Lightweight rapid forecasting system and method for typhoon and storm surge

Abstract

The invention belongs to the technical field of extreme ocean phenomenon forecasting, and discloses a typhoon and storm surge lightweight rapid forecasting system and method. The method comprises the following steps: preprocessing a multi-source marine environment data set and a typhoon moving path; constructing a typhoon and storm surge data set based on the preprocessed data; based on the constructed typhoon and storm surge data set, establishing a deep learning model VIT-TSS based on VIT; and comparing the output result of the deep learning model VIT-TSS with the open source data and the measured data to verify the forecast precision. The typhoon and storm surge intelligent forecasting model provided by the invention can realize typhoon path forecasting and regional water increase forecasting. According to the VIT-TSS model provided by the invention, the prediction time of the model is greatly shortened while the calculation precision is improved. The system provides key information for emergency management and navigation.

Description

Technical Field

[0001] The present invention belongs to the technical field of extreme ocean phenomenon forecasting, and in particular relates to a typhoon and storm surge lightweight rapid forecasting system and method. Background Art

[0002] Storm surge and typhoon are two extreme ocean phenomena that have a serious impact on coastal areas and maritime traffic safety. A typhoon is a tropical cyclone, a strong storm caused by the energy released by large-scale evaporation on tropical oceans. Storm surge refers to the ocean tidal phenomenon formed by wind. When strong winds blow towards the coastline, the wind exerts pressure on the sea surface, pushing the sea water toward the land. This wind-induced water movement will cause the sea level to rise and form a high water level phenomenon in coastal areas, namely storm surge. Typhoons and the storm surges they bring not only pose a serious threat to the economic development of coastal cities, but also pose a huge risk to the safety of people's lives and property. Therefore, timely and effective forecasting and warning of typhoons and storm surges are of great significance to improving coastal disaster prevention and mitigation capabilities and protecting the lives and property of coastal people.

[0003] Based on the actual needs of early warning of typhoons and storm surges and ensuring coastal safety, the present invention proposes a lightweight rapid forecasting system for typhoons and storm surges based on the VIT model. Based on machine learning, rapid forecasting of typhoon paths and storm surges is achieved, providing technical means for ensuring the safety of life and property in coastal areas.

[0004] Currently, there are methods based on parameter methods, LSTM, GRU and CNN model architectures to predict the typhoon path at the current moment or in the future through historical data. However, the parameter method model is relatively simple and has low accuracy; CNN, GRU and other models are relatively weak in capturing the spatial features of images; LSTM has low computational efficiency, especially when processing large-scale data, the training and inference speeds may not meet real-time requirements.

[0005] In addition, there are currently single storm surge or typhoon intelligent forecasts, but due to the strong correlation between typhoons and storm surges, single forecasts ignore the correlation and physical process between typhoons and storm surges. There is currently a lack of research on the joint intelligent forecast of typhoons and storm surges, and it is impossible to make accurate warnings for extreme disasters.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows: Regarding the problem of the scope of the sea area for storm surge water increase forecast, the current water increase forecast is mostly carried out for a single location, and there is a lack of regional water increase forecast, which cannot meet the demand for water increase forecast in nearshore waters.

[0007] For the problem of intelligent forecasting of typhoon paths and storm surges, there are currently methods for intelligent forecasting based on LSTM, GRU and CNN model architectures, but they are relatively weak in capturing the spatial features of images. Secondly, the computational efficiency of existing forecasting models is low, especially when processing large-scale data, the speed of training and inference may not meet the real-time requirements.

[0008] Finally, the existing extreme ocean phenomenon forecasting system forecasts typhoon paths and storm surges separately, and storm surges are caused by typhoons. Considering the forecast of one phenomenon alone cannot fully assess potential risks. Summary of the invention

[0009] In order to overcome the problems existing in the related art, the disclosed embodiments of the present invention provide a lightweight rapid forecasting system and method for typhoons and storm surges, and specifically relate to a lightweight rapid forecasting system for typhoons and storm surges based on a VIT model.

[0010] The technical solution is as follows: A lightweight rapid forecasting method for typhoons and storm surges, comprising the following steps: S1, preprocessing of multi-source ocean environment data sets and typhoon movement paths; S2, construct typhoon and storm surge dataset based on preprocessed data; S3, based on the constructed typhoon and storm surge dataset, establish the VIT-TSS deep learning model based on VIT to achieve the joint forecast of typhoon path and storm surge water level changes; S4, compare the output results of the deep learning model VIT-TSS with open source data and measured data to verify the forecast accuracy.

[0011] In step S1, the multi-source ocean environment data set and typhoon movement path are preprocessed, including: unifying the temporal and spatial resolution of the input data through interpolation processing; The input multi-source marine environment data sets include: gridded meteorological data of wind speed, temperature, and geopotential height; The input typhoon moving path includes: the longitude and latitude information of the typhoon center.

[0012] In step S2, a typhoon data set is constructed, including: historical longitude and latitude information of typhoon centers over many years and environmental information around the typhoon centers; Construct a storm surge dataset, including sea surface height data and gridded meteorological data. The sea surface height data contains two sets of different spatial resolutions. One set of coarse grid datasets contains hourly average sea surface height data with a spatial resolution of 0.5°×0.5°; the other set of high-precision datasets contains hourly average sea surface height data with a spatial resolution of 1 / 12°×1 / 12°. The meteorological data include hourly average wind speed and air pressure data with a spatial resolution of 0.25°×0.25°.

[0013] In step S3, a VIT-TSS deep learning model based on VIT is established, including: (1) Using the self-attention mechanism VIT to dynamically calculate the attention weights between different positions, adjust the representation of each position according to the importance of different positions in the input sequence, and capture the long-range dependencies and important information in the sequence; (2) Based on the self-attention mechanism VIT, a deep learning model VIT-TSS is built to enhance the VIT model's ability to learn the relationships between input features. The deep learning model VIT-TSS is applied to typhoon and storm surge intelligent forecasting. The encoding block of the deep learning model VIT-TSS is used to process typhoon path, sea surface height and wind speed meteorological data, and the above data are converted into useful feature representations.

[0014] Step (1) specifically includes: (1.1) Perform input transformation, input vector After three trainable weight matrices Transformation, respectively obtain the query vector query, key vector key, value vector value; the above vectors are used to calculate the attention weight and the final output; (1.2) Calculate the attention weight; for each query vector, use the dot product calculation method to calculate the similarity with all key vectors, through The function converts these similarities into attention weights; (1.3) The outputs of each of the multiple attention heads are concatenated together by multiplying them by another learnable weight matrix , and get the final output.

[0015] In step (1.1), the input data of typhoon path and sea surface height are transformed to learn the data in different ways. The process is expressed as: ; In the formula, is the query vector, is the key vector, is a value vector.

[0016] In step (1.2), to prevent The gradient disappears, Before converting, divide the value by the vector Dimensions The square root of and the output vector , the process is expressed as: ; ; In the formula, is the input vector For the output vector The attention weight, The normalization function transforms an arbitrary set of real numbers into a probability distribution such that each weight is between 0 and 1 and the sum of all weights is equal to 1. is the query vector, is the key vector, is the key vector The dimension represents a single vector The size of the feature dimension; is the output vector; is the input vector, representing the query vector With key vector The similarity between After normalization of the function, we get: is a value vector used for weighted summation to generate the final output; In step (1.3), the final output is: ; In the formula, is the overall output of the multi-head self-attention mechanism, through multiple attention heads After the parallel calculations are completed, the results are stitched together and passed through a learned weight matrix Perform linear transformation to obtain the final output; The three matrices generated after the linear transformation of the input feature vector are the core inputs of the attention mechanism and are defined as follows: Query, which indicates the target vector or query vector to be calculated for attention; is the key, which indicates the source vector to be matched or compared; is a value, which represents the value vector that the input data actually needs to obtain after matching is completed; Will The output of an attention head Splice along the feature dimension, ,in, is the number of heads, The dimensions of each head output, is the set of real numbers, containing all real values.

[0017] In step (2), building the deep learning model VIT-TSS includes: (2.1) The attention mechanism and convolution operation are integrated for data encoding and decoding. First, the input data, including typhoon path, sea surface height, and wind speed, are preprocessed. The convolution layer of the model is used to extract the features of the image blocks. The local information in the image is captured through the sliding operation of the convolution kernel. Subsequently, the sequential information of the elements in the data is introduced through position encoding to enhance the understanding of the time series. (2.2) The regularized data is fed into the VIT-TSS model encoding module to further process and encode the features. After multi-head attention processing, the regularization and normalization steps are applied again to maintain the stability of network training and help learn robust feature representations. (2.3) The processed data is reshaped into the format required for forecasting typhoon path information and storm surge height, and further transposed convolution is performed to adjust the size of the feature map and output it.

[0018] In step S4, before training, the original data is normalized, and according to the characteristics of typhoons and storm surges, the root mean square error (RMSE) and mean relative error (MRE) are used to evaluate the effectiveness of the VIT-TSS model.

[0019] Another object of the present invention is to provide a typhoon and storm surge lightweight rapid forecasting system, which implements the typhoon and storm surge lightweight rapid forecasting method, and the system comprises: Data preprocessing module, used to preprocess multi-source ocean environment data sets and typhoon movement paths; Typhoon and storm surge dataset construction module, used to construct typhoon and storm surge datasets based on preprocessed data; The deep learning model building module is used to build a VIT-TSS deep learning model based on the constructed typhoon and storm surge dataset, and realize the joint forecast of typhoon path and storm surge water increase changes; The forecast accuracy verification module is used to compare the output results of the deep learning model VIT-TSS with open source data and measured data to verify the forecast accuracy.

[0020] In combination with all the above technical solutions, the beneficial effects of the present invention are as follows: (1) In view of the current problems of untimely and inaccurate forecasting of extreme ocean events, a lightweight and rapid forecasting system for typhoons and storm surges is developed. Based on the deep learning method, an intelligent forecasting model for extreme ocean events such as storm surges and typhoons is studied, and the model sensitivity and model calculation performance are analyzed. A lightweight and rapid joint forecasting system including storm surges and typhoon extreme events is developed to achieve rapid and accurate forecasting of extreme ocean events. It can provide strong data support for ocean-related events and disaster warnings. The typhoon and storm surge intelligent forecasting model proposed in the present invention can realize regional water increase forecasting. The VIT-TSS model proposed in the present invention greatly shortens the forecasting time of the model while improving the calculation accuracy. The present invention realizes the comprehensive forecasting of typhoon paths and storm surge water increase, provides comprehensive warning of extreme disasters caused by typhoons, and provides key information for emergency management and navigation.

[0021] (2) The present invention significantly improves the timeliness and accuracy of forecasting extreme marine events. The lightweight rapid forecasting system established by deep learning methods can achieve efficient and intelligent forecasting of extreme marine events such as typhoons and storm surges, significantly shorten the forecasting time, improve the forecasting accuracy, and provide real-time and reliable forecasting information for decision-making departments, emergency management agencies and marine-related industries. The system is lightweight and computationally efficient, and can be easily deployed on a variety of hardware platforms, suitable for scientific research institutions and corporate users of different sizes.

[0022] (3) Aiming at the problem that the current storm surge water increase forecast cannot meet the regional needs, the present invention proposes an innovative solution to realize regional water increase forecast, breaking through the limitation of the previous water increase forecast only for a single location, and meeting the actual needs of regional water increase forecast in nearshore waters. In addition, the present invention realizes the comprehensive intelligent forecast of typhoon path and storm surge water increase, making up for the shortcomings of the existing forecasting system that independently forecasts typhoon path and storm surge and cannot comprehensively assess the potential disaster risks caused by typhoons, providing key support for comprehensive early warning of extreme marine disasters.

[0023] (4) Existing intelligent forecasting models such as LSTM, GRU and CNN are not good enough in capturing image spatial features, and it is difficult to fully explore the complex spatiotemporal relationship between storm surge and typhoon data. At the same time, when faced with large-scale data, the computational efficiency of these models is low and cannot meet the real-time forecasting needs of extreme ocean events. The VIT-TSS model proposed in the present invention effectively improves the ability to capture spatial features by integrating the visual Transformer structure, achieves higher calculation accuracy, and significantly shortens the model training and reasoning time, thereby meeting the real-time requirements. In addition, the present invention incorporates typhoon paths and storm surge water levels into a unified intelligent forecasting framework to achieve comprehensive assessment and accurate early warning of extreme ocean events. This technological breakthrough is an effective improvement and innovation of the current single forecasting model.

[0024] (5) The present invention integrates the correlation between typhoon path and storm surge. In the prior art, the forecast of typhoon path and storm surge is usually carried out independently, ignoring the direct correlation between typhoon and storm surge, resulting in incomplete forecast results. The present invention incorporates typhoon path and storm surge into a unified intelligent forecast model, realizes coupling analysis and comprehensive forecast between the two, thereby eliminating the one-sidedness and deviation caused by single factor forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure; Figure 1 This is a flow chart of a lightweight rapid forecasting method for typhoons and storm surges provided by an embodiment of the present invention; Figure 2 It is a schematic diagram of the VIT-TSS model structure and data forward propagation process provided by an embodiment of the present invention; Figure 3 is a schematic diagram of a lightweight rapid forecasting system for typhoons and storm surges provided by an embodiment of the present invention; In the figure: 1. Data preprocessing module; 2. Typhoon and storm surge dataset construction module; 3. Deep learning model establishment module; 4. Forecast accuracy verification module. DETAILED DESCRIPTION

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0027] In response to the current problems of untimely and inaccurate forecasts of extreme marine events, the present invention proposes a lightweight rapid forecasting system for typhoons and storm surges based on the Transformer deep learning model. Compared with other intelligent forecasting models, the present invention proposes a VIT-TSS model, which fully considers the temporal and spatial variation characteristics of typhoons and storm surges, while improving the calculation accuracy and greatly shortening the forecasting time of the model. In addition, the present invention also considers the strong correlation between typhoons and storm surges, and realizes the rapid joint forecasting of typhoon paths and storm surge water levels, providing key information for emergency management and navigation, minimizing casualties and losses from marine disasters, and reducing the adverse effects of extreme marine events on society and the economy.

[0028] Embodiment 1, as Figure 1As shown, the typhoon and storm surge lightweight rapid forecasting method provided by the embodiment of the present invention carries out typhoon and storm surge intelligent forecasting based on the deep learning method, including the following steps: S1, preprocessing of multi-source ocean environment data sets and typhoon movement paths; Including but not limited to changing the resolution of the data set through interpolation, dividing the time and space, and improving the accuracy of the data set through measured data. The data set for deep learning model training is obtained by preprocessing the data source; S2, construct typhoon and storm surge dataset based on preprocessed data; S3, based on the constructed typhoon and storm surge dataset, establish the VIT-TSS deep learning model based on VIT to achieve the joint forecast of typhoon path and storm surge water level changes; S4, compare the output results of the deep learning model VIT-TSS with open source data and measured data to verify the forecast accuracy.

[0029] Among them, the core idea of ​​the deep learning model VIT model is to process input and output sequences completely based on the self-attention mechanism. The advantage of this architecture is that it can process elements in the sequence in parallel, thereby speeding up training.

[0030] The self-attention mechanism is an attention mechanism used in machine learning and natural language processing. It allows the model to dynamically focus on elements at different positions in the sequence when processing sequence data, so as to better understand the dependencies within the sequence. A key advantage of the self-attention mechanism is that it can capture long-range dependencies between any two positions in the input sequence without explicitly defining the structure of the sequence like recurrent neural networks (RNNs) or convolutional neural networks (CNNs). This makes the self-attention mechanism perform well when processing long sequences. In the intelligent forecasting of typhoons and storm surges, the role of the self-attention mechanism is mainly reflected in its ability to capture and process long-range dependencies in typhoon path sequence data. In forecasting, this mechanism can help the model identify and learn the complex relationships between various time points in time series data, thereby improving the accuracy of typhoon and storm surge forecasts.

[0031] In the self-attention mechanism, for the input of a certain vector, the propagation process is as follows: First, the input transformation is performed, the input vector After three trainable weight matrices Transformation, respectively obtain query vector (query), key vector (key), value vector (value). These vectors are used to calculate attention weights and the final output. Input transformation processing of input data such as typhoon path and sea surface height in the present invention can enable the model to learn data in different ways and improve the model's ability to handle complex relationships. The innovative proposal of the present invention, this process can be expressed as: ; In the formula, is the query vector, is the key vector, is a value vector.

[0032] Then calculate the attention weight. For each query vector, calculate its similarity with all key vectors, usually using dot product calculation, and then The function converts these similarities into attention weights. These weights represent the importance of different positions to the current position, and the larger the weight, the more important it is. The gradient disappears, Before converting, divide the value by the vector Dimensions The square root of . In the present invention, the calculation of attention weights can focus on the most important information for typhoon and storm surge forecasts. The typhoon path information, sea surface height, wind speed and other data with the highest attention weights contribute more to the accuracy of the model output results, which enables the model to concentrate resources on the most critical parts. In addition, the attention weights can connect elements that are far apart in the time series, thereby helping the model learn long-distance dependencies and realize long-time series forecasts of typhoons and storm surges. For the input vector and the output vector , the innovative proposal of the present invention, the process is expressed as: ; ; In the formula, is the input vector For the output vector The attention weight, The normalization function transforms an arbitrary set of real numbers into a probability distribution such that each weight is between 0 and 1 and the sum of all weights is equal to 1. is the query vector, is the key vector, is the key vector The dimension represents a single vector The size of the feature dimension; is the output vector; is the input vector, representing the query vector With key vector The similarity between After normalization of the function, we get: is a value vector used for weighted summation to generate the final output; Finally, in order to increase the representation power and stability of the model, the self-attention mechanism usually uses multiple attention heads (i.e., multiple sets of independent query, key, and value weight matrices) and concatenates the outputs of each head together. Finally, by multiplying by another learnable weight matrix In the present invention, since the regional marine environment data is complex and the data volume is huge, the introduction of multi-head self-attention can process multiple subtasks in parallel, thereby improving the efficiency of the model and the generalization ability of the model.

[0033] ; In the formula, is the overall output of the multi-head self-attention mechanism, through multiple attention heads After the parallel calculations are completed, the results are stitched together and passed through a learned weight matrix Perform linear transformation to obtain the final output; The three matrices generated after the linear transformation of the input feature vector are the core inputs of the attention mechanism and are defined as follows: Query, which indicates the target vector or query vector to be calculated for attention; is the key, which indicates the source vector to be matched or compared; is a value, which represents the value vector that the input data actually needs to obtain after matching is completed; Will The output of an attention head Splice along the feature dimension, ,in, is the number of heads, The dimensions of each head output, is the set of real numbers, containing all real values.

[0034] In general, the self-attention mechanism dynamically calculates the attention weights between different positions, allowing the model to flexibly adjust the representation of each position according to the importance of different positions in the input sequence, thereby better capturing the long-range dependencies and important information in the sequence.

[0035] Exemplarily, the present invention proposes a new architecture VIT-TSS based on the VIT framework, which integrates attention mechanism and convolution operation for data encoding and decoding. First, input data, including typhoon path, sea surface height, wind speed, etc., pre-process the data, and use the convolution layer of the model to extract the features of the image block, and effectively capture the local information in the image through the sliding operation of the convolution kernel. Subsequently, the sequential information of the elements in the data is introduced through position encoding to enhance the model's understanding of the time series. Since the marine environment data input to the model is relatively complex, the model may learn noise or unimportant features, so regularization technology is used to prevent the model from overfitting.

[0036] Secondly, the regularized data is input into the VIT-TSS model encoding module. These features are further processed and encoded for VIT-TSS model analysis. To further stabilize the training process, normalization techniques are applied to the encoding block to standardize the input of the layer, thereby speeding up learning and improving the stability of the model. The multi-head attention mechanism, as a key component of the encoding block, enhances the ability of the VIT-TSS model to capture complex dependencies by allowing the VIT-TSS model to process information in multiple subspaces in parallel. After the multi-head attention process, the regularization and normalization steps are applied again to maintain the stability of the network training and help the VIT-TSS model learn more robust feature representations.

[0037] Exemplarily, the normalized data is input into a multi-layer perceptron block, which consists of multiple fully connected layers and introduces the nonlinear ability of the network through nonlinear activation functions to capture the complex relationships in the input data. In order to improve the generalization ability of the VIT-TSS model and avoid overfitting, regularization techniques are applied after the MLP, and then multiple linear layers further extract and transform the data, and the regularization step applied again ensures the stability of the VIT-TSS model during the training process.

[0038] Finally, the processed data is reshaped into the format required for forecasting typhoon path information and storm surge height, and further transposed convolution is performed to adjust the size of the feature map and output.

[0039] In general, the design of the VIT-TSS model structure not only enhances the VIT-TSS model's ability to learn the relationship between input features, but also helps the network to show better robustness when facing new data. In the typhoon and storm surge intelligent forecasting system, the encoding block can process meteorological data such as typhoon path, sea surface height and wind speed, and convert these data into useful feature representations to support the prediction tasks of the VIT-TSS model, thereby improving the accuracy and efficiency of typhoon and storm surge forecasts.

[0040] In the VIT-TSS model structure and data forward propagation process, the VIT-TSS model performs convolution, regularization, normalization and other processing on the input wind speed, sea surface height, typhoon path and other information, and finally outputs the typhoon path, sea surface height and other information at the future time.

[0041] Specifically, the input wind speed, sea surface height, typhoon path and other information are subjected to two-dimensional convolution, position encoding, regularization, encoding module setting, normalization, reshaping, and transposed convolution to finally output the typhoon path, sea surface height and other information at the future moment;

[0042] Exemplarily, the number of encoder block layers, the number of attention heads, the number of convolution kernels of the embedding layer, the batch size of training samples, the number of iterations, etc. of the encoding module in the VIT-TSS model are set. Among them, the typhoon forecast model uses wind speed, temperature and other data as the input samples of the forecast model, and the typhoon path data in the next 3, 12, and 24 hours as the output samples of the forecast model. The storm surge forecast model uses the meteorological elements and sea surface height data of the past 3 hours as the input samples of the forecast model, and the sea surface height data in the next 1 hour as the output samples of the forecast model.

[0043] Exemplarily, in the encoding module setting, from bottom to top, it is set as follows: first normalization layer, multi-head attention layer, first regularization layer, second normalization layer, multi-layer perceptron block, second regularization layer; In the multilayer perceptron block setting, from bottom to top, it is set as: 1st linear layer, 1st regularization layer, 2nd linear layer, 2nd regularization layer.

[0044] In order to avoid differences in the numerical ranges of various data, the original data is normalized before training. In order to effectively evaluate the pros and cons of the training and verification results in different modes, the root mean square error (RMSE) and mean relative error (MRE) are used to evaluate the effectiveness of the model according to the characteristics of typhoons and storm surges.

[0045] The root mean square error represents the error between the VIT-TSS model forecast and the actual value. The smaller the error, the more accurate the forecast result. The specific calculation formula is as follows: ; The average relative error is used to evaluate the ratio between the VIT-TSS model forecast value and the actual value. The smaller the ratio, the more accurate the forecast result. The specific calculation formula is as follows: ; The average relative error is used to evaluate the ratio between the VIT-TSS model forecast value and the actual value. The smaller the ratio, the more accurate the forecast result. The specific calculation formula is as follows: ; It can be seen from the above embodiments that the typhoon and storm surge forecasting method based on the integrated deep learning model of the present invention can accurately predict the typhoon route in the next 3, 12, and 24 hours according to the typhoon center longitude and latitude information, wind speed, temperature, potential height and other data, and reflect the typhoon path situation; and at the same time, accurately predict the sea surface height in the next hour to reflect the storm surge water situation. Compared with other current intelligent forecasting methods, the model proposed in the present invention fully considers the spatiotemporal variation characteristics of the marine environment, and can well predict the typhoon path and sea surface height changes at different lead times and different locations, and effectively improves the prediction accuracy. The present invention can provide more accurate and reliable data support for the prediction of extreme marine events, thereby helping to take corresponding measures for extreme marine events as soon as possible and minimize casualties and property losses caused by disasters. At the same time, it also provides valuable information for emergency management and navigation safety, helps to organize timely rescue operations, ensure that ships safely avoid dangerous areas, and thus reduce the adverse social and economic effects when natural disasters occur.

[0046] Example 2, as another implementation of the present invention, the lightweight rapid forecasting method for typhoons and storm surges provided in the embodiment of the present invention includes: (1) Constructing a typhoon dataset. First, preprocess the data and unify the temporal and spatial resolution of the input data through interpolation. The model input is gridded meteorological data such as wind speed, temperature, geopotential height, and the longitude and latitude information of the typhoon center.

[0047] (2) Construct a storm surge dataset.

[0048] The model inputs are sea surface height data, typhoon path information and gridded meteorological data. In order to fully explore the temporal and spatial correlation characteristics of storm surge and improve the forecast accuracy of storm surge, the present invention also considers the impact of typhoon on storm surge.

[0049] (3) Establish a deep learning model.

[0050] The present invention proposes a lightweight typhoon and storm surge rapid forecasting model based on the Transformer deep learning method. The Multi-head Attention mechanism of Vision Transformer (VIT) is used to extract the data features of wind and wave fields, and Position Encoding is used to record the position information of local areas, so that the model has good applicability in typhoon path forecasting and sea surface height forecasting.

[0051] (4) Conduct model sensitivity analysis.

[0052] The influence of multi-time comparison and influencing factors on the model forecast accuracy is studied separately. In the typhoon path forecast, the multi-time comparison selects the typhoon path information for the next 3, 12, and 24 hours. The influencing factors finally select gridded meteorological data such as wind speed, temperature, and potential height, and the longitude and latitude information of the typhoon center as the input variables of the model. In the storm surge forecast, the multi-time comparison finally selects the input data of the previous 3 hours to forecast the sea surface height for the next 1 hour. The influencing factors finally select the sea surface height data and typhoon path information as the input variables of the model.

[0053] (5) Analyze the calculation accuracy and efficiency of the model.

[0054] The model output results are compared with open source data and measured data to verify the forecast accuracy, and the forecast time of the intelligent model is compared with the traditional numerical model method to verify the computational efficiency of the model.

[0055] Exemplarily, in step (1), a typhoon dataset is constructed using the historical longitude and latitude information of continuous typhoon centers from 1959 to 2014 and environmental information around the typhoon centers, with a spatial resolution of 0.1°×0.1°.

[0056] Exemplarily, in step (2), a storm surge data set, sea surface height data and gridded meteorological data are constructed. The sea surface height data includes two sets of different spatial resolutions. One set of coarse grid data sets includes hourly average sea surface height data with a spatial resolution of 0.5°×0.5°. The other set of high-precision data sets includes hourly average sea surface height data with a spatial resolution of 1 / 12°×1 / 12°.

[0057] Exemplarily, in step (2), the meteorological data includes hourly average wind speed and air pressure data with a spatial resolution of 0.25°×0.25°.

[0058] Exemplarily, in step (2), the storm surge forecast process uses a sliding window mode to achieve rolling forecasts of sea surface height under different forecast periods, sliding along the time axis to the end of the 48th hour of the single storm surge process data. The input variables are the previous meteorological data and sea surface height characteristics. The storm surge forecast model simulates the nonlinear relationship between meteorological data and sea surface height, and finally predicts the change process of sea surface height during the storm surge.

[0059] For example, in step (3), the core mechanism of Transformer is the Self-Attention mechanism. However, a single Self-Attention can only capture part of the features of the input data. In order to obtain more comprehensive features of the input data, Transformer adopts a Multi-head Self-Attention mechanism with multiple Self-Attentions. When using the Multi-head Self-Attention mechanism, the input vector is divided into Parts ( is the number of attention heads), and then the calculation results of each attention head are concatenated, and finally a trainable weight matrix is ​​introduced Multiply with the concatenated matrix.

[0060] Further, in the step (3), the Muti-head Attention mechanism is used to extract the data features of the wind field and the wave field, and Position Encoding is used to record the position information of the local area. At the same time, in order to make the Vision Transformer model more suitable for the regional sea surface height forecasting task, the present invention removes the Class Token in the VIT model, and changes the output end of the model to the Upsampling module, and restores the matrix to the same dimension as the target matrix by deconvolution, so as to realize the forward transmission of the model data and the reverse propagation of the error.

[0061] In step (4), a multi-moment comparative analysis of storm surge forecast is performed. The multi-moment comparative analysis study includes two parts: the previous moment comparative analysis and the forecast moment comparative analysis. In the previous moment comparative test, the accuracy of the forecast results is compared by changing different previous input hours. In the forecast moment comparative test, the accuracy of the forecast results is compared by changing different forecast future hours, so as to select the best moment input and output model.

[0062] In step (4), the influencing factors of storm surge forecast are analyzed by inputting sea surface height, sea surface height and wind speed, sea surface height and air pressure, and sea surface height, wind speed and air pressure. The four groups of input factors are subjected to control experiments respectively, and the one with the smallest error is selected as the final input of the model.

[0063] In step (5), the computational accuracy and efficiency of the model are analyzed. By comparing the simulated data with the results of the measured data in the same time period, the minimum prediction error of the model is obtained. The running time of the model is recorded and compared with the traditional numerical model to verify the superiority of the intelligent model in computational efficiency.

[0064] For example, in order to avoid differences in the numerical ranges of various data, the original data is normalized before training. In order to effectively evaluate the pros and cons of the training and verification results in different modes, the root mean square error and mean relative error are used to evaluate the effectiveness of the model according to the characteristics of the storm surge of the research object. The first indicator is the root mean square error (RMSE) between the model predicted sea surface height and the actual sea surface height: ; The second indicator is the mean relative error (MRE) between the model predicted sea surface height and the actual sea surface height: ; Exemplarily, the present invention develops models based on the Pytorch framework.

[0065] Embodiment 3, as Figure 3 As shown, a typhoon and storm surge lightweight rapid forecasting system, the system comprises: Data preprocessing module 1, used to preprocess multi-source ocean environment data sets and typhoon movement paths; Typhoon and storm surge dataset construction module 2, used to construct typhoon and storm surge dataset based on preprocessed data; Deep learning model building module 3 is used to build a VIT-based deep learning model VIT-TSS based on the constructed typhoon and storm surge dataset; Verification of forecast accuracy module 4 is used to compare the output results of the deep learning model VIT-TSS with open source data and measured data to verify the forecast accuracy.

[0066] To further illustrate the effects of the embodiments of the present invention, the following experiment was conducted.

[0067] Typhoon No. 1909 "Li Mouma", which landed around August 10, 2019, stayed on land for up to 60 hours, with strong winds and maximum gusts of 17. The storm surge caused by Typhoon Li Mouma has caused huge economic losses to the coast of the East China Sea and is a typical typhoon storm surge event in recent years. As a specific embodiment, for the storm surge data file of Typhoon Li Mouma, a time series training set, test set and verification set are constructed. The typhoon moving direction, moving speed, longitude and latitude of the typhoon center, wind speed, air pressure, and sea surface height are a set of input parameters. If the current time is t, the longitude and latitude of the typhoon center at t+3, t+12, and t+24 hours and the storm surge water increase at t+1 hour are the corresponding output data.

[0068] The present invention uses the VIT-TSS model to analyze the typhoon Limouma's path and storm surge water increase forecast and measured results, and the results include: comparison of the typhoon "Limouma" path forecast 3h / 12h / 24h in advance with the measured results, and comparison of the storm surge process forecast caused by the typhoon "Limouma" with the measured results. The results show that the path forecast of the typhoon "Limouma" is relatively accurate, and the path forecast results 3h / 12h / 24h in advance are not much different from the actual path of the typhoon. At the same time, the storm surge process forecast caused by the typhoon is relatively accurate, and the sea surface height forecast results are not much different from the actual results.

[0069] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A lightweight and rapid forecasting method for typhoons and storm surges, characterized in that: The method comprises the following steps: S1, preprocessing of multi-source ocean environment data sets and typhoon movement paths; S2, construct typhoon and storm surge dataset based on preprocessed data; S3, based on the constructed typhoon and storm surge dataset, establish the VIT-TSS deep learning model based on VIT to achieve the joint forecast of typhoon path and storm surge water level changes; S4, the output results of the deep learning model VIT-TSS are compared with open source data and measured data to verify the forecast accuracy.

2. The lightweight rapid forecasting method for typhoons and storm surges according to claim 1 is characterized in that: In step S1, the multi-source ocean environment data set and typhoon movement path are preprocessed, including: unifying the temporal and spatial resolution of the input data through interpolation processing; The input multi-source marine environment data sets include: gridded meteorological data of wind speed, temperature, and geopotential height; The input typhoon moving path includes: the longitude and latitude information of the typhoon center.

3. The lightweight rapid forecasting method for typhoons and storm surges according to claim 1 is characterized in that: In step S2, a typhoon data set is constructed, including: historical longitude and latitude information of typhoon centers over many years and environmental information around the typhoon centers; Construct a storm surge dataset, including sea surface height data and gridded meteorological data. The sea surface height data contains two sets of different spatial resolutions. One set of coarse grid datasets contains hourly average sea surface height data with a spatial resolution of 0.5°×0.5°; the other set of high-precision datasets contains hourly average sea surface height data with a spatial resolution of 1 / 12°×1 / 12°. The meteorological data include hourly average wind speed and air pressure data with a spatial resolution of 0.25°×0.25°.

4. The lightweight rapid forecasting method for typhoons and storm surges according to claim 1 is characterized in that: In step S3, a VIT-TSS deep learning model based on VIT is established, including: (1) Using the self-attention mechanism VIT to dynamically calculate the attention weights between different positions, adjust the representation of each position according to the importance of different positions in the input sequence, and capture the long-range dependencies and important information in the sequence; (2) Based on the self-attention mechanism VIT, a deep learning model VIT-TSS is built to enhance the VIT model's ability to learn the relationships between input features. The deep learning model VIT-TSS is applied to typhoon and storm surge intelligent forecasting. The encoding block of the deep learning model VIT-TSS is used to process typhoon path, sea surface height and wind speed meteorological data, and the above data are converted into useful feature representations.

5. The lightweight rapid forecasting method for typhoons and storm surges according to claim 4 is characterized in that: Step (1) specifically includes: (1.1) Perform input transformation, input vector After three trainable weight matrices Transformation, respectively obtain the query vector query, key vector key, value vector value; the above vectors are used to calculate the attention weight and the final output; (1.2) Calculate the attention weight; for each query vector, use the dot product calculation method to calculate the similarity with all key vectors, through The function converts these similarities into attention weights; (1.3) The outputs of each of the multiple attention heads are concatenated together by multiplying them by another learnable weight matrix , and get the final output.

6. The lightweight rapid forecasting method for typhoons and storm surges according to claim 5 is characterized in that: In step (1.1), the input data of typhoon path and sea surface height are transformed to learn the data in different ways. The process is expressed as: ； In the formula, is the query vector, is the key vector, is a value vector.

7. The lightweight rapid forecasting method for typhoons and storm surges according to claim 5, characterized in that: In step (1.2), to prevent The gradient disappears, Before converting, divide the value by the vector Dimensions The square root of and the output vector , the process is expressed as: ； ； In the formula, is the input vector For the output vector The attention weight, is a normalization function that converts a set of arbitrary real numbers into a probability distribution so that each weight is between 0 and 1 and the sum of all weights is equal to 1; is the query vector, is the key vector, is the key vector The dimension represents a single vector The size of the feature dimension; is the output vector; is the input vector, representing the query vector With key vector The similarity between After normalization of the function, we get: is a value vector used for weighted summation to generate the final output; In step (1.3), the final output is: ； In the formula, is the overall output of the multi-head self-attention mechanism, through multiple attention heads After the parallel calculations are completed, the results are stitched together and passed through a learned weight matrix Perform linear transformation to obtain the final output; The three matrices generated after the linear transformation of the input feature vector are the core inputs of the attention mechanism and are defined as follows: Query, which indicates the target vector or query vector to be calculated for attention; is the key, which indicates the source vector to be matched or compared; is a value, which represents the value vector that the input data actually needs to obtain after matching is completed; Will The output of an attention head Splice along the feature dimension, ,in, is the number of heads, The dimensions of each head output, is the set of real numbers, containing all real values.

8. The lightweight rapid forecasting method for typhoons and storm surges according to claim 4 is characterized in that: In step (2), building the deep learning model VIT-TSS includes: (2.1) The attention mechanism and convolution operation are integrated for data encoding and decoding. First, the input data, including typhoon path, sea surface height, and wind speed, are preprocessed. The convolution layer of the model is used to extract the features of the image blocks. The local information in the image is captured through the sliding operation of the convolution kernel. Subsequently, the sequential information of the elements in the data is introduced through position encoding to enhance the understanding of the time series. (2.2) The regularized data is fed into the VIT-TSS model encoding module to further process and encode the features. After multi-head attention processing, the regularization and normalization steps are applied again to maintain the stability of network training and help learn robust feature representations. (2.3) The processed data is reshaped into the format required for forecasting typhoon path information and storm surge height, and further transposed convolution is performed to adjust the size of the feature map and output it.

9. The lightweight rapid forecasting method for typhoons and storm surges according to claim 1, characterized in that: In step S4, before training, the original data is normalized, and according to the characteristics of typhoons and storm surges, the root mean square error (RMSE) and mean relative error (MRE) are used to evaluate the effectiveness of the VIT-TSS model.

10. A lightweight rapid forecasting system for typhoons and storm surges, characterized in that: The system implements the typhoon and storm surge lightweight rapid forecasting method as claimed in any one of claims 1 to 9, and the system comprises: A data preprocessing module (1) is used to preprocess multi-source ocean environment data sets and typhoon movement paths; A typhoon and storm surge dataset construction module (2) is used to construct a typhoon and storm surge dataset based on the preprocessed data; A deep learning model building module (3) is used to build a VIT-based deep learning model VIT-TSS based on the constructed typhoon and storm surge dataset, so as to realize the joint forecast of typhoon path and storm surge water level change; The forecast accuracy verification module (4) is used to compare the output results of the deep learning model VIT-TSS with open source data and measured data to verify the forecast accuracy.

Images