Short and temporary rainfall prediction method

Abstract

The invention discloses a short-term rainfall prediction method, and relates to the technical field of short-term rainfall prediction.The short-term rainfall prediction method comprises the steps that text features are extracted from a radar echo map in a partitioned mode, semantic information implied in the echo map is fully mined, multi-modal information fusion is achieved in combination with image features, the expression ability of a model for the echo features is improved, and the prediction accuracy of the short-term rainfall is improved. The prediction precision can be improved; a large language model is introduced to learn historical text features and predict future text features, semantic guidance is provided for LSTM prediction, the capture ability of the model for echo sequence semantic association is enhanced, and the generalization ability of the model in a complex weather scene is improved; the text similarity is added into the loss function, the difference of pixel values is considered, the consistency of predicted echoes and real echoes in the semantic level is guaranteed, and the accuracy and reliability of the prediction result are further improved.

Description

Technical Field

[0001] This invention belongs to the field of short-term rainfall prediction technology, specifically a short-term rainfall prediction method. Background Art

[0002] Short-term rainfall forecasting is crucial for disaster prevention and mitigation. Radar echo extrapolation, as a mainstream method for short-term forecasting, primarily employs an architecture that integrates convolutional neural networks (CNNs) and recurrent neural networks (RNNs / LSTMs). A typical implementation involves inputting historical radar echo sequences, extracting spatial features using CNNs, and then modeling temporal dependencies using LSTMs to predict future echo maps. However, this approach has significant drawbacks. The model over-relies on pixel-level low-level features and fails to integrate spatial semantic descriptions of the echoes (such as the distribution of strong echo locations and regional motion characteristics), weakening the physical plausibility of the prediction results. Finally, existing methods severely underutilize multimodal prior knowledge, particularly lacking the integration and modeling of textual modal features (such as radar echo location descriptions and regional dynamic evolution patterns), limiting the model's ability to understand complex meteorological evolution mechanisms. These shortcomings collectively result in prediction accuracy and practicality that fail to meet the needs of disaster prevention decision-making.

[0003] The application "A Cascaded LSTM Short-Term Rainfall Prediction Method Based on Attention Mechanism" (Application Announcement No. CN116381690A) discloses a cascaded LSTM short-term rainfall prediction method based on attention mechanism, comprising: inputting a denoised graph dataset into a SAC-LSTM model to obtain a predicted radar echo map dataset at time t+1; comparing the loss of the real graph dataset and the predicted graph dataset at time t+1 to obtain a first comparison result; inputting the real radar echo map dataset and the predicted radar echo map dataset at time t+1 into a generative adversarial network to obtain a first feature map dataset and a second feature map dataset; comparing the loss of the first feature map dataset and the second feature map dataset to obtain a second comparison result; updating the loss function of the SAC-LSTM model, using the updated SAC-LSTM model to predict the current input graph dataset and then performing a ZR transform; and using the transform result to predict rainfall. This design is mainly based on image features, and does not make full use of the implicit textual semantic information in radar echo images. The feature mining of the data is not comprehensive enough, and no external knowledge is introduced to assist in prediction, resulting in limited generalization ability of the model in complex weather scenarios. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a short-term rainfall prediction method, comprising: The radar echo map is divided into several grid partitions with semantic labels according to spatial location; Areas with echo values ​​exceeding a preset value are marked as high echo regions, and text features describing the high echo distribution are generated; then, the text features are encoded to generate text feature vectors. Extract image features from radar echo images and fuse these image features with text feature vectors; The fused feature sequences are input into an LSTM network for spatiotemporal correlation learning and rainfall prediction. Construct a text feature sequence of historical radar echo sequences, input the historical text feature sequence into a pre-trained large language model, predict future text features, and input them into an LSTM network to provide semantic guidance; Divide the dataset, train the model using the training set with a pre-designed loss function, and evaluate the prediction results using the test set.

[0005] Furthermore, the acquired radar echo image dataset needs to be preprocessed before dividing it into grid partitions.

[0006] Furthermore, the grid is divided into 9×9 grids, and the semantic labels include upper left, upper middle, upper right, middle left, middle, middle right, lower left, lower middle, and lower right.

[0007] Furthermore, the high echo value in the high echo region exceeds 10%, which is defined as an echo value exceeding 40 dBZ.

[0008] Furthermore, the encoding method for text features is as follows: Perform word segmentation on the text; The word segmentation results are converted into feature vectors using a word vector model. The calculation formula is as follows: .

[0009] Furthermore, the loss function is designed as follows: Calculate the mean square error between the predicted radar echo and the actual echo; Calculate the cosine similarity between the feature vectors of the real and predicted texts; The total loss function is obtained by weighting the mean squared error and the cosine similarity.

[0010] Furthermore, the cosine similarity is calculated using the following formula: ; In the formula, , These are the real and predicted text feature vectors, respectively.

[0011] Furthermore, the total loss function is: ; In the formula, The weight parameters are dynamically adjusted using the validation set; L MSE This represents the mean square error.

[0012] Furthermore, the pre-trained large language model is a GPT series model, which is fine-tuned to predict future text features.

[0013] Furthermore, convolutional neural networks are used to extract image features from radar echo maps, and feature stitching or attention mechanisms are employed for the fusion operation.

[0014] Compared with the prior art, the present invention has the following beneficial effects: By extracting text features from radar echo image partitions, the implicit semantic information in the echo images is fully explored. Combined with image features, multimodal information fusion is achieved, which improves the model's ability to express echo features and helps to improve prediction accuracy. The introduction of a large language model to learn historical text features and predict future text features provides semantic guidance for LSTM prediction, enhances the model's ability to capture semantic associations in echo sequences, and improves the model's generalization ability in complex weather scenarios. By incorporating text similarity into the loss function, not only are pixel value differences considered, but the semantic consistency between predicted and actual echoes is also ensured, further improving the accuracy and reliability of the prediction results. By combining the architecture of LSTM and a large language model, the processing capabilities of recurrent neural networks for temporal features and the semantic understanding capabilities of large language models are fully utilized to achieve joint learning of the spatiotemporal and semantic features of radar echo sequences, providing a more comprehensive technical solution for short-term rainfall prediction. Attached Figure Description

[0015] Figure 1 This is a structural diagram of the model of the present invention; Figure 2 This invention describes the prediction process for the CIKM AnalytiCup 2017 competition dataset. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-2 As shown, this application provides a method for predicting short-term rainfall, including: Step 1: Obtain the radar echo image dataset and perform preprocessing such as missing value handling and cleaning; Step 2: Extract text features from the radar echo image by region. The specific steps are as follows: Step 2.1: Divide the radar echo image into a 9×9 grid according to spatial location, and attach semantic information to each grid, namely: upper left, upper middle, upper right, middle left, middle, middle right, lower left, lower middle and lower right; Step 2.2: Calculate the proportion of high echo values ​​within each region. High echo values ​​refer to echo values ​​greater than 40 dBZ. When the proportion of high echo values ​​in a region is greater than 10%, it is considered a high echo region. Text features are extracted in this way. The text example is: "High echo regions are mainly distributed in the Upper left, Uppermiddle, Upper right, Middle left, Middle and Lower left of the image." Step 2.3: Encode the text features obtained from the partitioning to generate text feature vectors. First, segment the text into words to obtain [“High”, “echo”, “regions”, “are”, “mainly”, “distributed”, “in”, “the”, “Upper”, “left”, “,”, “Upper”, “middle”, “,”, “Upper”, “right”, “,”, “Middle”, “left”, “,”, “Middle”, “,”, “and”, “Lower”, “left”, “of”, “the”, “image”, “.”]. Then, use a word vector model to convert it into feature vectors, using the following formula: ; Step 3: Input the radar echo image as the image feature into the convolutional network to extract image features, and fuse the text feature vector obtained in Step 2 with the image features. Step 4: Build an LSTM prediction network, input the fused feature sequence into the LSTM network, learn the spatiotemporal correlation, and perform rainfall prediction. Step 5: Perform text feature prediction using a large language model. The specific steps are as follows: Step 5.1: Collect text features from historical radar echo sequences and construct a text feature sequence; Step 5.2: Input the historical text feature sequence into a pre-trained large language model (such as the GPT series) to predict the text features of future echo sequences; Step 5.3: Input the future text features predicted by the large language model into the LSTM network to provide semantic guidance for subsequent predictions; Step 6: Design the loss function, as follows: Step 6.1: Calculate the mean square error (MSE) between the predicted radar echo and the actual echo. The formula is: , Where n is the number of samples. It is the ground truth value. It is a predicted value; Step 6.2: Calculate the cosine similarity (text similarity) between the true echo text features and the predicted echo text features. The formula is: , in , These are the real and predicted text feature vectors, respectively; Step 6.3, the total loss function is: ; in These are weight parameters; Step 7: Divide the dataset into training set, validation set and test set according to a certain ratio; Step 8: Train the model using the training set and optimize the model parameters through backpropagation, including LSTM network parameters, large language model fine-tuning parameters, and loss function weights, etc. Step 9: Use the test set to evaluate the trained model, obtain the prediction results, and complete the rainfall prediction.

[0018] In practical applications, this application is as follows: Taking the CIKM AnalytiCup 2017 competition dataset as an example, this dataset was collected and cleaned by the Shenzhen Meteorological Bureau, spanning three years, with the data from the first two years used for training and the data from the third year used for testing. Each radar image in the dataset covers the radar reflectivity of a 101 km × 101 km area around the target site. The radar map sequence contains 15 time spans, spaced 6 minutes apart, measured at 4 different altitudes, from 0.5 km to 3.5 km, spaced 1 km apart, providing 60 historical radar images for one sample. Our task is to predict the next 5 images for the following half hour based on the first 10 images from the 15 radar images at the same altitude for each sample.

[0019] The specific application process is as follows: Each radar image is divided into a 9×9 grid. Text features are extracted from each echo image using pixel calculations and encoded as a 128-dimensional vector. Simultaneously, image features are extracted using a convolutional network, fused with text features, and then input into an LSTM network. Text features from the first 5 echo frames are collected and input into a pre-trained large language model to predict text features for the next 10 frames. The text features predicted by the large language model are then input into the LSTM network, combined with image features, to predict echoes for the next 10 frames. The loss function combines MSE and text similarity, with weights... Set it to 100, and adjust the parameter using the validation set during training.

[0020] The specific changes in the feature map during the model prediction process are as follows: Figure 2 As shown.

[0021] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for predicting short-term rainfall, characterized in that, include: The radar echo map is divided into several grid partitions with semantic labels according to spatial location; Areas with echo values ​​exceeding a preset value are marked as high echo regions, and text features describing the high echo distribution are generated; then, the text features are encoded to generate text feature vectors. Extract image features from radar echo images and fuse these image features with text feature vectors; The fused feature sequences are input into an LSTM network for spatiotemporal correlation learning and rainfall prediction. Construct a text feature sequence of historical radar echo sequences, input the historical text feature sequence into a pre-trained large language model, predict future text features, and input them into an LSTM network to provide semantic guidance; Divide the dataset, train the model using the training set with a pre-designed loss function, and evaluate the prediction results using the test set.

2. The method for predicting short-term rainfall according to claim 1, characterized in that, Before dividing the data into grid partitions, the acquired radar echo image dataset needs to be preprocessed.

3. The method for predicting short-term rainfall according to claim 1, characterized in that, The grid is divided into 9×9 grids, and the semantic labels include upper left, upper middle, upper right, middle left, middle, middleright, lower left, lower middle, and lower right.

4. The method for predicting short-term rainfall according to claim 1, characterized in that, The high echo value in the high echo region exceeds 10%, which is defined as an echo value exceeding 40 dBZ.

5. The method for predicting short-term rainfall according to claim 1, characterized in that, The encoding method for text features is as follows: Perform word segmentation on the text; The word segmentation results are converted into feature vectors using a word vector model. The calculation formula is as follows: 。 6. The method for predicting short-term rainfall according to claim 1, characterized in that, The loss function is designed as follows: Calculate the mean square error between the predicted radar echo and the actual echo; Calculate the cosine similarity between the feature vectors of the real and predicted texts; The total loss function is obtained by weighting the mean squared error and the cosine similarity.

7. The method for predicting short-term rainfall according to claim 6, characterized in that, Cosine similarity is calculated using the following formula: ； In the formula, , These are the real and predicted text feature vectors, respectively.

8. A short-term rainfall prediction method according to claim 6, characterized in that, The total loss function is: ； In the formula, The weight parameters are dynamically adjusted using the validation set; L MSE This represents the mean square error.

9. The method for predicting short-term rainfall according to claim 1, characterized in that, The pre-trained large language model is a GPT series model, which is fine-tuned to predict future text features.

10. A short-term rainfall prediction method according to claim 1, characterized in that, Image features of radar echo maps are extracted using convolutional neural networks, and feature stitching or attention mechanisms are used for fusion operations.

Images