Mode rainfall forecast correction method based on deep learning and wind cloud satellite

Abstract

The invention relates to the technical field of meteorological satellite observation, and discloses a mode rainfall forecast correction method based on deep learning and a wind cloud satellite, and the method comprises the steps: constructing a rainfall correction main model, collecting multi-source data, and carrying out the primary processing of the collected multi-source data, and obtaining a fusion input data set A; carrying out feature extraction and generating a satellite feature tensor B, generating a mode physical feature tensor C, generating a fusion feature tensor D based on inter-modal correlation, generating a corrected rainfall prediction tensor E, and generating a radar reflectivity prediction tensor F and an uncertainty estimation tensor G; and deploying the trained model into the system, and inputting a fusion data set A at the current time. By introducing the physical consistency loss and combining with the physical constraint to optimize the rainfall prediction model, strict conformity between the rainfall prediction result and the meteorological physical law is realized, the rainfall field better conforming to the physical reality is obtained, unreasonable fluctuations and errors in the prediction result are effectively reduced, and the physical consistency of rainfall prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological satellite observations, and specifically to a method for correcting model precipitation forecasts based on deep learning and Fengyun satellites. Background Art

[0002] Data-driven precipitation prediction models are widely used in the field of meteorological forecasting. Especially under the impetus of deep learning technology, certain achievements have been made in using remote sensing images, radar reflectivity, and numerical weather prediction products for precipitation regression prediction. These models can achieve relatively high prediction accuracy in specific scenarios and have good spatio-temporal expression capabilities. However, since most models mainly aim to minimize statistical errors and lack explicit constraints on meteorological physical laws, they often ignore basic physical principles such as energy conservation, water vapor balance, and spatial continuity that the precipitation process should follow.

[0003] This lack of physical consistency easily leads to problems in the model output results, such as precipitation exceeding the upper limit of the theoretical water vapor flux, obvious jumps in spatial distribution, or imbalance in energy conservation. As a result, the credibility and interpretability of precipitation prediction results are affected, and especially the stability and promotion in practical applications are severely restricted. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for correcting model precipitation forecasts based on deep learning and Fengyun satellites, which solves the problems of the lack of an effective modeling mechanism for meteorological physical constraints, the inability to ensure the physical rationality and consistency of prediction results, the possible deviation of precipitation prediction results from the evolution law of the actual weather system, and the reduction of the credibility and generalization ability of the model.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for correcting model precipitation forecasts based on deep learning and Fengyun satellites, including the following steps:

[0006] First, build a main precipitation correction model and collect multi-source data, and perform preliminary processing on the collected multi-source data to obtain a fused input data set A, which is combined with ground truth data for supervised training;

[0007] Then, input the data set A into multi-modal feature extraction to perform feature extraction and generate a satellite feature tensor B;

[0008] At the same time, input the numerical model variable field into a one-dimensional convolutional network or a multi-layer perceptron for pixel-by-pixel feature extraction to generate a model physical feature tensor C;

[0009] Subsequently, input the satellite feature tensor B and the model feature tensor C into guided modal fusion to generate a fused feature tensor D based on the correlation between modalities;

[0010] Input the fused feature tensor D into the main model through precipitation correction to generate a corrected precipitation prediction tensor E;

[0011] Meanwhile, input the fused feature tensor D into the multi-task auxiliary decoder to generate a radar reflectivity prediction tensor F and an uncertainty estimation tensor G;

[0012] Deploy the trained model to the system, input the fused data set A at the current time step, and output the future precipitation correction tensor E and its uncertainty tensor G, which are used to replace or correct the precipitation forecast results of the original numerical model.

[0013] Preferably, the multi-source data includes multi-source raw data of FY satellite remote sensing observation data, numerical weather prediction model output data, and ground truth precipitation data. The preliminary processing includes spatial interpolation and time registration processing to unify the data format with consistent grid resolution and time step;

[0014] The data set A includes:

[0015] A sequence of FY satellite remote sensing images, with a shape of T, H, W, C1; the sequence of FY satellite remote sensing images includes brightness temperature, cloud top height, cloud phase state, and water vapor path channels, with a time resolution of 5 to 15 minutes, where T is the length of the time series, H and W are the spatial grid dimensions, and C1 is the number of remote sensing channels. The number of remote sensing channels includes TBB, cloud height, and water vapor;

[0016] A numerical model variable field, with a shape of T, H, W, C2, where C2 is the number of meteorological variables. The number of meteorological variables includes precipitation, humidity, temperature, and wind vector.

[0017] Preferably, the feature extraction is used to apply 3D convolution or convolutional long short-term memory network to the sequence of FY satellite images to extract spatio-temporal evolution features;

[0018] The feature tensor B is represented as H, W, D1;

[0019] where D1 is the number of channels of the extracted deep semantic features of the satellite, representing the dynamic feature encoding of the satellite images at each spatial position in the time dimension.

[0020] Preferably, the feature tensor C is represented as H, W, D2;

[0021] where D2 is the high-dimensional representation of the extracted model variables, including the semantic compression of the temperature, humidity, wind field, and convective parameter information.

[0022] Preferably, the guided modal fusion is used to perform a cross-attention mechanism. The cross-attention mechanism in the modal fusion module is constructed based on a multi-head attention network. Taking the satellite feature tensor B as the Query, and the pattern feature tensor C as the Key and Value respectively, a guided feature selection and fusion mechanism is constructed to dynamically regulate the importance of different modal features;

[0023] The fused feature tensor D is H, W, D3;

[0024] The tensor D represents the deep semantic representation of each spatial pixel integrating multi-modal information, and is used to guide the precipitation correction output.

[0025] Preferably, the precipitation correction main model is used for multi-scale feature encoding and decoding operations;

[0026] The tensor E is H, W, 1, representing the pixel-by-pixel precipitation correction result for the future time step.

[0027] Preferably, the multi-task auxiliary decoder includes two parallel branch networks, which are respectively used to predict the radar reflectivity tensor F and the precipitation uncertainty tensor G, and jointly optimize the loss with the main task output tensor E during the training process;

[0028] The tensor F is H, W, 1, representing the reflectivity factor map;

[0029] The tensor G is H, W, 1, representing the precipitation uncertainty corresponding to each pixel. The precipitation uncertainty includes standard deviation, confidence interval or confidence score.

[0030] Preferably, during the training stage of the precipitation correction main model, a multi-task loss function is constructed based on the tensors E, F, G and the ground truth data, and the prediction error and physical consistency loss are jointly minimized. The physical constraints include water vapor mass conservation, energy balance, and smoothness of the spatial gradient of the precipitation field, and a total loss function is formed to optimize the model parameters.

[0031] Preferably, the physical consistency loss includes:

[0032] The water vapor conservation loss based on vertical integration, which is used to constrain the predicted precipitation not to exceed the water vapor flux;

[0033] The spatial gradient regularization term, which is used to improve the spatial continuity of precipitation prediction;

[0034] The moist static energy balance term, which is used to maintain the energy closure of the physical field.

[0035] Preferably, in the model training stage, a meta-learning algorithm based on task division is used to train and optimize the task set composed of different regions or weather types, and the initial parameters of the model with fast adaptation ability are obtained. The initial parameters are used to enable the model to complete effective migration through a small number of samples in a new region or new scenario.

[0036] The present invention provides a method for correcting model precipitation forecasts based on deep learning and Fengyun satellites. It has the following beneficial effects:

[0037] 1. By introducing physical consistency loss and optimizing the precipitation prediction model in combination with physical constraints, the present invention realizes the strict conformity of the precipitation prediction result with meteorological physical laws, obtains a precipitation field that is more in line with physical reality, effectively reduces the unreasonable fluctuations and errors in the prediction result, and improves the physical consistency of precipitation prediction.

[0038] 2. Through the meta-learning algorithm based on task division, the present invention trains and optimizes the task sets of different regions or weather types, and obtains the initial parameters of the model with fast adaptation ability, enabling the model to perform effective migration and adjustment through only a small number of samples in a new region or new scenario, thereby realizing a wider application scenario adaptability and significantly improving the generalization ability of the model.

[0039] 3. Through multi-modal feature fusion and multi-scale feature encoding and decoding networks, the present invention effectively fuses various information from remote sensing data and numerical weather forecasts, captures the spatial and temporal distribution characteristics of precipitation, realizes the accurate prediction of extreme weather events, improves the model's ability to capture complex meteorological phenomena, and enhances the model's prediction ability in extreme weather events.

[0040] 4. Through the multi-task auxiliary decoder, the present invention introduces the modeling of precipitation uncertainty scoring, enabling the model to not only output the precipitation prediction result but also quantify the uncertainty of the prediction. It provides a basis for credibility analysis and risk assessment of the precipitation prediction result, plays an important auxiliary role in meteorological forecast decision-making, and enhances the user's trust in the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flowchart of the method for correcting model precipitation forecasts based on deep learning and Fengyun satellites of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the drawings 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. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] Please refer to the appendix Figure 1 , the embodiment of the present invention provides a method for correcting model precipitation forecasts based on deep learning and Fengyun satellites, including the following steps:

[0044] First, build a precipitation correction main model and collect multi-source data, and perform preliminary processing on the collected multi-source data to obtain a fused input data set A, which is combined with ground truth data for supervised training;

[0045] Then, input the data set A into multi-modal feature extraction to perform feature extraction and generate a satellite feature tensor B;

[0046] At the same time, input the numerical model variable field into a one-dimensional convolutional network or a multi-layer perceptron for pixel-by-pixel feature extraction to generate a model physical feature tensor C;

[0047] Subsequently, input the satellite feature tensor B and the model feature tensor C into guided modal fusion to generate a fused feature tensor D based on the correlation between modalities;

[0048] Input the fused feature tensor D into the model through the precipitation correction main model to generate a corrected precipitation prediction tensor E;

[0049] At the same time, input the fused feature tensor D into a multi-task auxiliary decoder to generate a radar reflectivity prediction tensor F and an uncertainty estimation tensor G;

[0050] Deploy the trained model to the system, input the fused data set A at the current time step, and output the future precipitation correction tensor E and its uncertainty tensor G, which are used to replace or correct the precipitation forecast results of the original numerical model.

[0051] The multi-source data includes multi-source raw data of Fengyun satellite remote sensing observation data, numerical weather prediction model output data, and ground truth precipitation data. The preliminary processing includes spatial interpolation and time registration processing to unify the data format with consistent grid resolution and time step;

[0052] The data set A includes:

[0053] A sequence of Fengyun satellite remote sensing images, with a shape of T, H, W, C1; the sequence of Fengyun satellite remote sensing images includes brightness temperature, cloud top height, cloud phase state, and water vapor path channels, with a time resolution of 5 minutes to 15 minutes, where T is the length of the time series, H and W are the spatial grid dimensions, and C1 is the number of remote sensing channels. The number of remote sensing channels includes TBB, cloud height, and water vapor;

[0054] The numerical model variable field, with a shape of T, H, W, C2, where C2 is the number of meteorological variables. The number of meteorological variables includes precipitation, humidity, temperature, and wind vector.

[0055] Specifically, by constructing a deep neural network model that fuses Fengyun satellite remote sensing data and numerical weather prediction model variables, refined correction of numerical model precipitation forecasts is achieved, and an uncertainty assessment mechanism is introduced at the output end, which has a significant improvement compared to traditional correction methods such as ground regression and grid bias correction. By constructing a multi-modal feature extraction and fusion structure, the model can simultaneously perceive the cloud mass evolution information in satellite images and the meteorological variable evolution process in numerical models, capture more abundant spatio-temporal features, and overcome the problem of incomplete information caused by a single data source. The fused feature tensor contains continuous textures from remote sensing images and numerical gradients of physical variables at the spatial scale, and is mapped to a precipitation prediction map through a deep decoder, effectively correcting the precipitation misalignment and deviation phenomena in numerical model forecasts, and improving the spatial accuracy. By introducing an auxiliary output branch, the credibility of each spatial precipitation prediction is modeled, which helps the downstream decision-making system evaluate the risk grading and emergency response level of extreme precipitation events. During the training process, water vapor conservation, energy balance, and spatial continuity loss functions are introduced to ensure that the model prediction results not only pursue the minimization of errors but also satisfy the basic meteorological physical laws, enhancing the generalization ability and credibility.

[0056] Feature extraction is used to apply three-dimensional convolution or convolutional long short-term memory network to the Fengyun satellite image sequence to extract spatio-temporal evolution features;

[0057] The feature tensor B is represented as H, W, D1;

[0058] Where D1 is the number of channels of the satellite's deep semantic features extracted, representing the dynamic feature encoding of the satellite image at each spatial position in the time dimension.

[0059] Specifically, the feature tensor C is represented as H, W, D2;

[0060] Where D2 is the high-dimensional representation of the extracted model variables, including the semantic compression of the temperature, humidity, wind field, and convective parameter information.

[0061] Specifically, first, for the Fengyun satellite remote sensing image sequence, considering its natural dual dimensions of space (H×W) and time (T), in order to fully extract the dynamic process features such as cloud system evolution, water vapor movement, and radiation field change in satellite observations, the present invention uses a three-dimensional convolutional neural network (3DCNN) or a convolutional long short-term memory network (ConvLSTM) as the main extraction structure. 3DCNN introduces a time dimension to the convolutional kernel, which can uniformly model continuous image frames in a local area, retain the spatio-temporal coupling structure, and is suitable for static training data sets; while ConvLSTM introduces a cyclic structure and a gating mechanism, which is suitable for modeling dynamic changes in real-time prediction scenarios;

[0062] The dimension of the satellite image input tensor is Where T represents the time step, H and W are spatial dimensions, and C1 is the number of remote sensing channels, which usually includes brightness temperature (TBB), cloud top height, cloud phase, water vapor path, etc. The 3D convolution structure can be defined as follows:

[0063] F s =σ(W* 3D S+b)

[0064] in:

[0065] * 3D Represents a three-dimensional convolution operation;

[0066] is the convolution kernel weight, where D1 is the number of output channels and σ(·) is the activation function such as ReLU; Represents the output feature tensor B;

[0067] The feature tensor B encodes the dynamic features extracted at each spatial location. The deep channel number D1 reflects the high-dimensional abstract expression of the satellite's spatiotemporal characteristics and can effectively characterize the cloud formation, movement, and development process related to precipitation. If a ConvLSTM structure is used, the image frame at each moment is input to the shared convolutional gate unit, and the state is updated through the time propagation unit:

[0068] i t =σ(W xi *X t +W hi *H t-1 +b i )

[0069] f t =σ(W xf *X t +W hf *H t-1 +b f )

[0070] o t =σ(W xo *X t +W ho *H t-1 +b o )

[0071]

[0072] H t =o t ⊙tanh(C t )

[0073] in, is the input image at time t, Output the feature map at the current moment. All convolutional operations are two-dimensional spatial convolutions, and the weights W are shared in time.

[0074] Secondly, for the output variables of the numerical weather forecast model, due to their large variable dimensions, strong physical meanings but lack of image structures, the present invention uses one-dimensional convolution (1DConv) of pixel elements or multi-layer perceptron (MLP) to perform feature compression and representation learning on them. The input tensor is M∈R^(T×H×W×C_2), where C_2 is the number of physical variable channels, including precipitation, temperature and humidity, wind speed, convective parameters (such as CAPE, CIN), etc. The extraction process is as follows:

[0075] F m (ij)=φ(W m ·Flatten(M :i,j, )+b m )

[0076] Where: - M :i,j, ∈R^(T×C_2) represents the time series physical variables of the (i,j) pixel; - Flatten(·) represents the flattening operation; is the output feature tensor C; - φ(·) is the non-linear activation function.

[0077] This feature tensor C contains the compressed representations of each spatial position in the time dimension and variable dimension. D2 reflects the deep feature encodings of the evolution trend of the wet condensation wind field, convective latent heat and boundary layer processes, etc., and can provide physical priors for subsequent feature fusion.

[0078] Thus, the dynamic perception of remote sensing data and the physical abstraction of numerical model data are realized during the use process, enabling the two heterogeneous modalities to be aligned and fused in the same dimensional space after being transformed by the deep network. In the initial stage of the precipitation system development, the strong convective generation stage and the precipitation center displacement scenario, this dual-channel feature extraction structure can improve the spatial accuracy of precipitation correction and the ability to identify extreme events compared with the traditional single-mode input method, effectively enhancing the spatio-temporal perception and forecasting ability of the model.

[0079] Guided modality fusion is used to execute the cross-attention mechanism. The cross-attention mechanism in the modality fusion module is constructed based on the multi-head attention network. Respectively, taking the satellite feature tensor B as the Query, and the model feature tensor C as the Key and Value, a guided feature selection and fusion mechanism is constructed to dynamically regulate the importance of different modality features;

[0080] The fused feature tensor D is H, W, D3;

[0081] The tensor D represents the deep semantic representation of each spatial pixel integrating multi-modal information and is used to guide the precipitation correction output.

[0082] Specifically, after being processed by the pre - placed multi - modal feature extraction module, the satellite feature tensor \(B\in\mathbb{R}^{H\times W\times D}\) and the model feature tensor \(C\in\mathbb{R}^{H\times W\times d_2}\) are obtained, which respectively represent the satellite spatio - temporal semantic features and numerical model physical variable features extracted on the spatial grid \((H, W)\). To achieve information guidance and fusion in the model state, the satellite feature tensor \(B\) is regarded as the main modality (main guiding modality) and used as the Query in the interactive attention mechanism; the model feature tensor \(C\) is regarded as the auxiliary modality and used as the Key and Value.

[0083] With this setting, the model can selectively focus on and weighted - integrate the multi - variable information in the model data based on the strong observation features of the satellite.

[0084] Specifically, the cross - attention mechanism adopts the standard multi - head attention (Multi - Head Attention) structure, and its core calculation formula is as follows:

[0085]

[0086] Among them,

[0087] \(Q = BW_Q\in\mathbb{R}^{l\times d_k}\) is the query matrix, obtained by linearly transforming the satellite feature \(B\);

[0088] \(K = CW_K\in\mathbb{R}^{l\times d_k}\) is the key matrix, generated from the model feature \(C\);

[0089] \(V = CW_V\in\mathbb{R}^{l\times d_s}\) is the value matrix, also generated from the model feature;

[0090] \(d_k\) is the dimension of the key / query, usually 64 or 128;

[0091] \(W_Q, W_K, W_V\) are learnable parameter matrices.

[0092] The core of this mechanism is to weighted - retrieve each variable dimension in the model feature through the query (Query) vector proposed by the satellite feature, so as to achieve "attention" to relevant regions / variables

[0093] and strengthen its weight in the fusion process. At the same time, the multi - head mechanism (h independent attention heads) is introduced, which allows the model to independently calculate the attention results in multiple semantic spaces, and then obtain the final fusion representation through concatenation and linear mapping:

[0094] \(MultiHead(Q, K, V)=Concat(head_1,\cdots,head_h)W\) h )W O ;

[0095] where each The final mapping matrix \(W\) OAggregate the multi-head results into a unified dimension D3 and output the fused feature tensor The core function of this guided cross-attention fusion module is to achieve adaptive information fusion between modalities. Different from simple feature concatenation or weighted average methods, its fusion process is guided by the main modality, highlighting the importance of regions or variables highly correlated with the main modality through the attention mechanism. The fusion result at each spatial position changes dynamically according to the input features, enhancing the expression diversity, establishing a semantic correspondence between remote sensing observations and numerical models, and providing a higher-level deep semantic input for subsequent precipitation prediction.

[0096] The fused feature tensor D is used as the final fusion result, and its shape is where D3 is usually taken as 128 or 256, representing the semantic channel dimension after integrating modal information. This tensor is directly input into the subsequent main precipitation correction model as an important basis for the model's predicted output, thus significantly enhancing the model's learning ability for non-linear precipitation structures such as complex weather systems and local convective processes.

[0097] The main precipitation correction model is used for multi-scale feature encoding and decoding operations;

[0098] The tensor E is H, W, 1, representing the pixel-by-pixel precipitation correction result for future time steps.

[0099] Specifically, hierarchically encode the input multi-modal fusion feature D to extract features from low-level detailed information to high-level semantic information. To effectively capture precipitation patterns and atmospheric dynamics processes at different scales, a multi-layer convolutional network (Convolutional Neural Networks, CNNs) or a variant convolutional neural network (such as the U-Net architecture or ResNet structure) is used for feature extraction. Specifically, spatial features are gradually extracted through multiple convolutional layers, the size of the feature map decreases layer by layer, and the number of channels gradually increases.

[0100] The convolutional operations for multi-scale encoding use different strides and convolutional kernel sizes to achieve feature extraction at different scales. The formula is:

[0101] FeatureMap i =Conv(D, W i , b i )

[0102] where D is the input multi-modal fusion feature tensor, W i and b i are the convolutional kernel and bias term of the i-th layer respectively, and the size of the obtained feature map changes gradually through the convolutional operation;

[0103] The decoding process gradually restores the low-level abstract features to the original spatial resolution through deconvolution or upsampling operations, and then obtains the prediction results of pixel-by-pixel precipitation. The decoding operation can be achieved through upsampling and deconvolution, and its core formula is:

[0104] UpsampleFeatureMap i =Deconv(FeatureMap i ,W i ,b i )

[0105] Among them, Deconv represents the deconvolution operation, which gradually restores the low-dimensional features to the target spatial size.

[0106] Finally, through the decoded feature map, the pixel-by-pixel precipitation correction result with a spatial resolution of H×W is obtained. The tensor where each spatial pixel value represents the precipitation amount at the corresponding time step. In order to make the predicted value consistent with the actual observation, a linear layer (FullvConnectedLaver) is used for the final mapping to ensure that the output result conforms to the range of precipitation intensity in terms of physical meaning. The finally output precipitation correction result is obtained through:

[0107] E=FC(UpsampleFeatureMap n )

[0108] Among them, FC is the fully connected layer operation, and UpsampleFeatureMap n is the last layer feature map output by the decoder.

[0109] The multi-task assisted decoder includes two parallel branch networks, which are respectively used to predict the radar reflectivity tensor F and the precipitation uncertainty tensor G, and jointly optimize the combined loss with the main task output tensor E during the training process;

[0110] The tensor F is H,W,1, representing the reflectivity factor map;

[0111] The tensor G is H,W,1, representing the precipitation uncertainty corresponding to each pixel. The precipitation uncertainty includes standard deviation, confidence interval or confidence score.

[0112] Specifically, in the training stage of the precipitation correction main model, a multi-task loss function is constructed based on the tensors E, F, G and the ground truth data, and the prediction error and physical consistency loss are jointly minimized. Among them, the physical constraints include water vapor mass conservation, energy balance, and smoothness of the spatial gradient of the precipitation field, and a total loss function is formed to optimize the model parameters.

[0113] Specifically, to further enhance the discriminative ability and output expressiveness of the model, the present invention introduces a multi-task auxiliary decoder structure after the precipitation correction main model decoding module to improve the physical consistency and uncertainty expression ability of the prediction results. This structure includes two parallel auxiliary task branches, which are respectively used to predict the radar reflectivity factor image (tensor F) and the precipitation uncertainty score image (tensor G), and jointly construct a multi-task learning framework with the main task output tensor E to perform end-to-end joint loss optimization during the model training process;

[0114] Fusion feature tensor Generate precipitation correction output through the main decoding module After that, the intermediate semantic features of the decoding process are passed into the multi-task auxiliary decoder. This auxiliary decoder structure includes two parallel branches with complementary functions:

[0115] Branch 1: Radar reflectivity factor prediction network (output tensor F): This branch aims to predict the radar reflectivity factor map (tensor F) of the future time step using remote sensing features directly related to precipitation, and its shape is

[0116] Branch 2: This branch is used to learn and predict the uncertainty score of the main task output E, and the output tensor, represents the standard deviation, confidence interval width or confidence score of the precipitation prediction result on each pixel. This output not only improves the robustness of the model to data anomalies, but also provides model-based credibility auxiliary information for meteorological operations to assist the decision support system in optimizing the precipitation alert level.

[0117] This branch uses a shallow network with the same structure as the main branch, but the output layer uses the Softplus or Exponential activation function to ensure that the uncertainty value is a non-negative real number. If the prediction standard deviation is the target (commonly used in Aleatoric uncertainty modeling), then it has the following form:

[0118] where σ = G

[0119] The model prediction result E is the mean value, G represents the predicted standard deviation, and is regressed by this branch network.

[0120] During the model training process, this paper uses a multi-task joint loss function to perform end-to-end optimization on the main task (precipitation correction) and the two auxiliary tasks (radar reflectivity prediction and uncertainty estimation). The joint loss function is defined as:

[0121] L total = λ1·L rain (E, Y)+λ2·L refl (F, R)+λ3·Luncert (E, Y, G)

[0122] Among them:

[0123] L rain is the precipitation correction loss function for the main task, and weighted MSE, CRPS or combined loss can be adopted;

[0124] L refl is the reflectivity prediction loss function, usually the standard MSE loss;

[0125] L uncert is the negative log-likelihood loss based on uncertainty, and its form is:

[0126]

[0127] This item assumes that the prediction error follows a Gaussian distribution and is optimized in the maximum entropy framework, which can achieve physically consistent modeling for uncertainty scoring.

[0128] λ1, λ2, λ3 are task weighting factors, and the hyperparameters are set according to the importance of the tasks, and usually take values in the ratio of 1:0.5:0.2, etc.

[0129] The physical consistency loss includes:

[0130] The water vapor conservation loss based on vertical integration, which is used to constrain the predicted precipitation not to exceed the water vapor flux;

[0131] The spatial gradient regularization term, which is used to improve the spatial continuity of precipitation prediction;

[0132] The moist static energy balance term, which is used to maintain the energy closure of the physical field.

[0133] Specifically, through physical constraints such as water vapor conservation and moist static energy balance, it is ensured that the precipitation prediction conforms to meteorological physical laws, avoiding prediction results that do not conform to actual meteorological phenomena. The spatial gradient regularization term helps to improve the spatial smoothness of the precipitation field, reduce the noise and fluctuations in the prediction results, and make the precipitation distribution more in line with the characteristics of continuity in nature;

[0134] And through the physical consistency loss, the model not only has good fitting ability statistically, but also can generate more realistic precipitation predictions in terms of physical meaning, improving the credibility of the prediction results. The physical consistency loss can help the model maintain good performance in different meteorological scenarios, especially for complex regional precipitation prediction and extreme weather event forecasting, and can effectively improve the generalization ability.

[0135] During the model training phase, a meta-learning algorithm based on task partitioning is used to train and optimize the task sets composed of different regions or weather types, obtaining the initial parameters of the model with fast adaptation ability. The initial parameters are used to enable the model to complete effective migration with a small number of samples in new regions or new scenarios.

[0136] Specifically, task partitioning means dividing the precipitation prediction task according to factors such as different regions, weather types, or seasons. Each task represents a specific meteorological scenario. There are certain similarities among these tasks, but also obvious differences. Therefore, each task needs to be optimized independently to enhance the generalization ability of the model. Through the meta-learning algorithm, the model can be trained on multiple tasks and learn an ability to quickly adapt to new tasks.

[0137] The training dataset is divided into multiple sub-task sets T1, T2, …, T N , where each sub-task set corresponds to a different region or weather type (e.g., urban weather, mountain precipitation, tropical cyclone, etc.). The samples within each task set are all from precipitation data under specific regions or weather conditions;

[0138] A task T randomly sampled from all available task sets i . Each task set T i contains a set of training samples and corresponding labels, representing the precipitation data of a specific region or weather type;

[0139] For each task T i it is trained by standard gradient descent methods (such as SGD or Adam), calculating the loss function and updating the model parameters θ.

[0140] The process of task training is as follows:

[0141]

[0142] where θ is the shared initial model parameter, α is the learning rate, is the loss function of task T i .

[0143] Task evaluation: For each task T i , after training, the training effect is evaluated using the task validation set and the corresponding validation loss is calculated

[0144] Meta-gradient update: The validation losses of all tasks are aggregated, and the shared model parameter θ is updated through meta-gradient to optimize the generalization ability on multiple tasks. The parameter update is performed through the following formula:

[0145]

[0146] Among them, β is the learning rate of meta-learning, which controls the meta-update pace.

[0147] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for correcting model precipitation forecasts based on deep learning and Fengyun satellites, characterized in that, It includes the following steps: First, build a precipitation correction main model, collect multi-source data, and perform preliminary processing on the collected multi-source data to obtain a fused input data set A, which is combined with ground truth data for supervised training; Then, input the data set A into multi-modal feature extraction to perform feature extraction and generate a satellite feature tensor B; At the same time, input the numerical model variable field into a one-dimensional convolutional network or a multi-layer perceptron for pixel-by-pixel feature extraction to generate a model physical feature tensor C; Subsequently, input the satellite feature tensor B and the model feature tensor C into guided modal fusion to generate a fused feature tensor D based on the inter-modal correlation; Input the fused feature tensor D into the model through the precipitation correction main model to generate a corrected precipitation prediction tensor E; At the same time, input the fused feature tensor D into a multi-task auxiliary decoder to generate a radar reflectivity prediction tensor F and an uncertainty estimation tensor G; Deploy the trained model into the system, input the fused data set A at the current time step, and output the future precipitation correction tensor E and its uncertainty tensor G, which are used to replace or correct the precipitation forecast results of the original numerical model.

2. The method for correcting model precipitation forecast based on deep learning and Fengyun satellite according to claim 1, characterized in that: The multi-source data includes multi-source raw data of Fengyun satellite remote sensing observation data, numerical weather prediction model output data, and ground truth precipitation data. The preliminary processing includes spatial interpolation and time registration processing to unify the data format with consistent grid resolution and time step; The data set A includes: A sequence of Fengyun satellite remote sensing images with a shape of T, H, W, C1; the sequence of Fengyun satellite remote sensing images includes brightness temperature, cloud top height, cloud phase state, and water vapor path channels, with a time resolution of 5 minutes to 15 minutes. Here, T is the length of the time series, H and W are the spatial grid dimensions, and C1 is the number of remote sensing channels. The number of remote sensing channels includes TBB, cloud height, and water vapor; A numerical model variable field with a shape of T, H, W, C2, where C2 is the number of meteorological variables. The number of meteorological variables includes precipitation, humidity, temperature, and wind vector.

3. The method for correcting model precipitation forecast based on deep learning and Fengyun satellites according to claim 1, characterized in that: The feature extraction is used to apply 3D convolution or convolutional long short-term memory network to the sequence of Fengyun satellite images to extract spatio-temporal evolution features; The feature tensor B is represented as H, W, D1; where D1 is the number of satellite deep semantic feature channels extracted, representing the dynamic feature encoding of satellite images at each spatial position in the time dimension.

4. The method for correcting pattern precipitation forecast based on deep learning and Fengyun satellites according to claim 1, wherein: The feature tensor C is represented as H, W, D2; where D2 is the high-dimensional representation of the extracted model variables, including the semantic compression of the temperature, humidity, wind field, and convective parameter information.

5. The method for correcting model precipitation forecast based on deep learning and Fengyun satellite according to claim 1, characterized in that: The guided modal fusion is used to perform a cross-attention mechanism. The cross-attention mechanism in the modal fusion module is constructed based on a multi-head attention network, using the satellite feature tensor B as the Query, and the model feature tensor C as the Key and Value respectively to construct a guided feature selection and fusion mechanism to dynamically regulate the importance of different modal features; The fused feature tensor D is H, W, D3; Tensor D represents the deep semantic representation of each spatial pixel integrating multi-modal information, which is used to guide the precipitation correction output.

6. The method for correcting model precipitation forecast based on deep learning and Fengyun satellite according to claim 1, characterized in that: The precipitation correction main model is used for multi-scale feature encoding and decoding operations; The tensor E is of size H, W, 1, representing the pixel-by-pixel precipitation correction result for future time steps.

7. The method for correcting model precipitation forecast based on deep learning and Fengyun satellite according to claim 1, characterized in that: The multi-task auxiliary decoder includes two parallel branch networks, which are respectively used to predict the radar reflectivity tensor F and the precipitation uncertainty tensor G, and jointly optimize the loss with the main task output tensor E during the training process; The tensor F is of size H, W, 1, representing the reflectivity factor map; The tensor G is of size H, W, 1, representing the precipitation uncertainty corresponding to each pixel, and the precipitation uncertainty includes standard deviation, confidence interval or confidence score.

8. The method for correcting model precipitation prediction based on deep learning and Fengyun satellites according to claim 1, characterized in that: In the training stage of the precipitation correction main model, a multi-task loss function is constructed based on the tensors E, F, G and the ground truth data, and the prediction error and the physical consistency loss are jointly minimized, where the physical constraints include water vapor mass conservation, energy balance, and smoothness of the spatial gradient of the precipitation field, and a total loss function is formed to optimize the model parameters.

9. The method for correcting model precipitation prediction based on deep learning and Fengyun satellites according to claim 1, characterized in that: The physical consistency loss includes: The water vapor conservation loss based on vertical integration, which is used to constrain the predicted precipitation not to exceed the water vapor flux; The spatial gradient regularization term, which is used to enhance the spatial continuity of precipitation prediction; The moist static energy balance term, which is used to maintain the energy closure of the physical field.

10. The method for correcting the model precipitation forecast based on deep learning and Fengyun satellites according to claim 1, wherein: In the model training stage, a meta-learning algorithm based on task division is used to train and optimize the task set composed of different regions or weather types, and the initial parameters of the model with fast adaptation ability are obtained. The initial parameters are used to enable the model to complete effective migration through a small number of samples in a new region or new scenario.

Images