Summer precipitation sub-seasonal prediction method and system fusing multi-scale deep learning

Abstract

The application discloses a kind of summer precipitation subseasonal prediction method and system of fusion multi-scale deep learning, comprising: collecting multi-source meteorological forecast data and observation data, and obtaining precipitation main mode and its mode sequence by carrying out experience orthogonal decomposition to observation data;Multi-scale signal extraction is carried out to observation data, and correlation analysis is carried out with mode sequence to obtain respective weight field;Deep learning model that is fused with multi-pole attention mechanism and time decomposition is constructed and trained;The trained model is input into the forecast data to carry out transfer learning, and the model is optimized;The forecast data of preset time is substituted into the trained model, and high-quality summer precipitation subseasonal forecast product is generated.The application fully considers the synergies of sea, land and air and the interaction of multi-scale signals, and constructs a model based on artificial intelligence methods and numerical model forecast products, effectively improves the subseasonal forecast skill of summer precipitation, and plays an important role in disaster prevention and reduction.

Description

Technical Field

[0001] This invention relates to a method for predicting summer precipitation in the next season, and more particularly to a method and system for predicting summer precipitation in the next season that integrates multi-scale deep learning. Background Technology

[0002] High-quality subseasonal precipitation forecasts play a crucial role in disaster prevention and mitigation. Dynamical models are one of the most important tools in weather forecasting, but numerous studies have shown that current dynamical models still have very limited capabilities for subseasonal precipitation forecasting, requiring further exploration and improvement.

[0003] Besides further optimizing dynamical models, constructing statistical models based on leading correlations in long-term observational data is an important means to improve sub-seasonal forecasting accuracy. However, such statistical models often suffer from instability. In recent years, researchers have found that although dynamical models have limited sub-seasonal forecasting accuracy for precipitation, they still possess some forecasting accuracy for atmospheric circulation signals over long forecast lead times. Therefore, constructing dynamical statistical models based on contemporaneous signals from dynamical models has become an important approach to improving sub-seasonal forecasting accuracy. Currently, dynamical statistical models have been applied to some extent, but most of these models only consider signals in the atmosphere, neglecting the synergistic effects of the ocean, land, and atmosphere. On the other hand, most current dynamical statistical models are based on linear methods or simple deep learning methods, which struggle to capture the interactions between multi-scale signals, thus their forecasting accuracy remains significantly limited. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method and system for predicting summer precipitation in the next season by integrating multi-scale deep learning, so as to improve the sub-seasonal prediction skills of summer precipitation and thus serve disaster prevention and mitigation work.

[0005] Technical solution: The method described in this invention includes the following steps:

[0006] Multi-source meteorological forecast data and observational data were collected, and empirical orthogonal decomposition of the observational data was performed to obtain the main precipitation modes and their mode sequences;

[0007] Multi-scale signal extraction was performed on ocean, land, and atmosphere observation data, and correlation analysis was conducted with modal sequences to obtain their respective weight fields;

[0008] A deep learning model integrating multi-level attention mechanism and temporal decomposition is constructed and trained based on processed observation data. The model includes a first prediction path and a second prediction path. The first prediction path includes: firstly, multiplying the obtained weight field with the extracted multi-scale signal point by point to obtain a preliminary prediction factor field; then, feeding the preliminary prediction factor field into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information; finally, obtaining the regression mode sequence through regression mapping. The second prediction path includes: temporally decomposing the mode sequence. The model is decomposed into a trend term and a seasonal term. The trend term is extrapolated and predicted using the Transformer module to characterize the overall direction of change of the mode. The seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure. The extrapolated results of the trend term and the seasonal extrapolated results are adaptively combined by a gated fusion unit to generate a complete extrapolated mode sequence. The regression mode sequence and the extrapolated mode sequence are integrated to obtain the future mode sequence prediction result. The future mode sequence is reconstructed with the main precipitation mode to generate the corresponding precipitation field, thereby obtaining the final summer precipitation subseasonal forecast product.

[0009] Forecast data is input into the trained model to perform transfer learning and optimize the model;

[0010] By inputting the forecast data at a preset time into the trained model, a high-quality sub-season forecast product for summer precipitation is generated.

[0011] Furthermore, outlier removal and spatiotemporal interpolation preprocessing are required for the collected multi-source meteorological forecast data and observation data. In the time dimension, linear interpolation is used to process missing or discontinuous data; in the spatial dimension, optimal interpolation is used to spatially interpolate data from different resolutions or different observation stations to obtain a spatiotemporally continuous and consistent data field.

[0012] Furthermore, multi-scale signal extraction is performed on ocean, land, and atmospheric observation data, including:

[0013] The signals are divided into three time scales: seasonal scale signals, 10–30 day scale signals, and 30–90 day scale signals.

[0014] Among them, the seasonal scale signal uses the 30-day moving average with a window of 90 days prior to the reporting date as the representative quantity to extract the main components of background variability; for the 30-90 day scale signal, the daily anomaly data is first obtained, then the average of the past 45 days is subtracted, and a moving average of the previous 15 days is performed; for the 10-30 day scale signal, the daily anomaly data is first obtained, then the average of the past 15 days is subtracted, and a moving average of the previous 5 days is performed; thus, signal fields of three different time scales are constructed respectively.

[0015] Furthermore, correlation analysis was performed with the modal sequences to obtain their respective weight fields, including:

[0016] For each time scale, the correlation coefficient distribution with each mode is calculated, and significance is screened based on a 90% confidence level: the weight is assigned to 1 at grid points that meet the significance, and otherwise assigned to 0; thus forming a weight field.

[0017] Furthermore, the preliminary forecast factor field is fed into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information. Specifically:

[0018] Preliminary forecast factor fields as input features By downsampling operator Feature maps at different scales were obtained:

[0019] ;

[0020] in, This is a scale feature map. For scale indexing, The total number of multi-scale branches;

[0021] At each scale Above, the attention weight matrix is ​​calculated based on the local neighborhood:

[0022] ;

[0023] in, For the first Attention weight matrix at each scale For normalization function, For the feature vector dimension, For the Query matrix, For the key matrix, This is the matrix transpose symbol. , From scale feature maps respectively The attention is obtained through linear mapping, and it is computed only within a preset neighborhood window to extract the local dependency structure at that scale.

[0024] The attention matrix at each scale is upsampled using an operator. Return to the original resolution and adjust by weighting coefficients. The fusion yields a multi-polar attention matrix. :

[0025] ;

[0026] Using the above multi-pole attention matrix for input features Feature aggregation yields:

[0027] ;

[0028] in, This represents a high-dimensional representation after fusing multi-scale dependencies; ultimately, a regression mapping is used to... Convert to a regression mode sequence.

[0029] Furthermore, the seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure; specifically:

[0030] The seasonal items are rearranged according to their main period length to construct a periodic view with a two-dimensional time-phase layout. Each column corresponds to the same phase position of the sequence, and different columns reflect the progression relationship of the cycle. The constructed two-dimensional periodic view contains a visible historical segment and the region where the future segment to be predicted is located. When constructing the input, several columns at the end of the periodic view corresponding to the future time period are completely occluded to form a masked segment. The masked periodic image and its corresponding historical part are input into a general large-scale vision model. This model extracts local patterns and phase correlations in the periodic image through an encoder-decoder structure and reconstructs the occluded region under the constraint of the historical segment. The reconstruction process is regarded as a mapping of the periodic view by the large-scale vision model.

[0031] ;

[0032] in, For large-scale visual models, This is a periodic view after masking. The model recovers a complete periodic image, with the end portion corresponding to the extrapolated estimate of the seasonal component.

[0033] The system of the present invention includes:

[0034] The first data processing unit is used to collect multi-source meteorological forecast data and observation data, and to perform empirical orthogonal decomposition on the observation data to obtain the main precipitation modes and their mode sequences;

[0035] The second data processing unit is used to extract signals from ocean, land, and air observation data at multiple scales and perform correlation analysis with modal sequences to obtain their respective weight fields.

[0036] A model building and training unit is used to build a deep learning model that integrates a multi-level attention mechanism and temporal decomposition, and to train the model based on processed observation data. The model includes a first prediction path and a second prediction path. The first prediction path includes: firstly, multiplying the obtained weight field with the extracted multi-scale signal point by point to obtain a preliminary prediction factor field; then, feeding the preliminary prediction factor field into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information; finally, obtaining a regression mode sequence through regression mapping. The second prediction path includes: processing the mode sequence... The time series is decomposed into a trend term and a seasonal term. The trend term is extrapolated and predicted using the Transformer module to characterize the overall direction of mode change. The seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure. The extrapolated results of the trend term and the seasonal extrapolated results are adaptively combined by a gated fusion unit to generate a complete extrapolated mode sequence. The regression mode sequence and the extrapolated mode sequence are integrated to obtain the future mode sequence prediction result. The future mode sequence is reconstructed with the main precipitation mode to generate the corresponding precipitation field, thus obtaining the final summer precipitation subseasonal forecast product.

[0037] The model optimization unit is used to input forecast data into the trained model to carry out transfer learning and optimize the model;

[0038] The forecasting unit is used to input forecast data at a preset time into a trained model to generate high-quality summer precipitation subseasonal forecast products.

[0039] The present invention also provides an electronic device, comprising:

[0040] Memory, used to store computer programs;

[0041] A processor for executing the computer program to implement the method.

[0042] The present invention also provides a non-volatile storage medium for storing a computer program, wherein the computer program implements the method when executed by a processor.

[0043] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described.

[0044] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are: (1) It fully considers the synergistic effect of sea, land and air on precipitation generation, providing more forecast factors for the second season of precipitation prediction; (2) The constructed deep learning model effectively integrates multi-scale complex signals through multi-pole attention mechanism and time series decomposition, thereby improving the second season of precipitation prediction skills; (3) The present invention first trains a preliminary model based on observation data, and then can be transferred to any dynamic model. It only needs to transfer the data of the dynamic model for learning, so it can be compatible with the output of multiple dynamic models and can be widely used in operational forecasting and disaster risk management, providing important technical support for disaster prevention and mitigation and climate services. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention;

[0046] Figure 2 A schematic diagram of a deep learning model structure that integrates multi-level attention mechanism and temporal decomposition;

[0047] Figure 3 A schematic diagram of the multi-level attention mechanism;

[0048] Figure 4 This is a schematic diagram of a periodic view predictor. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0050] This invention considers the synergistic effects of the sea, land, and air, and involves a deep learning model that can fuse multi-scale signals to construct a dynamic statistical model, which can further improve the sub-seasonal forecasting skills for summer precipitation, thereby serving disaster prevention and mitigation efforts.

[0051] like Figure 1 As shown, the method of the present invention includes the following steps:

[0052] Step 1: Collect multi-source meteorological forecast data and observational data. After data preprocessing, perform empirical orthogonal decomposition on the observational data to obtain the main precipitation modes and their mode sequences. Specifically, this includes the following steps:

[0053] Step 1.1: Collect multi-source meteorological forecast data and observational data, mainly including: outward longwave radiation, geopotential height fields at 200 hPa, 500 hPa, and 850 hPa, wind field, specific humidity field, sea surface temperature, soil moisture, and precipitation. Sea surface temperature characterizes the influence of ocean thermal conditions on atmospheric circulation and precipitation, while soil moisture reflects the modulating effect of land surface moisture conditions on precipitation. The above data undergoes unified preprocessing, mainly including outlier removal and spatiotemporal interpolation. Specifically, linear interpolation is used in the temporal dimension to handle missing or discontinuous data; optimal interpolation is used in the spatial dimension to spatially interpolate data from different resolutions or observation stations to obtain a spatiotemporally continuous and consistent data field, providing high-quality input for subsequent modality recognition and modeling.

[0054] Optimal interpolation is an objective analytical method that utilizes the statistical characteristics of errors between the background field and the observed field. It minimizes the variance of the analytical error while ensuring unbiasedness and is widely used in meteorological data assimilation and gridding of observational data. For the location to be estimated... Analytical value of optimal interpolation Expressed as a linear correction for background field and observation bias:

[0055] ;

[0056] in, Background field at the point to be estimated The value, For observation points The value, Background field at observation point The value, The number of observation points. The optimal interpolation weights are determined by the background error covariance and the observation error covariance matrix:

[0057] ;

[0058] in, This represents the background error covariance between the point to be estimated and each observation point. Represented as the background error covariance matrix between observation points, Let be the observation error covariance matrix. In practical meteorological field applications, the background error covariance can often be approximated by a spatial distance function. For example, an exponential correlation function can be used:

[0059] ;

[0060] in, Spatial distance The background error covariance, Indicates spatial distance. For the relevant length scale, This represents the background error variance.

[0061] Step 1.2: Extract the mean field from precipitation observation data, and then use the empirical orthogonal method to extract its main modes and mode sequences.

[0062] For by spatial points Hou Anomaly Field Composed of Sub-observations Represented in matrix form:

[0063] ;

[0064] in, For the first The spatial point The second observation value, , .

[0065] Jiang Houjupingchang It can be considered as a linear combination of several main precipitation modes and their corresponding mode sequences:

[0066] ;

[0067] in, It is a modal sequence. The main precipitation mode is represented in matrix form as follows:

[0068] ;

[0069] ;

[0070] in, For the first The main precipitation mode is in the first The value at a spatial point For the first The main precipitation mode is in the first The time coefficient for each observation (or time point). .

[0071] For Hou Jupingchang The expression can be multiplied on the right. ,get:

[0072] ;

[0073] in, For matrix transpose, for transpose, for transpose, Since it is a real symmetric matrix, it can be solved using the Jacobi method. eigenvalues ​​∧ of the matrix and main precipitation modes .and Given the identity matrix, the modal sequence can be further obtained:

[0074] ;

[0075] In this way, the most important spatial modes and mode sequences in the meteorological variable field can be extracted. In empirical orthogonal decomposition, the significance of the first few extracted spatial modes and their corresponding mode sequences is usually tested to ensure that the extracted models are statistically significant. A commonly used significance test method is the one proposed by North et al. (1982), which evaluates the significance of the models through randomized experiments.

[0076] Step 2 involves multi-scale signal extraction from ocean, land, and atmospheric observation data, followed by correlation analysis with modal sequences to obtain their respective weight fields. This step aims to identify physical signals that significantly contribute to the target mode, thus providing a basis for subsequent factor selection and model construction. In terms of time scale division, this invention mainly considers three typical intervals: seasonal scale signals, 10–30 day scale signals, and 30–90 day scale signals. To meet the real-time requirements of operational forecasting, this invention does not employ traditional bandpass filtering but instead constructs a non-bandpass decomposition method based on moving average and background field estimation to improve computational efficiency and avoid instability caused by filter endpoint effects.

[0077] For seasonal-scale signals, a 30-day moving average with a 90-day window as the reporting date is used as the representative quantity to extract the main components of background variability. For signals at 30–90 days, daily anomalies are first calculated, then the average of the past 45 days is subtracted, and a moving average of the preceding 15 days is applied. For signals at 10–30 days, daily anomalies are first calculated, then the average of the past 15 days is subtracted, and a moving average of the preceding 5 days is applied. Using these methods, signal fields at three different time scales can be constructed.

[0078] After multi-scale signal extraction, correlation tests are performed between the extracted signals and the main modal sequences obtained in step 1 to construct corresponding weight fields. Specifically, for each time scale, the correlation coefficient distribution with each modality is calculated, and significance is screened based on a 90% confidence level: weights of 1 are assigned to grid points that meet the significance criteria, and 0 is assigned otherwise. The resulting weight field highlights regional signals that are significantly correlated with the target modality while effectively reducing the influence of noise fields, making the subsequent prediction factor construction more robust and efficient.

[0079] Step 3: Construct a deep learning model that integrates multi-level attention mechanism and temporal decomposition, and train the model based on processed observation data. For example... Figure 2 As shown, firstly, the weight field obtained in step 2 is multiplied point-by-point with the ocean, land surface, and atmospheric elements after time-series decomposition (i.e., multi-scale signal extraction from ocean, land, and atmosphere observation data) to obtain a preliminary forecast factor field, which is used to characterize the synergistic effect of multi-sphere climate signals at different time scales. Subsequently, the preliminary forecast factor field is fed into a multi-polar attention module (its structure is shown in...). Figure 3 This module constructs local attention through multi-scale features, thereby obtaining an attention representation that combines both fine-scale and long-scale information. Specifically, it uses the preliminary prediction factor field as the input feature. By downsampling operator Feature maps at different scales were obtained:

[0080] ;

[0081] in, This is a scale feature map. For scale indexing, The total number of multi-scale branches.

[0082] At each scale Above, the attention weight matrix is ​​calculated based on the local neighborhood:

[0083] ;

[0084] in, For the first Attention weight matrix at each scale For normalization function, For the feature vector dimension, For the Query matrix, For the key matrix, This is the matrix transpose symbol. , From scale feature maps respectively The attention matrix is ​​obtained through linear mapping, and is computed only within a preset neighborhood window to extract the local dependency structure at that scale. Subsequently, the attention matrix at each scale is upsampled using an upsampling operator. Return to the original resolution and adjust by weighting coefficients. The fusion yields a multi-polar attention matrix. :

[0085] ;

[0086] Using the above multi-pole attention matrix for input features Feature aggregation yields:

[0087] ;

[0088] in, This represents a high-dimensional representation after fusing multi-scale dependencies. Finally, a regression mapping is used to... The sequence is converted into a regression mode sequence, forming the first prediction path of the model.

[0089] At the same time, such as Figure 2 As shown in the lower part, the present invention also addresses existing modal sequences. The time series is decomposed into a trend component and a seasonal component. The trend component is extrapolated and predicted using the Transformer module to characterize the overall direction of modality change; while the seasonal component (i.e., the seasonal term) is fed into the cycle view predictor to capture its internal cyclical perturbation structure. To enable the model to more fully identify the cyclical patterns in the seasonal sequence, this invention rearranges the seasonal components according to their main cycle length, constructing a cycle view with a two-dimensional "time-phase" layout. Each column corresponds to the same phase position in the sequence, and different columns reflect the progression relationship of the cycle. The constructed two-dimensional cycle view is as follows: Figure 4 As shown on the left, the image contains a visible historical segment and the region where the future segment to be predicted is located. During input construction, several columns at the end of the periodic view corresponding to future time periods are completely masked, forming masked segments. The masked periodic image and its corresponding historical portion are then input into a general large-scale vision model. The model extracts local patterns and phase correlations in periodic images through an encoder-decoder structure and reconstructs occluded regions under the constraints of historical segments. Figure 4 The left side represents the "history-masking" structure, with the white area being the completely obscured portion; the result after reconstruction is as follows. Figure 4 As shown on the right, the blue area represents the future segment (forecast) generated by the model. The reconstruction process can be viewed as a mapping of a large visual model to a periodic view:

[0090] ;

[0091] in, This is a periodic view after masking. The model reconstructs a complete periodic image, with the end portion corresponding to the extrapolated estimates of the seasonal components. Through this two-dimensional reconstruction mechanism, the spatial organization characteristics of periodic sequences can be effectively utilized, enabling the model to capture key information such as the propagation direction of periodic perturbations and phase progression patterns.

[0092] Subsequently, the trend extrapolation results and the seasonal extrapolation results are adaptively combined by a gated fusion unit to generate a complete extrapolation mode sequence, which serves as the second prediction path of the model.

[0093] Finally, the regression mode sequence given by the first prediction path and the extrapolated mode sequence obtained by the second prediction path are further integrated to take into account both short-term variation characteristics and long-term trend characteristics, thereby obtaining the prediction result of the future mode sequence. This future mode sequence is then reconstructed with the main precipitation modes obtained in step 1 to generate the corresponding precipitation field, thus obtaining the final summer precipitation sub-seasonal forecast product.

[0094] The technical principle behind the aforementioned dual-path architecture is as follows: The first prediction path (regression path) focuses on "physical attribution in the spatial dimension," that is, it uses a multi-polar attention mechanism to mine the nonlinear forcing effect of multi-sphere environmental fields such as ocean, land surface, and atmosphere on precipitation modes. Its advantage lies in capturing abrupt changes in precipitation caused by environmental field anomalies. The second prediction path (extrapolation path) focuses on "evolutionary inertia in the temporal dimension," that is, it uses large-scale visual models and Transformers to capture the periodic phase locking and long-term trends of precipitation mode sequences themselves. Its advantage lies in using the system's memory for stable prediction. Existing technologies typically only use a single "environmental field regression" or "sequence extrapolation," making it difficult to simultaneously consider spatial forcing and temporal inertia. This invention, by constructing these two positively complementary feature extraction paths and performing gated fusion, can adaptively adjust the prediction strategy according to the dominant mechanism in different time periods, thereby effectively improving forecasting skills.

[0095] Step 4: Input the forecast data into the trained model to conduct transfer learning, further optimizing model performance. Since there are usually fixed biases or structural differences between forecast and observational data, the model at this stage readjusts some parameters to better adapt its internal representation to the characteristic distribution of the forecast data. Through this process, the model's stability and applicability under operational conditions are further improved.

[0096] Step 5: Input the forecast data at the preset time into the trained model to generate a high-quality summer precipitation sub-season forecast product.

[0097] Example 2: This invention also provides a summer precipitation subseason prediction system that integrates multi-scale deep learning, comprising:

[0098] The first data processing unit is used to collect multi-source meteorological forecast data and observation data, and to perform empirical orthogonal decomposition on the observation data to obtain the main precipitation modes and their mode sequences;

[0099] The second data processing unit is used to extract signals from ocean, land, and air observation data at multiple scales and perform correlation analysis with modal sequences to obtain their respective weight fields.

[0100] A model building and training unit is used to build a deep learning model that integrates a multi-level attention mechanism and temporal decomposition, and to train the model based on processed observation data. The model includes a first prediction path and a second prediction path. The first prediction path includes: firstly, multiplying the obtained weight field with the extracted multi-scale signal point by point to obtain a preliminary prediction factor field; then, feeding the preliminary prediction factor field into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information; finally, obtaining a regression mode sequence through regression mapping. The second prediction path includes: processing the mode sequence... The time series is decomposed into a trend term and a seasonal term. The trend term is extrapolated and predicted using the Transformer module to characterize the overall direction of mode change. The seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure. The extrapolated results of the trend term and the seasonal extrapolated results are adaptively combined by a gated fusion unit to generate a complete extrapolated mode sequence. The regression mode sequence and the extrapolated mode sequence are integrated to obtain the future mode sequence prediction result. The future mode sequence is reconstructed with the main precipitation mode to generate the corresponding precipitation field, thus obtaining the final summer precipitation subseasonal forecast product.

[0101] The model optimization unit is used to input forecast data into the trained model to carry out transfer learning and optimize the model;

[0102] The forecasting unit is used to input forecast data at a preset time into a trained model to generate high-quality summer precipitation subseasonal forecast products.

[0103] Example 3: The present invention also provides an electronic device, comprising:

[0104] Memory, used to store computer programs;

[0105] A processor for executing the computer program to implement the method.

[0106] Example 4: The present invention also provides a non-volatile storage medium for storing a computer program, wherein the computer program implements the method described when executed by a processor.

[0107] Example 5: The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described.

Claims

1. A method for predicting summer precipitation in the next season by integrating multi-scale deep learning, characterized in that, Includes the following steps: Multi-source meteorological forecast data and observational data were collected, and empirical orthogonal decomposition of the observational data was performed to obtain the main precipitation modes and their mode sequences; Multi-scale signal extraction was performed on ocean, land, and atmosphere observation data, and correlation analysis was conducted with modal sequences to obtain their respective weight fields; A deep learning model integrating multi-level attention mechanism and temporal decomposition is constructed and trained based on processed observation data. The model includes a first prediction path and a second prediction path. The first prediction path includes: firstly, multiplying the obtained weight field with the extracted multi-scale signal point by point to obtain a preliminary prediction factor field; then, feeding the preliminary prediction factor field into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information; finally, obtaining the regression mode sequence through regression mapping. The second prediction path includes: temporally decomposing the mode sequence. The model is decomposed into a trend term and a seasonal term. The trend term is extrapolated and predicted using the Transformer module to characterize the overall direction of change of the mode. The seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure. The extrapolated results of the trend term and the seasonal extrapolated results are adaptively combined by a gated fusion unit to generate a complete extrapolated mode sequence. The regression mode sequence and the extrapolated mode sequence are integrated to obtain the future mode sequence prediction result. The future mode sequence is reconstructed with the main precipitation mode to generate the corresponding precipitation field, thereby obtaining the final summer precipitation subseasonal forecast product. Forecast data is input into the trained model to perform transfer learning and optimize the model; By inputting the forecast data at a preset time into the trained model, a high-quality sub-season forecast product for summer precipitation is generated.

2. The method according to claim 1, characterized in that, It is also necessary to perform outlier removal and spatiotemporal interpolation preprocessing on the collected multi-source meteorological forecast data and observation data. In the time dimension, linear interpolation method is used to process missing or discontinuous data; in the spatial dimension, optimal interpolation method is used to perform spatial interpolation on data of different resolutions or different observation stations to obtain a spatiotemporally continuous and consistent data field.

3. The method according to claim 1, characterized in that, Multi-scale signal extraction from ocean, land, and atmosphere observation data, including: The signals are divided into three time scales: seasonal scale signals, 10–30 day scale signals, and 30–90 day scale signals. Among them, the seasonal scale signal uses the 30-day moving average with a window of 90 days prior to the reporting date as the representative quantity to extract the main components of background variability; for the 30-90 day scale signal, the daily anomaly data is first obtained, then the average of the past 45 days is subtracted, and a moving average of the previous 15 days is performed; for the 10-30 day scale signal, the daily anomaly data is first obtained, then the average of the past 15 days is subtracted, and a moving average of the previous 5 days is performed; thus, signal fields of three different time scales are constructed respectively.

4. The method according to claim 1, characterized in that, Correlation analysis was performed with the modal sequences to obtain their respective weight fields, including: For each time scale, the correlation coefficient distribution with each mode is calculated, and significance is screened based on a 90% confidence level: the weight is assigned to 1 at grid points that meet the significance, and otherwise assigned to 0; thus forming a weight field.

5. The method according to claim 1, characterized in that, The initial forecast factor field is fed into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information. Specifically: Preliminary forecast factor fields as input features By downsampling operator Feature maps at different scales were obtained: ； in, This is a scale feature map. For scale indexing, The total number of multi-scale branches; At each scale Above, the attention weight matrix is ​​calculated based on the local neighborhood: ； in, For the first Attention weight matrix at each scale For normalization function, For the feature vector dimension, For the Query matrix, For the key matrix, This is the matrix transpose symbol. , From scale feature maps respectively The attention is obtained through linear mapping, and it is computed only within a preset neighborhood window to extract the local dependency structure at that scale. The attention matrix at each scale is upsampled using an operator. Return to the original resolution and adjust by weighting coefficients. The fusion yields a multi-polar attention matrix. : ； Using the above multi-pole attention matrix for input features Feature aggregation yields: ； in, This represents a high-dimensional representation after fusing multi-scale dependencies; ultimately, a regression mapping is used to... Convert to a regression mode sequence.

6. The method according to claim 1, characterized in that, The seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure; specifically: The seasonal items are rearranged according to their main period length to construct a periodic view with a two-dimensional time-phase layout. Each column corresponds to the same phase position of the sequence, and different columns reflect the progression relationship of the cycle. The constructed two-dimensional periodic view contains a visible historical segment and the region where the future segment to be predicted is located. When constructing the input, several columns at the end of the periodic view corresponding to the future time period are completely occluded to form a masked segment. The masked periodic image and its corresponding historical part are input into a general large-scale vision model. This model extracts local patterns and phase correlations in the periodic image through an encoder-decoder structure and reconstructs the occluded region under the constraint of the historical segment. The reconstruction process is regarded as a mapping of the periodic view by the large-scale vision model. ； in, For large-scale visual models, This is a periodic view after masking. The model recovers a complete periodic image, with the end portion corresponding to the extrapolated estimate of the seasonal component.

7. A summer precipitation subseasonal prediction system integrating multi-scale deep learning, characterized in that, include: The first data processing unit is used to collect multi-source meteorological forecast data and observation data, and to perform empirical orthogonal decomposition on the observation data to obtain the main precipitation modes and their mode sequences; The second data processing unit is used to extract signals from ocean, land, and air observation data at multiple scales and perform correlation analysis with modal sequences to obtain their respective weight fields. A model building and training unit is used to build a deep learning model that integrates a multi-level attention mechanism and temporal decomposition, and to train the model based on processed observation data. The model includes a first prediction path and a second prediction path. The first prediction path includes: firstly, multiplying the obtained weight field with the extracted multi-scale signal point by point to obtain a preliminary prediction factor field; then, feeding the preliminary prediction factor field into a multi-level attention module, which constructs local attention through multi-scale features to obtain an attention representation that combines fine-scale and long-scale information; finally, obtaining a regression mode sequence through regression mapping. The second prediction path includes: processing the mode sequence... The time series is decomposed into a trend term and a seasonal term. The trend term is extrapolated and predicted using the Transformer module to characterize the overall direction of mode change. The seasonal term is fed into the periodic view predictor to capture its internal periodic perturbation structure. The extrapolated results of the trend term and the seasonal extrapolated results are adaptively combined by a gated fusion unit to generate a complete extrapolated mode sequence. The regression mode sequence and the extrapolated mode sequence are integrated to obtain the future mode sequence prediction result. The future mode sequence is reconstructed with the main precipitation mode to generate the corresponding precipitation field, thus obtaining the final summer precipitation subseasonal forecast product. The model optimization unit is used to input forecast data into the trained model to carry out transfer learning and optimize the model; The forecasting unit is used to input forecast data at a preset time into a trained model to generate high-quality summer precipitation subseasonal forecast products.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1-6.

9. A non-volatile storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 1-6.

Images