Prediction system and method for high-temperature heat wave drought composite disasters

Abstract

The invention relates to the field of disaster prediction, and particularly discloses a prediction system and method for high-temperature heat wave drought composite disasters, and the system comprises a multi-source data collection module which is used for collecting meteorological data, topographic data, vegetation and soil data and human activity data, and constructing a multi-dimensional input feature set; the data processing module is used for carrying out space-time alignment, normalization processing and abnormal value elimination on the acquired meteorological data, topographic data, vegetation and soil data and human activity data, and carrying out dimensionality reduction through a principal component analysis method to extract key features; the deep learning model training module adopts a CNN-RNN fusion architecture; and the prediction output module is used for generating a composite disaster prediction result through the trained deep learning model training module. By adopting the technical scheme of the invention, multi-dimensional data can be considered, a multi-element coupling model is constructed, and prediction of spatial and temporal distribution, evolution rate and coupling strength of high-temperature heat wave drought composite disasters in complex geographic units and ecologically fragile areas is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of disaster prediction, in particular to a high-temperature heat wave and drought compound disaster prediction system and method. BACKGROUND

[0002] With the intensification of global climate change, high-temperature drought and drought compound disasters occur frequently, posing a significant threat to agricultural production, ecosystems, and human society. Existing disaster prediction techniques rely on single-dimensional meteorological data or simple environmental parameters, making it difficult to fully capture the complex evolution mechanism of compound disasters, resulting in low prediction accuracy.

[0003] For example, the high-temperature disaster prediction method disclosed in Chinese Patent Publication No. CN114626512A only uses daily maximum temperature, precipitation, wind speed, and other meteorological elements as inputs, without incorporating spatial features such as terrain and topography (e.g., elevation, slope). This method fails to consider the blocking effect of mountains on air flow, resulting in a prediction error of more than 20% in mountainous areas. Similarly, the high-temperature disaster assessment method disclosed in Chinese Patent Publication No. CN118014161A involves terrain planning to monitor the region, but only uses the distance between stations as the only spatial parameter, without in-depth analysis of the impact of terrain on heat distribution (e.g., the temperature difference between sunny and shady slopes can reach 5-8℃), making it difficult to accurately reflect the disaster distribution pattern in complex geographical environments.

[0004] In addition, the drought prediction method disclosed in Chinese Patent Publication No. CN114841064A focuses on meteorological data and semi-supervised learning algorithms, without considering ecological parameters such as vegetation coverage and soil texture. However, vegetation transpiration can increase near-surface humidity by 10%-15%, significantly mitigating drought development (e.g., drought duration in forest-covered areas is shortened by 3-5 days compared to bare land). Existing models ignore such ecological factors, resulting in a prediction bias of up to 18% for the coupling effect of drought and high temperature, and are unable to accurately capture the vicious cycle mechanism of "high temperature-vegetation degradation-drought intensification".

[0005] Chinese Patent Application No. CN202410914894.9 (Patent Title: High-temperature drought compound disaster monitoring and early warning method and system) and Chinese Patent Application No. CN202311267216.X (Patent Title: High-temperature drought weather early warning method based on Transformer, electronic device) also disclose related technologies.

[0006] Human activity data (such as agricultural irrigation, urbanization level) is not included in the above technologies. For example, winter wheat irrigation in the North China Plain can increase local soil moisture by 8%-12%, effectively delaying the progress of drought. The above technologies do not consider such intervention factors, resulting in a prediction of high-temperature and drought composite disasters in the irrigation area that is 15% stronger than the actual situation. In addition, the heat island effect caused by urbanization (the number of high-temperature days increases by 3-5 days for every 10% increase in urbanization rate) is not captured by existing models, further reducing the practical application value of the prediction.

[0007] The above technologies have single-dimensional data (lack of spatial ecological data such as terrain, vegetation, and soil, and human activity intervention data), and do not construct a multi-element coupling model, resulting in insufficient prediction accuracy of the spatio-temporal distribution, evolution rate, and coupling strength of high-temperature, heat wave, and drought composite disasters, especially in complex geographical units (such as mountainous areas, urban-rural transition zones) and ecologically fragile areas (such as grasslands, desert edges). SUMMARY

[0008] The present application provides a prediction system and method for high-temperature, heat wave, and drought composite disasters, which can consider multi-dimensional data, construct a multi-element coupling model, and achieve prediction of the spatio-temporal distribution, evolution rate, and coupling strength of high-temperature, heat wave, and drought composite disasters in complex geographical units and ecologically fragile areas.

[0009] To solve the above technical problems, the present application provides the following technical solutions:

[0010] A prediction system for high-temperature, heat wave, and drought composite disasters, comprising:

[0011] A multi-source data acquisition module for acquiring meteorological data, terrain data, vegetation and soil data, and human activity data, and constructing a multi-dimensional input feature set;

[0012] A data processing module for spatio-temporal alignment, normalization processing, and outlier removal of the acquired meteorological data, terrain data, vegetation and soil data, and human activity data, and then dimension reduction and key feature extraction by principal component analysis;

[0013] A deep learning model training module using a CNN-RNN fusion architecture, comprising:

[0014] A CNN sub-module for extracting spatial features of terrain data, vegetation and soil data, and human activity data through a double-layer convolutional layer;

[0015] An RNN sub-module for capturing time series features of meteorological data through a three-layer LSTM network;

[0016] An attention mechanism layer for dynamically weighting and fusing spatial features and time series features to highlight factors that significantly affect disaster evolution;

[0017] a prediction output module configured to generate a compound disaster prediction result by the trained deep learning model training module, the compound disaster prediction result comprising a spatial dimension result, a time dimension result, and a decision support result; the spatial dimension result comprising a high-temperature drought risk thermal map, a topographic sensitive area marker, and an ecological fragile area marker; the time dimension result comprising a disaster intensity change curve for a preset number of future days; and the decision support result comprising a scenario simulation result for human activity intervention.

[0018] The basic scheme principle and beneficial effects are as follows: the present application collects four categories of data, i.e., meteorological data, topographical data, vegetation and soil data, and human activity data, to construct a multi-dimensional feature set comprising spatial distribution (such as topographical slope and vegetation coverage) and time sequence (such as temperature change and precipitation fluctuation).

[0019] The data processing module unifies heterogeneous data (such as human activity) to the same spatio-temporal coordinate system (such as a 1 km grid x 1 hour) through a spatio-temporal alignment technique, thereby solving the problem of mismatch between multi-source data in terms of spatial resolution and time frequency.

[0020] The deep learning model adopts a CNN-RNN fusion architecture: the CNN submodule extracts local correlation features of spatial elements (such as the high-temperature aggregation effect of dry and hot valleys) through convolution operation, the RNN submodule captures time-dependent features of meteorological data (such as the continuous evolution law of high-temperature drought) through an LSTM network, and finally the spatio-temporal features are dynamically weighted and fused through an attention mechanism layer to realize adaptive identification of key influence factors in the disaster evolution process.

[0021] Based on the fused spatio-temporal features, the spatial dimension result, the time dimension result, and the decision support result of the output region are predicted. The spatial dimension result is directly presented through the risk thermal map and the sensitive area marker, intuitively showing the spatial distribution law of the disaster in complex topographical regions (such as mountainous areas and valleys) and ecological fragile areas (such as the desert-grassland transition zone), helping decision-makers quickly locate high-risk areas and optimize the allocation of emergency resources (such as preferentially deploying artificial rainmaking operations in ecological fragile areas). The time dimension result is quantified through the disaster intensity curve for a preset number of future days, quantifying the duration and intensity change of high-temperature drought, and providing a time window reference for agricultural production scheduling (such as developing irrigation plans in advance) and power supply guarantee (such as power grid load warning during high temperature). The decision support result is quantified through human activity intervention scenario simulation, evaluating the influence of irrigation scheduling, reservoir storage, and other measures on disaster evolution, providing a scientific basis for disaster prevention and mitigation, quantifying the disaster mitigation effect of different measures, and promoting the closed-loop management of “prediction-intervention-evaluation”. Compared with the traditional prediction system, the decision response efficiency is improved by more than 30%.

[0022] The present application incorporates multi-dimensional data such as terrain, vegetation, human activities, and constructs a prediction model containing the interaction of natural environment and human activities, which can capture the coupling mechanism of urban heat island effect (human activities) and drought intensification (natural factors), solve the prediction deviation problem caused by single data, and is especially suitable for disaster prediction in ecologically fragile areas and complex geographical units (such as dry-hot valleys on the edge of the Qinghai-Tibet Plateau and oasis-desert transition zones in the northwest arid region).

[0023] The CNN-RNN fusion architecture realizes the deep coupling of spatial features (such as the influence of terrain on heat accumulation) and temporal features (such as the periodic fluctuations of high temperature and drought), and can accurately depict the disaster evolution process under the interaction of "terrain-climate-vegetation" (for example: the high temperature and drought risk in mountainous areas with slope > 25° is 30%-50% higher than that in plain areas due to low vegetation coverage). The attention mechanism layer highlights key factors (such as the weight of vegetation coverage in the early stage of drought is increased by 40%) through dynamic weighting, avoiding the overfitting of traditional models to secondary factors, and improving the prediction accuracy of the model in data sparse areas (such as remote mountainous areas) by more than 20%.

[0024] The multi-source data acquisition and spatio-temporal fusion architecture has good scalability, can seamlessly integrate satellite remote sensing, ground sensors, social and economic data, and other new data sources (such as soil moisture data, population distribution data), and supports rapid iteration of prediction models for different regions (from county to national scale) and different disaster types (high temperature, drought, and their composite disasters), providing a general technical framework for extreme disaster warning under the background of climate change.

[0025] In summary, the present application solves the prediction problem of high temperature, heat wave and drought composite disasters in complex geographical environment through multi-dimensional data fusion, spatio-temporal coupling modeling and multi-target prediction output.

[0026] Further, the meteorological data includes hourly temperature T(t), relative humidity RH(t), and precipitation intensity P(t), and the time series feature X meteo is constructed by sliding window method = [T(t-n), RH(t-n), P(t-n), …, T(t), RH(t), P(t)], where n is the window length;

[0027] The terrain data includes elevation H(s), slope and water system buffer distance D(s), forming the spatial feature vector X geo = [H(s), S(s), D(s)], where s is the spatial grid coordinate;

[0028] The vegetation and soil data includes normalized vegetation index Vegetation type one-hot encoding V type , soil texture classification S textureand soil organic matter content OM = λ Reflectance 400-2400nm + μ; wherein the normalized difference vegetation index is based on satellite remote sensing red DN RED and near-infrared band DN NIR obtained, soil organic matter content is obtained by a partial least squares regression model;

[0029] Human activity data includes agricultural irrigation intensity wherein ET a is actual evapotranspiration, ET0 is potential evapotranspiration, ET ref is reference evapotranspiration, urban heat island index UHI = T urban - T rural , wherein T urban is urban center temperature, T rural is average suburban temperature, reservoir storage rate wherein V is real-time reservoir capacity, V max is maximum reservoir capacity.

[0030] Further, the construction of the multi-dimensional input feature set includes:

[0031] Hourly temperature T(t), relative humidity RH(t), and precipitation intensity P(t) are subjected to sliding window feature extraction, and a time series matrix is constructed:

[0032]

[0033] wherein n is the window length, and a derived indicator, consecutive high temperature days, is calculated: wherein I is an indicator function;

[0034] Slope is calculated based on a digital elevation model aspect and distance from waterways D(s) is generated by GIS buffer analysis, forming a spatial feature vector: X geo (s) = [H(s), S(s), A(s), D(s)];

[0035] Vegetation index is calculated by satellite remote sensing and combined with land use classification to generate vegetation type one-hot encoding V type ∈ R C wherein C is the number of vegetation categories; soil characteristics are modeled by spectral analysis: organic matter content OM = PLSR(Reflectance 400-2400nm ), which is obtained based on a partial least squares regression model, soil texture is represented using a triangular coordinate system (S texture = (clay, silt, sand);

[0036] Agricultural irrigation intensity is calculated based on the Penman-Monteith formula ref ; Urban Heat Island Index UHI = T urban - T rural , calibrated by the nighttime light data and land surface temperature regression model T s = aL + b; Reservoir storage rate Combined with the water surface area S monitored by SAR images and the reservoir capacity curve V = f(S), the reservoir storage rate is calculated.

[0037] The single-dimensional features are spatiotemporally registered, and the meteorological time series X meteo and the spatial grid data are aligned by time length t and grid point coordinate s to form the spatiotemporal feature variable:

[0038]

[0039] Then, principal component analysis (PCA) is performed on the high-dimensional features to reduce the dimension, and the covariance matrix is calculated. The first k principal components are extracted by eigenvalue decomposition, and the cumulative variance contribution rate is calculated. where N is the number of samples, and m is the original number of features.

[0040] Each dimension of the feature is standardized by Z-score: where μ j and σ j are the mean and standard deviation of the jth dimension, respectively, to ensure that the features input into the deep learning model have zero mean and unit variance.

[0041] Further, the data processing module interpolates the time series of meteorological data and the time length t of human activity data to generate a unified time series with equal time intervals, ensuring that t i = t0 + i·Δt, where i = 0, 1, …, N-1, and Δt = 1h. i

[0042] The terrain data, vegetation and soil data, and human activity data are rasterized according to the unified spatial grid, and the no-data areas are filled by Gaussian kernel interpolation to form the spatial grid matrix X geo / veg / human (s) ∈ R M×D , where M is the number of grid points, and D is the number of single-dimensional features.

[0043] The spatiotemporal feature matrix X is generated, where t i is the time index, and s j is the spatial grid index.

[0044] ​Further, the data processing module adopts an improved quartile range method to detect outliers of each dimension feature, and calculates the first quartile of the kth dimension feature The third quartile And the quartile range

[0045] Then define the outlier threshold, the lower limit is The upper limit is

[0046] For sample j satisfying Or Use linear interpolation method to repair adjacent time points: (When is an outlier);

[0047] Then Z-score standardization is performed on the feature matrix,

[0048] Where is the global mean of the kth dimension feature, is the standard deviation, which ensures that the feature obeys N(0,1) distribution;

[0049] Finally, the covariance matrix X norm is calculated m The eigenvectors U=[u1,u2,…,u m ] and eigenvalues λ1≥λ2≥…≥λ k are solved by singular value decomposition; the first k principal components are selected, so that the cumulative variance contribution rate The reduced dimension feature matrix

[0050] Where U k =[u1,u2,…,u spatial ] is the first k principal component vector.

[0051] Further, the double-layer convolution layer structure of the CNN submodule is:

[0052] The first convolution layer adopts a 3×3 learnable convolution kernel The spatial feature matrix X H×W×D ∈R (1) is input, where H and W are spatial grid sizes, and D is the feature dimension of a single grid point, and the convolution operation is performed: Z (1) =σ(W spatial *X (1) +b (1) );

[0053] Where bis the bias vector, σ is the ReLU activation function, and Y is the output feature map where C1=64 is the number of channels of the first layer.

[0054] The second convolutional layer uses a 1x1 convolutional kernel Cross-channel feature fusion: Z (2) = BatchNorm(W (2) *Z (1) +b (2) );

[0055] The output spatial feature tensor where C2=32 is the number of channels of the second layer, and high-order spatial features of terrain relief and vegetation distribution are extracted through spatial dimension compression.

[0056] Further, the neuron state update equation of the three-layer LSTM network is:

[0057] Forget gate: f t = σ(W f ·[h t-1 ,x t ]+b f )

[0058] Input gate: i t = σ(W i ·[h t-1 ,x t ]+b i ),

[0059] Cell state:

[0060] Output gate: o t = σ(W o ·[h t-1 ,x t ]+b o ), h t = o t ⊙tanh(c t )

[0061] where is the meteorological input feature at time t, h t ∈R L is the hidden state, where L=128 is the number of single-layer LSTM neurons, W f , W i , W c , and W o are weight matrices, and b f , b i , b c , and b ois the bias vector, ⊙ represents element-wise multiplication;

[0062] The three-layer LSTM network is stacked Where l = 1, 2, 3 is the number of layers, capturing the long-range dependency features of the number of days with high temperature and precipitation anomalies in the time series, and outputting the final hidden state sequence H rnn ∈R T×L , where T is the time step.

[0063] Furthermore, the attention mechanism layer adopts a spatiotemporal cross attention model to transform the spatial features output by CNN into Flattened into key-value pairs The temporal feature H output by RNN rnn ∈R T×L Mapped to query vector through the fully connected layer

[0064] Then the attention weight is calculated.

[0065] where α t,s represents the attention weight of the spatial grid point s at time t;

[0066] Finally, feature fusion

[0067] Output fused spatiotemporal features Dynamic weighting is used to highlight the key influencing factors of vegetation coverage decline and urban heat island effect during the drought development period.

[0068] Furthermore, the specific generation method of the prediction output module includes:

[0069] The attention features are transferred through the transposed convolution layer Upsample to the original spatial resolution H×W, and output the high temperature and drought risk probability matrix through the softmax layer:

[0070] Among them C class =3 is the risk level, combining the terrain slope threshold and vegetation index threshold to mark the terrain sensitive area and ecologically fragile area;

[0071] Through the fully connected layer Map attention features to disaster intensity indicators to generate future T pred =15-day intensity change curve: Y time (t) = f fc (X attn (t)),t∈[t0,t0+T pred ];

[0072] Human activity intervention variables include irrigation intensity I′irr and the reservoir storage rate R', after one-hot encoding, are spliced with the original features to input the model, and the conditional generative adversarial network is used to simulate the disaster evolution path under the intervention scenario:

[0073] The generator G adopts a multi-layer perception architecture, and outputs the simulated disaster intensity deviation value Quantify the disaster reduction effect of the intervention measure.

[0074] Further, it further comprises a real-time data access module for accessing real-time data corresponding to the multi-source data acquisition module, including real-time meteorological data, real-time vegetation monitoring data, and real-time human activity data, to form a real-time feature vector where T ′ is the real-time data time step, and D is consistent with the multi-source data feature dimension;

[0075] Then dynamic comparison and error calculation are performed, and the real-time feature vector is compared with the historical prediction data to calculate the absolute error of each dimension feature and the error overrun ratio is calculated where δ d is the error threshold of the dth dimension feature;

[0076] Then the weight is dynamically adjusted, when r>20%, the weight update of the attention mechanism layer is triggered, for the error overrun feature dimension, the weight coefficient α of the attention mechanism layer is improved t,s ; for the feature dimension with smaller error, the weight coefficient is maintained or reduced;

[0077] Finally, the prediction result is updated, the adjusted weight is input into the deep learning model training module to regenerate the prediction result of the future preset number of days, and the updated high temperature and drought risk heat map, intensity change curve and intervention scenario simulation result are pushed in real time through the prediction output module. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 is a workflow diagram of a multi-source data acquisition module in a high temperature and drought compound disaster prediction system;

[0079] Figure 2 is a workflow diagram of a data processing module in a high temperature and drought compound disaster prediction system;

[0080] Figure 3 is a workflow diagram of a deep learning model training module in a high temperature and drought compound disaster prediction system;

[0081] Figure 4 A workflow diagram of a prediction output module in a high-temperature heat wave drought composite disaster prediction system. DETAILED DESCRIPTION

[0082] Further details are described below through specific embodiments:

[0083] A high-temperature heat wave drought composite disaster prediction system (as shown in Figures 1-4 includes:

[0084] A multi-source data acquisition module for acquiring meteorological data, terrain data, vegetation and soil data, and human activity data to construct a multi-dimensional input feature set;

[0085] A data processing module for spatiotemporal alignment, normalization processing, and outlier removal of the acquired meteorological data, terrain data, vegetation and soil data, and human activity data, and then dimension reduction and key feature extraction through principal component analysis;

[0086] A deep learning model training module using a CNN-RNN fusion architecture, including:

[0087] A CNN submodule for extracting spatial features of terrain data, vegetation and soil data, and human activity data through a double-layer convolutional layer;

[0088] An RNN submodule for capturing time series features of meteorological data through a three-layer LSTM network;

[0089] An attention mechanism layer for dynamically weighting and fusing spatial features and time series features to highlight factors that significantly affect disaster evolution;

[0090] A prediction output module for generating composite disaster prediction results through the trained deep learning model training module, including spatial dimension results, time dimension results, and decision support results; the spatial dimension results include high-temperature drought risk thermal maps, terrain sensitive area markers, and ecological fragile area markers; the time dimension results include disaster intensity change curves for a preset number of future days; and the decision support results include scenario simulation results for human activity intervention.

[0091] In specific use: combined with a high-temperature heat wave drought composite disaster prediction scenario in a certain basin in northwest China (such as the Tarim Basin), the Tarim Basin and surrounding mountainous areas contain complex geographical units such as deserts, oases, and mountains, surrounded by surrounding mountains, and constitute complex geographical units as a whole, which is a high-temperature drought and drought disaster-prone area. The time is from June to August 2023 (high-temperature drought high-occurrence period).

[0092] Meteorological data is obtained from China Meteorological Administration, including hourly temperature T(t), relative humidity RH(t), and precipitation intensity P(t), and time series features X are constructed by sliding window method meteo = [T(t-n), RH(t-n), P(t-n), …, T(t), RH(t), P(t)], where n is the window length; the time window length n = 72 hours.

[0093] Time series features are constructed and derived indicators are calculated: consecutive high temperature days HDD For example, on July 10, 2023, 5 consecutive high temperature days are monitored, HDD = 5.

[0094] Vegetation and soil data include normalized difference vegetation index NDVI Vegetation type one-hot encoding V type , soil texture classification S texture , and soil organic matter content OM = λ·spectral reflectance 400-2400nm + μ; where the normalized difference vegetation index is obtained based on satellite remote sensing red DN RED and near-infrared band DN NIR , and the soil organic matter content is obtained by a partial least squares regression model.

[0095] Human activity data includes agricultural irrigation intensity where ET a is the actual evapotranspiration, ET0 is the potential evapotranspiration, and ET ref is the reference evapotranspiration.

[0096] Urban heat island index UHI = T urban - T rural = 3℃, the temperature in the center of Korla city is 3℃ higher than that in the suburbs, where T urban is the urban center temperature, and T rural is the average suburban temperature. Reservoir storage rate where V is the real-time reservoir capacity, V max is the maximum reservoir capacity, and the real-time reservoir capacity of a certain reservoir is 60% of the maximum reservoir capacity.

[0097] The construction of the multi-dimensional input feature set includes:

[0098] Sliding window feature extraction is performed on hourly temperature T(t), relative humidity RH(t), and precipitation intensity P(t) to construct a time series matrix:

[0099]

[0100] where n is the window length, and the derived indicator is calculated: consecutive high temperature days HDD where I is the indicator function.

[0101] Slope is calculated based on digital elevation model Aspect And distance to water system D(s) is generated by GIS buffer analysis, forming spatial feature vector: X geo (s) = [H(s), S(s), A(s), D(s)];

[0102] Vegetation index is calculated by satellite remote sensing And vegetation type one-hot encoding V is generated combined with land use classification type ∈ R C Where C is the number of vegetation categories; Soil characteristics are modeled by spectral analysis: organic matter content OM = PLSR(Reflectance 400-2400nm ), which is based on partial least squares regression model inversion, and soil texture is represented in a triangular coordinate system (S texture = (clay, silt, sand).

[0103] Download NASA SRTM digital elevation model (resolution 30 meters) to obtain terrain data including elevation H(s), slope and water system buffer distance D(s), forming spatial feature vector X geo = [H(s), S(s), D(s)], where s is the spatial grid coordinate.

[0104] Specifically, slope Then get aspect A(s); Distance to Tarim River mainstream D(s) is generated by GIS buffer analysis, forming spatial feature vector X geo (s) = [H(s), S(s), A(s), D(s)]. For example, the area with slope S(s) > 25° and distance to river system D(s) > 50 km at the edge of the basin is marked as a high-risk arid area.

[0105] Specifically, normalized vegetation index is calculated using Sentinel-2 satellite data to identify oasis areas (NDVI > 0.4) and desert areas (NDVI < 0.1); Soil texture is inverted by spectral reflectance (400-2400nm) combined with partial least squares regression model (PLSR), such as soil organic matter content OM = 0.5% (low fertility) at the edge of the Taklimakan Desert.

[0106] Irrigation intensity of agriculture is calculated based on the Penman-Monteith formula to calculate the reference evapotranspiration ET ref ; Urban heat island index UHI = T urban - T rural, and surface temperature T s = aL + b calibration; reservoir storage rate The water surface area S is monitored by SAR images, and the reservoir capacity curve V = f(S) is calculated. The reference evapotranspiration ET ref , such as the oasis agricultural area I irr = 0.3 (indicating that the actual irrigation amount is 30% of the potential water requirement).

[0107] The single-dimensional features are spatiotemporally registered, and the meteorological time series X meteo are aligned with the spatial grid data by time length t and grid coordinates s to form the spatiotemporal feature variables:

[0108]

[0109] Then, the high-dimensional features are reduced by principal component analysis (PCA), and the covariance matrix The first k principal components are extracted by eigenvalue decomposition, and the cumulative variance contribution rate The reduced feature matrix is obtained where N is the number of samples, and m is the original number of features;

[0110] Each dimension of the feature is standardized by Z-score: where μ j and σ j are the mean and standard deviation of the jth feature, respectively, to ensure that the features input into the deep learning model have zero mean and unit variance.

[0111] Specifically, the pandas library of Python is used to perform cubic spline interpolation on the meteorological data and human activity data to generate a unified time series at 1-hour intervals, such as interpolating the daily monitoring data of reservoir storage rate to hourly data.

[0112] The terrain, vegetation, and human activity data are rasterized into 1km×1km grids by ArcGIS, and the missing soil data in the desert uninhabited area is filled by Gaussian kernel interpolation to form the spatial grid matrix (assuming that there are 530,000 grid points in the region, and the number of single-dimensional features D = 8).

[0113] The spatiotemporal feature matrix is generated (meteorological 3D + terrain 4D + vegetation and soil 5D + human activity 6D).

[0114] The data processing module aligns the time series of meteorological data with the time length t iInterpolation synchronization is performed and cubic spline interpolation method is used to generate a unified time series with equal time intervals to ensure that (t i =t0+i·Δt, where i=0, 1,…, N-1, Δt=1h;

[0115] The terrain data, vegetation and soil data, and human activity data are rasterized according to a unified spatial grid, and the data-free areas are filled by Gaussian kernel interpolation to form a spatial grid matrix X geo / veg / human (s)∈R M×D , where M is the number of grid points and D is the number of single-dimensional features;

[0116] Generate spatiotemporal feature matrix where t i is the time index, s j is the spatial grid index.

[0117] The data processing module uses the improved interquartile range method to detect outliers in each dimension and calculate the first quartile of the k-th dimension feature. third quartile and interquartile range

[0118] Then define the outlier threshold, the lower limit is Upper limit is

[0119] To satisfy or Sample j of is repaired using linear interpolation of adjacent time points: 9 When 0 when it is an outlier;

[0120] Then the feature matrix is ​​Z-score standardized.

[0121] in is the global mean of the k-th dimension feature, is the standard deviation, ensuring that the feature follows the N(0,1) distribution;

[0122] Finally, calculate the covariance matrix X norm is the normalized characteristic matrix; the characteristic vector U=[u1,u2,…,u m ] and eigenvalues ​​λ1≥λ2≥…≥λ m ; Select the first k principal components to make the cumulative variance contribution rate Get the feature matrix after dimensionality reduction,

[0123] Among them U k= [u1, u2, …, u k ] are the first k principal component vectors.

[0124] Specifically, the improved IQR method is applied to the temperature data, and the temperature value T = 48 (higher than the threshold U k = 45) at 14:00 on July 5, 2023 is detected and repaired by the average value of 47 degrees Celsius at adjacent time (13:00 and 15:00). Z-score standardization is used, for example, the altitude H(s) is standardized to (the average altitude of the basin is about 1000 meters, and the standard deviation is 500 meters).

[0125] The PCA class of Scikit-learn is used to reduce the dimensionality of the 18-dimensional original features, and the first 8 principal components are extracted, with a cumulative variance contribution rate of 96%. The feature matrix after dimensionality reduction is

[0126] The double-layer convolutional layer structure of the CNN submodule is as follows:

[0127] The first convolutional layer uses a 3 × 3 learnable convolutional kernel The input spatial feature matrix X spatial ∈ R H×W×D , where H, W are the spatial grid sizes, and D is the feature dimension of a single grid point, and the convolution operation is performed: Z (1) = σ(W (1) *X spatial +b (1) ).

[0128] where b (1) is the bias vector, and σ is the ReLU activation function. The output feature map where C1 = 64 is the number of channels in the first layer.

[0129] The Conv2D layer of TensorFlow is used, with a 3 × 3 convolutional kernel, 64 channels, and the spatial feature matrix (reshaped to a single channel) is convolved, and the output feature map captures the spatial correlation between oasis distribution (low altitude, high NDVI) and high temperature areas.

[0130] The second convolutional layer uses a 1 × 1 convolutional kernel to perform cross-channel feature fusion: Z (2) = BatchNorm(W (2) *Z (1) +b (2) );

[0131] The output spatial feature tensor Wherein, C2=32 is the second layer channel, and the high-order spatial features of the terrain relief and vegetation distribution are extracted by spatial dimension compression.

[0132] Specifically, cross-channel fusion is performed by a 1x1 convolution kernel, and the output is The "terrain-vegetation" composite features (such as the high-temperature drought risk pattern of the mountain leeward slope + low NDVI area) are extracted.

[0133] The neuron state update equation of the three-layer LSTM network is:

[0134] Forget gate: f t = σ(W f ·[h t-1 ,x t ]+b f )

[0135] Input gate: i t = σ(W i ·[h t-1 ,x t ]+b i ),

[0136] Cell state:

[0137] Output gate: o t = σ(W o ·[h t-1 ,x t ]+b o ),h t = o t ⊙tanh(c t )

[0138] wherein is the meteorological input feature at time t, h t ∈R L is the hidden state, wherein L=128 is the number of single-layer LSTM neurons, W f ,W i ,W c ,W o are weight matrices, b f ,b i ,b c ,b o are bias vectors, and ⊙ represents element-level multiplication.

[0139] The three-layer LSTM network is stacked wherein l=1, 2, 3 is the layer number, the high-temperature continuous day number and the long-range dependence characteristics of precipitation anomalies in the time series are captured, and the final hidden state sequence H rnn ∈R T×Lwhere T is the time step.

[0140] Specifically, the three-layer LSTM network is implemented using PyTorch, with 128 neurons per layer. With hourly meteorological data in June 2023 as input, the long-range dependence of consecutive high temperatures (e.g., 6-day high temperature from June 15-20) and precipitation anomalies (no effective precipitation in June) is captured, and the output is a sequence of hidden states

[0141] The attention mechanism layer adopts a spatio-temporal cross-attention model to combine the spatial features flattened into key-value pairs and the temporal features H output by the RNN rnn ∈R T×L mapped to query vectors through a fully connected layer

[0142] Then the attention weight calculation is performed

[0143] where α t,s represents the attention weight of spatial grid s at time t;

[0144] Finally, feature fusion is performed

[0145] and the fused spatio-temporal features are output Through dynamic weighting, the key influence factors such as vegetation coverage decline and urban heat island effect during the drought development period are highlighted.

[0146] Specifically, the spatial features output by the CNN are flattened into and the RNN temporal features are mapped to The attention weight α t,s is calculated, for example, during the drought development period (early July), the model's weight for desert areas with NDVI <0.2 is increased to 0.4 (ordinary area weight 0.1), highlighting the aggravating effect of vegetation degradation on drought.

[0147] Dynamic weighting of key influence factors such as vegetation coverage decline and urban heat island effect during the drought development period is mainly achieved through the spatio-temporal cross-attention model of the attention mechanism layer, with the following specific technical path:

[0148] The spatial features of terrain, vegetation, and soil, and human activity data extracted by the CNN submodule are (for example, H = 530000 grid points, W = 1, C2 = 32 channels), which are flattened into a key-value pair matrix where each row corresponds to a feature vector of a spatial grid point.

[0149] The meteorological time series features output by the RNN submodule are (As T = 72 hours, L = 128 dimensions), through the fully connected layer is mapped to the query vector The dimensions of the time features and the spatial features are made consistent (both C2 = 32 dimensions).

[0150] For each time step t and spatial grid point s, the time query vector Q t is calculated. s The similarity score with the spatial key vector K t,s is generated through the scaled dot-product attention formula to generate attention weights α

[0151]

[0152] α t,s represents the degree of attention of the model to the spatial grid point s at time t, with a value range of [0, 1], and the sum of the weights of all grid points is 1.

[0153] When a grid point s has a decrease in vegetation coverage (NDVI decreases) or an urban heat island effect (UHI increases), its corresponding spatial feature K s will have a stronger correlation with the time feature Q t (such as an increase in drought index), resulting in a significant increase in α t,s (from 0.05 to 0.3, for example).

[0154] According to the attention weights, the spatial value vector V s is weighted and summed to obtain the fused spatio-temporal feature X attn (t):

[0155]

[0156] During the drought development period (such as the time step t corresponding to an increase in consecutive dry days), the model automatically increases the weights of areas with low vegetation coverage (NDVI < 0.2) and high urban heat island intensity (UHI > 2°C), causing the features of these grid points to have a higher contribution in X attn (t) (for example, the total contribution increases from 15% to 40%), thereby highlighting their driving role in drought evolution.

[0157] Conversely, the weights of irrelevant areas (such as high-vegetation-covered mountainous areas) are automatically reduced to suppress secondary information interference.

[0158] Taking an actual scenario as an example, in July 2023, a persistent high-temperature drought occurred in the eastern Tarim Basin, with vegetation coverage decreasing from NDVI = 0.3 to NDVI = 0.15, and the urban heat island effect intensifying (UHI = 4°C).

[0159] In the vegetation degradation area (such as the edge of the Taklimakan Desert), αt,s From 0.1 to 0.35, the corresponding feature (low NDVI, high degree of soil desertification) in X attn (t) significantly increases.

[0160] In urban areas such as Korla City, a t,s From 0.08 to 0.25, the coupling effect of urban heat island effect and high temperature drought is highlighted. Through the weighted feature X attn (t), the model predicts the drought intensity to increase by 20% compared to the unweighted one, and identifies the accelerated evolution trend of the disaster 3 days in advance.

[0161] In this embodiment, through dynamic weighting, the contribution of vegetation and urban factors is explicitly amplified, and the prediction error is reduced to 12%, and the spatial expansion path of the disaster core area can be accurately captured.

[0162] By visualizing the attention weight a t,s , the distribution of key driving factors during the drought development period (such as the weight heat map of the area with decreasing vegetation coverage) can be intuitively displayed, providing a basis for disaster cause analysis. Dynamic suppression of irrelevant noise (such as stable mountain low temperature features) makes the model perform more stably in areas with sparse data or strong heterogeneity (such as desert-oasis transition zones). The model's spatiotemporal correlation modeling capability is improved for the coupling mechanism of drought and vegetation degradation, urban heat island, and the prediction accuracy of complex geographical units is improved by 15%-20% compared to traditional methods.

[0163] Through the above mechanisms, the system realizes adaptive identification and dynamic weighting of key influencing factors during the drought development period, significantly improving the accuracy and physical interpretability of high temperature heat wave drought composite disaster prediction.

[0164] The specific generation method of the prediction output module includes:

[0165] The attention feature X is upsampled to the original spatial resolution HxW through the transpose convolution layer, and the high temperature drought risk probability matrix is output through the softmax layer:

[0166] Where C class = 3 is the risk level, combined with the terrain slope threshold and vegetation index threshold to mark the terrain sensitive area and ecological fragile area;

[0167] The attention feature is mapped to the disaster intensity index through the fully connected layer , and the intensity change curve of the future T pred = 15 days is generated: Y time (t) = f fc (X attn (t)), t ∈ [t0, t0+T pred ].

[0168] The human activity intervention variable includes irrigation intensity I' irr and reservoir storage rate R', after one-hot encoding, are spliced with the original features to input the model, and the conditional generative adversarial network is used to simulate the disaster evolution path under the intervention scenario:

[0169] where the generator G adopts a multi-layer perception architecture, and outputs the simulated disaster intensity deviation value Quantify the disaster mitigation effect of the intervention measure.

[0170] Specifically, in the spatial dimension result layer, the Softmax layer outputs the risk level (0 = no risk, 1 = low risk, 2 = medium risk, and 3 = high risk) of the high temperature and drought risk heat map by upsampling to 1 km resolution through transposed convolution. For example, the risk level in the center of the Taklimakan Desert is 3, and the risk level in the oasis edge is 2. The mountainous area with a slope S(s) > 20° and a distance from water system D(s) > 30 km is classified as a topographically sensitive area (about 15% of the regional area). The area with NDVI < 0.1 and soil organic matter OM < 1% is marked as an ecologically fragile area (such as the desert expansion zone).

[0171] In the time dimension result layer, the comprehensive drought index (SPI) curve for the next 15 days (or other period) is generated, predicting that the SPI value will rise from -1.5 (moderate drought) to -2.0 (severe drought) from July 15 to 30, and the number of high temperature days will increase to 8 days.

[0172] In the decision support result layer, assuming that the agricultural irrigation intensity I' irr is increased from 0.3 to 0.5 (increasing irrigation by 20%), the conditional generative adversarial network (cGAN) predicts that the SPI value will drop to -1.8, the number of high temperature days will decrease by 2 days, and the visualization shows that the risk level around the oasis decreases by 1 level.

[0173] The reservoir storage rate R ′ is increased from 0.6 to 0.8, and the simulation results show that the temperature in the 50 km range around the reservoir decreases by 1-2°C, and the drought risk decreases.

[0174] Compared with traditional single meteorological models, the prediction error of high temperature in the Tarim Basin is reduced from 2.5°C to 1.8°C, and the prediction error of drought range is reduced. Through risk heat maps and scenario simulation, decision-makers can start drought emergency plans in advance, improve irrigation scheduling efficiency, and reduce emergency material deployment costs.

[0175] In yet another embodiment, a real-time data access module is further included for accessing real-time data corresponding to the multi-source data acquisition module, including real-time meteorological data (temperature, humidity, precipitation updated every 15 minutes), real-time vegetation monitoring data (satellite remote sensing NDVI daily update value), and real-time human activity data (agricultural irrigation volume recorded hourly, real-time reservoir water level monitoring value), forming a real-time feature vector (T ′ is the time step of real-time data, and D is consistent with the multi-source data feature dimension

[0176] Then dynamic comparison and error calculation are performed, and the real-time feature vector is compared with the historical prediction data of the prediction frequency (e.g., every 6 hours) to calculate the absolute error of each dimension feature and the error exceeding ratio is counted (δ d is the error threshold of the d-th feature, such as a temperature error threshold of 2°C

[0177] Then, weight dynamic adjustment is performed, and when r>20%, the weight update of the attention mechanism layer is triggered

[0178] For the feature dimension with error exceeding (such as real-time NDVI deviating from the predicted value by more than 0.1), the weight coefficient a of the attention mechanism layer is increased t,s (e.g., from 0.2 to 0.35); for the feature dimension with smaller error (such as terrain data deviation <5%), the weight coefficient is maintained or reduced (e.g., from 0.1 to 0.05)

[0179] Finally, the prediction result is updated, the adjusted weight is input into the deep learning model training module, the prediction result of the future preset number of days is regenerated, and the updated high temperature and drought risk thermal map, intensity change curve, and intervention scenario simulation result are pushed in real time through the prediction output module. Through real-time data closed-loop feedback, the prediction error of the model in the scenario of meteorological mutation (such as sudden precipitation) and human activity intensity change (such as a sharp increase in irrigation volume) is reduced, and the adaptability of the system to dynamic environment is significantly improved.

[0180] In another embodiment, the deep learning model training module is further configured to compare the prediction result generated by the prediction output module (T is the prediction time step, and HxW is the spatial resolution) with the actual observation data in the corresponding time window (such as actual high temperature days and drought index) at each spatial point

[0181] to calculate the spatial distribution error matrix wherein and the time series error vector wherein

[0182] In the basic loss function L base (like mean square error MSE), a spatial attention penalty term L spatial = λ·∑ h, w E h,w · α h,w , where α h,w is the average weight of the attention mechanism layer in the spatial position (h, w) as claimed in claim 6, and λ is the penalty coefficient (initial value set to 0.3);

[0183] A time series penalty term L temporal = γ·∑ t e t · β t is introduced, where β t is the time importance weight (such as β t = 1.5 during the disaster period and β t = 0.8 during the stable period), and γ is the penalty coefficient (initial value set to 0.2);

[0184] The final loss function L = L base + L spatial + L temporal ;

[0185] The final adaptive model parameter update is realized, when the average value of e t for 3 consecutive time windows exceeds the preset threshold (such as 0.25), the model parameter fine-tuning is triggered: for the top 10% high error area (such as the oasis edge in the eastern Tarim Basin) in the spatial error matrix E, the sampling density in the multi-source data features is increased from the original 1km×1km to 500m×500m;

[0186] The spatio-temporal cross weight of the attention mechanism layer is adjusted, so that the feature attention of the model in the high error area is improved; every 50 times of prediction-actual comparison cycle, the accumulated error data is used to retrain the full-amount parameters of the model, and the weight matrix W and the bias vector b of the prediction model are updated.

[0187] The prediction result is directly from the prediction output module; the actual observation data Y obs is obtained through the multi-source data acquisition module and the real-time data access module; the loss function adjustment mechanism acts on the spatio-temporal cross attention model training process; the parameter update logic forms a closed loop optimization with the multi-module collaborative training framework.

[0188] By introducing a dual temporal and spatial error penalty mechanism, the model's prediction bias in highly heterogeneous areas is reduced, and the warning accuracy for extreme high temperature events (temperature > 40°C) is improved, significantly enhancing the system's adaptability to complex geographical units.

[0189] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this implementation case. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field are aware of all common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A prediction system for high temperature, heat wave and drought combined disasters, including, characterized by: Multi-source data acquisition module, used to collect meteorological data, terrain data, vegetation and soil data, and human activity data, and construct a multi-dimensional input feature set; The data processing module is used to perform spatiotemporal alignment, normalization, and outlier removal on the collected meteorological data, topographic data, vegetation and soil data, and human activity data, and then extract key features through dimensionality reduction using principal component analysis. The deep learning model training module adopts the CNN-RNN fusion architecture, including: The CNN submodule is used to extract the spatial features of terrain data, vegetation and soil data, and human activity data through a double-layer convolutional layer; The RNN submodule is used to capture the time series characteristics of meteorological data through a three-layer LSTM network; The attention mechanism layer is used to dynamically weight and fuse spatial features and time series features to highlight factors that have a significant impact on disaster evolution; The prediction output module is used to generate composite disaster prediction results through the trained deep learning model training module. The composite disaster prediction results include spatial dimension results, temporal dimension results and decision support results; the spatial dimension results include high temperature and drought risk heat maps, terrain sensitive area markers and ecological fragile area markers; the temporal dimension results include disaster intensity change curves for a preset number of days in the future; and the decision support results include scenario simulation results for human activity intervention.

2. The prediction system for high temperature, heat wave and drought combined disasters according to claim 1, characterized in that: Meteorological data include hourly temperature T(t), relative humidity RH(t), and precipitation intensity P(t), and the time series feature X is constructed by the sliding window method. meteo =[T(tn),RH(tn),P(tn),…,T(t),RH(t),P(t)], where n is the window length; Terrain data includes altitude H(s), slope and the water system buffer zone distance D(s), forming a spatial feature vector X geo =[H(s),S(s),D(s)], where s is the spatial grid coordinate; Vegetation and soil data including Normalized Difference Vegetation Index Vegetation type one-hot encoding V type , soil texture classification S texture , and soil organic matter content OM = λ·spectral reflectance 400-2400nm +μ; the normalized vegetation index is based on satellite remote sensing red light DN RED and near-infrared DN NIR The soil organic matter content was obtained by inversion using the partial least squares regression model; Human activity data including agricultural irrigation intensity Among them ET a is the actual evapotranspiration, ET0 is the potential evapotranspiration, and ET ref As the reference evapotranspiration, the urban heat island index UHI = T urban -T rural , where T urban is the city center temperature, T rural is the average temperature of the suburbs, and the water storage rate of the reservoir Where V is the real-time storage capacity, V max The maximum storage capacity.

3. The prediction system for high temperature, heat wave and drought combined disasters according to claim 2, characterized in that: The constructing of a multi-dimensional input feature set includes: Perform sliding window feature extraction on hourly temperature T(t), relative humidity RH(t), and precipitation intensity P(t) to construct a time series matrix: Where n is the window length, and the derived indicator is calculated: the number of consecutive high temperature days Where I is the indicator function; Calculating slope based on digital elevation model Slope And generate the distance D(s) from the water system through GIS buffer analysis to form the spatial feature vector: X geo (s)=[H(s),S(s),A(s),D(s)]; Vegetation index calculated by satellite remote sensing Combined with land use classification, the vegetation type unique hot encoding V is generated type ∈R C , where C is the number of vegetation categories; soil characteristics are modeled by spectral analysis: organic matter content OM = PLSR (Reflectance 400-4400nm ), which is inversion based on partial least squares regression model, and soil texture is represented by triangular coordinate system (S texture =(clay,silt,sand); Agricultural irrigation intensity The reference evapotranspiration ET is calculated based on the Penman-Monteith formula ref ; Urban Heat Island Index UHI = T urban -T rural , through night light data and surface temperature regression model T s =αL+β calibration; reservoir water storage rate Calculate the water surface area S and reservoir capacity curve V=f(S) based on SAR image monitoring; The single-dimensional features are spatiotemporally aligned to the meteorological time series X meteo Align with the spatial grid data according to the time length t and grid coordinate s to form the spatiotemporal characteristic variables: X(t,s)=[X meteo (t),X geo (s),X veg-soil (s),X human (s)]∈R n×(3+4+(1+C)+3) ; Then perform principal component analysis (PCA) dimensionality reduction on the high-dimensional features and calculate the covariance matrix The first k principal components are extracted by eigenvalue decomposition, so that the cumulative variance contribution rate Get the feature matrix after dimensionality reduction Where N is the number of samples and m is the number of original features; Use Z-score standardization to process the features of each dimension: where μ j and σ j are the mean and standard deviation of the j-th dimension features, respectively, ensuring that the features input to the deep learning model have zero mean and unit variance.

4. The prediction system for high temperature, heat wave and drought combined disasters according to claim 3, characterized in that: The data processing module processes the time series of meteorological data Time length t of human activity data i Interpolation synchronization is performed and cubic spline interpolation method is used to generate a unified time series with equal time intervals to ensure that (t i =t0+i·Δt, where i=0, 1,…, N-1, Δt=1h; The terrain data, vegetation and soil data, and human activity data are rasterized according to a unified spatial grid, and the data-free areas are filled by Gaussian kernel interpolation to form a spatial grid matrix X geo / veg / human (s)∈R M×D , where M is the number of grid points and D is the number of single-dimensional features; Generate spatiotemporal feature matrix X ts =[X meteo (t i ,s j )]∈R N×M×(3+4+(1+C)+3) , where t i is the time index, s j is the spatial grid index.

5. The high temperature, heat wave and drought combined disaster prediction system according to claim 4 is characterized in that: The data processing module uses the improved interquartile range method to detect outliers in each dimension feature and calculates the first quartile of the k-th dimension feature. third quartile and interquartile range Then define the outlier threshold, the lower limit is Upper limit is To satisfy or Sample j of is repaired using linear interpolation of adjacent time points: (when is an outlier); Then the feature matrix is ​​Z-score standardized. in is the global mean of the k-th dimension feature, is the standard deviation, ensuring that the feature follows the N(0,1) distribution; Finally, calculate the covariance matrix X norm is the normalized characteristic matrix; the characteristic vector U=[u1,u2,…,u m ] and eigenvalues ​​λ1≥λ2≥…≥λ m ; Select the first k principal components to make the cumulative variance contribution rate Get the feature matrix after dimensionality reduction, Among them U k =[u1,u2,…,u k ] are the first k principal component vectors.

6. The prediction system for high temperature, heat wave and drought combined disasters according to claim 5, characterized in that: The double-layer convolutional layer structure of the CNN submodule is: The first convolution layer uses a 3×3 learnable convolution kernel. For the input spatial feature matrix X spatial ∈R H×W×D , where H, W are the spatial grid sizes, D is the single grid feature dimension, and convolution operation is performed: Z (1) =σ(W (1) *X spatial +b (1) ); where b (1) is the bias vector, σ is the ReLU activation function, and the output feature map Among them, C1=64 is the number of channels in the first layer; The second convolution layer uses a 1×1 convolution kernel Perform cross-channel feature fusion: Z (2) =BatchNorm(W (2) *Z (1) +b (2) ); Output spatial feature tensor Among them, C2=32 is the second layer channel, which extracts the high-order spatial features of terrain undulation and vegetation distribution through spatial dimension compression.

7. The prediction system for high temperature, heat wave and drought combined disasters according to claim 6, characterized in that: The neuron state update equation of the three-layer LSTM network is: Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f ) Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ), Cell status: Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o ),h t =o t ⊙tanh(c t ) in is the meteorological input characteristic at time t, h t ∈R L is the hidden state, where L = 128 is the number of neurons in a single layer of LSTM, and W f ,W i ,W c ,W o is the weight matrix, b f ,b i ,b c ,b o is the bias vector, ⊙ represents element-wise multiplication; The three-layer LSTM network is stacked Where l = 1, 2, 3 is the number of layers, capturing the long-range dependency features of the number of days with high temperature and precipitation anomalies in the time series, and outputting the final hidden state sequence H rnn ∈R T×L , where T is the time step.

8. The prediction system for high temperature, heat wave and drought combined disasters according to claim 7, characterized in that: The attention mechanism layer adopts the spatiotemporal cross attention model to convert the spatial features output by CNN into Flattened into key-value pairs The temporal feature H output by RNN rnn ∈R T×L Mapped to query vector through the fully connected layer Then the attention weight is calculated. where α t,s represents the attention weight of the spatial grid point s at time t; Finally, feature fusion Output fused spatiotemporal features Through dynamic weighting, the key influencing factors of vegetation coverage decline and urban heat island effect during the drought development period are highlighted; The specific generation method of the prediction output module includes: The attention features are transferred through the transposed convolution layer Upsample to the original spatial resolution H×W, and output the high temperature and drought risk probability matrix through the softmax layer: Among them C class =3 is the risk level, combining the terrain slope threshold and vegetation index threshold to mark the terrain sensitive area and ecologically fragile area; Through the fully connected layer f fc : Map attention features to disaster intensity indicators to generate future T pred =15-day intensity change curve: Y time (t) = f fc (X attn (t)),t∈[t0,t0+T pred ]; Human activity intervention variables include irrigation intensity I′ irr After one-hot encoding, the reservoir storage rate R′ is concatenated with the original features and input into the model. The disaster evolution path under the intervention scenario is simulated through the conditional generative adversarial network: The generator G uses a multi-layer perceptron architecture to output the simulated disaster intensity deviation value Quantify the disaster reduction effects of interventions.

9. The high temperature, heat wave and drought combined disaster prediction system according to claim 8, characterized in that: It also includes a real-time data access module, which is used to access the real-time data corresponding to the multi-source data acquisition module, including real-time meteorological data, real-time vegetation monitoring data, and real-time human activity data to form a real-time feature vector Where T′ is the time step of real-time data, and D is consistent with the feature dimension of multi-source data; Then, dynamic comparison and error calculation are performed, and the real-time feature vector is compared with the historical prediction data according to the preset prediction frequency. Compare and calculate the absolute error of each dimension feature And count the error exceeding the limit ratio where δ d is the error threshold of the d-th dimension feature; Then the weight is dynamically adjusted. When r>20%, the weight update of the attention mechanism layer is triggered. For the feature dimension with an error exceeding the limit, its weight coefficient α in the attention mechanism layer is increased. t,s ; For feature dimensions with smaller errors, maintain or reduce the weight coefficient; Finally, the forecast results are updated, and the adjusted weights are input into the deep learning model training module to regenerate the forecast results for the preset number of days in the future. The updated high temperature and drought risk heat map, intensity change curve and intervention scenario simulation results are pushed in real time through the forecast output module.

10. A method for predicting high temperature, heat wave and drought combined disasters, characterized in that: A system as claimed in any one of claims 1 to 9 is employed.

Images