A precipitation probability prediction method based on low-frequency weather map

Abstract

The application discloses a precipitation probability prediction method based on a low-frequency weather map, and comprises the following steps: collecting forecast data and historical precipitation real-time data of a meteorological numerical model; filtering the forecast data to obtain a low-frequency forecast field; converting daily precipitation into precipitation type one-hot encoding according to meteorological standards; pairing and aligning the low-frequency forecast field and the precipitation type one-hot code according to time as a data set; constructing a precipitation type probability prediction model by using a convolutional neural network; training the precipitation type probability prediction model by using the data set; and outputting a precipitation type probability prediction result based on the meteorological numerical model by using the trained precipitation type probability prediction model. The application takes the circulation prediction information of the meteorological numerical model as the starting point, takes the occurrence probability of the precipitation type as the prediction object, uses a time scale separation method to take the low-frequency circulation prediction information and the precipitation type as input quantities, and then constructs a prediction model based on a deep learning method to output the precipitation type probability prediction result.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to a method for predicting precipitation probability based on low-frequency weather maps. Background Technology

[0002] In meteorological departments, forecasts with a lead time of 10-30 days are called extended-range forecasts. Their time scale is between short-term weather forecasts of 0-10 days and short-term climate predictions of more than a month. They are a crucial link in building a refined and seamless weather forecasting system, but also a weak link.

[0003] Due to the chaotic nature of the atmosphere, subseasonal forecasts lack predictability, resulting in very limited predictive signals at subseasonal timescales. Although the timescale of subseasonal forecasts exceeds the theoretical upper limit of daily weather forecasts, atmospheric motion remains predictable, and this predictability is related to spatiotemporal scales. Meteorologists have conducted various attempts and studies on subseasonal atmospheric motions. The Madden Julian Oscillation (MJO) is the most important source of predictive techniques at subseasonal timescales. Some researchers have used the spatiotemporal information of the MJO as predictive factors to establish an empirical model for predicting spring precipitation in southern China at the subseasonal timescale. To accelerate research progress in subseasonal to seasonal forecasting and bridge the timescale gap between synoptic-scale forecasts and short-term climate forecasts, the World Weather Research Programme (WWRP) and the World Climate Research Programme (WCRP) jointly launched a five-year research project called "Seasonal to Seasonal (S2S) Prediction," aiming to improve subseasonal forecasting capabilities and enhance understanding of the sources of seasonal to subseasonal predictability.

[0004] Building upon research on predictability, researchers have developed a method for scale separation to extract low-frequency signals from the atmosphere over 10-60 days without using bandpass filters. This developed STPM method has shown good performance in predicting subseasonal precipitation in South China. Furthermore, researchers have developed a practical method and prediction technique for extracting predictable components from numerical models by extracting predictable components at the subseasonal timescale and referencing the Conditional Nonlinear Optimal Perturbation (CNOP) related algorithm.

[0005] Because weather and climate systems are typically nonlinear systems, the sheer volume and complexity of meteorological data make accurate predictions difficult. Artificial intelligence (AI) technology's ability to effectively learn and capture features from massive datasets has been widely applied across various fields. Therefore, deep learning technology has been extensively used in the meteorological industry in recent years, becoming a research hotspot. From cyclones to fronts to ENSO, the capabilities of machine learning in weather forecasting have become apparent. Summary of the Invention

[0006] To address the problems of current medium- and long-term precipitation forecast products being limited in variety, having low accuracy, and lacking uncertain information, this invention proposes a precipitation probability prediction method based on low-frequency weather maps. By utilizing the low-frequency characteristics of various meteorological forecast circulation fields and combining them with deep learning, the traditional precipitation forecast, which is a regression problem, is transformed into a precipitation type classification problem. This improves the predictability and accuracy of medium- and long-term time-leading precipitation forecasts. At the same time, the uncertainty descriptions provided can offer users more comprehensive information as a basis for behavioral decisions.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A precipitation probability prediction method based on low-frequency weather maps includes the following steps:

[0009] Collect daily forecasts of various meteorological circulations at multiple levels from 10 to 30 days using numerical models, as well as historical precipitation data corresponding to the forecast lead time.

[0010] By filtering the meteorological circulation data from the numerical model, a low-frequency circulation forecast field of meteorological elements for the next 10-30 days is obtained.

[0011] Based on the historical precipitation data, and according to the precipitation types classified by meteorological standards, the daily precipitation is converted into unique thermal codes for precipitation types, which are divided into no rain, light rain, moderate rain, heavy rain, and rainstorm.

[0012] The unique thermal codes of the low-frequency forecast field and precipitation type are paired and aligned by time to form a dataset.

[0013] A probabilistic prediction model for 10-30 day meteorological precipitation types was constructed using a convolutional neural network.

[0014] The dataset was used to train a probabilistic prediction model for the 10-30 day meteorological precipitation type.

[0015] Using the probabilistic prediction model of the 10-30 day meteorological precipitation type trained by the model, a probabilistic prediction of the precipitation type for the next 10-30 days based on the circulation forecast of the meteorological numerical model is completed.

[0016] Preferably, the method for filtering the circulation data of the 10-30 day meteorological numerical forecast is as follows:

[0017] By using a moving average, seasonal components longer than 90 days are removed from daily data, and high-frequency synoptic-scale components within 10 days are also removed, thus preserving the seasonal variation components.

[0018] Preferably, the specific process for filtering the 10-30 day meteorological circulation forecast data is as follows:

[0019] Remove slowly changing annual climate cycles and high-frequency cycles within 10 days;

[0020] The seasonal portion is kept between 10 and 60 days;

[0021] After filtering, the various elements are normalized.

[0022] Preferably, the method for removing slowly changing climate annual cycles is to subtract the climatological 90-day low-pass filter component from the original data.

[0023]

[0024] Wherein, the anomaly field X' is the meteorological component with the seasonal and higher periodic scales removed, including the high-frequency component of synoptic scale changes and the low-frequency component of intra-seasonal periodic scales; X is the original data; and X is the 90-day low-pass filter component of climatology.

[0025] Preferably, the method for retaining seasonal components for 10 to 60 days is as follows:

[0026] By subtracting the moving average of the last 30 days, all low-frequency signals are removed, and the residual is obtained.

[0027] The residual expression is: The moving average over the last 30 days is represented by X”, where X” is the residual.

[0028] The residuals are then averaged forward over 5 days, and the synoptic-scale component is removed. The expression is as follows: The forward 5-day moving average;

[0029] The final filtered components for 10-60 days are obtained.

[0030] Preferably, the method for normalizing the various elements after filtering is the minimax method, and its formula is as follows:

[0031]

[0032] in, x represents j The minimum value in the column. x represents j The maximum value in the column, x i,j Represents the original data. This represents the result obtained by normalizing through the maximum and minimum values.

[0033] Preferably, the method for training the 10-30 day meteorological precipitation probability prediction model using the dataset is as follows:

[0034] The forecast field of the low-frequency circulation is used as the input of the 10-30 day meteorological precipitation probability prediction model, and the output layer outputs the rainfall pattern.

[0035] The relationship between the output of the 10-30 day meteorological precipitation probability prediction model and the true value of precipitation type was calculated using the Softmax activation function, and the loss function was obtained.

[0036] The parameters of each neuron in the 10-30 day meteorological precipitation type probability prediction model are continuously trained by the back gradient propagation method until the accuracy of the 10-30 day meteorological precipitation type probability prediction model reaches the optimal level and meets the requirements.

[0037] Preferably, the method for predicting precipitation probability based on circulation forecasting using the trained 10-30 day meteorological precipitation probability prediction model is as follows:

[0038] The circulation forecast results based on the 10-30 day meteorological numerical model are filtered to obtain the low-frequency forecast field of meteorological elements, which is then input into the trained 10-30 day meteorological precipitation probability prediction model. The trained 10-30 day meteorological precipitation probability prediction model is used to perform forward calculations and output the probability prediction results of various precipitation types.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention utilizes the low-frequency characteristics of various meteorological forecast elements and combines them with deep learning to transform traditional precipitation forecasting as a regression problem into a classification problem of rainfall types with limited results. This improves the predictability and accuracy of medium- and long-term precipitation forecasts, while providing uncertainty descriptions to offer users more comprehensive information as a basis for behavioral decisions. Attached Figure Description

[0041] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of a precipitation probability prediction method based on low-frequency weather maps provided in an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of one-hot encoding provided in an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the network structure of the probabilistic prediction model for medium- and long-term precipitation types provided in an embodiment of the present invention.

[0045] Figure 4 A training and modeling flowchart provided for embodiments of the present invention;

[0046] Figure 5 This is a comparison chart of the hit rate of the method described in this embodiment of the invention and the CFS model for 10-30 days of precipitation type in the region in 2022. Detailed Implementation

[0047] 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 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.

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Example 1

[0050] Because 10-30 day weather forecasts have long lead times, the high-frequency components of precipitation have low predictability, while the low-frequency components of atmospheric motion are related to long-wave motion and exhibit a certain periodicity, thus having relatively higher predictability. This invention filters both the predictor and the forecast object separately to remove high-frequency signals, and uses deep learning methods to establish an end-to-end connection between the low-frequency information, thereby constructing a probabilistic prediction model for precipitation types based on low-frequency weather maps.

[0051] like Figure 1 As shown, this invention provides a precipitation probability prediction method based on low-frequency weather maps, comprising the following steps:

[0052] Collect daily forecasts of various meteorological circulations at multiple levels from 10 to 30 days using numerical models, as well as historical precipitation data corresponding to the forecast lead time.

[0053] By filtering the meteorological circulation data from the numerical model, a low-frequency forecast field of meteorological elements for the next 10-30 days is obtained.

[0054] Based on historical precipitation data, and according to the precipitation types classified by meteorological standards, daily precipitation is converted into unique thermal codes for precipitation types.

[0055] The unique thermal codes of low-frequency forecast fields and precipitation types are paired and aligned by time to form a dataset.

[0056] A probabilistic prediction model for 10-30 day meteorological precipitation types was constructed using a convolutional neural network.

[0057] A probabilistic prediction model for 10-30 day precipitation types was trained using a dataset.

[0058] Using a probabilistic prediction model of 10-30 day meteorological precipitation types trained on a meteorological numerical model, a probabilistic prediction of precipitation types for the next 10-30 days based on the circulation forecast of the meteorological numerical model is completed.

[0059] In this embodiment, the collected 10-30 day weather forecast results from numerical models include: collecting long-term meteorological data from multiple models, such as the US Climate Forecast Center (CFS), with the forecast spatial range covering all regions of the globe, and the forecast results including multiple meteorological elements such as temperature, pressure, humidity, and wind; and collecting historical daily precipitation data from forecast stations.

[0060] In this embodiment, the method for filtering 10-30 day weather forecast data is as follows:

[0061] The filtering method employed by Hsu et al. (2012) utilizes a non-filtering approach to extract intraseasonal oscillation signals of precipitation and atmospheric circulation. By using a moving average, annual cycle components longer than 90 days are removed from daily data, followed by the removal of components within 10 days at preset high-frequency and low-frequency synoptic scales, retaining the intraseasonal variation components. The advantage of this method is its ability to extract intraseasonal precipitation oscillation signals in real time, and it has already been applied to the extraction of intraseasonal precipitation components in tropical and subtropical regions.

[0062] In this embodiment, the extraction of the 10-60 day filtered components using a non-traditional filtering method mainly includes the following three steps: (10-30 days refers to the forecast lead time, and the filtering here refers to removing the low-frequency components after removing high-frequency and very low-frequency components)

[0063] Remove slowly changing (cycles longer than 90 days) annual climate cycles;

[0064] Seasonal ingredients are retained for 10 to 60 days;

[0065] After filtering, the various elements are normalized.

[0066] In this embodiment, the method for removing slowly changing annual climate cycles is as follows: it can be achieved by subtracting the climatological 90-day low-pass filter component from the original data (X); as shown in the following formula:

[0067]

[0068] Among them, the anomaly field X' contains a subseasonal component with a period between 10 and 60 days, a high-frequency component including synoptic-scale variations, and a low-frequency component including interannual and interdecadal anomalies; X is the raw data. This is the 90-day low-pass filter component for climatology.

[0069] In this embodiment, the method for retaining seasonal components for 10 to 60 days is as follows:

[0070] By subtracting the moving average of the last 30 days (i.e., from day -30 to day 0, denoted by 30d), all low-frequency signals are removed, and the residual is obtained.

[0071] The residuals are forward 5-day moving averages (from day -5 to day 0, denoted as 5d), and the synoptic scale components are removed to obtain the 10-60 day filtered components.

[0072] The residual expression is as follows: The moving average over the last 30 days is represented by X”, where X” is the residual.

[0073] The expression for the 10-60 day filter component is: This is a forward 5-day moving average.

[0074] In this embodiment, the method of normalizing each element after filtering to prevent low modeling efficiency and poor model generalization ability caused by different units is as follows: the minimax method is used; the formula is shown below:

[0075]

[0076] in, x represents j The minimum value in the column. x represents j The maximum value in the column, x i,j Represents the original data. This represents the result obtained by normalizing through the maximum and minimum values.

[0077] Maximum and minimum value normalization is to standardize the data using the maximum and minimum values ​​in the data column. The standardized values ​​are between [0,1]. The calculation method is to take the difference between the data and the minimum value of the column, and then divide by the range.

[0078] According to the national standard for precipitation levels (GB / T 28592-2012), the historical daily precipitation at the prediction stations is converted into daily precipitation type (no rain, light rain, moderate rain, heavy rain, and rainstorm), thus transforming the regression problem with infinite results into a classification problem with finite results. The classification results are then converted into one-hot encodings for model training. See the appendix for details. Figure 2 .

[0079] In this embodiment, the processed 10-30 day meteorological numerical forecast data (i.e., low-frequency forecast field) and precipitation data (rainfall pattern) are paired and aligned by time to serve as the dataset required for training the prediction model. Specifically, the low-frequency forecast field serves as the model input, with data dimensions of (number of samples, longitude, latitude, meteorological elements), and the rainfall pattern is the dependent variable, with data dimensions of (number of samples, no rain, light rain, moderate rain, heavy rain, torrential rain). The dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0080] In this embodiment, the method for training a probabilistic prediction model for 10-30 day precipitation types using a dataset is as follows:

[0081] The model uses a deep convolutional neural network (see attached diagram for structural schematic). Figure 3 This model uses preprocessed 10-30 day meteorological numerical forecast data (i.e., low-frequency forecast field) as input and outputs rainfall patterns. The Softmax activation function is used to calculate the relationship between the model output and the ground truth rainfall pattern, yielding the loss function (in this case, cross-entropy). Backward gradient propagation is then used to continuously train the parameters of each neuron until the model accuracy reaches its optimal level. The model continuously downsamples and abstracts the features of the low-frequency forecast field through convolutional layers, pooling layers, and batch normalization layers, achieving key feature extraction and dimensionality reduction (upscaling), thus enabling end-to-end prediction from low-frequency weather maps to rainfall patterns. The model training process is detailed in the appendix. Figure 4 .

[0082] In this embodiment, the method for predicting the probability of precipitation type in the 10-30 day meteorological numerical forecast using a trained probabilistic prediction model of 10-30 day precipitation types is as follows:

[0083] The latest 10-30 day meteorological numerical forecast results are filtered to obtain the latest low-frequency forecast field of meteorological elements, which is then input into the trained 10-30 day meteorological precipitation probability prediction model. The trained 10-30 day meteorological precipitation probability prediction model performs forward calculations and outputs the probability prediction results of each precipitation type.

[0084] Figure 5This is a comparison chart of the hit rate of the method described in the embodiments of the present invention and the CFS model for precipitation type in the region for 10-30 days in 2022 (the hit rate is calculated by comparing the precipitation type forecast results of each station with the actual precipitation type data of the same period). In the 10-30 day forecast lead time period, the average hit rate of the 10-30 day method of the embodiments of the present invention (CNN) is higher than that of the CFS model. Although the forecast effect for 10-14 days is slightly lower than that of the CFS model, the forecast effect for 18-30 days is consistently higher than that of the CFS model, indicating that the forecast effect of the embodiments of the present invention is improved compared with the results of the CFS model.

[0085] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A precipitation probability prediction method based on low-frequency weather maps, characterized in that, Includes the following steps: Collect daily forecasts of various meteorological circulations at multiple levels from 10 to 30 days using numerical models, as well as historical precipitation data corresponding to the forecast lead time. By filtering the meteorological circulation data from the numerical model, a low-frequency circulation forecast field of meteorological elements for the next 10-30 days is obtained. Based on the historical precipitation data, and according to the precipitation types classified by meteorological standards, the daily precipitation is converted into unique thermal codes for precipitation types, which are divided into no rain, light rain, moderate rain, heavy rain, and rainstorm. The unique thermal codes of the low-frequency circulation forecast field and precipitation type are paired and aligned by time to form a dataset. A probabilistic prediction model for 10-30 day meteorological precipitation types was constructed using a convolutional neural network. The dataset was used to train a probabilistic prediction model for the 10-30 day meteorological precipitation type. Using the trained probabilistic prediction model of the 10-30 day meteorological precipitation type, a probabilistic prediction of the precipitation type for the next 10-30 days based on the circulation forecast of the meteorological numerical model is completed. According to the national standard GB / T 28592-2012 for precipitation levels, the historical daily precipitation of the predicted stations is converted into daily precipitation type: no rain, light rain, moderate rain, heavy rain, and rainstorm. This transforms the regression problem with infinite results into a classification problem with finite results, and the classification results are converted into one-hot encodings for model training. The processed 10-30 day meteorological numerical forecast data (i.e., low-frequency forecast field) and precipitation data (i.e., rain pattern) were paired and aligned by time to serve as the dataset required for training the prediction model. The low-frequency forecast field was used as the input to the model, with data dimensions of: number of samples, longitude, latitude, and meteorological elements. The rain pattern was used as the dependent variable, with data dimensions of: number of samples, no rain, light rain, moderate rain, heavy rain, and rainstorm. The dataset was randomly divided into training set, validation set, and test set in a ratio of 8:1:

1. The method for training a probabilistic prediction model of 10-30 day precipitation types using a dataset is as follows: The model employs a deep convolutional neural network, using preprocessed 10-30 day meteorological numerical forecast data (i.e., low-frequency forecast fields) as input and outputting rainfall patterns. The Softmax activation function is used to calculate the relationship between the model output and the ground truth rainfall pattern, yielding the loss function. Backward gradient propagation is then used to continuously train the parameters of each neuron until the model achieves optimal accuracy and meets the requirements. The model continuously downsamples and abstracts the features of the low-frequency forecast field through convolutional layers, pooling layers, and batch normalization layers, achieving key feature extraction and dimensionality reduction, thus enabling end-to-end prediction from low-frequency weather maps to rainfall patterns. The method for predicting the probability of precipitation type in the 10-30 day meteorological numerical forecast using a trained probabilistic prediction model of precipitation type is as follows: The 10-30 day meteorological numerical forecast results are filtered to obtain the latest low-frequency forecast field of meteorological elements, which is then input into the trained 10-30 day meteorological precipitation probability prediction model. The trained 10-30 day meteorological precipitation probability prediction model performs forward calculations and outputs the probability prediction results of each precipitation type.

2. The precipitation probability prediction method based on low-frequency weather maps according to claim 1, characterized in that, The method for filtering the circulation data from the 10-30 day meteorological numerical forecast is as follows: By using a moving average, seasonal components longer than 90 days are removed from daily data, and high-frequency synoptic-scale components within 10 days are also removed, thus preserving the seasonal variation components.

3. The precipitation probability prediction method based on low-frequency weather maps according to claim 2, characterized in that, The specific process of filtering the 10-30 day meteorological circulation forecast data is as follows: Remove slowly changing annual climate cycles and high-frequency cycles within 10 days; The seasonal portion is kept between 10 and 60 days; After filtering, the various elements are normalized.

4. The precipitation probability prediction method based on low-frequency weather maps according to claim 3, characterized in that, The method for removing slowly changing annual climate cycles is achieved by subtracting the climatological 90-day low-pass filter component from the original data. Among them, the abnormal field To remove meteorological components at periodic scales above the seasonal level, including high-frequency components of synoptic-scale changes and low-frequency components at intra-seasonal periodic scales. This is the original data; This is the 90-day low-pass filter component for climatology.

5. The precipitation probability prediction method based on low-frequency weather maps according to claim 3, characterized in that, The method for retaining seasonal components for 10 to 60 days is as follows: By subtracting the moving average of the last 30 days, all low-frequency signals are removed, and the residual is obtained. The residual expression is: , The moving average over the last 30 days. For residuals; The residuals are then averaged forward over 5 days, and the synoptic-scale component is removed. The expression is as follows: , The forward 5-day moving average; The final filtered components for 10-60 days are obtained.

6. The precipitation probability prediction method based on low-frequency weather maps according to claim 3, characterized in that, The method for normalizing the various elements after filtering is the minimax method, and its formula is as follows: in, express The minimum value in the column. express The maximum value in the column. Represents the original data. This represents the result obtained by normalizing through the maximum and minimum values.

Images