, it can further accurately identify the material composition and microstructural characteristics of the cloud system, providing richer data dimensions for precipitation prediction. At the same time, the low-altitude UAV is equipped with a high-definition camera with adjustable focal length and a micro-meteorological sensor, which can automatically adjust the flight path and monitoring parameters according to the height and shape of the cloud system, thereby realizing refined data collection of specific cloud areas.

4. The precipitation prediction system based on cloud parameter monitoring according to claim 1, characterized in that: The data transmission module adopts data encryption and compression technology based on the integration of 5G and satellite communications; through advanced encryption algorithms, it ensures the security of data during transmission and prevents data from being stolen or tampered with; using efficient data compression algorithms, it compresses the large amount of collected data to reduce transmission bandwidth requirements and improve transmission efficiency, thus ensuring stable and fast data transmission even in complex communication environments.

5. The precipitation prediction system based on cloud parameter monitoring according to claim 1, characterized in that: In the preprocessing unit of the data processing center, the adaptive noise identification algorithm based on deep learning adopts transfer learning technology, which can quickly adapt to the data noise characteristics under different monitoring equipment and environments, and can achieve efficient noise removal of new data without the need for retraining a large number of samples; at the same time, the new missing value filling algorithm based on spatiotemporal correlation is combined with the Bayesian inference method, which can more accurately estimate missing values ​​according to the uncertainty of existing data, thereby improving the integrity and quality of the data.

6. The precipitation prediction system based on cloud parameter monitoring according to claim 1, characterized in that: The innovative algorithm combining the generative adversarial network (GAN) and the convolutional neural network (CNN) in the feature extraction unit introduces an attention mechanism, which can automatically focus on the most critical areas and features for precipitation prediction in cloud images, the edges of clouds and internal convection structures, further enhancing the accuracy and pertinence of feature extraction; moreover, the algorithm can adaptively adjust the feature extraction strategy according to different precipitation types such as heavy rain, light rain and snowfall, thereby improving the effectiveness of features.

7. The precipitation prediction system based on cloud parameter monitoring according to claim 1, characterized in that: The multimodal deep learning neural network model that integrates the attention mechanism in the model prediction unit adopts incremental learning technology, which can automatically update the model parameters and continuously improve the prediction accuracy as new historical cloud parameter data and precipitation data continue to accumulate. At the same time, the model also combines geographic information system (GIS) technology, which can incorporate terrain and geomorphic factors into the precipitation prediction model, consider the impact of terrain on cloud movement and precipitation formation, and improve the accuracy of precipitation prediction in different geographical regions.

8. The precipitation prediction method based on cloud parameter monitoring according to claim 2 is characterized in that: In the step 1, personalized data collection strategies are formulated for different application scenarios, such as urban meteorological disaster warning, agricultural irrigation water scheduling, and water conservancy project flood control; in the urban meteorological disaster warning scenario, the focus is on increasing the monitoring density of cloud systems and ground meteorological parameters in the urban surrounding areas; in the agricultural irrigation water scheduling scenario, cloud systems and meteorological data above farmland are collected in a targeted manner according to the distribution area and growth cycle of crops to meet the special needs of precipitation prediction in different scenarios.

9. The precipitation prediction method based on cloud parameter monitoring according to claim 2, characterized in that: After step three, data fusion is added to deeply fuse data from different sources and types, and to fuse satellite remote sensing data, ground radar data, and drone monitoring data, so as to fully utilize the advantages of each data source, eliminate inconsistencies and redundancies between data, and form a more comprehensive and accurate integrated data set, providing a better data foundation for subsequent feature extraction and model prediction.

10. The precipitation prediction method based on cloud parameter monitoring according to claim 2, characterized in that: The visualization report in step six not only uses augmented reality AR and virtual reality VR technology, but also combines artificial intelligence voice interaction technology; users can obtain detailed precipitation forecast information for a specific area and a specific time through voice commands, and realize natural interaction with the visualization report; at the same time, the visualization report can display precipitation forecast results in different forms according to user needs, dynamic charts, and three-dimensional models, so as to facilitate users to intuitively understand and analyze precipitation forecast information.

Images