Half-year drought event prediction method and system based on sea and land gas physical quantity driving deep learning model

Abstract

The invention discloses a half-year drought event prediction method and system based on a sea-land gas physical quantity driving deep learning model. The method comprises the following steps: acquiring global reanalysis weft wind U, warp wind V, vertical wind speed OMEGA, surface pressure SP, sea temperature SST, temperature T, potential height GPH, outward long-wave radiation OLR, second-layer soil humidity SWVL and sea ice coverage SICOC data; the reanalysis data are integrated into day-by-day data, an atmosphere apparent heat source Q1 is calculated, and all variables are processed into standard distance flatness; extracting SST, SWVL, Q1, OLR, 200hPa RV and 100hPa GPH combination as predictive factors day by day in the past 180 days to hitherto; inputting the integrated prediction factors into a trained SimVP model, and predicting a global grid scale drought index in the next 180 days; inputting the drought index of the next 180 days into the three-dimensional DBSCAN model to generate a three-dimensional drought event drought body; and extracting regional drought events, and generating a drought frequency diagram. According to the invention, real-time prediction of the drought event can be realized more quickly and more accurately.

Description

Technical Field

[0001] This invention relates to a method and system for predicting drought events, and more particularly to a method and system for predicting six-month drought events based on a deep learning model driven by land, sea and air physical quantities. Background Technology

[0002] Against the backdrop of global warming, the frequency, intensity, and complexity of drought events have increased significantly, thus necessitating progress in research on ultra-long-term (six-month) drought prediction. In terms of dynamic-statistical coupled models, combining atmospheric circulation models (such as ECMWF SEAS5) with statistical correction techniques typically extends the drought prediction period to three months or more. Sea surface temperature anomalies (such as ENSO / IOD phases) can also be used as precursor signals to predict regional drought changes. In the field of deep learning applications, dynamic-statistical hybrid models based on LSTM (Long Short-Term Memory) outperform pure dynamic models in prediction periods of 30–90 days. By mining the statistical relationships between historical drought events and circulation factors, they significantly improve prediction accuracy. Furthermore, breakthroughs have been achieved in three-dimensional drought event identification technology. For example, the Global Hydro-Climate Extreme Event Set (Glo3DHydroClimEventSet) constructed based on the three-dimensional DBSCAN algorithm has achieved accurate spatiotemporal clustering of large-scale drought events, providing a sample basis for mechanism research and event-scale prediction.

[0003] Despite some progress, existing methods still have three limitations. First, there is insufficient analysis of physical processes. Most models rely on statistical correlations and lack quantitative embedding of drought-causing physical chains (such as multi-sphere interactions between the ocean, land, and atmosphere). For example, while drought predictions under ENSO forcing can identify sea surface temperature anomalies, they do not explicitly couple local thermodynamic processes such as atmospheric apparent heat sources and relative vorticity, resulting in weak physical interpretability. Second, there is a lack of prediction of three-dimensional drought events. Existing research focuses on drought index (such as SPI) predictions, ignoring the evolutionary characteristics of drought as a "three-dimensional spatiotemporal continuum." Traditional methods cannot output the spatial morphology of future drought events (such as migration paths and three-dimensional structures), which restricts disaster chain prevention and control decisions. Finally, there is a bottleneck in the accuracy of ultra-long-term predictions. When the forecast period exceeds 90 days, the prediction skill of dynamical model-based prediction models drops sharply due to amplified initial field errors.

[0004] To address the aforementioned shortcomings, deep learning models integrating physical quantities from multiple ocean, land, and atmosphere spheres have emerged as a breakthrough direction. Drought is essentially a product of ocean-land-atmosphere interactions. Research shows that ocean forcing (such as the Indo-Pacific sea surface temperature anomaly) modulates atmospheric circulation, triggering the mid-latitude Rossby wave train, leading to enhanced subsidence and precipitation deficit; land surface processes (such as soil moisture-atmospheric feedback) can prolong drought duration. Ignoring any of these factors amplifies prediction uncertainty. Therefore, standardizing dynamic / thermodynamic variables such as Q1, OLR, and RV as input factors can physically constrain deep learning models and improve the interpretability of the mechanisms. Employing video prediction architectures such as SimVP, which encode multi-scale processes using a four-dimensional tensor (time × variable × longitude × latitude), can simultaneously capture the spatiotemporal propagation characteristics of drought. Furthermore, SPAI based on 3D DBSCAN clustering prediction can directly output the spatiotemporal boundaries of the "drought event body" (such as the scope of impact and duration), supporting precise disaster chain prevention and control. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to propose a method and system for predicting six-month drought events based on a deep learning model driven by land, sea and air physical quantities. This method can more quickly and accurately predict drought events in real time, determine the future drought-affected areas, and improve the ability to proactively combat drought.

[0006] Technical solution: This invention includes the following steps:

[0007] S1. Acquire global reanalysis data including zonal wind (U), meridional wind (V), vertical wind speed (OMEGA), surface air pressure (SP), sea surface temperature (SST), temperature (T), geopotential height (GPH), outward longwave radiation (OLR), second-layer soil moisture (SWVL), and sea ice coverage (SICOC).

[0008] S2. Integrate the reanalysis data into daily data and calculate the atmospheric apparent heat source Q1, and process all variables into standardized anomalies;

[0009] S3. Extract daily SST, SWVL, Q1, OLR, 200hPa RV, and 100hPa GPH from the past 180 days to the present, and combine them as predictive factors;

[0010] S4. Input the integrated predictive factors into the trained SimVP model to predict the global grid-scale drought index for the next 180 days.

[0011] S5. Input the drought index for the next 180 days into the trained 3D DBSCAN model to generate a 3D drought event drought body.

[0012] S6. Extract regional drought events and generate a drought frequency map.

[0013] Step S2 specifically involves: averaging the downloaded hourly ERA5 reanalysis data for the past 365 days on a daily scale to calculate the daily average values ​​of land, sea, and air physical quantities for the past 365 days, and then calculating Q1 step by step based on the thermodynamic equation.

[0014] The predictor is a five-dimensional [time, variable, anomaly, longitude, latitude] predictor.

[0015] Step S4 specifically involves inputting the integrated five-dimensional predictive factors into the trained SimVP model to automatically predict the three-dimensional [time, longitude, latitude] global grid scale SPAI for the next 180 days.

[0016] Step S5 specifically involves: setting a drought threshold based on the SPAI nc file for the next 180 days, and extracting drought grid points; performing DBSCAN clustering on the drought grid points in the region, filtering each clustered event, converting the filtered events into nc files, with each file representing a drought event and containing the spatial distribution of the event at each time step; for each drought event, fusing grid points around the drought core area that meet the conditions to expand the scope of the drought event, and finally outputting the nc file of the drought event for each event.

[0017] Step S6 specifically involves: reading the nc file of each drought event, extracting drought events in the affected area, setting multiple different forecast periods, statistically analyzing the temporal and spatial frequency distribution of each drought event, and plotting the spatial drought frequency distribution for each segment.

[0018] A prediction system for a six-month drought event prediction method based on a deep learning model driven by land, sea and air physical quantities includes: a reanalysis data acquisition module, a reanalysis data preprocessing module, a predictor factor construction module, a drought index spatiotemporal prediction module, a three-dimensional drought event volume extraction module, and a regional drought frequency map generation module.

[0019] The predictor construction module is used to extract the 30-day and 90-day standardized anomalies of six variables—SST, SWVL, Q1, OLR, 200hPa RV, and 100hPa GPH—from the past 180 days to the present at the daily global grid scale, and combine them into a five-dimensional [time, variable, anomaly, longitude, latitude] predictor.

[0020] The drought index spatiotemporal prediction module is used to input the integrated five-dimensional predictive factors into the trained SimVP model to predict the global grid-scale cumulative precipitation anomaly index for the next 180 days.

[0021] The three-dimensional drought event volume extraction module is used to input the generated global grid-scale SPAI for the next 180 days into the trained 3D DBSCAN model, and cluster the SPAI grid points with values ​​less than the drought threshold into three-dimensional drought event volumes.

[0022] Beneficial effects: This invention can achieve drought prediction at a daily grid scale across the country, and can obtain a forecast of the drought situation for the next 180 days; it can achieve three-dimensional drought event volume prediction for the next 180 days across the country, breaking through the two-dimensional drought prediction that relies solely on the drought index; it can achieve rolling prediction, with ERA5 reanalysis data updated daily, thus enabling rolling operation to predict drought events for the next 180 days. Attached Figure Description

[0023] Figure 1 is a flowchart of the present invention;

[0024] Figure 2 shows the time-longitude evolution of the SimVp+3D DBSCAN model in the embodiment of the present invention during the 2022 drought in the Yangtze River Basin, based on actual measurements and predictions. Detailed Implementation

[0025] The invention will now be further described with reference to the accompanying drawings.

[0026] Example 1

[0027] like Figure 1 As shown in this embodiment, the six-month drought event prediction method based on a deep learning model driven by ocean, land, and air physical quantities uses the standardized anomalies of the ocean, atmosphere, and surface over the previous 180 days to drive the SimVP deep learning model, thereby predicting the SPAI index for the next 180 days. Then, the 3D DBSCAN algorithm is used to identify the three-dimensional drought event volume, enabling faster and more accurate real-time prediction of drought events, determining future drought areas, and improving proactive drought mitigation capabilities. Specifically, it includes the following steps:

[0028] S1. Acquire global reanalysis data for zonal wind (U), meridional wind (V), vertical wind speed (OMEGA), surface air pressure (SP), sea surface temperature (SST), temperature (T), geopotential height (GPH), outward longwave radiation (OLR), second-layer soil moisture (SWVL), and sea ice cover (SICOC).

[0029] Using Python code, the system automatically downloads 365 days of ERA5 reanalysis data for 6 hours at a time, covering global (90°N~60°S, -180°~180°S) data in 1°×1° grids, including single-layer variables (SP, OLR, SWVL, SICOC, SST), 32-layer variables (U, V, OMEGA) in the troposphere from 10 to 1000 hPa, and 27-layer variables (T, GPH) in the troposphere from 100 to 1000 hPa. All data is saved as nc files for further processing.

[0030] S2. The reanalysis data is integrated into daily data, and atmospheric apparent heat source (Q1) is calculated using variables such as temperature and wind field. All variables are processed into standardized anomalies.

[0031] The downloaded 6-hourly ERA5 reanalysis data for the past 365 days were averaged on a daily scale to calculate the daily average values ​​of land, sea and air physical quantities for the past 365 days, and then Q1 was calculated step by step according to the thermodynamic equation.

[0032] First, calculate the potential temperature of each pressure layer. :

[0033]

[0034] In the formula, This represents the actual temperature of each pressure layer (unit: K). The reference pressure is usually taken as 1000 hPa, which is the standard sea level pressure. It is the actual gas pressure of the gas block, which is GPH (unit: nitrile) in this invention. Same, usually hPa). It is the gas constant. Compared with specific heat capacity at constant pressure The ratio (for dry air, =287J / (kg·K), =1004J / (kg·K), therefore ).

[0035] Calculate each pressure layer based on thermodynamic equations. :

[0036]

[0037] In the formula, This represents the local variation of potential temperature. For the potential temperature advection term, For vertical transport items, Represents the UV wind field vector. For Hamiltonian operators.

[0038] For all pressure layers Vertical integration is performed to obtain ViQ1. Finally, the 90-day standardized anomaly at the global grid scale is calculated for six variables: SST, SWVL, ViQ1, OLR, RV, and GPH.

[0039] S3. Extract six sets of variables from the past 180 days to the present: SST, SWVL, Q1, OLR, 200hPa RV (relative vorticity at 200 hPa tropospheric height), and 100hPa GPH. Combine them into predictive factors.

[0040] The trained SimVP model predicts global (90°N~60°S, -180°~180°S) 1°×1° grid single-layer variables (SST, SWVL, OLR), 200hPa RV and 100hPa GPH extracted from the troposphere, and 30-day and 90-day standardized anomalies of the whole-layer integral ViQ1. The above six types of variables are extracted every 5 days from the past 180 days to the present and combined into a five-dimensional [time, variable, anomaly, longitude, latitude] predictor.

[0041] S4. Input the integrated predictive factors into the trained SimVP model to predict the global grid-scale drought index for the next 180 days.

[0042] The integrated five-dimensional predictive factors are input into the trained SimVP model to automatically predict the global grid-scale SPAI for the next 180 days, i.e., the three-dimensional [time, longitude, latitude] SPAI. The SimVP model architecture mainly consists of an Encoder, a Decoder, Mid_Xnet, and Fusion. The Encoder downsamples and encodes the input data (such as meteorological variables) using the ConvSC module (with alternating strides of 1 / 2). The Decoder upsamples and reconstructs the prediction results, fusing shallow features through transposed convolutions and skip connections. Mid_Xnet is the core of temporal modeling, using the Inception module (multi-scale convolution) to process temporal features. Fusion fuses multivariate features through CBAM (channel + spatial attention) and LayerNorm.

[0043] S5. Input the drought index for the next 180 days into the pre-trained 3D DBSCAN model to generate a 3D drought event drought body.

[0044] Based on the nc file of SPAI for the next 180 days, set the drought threshold to -1.5 and extract the drought grid points (i.e., grid points where SPAI ≤ -1.5). Perform DBSCAN clustering on the drought grid points in the Chinese region (three dimensions: longitude, latitude, time). Filter each clustering event (according to thresholds such as coverage area, land area, duration, etc.). Convert the filtered events into nc files, with each file representing a drought event and containing the spatial distribution at each time step of that event (binary: 1 represents drought, NaN represents non-drought). For each drought event, incorporate the grid points around the drought core area (SPAI ≤ -1.5) that meet the condition (-1.5 < SPAI ≤ -1.0) to expand the scope of the drought event, and finally output the nc files of the场次 drought events.

[0045] S6. Extract the drought events in the Chinese region and generate a drought frequency map.

[0046] Read each nc file of the场次 drought events, extract the drought events affecting the Chinese region, set three different lead time periods of 1 - 60 days, 61 - 120 days, and 121 - 180 days, and statistically analyze the frequency distribution of each drought event in time - space (i.e., the number of times each grid point is identified as drought at each time step), and plot the spatial drought frequency distribution for each period.

[0047] Embodiment 2

[0048] The semi - annual drought event prediction system based on the physical quantities of sea, land and atmosphere driving a deep learning model in this embodiment includes:

[0049] Re - analysis data acquisition module: Download the global grid - scale single - layer variables (SP, OLR, SWVL, SICOC, SST) every 6 hours and tropospheric multi - layer variables (U, V, OMEGA, T, GPH) for the past 365 days from the ECMWF ERA5 dataset.

[0050] Re - analysis data pre - processing module: Calculate the daily average of the global grid - scale re - analysis data every hour for the past 365 days, and then calculate the local change term, advection term and vertical convection term step by step based on the thermodynamic equation, sum the three terms to obtain the atmospheric apparent heat source (Q1), and finally calculate the 30 - day and 90 - day standardized anomalies of all sea, land and atmosphere variables for the past 365 days.

[0051] Predictor construction module: Extract the 30 - day and 90 - day standardized anomalies of six variables, namely SST, SWVL, Q1, OLR, 200hPa RV, and 100hPa GPH, at the global grid - scale on a daily basis from the past 180 days to date (every 5 days), and combine them into a five - dimensional [time, variable, anomaly, longitude, latitude] predictor.

[0052] Drought Index Spatiotemporal Prediction Module: The integrated five-dimensional predictor factors are input into the trained SimVP model to predict the global grid-scale cumulative precipitation anomaly index (SPAI) for the next 180 days.

[0053] 3D drought event volume extraction module: The generated global grid-scale SPAI for the next 180 days is input into the pre-trained 3D DBSCAN model, and SPAI grid points with values ​​less than the drought threshold are clustered into 3D drought event volumes.

[0054] Regional drought frequency map generation module: Extracts drought events affecting China, and statistically analyzes grid points for drought occurrence within different forecast periods; finally, it exports the image.

[0055] This invention applies the SimVP and DBSCAN models to the field of drought prediction, improving the accuracy of six-month drought event prediction. It utilizes global land-sea-ocean variables as predictive factors to achieve accurate prediction of the national three-dimensional drought event volume.

[0056] As shown in Figure 2, the SimVP+3D DBSCAN model, in response to the severe summer-autumn drought in the Yangtze River in 2022 (August 11 to December 4), was able to effectively predict the occurrence of the drought event 75 days (approximately 2.5 months) in advance during the control forecasting experiment. Although the predicted drought intensity was mild to moderate, it could robustly capture the seasonal drought event process and its location. It could also accurately predict the spatial distribution of the drought mitigation phase 5 months in advance. For the summer drought in the middle reaches of the Yellow River in 2023, the SimVP+3D DBSCAN model was able to effectively predict the location and course of the drought event at each time point, starting from the moment the drought began.

Claims

1. A method for predicting semi-annual drought events based on a land-sea-air physical quantity driven deep learning model, characterized in that, Comprise the following steps: S1, obtain global reanalysis zonal wind U, meridional wind V, vertical wind speed OMEGA, surface pressure SP, sea surface temperature SST, temperature T, geopotential height GPH, outward long wave radiation OLR, second layer soil moisture SWVL, sea ice coverage SICOC data; S2, integrate the reanalysis data into daily, and calculate the atmospheric apparent heat source Q1, and process all variables into standardized anomaly; S3, extract daily SST, SWVL, Q1, OLR, 200hPa RV, 100hPa GPH from the past 180 days to date, and combine them into prediction factors; S4, input the integrated prediction factors into the trained SimVP model to predict the global grid scale drought index in the next 180 days; S5, input the drought index in the next 180 days into the trained three-dimensional DBSCAN model to generate a three-dimensional drought event drought body; S6, extract regional drought events and generate a drought frequency map.

2. The method of claim 1, wherein the method is characterized by, The step S2 is specifically: averaging the downloaded hourly ERA5 reanalysis data of the past 365 days on a daily scale to calculate the daily mean of the sea-land atmospheric physical quantity in the past 365 days, and then calculating Q1 step by step according to the thermodynamic equation.

3. The method of claim 1, wherein the method is characterized by: The prediction factor is a five-dimensional [time, variable, anomaly, longitude, latitude] prediction factor.

4. The method of claim 3, wherein the method is characterized by, The step S4 is specifically: inputting the integrated five-dimensional prediction factor into the trained SimVP model to automatically predict the three-dimensional [time, longitude, latitude] global grid scale SPAI in the next 180 days.

5. The method of claim 4, wherein the method is characterized by: The step S5 is specifically: based on the nc file of the SPAI in the next 180 days, setting a drought threshold to extract drought grid points; performing DBSCAN clustering on the drought grid points of a region, filtering each clustering event, converting the filtered event into an nc file, and each file representing a drought event containing the spatial distribution of the event at each time step; for each drought event, fuse the grid points around the drought core area that meet the conditions to expand the range of the drought event, and finally output the nc file of the field drought event.

6. The method of claim 5, wherein the method is characterized by: The step S6 is specifically: reading each field drought event nc file, extracting the drought events affecting the region, setting a plurality of different forecast period time periods, and counting the frequency distribution of each drought event in time and space, and drawing the spatial drought frequency distribution of each period.

7. A prediction system suitable for the semi-annual drought event prediction method based on the deep learning model driven by the sea-land-air physical quantity of any one of claims 1-6, characterized in that, Comprise: The reanalysis data acquisition module, the reanalysis data preprocessing module, the prediction factor construction module, the drought index space-time prediction module, the three-dimensional drought event body extraction module and the regional drought frequency map generation module.

8. The system for predicting semi-annual drought events based on a deep learning model driven by land-air-sea physical quantities according to claim 7, wherein, The prediction factor construction module is used for extracting 30-day and 90-day standardized anomalies of six variables of daily global grid scale SST, SWVL, Q1, OLR, 200hPa RV and 100hPa GPH from the past 180 days to date, and combining them into five-dimensional [time, variable, anomaly, longitude, latitude] prediction factors.

9. The system for predicting semi-annual drought events based on a deep learning model driven by land-air-sea physical quantities according to claim 7, wherein, The drought index spatiotemporal prediction module is configured to input the integrated five-dimensional prediction factors into a trained SimVP model to predict a global grid scale accumulated precipitation anomaly index in the next 180 days.

10. The system according to claim 7, wherein the system is characterized by: The three-dimensional drought event body extraction module is configured to input the generated global grid scale SPAI in the next 180 days into a trained 3D DBSCAN model to cluster SPAI grid points with values less than a drought threshold into three-dimensional drought event bodies.

Images