A method and apparatus for predicting raindrop spectral distribution based on time series networks

Abstract

This application relates to a method and apparatus for predicting raindrop spectral distribution based on a time-series network. The method includes: acquiring historical raindrop spectral observation data to obtain time series data of shape factor, normalized intercept, and mass-weighted average diameter; quantitatively evaluating the predictability and interrelationships of each parameter using information theory analysis; constructing a two-branch time-series network based on the evaluation results, where the first branch independently predicts the highly complex shape factor, and the second branch jointly predicts the strongly correlated normalized intercept and mass-weighted average diameter; and finally, performing prediction using the trained network. This method achieves more accurate and physically consistent time-series predictions of raindrop spectral parameters through information theory-guided heterogeneous network design and the construction of a physically constrained loss function.

Description

Technical Field

[0001] This application relates to the fields of meteorological forecasting and artificial intelligence technology, and in particular to a method and apparatus for predicting raindrop spectral distribution based on time series networks. Background Technology

[0002] Accurate precipitation prediction is crucial for disaster prevention and mitigation, water resource management, and agricultural production. As a core operational technology, radar quantitative precipitation estimation relies heavily on the accurate description of raindrop microphysical properties, with the raindrop size distribution (DSD) being key to characterizing the microphysical structure of precipitation. The DSD is typically characterized by several key parameters, such as the shape factor μ and the normalized intercept parameter. (usually its logarithm is taken) ) and mass-weighted average diameter .

[0003] In recent years, deep learning technology has made significant progress in DSD parameter inversion, enabling efficient estimation of DSD parameters from multi-source observation data such as radar. However, most existing research focuses on static parameter inversion, with relatively few studies on temporal prediction of future changes in DSD parameters. Only a few studies have attempted to apply temporal models such as Long Short-Term Memory networks to DSD parameter prediction, but these methods typically involve all parameters (μ, μ, μ) in a single step. , Treating time series as independent and forecasting them simultaneously ignores the physical dependencies between parameters. This independent forecasting strategy may not only violate microphysical constraints, leading to physically unreasonable forecasts, but also fails to differentiate the modeling of the different predictability of different parameters on the time series. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and apparatus for predicting raindrop spectral distribution based on time series networks to address the aforementioned technical problems.

[0005] Firstly, this application provides a method for predicting raindrop spectral distribution based on time series networks. The method includes:

[0006] Raindrop spectrum observation data of the target area within a historical preset time period are obtained to obtain a continuous time series of raindrop spectrum parameters; the raindrop spectrum parameters include mass-weighted average diameter, normalized intercept parameter and shape factor.

[0007] Based on information theory analysis methods, the predictability of the time series of the raindrop spectral parameters is evaluated and the inter-parameter dependencies are analyzed to obtain the analysis results.

[0008] Based on the analysis results, a two-branch time series network is constructed; wherein, the first branch of the two-branch time series network is used to predict the shape factor, and the second branch is used to jointly predict the normalized intercept and the mass-weighted average diameter.

[0009] The dual-branch time series network is trained using samples constructed from the time series of the raindrop spectral parameters to obtain the trained dual-branch time series network.

[0010] The raindrop spectral parameters time series to be predicted are input into a trained dual-branch time series network, which outputs predicted values ​​of raindrop spectral parameters for a future preset time period.

[0011] In one embodiment, the information theory analysis method includes permutation entropy analysis and cross-correlation analysis. Specifically, based on the results of the permutation entropy analysis, it is determined that the temporal complexity of the shape factor μ is higher than that of the mass-weighted average diameter and the normalized intercept. Based on the results of the cross-correlation analysis, a quantitative negative correlation is determined between the normalized intercept and the mass-weighted average diameter.

[0012] In one embodiment, prior to constructing the dual-branch time series network, the method further includes: performing monthly analysis on the time series of the raindrop spectral parameters to extract seasonal physical features of the raindrop spectral parameters. Each branch of the dual-branch time series network includes a seasonal feature fusion module, used to encode the seasonal physical features into learnable embedding vectors and fuse them with standard positional encodings to generate time series features that incorporate seasonal physical laws as input to the dual-branch time series network.

[0013] In one embodiment, the second branch includes an intervariate interaction module. This intervariate interaction module models the correlation between the normalized intercept and the mass-weighted average diameter through feature concatenation and linear transformation.

[0014] In one embodiment, the step of training the dual-branch time series network includes: employing independently optimized training strategies for the first branch and the second branch. Specifically, for the prediction target of the first branch, a two-stage training strategy is used: first, pre-training is performed using mean squared error loss, then switching to a loss function weighted by extreme values ​​for fine-tuning. For the joint prediction target of the second branch, end-to-end training is employed, using a loss function that integrates a high-value weighting mechanism.

[0015] In one embodiment, the loss function of the first branch is:

[0016] ;

[0017] in, The time series representing the raindrop spectral parameters True values ​​of shape factors, The time series representing the raindrop spectral parameters One shape factor predicted value, This represents the total number of shape factors in the time series of raindrop spectral parameters. The time series representing the raindrop spectral parameters An extreme value attention coefficient is defined when the true value of the shape factor is at a preset normal value. Set to 1, when the true value of the shape factor is at a preset extreme value, the extreme value attention coefficient is... Set to be significantly greater than 1.

[0018] In one embodiment, the loss function of the second branch is:

[0019] ;

[0020] in, The time series representing the raindrop spectral parameters The true value of the mass-weighted average diameter. The time series representing the raindrop spectral parameters A mass-weighted average diameter prediction value, The first term represents the time series of raindrop spectral parameters with respect to the mass-weighted average diameter. One error penalty coefficient, The normalized intercept is the error penalty coefficient when the true value of the mass-weighted average diameter is greater than a preset value. The error penalty coefficient is set to be greater than 1 when the true value of the mass-weighted average diameter is less than or equal to a preset value. Set to 1.

[0021] In one embodiment, the time series feature extractors of the first branch and the second branch in the dual-branch time series network both employ gated recurrent unit networks.

[0022] Secondly, this application provides a raindrop spectrum distribution prediction device based on time series networks, the device comprising:

[0023] The historical data acquisition module is used to acquire raindrop spectrum observation data of the target area within a preset historical time period, and obtain a continuous time series of raindrop spectrum parameters.

[0024] The dependency analysis module is used to perform predictability assessment and inter-parameter dependency analysis on the time series of the raindrop spectral parameters based on information theory analysis methods, and obtain analysis results.

[0025] The network construction module is used to construct a two-branch time series network based on the analysis results.

[0026] The training module is used to train the dual-branch time series network using samples constructed from the time series of the raindrop spectral parameters, so as to obtain the trained dual-branch time series network.

[0027] The prediction module is used to input the time series of raindrop spectral parameters to be predicted into a trained dual-branch time series network and output the predicted values ​​of raindrop spectral parameters for a future preset period.

[0028] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the above embodiments.

[0029] The aforementioned raindrop spectral distribution prediction method and apparatus based on time-series networks quantitatively assesses the predictability of each DSD parameter and the dependencies between parameters by introducing information theory analysis. Based on this quantitative analysis, the design of the neural network architecture is guided, constructing independent prediction branches for highly complex parameters and joint prediction branches for strongly correlated parameter pairs. This heterogeneous dual-branch architecture can more effectively capture the characteristics of different parameters. This method overcomes the homogenization of prediction strategies and the neglect of physical constraints in existing technologies, thereby achieving more accurate and physically consistent time-series predictions of DSD parameters, providing an effective microphysical process prediction tool for improving quantitative precipitation estimation and forecasting. Attached Figure Description

[0030] Figure 1 This is an application scenario diagram of a raindrop spectrum distribution prediction method based on time series networks in one embodiment;

[0031] Figure 2 This is a flowchart illustrating a raindrop spectral distribution prediction method based on time series networks in one embodiment.

[0032] Figure 3 This is a structural block diagram of a raindrop spectrum distribution prediction device based on a time series network in one embodiment;

[0033] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0035] The raindrop spectral distribution prediction method based on time series networks provided in this application can be applied to, for example... Figure 1 In the application environment shown, multiple raindrop spectrum observation stations (such as OTT Parsivel²) 102 deployed in the target area continuously collect data on the size and velocity distribution of precipitation particles. This raw observation data is transmitted via a network to a computing device 104, which can be a data processing server or a local computing terminal. The computing device 104 performs quality control and parameter calculations on the raw data to obtain continuous DSD parameters (μ, , The time series is then processed. The computing device 104 executes the method of this application: based on information theory analysis, it constructs and trains a two-branch time series prediction network, and uses the trained model to predict the DSD parameters for future periods. The prediction results can be further used to drive quantitative precipitation estimation algorithms or to provide decision support for weather forecasters. The local computing terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, workstations, or servers. The data processing server can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0036] In one embodiment, such as Figure 2 As shown, a method for predicting raindrop spectral distribution based on time series networks is provided, and this method is applied to... Figure 1 Taking a computing device as an example, the explanation includes the following steps:

[0037] Step 202: Obtain raindrop spectrum observation data of the target area within a historical preset time period to obtain a continuous time series of raindrop spectrum parameters.

[0038] Specifically, raindrop spectrometer observation data for a target region over several years are collected. The target region could be an island affected by a monsoon climate, and the data can be two, four, or five years' worth of observations, with no specific restrictions. Rigorous quality control is performed on the raw observation data, including removing low signal-to-noise ratio data, filtering physically unreasonable raindrop velocities, and excluding samples with insufficient rainfall intensity or raindrop count. Then, data is extracted from consecutive precipitation events at a fixed high temporal resolution, which can be selected as one minute. Three key DSD parameters for each time point are calculated using the moment method: shape factor μ, normalized intercept parameter... (and take its common logarithm) ) and mass-weighted average diameter Ultimately, a high-quality μ over a long period of time is obtained. , Time series dataset.

[0039] Step 204: Based on information theory analysis methods, predictability assessment and inter-parameter dependency analysis are performed on the time series of raindrop spectral parameters to obtain analysis results.

[0040] In existing technologies, the prediction of DSD parameters often lacks a quantitative assessment of the predictability of each parameter and their interrelationships, leading to blind model design. This step aims to systematically address this problem. A multivariate information theory framework is adopted, primarily including the following analyses:

[0041] Permutation entropy analysis: used to quantify the complexity and regularity of time series. The lower the permutation entropy value, the more regular the sequence and the higher its predictability. This is achieved by analyzing μ, ... , The permutation entropy of each of the three parameters is calculated separately, allowing for a quantitative comparison of their predictability. Analysis typically shows that μ has the highest permutation entropy, indicating that its temporal behavior is the most complex and difficult to predict. The permutation entropy is relatively the lowest, and the predictability is the highest; In between.

[0042] Detrending volatility analysis: used to detect whether time series exhibit long-term memory effects, i.e., the persistent influence of historical states on the future. By calculating the scaling exponent α, if α > 1, it indicates that the series has long-term memory, providing a theoretical basis for using models such as recurrent neural networks that excel at capturing long-term dependencies. Analysis typically confirms μ, , All three sequences are persistent.

[0043] Cross-correlation analysis: used to quantify the linear correlation between different parameters and their potential lag effects. Calculation and μ and The cross-correlation coefficient between pairs of equal parameters. The key finding is... and There is a significant quantitative negative correlation between μ and other parameters (e.g., correlation coefficient r≈-0.777), while μ has a weaker correlation with other parameters. This reveals a strong physical constraint between the parameters.

[0044] Conditional entropy analysis (optional but beneficial): This further quantifies the uncertainty of one parameter given another parameter from an information theory perspective, thereby identifying the optimal predictor. For example, the analysis might show that, given... Under the conditions, The uncertainty has been significantly reduced, that is yes A strong predictor.

[0045] These quantitative analysis results provide direct and objective guidance for the differentiated design of subsequent network architectures.

[0046] Step 206: Based on the analysis results, construct a dual-branch time series network.

[0047] Based on the quantitative analysis conclusions of step 204, we abandon the traditional single-branch prediction architecture that treats all parameters equally and design a heterogeneous two-branch time series network. Its core design logic is as follows:

[0048] Branch 1 (μ Branch): Since permutation entropy analysis shows that μ has the highest temporal complexity and relatively independent behavior (weak correlation with other parameters), a separate prediction branch is designed for it. This branch is specifically responsible for learning the complex temporal evolution pattern of μ itself.

[0049] Branch Two ( - Joint branch): Since both cross-correlation and conditional entropy analysis show... and There is a strong negative correlation between them, and right Since these two parameters contribute significantly to the information, they are jointly modeled within the same branch. This branch explicitly learns and utilizes this information through a specific intervariate interaction module. and The correlation between them makes Predictive features can assist The predictions are accurate, and vice versa, thus ensuring the physical consistency of the prediction results.

[0050] The two branches work in parallel to jointly predict the three parameters. Each branch typically contains a temporal feature extractor, necessary intermediate layers, and an output layer.

[0051] Step 208: Use the samples constructed from the time series of raindrop spectral parameters to train the dual-branch time series network, and obtain the trained dual-branch time series network.

[0052] First, the continuous DSD parameter time series is preprocessed, including constructing historical observation-future prediction sample pairs using the sliding window method. For example, using data from the past 24 minutes to predict data from the next 12 minutes, and dividing the data into training, validation and test sets in chronological order. All parameters need to be normalized.

[0053] Secondly, a training strategy was designed. Since the two branches have different task characteristics, an independent optimization strategy was adopted:

[0054] For the μ branch, a two-stage training approach is adopted: the first stage uses standard mean squared error loss for pre-training, enabling the model to initially grasp the basic variation law of μ; the second stage switches to an extreme value weighted loss function tailored for μ for fine-tuning. This loss function assigns higher penalty weights to rare but important extreme values ​​(extremely low or extremely high values) in the true value of μ, forcing the model to better learn the long-tail characteristics of the right-skewed distribution.

[0055] for - The joint branch employs end-to-end training and uses a joint loss function. This joint loss function comprises at least two parts: one is a set of... The predicted loss can be in a standard or weighted form, or secondly, for... The predicted high-value weighted loss function. (For...) The distribution also exhibits a right-skewed characteristic; therefore, this loss function is more effective for predicting larger values. The samples are given higher weights to improve the model's ability to predict large raindrop events.

[0056] The two branches perform forward propagation synchronously during training, but backpropagation and parameter updates can be performed independently, each with its own independent optimizer and learning rate scheduler.

[0057] Step 210: Input the time series of raindrop spectral parameters to be predicted into the trained dual-branch time series network, and output the predicted values ​​of raindrop spectral parameters for the future preset time period.

[0058] In the application phase, the pre-processed DSD parameter time series to be predicted is input into the pre-trained dual-branch network. The network's forward propagation process simultaneously computes data through both branches, ultimately outputting μ for a future time period. and The predicted value sequence. These predicted DSD parameters can be directly used to select or adjust the optimal radar quantitative precipitation estimation algorithm, thereby generating more accurate quantitative precipitation estimates or short-term forecast products.

[0059] In the aforementioned raindrop spectral distribution prediction method based on time series networks, information theory tools such as permutation entropy and cross-correlation are introduced to systematically and quantitatively evaluate the predictability and interrelationships of DSD parameters for the first time. Based on this evaluation, a heterogeneous dual-branch network architecture is innovatively designed, independently modeling the highly complex and weakly correlated parameter μ, while modeling the strongly correlated parameter pairs (…). , This approach employs joint modeling and introduces a cross-variable interaction mechanism. Simultaneously, by designing a physical constraint loss function that integrates data distribution characteristics, it guides the model to more accurately learn the physical laws of the real world. From the guiding principles of model design to the specific architecture and optimization objectives, this method closely revolves around the physical and statistical characteristics of the DSD parameters themselves, overcoming the shortcomings of existing methods that rely on a single prediction strategy and neglect physical constraints. This results in higher accuracy and physical consistency in DSD parameter time series prediction.

[0060] In one embodiment, the information theory analysis method includes: permutation entropy analysis and cross-correlation analysis; wherein, based on the results of permutation entropy analysis, it is determined that the temporal complexity of the shape factor is higher than that of the mass-weighted average diameter and the normalized intercept; and based on the results of cross-correlation analysis, it is determined that there is a quantitative negative correlation between the normalized intercept and the mass-weighted average diameter.

[0061] In this embodiment, the information theory analysis method in step 204 specifically includes permutation entropy analysis and cross-correlation analysis. Permutation entropy is calculated based on the ordinal pattern distribution of the time series, and its value directly reflects the randomness and complexity of the sequence. Calculations using five years of observation data from a certain island revealed that the average permutation entropy of the shape factor μ is the highest, indicating that its temporal variation pattern is the most irregular and its predictability is the lowest; while the normalized intercept... and mass-weighted average diameter The permutation entropy is relatively low, especially The predictability is the highest. This quantitative result provides an objective basis for why special treatment of μ is necessary. Cross-correlation analysis reveals the linear dependence between different parameters by calculating the correlation coefficients of different parameters at different time lags. The analysis results show that... and A strong negative correlation exists near zero hysteresis, which physically reflects the inverse relationship between particle concentration and average size in the raindrop spectrum. And μ and... or The correlation is much weaker. This finding directly supports the conclusion that... and Design decisions that involve joint modeling within the same branch are better made because joint modeling naturally preserves and utilizes this strong correlation, while independent predictions may destroy it.

[0062] In one embodiment, before constructing the dual-branch time series network, the method further includes: performing monthly analysis on the time series of raindrop spectral parameters to extract seasonal physical features of the raindrop spectral parameters; each branch of the dual-branch time series network contains a seasonal feature fusion module, which encodes the seasonal physical features into learnable embedding vectors and fuses them with standard positional encodings to generate time series features that fuse seasonal physical laws as input to the dual-branch time series network.

[0063] In this embodiment, before constructing the dual-branch time series network in step 206, there is an important preliminary step: monthly statistical analysis of the time series of raindrop spectral parameters to extract seasonal physical characteristics characterizing the raindrop spectral parameters. Monthly analysis of data from a certain island reveals the seasonal cyclical variation characteristics of the DSD parameters: during the East Asian winter monsoon (approximately November to March of the following year), μ and... The values ​​for are relatively high, while the values ​​for Dm are low, which corresponds to stratiform precipitation with small and dense raindrops; during the East Asian summer monsoon (typhoon season, approximately June to September), the opposite mode is observed. Reaching the maximum, μ and The lower value reflects the characteristic of strong convective precipitation producing large but few raindrops. To incorporate this important climatological law into the model, a seasonal feature fusion module is integrated into each branch of the constructed two-branch time series network. This module operates as follows: First, based on the month corresponding to the input sample, it searches and extracts the corresponding seasonal embedding vector from a learnable embedding vector table (e.g., setting different learnable vectors for the summer monsoon, winter monsoon, and transitional periods). Then, it fuses this seasonal embedding vector with a standard sinusoidal positional code used to represent the absolute and relative positions of elements in the sequence, thereby generating an enhanced positional code that contains both temporal positional information and seasonal climatic background. This fused feature serves as input to time series encoders such as GRU, enabling the model to learn to adopt different prediction patterns in different seasons, for example, focusing more on [specific prediction patterns] in summer. The high predicted values ​​improve the model's ability to adapt to climate change.

[0064] In one embodiment, the second branch includes: a cross-variable interaction module; the cross-variable interaction module models the correlation between the normalized intercept and the mass-weighted average diameter through feature concatenation and linear transformation.

[0065] Specifically, the second branch contains a key intervariate interaction module. This module is designed to explicitly model the dependencies between parameters revealed by the cross-correlation analysis in step 204. Its specific implementation could be as follows: the GRU encoder in the branch extracts... and After extracting the initial temporal latent features of each feature group, these features are concatenated along the feature dimension and then interactively fused through one or more linear transformation layers. This linear transformation process learns how to extract relevant information from the concatenated features and redistribute it to specific applications. and On a separate decoding path. For example, this module can learn when When the characteristic indicates that its value will increase, it should be adjusted accordingly. The predicted features are adjusted towards a lower value, thus maintaining the negative correlation constraint between the two at the predicted value level. This modular design allows the physical relationship between parameters to be structurally embedded in the network, rather than relying solely on the loss function for soft constraints. Ablation experiments show that removing this module leads to... The prediction accuracy decreased significantly, specifically with a 26.6% increase in root mean square error, while the prediction accuracy for... The predicted impact is small, proving that the module effectively utilizes... Information to improve Predictions are one of the core components that enable the advantages of joint modeling to be realized.

[0066] In one embodiment, the steps of training the dual-branch time series network include: employing an independent optimization training strategy for the first branch and the second branch; wherein, for the prediction target of the first branch, a two-stage training strategy is adopted, first using mean squared error loss for pre-training, and then switching to a loss function weighted by extreme values ​​for fine-tuning; for the joint prediction target of the second branch, end-to-end training is adopted, and a loss function integrating a high-value weighting mechanism is used.

[0067] Specifically, a differentiated training strategy was adopted for the two-branch network. For the first branch, since its target μ distribution is right-skewed and extreme values ​​are rare, directly using the mean squared error loss could easily lead the model to overemphasize the numerous common values ​​while neglecting extreme μ events that may be more important for disaster prevention and mitigation. Therefore, a two-stage training strategy was adopted: In the first stage, sufficient pre-training rounds were performed using the standard mean squared error loss to allow the model to grasp the basic pattern of μ changes. In the second stage, a specially designed weighted mean squared error loss function was switched for fine-tuning. For the second branch, since its task involves two parameters, and It also exhibits a right-skewed distribution, therefore an end-to-end training method is adopted.

[0068] In one embodiment, the loss function for the first branch is:

[0069] ;

[0070] in, The time series representing the raindrop spectral parameters indivual The truth value of , The time series representing the raindrop spectral parameters The predicted value of μ, In the time series representing raindrop spectral parameters The total number, The time series representing the raindrop spectral parameters The extreme value attention coefficient, when When the true value is within the preset normal value, the extreme value attention coefficient Set to 1, when When the true value is at a preset extreme value, the extreme value attention coefficient Set to be significantly greater than 1.

[0071] Specifically, it can be set when At (extremely low values), ,when (At extremely high values) ,when (Normal value) The extreme value attention coefficients of 25.0 and 30.0 are experimentally validated hyperparameters that effectively balance the learning intensity of common and extreme values. Through this explicit, asymmetric weighting scheme, the model generates a much larger gradient signal for incorrectly predicted extreme μ-value samples than for normal value samples when calculating gradients. This drives targeted adjustments to the model parameters, significantly improving the predictive ability for rare but potentially physically important extreme precipitation microphysical states. This loss function design is a classic example of transforming statistical data characteristics into model optimization constraints.

[0072] In another embodiment, the loss function for the second branch is:

[0073] ;

[0074] in, The time series representing the raindrop spectral parameters The true value of the mass-weighted average diameter. The time series representing the raindrop spectral parameters A mass-weighted average diameter prediction value, The first term represents the time series of raindrop spectral parameters with respect to the mass-weighted average diameter. One error penalty coefficient, This represents the normalized intercept, and the error penalty coefficient when the true value of the mass-weighted average diameter is greater than a preset value. Set to greater than 1, the error penalty coefficient is set when the true value of the mass-weighted average diameter is less than or equal to the preset value. Set to 1.

[0075] Specifically, when hour ;when hour This means that if the model predicts a high If a prediction is made with a high value but is incorrect, it will be penalized three times more than a normal error; if it predicts correctly, it will receive the same reward as a normal sample. This mechanism encourages the model to make high-value predictions while striving for accuracy, thus effectively improving the model's ability to capture large raindrop events and its prediction accuracy. This weighted strategy based on predicted values ​​complements the aforementioned weighted strategy based on true values, together forming a physical constraint training framework to cope with non-uniform data distributions.

[0076] In one embodiment, in the dual-branch time series network, both the first and second branches employ gated recurrent unit (GRU) networks as their temporal feature extractors. Specifically, GRU is a type of recurrent neural network, similar to Long Short-Term Memory (LSTM) networks. It effectively captures long-term dependencies in time series data through a gating mechanism, and is simpler in structure and has fewer parameters than LSTM, potentially offering higher training efficiency. The selection of GRU as the basic building block is based on the results of the detrended volatility analysis in step 204—this analysis confirms that μ, , All three time series exhibit long-term memory properties (scaling exponent α > 1.0). This means that the parameter values ​​at the current moment are influenced by their historical states over a long period, thus requiring a sequence model capable of effectively modeling long-range dependencies. GRU possesses this capability. In each branch, the GRU encoder receives location codes incorporating seasonal features and other possible inputs, processing them step-by-step. The final output is a context vector summarizing the historical information of the entire input sequence or the hidden state of the last time step. This state is then fed into subsequent decoding or prediction layers to generate parameter sequences for future periods. Using GRU ensures the model has the fundamental ability to handle the time-series memory effect of DSD parameters.

[0077] To further verify the effectiveness of the present invention, prediction experiments and ablation experiments were conducted as follows:

[0078] Model performance was quantified using Pearson correlation coefficient (R), root mean square error (RMSE), mean absolute error (MAE), and normalized root mean square error (NRMSE).

[0079] The optimized two-branch time series network model (HVTS-Net) was applied to the test set, and its performance was quantified using evaluation metrics. The two-branch time series network demonstrated good predictive performance on all three DSD parameters, with Pearson correlation coefficients all above 0.74 and RMSE values ​​ranging from 0.0252 to 0.0296. It exhibits the highest prediction accuracy (R=0.9315) and low error indices (RMSE=0.0290, MAE=0.0214). The prediction results for parameter μ are RMSE=0.0296 and MAE=0.0197, reflecting the complexity of μ. The RMSE of all parameters is below 5%. and The prediction results maintained their inherent negative correlation (r=-0.777) while maintaining high prediction accuracy, which verifies the ability of the dual-branch architecture to handle independent and interdependent parameter features.

[0080] To quantify the individual contributions of key components in the dual-branch time series network model and validate the design rationale, we conducted ablation experiments. Considering the dual-branch structure (joint...) of the DSD parameter prediction task... - The ablation experiments were conducted on key modules of both branches (prediction branch and independent μ prediction branch). Each ablation experiment maintained the same training strategy and hyperparameters to ensure the reliability of the comparisons. The performance difference between the two-branch time series network model and the ablation model quantified the contribution of each module to the prediction accuracy.

[0081] exist - In the joint prediction branch, the three modules each contribute different functionalities. Removing the temporal feature extraction module leads to a comprehensive performance degradation; this module originally extracted monsoon seasonal features by combining learnable embedding vectors with location encoding. The correlation index decreased significantly, and various errors increased substantially, demonstrating that encoding monsoon climate features as temporal features is a necessary condition for accurately predicting DSD parameters.

[0082] The physical information removal loss function significantly impacts both correlation and error metrics. This function addresses data imbalance by assigning higher weights to extreme values ​​and bimodal regions; its absence suggests that this weighting mechanism effectively helps the model learn the actual right skewness and bimodal characteristics present in the DSD parameter distribution.

[0083] The performance degradation caused by removing the parameter correlation modeling module is the most significant, and it severely impacts performance. The prediction accuracy, and The predictions showed only a slight degradation (root mean square error increased by 7.9%). This module originally modeled the correlation between parameters through feature concatenation and linear transformation. This result demonstrates that the correlation constraint strategy successfully utilized... Information enhancement The accuracy of predictions, when this constraint is removed, Predictions were severely damaged, and The forecast is almost unaffected:

[0084]

[0085] exist In the prediction branch, the ablation experiments showed different results. Removing the temporal coding module led to... The predicted correlation coefficient decreased from 0.7488 to 0.5082, while the normalized root mean square error increased by 5.7%. This indicates that time series modeling is crucial for maintaining high Pearson correlation. Removing the weighted loss function increased the root mean square error by 7.4%, demonstrating that without the weighting mechanism of the loss function, The branch cannot effectively learn the right-skewed μ distribution. The complete model maintains a balanced root mean square error to mean absolute error ratio across all parameters. , , ):

[0086]

[0087] To evaluate the performance of the two-branch time series network, it was compared with representative time series prediction models, including LSTM, Transformer, Informer, and PatchTST. All models used the same data partitioning scheme and preprocessing strategy to ensure a fair comparison. Additionally, a control variant model—the joint two-branch time series network—was designed. This model retains the feature modules of the two-branch time series network but integrates the predictions into a single output layer and uses a unified loss function.

[0088] ;

[0089] This strategy provides a control experiment to distinguish the effects of joint optimization versus independent optimization.

[0090] Comparing the Pearson correlation coefficient reveals no direct positive correlation between model complexity and predictive performance. Transformer and Informer models employing attention mechanisms perform excellently in natural language processing, but their performance in DSD parameter time series prediction is inferior to LSTM models. The predictive performance of the joint two-branch time series network model is also lower than that of the two-branch time series network, particularly in terms of μ... and The Pearson R values ​​for the two branches were 0.5350, 0.8385 and 0.7592, respectively, while the corresponding results for the two-branch time series network were 0.7488, 0.9315 and 0.8761.

[0091] The low performance of the joint two-branch time series network indicates that uniform optimization of all parameters ignores their different predictability and correlation characteristics. This confirms the necessity of the two-branch design: the high-complexity μ needs to be optimized independently, while the correlated... and This necessitates joint optimization.

[0092] A DSD (Distributed Drop Spectrum) observation station in a certain county was selected for model application. The precipitation event occurred from 9:12 AM to 10:11 AM on July 17, 2023. The DSD parameter data for the first 24 minutes were input into the model to predict the parameter values ​​for the following 12 minutes. The predicted 12-minute data was then fed back into the model as input for continuous prediction, ultimately generating a 36-minute forecast. The predicted values ​​of the three parameters were compared with the actual observed values. It can be seen that, compared with other existing models, the raindrop spectrum distribution prediction model (HVTS-Net) proposed in this invention can more accurately predict the changing trends of each parameter, indicating that it successfully captures the temporal variation characteristics of the parameters.

[0093] It should be understood that, although Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0094] In one embodiment, such as Figure 3 As shown, a raindrop spectrum distribution prediction device based on time series networks includes: a historical data acquisition module 302, a dependency analysis module 304, a network construction module 306, a training module 308, and a prediction module 310, wherein:

[0095] The historical data acquisition module 302 is used to acquire raindrop spectrum observation data of the target area within a preset historical time period to obtain a continuous time series of raindrop spectrum parameters; the raindrop spectrum parameters include mass-weighted average diameter, normalized intercept, and shape factor.

[0096] The dependency analysis module 304 is used to perform predictability assessment and inter-parameter dependency analysis on the time series of the raindrop spectral parameters according to the information theory analysis method, and obtain the analysis results.

[0097] The network construction module 306 is used to construct a two-branch time series network based on the analysis results; wherein, the first branch of the two-branch time series network is used to predict the shape factor, and the second branch is used to jointly predict the normalized intercept and the mass-weighted average diameter.

[0098] Training module 308 is used to train the dual-branch time series network using samples constructed from the time series of the raindrop spectral parameters, so as to obtain the trained dual-branch time series network.

[0099] The prediction module 310 is used to input the time series of raindrop spectral parameters to be predicted into a trained dual-branch time series network and output the predicted values ​​of raindrop spectral parameters for a future preset period.

[0100] Specific limitations regarding the raindrop spectral distribution prediction device based on time-series networks can be found in the limitations of the raindrop spectral distribution prediction method based on time-series networks mentioned above, and will not be repeated here. Each module in the aforementioned raindrop spectral distribution prediction device based on time-series networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0101] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores raindrop spectrum observation data, intermediate analysis results, and model parameters. The network interface communicates with external data acquisition terminals or user terminals via a network connection. When executed by the processor, the computer program implements a raindrop spectrum distribution prediction method based on time-series networks.

[0102] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0103] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0104] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting raindrop spectral distribution based on time series networks, characterized in that, The method includes: Raindrop spectrum observation data of the target area within a historical preset time period are obtained to obtain a continuous time series of raindrop spectrum parameters; the raindrop spectrum parameters include mass-weighted average diameter, normalized intercept, and shape factor; Based on information theory analysis methods, the predictability of the time series of the raindrop spectral parameters is evaluated and the inter-parameter dependency is analyzed to obtain the analysis results; Based on the analysis results, a two-branch time series network is constructed; wherein, the first branch of the two-branch time series network is used to predict the shape factor, and the second branch is used to jointly predict the normalized intercept and the mass-weighted average diameter; the second branch includes: an intervariate interaction module; the intervariate interaction module models the correlation between the normalized intercept and the mass-weighted average diameter through feature concatenation and linear transformation; The dual-branch time series network is trained using samples constructed from the time series of the raindrop spectral parameters to obtain a trained dual-branch time series network. Input the time series of raindrop spectral parameters to be predicted into the trained dual-branch time series network, and output the predicted values ​​of raindrop spectral parameters for the future preset time period; The steps for training the dual-branch time series network include: Independent optimization training strategies are used for the first branch and the second branch; Specifically, for the prediction target of the first branch, a two-stage training strategy is adopted: first, pre-training is performed using mean squared error loss, and then switching to a loss function weighted by extreme values ​​for fine-tuning. For the joint prediction objective of the second branch, end-to-end training is adopted, and a loss function with integrated high-value weighting mechanism is used.

2. The method according to claim 1, characterized in that, The information theory analysis method includes: Permutation entropy analysis and cross-correlation analysis; Specifically, based on the results of the permutation entropy analysis, the temporal complexity of the shape factor is determined to be higher than that of the mass-weighted average diameter and the normalized intercept. Based on the results of the cross-correlation analysis, a quantitative negative correlation was determined between the normalized intercept and the mass-weighted average diameter.

3. The method according to claim 1, characterized in that, Prior to the step of constructing the dual-branch time series network, the method further includes: Monthly analysis was performed on the time series of the raindrop spectral parameters to extract the seasonal physical characteristics of the raindrop spectral parameters; Each branch of the dual-branch time series network contains a seasonal feature fusion module, which encodes the seasonal physical features into learnable embedding vectors and fuses them with standard positional encodings to generate time series features that incorporate seasonal physical laws as input to the dual-branch time series network.

4. The method according to claim 1, characterized in that, The loss function for the first branch is: in, The time series representing the raindrop spectral parameters i True values ​​of shape factors, The time series representing the raindrop spectral parameters i One shape factor predicted value, n This represents the total number of shape factors in the time series of raindrop spectral parameters. The time series representing the raindrop spectral parameters i An extreme value attention coefficient is defined when the true value of the shape factor is at a preset normal value. Set to 1, when the true value of the shape factor is at a preset extreme value, the extreme value attention coefficient is... Set to be significantly greater than 1.

5. The method according to claim 1, characterized in that, The loss function for the second branch is: in, The time series representing the raindrop spectral parameters j The true value of the mass-weighted average diameter. The time series representing the raindrop spectral parameters j A mass-weighted average diameter prediction value, The first term represents the time series of raindrop spectral parameters with respect to the mass-weighted average diameter. j One error penalty coefficient, The normalized intercept is the error penalty coefficient when the true value of the mass-weighted average diameter is greater than a preset value. The error penalty coefficient is set to be greater than 1 when the true value of the mass-weighted average diameter is less than or equal to a preset value. Set to 1.

6. The method according to claim 1, characterized in that, In the dual-branch time series network, the time series feature extractors of the first branch and the second branch both adopt gated recurrent unit networks.

7. A raindrop spectral distribution prediction device based on time series networks, used to implement the raindrop spectral distribution prediction method based on time series networks as described in any one of claims 1 to 6, characterized in that, The device includes: The historical data acquisition module is used to acquire raindrop spectrum observation data of the target area within a preset historical time period to obtain a continuous time series of raindrop spectrum parameters; the raindrop spectrum parameters include mass-weighted average diameter, normalized intercept, and shape factor; The dependency analysis module is used to perform predictability assessment and inter-parameter dependency analysis on the time series of the raindrop spectral parameters based on information theory analysis methods, and obtain analysis results. A network construction module is used to construct a two-branch time series network based on the analysis results; wherein, the first branch of the two-branch time series network is used to predict the shape factor, and the second branch is used to jointly predict the normalized intercept and the mass-weighted average diameter; The training module is used to train the dual-branch time series network using samples constructed from the time series of the raindrop spectral parameters, so as to obtain the trained dual-branch time series network. The prediction module is used to input the time series of raindrop spectral parameters to be predicted into a trained dual-branch time series network and output the predicted values ​​of raindrop spectral parameters for a future preset period.

Images