Extreme rainfall prediction method based on parameter adaptation
Through the extreme precipitation prediction method based on parameter adaptation, decomposing and fine-tuning the parameters of the spatiotemporal sequence prediction model, the problem of existing models performing poorly in extreme precipitation prediction is solved, and higher accuracy in extreme precipitation prediction is achieved.
Patent Information
- Application Number
- CN202510388683.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-31
- Publication Date
- 2025-06-17
AI Technical Summary
The existing spatiotemporal sequence prediction models often perform poorly when dealing with extreme precipitation predictions, because extreme precipitation events are rare, resulting in the model being affected by conventional precipitation samples during training and ignoring the characteristics of extreme precipitation.
By designing an extreme precipitation prediction method based on parameter adaptation, the model parameters are decomposed into conventional and extreme precipitation subspaces using singular value decomposition, and the extreme precipitation subspace parameters are fine-tuned using a reweighted loss function to optimize the extreme precipitation prediction performance.
This method can eliminate the performance deviation of spatiotemporal sequence prediction models on conventional and extreme precipitation prediction, greatly improve the accuracy of extreme precipitation prediction, while retaining the performance of conventional precipitation prediction.
Smart Images
Figure CN120161545A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of precipitation prediction, and specifically, relates to an extreme precipitation prediction method based on parameter adaptation. Background Art
[0002] Extreme precipitation (such as heavy rain, extremely heavy rain, etc.) is usually accompanied by a series of serious natural disasters, such as floods and landslides, which have a significant impact on daily life. Therefore, it is crucial to improve the accuracy of extreme precipitation prediction. Currently, there are very few studies directly targeting extreme precipitation prediction. Extreme precipitation prediction still relies on ordinary precipitation prediction problems and adopts the same research methods, appearing as a part in the results of ordinary precipitation prediction.
[0003] In the problem of precipitation prediction, radar images are important tools in meteorology for monitoring and predicting precipitation. Radar detects precipitation particles in the atmosphere by emitting electromagnetic waves and receiving their reflected signals, and the echo intensity can reflect the intensity of precipitation. Therefore, the precipitation prediction problem can extrapolate the radar echo sequence in the future for a certain period through the historical radar echo sequence, and then convert the predicted radar echo value into rainfall intensity through the Z-R relationship, thereby realizing precipitation prediction. Therefore, the precipitation prediction problem essentially belongs to the spatio-temporal sequence prediction problem.
[0004] Currently, there are various methods for realizing precipitation prediction by extrapolating radar images. Most traditional methods rely on numerical weather prediction (NWP). These methods are based on aerodynamics, identify the atmospheric quantities related to rainfall, and establish numerical equations to predict future precipitation. However, the drawback of this method is that the typical characteristics of atmospheric motion are complexity and variability, so it is extremely difficult to solve these numerical equations and requires a large amount of computational cost. In recent years, with the rise of deep learning, more and more studies use neural network models to extrapolate radar echo images. Such models are called spatio-temporal sequence prediction models in the field of deep learning. For example, DGMR is a deep generative model that learns the probability distribution of radar data and simulates future radar images. SmaAt-UNet is a U-Net architecture model configured with an attention module and depthwise separable convolution, which is used to capture the key features of radar images while reducing the number of parameters to improve the prediction speed of the model. The development of these models shows that deep learning methods not only help to model the spatio-temporal multi-scale features of the nonlinear precipitation process, but also significantly reduce the computational time and resource requirements. Therefore, deep learning methods have gradually become one of the mainstream solutions for nowcasting of precipitation.
[0005] However, common deep learning methods often perform poorly when dealing with extreme precipitation prediction problems. This is because extreme precipitation events are relatively rare in nature, resulting in fewer relevant samples in the radar echo dataset, which is a typical long-tail distribution problem. Existing spatiotemporal series prediction models (such as ConvGRU, PhyDNet, SimVP, etc.) are significantly affected by the large number of samples during training. Therefore, they tend to fit the distribution of conventional precipitation and ignore the characteristics of extreme precipitation. Therefore, in the prediction results, it can be clearly found that these models are significantly less effective than ordinary precipitation prediction problems when facing extreme precipitation prediction problems. If only extreme precipitation samples are used as training data, the small number of samples will lead to underfitting of the model and low performance in extreme precipitation prediction. Summary of the invention
[0006] In view of the problems existing in the prior art, the present invention proposes an extreme precipitation prediction method based on parameter adaptation. Through the parameter adaptation framework designed by the present invention, the performance deviation of the spatiotemporal series prediction model in the prediction of conventional precipitation (such as drizzle, light rain, moderate rain) and extreme precipitation can be eliminated, and the accuracy of extreme precipitation prediction can be greatly improved. At the same time, the framework is model-independent and can be integrated into most existing spatiotemporal series prediction models.
[0007] The present invention is achieved through the following technical solutions:
[0008] The extreme precipitation prediction method based on parameter adaptation specifically comprises the following steps:
[0009] Step 1: Train the spatiotemporal series prediction model on the precipitation dataset to obtain mixed features including regular precipitation and extreme precipitation;
[0010] Step 2: split the spatiotemporal series prediction model of step 1 and find the model structure that contributes most to extreme precipitation prediction;
[0011] Step 3, decompose the parameters of the structure selected in step 2 into normal and extreme precipitation subspaces using singular value decomposition;
[0012] Step 4: Use the reweighted loss function to fine-tune the extreme precipitation subspace parameters to optimize the extreme precipitation prediction performance while retaining the conventional precipitation prediction capability.
[0013] Furthermore, in step 1,
[0014] The precipitation data set is radar echo data including normal precipitation and extreme precipitation.
[0015] Further, step 2 includes:
[0016] Step 2.1, train a spatio-temporal sequence prediction model on the dataset containing all precipitation samples, and debug the model until the overall precipitation prediction effect is optimal;
[0017] Step 2.2, retain the knowledge related to regular precipitation, conduct experiments on each structure in the model one by one, and freeze the remaining structures;
[0018] Step 2.3, observe the changes in the model's performance in extreme precipitation prediction in different experiments, and screen out the model structure that contributes the most to the improvement of extreme precipitation performance.
[0019] Furthermore, in Step 3,
[0020] Step 3.1, fold the parameter matrix and perform singular value decomposition: Let be the parameters of a convolutional layer, with C o output channels, C i input channels, and the convolutional kernel size is C k ×C k ;
[0021] First, fold W into a matrix W' of size , and decompose it through SVD into:
[0022] W' = USV T
[0023] where S ∈ R R×R , and the rank According to the mathematical meaning of SVD, the singular matrices U and V contain the eigeninformation of W'; S is an importance matrix, and the values on its diagonal represent the importance of the singular vectors in U and V, and these values gradually decrease;
[0024] Step 3.2, subspace partitioning: Divide U, S, and V into two groups, namely (U1, S1, V1) and (U2, S2, V2),
[0025] where (U1, S1, V1) represents the regular subspace, retaining the regular precipitation features; (U2, S2, V2) represents the extreme subspace, corresponding to the extreme precipitation features.
[0026] Furthermore, in Step 4,
[0027] Step 4.1, freeze U1, S1, V1, and S2, and only fine-tune the features in the extreme precipitation parameter subspace, that is, fine-tune U2 and V2;
[0028] Step 4.2, use the reweighted loss (BMSE) and the misclassification loss to form the class reweighted loss;
[0029] Among them, BMSE assigns weights according to precipitation intensity, with higher weights for extreme precipitation;
[0030] The misclassification loss increases the penalty for misclassifying normal precipitation as extreme precipitation.
[0031] An extreme precipitation prediction system based on parameter adaptation:
[0032] The prediction system includes a pre-training module, a model splitting module, a model decomposition module, and a parameter adjustment module:
[0033] The pre-training module trains a spatio-temporal sequence prediction model on a precipitation dataset to obtain mixed features including normal precipitation and extreme precipitation;
[0034] The model splitting module splits the spatio-temporal sequence prediction model of the pre-training module to find the model structure that contributes the most to extreme precipitation prediction;
[0035] The model decomposition module uses singular value decomposition to decompose the parameters of the selected structure of the model splitting module into normal and extreme precipitation subspaces;
[0036] The parameter adjustment module uses a reweighted loss function to fine-tune the parameters of the extreme precipitation subspace, optimize the extreme precipitation prediction performance, and at the same time retain the normal precipitation prediction ability.
[0037] An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.
[0038] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.
[0039] Advantages of the present invention
[0040] The present invention can eliminate the performance bias of the spatio-temporal sequence prediction model in predicting normal precipitation (such as drizzle, light rain, and moderate rain) and extreme precipitation, and can greatly improve the extreme precipitation prediction performance on the premise of well retaining the prediction performance of normal precipitation.
[0041] The parameter freezing scheme of the present invention can not only avoid potential damage to the majority class information that may be caused when fine-tuning all structures, thus maintaining the performance of normal precipitation; but also enable the parameters sensitive to extreme precipitation to learn more features of extreme precipitation samples, greatly improving the extreme precipitation prediction effect.
[0042] The present invention only needs to fine-tune some parameters, greatly accelerating the model optimization process. At the same time, the framework of the present invention is model-independent and can be integrated into most existing spatio-temporal sequence prediction models. Description of the drawings
[0043] Figure 1 This is a schematic diagram of the overall framework of the method of the present invention.
[0044] Figure 2 These are schematic diagrams of each structure in the individual fine-tuning model.
[0045] Figure 3 This is the decoupling of the parameter space.
[0046] Figure 4 This is a schematic diagram of the structure-guided parameter adaptation method of the present invention. Specific implementation manners
[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0048] The experimental methods used in the following embodiments are all conventional methods unless otherwise specified. The materials, reagents, methods, and instruments used, unless otherwise specified, are all conventional materials, reagents, methods, and instruments in the art, and those skilled in the art can obtain them through commercial channels.
[0049] Combined with Figures 1 to 4 , the extreme precipitation prediction method based on parameter adaptation of the present invention;
[0050] The method consists of two training stages, namely the pre-training stage and the fine-tuning stage;
[0051] In the pre-training stage, the spatio-temporal sequence precipitation prediction model, such as ConvGRU, PhyDNet, SimVP, etc., is still trained on the entire precipitation data set in the original training manner. The purpose of this step is to learn the implicit features of precipitation of various intensities. Compared with directly training the model on a small number of extreme precipitation samples, more general features of precipitation can be obtained in this way.
[0052] The precipitation data set includes radar echo data of conventional precipitation (such as drizzle, light rain, moderate rain) and extreme precipitation;
[0053] The fine-tuning stage is the core stage, aiming to eliminate the performance deviation between ordinary precipitation prediction and extreme precipitation prediction in a parameter-adaptive manner. In the previous pre-training stage, the parameters of the model will learn the features of these two different types of samples, namely conventional precipitation and extreme precipitation.
[0054] The main idea of the fine-tuning stage is as follows: First, decouple the model parameters obtained in the pre-training stage into two parts, namely the specific parameters for normal precipitation and the specific parameters for extreme precipitation. Second, freeze the parameter part corresponding to normal precipitation and fine-tune the specific parameters corresponding to extreme precipitation to eliminate the performance bias in extreme precipitation.
[0055] The advantage of this is that it can improve the prediction performance of extreme precipitation on the premise of well retaining the prediction performance of normal precipitation.
[0056] As Figure 1 shown: The structure-guided parameter adaptation method in the fine-tuning stage includes:
[0057] (1) Find the model structure that contributes the most to the improvement of extreme precipitation performance;
[0058] First, it is necessary to split the structure of the spatio-temporal sequence prediction model to find the model structure that contributes the most to the improvement of extreme precipitation performance. There are mainly two architectures for spatio-temporal prediction models, namely the encoder-predictor framework (such as ConvGRU) and the encoder-translator-decoder framework (such as SimVP). Regardless of which framework, they all contain a structure for encoding images, a structure for predicting future feature vectors, and a structure for decoding features into images. Each structure learns information from samples of normal precipitation and extreme precipitation. To find the structure that has a greater impact on the prediction of extreme precipitation, the following three sub-steps are required:
[0059] i) Train the spatio-temporal sequence prediction model on the dataset containing all precipitation samples in the original way, and debug the model to the state where the overall effect of precipitation prediction is optimal.
[0060] ii) Conduct experiments on each structure in the model one by one: Freeze the remaining structures and fine-tune the current structure on extreme precipitation samples using a reweighted loss function, as Figure 2 shown. (Freezing the remaining structures can retain the knowledge related to normal precipitation)
[0061] iii) Observe the changes in the performance of the model in extreme precipitation prediction in different experiments, and find the model structure that contributes the most to the improvement of extreme precipitation performance.
[0062] (2) Decompose the parameter space in this structure into a normal precipitation parameter subspace and an extreme precipitation parameter subspace;
[0063] Second, it is to decompose the parameter space in this structure, as Figure 3As shown. When a model or a certain structure in the model is decoupled to the smallest granularity, it can be regarded as a set of parameters. Each parameter continuously learns the information contained in the normal precipitation and extreme precipitation samples through backpropagation. Therefore, each parameter space can be decoupled into a normal precipitation parameter subspace and an extreme precipitation parameter subspace. The present invention uses singular value decomposition (SVD) to decompose the parameter space.
[0064] Suppose is the parameter of a convolutional layer, with C o output channels, C i input channels, and the convolutional kernel size is C k ×C k . First, fold W into a matrix W’ of size , and decompose it by SVD as:
[0065] W’ = USV T
[0066] where S ∈ R R×R , and the rank According to the mathematical meaning of SVD, the singular matrices U and V contain the eigeninformation of W’. S is an importance matrix, and the values on its diagonal represent the importance of the singular vectors in U and V, and these values gradually decrease.
[0067] In the dataset used in the embodiment, most samples are normal precipitation, and only a small part is extreme precipitation. Therefore, the information related to normal precipitation in the pre-trained model dominates the parameter space, which means that this information should be preserved in the leading vectors of U and V. At the same time, the information of extreme precipitation is stored in the trailing vectors of U and V. Therefore, further divide U, S, V into two groups, namely (U1, S1, V1) and (U2, S2, V2), as Figure 4 shown. They represent the feature matrices of normal precipitation and extreme precipitation respectively.
[0068] The method of decomposing (U, S, V) is as follows: Suppose the information of extreme precipitation is located in the last K of the R singular values, then we will get:
[0069] S1 ∈ R (R-K)×(R-K) ,
[0070]
[0071] S2 ∈ R K×K ,
[0072] In a specific implementation, the modules (U1, S1, V1) and (U2, S2, V2) each consist of three consecutive layers. For example, U2, S2, V2 are respectively: (1) a convolutional layer of size C o ×K×1×1, (2) a scaling layer of size K×K, (3) a convolutional layer of size K×C i ×C k ×C k ×C. The specific value of K varies depending on the model and needs to be determined through experiments.
[0073] (3) Fine-tune the extreme precipitation parameter subspace using a reweighted loss function.
[0074] In the last step, in order to increase the knowledge of extreme precipitation without affecting the accuracy of regular precipitation prediction, on the extreme precipitation sample dataset, use a reweighted loss function to fine-tune the features in the extreme precipitation parameter subspace, as Figure 4 shown. That is, fine-tune U2 and V2 while keeping U1, S1, V1, and S2 frozen. This can avoid potential damage to the majority class information that may occur when fine-tuning all structures, thus maintaining the performance of regular precipitation; and it can also enable the parameters sensitive to extreme precipitation to learn more features of extreme precipitation samples, significantly improving the extreme precipitation prediction effect. In addition, this method greatly speeds up the model optimization process by reducing the number of parameters that need to be fine-tuned.
[0075] Solve the problem of the decline in the prediction performance of regular precipitation through class reweighted loss;
[0076] Because even if the pixel values of radar echo values belonging to the regular precipitation range (0 - 15 dBZ) are mispredicted as moderate rain or even extreme precipitation, due to their small weight in BMSE, the loss value calculated by the model will be very small. On the contrary, even if the model has a slight error in predicting extreme precipitation values, due to the large weight assigned to them, the loss value will be large, resulting in the model paying too much attention to extreme precipitation. This will lead to a decline in the prediction performance of regular precipitation.
[0077] The class reweighted loss consists of a simple reweighted loss (BMSE) and a misclassification loss. Among them, BMSE is formulated based on the distribution of common rainfall datasets, that is:
[0078]
[0079] The misclassification loss is specifically formulated for the case of regular precipitation. For any pixel i whose true radar echo value belongs to the regular precipitation echo value, calculate the degree to which its predicted value is misclassified. If its predicted value is also within the regular precipitation range, the weight is set to 0. Otherwise, as the degree of misclassification deviation increases, the weight assigned to this pixel will also increase, specifically as follows:
[0080]
[0081] In summary, the BMSE in the class reweighted loss ensures that the model can capture more information about extreme precipitation, while the misclassification loss alleviates the problem of the decline in prediction performance caused by the too small weight of normal precipitation.
[0082] An extreme precipitation prediction system based on parameter adaptation:
[0083] The prediction system includes a pre-training module, a model splitting module, a model decomposition module, and a parameter adjustment module:
[0084] The pre-training module trains a spatio-temporal sequence prediction model on a precipitation data set to obtain mixed features including normal precipitation and extreme precipitation;
[0085] The model splitting module splits the spatio-temporal sequence prediction model of the pre-training module to find the model structure that contributes the most to extreme precipitation prediction;
[0086] The model decomposition module uses singular value decomposition to decompose the parameters of the structure selected by the model splitting module into normal and extreme precipitation subspaces;
[0087] The parameter adjustment module uses a reweighted loss function to finely tune the parameters of the extreme precipitation subspace, optimize the extreme precipitation prediction performance, and at the same time retain the normal precipitation prediction ability.
[0088] An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.
[0089] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.
[0090] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the method described in the present invention is intended to include but not limited to these and any other suitable types of memory.
[0091] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from a website, computer, server, or data center to another website, computer, server, or data center by wire, such as coaxial cable, fiber optic, digital subscriber line (DSL), or wirelessly, such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes one or more integrated available media. The available media can be magnetic media, such as floppy disks, hard disks, magnetic tapes, optical media, such as high-density digital video discs (DVDs), or semiconductor media, such as solid state discs (SSDs), etc.
[0092] In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as being executed by the hardware processor, or executed by the combination of the hardware and software modules in the processor. The software module can be located in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.
[0093] It should be noted that the processor in the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiments can be completed by the integrated logic circuit in the hardware of the processor or instructions in the form of software. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor, or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art 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, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.
[0094] The above has introduced in detail the method for extreme precipitation prediction based on parameter adaptation proposed by the present invention, and elaborated on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.
Claims
1. The extreme precipitation prediction method based on parameter adaptation is characterized by: The method specifically comprises the following steps: Step 1: Train the spatiotemporal series prediction model on the precipitation dataset to obtain mixed features including regular precipitation and extreme precipitation; Step 2: split the spatiotemporal series prediction model of step 1 and find the model structure that contributes most to extreme precipitation prediction; Step 3, decompose the parameters of the structure selected in step 2 into normal and extreme precipitation subspaces using singular value decomposition; Step 4: Use the reweighted loss function to fine-tune the extreme precipitation subspace parameters to optimize the extreme precipitation prediction performance while retaining the conventional precipitation prediction capability.
2. The extreme precipitation prediction method according to claim 1, characterized in that: In step 1, The precipitation data set is radar echo data including normal precipitation and extreme precipitation.
3. The extreme precipitation prediction method according to claim 2, characterized in that: In step 2 include, Step 2.1, train the spatiotemporal series prediction model on the data set containing all precipitation samples, and debug the model until the precipitation prediction presents the best overall effect; Step 2.2, retain the knowledge related to conventional precipitation, conduct experiments on each structure in the model one by one, and freeze the remaining structures; Step 2.3: Observe the changes in the model's performance in extreme precipitation prediction in different experiments, and select the model structure that contributes most to the improvement of extreme precipitation performance.
4. The extreme precipitation prediction method according to claim 3, characterized in that: In step 3, Step 3.1, fold the parameter matrix and perform singular value decomposition: Let is a parameter of a convolutional layer with C o Output channels, C i input channels, and the convolution kernel size is C k ×C k ; First, fold W into a The matrix W' is decomposed into: In'=USV T in S∈R R×R , And order According to the mathematical meaning of SVD, the singular matrices U and V contain the characteristic information of W'; S is an importance matrix whose diagonal values represent the importance of singular vectors in U and V, and these values gradually decrease; Step 3.2, subspace partitioning: divide U, S, and V into two groups, namely (U1, S1, V1) and (U2, S2, V2). Among them, (U1, S1, V1) represents the conventional subspace, which retains the conventional precipitation characteristics; (U2, S2, V2) represents the extreme subspace, corresponding to the extreme precipitation characteristics.
5. The extreme precipitation prediction method according to claim 4, characterized in that: In step 4, Step 4.1, freeze U1, S1, V1 and S2, and only fine-tune the features in the extreme precipitation parameter subspace, that is, fine-tune U2 and V2; Step 4.2, use the reweighted loss (BMSE) and misclassification loss to form the category reweighted loss; Among them, BMSE assigns weights according to precipitation intensity, with extreme precipitation having a higher weight; The misclassification loss adds a penalty for cases where regular precipitation is misclassified as extreme precipitation.
6. A prediction system for executing the extreme precipitation prediction method based on parameter adaptation according to any one of claims 1 to 5, characterized in that: The prediction system includes a pre-training module, a model splitting module, a model decomposition module and a parameter adjustment module: The pre-training module trains a spatiotemporal series prediction model on a precipitation dataset to obtain mixed features including regular precipitation and extreme precipitation; The model splitting module splits the spatiotemporal sequence prediction model of the pre-training module to find the model structure that contributes most to extreme precipitation prediction; The model decomposition module decomposes the parameters of the structure selected by the model splitting module into normal and extreme precipitation subspaces using singular value decomposition; The parameter adjustment module uses a reweighted loss function to fine-tune the extreme precipitation subspace parameters to optimize the extreme precipitation prediction performance while retaining the conventional precipitation prediction capability.
7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.
8. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.