Satellite and radar fusion short-term imminent forecasting method and device based on depth model compression

Abstract

The invention provides a satellite and radar fusion short-term imminent forecasting method and device based on depth model compression, and the method comprises the steps: generating meteorological element image data and meteorological element key features based on historical satellite data and historical radar data corresponding to a target region; a pre-constructed short-term and imminent forecasting model is obtained, depth model compression processing is carried out on the short-term and imminent forecasting model, and the short-term and imminent forecasting model comprises a long time sequence feature processing structure and a short time sequence feature processing structure which are connected in sequence; and training the short-term impending forecasting model after the depth model compression processing by using the ground observation data, the meteorological element image data and the meteorological element key features corresponding to the target area, the short-term impending forecasting model obtained by training being used for performing short-term impending forecasting on the target area based on the current satellite data and the current radar data. According to the method, the calculation overhead and the storage demand of the short-term and imminent forecasting model can be reduced on the premise of ensuring the forecasting precision.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a satellite and radar fusion short-term forecasting method and device based on deep model compression. Background Art

[0002] Traditional numerical weather prediction (NWP) simulates atmospheric processes based on physical models and can provide high-precision results for medium- and long-term weather forecasts. However, the complex calculation process of NWP requires high-performance computing resources and responds slowly to rapid weather changes in a short period of time (such as heavy rain, strong winds, thunderstorms, etc.). Short-term weather forecasts (i.e., short-term forecasts) aim to provide accurate predictions of weather conditions within the next few hours, especially for applications in sudden weather events. However, due to the rapid changes in short-term weather, the performance of traditional NWP models in short-term forecasts is not ideal.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially deep learning technology, more and more research has begun to apply it to the field of weather forecasting. Deep learning models can automatically extract spatiotemporal features in data by training massive historical data, and make accurate short-term weather forecasts based on these features. However, the high computational complexity and large storage requirements of deep learning models limit their application in resource-constrained environments, especially in scenarios where they are deployed on mobile devices and embedded systems. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a satellite and radar fusion short-term and impending forecast method and device based on deep model compression, which can reduce the computational overhead and storage requirements of the short-term and impending forecast model while ensuring the forecast accuracy.

[0005] In a first aspect, the present invention provides a satellite and radar fusion short-term forecasting method based on deep model compression, comprising:

[0006] Generate meteorological element image data and key features of meteorological elements based on historical satellite data and historical radar data corresponding to the target area;

[0007] Acquire a pre-built short-term forecast model, and perform deep model compression processing on the short-term forecast model, wherein the short-term forecast model includes a long-time series feature processing structure and a short-time series feature processing structure connected in sequence;

[0008] The short-term prediction model after deep model compression processing is trained using the ground observation data, meteorological element image data and key features of meteorological elements corresponding to the target area. The trained short-term prediction model is used to make short-term predictions for the target area based on current satellite data and current radar data.

[0009] In one embodiment, the meteorological element image data and the key features of the meteorological elements are generated based on the historical satellite data and the historical radar data corresponding to the target area, including:

[0010] Preprocess the historical satellite data and historical radar data corresponding to the target area to obtain the historical satellite data and historical radar data aligned in time and space;

[0011] The multi-source data fusion model obtained through pre-training generates meteorological element image data based on the historical satellite data and historical radar data aligned in time and space. The resolution of the meteorological element image data is higher than that of the historical satellite data and historical radar data.

[0012] In addition, through the pre-trained feature extraction model, feature extraction is performed based on the historical satellite data and historical radar data that are aligned in time and space to obtain the key features of meteorological elements.

[0013] In one embodiment, the multi-source data fusion model adopts a deep downscaling network, and the feature extraction model adopts a deep convolutional neural network.

[0014] In one implementation, performing deep model compression processing on the short-term forecast model includes:

[0015] The redundant connections of the short-term forecast model are pruned, and the floating-point weights of the short-term forecast model are quantized.

[0016] In one implementation, the short-term forecast model after the deep model compression processing is trained using the ground observation data, meteorological element image data and meteorological element key features corresponding to the target area, including:

[0017] The meteorological element image data and key features of meteorological elements are input into the short-term forecast model after deep model compression processing to generate the short-term forecast results corresponding to the target area;

[0018] Based on the short-term forecast results and the ground observation data corresponding to the target area, the short-term forecast model is trained, and the knowledge distillation technology is used to guide the forecast process of the short-term forecast model.

[0019] In one implementation, the long time series feature processing structure adopts a Transformer structure, and the short time series feature processing structure adopts a PredRNN structure.

[0020] In one embodiment, the meteorological element image data and the key features of the meteorological elements are input into the short-term forecast model after the deep model compression processing to generate the short-term forecast result corresponding to the target area, including:

[0021] The Transformer structure is used to capture the long-term series features in the meteorological element image data and the key features of the meteorological elements, and the PredRNN structure is used to generate the short-term forecast results corresponding to the target area based on the long-term series features.

[0022] In a second aspect, the present invention further provides a satellite and radar fusion short-term forecasting device based on deep model compression, comprising:

[0023] Satellite and radar fusion module, used to generate meteorological element image data and key features of meteorological elements based on historical satellite data and historical radar data corresponding to the target area;

[0024] A deep model compression module is used to obtain a pre-built short-term forecast model and perform deep model compression processing on the short-term forecast model, where the short-term forecast model includes a long-time series feature processing structure and a short-time series feature processing structure connected in sequence;

[0025] The model training module is used to train the short-term forecast model after the deep model compression processing using the ground observation data, meteorological element image data and key features of the meteorological elements corresponding to the target area. The trained short-term forecast model is used to make short-term forecasts for the target area based on the current satellite data and the current radar data.

[0026] In a third aspect, the present invention further provides an electronic device, comprising a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement any one of the methods provided in the first aspect.

[0027] In a fourth aspect, the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement any one of the methods provided in the first aspect.

[0028] The present invention provides a satellite and radar fusion short-term and impending forecast method and device based on deep model compression. First, based on historical satellite data and historical radar data corresponding to a target area, meteorological element image data and key features of meteorological elements are generated; then a pre-constructed short-term and impending forecast model is obtained, and deep model compression processing is performed on the short-term and impending forecast model, wherein the short-term and impending forecast model includes a long time series feature processing structure and a short time series feature processing structure connected in sequence; finally, the short-term and impending forecast model after the deep model compression processing is trained using ground observation data, meteorological element image data and key features of meteorological elements corresponding to the target area, and the trained short-term and impending forecast model is used to perform short-term and impending forecast on the target area based on current satellite data and current radar data. The above method combines deep model compression with short-term forecasting. While ensuring the forecast accuracy of the short-term forecasting model, it significantly improves the real-time reasoning capability of the short-term forecasting model, meets the timeliness requirements of short-term forecasting, and reduces the computational overhead and storage requirements of the short-term forecasting model. In addition, through the joint modeling of long and short time series features, the short-term forecasting model can make up for the shortcomings of a single model in dealing with complex meteorological systems, ensuring that the short-term forecasting model can retain global information while focusing on local details when capturing complex spatiotemporal characteristics, so as to further improve the forecast accuracy of the short-term forecasting model.

[0029] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0030] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 A schematic flow chart of a satellite and radar fusion short-term and impending forecast method based on deep model compression provided by an embodiment of the present invention;

[0033] Figure 2 A technical framework diagram of a satellite and radar fusion short-term forecasting method based on deep model compression provided by an embodiment of the present invention;

[0034] Figure 3 A network framework diagram of a short-term forecast model provided by an embodiment of the present invention;

[0035] Figure 4 A Transformer+PredRNN network framework and model compression flow graph provided for an embodiment of the present invention;

[0036] Figure 5 A schematic diagram of the structure of a satellite and radar fusion short-term forecasting device based on deep model compression provided by an embodiment of the present invention;

[0037] Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are 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.

[0039] In recent years, with the rapid development of deep learning technology, the field of weather forecasting has gradually shifted from traditional numerical weather prediction models (NWP) to data-driven deep learning models. Deep learning models have powerful feature extraction capabilities and can capture hidden meteorological patterns from massive historical data and use them to predict future changes in meteorological elements. Especially in short-term nowcasting, traditional numerical forecasting models often have limitations and cannot respond quickly to short-term changes in weather systems. Deep learning models can generate high-precision and fast forecast results with the help of multi-source heterogeneous data such as radar, satellites, and ground observations. Therefore, deep learning models have broad application prospects in short-term nowcasting.

[0040] However, the complexity of deep learning models brings challenges in computational cost and storage requirements. Especially for short-term forecasts that integrate satellite and radar, the model must process a large number of spatiotemporal features, which requires huge computation. When deployed in resource-constrained environments (such as mobile devices, embedded systems, or edge computing nodes), the computational and storage requirements of large-scale models become obstacles. Therefore, how to optimize deep learning models to adapt to these environments while ensuring forecast accuracy has become an important research topic in the field of meteorological forecasting.

[0041] Specifically, the existing short-term forecasting technology has at least the following problems:

[0042] Problem 1: Deep learning architectures in weather forecast models, especially networks used to process complex spatiotemporal data, usually contain a large number of parameters and hierarchical structures, which results in extremely large computational workloads. Short-term nowcasting requires models to produce results in real time or near real time, so inference speed is critical. If a model cannot generate forecast results within a reasonable time, even if its forecast accuracy is high, it will not meet the needs of practical applications.

[0043] Problem 2: The storage requirement of deep learning models is also an important factor affecting their actual deployment. In particular, the fusion processing of satellite and radar data requires a huge network structure, and the model parameters often reach millions or even tens of millions, resulting in huge consumption of storage space. Many mobile devices and embedded systems cannot carry such a large model due to limited hardware resources. Therefore, how to reduce the storage requirements of the model and ensure that it can be deployed in various restricted environments is another key problem solved by the embodiments of the present invention.

[0044] Question 3: Although model compression technology can effectively reduce computing and storage requirements, it is usually accompanied by a certain loss of accuracy. Especially in complex tasks such as weather forecasting, any slight forecast deviation may lead to errors in practical applications.

[0045] Question 4: With the rapid development of the Internet of Things (IoT) and edge computing, weather forecast models need to be deployed in a wider range of environments, such as sensor networks for weather monitoring, mobile devices, or edge computing nodes. These devices usually have limited computing resources and insufficient storage capacity, and cannot carry complex large-scale deep learning models.

[0046] In order to solve these problems, the present invention proposes a satellite and radar fusion short-term forecasting method based on deep model compression, which aims to reduce the computational overhead and storage requirements of the model through model compression technology, while maintaining high-precision forecasting performance and improving the deployment feasibility of the model in resource-constrained environments.

[0047] To facilitate understanding of this embodiment, firstly, a satellite and radar fusion short-term forecasting method based on deep model compression disclosed in an embodiment of the present invention is described in detail, see Figure 1 The flowchart of a satellite and radar fusion short-term forecast method based on deep model compression is shown, and the method mainly includes the following steps S102 to S106:

[0048] Step S102, based on the historical satellite data and historical radar data corresponding to the target area, meteorological element image data and meteorological element key features are generated, and the meteorological element key features can be key features of meteorological elements such as precipitation, temperature, wind speed, etc. In one example, the historical satellite data and historical radar data corresponding to the target area can be pre-processed by cleaning and integration, and the pre-processed data can be fused by multi-source data using a deep downscaling network to obtain meteorological element image data, and the pre-processed data can be feature extracted using a deep convolutional neural network to obtain meteorological element key features.

[0049] Step S104, obtain a pre-built short-term forecast model, and perform deep model compression processing on the short-term forecast model, the short-term forecast model includes a long-term feature processing structure and a short-term feature processing structure connected in sequence, the long-term feature processing structure is also a network structure that is good at capturing and processing long-term features, such as a Transform structure, and the short-term feature processing structure is also a network structure that is good at capturing and processing short-term features, such as a PredRNN structure. Deep model compression processing may include model compression technologies such as pruning, quantization, and knowledge distillation.

[0050] Step S106, using the ground observation data, meteorological element image data and meteorological element key features corresponding to the target area, the short-term forecast model after the deep model compression processing is trained, and the trained short-term forecast model is used to make short-term forecasts for the target area based on the current satellite data and the current radar data. In one example, the ground observation data is used as a label, and the meteorological element image data and the meteorological element key features are used as input to train the short-term forecast model. After the short-term forecast model is trained, the current satellite data and the current radar data corresponding to the target area can be obtained, and they are processed according to the aforementioned step S104 to obtain the corresponding meteorological element image data and the meteorological element key features, and then the trained short-term forecast model is used to make short-term forecasts for the target area based on the meteorological element image data and the meteorological element key features.

[0051] The satellite and radar fusion short-term forecasting method based on deep model compression provided by the embodiment of the present invention combines deep model compression with short-term forecasting, and while ensuring the forecasting accuracy of the short-term forecasting model, significantly improves the real-time reasoning capability of the short-term forecasting model, meets the timeliness requirements of short-term forecasting, and reduces the computational overhead and storage requirements of the short-term forecasting model; in addition, through the joint modeling of long and short time series features, the short-term forecasting model can make up for the shortcomings of a single model in processing complex meteorological systems, ensuring that the short-term forecasting model can retain global information while focusing on local details when capturing complex spatiotemporal characteristics, so as to further improve the forecasting accuracy of the short-term forecasting model.

[0052] Current weather forecasting faces the problem of processing multi-source heterogeneous data (such as satellite inversion data, radar echo data, and ground observation data), which have significant differences in temporal and spatial resolution. How to effectively fuse these data and combine them with advanced deep learning algorithms to provide accurate and fast short-term weather forecasts has become a key challenge in meteorology. The embodiments of the present invention optimize the data fusion process through deep learning technology, and combine model compression technology to solve the problem of high computational complexity, thereby providing an innovative solution for improving the accuracy and speed of weather forecasting.

[0053] For ease of understanding, the present invention provides a specific implementation of a satellite and radar fusion short-term weather forecast method based on deep model compression, which aims to improve the efficiency and accuracy of satellite and radar fusion short-term weather forecast through deep learning model compression technology and multi-source data fusion method. Figure 2 The technical framework diagram of a satellite and radar fusion short-term forecasting method based on deep model compression is shown, including multiple processes such as data collection, data cleaning and integration, data standardization, feature extraction and fusion, feature selection, deep learning model construction, model training and evaluation, and real-time weather forecast.

[0054] 1. Data collection:

[0055] (1.1) Satellite data: Obtain meteorological data from different satellite sensors, such as MODIS, GOES, GPM, etc., with a focus on band data related to precipitation (such as microwave, infrared and visible light bands). These data can provide a wide range of ground observation information.

[0056] (1.2) Radar data: Use weather radar to capture precipitation echo data, including reflectivity and velocity data, to obtain the temporal and spatial distribution of precipitation.

[0057] (1.3) Ground observation data: Real-time ground meteorological data such as temperature, humidity, wind speed, air pressure, and precipitation are collected from weather stations. These data usually have high temporal resolution but limited spatial coverage.

[0058] (ii) Data cleaning and integration: Preprocess the historical satellite data and historical radar data corresponding to the target area to obtain historical satellite data and historical radar data that are aligned in time and space. The preprocessing includes data cleaning and integration processing.

[0059] (2.1) Data cleaning: Process the collected satellite data and radar data to remove missing values ​​and outliers. Use interpolation methods (such as linear interpolation or spline interpolation) to fill in missing data, and use statistical methods to identify and process outliers to ensure the integrity and accuracy of the data.

[0060] (2.2) Data integration: Align satellite data and radar data from different sources in space and time. Since radar data and satellite data have different resolutions, they need to be resampled. The radar data can be spatially resampled using the nearest neighbor interpolation or bilinear interpolation method to make it consistent with the spatial resolution of the satellite data. At the same time, ensure that the timestamps of each data source are aligned, and unify the time step by interpolation or time windowing.

[0061] Specifically, by standardizing and time-aligning multi-source data, the embodiment of the present invention ensures that data from different sources are effectively integrated under the same framework, thereby enhancing the comprehensive performance of the model. The expression for data integration is as follows:

[0062]

[0063] Dynamically adjust the weight coefficient w i , to adapt to the importance of different data sources and improve the accuracy of fused data; F i For multi-source data, F fused The integrated data.

[0064] (III) Feature extraction and fusion:

[0065] (3.1) Feature extraction: The multi-source data fusion model obtained through pre-training generates meteorological element image data based on the historical satellite data and historical radar data that are aligned in time and space. The resolution of the meteorological element image data is higher than that of the historical satellite data and historical radar data. The multi-source data fusion model uses a deep downscaling network (such as ESRGAN) and an improved deep downscaling network to fuse data from different sources into a unified high-resolution data set (i.e., meteorological element image data). The generator part of the network generates high-resolution meteorological element image data by inputting low-resolution satellite data and radar data. During training, the goal is to minimize the loss between the meteorological element image data and the true high-resolution image.

[0066] Specifically, the embodiment of the present invention uses a deep learning model (such as ESRGAN) to reconstruct low-resolution satellite data to generate high-resolution meteorological element products. This method can greatly improve the temporal and spatial resolution of the data, thereby enhancing the model's ability to respond to meteorological changes.

[0067] F HR =G(Concat(F LR , C geo ));

[0068] By combining geographic information features, the reconstruction accuracy is improved, ensuring that the generated meteorological element data is more accurate and detailed. HRis high-resolution meteorological element image data, F LR For low-resolution satellite and radar data, G geo The data is from GEO (Group on Earth Observations).

[0069] The embodiment of the present invention uses a deep learning model (such as ESRGAN) to perform fusion processing of multi-source data, reconstruct satellite, radar, ground observation and other data with high resolution, and improve the accuracy of the forecast model in capturing key meteorological elements such as precipitation and wind speed. The fused data product reaches a spatial resolution of 100 meters, ensuring that information from different data sources can be effectively integrated and improve forecast accuracy.

[0070] (3.2) Multi-source data fusion: The feature extraction model obtained in advance is used to extract features based on the historical satellite data and historical radar data aligned in time and space to obtain the key features of meteorological elements. The feature extraction model uses a deep convolutional neural network (CNN) to extract features from the satellite data and radar data aligned in time and space. A specific convolutional layer is designed to extract the key features of meteorological elements such as precipitation, temperature, and wind speed, and feature selection is performed to reduce redundant information.

[0071] Specifically, the embodiment of the present invention utilizes a convolutional neural network (CNN) to extract key spatiotemporal features from multi-source data, ensuring that the model can capture the relevance and importance of data from different sources.

[0072] F = ReLU(W*X+b)+γ;

[0073] The introduction of residual connections not only speeds up the training process, but also improves the performance of the model when dealing with complex meteorological phenomena.

[0074] (IV) Deep learning model construction:

[0075] According to the needs of weather forecasting, a suitable deep learning model is selected to build a short-term forecast model, such as Figure 3 A network framework diagram of a short-term forecast model is shown. Exemplarily, the long-term feature processing structure in the short-term forecast model adopts the Transformer structure, and the short-term feature processing structure adopts the PredRNN structure. The present invention realizes efficient modeling of the spatiotemporal dependency of meteorological elements by combining the Transformer model and the PredRNN model. The Transformer model processes long-term features through a self-attention mechanism, and the PredRNN processes short-term complex spatial features through a spatiotemporal long short-term memory network (LSTM) network, which solves the shortcoming that traditional models are difficult to handle complex meteorological systems. Specifically:

[0076] (4.1) Transformer structure: This model is good at processing long time series data and can capture global dependencies. Through the multi-head self-attention mechanism, different parts of the input sequence can be paid attention to in parallel.

[0077] The Transformer structure uses multiple encoder layers, each of which consists of a self-attention mechanism and a feedforward neural network. Residual connections and layer normalization are used to enhance the stability and training effect of the model. Position encoding is used to retain the position information of the input sequence to ensure that the model can capture the time order.

[0078] The Transformer model uses the self-attention mechanism to model long-term time series features. By modeling the global correlation of each input in the sequence, the model's ability to capture complex time series data is improved.

[0079]

[0080] By introducing the regularization term ∈, the model is more stable when processing long sequences and alleviates the problem of information loss.

[0081] (4.2) PredRNN structure: This model combines convolutional neural networks (CNN) and spatiotemporal long short-term memory networks (LSTM), which can effectively process complex time series data and capture both short-term and long-term meteorological characteristics.

[0082] Among them, the PredRNN structure: convolution operations are designed to extract local features, and long short-term memory units (LSTM) are combined to process the dependencies of time series. Causal convolution is used to ensure the correctness of the time sequence. In the output layer of the model, appropriate activation functions (such as ReLU or Tanh) are designed to map the predicted meteorological elements to actual values.

[0083] The PredRNN model combines spatiotemporal long and short-term memory networks, which can handle temporal and spatial dependencies at the same time and is suitable for capturing the complex dynamic changes of meteorological elements.

[0084] h t =σ(W h h t-1 +W x x t +b+δ);

[0085] By introducing the adjustment factor δ, the model's ability to capture time dependence is enhanced, making the model more sensitive when dealing with meteorological elements.

[0086] The short-term forecast model is subjected to deep compression processing, such as pruning redundant connections of the short-term forecast model and quantizing floating-point weights of the short-term forecast model. Figure 4 The Transformer+PredRNN network framework and model compression flow chart shown in the figure, the embodiment of the present invention combines deep model compression technologies such as pruning, quantization, and knowledge distillation with the weather forecast model for the first time, so that the complex satellite and radar fusion weather forecast model can run efficiently on low-resource devices. Pruning reduces redundant connections, quantization reduces floating-point calculation overhead, and knowledge distillation maintains forecast accuracy. Specifically:

[0087] (a) Pruning: By calculating the weight importance of each connection, unimportant connections are identified and removed to reduce the number of model parameters. After pruning, the accuracy is restored by fine-tuning the model to ensure that the accuracy is not significantly affected.

[0088] The embodiment of the present invention effectively deletes those connections that have little impact on model performance by introducing a dynamic pruning strategy and combining the setting of weight thresholds. This strategy allows the model to dynamically adjust the pruning threshold during the training phase, making the pruning process more flexible and efficient.

[0089] Pruned Weight=w·Step(|w|-θ);

[0090] Among them, Prunen Weight is the weight vector after pruning; Step(x) is a threshold function that can dynamically adjust the weight according to real-time feedback to ensure that there is no excessive loss of accuracy during the pruning process; w is the original weight vector; θ is the weight threshold.

[0091] (b) Quantization: Convert the floating-point weights in the model into low-order integers to reduce storage requirements and computational overhead. A combination of static quantization and dynamic quantization is used to meet the needs of different inference stages.

[0092] In order to reduce storage requirements and improve inference speed, the embodiment of the present invention adopts a quantization method to convert floating point weights and activation values ​​into low-order integers. This method is particularly suitable for running on resource-constrained devices.

[0093]

[0094] Among them, q is a low-order integer, the round function is used to round floating-point numbers, and x is a floating-point weight or activation value. By standardizing data, the quantization process can effectively reduce the error during model reasoning and improve the stability of prediction.

[0095] (c) Knowledge distillation: During the training process, a complex teacher model is used to guide the learning of a simple student model, ensuring that the student model can effectively capture the complex characteristics of meteorological data and reduce accuracy loss.

[0096] The embodiment of the present invention utilizes knowledge distillation technology to introduce the knowledge of the teacher model in the process of training the small model, so that the small model can learn complex meteorological patterns without significant loss of accuracy.

[0097] L=αL hard +β·T 2 L soft ;

[0098] Among them, L is the loss value, α and β are weights, and L hard is the hard label loss value, L soft is the soft label loss value, and T is the temperature parameter, which is used to control the smoothness of the soft label to better capture the knowledge of the teacher model.

[0099] (V) Model training and evaluation:

[0100] (5.1) Data partitioning: The integrated dataset is divided into training set, validation set and test set. Usually, 70% of the data is used as the training set, 15% as the validation set, and 15% as the test set. This partitioning ensures that the model can be evaluated and tuned on different data.

[0101] (5.2) Training process:

[0102] Training parameter settings: Set hyperparameters such as learning rate and batch size. Select an appropriate loss function, such as mean square error (MSE) for regression tasks and cross entropy loss for classification tasks. Dynamically adjust the weight of the loss function according to the characteristics of meteorological elements to improve the prediction ability of important meteorological elements.

[0103] Furthermore, a multi-task learning strategy is adopted to simultaneously predict multiple meteorological elements, thereby improving the overall prediction performance. By modeling the interdependencies between tasks, the generalization ability of the model is improved. The loss function of the short-term forecast model is as follows:

[0104]

[0105] L i , L j is the loss function of different tasks, and λ is the weight. By considering the covariance between tasks, the model's prediction ability for multiple meteorological elements is optimized, making it more comprehensive and accurate. These joint modeling strategies enable the present invention to better capture the complex changes of meteorological elements and improve the accuracy of short-term forecasts.

[0106] Model training: The meteorological element image data and the key features of meteorological elements are input into the short-term forecast model after the deep model compression processing, so as to capture the long time series features in the meteorological element image data and the key features of meteorological elements through the Transformer structure, and generate the short-term forecast results corresponding to the target area based on the long time series features through the PredRNN structure; based on the short-term forecast results and the ground observation data corresponding to the target area, the short-term forecast model is trained, and the aforementioned knowledge distillation technology is used to guide the forecast process of the short-term forecast model.

[0107] Train the model on the training set and monitor the performance of the model on the validation set to prevent overfitting. Use early stopping to select the best model, set appropriate monitoring indicators (such as validation loss), and terminate training when the model performance no longer improves.

[0108] (5.3) Evaluation and tuning:

[0109] Model evaluation: Use the test set to evaluate the model performance and calculate key indicators such as root mean square error (RMSE), accuracy, recall, and F1 score. Based on the evaluation results, generate a visual report to facilitate the analysis of the model's performance under different meteorological conditions.

[0110] Model tuning: Based on the evaluation results, fine-tune the model to optimize hyperparameters and model structure. Use cross-validation to ensure the generalization ability of the model, especially in extreme weather conditions.

[0111] (VI) Real-time weather forecast:

[0112] (6.1) Short-term and nowcasting: Apply the trained model to the actual short-term and nowcasting task. Input the satellite data and radar data acquired in real time, and use the model to predict factors such as precipitation, temperature, and wind speed. The model should be able to process data in real time and generate forecast results within a few minutes to support meteorological decision-making.

[0113] (6.2) Real-time monitoring and early warning: Combine mobile devices and edge computing nodes to deploy compressed deep learning models for real-time weather monitoring and early warning. The model should be able to make timely predictions and provide alerts for extreme weather events such as heavy precipitation and storms. This can be achieved by connecting to the meteorological monitoring network to receive and analyze sensor data in real time.

[0114] (6.3) Data feedback and model updating: Continuously update and optimize the model by collecting feedback on actual weather data and forecast results. Use new data to retrain the model to improve its long-term forecast performance. Set up a feedback mechanism to compare actual observations with model outputs and adjust model parameters or structure to adapt to changing meteorological conditions.

[0115] Furthermore, the embodiments of the present invention can be efficiently deployed in weather forecasting in resource-constrained environments, specifically:

[0116] Through the application of model compression technology, the present invention enables complex deep learning weather forecast models to be efficiently deployed on mobile devices or embedded systems, providing real-time short-term weather forecast capabilities for devices with limited resources, greatly expanding the application scenarios of weather forecast models. The embodiments of the present invention achieve lightweight deployment by simplifying the network structure (model lightweighting) and reducing the number of model layers or neurons.

[0117] ELOPs reduced =FLOPs original (1-p);

[0118] Where p is the pruning ratio, which quantifies the significant reduction in computation during lightweighting, allowing the model to reason quickly. Soriginal For the simplified model, FLOP Sredeced This is a simplified model.

[0119] Furthermore, in order to optimize the reasoning process on the low-power device, the embodiment of the present invention applies an efficient reasoning framework (low-power reasoning algorithm) to reduce the overall power consumption by adjusting the current and voltage.

[0120] Power optimized =V·I optimized with I optimized =I base (1-q);

[0121] By optimizing the current I optimized , reducing power consumption and ensuring efficient operation in power-constrained environments. V is voltage; I base is the basic current; q is the charge.

[0122] Furthermore, a dynamic reasoning strategy is adopted to select a suitable model for reasoning according to the resource status of the current device, ensuring high efficiency in a changing environment.

[0123]

[0124] R is the current resource volume. Through this dynamic adjustment mechanism, the model can flexibly respond to different environmental requirements and maintain efficient real-time forecasting capabilities. t hresh olr It is the two-point threshold of the current resource. Full Model represents the complete model, and Compressed Model represents the compressed model.

[0125] The specific implementation of the present invention solves the deployment problem of deep models in resource-constrained environments by systematically collecting and processing multi-source meteorological data, combined with the construction and compression technology of deep learning models, and realizes efficient and accurate short-term and near-term weather forecasts. During the implementation process, the various technologies and methods adopted ensure the availability of models and the efficiency of predictions in resource-constrained environments, and are particularly suitable for applications in mobile devices and embedded systems, providing a reliable solution for the practical application of weather forecasts.

[0126] The satellite and radar fusion short-term forecasting method based on deep model compression provided by the embodiment of the present invention can achieve the following effects:

[0127] In response to the aforementioned problem one, the embodiment of the present invention can improve the model calculation efficiency and optimize the reasoning speed. Specifically: The embodiment of the present invention adopts key point one: the combination of deep model compression and weather forecasting. Through compression technologies such as pruning, quantization, and weight sharing, the computational complexity of the model is effectively reduced. Pruning greatly reduces the amount of calculation by deleting unimportant neuron connections in the model, while quantization technology converts floating-point operations into low-order integer operations, further reducing the computational load. These technologies not only speed up the reasoning speed of the model, but are also particularly suitable for resource-constrained hardware environments such as mobile devices and embedded systems. Through this innovative model compression, the embodiment of the present invention significantly improves its real-time reasoning capabilities while ensuring the accuracy of model forecasts, meeting the timeliness requirements of short-term and near-forecasts.

[0128] In response to the aforementioned problem two, the embodiment of the present invention can reduce storage requirements and improve the flexibility of model deployment. Specifically: The embodiment of the present invention optimizes the multi-source data processing process through key point two: an innovative method of multi-source data fusion, combines weight sharing and compression coding technology, and reduces the storage requirements of the model. In the process of multi-source data fusion, a deep learning framework (such as ESRGAN) is used for efficient data processing, and through weight sharing technology, multiple neurons are forced to share the same parameters to significantly reduce the storage volume of the model. In particular, in the process of fusing satellite inversion data and radar echo data, the model can automatically learn the correlation between these data, and reduce storage requirements through compression technology, so that large-scale forecasting models can run efficiently in embedded systems and mobile devices.

[0129] In response to the aforementioned problem 2, the embodiment of the present invention can ensure a balance between forecast accuracy and model compression. Specifically: While compressing the model, the embodiment of the present invention pays special attention to how to maintain the forecast accuracy of the model. To this end, the embodiment of the present invention adopts the third key point: joint modeling of long and short time series features, which combines the advantages of different model architectures to ensure that the forecast accuracy is not significantly affected.

[0130] Specifically, the embodiment of the present invention combines the Transformer model and the PredRNN model. The Transformer model is used to capture long-distance spatiotemporal dependency features, while the PredRNN processes short-term dynamic changes through convolutional spatiotemporal LSTM units. This combination can make up for the shortcomings of a single model in dealing with complex meteorological systems, ensuring that the model can retain global information while focusing on local details when capturing complex spatiotemporal features. In addition, through the knowledge distillation technology, the knowledge of the complex model is transferred to the simplified model, ensuring that the compressed model can maintain good forecasting performance without significantly reducing the accuracy.

[0131] In response to the aforementioned problem 2, the embodiments of the present invention can adapt to the real-time forecasting needs in resource-constrained environments. Specifically: The ultimate goal of the embodiments of the present invention is to enable the deep learning weather forecast model to run efficiently in these resource-constrained environments through model compression technology, and provide real-time and accurate short-term weather forecasts.

[0132] Key point 4 of the embodiment of the present invention: Efficient deployment of weather forecasts in resource-constrained environments, through pruning, quantization, knowledge distillation and other technologies, the complex deep learning forecast model is compressed into a lightweight version, so that it can achieve efficient reasoning on mobile devices or embedded systems. In particular, for real-time meteorological monitoring and warning tasks, such as rainstorm warnings, strong wind forecasts, etc., the compressed model can respond quickly on low-power devices, thereby effectively improving the practical application capabilities of weather forecasts.

[0133] In summary, the aforementioned key points one to four not only improve the computational efficiency and storage requirements of the weather forecast model, but also ensure that it can run efficiently in a resource-constrained environment and maintain high-precision forecasting capabilities. The purpose of the embodiment of the present invention is to solve the problems of high model computational complexity, large storage requirements, and difficult deployment in weather forecasting by applying deep learning model compression technology to the short-term nowcasting task of satellite and radar data fusion. By adopting technologies such as pruning, quantization, and knowledge distillation, the embodiment of the present invention improves computing efficiency and reduces storage requirements while maintaining the high forecast accuracy of the model, enabling it to run efficiently in a resource-constrained hardware environment.

[0134] Based on the above embodiments, the present invention provides a satellite and radar fusion short-term forecasting device based on deep model compression, see Figure 5 The schematic diagram of the structure of a satellite and radar fusion short-term forecast device based on deep model compression is shown, and the device mainly includes the following parts:

[0135] The satellite and radar fusion module 502 is used to generate meteorological element image data and meteorological element key features based on historical satellite data and historical radar data corresponding to the target area;

[0136] A deep model compression module 504 is used to obtain a pre-built short-term forecast model and perform deep model compression processing on the short-term forecast model, wherein the short-term forecast model includes a long-term feature processing structure and a short-term feature processing structure connected in sequence;

[0137] The model training module 506 is used to train the short-term prediction model after the deep model compression processing using the ground observation data, meteorological element image data and key features of the meteorological elements corresponding to the target area. The trained short-term prediction model is used to perform short-term prediction on the target area based on the current satellite data and the current radar data.

[0138] The satellite and radar fusion short-term forecasting device based on deep model compression provided by the embodiment of the present invention combines deep model compression with short-term forecasting, and while ensuring the forecasting accuracy of the short-term forecasting model, significantly improves the real-time reasoning capability of the short-term forecasting model, meets the timeliness requirements of short-term forecasting, and reduces the computational overhead and storage requirements of the short-term forecasting model; in addition, through the joint modeling of long and short time series features, the short-term forecasting model can make up for the shortcomings of a single model in processing complex meteorological systems, ensuring that the short-term forecasting model can retain global information while focusing on local details when capturing complex spatiotemporal characteristics, so as to further improve the forecasting accuracy of the short-term forecasting model.

[0139] In one implementation, the satellite and radar fusion module 502 is specifically used to:

[0140] Preprocess the historical satellite data and historical radar data corresponding to the target area to obtain the historical satellite data and historical radar data aligned in time and space;

[0141] The multi-source data fusion model obtained through pre-training generates meteorological element image data based on the historical satellite data and historical radar data aligned in time and space. The resolution of the meteorological element image data is higher than that of the historical satellite data and historical radar data.

[0142] In addition, through the pre-trained feature extraction model, feature extraction is performed based on the historical satellite data and historical radar data that are aligned in time and space to obtain the key features of meteorological elements.

[0143] In one embodiment, the multi-source data fusion model adopts a deep downscaling network, and the feature extraction model adopts a deep convolutional neural network.

[0144] In one implementation, the deep model compression module 504 is specifically used to:

[0145] The redundant connections of the short-term forecast model are pruned, and the floating-point weights of the short-term forecast model are quantized.

[0146] In one embodiment, the model training module 506 is specifically used to:

[0147] The meteorological element image data and key features of meteorological elements are input into the short-term forecast model after deep model compression processing to generate the short-term forecast results corresponding to the target area;

[0148] Based on the short-term forecast results and the ground observation data corresponding to the target area, the short-term forecast model is trained, and the knowledge distillation technology is used to guide the forecast process of the short-term forecast model.

[0149] In one implementation, the long time series feature processing structure adopts a Transformer structure, and the short time series feature processing structure adopts a PredRNN structure.

[0150] In one embodiment, the model training module 506 is specifically used to:

[0151] The Transformer structure is used to capture the long-term series features in the meteorological element image data and the key features of the meteorological elements, and the PredRNN structure is used to generate the short-term forecast results corresponding to the target area based on the long-term series features.

[0152] The device provided in the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0153] An embodiment of the present invention provides an electronic device. Specifically, the electronic device includes a processor and a storage device. The storage device stores a computer program, and when the computer program is executed by the processor, it executes the method described in any one of the above-mentioned implementation methods.

[0154] Figure 6A structural diagram of an electronic device provided in an embodiment of the present invention, the electronic device 100 includes: a processor 60, a memory 61, a bus 62 and a communication interface 63, wherein the processor 60, the communication interface 63 and the memory 61 are connected via the bus 62; the processor 60 is used to execute an executable module stored in the memory 61, such as a computer program.

[0155] The memory 61 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 63 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0156] The bus 62 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0157] Among them, the memory 61 is used to store programs, and the processor 60 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiment of the present invention can be applied to the processor 60 or implemented by the processor 60.

[0158] The processor 60 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 60. The above processor 60 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present invention can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 61, and the processor 60 reads the information in the memory 61 and completes the steps of the above method in combination with its hardware.

[0159] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0160] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0161] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A satellite and radar fusion short-term forecasting method based on deep model compression, characterized in that: include: Generate meteorological element image data and key features of meteorological elements based on historical satellite data and historical radar data corresponding to the target area; Acquire a pre-built short-term forecast model, and perform deep model compression processing on the short-term forecast model, wherein the short-term forecast model includes a long time series feature processing structure and a short time series feature processing structure connected in sequence; The short-term prediction model after the deep model compression processing is trained using the ground observation data corresponding to the target area, the meteorological element image data and the key features of the meteorological elements. The trained short-term prediction model is used to perform short-term prediction on the target area based on current satellite data and current radar data.

2. The satellite and radar fusion short-term forecasting method based on deep model compression according to claim 1 is characterized in that: Based on the historical satellite data and historical radar data corresponding to the target area, meteorological element image data and key features of meteorological elements are generated, including: Preprocessing the historical satellite data and the historical radar data corresponding to the target area to obtain the historical satellite data and the historical radar data aligned in time and space; Generate meteorological element image data based on the spatiotemporally aligned historical satellite data and the historical radar data through a pre-trained multi-source data fusion model, wherein the resolution of the meteorological element image data is higher than that of the historical satellite data and the historical radar data; Furthermore, by using a pre-trained feature extraction model, feature extraction is performed based on the historical satellite data and the historical radar data that are aligned in time and space to obtain key features of meteorological elements.

3. The satellite and radar fusion short-term forecasting method based on deep model compression according to claim 2 is characterized in that: The multi-source data fusion model adopts a deep downscaling network, and the feature extraction model adopts a deep convolutional neural network.

4. The satellite and radar fusion short-term forecasting method based on deep model compression according to claim 1 is characterized in that: The short-term forecast model is subjected to deep model compression processing, including: The redundant connections of the short-term prediction model are pruned, and the floating-point weights of the short-term prediction model are quantized.

5. The satellite and radar fusion short-term forecasting method based on deep model compression according to claim 1 is characterized in that: Using the ground observation data corresponding to the target area, the meteorological element image data and the key features of the meteorological elements, the short-term forecast model after the deep model compression processing is trained, including: Inputting the meteorological element image data and the meteorological element key features into the short-term forecast model after the deep model compression processing to generate a short-term forecast result corresponding to the target area; Based on the short-term forecast results and the ground observation data corresponding to the target area, the short-term forecast model is trained, and the knowledge distillation technology is used to guide the forecast process of the short-term forecast model.

6. The satellite and radar fusion short-term forecasting method based on deep model compression according to claim 5 is characterized in that: The long time series feature processing structure adopts the Transformer structure, and the short time series feature processing structure adopts the PredRNN structure.

7. The satellite and radar fusion short-term forecasting method based on deep model compression according to claim 6 is characterized in that: Inputting the meteorological element image data and the meteorological element key features into the short-term forecast model after the deep model compression processing to generate the short-term forecast result corresponding to the target area, including: The Transformer structure is used to capture the long-term series features in the meteorological element image data and the key features of the meteorological elements, and the PredRNN structure is used to generate the short-term forecast results corresponding to the target area based on the long-term series features.

8. A satellite and radar fusion short-term forecast device based on deep model compression, characterized in that: include: Satellite and radar fusion module, used to generate meteorological element image data and key features of meteorological elements based on historical satellite data and historical radar data corresponding to the target area; A deep model compression module, used to obtain a pre-built short-term forecast model and perform deep model compression processing on the short-term forecast model, wherein the short-term forecast model includes a long time series feature processing structure and a short time series feature processing structure connected in sequence; The model training module is used to train the short-term forecast model after the deep model compression processing by using the ground observation data corresponding to the target area, the meteorological element image data and the key features of the meteorological elements. The short-term forecast model obtained by training is used to perform short-term forecast on the target area based on current satellite data and current radar data.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 7.

Images