A precipitation nowcasting method based on trend-to-detail synergistic residual conditional diffusion

Abstract

The application provides a precipitation nowcast method based on trend-to-detail synergistic residual condition diffusion, and belongs to the technical field of meteorological prediction and artificial intelligence, and comprises the following steps: S1, acquiring a radar echo sequence and near-surface meteorological element data; S2, performing time alignment processing on the near-surface meteorological element data; S3, generating a multi-channel meteorological element grid and a confidence map; S4, jointly training a deterministic trend prediction network and a residual condition diffusion model; S5, generating a prediction residual, generating a radar trend prediction field, fusing the prediction residual and the radar trend prediction field, and obtaining precipitation nowcast results at multiple future moments. The method synergistically models a trend prediction branch and a detail generation branch, fuses radar echo and near-surface meteorological element information in a unified framework, takes into account the translation trend of a large-scale precipitation system and the detail expression of a small-scale convective structure, improves the precision and stability of short-time precipitation nowcast, and has good engineering application value.

Description

Technical Field

[0001] This invention relates to the fields of meteorological forecasting and artificial intelligence technology, specifically to a precipitation nowcasting method based on trend-to-detail co-dependent residual condition diffusion. Background Technology

[0002] Short-duration heavy rainfall and convective storms can severely impact urban drainage, transportation, power and communications, and the safety of people's lives and property. Precipitation nowcasting on a 0-2 hour timescale is a crucial link between real-time monitoring and short-to-medium-term numerical forecasting, and is of great significance for disaster prevention and mitigation and refined meteorological services.

[0003] Current precipitation nowcasting mainly relies on numerical weather prediction models and empirical methods based on radar echo extrapolation. Numerical weather prediction models are limited by spatial resolution and update frequency, resulting in limited ability to characterize small- and medium-scale convective systems and difficulty in timely reflecting the rapid development and attenuation of strong convection. Traditional radar extrapolation methods are mostly based on translation assumptions, which are insufficient in characterizing changes in precipitation intensity, convection formation and dissipation, and are prone to problems such as smoothing of the forecast field and weakening of peak values. With the improvement of radar observation and computing capabilities, deep learning methods based on radar images have been introduced into precipitation nowcasting, but they have the following shortcomings: First, most methods only use radar echoes as a single mode input, failing to fully utilize the physical information contained in near-surface meteorological elements such as temperature, humidity, wind field, and surface air pressure, resulting in insufficient ability to distinguish the occurrence and evolution of strong convection; Second, commonly used deterministic prediction frameworks tend to output an "average field," making it difficult to simultaneously retain the translational trend of large-scale precipitation systems and the detailed features of small-scale convective structures; Third, loss functions often focus on pixel-level errors, which do not adequately constrain the physical consistency of prediction results in terms of temporal continuity, spatial smoothness, and overall precipitation conservation, and are prone to error accumulation and forecast field drift when extrapolating over long periods.

[0004] Therefore, there is an urgent need for a precipitation nowcasting method that can integrate multimodal information such as radar and near-surface meteorological elements within a unified framework, while taking into account both large-scale trend prediction and small-scale detail generation, and introducing physical constraints and reasonable long-term inference strategies, so as to improve the accuracy and stability of short-term heavy precipitation forecasts. Summary of the Invention

[0005] The purpose of this invention is to provide a precipitation nowcasting method based on trend-to-detail co-residual condition diffusion, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion, comprising the following steps: S1: Acquire radar echo sequences and near-surface meteorological element data from multiple meteorological observation stations for the target area within a preset time window; S2: Perform time alignment processing on the near-surface meteorological element data of multiple meteorological observation stations obtained in step S1; S3: Spatial rasterization is performed on the near-surface meteorological element data processed in step S2 to generate multi-channel meteorological element raster and confidence map; S4: Based on the radar echo sequence obtained in step S1 and the real future radar echo, jointly train a deterministic trend prediction network and a residual conditional diffusion model. S5: Generate prediction residuals using the residual conditional diffusion model trained in step S4, generate radar trend prediction fields using the deterministic trend prediction network trained in step S4, and fuse the prediction residuals with the radar trend prediction fields to obtain precipitation nowcast results for multiple future times.

[0007] Furthermore, the specific process of step S2, time alignment processing, is as follows: The near-surface meteorological element data of multiple meteorological observation stations obtained in step S1 are linearly interpolated in the time dimension according to the time corresponding to the radar echo between adjacent observation times to obtain station observation data aligned with the time of each radar echo frame.

[0008] Furthermore, the specific process of step S3 is as follows: S31: For any grid point on the radar grid Select several meteorological observation stations within a preset radius around the grid point, and record the distance between the grid point and the first... The distance between the meteorological observation stations is The meteorological element observation values ​​of this observation station are Meteorological element values ​​at grid points are calculated using weighted linear interpolation. The formula is as follows: , in, Represented as weights; Weight Distance The monotonically decreasing function is given by the formula: , in, For smoothing parameters; S32: Calculate the confidence score based on the weight information of the observation stations participating in the weighted average. And map the confidence scores to a set of values ​​according to a preset normalization rule. interval; S33: Perform steps S31 and S32 on multiple meteorological elements respectively to obtain multi-channel meteorological element grids and confidence maps.

[0009] Furthermore, the specific process of step S4 is as follows: S41: Input the radar echo sequence obtained in step S1 into the deterministic trend prediction network to obtain the radar trend prediction field at multiple future times. S42: Construct a residual sequence based on the real future radar echo and the radar trend prediction field obtained in step S41, and use the residual sequence as the learning target to train the residual conditional diffusion model.

[0010] Furthermore, the deterministic trend prediction network in step S41 includes an encoder module, a temporal mixing module, and a decoder module, wherein the temporal mixing module is disposed between the encoder module and the decoder module.

[0011] Furthermore, the specific process of step S41 is as follows: S411: Arrange the radar echo sequence obtained in step S1 in chronological order to obtain a continuous sequence. A sequence of radar echo frames, each frame of which is represented as a single-channel echo map. ; S412: Transfer each frame Stacking along the time dimension yields the input tensor. , , For time steps, For single-channel radar echo, , These are the space dimensions; S413: Encoder module includes Level 1 convolutional downsampling unit, the first Level downsampling unit for features from the previous layer Spatial downsampling and convolution operations are performed sequentially to output features. The formula is: , ,

[0012] in, Indicates the first Level convolution operation; Indicates spatial downsampling operation; S414: The final feature output by the encoder module from the time mixing module. Apply convolutional aggregation along the time dimension to obtain temporally blended features. The formula is: , in, For time indexing; This is the relative time offset; These are the weights of the temporal convolution kernel; The time convolution radius; This represents a two-dimensional convolution operation in the spatial dimension; S415: Decoder module includes Level upsampling unit, the first Level upsampling unit for features from the previous layer Upsampling ,Will Features corresponding to the level in the encoder Feature concatenation is performed, and output features are obtained through convolution operations. The formula is: , , in, ; Indicates the first Level-space upsampling operation; Indicates feature concatenation operation; Indicates the first Level convolution operation; S416: The last layer of the decoder module passes through... Convolution operation, outputting the future Radar trend prediction field at each time step ,in , To predict the number of time steps.

[0013] Furthermore, the loss function of the deterministic trend prediction network in step S41 includes a basic prediction error term and multiple physical constraint terms. These physical constraint terms constrain temporal smoothing, spatial smoothing, and overall intensity variation. Specifically, the physical constraint terms are combinations of at least one of total variation, adjacent-time-time difference penalty, and overall intensity difference penalty. The loss function of the deterministic trend prediction network... The formula is: , in, These are non-negative weighting coefficients; This is the pixel-level mean square error term between the trend prediction field and the target prediction radar echo; This is a range constraint term used to penalize the portion of the predicted value that is less than zero or greater than a preset upper limit; This is a time smoothing constraint term used to penalize excessive differences between adjacent forecast times; This is a constraint term for total spatial variation, used to suppress isolated spatial noise and severe spatial oscillations; This is an overall intensity constraint term used to penalize abnormal changes in overall reflectivity intensity between adjacent forecast times.

[0014] Furthermore, the specific process of step S42 is as follows: S421: Real Future Radar Echoes With radar trend prediction field The difference is used as the residual. The formula is:

[0015] S422: Residual calculated based on step S421 Construct the residual sequence, and obtain the residual domain based on the residual sequence; S423: Define a forward noise addition process in the residual domain, using the original residuals. Starting from, through Time step on residual Gaussian noise is injected step by step, with the conditional distribution at each step as follows: , The equivalent reparameterized form, i.e., the first The noise addition process is as follows: , in, ; To schedule noise levels in advance; To follow a standard normal distribution Random noise; Let be the identity matrix, representing independent noise addition in each dimension; S424: From the forward process, we can obtain... Regarding the original residual Closed expression: , That is: , in, , indicating the first Signal retention rate at each step; , indicating the preceding The cumulative signal retention rate of the step; S425: Constructing a residual diffusion generation network Generate a network by diffusion to residuals Input noise residual Time step index and conditional features Residual diffusion generation network Output noise estimation ; S426: Employing weighted diffusion loss The residual diffusion generation network is trained with the following loss form: , in, To keep pace with time The relevant non-negative weighting coefficients; Represents the variable and The expected operation.

[0016] Furthermore, in step S4, when jointly training the deterministic trend prediction network and the residual conditional diffusion model, the total loss function... for: , in, For deterministic losses; , These are non-negative weighting coefficients.

[0017] Furthermore, the forecasting process of the forecast results in step S5 adopts a segmented autoregressive sampling strategy. The specific process is as follows: the forecasting process is divided into several continuous forecast segments, and trend prediction and residual generation are performed for each forecast segment. The forecast result of the current segment is obtained based on the residual and trend prediction. The forecast result of the current segment is incorporated into the radar echo sequence as the input for the next forecast segment, until the entire forecasting process is completed.

[0018] Compared with the prior art, the present invention has the following technical effects: The precipitation nowcasting method provided by this invention integrates radar echo and near-surface meteorological element information under a unified framework through collaborative modeling of trend prediction branch and detail generation branch. It also takes into account the translation trend of large-scale precipitation systems and the detailed expression of small-scale convective structures, and introduces physical constraints and reasonable long-term inference strategies, thereby improving the accuracy and stability of short-term precipitation nowcasting and having good engineering application value. Attached Figure Description

[0019] Fig. 1 This is a flowchart of the precipitation nowcasting method according to an embodiment of the present invention; Fig. 2 This is a schematic diagram illustrating the working principle of the precipitation nowcasting method according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0021] In this article, terms such as "left," "right," "up," "down," "front," and "back" are established based on the positional relationships shown in the attached drawings. Depending on the attached drawings, the corresponding positional relationships may also change. Therefore, they should not be interpreted as an absolute limitation on the scope of protection.

[0022] Please see Figs. 1-2 This embodiment provides a precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion, wherein steps S1-S4 are the training phase and step S5 is the inference phase, including the following steps: S1: Acquire radar echo sequences and near-surface meteorological element data from multiple meteorological observation stations within a preset time window (a preset time interval) for the target area (i.e., the area near the forecast of precipitation to be predicted).

[0023] S2: Perform time alignment processing on the near-surface meteorological element data of multiple meteorological observation stations obtained in step S1.

[0024] Specifically, the time alignment process in step S2 is as follows: The near-surface meteorological element data of multiple meteorological observation stations obtained in step S1 are linearly interpolated in the time dimension according to the time corresponding to the radar echo between adjacent observation times to obtain station observation data aligned with the time of each radar echo frame.

[0025] S3: Spatial rasterization is performed on the near-surface meteorological element data processed in step S2 to generate multi-channel meteorological element raster and confidence map.

[0026] Specifically, the process of step S3 is as follows: S31: For any grid point on the radar grid Select several meteorological observation stations within a preset radius around the grid point, and record the distance between the grid point and the first... The distance between the meteorological observation stations is The meteorological element observation values ​​of this observation station are Meteorological element values ​​at grid points are calculated using weighted linear interpolation. The formula is as follows: , in, Represented as weights; Weight Distance The monotonically decreasing function is given by the formula: , in, For smoothing parameters; S32: Calculate the confidence score based on the weight information of the observation stations participating in the weighted average. ,in The cumulative amount of the weights is monotonically correlated. The confidence values ​​are then mapped to a predetermined normalization rule. Interval. Normalization rules include normalization based on a preset upper limit, the maximum possible value of the cumulative weight within the neighborhood, or a statistical measure, so that when the number of observation stations involved in the calculation decreases or the cumulative weight decreases, It then decreases.

[0027] S33: Perform steps S31 and S32 on multiple meteorological elements respectively to obtain multi-channel meteorological element raster and confidence map. The multi-channel meteorological element raster and confidence map are used as one of the conditional information for the residual conditional diffusion model.

[0028] S4: Based on the radar echo sequence obtained in step S1 and the real future radar echo, jointly train the deterministic trend prediction network and the residual conditional diffusion model.

[0029] Specifically, the process of step S4 is as follows: S41: Input the radar echo sequence obtained in step S1 into the deterministic trend prediction network to obtain the radar trend prediction field at multiple future times.

[0030] Specifically, the deterministic trend prediction network in step S41 includes an encoder module, a time mixing module, and a decoder module, with the time mixing module positioned between the encoder module and the decoder module.

[0031] Specifically, the process of step S41 is as follows: S411: Arrange the radar echo sequence obtained in step S1 in chronological order to obtain a continuous sequence. A sequence of radar echo frames, each frame of which is represented as a single-channel echo map. .

[0032] S412: Transfer each frame Stacking along the time dimension yields the input tensor. , , For time steps, For single-channel radar echo, , These refer to the spatial dimensions.

[0033] S413: Encoder module includes Level 1 convolutional downsampling unit, the first Level downsampling unit for features from the previous layer Spatial downsampling and convolution operations are performed sequentially to output features. The formula is: , ,

[0034] in, Indicates the first Level convolution operation; This indicates a spatial downsampling operation.

[0035] S414: The final feature output by the encoder module from the time mixing module. Apply convolutional aggregation along the time dimension to obtain temporally blended features. The formula is: , in, For time indexing; This is the relative time offset; These are the weights of the temporal convolution kernel; The time convolution radius; This represents a two-dimensional convolution operation in spatial dimensions.

[0036] S415: Decoder module includes Level upsampling unit, the first Level upsampling unit for features from the previous layer Upsampling ,Will Features corresponding to the level in the encoder Feature concatenation is performed, and output features are obtained through convolution operations. The formula is: , , in, ; Indicates the first Level-space upsampling operation; Indicates feature concatenation operation; Indicates the first Level convolution operation.

[0037] S416: The last layer of the decoder module passes through... Convolution operation, outputting the future Radar trend prediction field at each time step ,in , To predict the number of time steps.

[0038] Specifically, the loss function of the deterministic trend prediction network in step S41 includes a basic prediction error term and multiple physical constraint terms. These physical constraint terms constrain temporal smoothing, spatial smoothing, and overall intensity variation. The specific form of the physical constraint terms is at least one combination of total variation, adjacent-time-time difference penalty, and overall intensity difference penalty. Loss function of the deterministic trend prediction network Defined as the weighted sum of all losses, the formula is: , in, These are non-negative weighting coefficients; This is the pixel-level mean square error term between the trend prediction field and the target prediction radar echo; This is a range constraint term used to penalize the portion of the predicted value that is less than zero or greater than a preset upper limit; This is a time smoothing constraint term used to penalize excessive differences between adjacent forecast times; This is a constraint term for total spatial variation, used to suppress isolated spatial noise and severe spatial oscillations; This is an overall intensity constraint term used to penalize abnormal changes in overall reflectivity intensity between adjacent forecast times.

[0039] S42: Construct a residual sequence based on the real future radar echo and the radar trend prediction field obtained in step S41, and train the residual conditional diffusion model using the residual sequence as the learning objective. The real future radar echo is the observed radar echo during the forecast target period. The residual conditional diffusion model includes the radar echo sequence obtained in step S1, the radar trend prediction field obtained in step S416, and the multi-channel meteorological element grid and its confidence map obtained in step S3 as conditional information.

[0040] Specifically, the process of step S42 is as follows: The specific process of step S42 is as follows: S421: Real Future Radar Echoes With radar trend prediction field The difference is used as the residual. The formula is:

[0041] S422: Residual calculated based on step S421 Construct the residual sequence and obtain the residual domain based on the residual sequence.

[0042] S423: Define a forward noise addition process in the residual domain, using the original residuals. Starting from, through Time step on residual Gaussian noise is injected step by step, with the conditional distribution at each step as follows: , The equivalent reparameterized form, i.e., the first The noise addition process is as follows: , in, ; To schedule noise levels in advance; To follow a standard normal distribution Random noise; It is an identity matrix, representing that noise is added independently in each dimension.

[0043] S424: From the forward process, we can obtain... Regarding the original residual Closed expression: , That is: , in, , indicating the first Signal retention rate at each step; , indicating the preceding The cumulative signal retention rate of the step.

[0044] S425: Constructing a residual diffusion generation network Generate a network by diffusion to residuals Input noise residual Time step index and conditional features Residual diffusion generation network Output noise estimation .

[0045] Specifically, conditional features It is obtained by encoding the radar echo sequence in step S1, the trend prediction field in step S416, and the multi-channel meteorological element grid and confidence map in step S3.

[0046] S426: Employing weighted diffusion loss The residual diffusion generation network is trained with the following loss form: , in, To keep pace with time The relevant non-negative weighting coefficients are preferably obtained using a minimum signal-to-noise ratio weighting strategy. , For the first The signal-to-noise ratio of the step, This is the cutoff constant; Represents the variable and The expected operation.

[0047] Specifically, in step S4, when jointly training the deterministic trend prediction network and the residual conditional diffusion model, the total loss function... for: , in, For deterministic losses; , These are non-negative weighting coefficients.

[0048] S5: Generate prediction residuals using the residual conditional diffusion model trained in step S4, generate radar trend prediction fields using the deterministic trend prediction network trained in step S4, and fuse the prediction residuals with the radar trend prediction fields to obtain precipitation nowcast results for multiple future times.

[0049] Specifically, the forecasting process in step S5 adopts a segmented autoregressive sampling strategy. The specific process is as follows: the forecasting process is divided into several continuous forecast segments. For each forecast segment, trend prediction and residual generation are performed. Based on the residual and trend prediction, the forecasting result of the current segment is obtained. The forecasting result of the current segment is incorporated into the radar echo sequence as the input for the next forecast segment, until the entire forecasting process is completed.

[0050] Specifically, this precipitation nowcasting method integrates radar echo and near-surface meteorological element information within a unified framework through collaborative modeling of trend prediction branch and detail generation branch. It also takes into account the translational trend of large-scale precipitation systems and the detailed representation of small-scale convective structures, and introduces physical constraints and reasonable long-term inference strategies to improve the accuracy and stability of short-term precipitation nowcasting, thus having good engineering application value.

[0051] The above embodiments merely illustrate the basic principles and characteristics of the present invention, but are not limited to the above implementation schemes. It should be understood that those skilled in the art can make various changes and modifications to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A trend-to-detail co-residual conditional diffusion method for precipitation nowcasting, characterized in that, Includes the following steps: S1: Acquire radar echo sequences and near-surface meteorological element data from multiple meteorological observation stations for the target area within a preset time window; S2: Perform time alignment processing on the near-surface meteorological element data of multiple meteorological observation stations obtained in step S1; S3: Spatial rasterization is performed on the near-surface meteorological element data processed in step S2 to generate multi-channel meteorological element raster and confidence map; S4: Based on the radar echo sequence obtained in step S1 and the real future radar echo, jointly train a deterministic trend prediction network and a residual conditional diffusion model. S5: Generate prediction residuals using the residual conditional diffusion model trained in step S4, generate radar trend prediction fields using the deterministic trend prediction network trained in step S4, and fuse the prediction residuals with the radar trend prediction fields to obtain precipitation nowcast results for multiple future times.

2. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 1, characterized in that, The specific process of time alignment processing in step S2 is as follows: In step S1, near-surface meteorological element data from multiple meteorological observation stations are linearly interpolated between adjacent observation times based on the time corresponding to the radar echo, to obtain station observation data aligned with the time of each radar echo frame.

3. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 2, characterized in that, The specific process of step S3 is as follows: S31: For any grid point on the radar grid Select several meteorological observation stations within a preset radius around the grid point, and record the distance between the grid point and the first... The distance between the meteorological observation stations is The meteorological element observation values ​​of this observation station Meteorological element values ​​at grid points are calculated using weighted linear interpolation. The formula is as follows: ， in, Represented as weights; Weight Distance The monotonically decreasing function is given by the formula: ， in, For smoothing parameters; S32: Calculate the confidence score based on the weight information of the observation stations participating in the weighted average. And map the confidence scores to a set of values ​​according to a preset normalization rule. interval; S33: Perform steps S31 and S32 on multiple meteorological elements respectively to obtain multi-channel meteorological element grids and confidence maps.

4. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 1, characterized in that, The specific process of step S4 is as follows: S41: Input the radar echo sequence obtained in step S1 into the deterministic trend prediction network to obtain the radar trend prediction field at multiple future times. S42: Construct a residual sequence based on the real future radar echo and the radar trend prediction field obtained in step S41, and use the residual sequence as the learning target to train the residual conditional diffusion model.

5. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 4, characterized in that, The deterministic trend prediction network in step S41 includes an encoder module, a time mixing module, and a decoder module, with the time mixing module positioned between the encoder module and the decoder module.

6. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 5, characterized in that, The specific process of step S41 is as follows: S411: Arrange the radar echo sequence obtained in step S1 in chronological order to obtain a continuous sequence. A sequence of radar echo frames, each frame of which is represented as a single-channel echo map. ; S412: Transfer each frame Stacking along the time dimension yields the input tensor. , , For time steps, For single-channel radar echo, , These are the space dimensions; S413: Encoder module includes Level 1 convolutional downsampling unit, the first Level downsampling unit for features from the previous layer Spatial downsampling and convolution operations are performed sequentially to output features. The formula is: ， ， in, Indicates the first Level convolution operation; Indicates spatial downsampling operation; S414: The final feature output by the encoder module from the time mixing module. Applying convolutional aggregation along the time dimension yields temporally blended features. The formula is: ， in, For time indexing; This is the relative time offset; These are the weights of the temporal convolution kernel; The time convolution radius; This represents a two-dimensional convolution operation in the spatial dimension; S415: Decoder module includes Level upsampling unit, the first Level upsampling unit for features from the previous layer Upsampling ,Will Features corresponding to the level in the encoder Feature concatenation is performed, and output features are obtained through convolution operations. The formula is: ， ， in, ; Indicates the first Level-space upsampling operation; Indicates feature concatenation operation; Indicates the first Level convolution operation; S416: The last layer of the decoder module passes through... Convolution operation, outputting the future Radar trend prediction field at each time step ,in , To predict the number of time steps.

7. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 6, characterized in that, The loss function of the deterministic trend prediction network in step S41 includes a basic prediction error term and multiple physical constraint terms. These physical constraint terms constrain temporal smoothing, spatial smoothing, and overall intensity variation. Specifically, each physical constraint term is a combination of at least one of the following: total variation, adjacent-time difference penalty, and overall intensity difference penalty. The loss function of the deterministic trend prediction network... The formula is: ， in, These are non-negative weighting coefficients; This is the pixel-level mean square error term between the trend prediction field and the target prediction radar echo; This is a range constraint term used to penalize the portion of the predicted value that is less than zero or greater than a preset upper limit; This is a time smoothing constraint term used to penalize excessive differences between adjacent forecast times; This is a constraint term for total spatial variation, used to suppress isolated spatial noise and severe spatial oscillations; This is an overall intensity constraint term used to penalize abnormal changes in overall reflectivity intensity between adjacent forecast times.

8. Precipitation proximity based on trend-to-detail co-residual conditional diffusion as described in claim 7 Forecasting method, characterized in that, The specific process of step S42 is as follows: S421: Real Future Radar Echoes With radar trend prediction field The difference as residual The formula is: S422: Residual calculated based on step S421 Construct the residual sequence, and obtain the residual domain based on the residual sequence; S423: Define a forward noise addition process in the residual domain, using the original residuals. Starting from, through Time step on residual Gaussian noise is injected step by step, with the conditional distribution at each step as follows: ， The equivalent reparameterized form, i.e., the first The noise addition process is as follows: ， in, ; To schedule noise levels in advance; To follow a standard normal distribution Random noise; Let be the identity matrix, representing independent noise addition in each dimension; S424: From the forward process, we can obtain... Regarding the original residual Closed expression: ， That is: ， in, , indicating the first Signal retention rate at each step; , indicating the preceding The cumulative signal retention rate of the step; S425: Constructing a residual diffusion generation network Generate a network by diffusion to residuals Input noise residual Time step index and conditional features Residual diffusion generation network Output noise estimation ; S426: Employing weighted diffusion loss The residual diffusion generation network is trained with the following loss form: ， in, To keep pace with time The relevant non-negative weighting coefficients; Represents the variable and The expected operation.

9. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 8, characterized in that, In step S4, when jointly training the deterministic trend prediction network and the residual conditional diffusion model, the total loss function is... for: ， in, For deterministic losses; , These are non-negative weighting coefficients.

10. The precipitation nowcasting method based on trend-to-detail co-residual conditional diffusion according to claim 1, characterized in that, The forecasting process in step S5 adopts a segmented autoregressive sampling strategy. Specifically, the forecasting process is divided into several continuous forecast segments. For each forecast segment, trend prediction and residual generation are performed. Based on the residual and trend prediction, the forecasting result of the current segment is obtained. The forecasting result of the current segment is incorporated into the radar echo sequence as the input for the next forecast segment, until the entire forecasting process is completed.

Images