A subseasonal heat wave prediction method and storage medium

Abstract

The present invention provides a sub-seasonal heat wave prediction method, comprising: determining all optimal prediction factors associated with any temperature main mode time coefficient under a preset forecast time according to a pre-obtained temperature main mode of the target area and a temperature main mode time coefficient-forecast time effect-optimal prediction factor relationship table. Data of all optimal prediction factors are obtained by real-time non-bandpass filtering, input into a pre-trained sub-seasonal temperature forecast model, and a predicted temperature anomaly field under a preset time effect in the target area. Sub-seasonal heat wave events in the target area are predicted based on the predicted temperature anomaly field. The present invention utilizes the powerful data mining capabilities of convolutional neural networks to consider the interactions between multi-scale and multi-variable variables, fully considers the impact of interannual-scale signals on large-scale sub-seasonal variability, and explores the potential correlation between previous multi-scale signals and future temperatures.

Description

Technical Field

[0001] The invention relates to a sub-seasonal heat wave prediction method and a storage medium, belonging to the technical field of sub-seasonal climate prediction. Background Art

[0002] Extreme heat or heatwave events can cause considerable harm to human health, ecosystems, and social services. The Intergovernmental Panel on Climate Change's Sixth Assessment Report indicates that climate change has exacerbated extreme heat events worldwide. Improving the ability to predict extreme weather events and extending the forecast timeline beyond conventional weather scales (>10 days) is crucial for effective risk reduction and emergency preparedness. Given the difficulties that current operational numerical dynamical models face in accurately predicting extreme heat events on the subseasonal scale, the World Meteorological Organization has launched the Subseasonal-Seasonal (S2S) Prediction Project to improve the predictability of these high-impact events.

[0003] The Tropical Intraseasonal Oscillation (MJO) and the Extratropical Intraseasonal Variability (ETV) are the main modes of atmospheric subseasonal variability. These modes influence temperature changes over China by inducing anomalous circulation and thermal conditions and are therefore considered key sources of predictability for subseasonal heatwaves in China. Previous studies have used large-scale intraseasonal signals as predictors in statistical models to develop subseasonal forecast models, effectively improving subseasonal forecasts of extreme weather and climate events over China. Furthermore, disturbances associated with the MJO and the ETVV are influenced by background conditions associated with the El Niño–Southern Oscillation (ENSO) and other interannual sea surface temperature patterns. Therefore, incorporating interannual anomalous signals into statistical models can improve subseasonal forecast capabilities.

[0004] While the sources of subseasonal predictability in extreme weather events are well understood, effectively leveraging this information to improve forecasting capabilities remains a challenge. With the use of deep learning techniques, some studies have employed deep learning models such as convolutional neural networks (CNNs) and long short-term memory (LSTMs) to correct for systematic biases in dynamic forecasts of the MJO, precipitation, and temperature, and to use precursor signals to predict future regional precipitation and temperature changes. However, these models fail to consider the influence of interannual-scale signals on large-scale subseasonal variability, neglecting the potential predictability implications of multiscale interacting processes. This presents a critical yet unresolved scientific issue in subseasonal forecasting. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a sub-seasonal heat wave prediction method. By leveraging the powerful data mining capabilities of convolutional neural networks to consider the interactions between multi-scale and multi-variable variables, the modulation of interannual variability on atmospheric sub-seasonal variability is taken into account. By mining the potential correlation between previous multi-scale signals and future temperatures, future temperature changes are predicted, and sub-seasonal predictions of heat wave events are then made.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a sub-seasonal heat wave prediction method, comprising:

[0008] According to the pre-obtained temperature main mode of the target area and the temperature main mode time coefficient-forecast time effectiveness-optimal forecast factor component relationship table, all optimal forecast factors associated with any temperature main mode time coefficient under the preset forecast time effectiveness are determined to form an optimal forecast factor set;

[0009] Acquire data of each optimal predictor in the optimal predictor set, and extract the subseasonal component and the low-frequency background state (LFBS) component of each optimal predictor respectively through real-time non-bandpass filtering;

[0010] Inputting the subseasonal component and the low-frequency background state (LFBS) component of each optimal forecast factor into a pre-trained subseasonal temperature forecast model to obtain a predicted value of each main mode time coefficient of the temperature in the target area under the time effect;

[0011] Obtaining a predicted temperature anomaly field under the time effect according to the predicted value of each temperature main mode time coefficient and the temperature main mode reconstruction;

[0012] Predicting sub-seasonal heat wave events in the target area based on the predicted temperature anomaly field under the time validity period and the preset threshold standard;

[0013] The process of obtaining the predicted value of any temperature main mode time coefficient includes:

[0014] The sub-seasonal temperature forecast model performs feature extraction on the optimal forecast factor component corresponding to the main mode time coefficient of the temperature to obtain the deep features of the optimal forecast factor component;

[0015] The deep features of all the optimal predictor components corresponding to the main mode time coefficient of the temperature are fused to obtain the joint features containing multi-scale and multi-variable interaction information;

[0016] The predicted value of the main modal time coefficient of the temperature is obtained by combining the joint features and the current month information.

[0017] Furthermore, the sub-seasonal temperature forecast model also includes: a univariate forecast model; the univariate forecast model is used to provide the temperature modal time coefficient-forecast time-optimal forecast factor relationship table and a pre-trained convolution block for extracting deep features of each optimal forecast factor component during the training process of the sub-seasonal temperature forecast model.

[0018] Furthermore, the training process of the sub-seasonal temperature forecast model includes:

[0019] Determining a target area and determining candidate predictors influencing daily maximum temperature within the target area;

[0020] Obtain the daily maximum temperature data for a preset historical period in the target area and the data of the candidate forecast factors, construct a data set and perform preprocessing, and divide the processed data into a training set, a validation set, and a test set according to the year;

[0021] A non-bandpass filtering method is used to extract the subseasonal component and the low-frequency background state LFBS component of the candidate predictor;

[0022] The summer temperature data is obtained from the daily maximum temperature data in the training set, and the empirical orthogonal function (EOF) is used for analysis to obtain S main temperature modes and their corresponding time coefficients.

[0023] The full-year data of each candidate predictor is obtained from the training set. The subseasonal component and LFBS component of each candidate predictor are used as the input of the univariate prediction model, and the time coefficient of the S main temperature modes is used as the output of the univariate prediction model to train the univariate prediction model.

[0024] The subseasonal components or LFBS components of each candidate predictor in the validation set are input into the corresponding univariate prediction model, and S types of temperature main mode time coefficients are predicted separately at different time periods in the future. By evaluating and comparing the prediction skills of the univariate prediction models, the a subseasonal components of the predictors and b LFBS components of the predictors that perform best for each temperature main mode time coefficient at each time period are determined, forming a relationship table of temperature main mode time coefficient, forecast time period, and optimal predictor component; where a and b are determined based on sensitivity tests;

[0025] The data of the (a+b) optimal predictors corresponding to each main temperature mode time coefficient in the training set are used as input, and each main temperature mode time coefficient is used as the output of the sub-seasonal temperature forecast model. The training is performed in combination with the pre-trained convolution blocks provided by the univariate forecast model corresponding to the (a+b) optimal predictors;

[0026] Use the validation data to evaluate the trained subseasonal temperature forecast model and optimize the model structure and parameters;

[0027] The independent forecast performance of the subseasonal temperature forecast model was tested using data from the test set.

[0028] Furthermore, the univariate prediction model includes: 2 input layers, 1 convolutional block, 1 fully connected layer and 1 output layer;

[0029] For any candidate predictor:

[0030] The first input layer provides a corresponding feature map of the subseasonal component or LFBS component of the candidate predictor;

[0031] The convolution block is used to perform convolution and pooling operations on the feature map to extract corresponding deep features;

[0032] The second input layer provides information for initializing the current month;

[0033] The fully connected layer integrates the deep features and the monthly information, and outputs the time coefficient of the specified temperature main mode under the future specified forecast time through the output layer;

[0034] The sub-seasonal temperature forecast model includes: (a+b+1) input layers, (a+b) convolution blocks, 2 two-dimensional convolution layers, 1 fully connected layer and 1 output layer;

[0035] The (a+b) input layers are respectively arranged in (a+b) independent input channels; and are used to provide (a+b) optimal prediction factors corresponding to input feature maps;

[0036] The (a+b) convolution blocks are respectively used to extract deep features corresponding to the (a+b) optimal predictor factors;

[0037] The two-dimensional convolution layer is used to perform further convolution operations on the result of splicing the deep features of the (a+b) optimal forecast factors in the channel dimension direction. The convolution result and the monthly information provided by the last input layer are integrated through the fully connected layer, and the time coefficient of the specified main mode of temperature under the future specified forecast time is output through the output layer.

[0038] Furthermore, the convolution block includes three two-dimensional convolutional layers and two maximum pooling layers arranged at intervals.

[0039] Furthermore, the extracting of the subseasonal component and the LFBS component of the candidate predictor by using a non-bandpass filtering method includes:

[0040] Subtract the daily climate state from the raw daily data of the candidate predictor to eliminate the annual climate cycle and obtain the daily anomaly value; the daily climate state is the multi-year average of the same date from 1991 to 2020;

[0041] Subtract the average of the previous 45 days from the daily outlier to eliminate the signal of more than 90 days; get the outlier of less than 90 days;

[0042] A 5-day moving average is used for the abnormal values ​​less than 90 days to eliminate the high-frequency weather-scale disturbances less than 10 days, and the subseasonal component of the candidate forecast factor is obtained;

[0043] The LFBS component is the 30-day average of the daily anomalies.

[0044] Furthermore, the filter size of the two-dimensional convolutional layer is 3×3, the sliding step value is 1×1, and the activation function of the two-dimensional convolutional layer and the fully connected layer is a rectified linear unit with a leakage parameter; the filter size of the maximum pooling layer is 2×2, and the sliding step value is 2×2;

[0045] Before the two-dimensional convolutional layer performs the convolution operation, it also includes: padding the boundary of the feature map input to the two-dimensional convolutional layer with zero values ​​so that the size of the feature map remains unchanged before and after the operation.

[0046] Furthermore, the candidate prediction factors include: 850 hPa temperature T850, 200 hPa zonal wind U200, 500 hPa zonal wind U500, 850 hPa zonal wind U850, 200 hPa meridional wind V200, 500 hPa meridional wind V500, 850 hPa meridional wind V850, 200 hPa geopotential height field H200, 500 hPa geopotential height field H500, 850 hPa geopotential height field H850, 700 hPa relative humidity RH700, precipitable water PW, sea level pressure SLP, surface soil moisture SSM and outward longwave radiation OLR.

[0047] Furthermore, during the training process of the sub-seasonal temperature forecast model, the Adam optimizer was used to optimize the training parameters, and the initial learning rate of the Adam optimizer was 1×10 −3 ;Use mean square error as loss function;

[0048] A dropout method and an early stopping method are used to prevent the model from overfitting. The dropout method includes: adding a dropout layer with a preset probability of discarding after each two-dimensional convolutional layer during the model training iteration;

[0049] The early stopping method includes: in response to no improvement in skill over T consecutive training iterations on the validation set, stopping training to ensure effective utilization of model weights corresponding to the minimum validation loss, where T is a preset value.

[0050] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting sub-seasonal heat waves as described in the first aspect is implemented.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) The subseasonal temperature forecast model provided by the present invention utilizes the powerful data mining capabilities of convolutional neural networks to consider the interactions between multi-scale and multi-variable variables. Non-bandpass filtering is used to extract the subseasonal component and LFBS component of the optimal predictor as model input. The LFBS component of the optimal predictor contains interannual-scale signals, allowing the model to not only learn the predictability information provided by large-scale subseasonal variability, but also consider the impact of interannual-scale signals on large-scale subseasonal variability.

[0053] (2) During the training process of the sub-seasonal temperature forecast model provided by the present invention, a univariate forecast model is trained based on the sub-seasonal component or LFBS component of a single candidate forecast factor. The optimal forecast factor is determined based on the performance of multiple candidate forecast factors during the validation period. The convolutional block of the univariate forecast model corresponding to the optimal forecast factor is used to initialize the trained sub-seasonal temperature forecast model, thereby improving its sub-seasonal forecast capability for extreme heat waves. The number of optimal forecast factors is determined based on sensitivity testing to ensure that the optimal forecast factor configuration has the most robust and proficient forecast performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Results of applying the empirical orthogonal function (EOF) analysis to China's summer temperature data, provided in Example 1 of the present invention; in the figure, af are the eigenvectors of the first six major EOF modes of summer temperature anomalies in China from 1991 to 2010, g is the corresponding explained variance (black curve; left y-axis; unit: %) and error range (red line), and the blue curve in g is the PCC (right y-axis) between the temperature anomalies reconstructed using the first N (1-8) major EOF modes from 2016 to 2021 and the observed values;

[0055] Figure 2 The spatial distribution of the lead-lag correlation coefficients between PC1 of the summer temperature anomaly in China during 1979–2009 and the antecedent (af) PW and (gl) H500 subseasonal components and the (mr) SLP and (sx) H850 LFBS components provided in Example 1 of the present invention;

[0056] Figure 3 The model architecture of the univariate forecast model (a) and the sub-seasonal temperature forecast model (b) provided in Example 1 of the present invention;

[0057] Figure 4 The TCC skill distribution of the 10-30 day (2-6 pentad) forecasts of the (ae) CNN, (fj) CMA, and (ko) ECMWF models during the test period provided in Example 2 of the present invention;

[0058] Figure 5 The PCC skills of the three models for the spatial mode of temperature at different forecast times during the test period provided in Example 2 of the present invention;

[0059] Figure 6 The RMSE technique for predicting temperature anomalies provided in Example 2 of the present invention;

[0060] Figure 7 The EDI technique for predicting heat waves provided in Example 2 of the present invention;

[0061] Figure 8 The attribution graph corresponding to the accurate prediction of the PC1 high value period 20 days in advance during the test period provided in Example 2 of the present invention;

[0062] Figure 9 During the test period provided by Example 2 of the present invention, the CNN model accurately predicted the large-scale subseasonal disturbance input field composite map corresponding to the initial reporting sample during the PC1 high value period (i.e., exceeding the 90th percentile threshold) 20 days in advance;

[0063] Figure 10 The distribution diagram of the lead-lag correlation coefficient between the LFBS field three months ahead and PC1 of the temperature anomaly in China during the test period provided in Example 2 of the present invention;

[0064] Figure 11 The model architecture of the sensitivity experiment provided in Example 1 of the present invention;

[0065] Figure 12 The prediction skills of CNN models based on different architectures for temperature and heat waves during the test period provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0066] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments.

[0067] Example 1

[0068] The present invention provides a sub-seasonal heat wave prediction method that predicts the spatial distribution of temperature anomalies in a target area for the next 10-30 days based on a pre-trained sub-seasonal temperature forecast model, and monitors the occurrence of heat wave events based on the predicted temperature anomalies.

[0069] In this embodiment, the application scenario of sub-seasonal heat wave prediction is China. This embodiment provides a process of constructing a sub-seasonal temperature forecast model based on CNN and a training process of the sub-seasonal temperature forecast model.

[0070] First, empirical orthogonal function (EOF) analysis was applied to Chinese summer temperature data to obtain the main temperature modes and their temporal coefficients over China. Candidate predictors were then selected based on previous research on pre-seasonal signals associated with heatwaves in China.

[0071] Specifically, EOF analysis was performed using China's summer temperature data from 1991 to 2010. Figure 1 The eigenvectors (a–f) of the first six main EOF modes of summer temperature anomalies over China during 1991–2010 are shown. Figure 1 (g) shows the corresponding explained variance (black curve; left y-axis; unit: %) and error range (red line). The blue curve in (g) is the spatial correlation coefficient (PCC) between the temperature anomaly reconstructed using the first N (1–8) (x-axis) main EOF modes and the observed value from 2016 to 2021, which is given by Figure 1 It can be seen that the predicted temperature anomaly field in China can be reconstructed from the first six main temperature modes and the corresponding time coefficients (PC1-PC6).

[0072] The large-scale fields that serve as candidate predictors include: 850 hPa temperature T850, 200 hPa zonal wind U200, 500 hPa zonal wind U500, 850 hPa zonal wind U850, 200 hPa meridional wind V200, 500 hPa meridional wind V500, 850 hPa meridional wind V850, 200 hPa geopotential height field H200, 500 hPa geopotential height field H500, 850 hPa geopotential height field H850, 700 hPa relative humidity RH700, precipitable water PW, sea level pressure SLP, surface soil moisture SSM, and outward longwave radiation OLR.

[0073] In order to verify the influence of the selected candidate predictors on the temperature anomaly in China, the subseasonal component and low-frequency background state LFBS component of the candidate predictors were extracted by non-bandpass filtering method and experimental analysis was carried out. Figure 2 The correlation between the variability of the main mode time coefficient PC1 of China's temperature and the subseasonal component of the previous PW, the subseasonal component of H500, the LFBS component of SLP and the LFBS component of H850 is shown.

[0074] similar Figure 2 Lead-lag correlation analysis of other PCs revealed that three main pathways of subseasonal disturbances are crucial for regulating summer temperature variability in China: 1) a southeasterly signal originating in northwestern Eurasia, 2) a southeasterly signal originating in Northeast Asia, and 3) a northerly signal originating in the tropical western Pacific and tropical Indian Oceans. Key early-stage LFBS signals influencing temperature are also important over Eurasia and the tropical oceans. Therefore, in this example, the target region for constructing the subseasonal temperature forecast model is defined as Eurasia and the tropics (2.5°S-75.0°N, 40°-167.5°E).

[0075] The daily maximum temperature data in China from 1979 to 2021 and the data of candidate forecast factors in the target area were obtained, a data set was constructed and preprocessed, and the processed data were divided into training set, validation set and test set according to the year.

[0076] Specifically, the SSM and OLR are derived from reanalysis data from the National Centers for Environmental Prediction / National Center for Atmospheric Research and polar-orbiting satellites from the National Oceanic and Atmospheric Administration, respectively. The remaining variables are derived from reanalysis data from the National Centers for Environmental Prediction / Department of Energy. The SSM uses a T62 Gaussian grid format (192 × 94), and the horizontal resolution of the remaining large-scale variables is 2.5° × 2.5°. Before model training, the SSM field was interpolated to a horizontal resolution of 2.5° × 2.5°.

[0077] For real-time applications, a non-bandpass filtering method was used to extract the subseasonal component (10–90 days) and low-frequency background state (LFBS) component of the candidate predictors. Specifically, for each variable, the daily climate state was first subtracted from the original daily data of the candidate predictor to eliminate the annual climate cycle, resulting in a daily anomaly. The daily climate state was then averaged over the same date from 1991 to 2020. The average of the preceding 45 days was then subtracted from the daily anomaly to eliminate signals exceeding 90 days, resulting in anomalies under 90 days. These anomalies under 90 days were then subjected to a leading 5-day running average to eliminate high-frequency weather-scale disturbances under 10 days, resulting in the subseasonal component of the candidate predictor. The LFBS component was then calculated as the 30-day average of the daily anomaly.

[0078] The historical data are divided into three stages: 1979-2009 is the training period, and relevant samples are used to train the model; 2010-2015 is the validation period, which is used to determine the optimal prediction factor components and model tuning; 2016-2021 is the testing period, which is used to detect the actual forecasting ability of the model.

[0079] A sub-seasonal temperature forecast model is constructed based on CNN. This model is a multivariate forecast model, and its construction process involves the construction of a univariate forecast model. The univariate forecast model is used to determine the relationship table of the main mode time coefficient of the temperature - forecast time - optimal forecast factor component, and provides a convolution block for extracting the deep features of the optimal forecast factor component. The convolution block is used to train the sub-seasonal temperature forecast model. Figure 3 The specific structures of the univariate forecast model and the sub-seasonal temperature forecast model considering multivariate information are shown.

[0080] like Figure 3 As shown in the figure, the univariate forecast model consists of two input layers, a convolutional block consisting of three two-dimensional convolutional (Conv2d) layers and two maximum pooling (MP) layers, a fully connected (FC) layer, and an output layer. The first input layer (orange box) provides the input features corresponding to the forecast factors, while the second input layer (red box) provides the encoded initialization information for the current month, which guides the model to consider the annual cycle of subseasonal variability.

[0081] The subseasonal temperature forecast model that considers multivariate information consists of eight input layers, seven convolutional blocks consisting of three Conv2d layers and two MP layers, two two-dimensional convolutional layers, one fully connected layer, and one output layer. Seven of the input layers are arranged in seven independent input channels to consider the predictable information provided by four subseasonal predictors and three LFBS predictors. Another input layer (red box) provides the encoded initial information for the current month.

[0082] First, a univariate forecast model is trained using each candidate forecast factor. This model is trained using data from the entire year, not just the summer samples. This method yields a better model than the one trained using only the summer samples. The training of the univariate forecast model includes:

[0083] For any candidate predictor: obtain the full-year data of the candidate predictor from the training set, and use the non-bandpass filtering method to extract the subseasonal component and low-frequency background state LFBS component of the candidate predictor; use the subseasonal component or LFBS component of the candidate predictor as the input of the univariate prediction model, specify the main mode time coefficient of temperature as the output of the univariate prediction model, and train the univariate prediction model.

[0084] Specifically, for any candidate forecast factor: the first input layer provides the corresponding feature map of the subseasonal component or LFBS component of the candidate forecast factor; the convolution block is used to perform convolution and pooling operations on the input layer feature map to obtain the deep features provided by the candidate forecast factor component; the second input layer provides the information for initializing the current month; the fully connected layer integrates the deep features and the monthly information, and outputs the time coefficient of the main mode of temperature under the future specified forecast time through the output layer.

[0085] Furthermore, the univariate forecasting model constructed based on the subseasonal component or LFBS component of each candidate forecasting factor on the validation set was evaluated for its forecasting skill for the six main temperature mode time coefficients 10 days, 15 days, 20 days, 25 days and 30 days in advance, and the four subseasonal forecasting factors and three LFBS forecasting factors with the best forecasting performance for each main temperature mode time coefficient under each forecast time were determined, generating the relationship table of main temperature mode time coefficient-forecast time-optimal forecasting factor component shown below.

[0086]

[0087] Specifically, the feature map corresponding to the first input layer (orange box) has dimensions N × 32 × 52. For subseasonal predictors, this represents the large-scale subseasonal perturbation values ​​(N = 6) every 5 days over the preceding 30 days within the region of 2.5°S–75.0°N (32 grid points) and 40°–167.5°E (52 grid points). For LFBS predictors, this represents the LFBS components (N = 3) every 30 days over the preceding 90 days within the same region. The output layer (black box) is used to predict the time coefficients of the main mode of temperature for the next 10, 15, 20, 25, or 30 days (i.e., 2–6 pentads).

[0088] It should be noted that the number of subseasonal predictors and LFBS predictors used to construct the subseasonal temperature forecast model considering multivariate information is determined by sensitivity tests. Such a predictor configuration has the most robust and proficient forecast performance.

[0089] The model's impressive predictive skill is attributed to its integration of information from multiple scales and multivariate interactions. This is demonstrated by a series of sensitivity tests using different model structures:

[0090] (1) CNN-A Figure 11 a): Similar to the sub-seasonal temperature forecast model considering multivariate information in the present invention, but this model only uses the forecast information from four sub-seasonal forecast factors.

[0091] (2) CNN-L Figure 11 b): with CNN-A ( Figure 11 a) Similar, but using only information from the three LFBS predictors.

[0092] (3) CNN-A-t0 Figure 11 c): Based on CNN-A ( Figure 11 a), but only using information from the most recent previous time step of the subseasonal predictor (N = 1). The input layer feature map associated with each large-scale predictor has dimensions of 1 × 32 × 52.

[0093] (4) CNN-A-ENS: No additional convolution operations are added, and prediction is made by simply averaging the results of a single predictor.

[0094] (5) CNN-NP: Similar to the subseasonal temperature forecast model in this invention, but this model does not use the pre-trained weights of the convolutional blocks obtained by training the optimal predictor model.

[0095] Figure 12 The results of the sensitivity test are shown: when the model adopts the same CNN structure as the sub-seasonal temperature forecast model in the present invention, but only uses the sub-seasonal forecast factor ( Figure 11 a) or LFBS predictor ( Figure 11 b), the prediction skill decreased by 5%–15% and 25%–60% compared to the model that exploited information from both scales ( Figure 12 Introducing the spatiotemporal evolution of the predictors into the model can also improve the prediction ability. Compared with the model prediction skill using the subseasonal signal of the first 30 days (N = 6), the prediction skill of the model using only the first 5 days (i.e., N = 1; Figure 11 c) The model's skill in predicting temperature and heat waves has decreased by 5%–30% ( Figure 12Purple and green histograms). The sensitivity test results also emphasize the importance of adding additional convolution operations to integrate information from various predictors ( Figure 3 b) Without this feature fusion, the ensemble forecast obtained by averaging the outputs of the individual predictors shows lower skill ( Figure 11 In addition, not using pre-trained convolutional block weights to train the sub-seasonal temperature forecast model will also lead to a decrease in model prediction skills ( Figure 11 Orange and blue histograms), highlighting the advantages and necessity of pre-training in improving the generalization ability of the model.

[0096] The training of the sub-seasonal temperature forecast model includes: taking the data of the 7 optimal prediction factor components corresponding to each main temperature modal time coefficient in the training set as input, taking each main temperature modal time coefficient as the output of the sub-seasonal temperature forecast model, and combining the convolution blocks provided by the 7 optimal prediction factor components in the univariate forecast model to train the sub-seasonal temperature forecast model.

[0097] Specifically, the input features of the seven optimal predictor components are respectively extracted through corresponding convolution blocks to obtain the deep features of the seven optimal predictor components; the two-dimensional convolution layer is used to perform convolution processing on the channel connection results of the deep features of the seven optimal predictor components to obtain a joint feature, and the joint feature and the month information encoded by the last input layer are integrated through the fully connected layer, and the main modal time coefficient of the specified temperature under the future specified forecast time is output through the output layer.

[0098] When training the model, the pre-trained convolutional blocks from the univariate prediction model corresponding to the optimal predictor component are used ( Figure 3 a) for training. After processing through the convolutional blocks, the deep features extracted from multiple optimal predictors are concatenated along the channel dimension, followed by two convolution operations. This approach leverages the joint prediction information from the optimal predictors, allowing the model to fully account for the interaction between subseasonal and LFBS variability.

[0099] Use the validation data to evaluate the trained subseasonal temperature forecast model and optimize the model structure and parameters;

[0100] The independent forecast performance of the subseasonal temperature forecast model was tested using data from the test set.

[0101] The filter size of all two-dimensional convolutional layers in the sub-seasonal temperature forecast model and the univariate forecast model is 3×3, with a sliding step size of 1×1. The activation functions of the two-dimensional convolutional layers and the fully connected layers are rectified linear units with leakage parameters. The filter size of the maximum pooling layer is 2×2, with a sliding step size of 2×2. Before performing convolution operations on all two-dimensional convolutional layers, the boundaries of the feature maps input to the two-dimensional convolutional layer are padded with zero values ​​to ensure that the feature map size remains unchanged before and after the operation.

[0102] Preferably, during the training process of the sub-seasonal temperature forecast model, the Adam optimizer is used to optimize the training parameters, and the initial learning rate of the Adam optimizer is 1×10 −3 .

[0103] Preferably, the mean square error is used as the loss function.

[0104] Preferably, a dropout method and an early stopping method are used to prevent the model from overfitting; the dropout method includes: adding a dropout layer with a dropout probability of 50% after each two-dimensional convolutional layer during the model training iteration; the early stopping method includes: stopping training in response to no improvement in the skill of 15 consecutive training iterations on the validation set to ensure effective utilization of the model weights corresponding to the minimum validation loss.

[0105] Example 2

[0106] Based on Example 1, this embodiment provides a method for predicting sub-seasonal heat waves in China, including the following steps:

[0107] Step 1: Based on the pre-obtained main temperature mode of the target area and the temperature main mode time coefficient-forecast time effectiveness-optimal prediction factor component relationship table, determine all optimal prediction factors associated with any temperature main mode time coefficient under the preset forecast time effectiveness to form the optimal prediction factor set.

[0108] In this embodiment, the main components of the temperature mode in China are: PC1-PC6. From the relationship table of the main components of the temperature mode-forecast timeliness-optimal forecast factors, it can be obtained that when forecasting 20 days in advance, the corresponding optimal forecast factors of PC1 include: the subseasonal components of PW, T850, U850 and OLR, and the LFBS components of T850, PW and H500; the corresponding optimal forecast factors of PC2 include: the subseasonal components of U200, U500, H500 and V850, and the LFBS components of U200, V200 and T850; similarly, when forecasting PC3-PC6 20 days in advance, the corresponding optimal forecast factor components are obtained.

[0109] Step 2: Acquire data of each optimal predictor in the optimal predictor set, and extract the subseasonal component and low-frequency background state LFBS component of each optimal predictor respectively through real-time non-bandpass filtering.

[0110] Step 3: Input the subseasonal component and low-frequency background state LFBS component of each optimal forecast factor into the pre-trained subseasonal temperature forecast model to obtain the predicted value of each main mode time coefficient of the temperature in the target area under the time effect.

[0111] Among them, the process of obtaining the predicted value of any main modal time coefficient of temperature includes: the sub-seasonal temperature forecast model extracts features of the optimal forecast factor component corresponding to the main modal time coefficient of temperature to obtain the deep features of the optimal forecast factor component; the deep features of all optimal forecast factor components corresponding to the main modal time coefficient of temperature are integrated to obtain joint features containing multi-scale multivariate interaction information; and the predicted value of the main modal time coefficient of temperature is obtained by combining the joint features and the current month information.

[0112] Specifically, the values ​​of PC1-PC6 for the 20-day-ahead forecast for the Chinese region include:

[0113] For the main mode time coefficient PC1 of air temperature:

[0114] like Figure 3 As shown in the figure, the optimal predictor including the subseasonal components of PW, T850, U850 and OLR, and the LFBS components of T850, PW and H500 are input into the subseasonal temperature forecast model. The convolution blocks corresponding to the seven input layers respectively extract the deep features contained in the input features provided by the optimal predictor components; the deep features of the seven optimal predictor components corresponding to PC1 are stacked in the channel dimension and further subjected to two convolution operations to obtain the joint features of the optimal predictor; the joint features and the current month information are combined to predict the PC1 value.

[0115] The same method is used to predict the values ​​of PC2-PC6.

[0116] Step 4: Obtain the predicted temperature anomaly field under the time effect according to the predicted value of each temperature main mode time coefficient and the temperature main mode reconstruction;

[0117] In this embodiment, the predicted temperature anomaly field in China is obtained by reconstructing the six main temperature mode time coefficients (PC1-PC6) and the pre-obtained EOF main mode.

[0118] Step 5: Predicting sub-seasonal heat wave events in the target area based on the predicted temperature anomaly field under the time limit and the preset threshold standard;

[0119] Specifically, the sub-seasonal heat wave events in China are predicted based on the predicted temperature anomaly field reconstructed from the main temperature mode time coefficients (PC1-PC6) in step 2 and the pre-obtained main temperature mode.

[0120] It should be noted that in observations or forecasts, when the grid temperature anomaly exceeds the 90th percentile for at least three consecutive days, it is determined to be a heat wave event.

[0121] Example 3

[0122] This example demonstrates the performance of the proposed method in subseasonal heat wave forecasting. The temporal correlation coefficient (TCC) and spatial correlation coefficient (PCC) are used to evaluate the model's ability to predict the temporal evolution and spatial modal characteristics of temperature anomalies. The root mean square error (RMSE) measures the magnitude of the forecast error, and the extreme dependence index (EDI) assesses the accuracy of heat wave predictions.

[0123]

[0124] where H is the hit rate, defined as the ratio of correctly predicted observed heatwaves to the total observed heatwaves, F is the false alarm rate, calculated as the ratio of the number of false alarms to the total number of observed non-events, and EDI ranges from −1 to 1, with 1 indicating perfect prediction and 0 indicating no skill.

[0125] Figure 4 The TCC skill of the CNN model used in the present invention and the CMA and ECMWF models in predicting temperature anomalies is shown. This skill is used to measure the model's ability to predict the temporal evolution of temperature anomalies. The CNN model has a lower skill when 2-3 pentads in advance ( Figure 4 ab), is not as good as the dynamic model ( Figure 4 fl, but outperforms the CMA and ECMWF models in forecasts beyond 20 days ahead ( Figure 4 Even up to 6 months ahead, the CNN model still has proficient (i.e., TCC with 99% confidence level) prediction skills over most of China ( Figure 4 e), while the dynamical model mainly performs well in western and southwestern China ( Figure 4 jo).

[0126] Figure 5 The CNN model used in the present invention and the PCC technique of temperature anomaly prediction by the CMA and ECMWF models are shown. This technique is used to evaluate the prediction ability of each model for the spatial distribution of temperature anomalies in China. Figure 5a is the temporal average of the PCC for each forecast result. Similar to the TCC technique, the CNN model struggles to match the dynamical model within a 3-pentad lead. However, compared to the CMA and ECMWF models, they demonstrate superior forecasting capabilities at 4-6 pentad lead times. Figure 5 b shows the percentage of PCC prediction frequencies that pass the 99% confidence level test, further confirming the superior prediction capability of the CNN model used in the present invention at 4-6 pentad years in advance.

[0127] In addition, if Figure 6 As shown in Figure 2, when the forecast lead time exceeds 20 days, the RMSE value of the CNN model is smaller than that of the dynamic model, and the average RMSE of the Chinese region is always less than 1 standard deviation when the forecast lead time is 10-30 days. Figure 4 and Figure 6 It can be seen that regions with higher TCC generally exhibit smaller prediction errors.

[0128] Figure 7 We further compared the EDI between CNN-based and dynamical predictions. 2-6 pentads in advance, both CNN and dynamical models can effectively predict heat waves in most parts of China (EDI > 0). Within a shorter forecast horizon (3 pentads), the ECMWF model showed the highest skill among the three models ( Figure 7 kl), the CMA model also performs well in the 2-month-ahead forecast ( Figure 7 f). However, in long-term (more than 20 days) forecasts, the CNN model has a higher forecast accuracy for temperature ( Figure 4-6 ), which has better forecasting ability for heat waves than the dynamic model ( Figure 7 rows 3-5), especially in the densely populated eastern part of China.

[0129] Below, based on the Grad-CAM method, a physical explanation of the prediction using the CNN model is given.

[0130] Figure 8 The attribution map shows the period when the CNN model accurately predicts the main temperature mode corresponding to the high value period of the time coefficient PC1 (exceeding the 90th percentile threshold) 20 days in advance. Figure 8 ag shows the results of the last Conv2d layer of each convolutional block, highlighting the key features of the predictors located in Northwest Eurasia and the tropics, which are important for the model to produce accurate predictions.

[0131] Figure 9-10The evolution of these predictors is shown, with a focus on precursor signals from tropical and temperate regions. This demonstrates that the CNN model is able to identify the origin and propagation of the preceding subseasonal signal associated with periods of high PC1 values, and also demonstrates the model's ability to capture the impact of interannual variability on temperature changes. Figure 8 h) shows the performance of all predictors in the sub-seasonal temperature forecast model ( Figure 3 The results show that large-scale signals from Northwest Eurasia, Northeast Asia, the tropical Indian Ocean, and local areas of China provide a source of predictability for the 20-day-ahead forecast.

[0132] Example 4

[0133] According to a second aspect of the present invention, a computer storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the sub-seasonal heat wave prediction method as described in Example 1 is implemented.

[0134] The storage medium may include, for example, a storage component of a tablet computer, a hard disk of a computer, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disk read-only memory (CD-ROM), a USB memory, or any combination thereof. The computer storage medium may be any combination of one or more computer storage media.

[0135] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A sub-seasonal heat wave prediction method, characterized in that: include: According to the pre-obtained temperature main mode of the target area and the temperature main mode time coefficient-forecast time effectiveness-optimal forecast factor component relationship table, all optimal forecast factors associated with any temperature main mode time coefficient under the preset forecast time effectiveness are determined to form an optimal forecast factor set; Acquire data of each optimal predictor in the optimal predictor set, and extract the subseasonal component and the low-frequency background state (LFBS) component of each optimal predictor respectively through real-time non-bandpass filtering; Inputting the subseasonal component and the low-frequency background state (LFBS) component of each optimal forecast factor into a pre-trained subseasonal temperature forecast model to obtain a predicted value of each main mode time coefficient of the temperature in the target area under the time effect; Obtaining a predicted temperature anomaly field under the time effect according to the predicted value of each temperature main mode time coefficient and the temperature main mode reconstruction; Predicting sub-seasonal heat wave events in the target area based on the predicted temperature anomaly field under the time validity period and the preset threshold standard; The process of obtaining the predicted value of any temperature main mode time coefficient includes: The sub-seasonal temperature forecast model performs feature extraction on the optimal forecast factor component corresponding to the main mode time coefficient of the temperature to obtain the deep features of the optimal forecast factor component; The deep features of all the optimal predictor components corresponding to the main mode time coefficient of the temperature are fused to obtain the joint features containing multi-scale and multi-variable interaction information; The predicted value of the main modal time coefficient of the temperature is obtained by combining the joint features and the current month information.

2. The sub-seasonal heat wave prediction method according to claim 1, characterized in that: The sub-seasonal temperature forecast model also includes: a univariate forecast model; the univariate forecast model is used to provide the temperature modal time coefficient-forecast time-optimal forecast factor relationship table and a pre-training convolution block for extracting deep features of each optimal forecast factor component during the training process of the sub-seasonal temperature forecast model.

3. The sub-seasonal heat wave prediction method according to claim 2, characterized in that: The training process of the sub-seasonal temperature forecast model includes: Determining a target area and determining candidate predictors influencing daily maximum temperature within the target area; Obtain the daily maximum temperature data for a preset historical period in the target area and the data of the candidate forecast factors, construct a data set and perform preprocessing, and divide the processed data into a training set, a validation set, and a test set according to the year; A non-bandpass filtering method is used to extract the subseasonal component and the low-frequency background state LFBS component of the candidate predictor; The summer temperature data is obtained from the daily maximum temperature data in the training set, and the empirical orthogonal function (EOF) is used for analysis to obtain S main temperature modes and their corresponding time coefficients. The full-year data of each candidate predictor is obtained from the training set. The subseasonal component and LFBS component of each candidate predictor are used as the input of the univariate prediction model, and the time coefficient of the S main temperature modes is used as the output of the univariate prediction model to train the univariate prediction model. The subseasonal components or LFBS components of each candidate predictor in the validation set are input into the corresponding univariate prediction model, and S types of temperature main mode time coefficients are predicted separately at different time periods in the future. By evaluating and comparing the prediction skills of the univariate prediction models, the a subseasonal components of the predictors and b LFBS components of the predictors that perform best for each temperature main mode time coefficient at each time period are determined, forming a relationship table of temperature main mode time coefficient, forecast time period, and optimal predictor component; where a and b are determined based on sensitivity tests; The data of the (a+b) optimal predictors corresponding to each main temperature mode time coefficient in the training set are used as input, and each main temperature mode time coefficient is used as the output of the sub-seasonal temperature forecast model. The training is performed in combination with the pre-trained convolution blocks provided by the univariate forecast model corresponding to the (a+b) optimal predictors; Use the validation data to evaluate the trained subseasonal temperature forecast model and optimize the model structure and parameters; The independent forecast performance of the subseasonal temperature forecast model was tested using data from the test set.

4. The sub-seasonal heat wave prediction method according to claim 3, characterized in that: The univariate prediction model includes: 2 input layers, 1 convolutional block, 1 fully connected layer and 1 output layer; For any candidate predictor: The first input layer provides a corresponding feature map of the subseasonal component or LFBS component of the candidate predictor; The convolution block is used to perform convolution and pooling operations on the feature map to extract corresponding deep features; The second input layer provides information for initializing the current month; The fully connected layer integrates the deep features and the monthly information, and outputs the time coefficient of the specified temperature main mode under the future specified forecast time through the output layer; The sub-seasonal temperature forecast model includes: (a+b+1) input layers, (a+b) convolution blocks, 2 two-dimensional convolution layers, 1 fully connected layer and 1 output layer; The (a+b) input layers are respectively arranged in (a+b) independent input channels; and are used to provide (a+b) optimal prediction factors corresponding to input feature maps; The (a+b) convolution blocks are respectively used to extract deep features corresponding to the (a+b) optimal predictor factors; The two-dimensional convolution layer is used to perform further convolution operations on the result of splicing the deep features of the (a+b) optimal forecast factors in the channel dimension direction. The convolution result and the monthly information provided by the last input layer are integrated through the fully connected layer, and the time coefficient of the specified main mode of temperature under the future specified forecast time is output through the output layer.

5. The sub-seasonal heat wave prediction method according to claim 4, characterized in that: The convolution block includes three two-dimensional convolutional layers and two maximum pooling layers arranged at intervals.

6. The sub-seasonal heat wave prediction method according to claim 5, characterized in that: The extracting of the subseasonal component and the LFBS component of the candidate predictor by using a non-bandpass filtering method includes: Subtract the daily climate state from the raw daily data of the candidate predictor to eliminate the annual climate cycle and obtain the daily anomaly value; the daily climate state is the multi-year average of the same date from 1991 to 2020; Subtract the average of the previous 45 days from the daily outlier to eliminate the signal of more than 90 days; get the outlier of less than 90 days; A 5-day moving average is used for the abnormal values ​​less than 90 days to eliminate the high-frequency weather-scale disturbances less than 10 days, and the subseasonal component of the candidate forecast factor is obtained; The LFBS component is the 30-day average of the daily anomalies.

7. The sub-seasonal heat wave prediction method according to claim 5, characterized in that: The filter size of the two-dimensional convolution layer is 3×3, the sliding step value is 1×1, and the activation function of the two-dimensional convolution layer and the fully connected layer is a rectified linear unit with a leakage parameter; the filter size of the maximum pooling layer is 2×2, and the sliding step value is 2×2; Before the two-dimensional convolutional layer performs the convolution operation, it also includes: padding the boundary of the feature map input to the two-dimensional convolutional layer with zero values ​​so that the size of the feature map remains unchanged before and after the operation.

8. The sub-seasonal heat wave prediction method according to claim 5, characterized in that: The candidate prediction factors include: 850 hPa temperature T850, 200 hPa zonal wind U200, 500 hPa zonal wind U500, 850 hPa zonal wind U850, 200 hPa meridional wind V200, 500 hPa meridional wind V500, 850 hPa meridional wind V850, 200 hPa geopotential height field H200, 500 hPa geopotential height field H500, 850 hPa geopotential height field H850, 700 hPa relative humidity RH700, precipitable water PW, sea level pressure SLP, surface soil moisture SSM and outward longwave radiation OLR.

9. The sub-seasonal heat wave prediction method according to claim 5, characterized in that: During the training process of the sub-seasonal temperature forecast model, the Adam optimizer was used to optimize the training parameters. The initial learning rate of the Adam optimizer was 1×10 −3 ; The mean square error is used as the loss function; A dropout method and an early stopping method are used to prevent the model from overfitting. The dropout method includes: adding a dropout layer with a preset probability of discarding after each two-dimensional convolutional layer during the model training iteration; The early stopping method includes: in response to no improvement in skill over T consecutive training iterations on the validation set, stopping training to ensure effective utilization of model weights corresponding to the minimum validation loss, where T is a preset value.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the sub-seasonal heat wave prediction method according to any one of claims 1 to 9.

Images