Meteorological disaster dynamic monitoring method and system applied to real-time meteorological data

Abstract

The invention provides a meteorological disaster dynamic monitoring method and system applied to real-time meteorological data, and the method comprises the steps: firstly obtaining a real-time meteorological data set of a target region, and generating time change features and spatial distribution features of each meteorological data unit through spatial-temporal feature extraction processing; the method comprises the following steps: firstly, acquiring time change characteristics of a meteorological disaster, generating regional correlation characteristics according to the correlation between the time change characteristics and spatial distribution characteristics, then inputting the regional correlation characteristics into a pre-trained meteorological disaster prediction model for disaster risk assessment to generate a disaster prediction result, and finally, outputting the disaster prediction result. And generating a dynamic monitoring strategy including a monitoring equipment deployment scheme and an early warning information pushing rule according to the disaster prediction result, and feeding back the dynamic monitoring strategy to the meteorological monitoring platform to trigger a monitoring resource allocation operation, so that the meteorological disaster can be dynamically monitored by fully utilizing the real-time meteorological data, and the accuracy and timeliness of disaster early warning are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological disaster monitoring, and in particular to a method and system for dynamic monitoring of meteorological disasters applied to real-time meteorological data. Background Art

[0002] Meteorological disasters such as heavy rain, typhoons, and droughts have a significant impact on human production and life. Accurate and timely monitoring and early warning of meteorological disasters are of great significance for reducing disaster losses and protecting people's lives and property. Traditional meteorological disaster monitoring methods mainly rely on observation data from fixed meteorological stations. Although these data are accurate, they have problems such as limited spatial coverage and insufficient temporal resolution, making it difficult to comprehensively and dynamically reflect the evolution of meteorological disasters. In addition, traditional methods often focus on the analysis of single meteorological elements and lack in-depth exploration of the spatiotemporal correlations between meteorological elements, resulting in deficiencies in the accuracy and timeliness of disaster warnings. With the continuous development of meteorological observation technology, the acquisition of real-time meteorological data has become more convenient and abundant. However, how to efficiently process and analyze these data and extract valuable information for meteorological disaster monitoring remains a major challenge facing the current meteorological field. Summary of the Invention

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for dynamic monitoring of meteorological disasters applied to real-time meteorological data, the method comprising: Acquire a real-time meteorological data set of meteorological data units containing meteorological status information of a target area at multiple different time stamps; Performing spatiotemporal feature extraction processing on the real-time meteorological data set to generate temporal variation features and spatial distribution features of each meteorological data unit, wherein the temporal variation features reflect the evolution law of basic meteorological parameters at adjacent timestamps, and the spatial distribution features reflect the distribution pattern of basic meteorological parameters at different locations at the same timestamp; Generate a regional correlation feature based on the correlation between the time-varying feature and the spatial distribution feature, wherein the regional correlation feature includes a description of the influence of the time evolution law on the spatial distribution pattern, wherein the regional correlation feature is generated by extracting the time key features from the time-varying feature and the spatial key areas from the spatial distribution feature, establishing a timestamp mapping, and then calculating the Pearson correlation coefficient to construct a correlation relationship description vector, and fusing it with the time key features and the spatial key areas to generate the regional correlation feature; Calling a pre-trained meteorological disaster prediction model, inputting the regional correlation features into the meteorological disaster prediction model to perform disaster risk assessment processing, and generating a disaster prediction result indicating the potential probability of occurrence and impact range of meteorological disasters in the target area; A dynamic monitoring strategy including a monitoring equipment deployment plan and early warning information push rules is generated based on the disaster prediction results, and the dynamic monitoring strategy is fed back to the meteorological monitoring platform to trigger a monitoring resource allocation operation.

[0004] On the other hand, an embodiment of the present invention also provides a meteorological disaster dynamic monitoring system applied to real-time meteorological data, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0005] Based on the above aspects, the embodiment of the present invention realizes the comprehensive extraction and in-depth mining of the spatiotemporal characteristics of meteorological elements by comprehensively analyzing the meteorological data units containing meteorological status information of multiple different timestamps in the target area. It not only focuses on the temporal change characteristics of meteorological elements, but also deeply analyzes the spatial distribution characteristics of meteorological elements at different positions at the same timestamp, and further reveals the influence of time evolution laws on spatial distribution patterns, thereby constructing more accurate and comprehensive regional correlation features. Inputting these regional correlation features into a pre-trained meteorological disaster prediction model can significantly improve the accuracy and timeliness of disaster risk assessment and generate more accurate disaster prediction results. Based on these disaster prediction results, a monitoring strategy including monitoring equipment deployment plans and warning information push rules can be dynamically generated to achieve optimal allocation of monitoring resources, thereby more effectively responding to meteorological disasters and reducing disaster losses. Therefore, by integrating multiple links such as real-time meteorological data, spatiotemporal feature extraction, correlation feature generation and disaster prediction models, the ability of meteorological disaster monitoring and early warning is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 The present invention provides a method for dynamically monitoring meteorological disasters using real-time meteorological data.

[0007] Figure 2 Schematic diagram of exemplary hardware and software components of a meteorological disaster dynamic monitoring system applied to real-time meteorological data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0008] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a method for dynamic monitoring of meteorological disasters applied to real-time meteorological data provided by an embodiment of the present invention. The method for dynamic monitoring of meteorological disasters applied to real-time meteorological data is introduced in detail below.

[0009] Step S110: obtaining a real-time meteorological data set of meteorological data units containing meteorological status information of multiple different time stamps in the target area.

[0010] In this embodiment, the real-time meteorological data set is composed of multiple meteorological data units, each of which contains meteorological status information at a specific time stamp. The collection of meteorological data relies on various types of monitoring equipment, which are distributed in different locations in the target area.

[0011] Ground-based weather stations are a key data collection source. Equipped with a variety of sensors, they provide real-time measurements of basic meteorological parameters. For example, temperature sensors sense the temperature of the surrounding air and record the corresponding value. Humidity sensors detect the water vapor content in the air. Pressure sensors acquire atmospheric pressure data. Wind speed and direction sensors measure the speed and direction of air flow, respectively. These sensors record data at different time points, and each recorded meteorological data unit constitutes a meteorological data unit.

[0012] Satellite remote sensing equipment monitors the target area from high altitude, capturing information such as the distribution of clouds and the amount of water vapor in the atmosphere. The satellite scans and collects data at set intervals, with each data point being counted as a meteorological data unit.

[0013] Radar monitoring equipment focuses on monitoring the dynamic changes in meteorological phenomena such as precipitation and storms. It transmits and receives electromagnetic waves to analyze information such as the location, intensity, and movement speed of meteorological targets. Radar equipment continuously monitors the surrounding area and records relevant data at different times, forming meteorological data units.

[0014] Because different monitoring devices may have different data collection times and frequencies, the collected data needs to be consolidated. During this consolidation process, data accuracy and consistency must be considered. In the case of inconsistent data collection times, time synchronization is required to ensure that each meteorological data unit accurately corresponds to its corresponding timestamp. Data quality control is also required to check for missing values ​​or outliers. Missing values ​​can be supplemented through interpolation; outliers need to be corrected or removed.

[0015] Each monitoring device transmits collected data to the data center via wired or wireless communication networks. To protect data security and privacy, encryption technology is used during transmission to prevent data theft or tampering. The data center stores and manages the received meteorological data for subsequent analysis and processing.

[0016] Step S120: Perform spatio-temporal feature extraction processing on the real-time meteorological data set to generate the time-varying features and spatial distribution features of each meteorological data unit. The time-varying features reflect the evolution law of the basic meteorological parameters at adjacent timestamps, and the spatial distribution features reflect the distribution pattern of the basic meteorological parameters at different positions at the same timestamp.

[0017] Step S121: Perform time dimension alignment processing on the real-time meteorological data set to obtain a meteorological data sequence after time alignment.

[0018] Since there may be differences in the data collection times of different monitoring devices, in order to facilitate subsequent analysis and processing, it is necessary to perform time dimension alignment processing on the real-time meteorological data set. First, determine a unified time reference and time interval. The time reference can be a specific starting time point, and the time interval is set according to actual needs. For example, a shorter time interval can be selected to more accurately capture meteorological changes.

[0019] After determining the time reference and time interval, each meteorological data unit in the real-time meteorological data set needs to be arranged in chronological order. For data with timestamps not within the unified time interval, corresponding processing is required. If the timestamp of a meteorological data unit is between two unified time intervals, an interpolation method can be used to estimate the meteorological parameter value at that time point. There are various interpolation methods, such as linear interpolation, which assumes that the change in meteorological parameters between two adjacent known timestamps is linear. Specifically, if the meteorological parameter values corresponding to timestamps t1 and t2 are A1 and A2 respectively, and we want to estimate the meteorological parameter value A corresponding to timestamp t (t1 < t < t2), we can calculate the value of A according to the proportional relationship based on the relative positions of t, t1, and t2.

[0020] Through the above time dimension alignment processing, a meteorological data sequence after time alignment can be obtained. Each timestamp in this meteorological data sequence corresponds to a complete set of meteorological parameter values.

[0021] Step S122: Perform time series analysis processing on the meteorological data sequence after time alignment, calculate the change rate parameter and change direction parameter of each basic meteorological parameter at consecutive timestamps, and construct a time evolution curve based on the change rate parameter and change direction parameter as the time-varying feature.

[0022] Step S1221: Extract the basic meteorological parameter values at adjacent timestamps in the meteorological data sequence after time alignment, and calculate the difference between the parameter values at adjacent timestamps as the absolute change amount.

[0023] In this embodiment, after obtaining a time-aligned meteorological data sequence, in order to analyze the temporal evolution of each basic meteorological parameter, it is necessary to calculate the absolute change between adjacent timestamps. Taking temperature as an example, assuming that in the time-aligned meteorological data sequence, timestamp t1 corresponds to temperature value T1, and timestamp t2 (which immediately follows t1) corresponds to temperature value T2, then the absolute change between adjacent timestamps is the difference between T2 and T1. If T2 is greater than T1, the difference is positive, indicating an increase in temperature; if T2 is less than T1, the difference is negative, indicating a decrease in temperature.

[0024] The same method is used to calculate the absolute changes between adjacent timestamps for other basic meteorological parameters, such as humidity, air pressure, wind speed, etc. These absolute changes reflect the magnitude of change of each basic meteorological parameter between adjacent timestamps.

[0025] Step S1222: Calculate the change rate parameter per unit time according to the absolute change amount and the time interval length.

[0026] After calculating the absolute change between adjacent timestamps, we also need to consider the length of the time interval to determine the rate of change per unit time. The time interval is the time difference between two adjacent timestamps. For example, if the time interval between timestamps t1 and t2 is Δt, and the absolute temperature change corresponding to the adjacent timestamps is calculated as ΔT, then the rate of change per unit time is ΔT divided by Δt.

[0027] For other basic meteorological parameters, the same method is used to calculate the change rate parameter per unit time. The change rate parameter per unit time can more accurately reflect the speed of change of the basic meteorological parameters per unit time.

[0028] Step S1223: Determine a change direction parameter by comparing the magnitude relationship of adjacent timestamp parameter values, where the change direction parameter includes an upward trend indicator and a downward trend indicator.

[0029] Determining the direction of change parameter allows for further analysis of the changing trends of basic meteorological parameters. For each basic meteorological parameter, the parameter values ​​corresponding to adjacent timestamps are compared. If the parameter value at the subsequent timestamp is greater than the parameter value at the previous timestamp, it can be determined that the basic meteorological parameter exhibited an upward trend within that time interval, indicated by an upward trend indicator. If the parameter value at the subsequent timestamp is less than the parameter value at the previous timestamp, the basic meteorological parameter exhibited a downward trend, indicated by a downward trend indicator.

[0030] For example, for temperature parameters, if T2 is greater than T1, then the temperature change direction parameter indicates an upward trend; if T2 is less than T1, then the temperature change direction parameter indicates a downward trend. The same method applies to other basic meteorological parameters such as humidity, air pressure, and wind speed.

[0031] Step S1224: Perform sliding window smoothing on the change rate parameters and change direction parameters of continuous timestamps, arrange the smoothed change rate parameters and change direction parameters in timestamp order, and generate a time evolution curve that reflects the evolution law of basic meteorological parameters over time. The horizontal axis of the time evolution curve is the timestamp, the vertical axis is the change rate parameter, and the slope sign of the curve is determined by the change direction parameter.

[0032] To ensure that the rate and direction parameters more accurately reflect the true evolution of the underlying meteorological parameters, these parameters at consecutive timestamps must be smoothed using a sliding window. A sliding window is a fixed-size time interval that slides across the time series. Within each window, the rate of change parameters are smoothed, for example, using an average method.

[0033] Specifically, suppose there is a sliding window of size n. For timestamp ti, the time interval covered by the window is [ti-n / 2, ti+n / 2] (assuming n is an even number). The change rate parameters within the sliding window are V1, V2,…, Vn respectively. Then the change rate parameter Vi' after smoothing is the average value of these n change rate parameters.

[0034] For the change direction parameter, a majority vote can be used within the sliding window to determine the smoothed change direction parameter. That is, the number of rising trend indicators and falling trend indicators within the window is counted, and the indicator with the larger number is used as the smoothed change direction parameter.

[0035] Arranging the smoothed rate and direction parameters in timestamp order creates a time evolution curve. The abscissa of the time evolution curve represents the timestamp, and the ordinate represents the rate parameter. The sign of the slope of the curve is determined by the direction parameter: if the direction parameter indicates an upward trend, the slope is positive; if the direction parameter indicates a downward trend, the slope is negative. This time evolution curve intuitively reflects the temporal evolution of basic meteorological parameters.

[0036] Step S123: performing spatial position gridding processing on the real-time meteorological data set, dividing the target area into evenly distributed spatial grid units, and extracting the mean statistics and variance statistics of the basic meteorological parameters at the same time stamp in each spatial grid unit.

[0037] Step S1231: Determine the number of horizontal and vertical divisions of the spatial grid unit according to the geographic boundary range of the target area and the preset grid division accuracy.

[0038] In this embodiment, the number of horizontal and vertical divisions of the spatial grid cells is first determined based on the geographic boundary of the target area and the preset grid division accuracy. The geographic boundary of the target area can be obtained using tools such as a geographic information system (GIS), which clearly defines the spatial extent of the target area.

[0039] The preset grid division accuracy determines the size of the grid cells. If the grid division accuracy is high, the grid cells will be smaller, which can more finely reflect the changes in meteorological parameters in the target area; if the grid division accuracy is low, the grid cells will be larger, which can describe the meteorological parameters more comprehensively.

[0040] The number of horizontal divisions can be calculated based on the horizontal length of the target area and the preset horizontal grid unit size; similarly, the number of vertical divisions can be calculated based on the vertical length of the target area and the preset vertical grid unit size.

[0041] Step S1232: performing equidistant grid division processing on the target area based on the number of horizontal divisions and the number of vertical divisions to generate a set of spatial grid units covering the target area.

[0042] After determining the number of horizontal and vertical divisions of the spatial grid cell, the target area can be divided into equidistant grids. Starting from one boundary of the target area, the target area is divided into multiple spatial grid cells of equal size according to the number of horizontal and vertical divisions. These spatial grid cells are closely arranged to cover the entire target area, forming a spatial grid cell set.

[0043] Each spatial grid cell has its own unique position coordinates, through which the position of each grid cell in the target area can be accurately located.

[0044] Step S1233: Collect the basic meteorological parameter values ​​at the same time stamp in each spatial grid cell, and calculate the arithmetic mean of all basic meteorological parameter values ​​in the grid cell as the mean statistic.

[0045] For each spatial grid cell, at the same timestamp, the basic meteorological parameter values ​​recorded by all meteorological monitoring equipment within the grid cell are collected. For example, at a certain timestamp t, the temperature, humidity, air pressure and other parameter values ​​recorded by each meteorological station within the grid cell are collected.

[0046] Then, the arithmetic mean of all the basic meteorological parameter values ​​within the grid cell is calculated. Taking temperature as an example, if there are m weather stations in the grid cell recording temperature values ​​T1, T2, ..., Tm, then the mean statistic of temperature is the sum of these m temperature values ​​divided by m.

[0047] The same method is used to calculate mean statistics for other basic meteorological parameters, such as humidity and air pressure. These mean statistics reflect the average level of the basic meteorological parameters within the grid cell.

[0048] Step S1234: Calculate the average of the squares of the differences between the basic meteorological parameter values ​​and the mean statistic in the grid cell as the variance statistic, which reflects the degree of dispersion of the basic meteorological parameter in the grid cell.

[0049] After calculating the mean statistic, we also need to calculate the variance statistic to reflect the degree of dispersion of the basic meteorological parameters within the grid cell. For each basic meteorological parameter, we calculate the difference between each parameter value within the grid cell and the mean statistic, and then square these differences.

[0050] Taking temperature as an example, assuming that the mean temperature statistic is T_mean, and the temperature values ​​recorded by each weather station in the grid cell are T1, T2, ..., Tm, then the difference between each temperature value and the mean statistic is T1-T_mean, T2-T_mean, ..., Tm-T_mean, and squaring these differences gives (T1-T_mean)^2, (T2-T_mean)^2, ..., (Tm-T_mean)^2.

[0051] Next, the average of these squared differences is calculated, which is the variance statistic. A larger variance statistic indicates that the basic meteorological parameters are more dispersed within the grid cell and the distribution of meteorological parameters is more uneven; a smaller variance statistic indicates that the basic meteorological parameters are less dispersed within the grid cell and the distribution of meteorological parameters is more uniform.

[0052] Step S1235: Arrange the mean statistics and variance statistics according to the position coordinates of the spatial grid units to generate a statistics matrix with spatial position indexes.

[0053] Finally, the calculated mean and variance statistics are arranged according to the position coordinates of the spatial grid cells. Each spatial grid cell has its corresponding mean and variance statistics, and these statistics are sorted according to the position order of the grid cells to form a statistical matrix.

[0054] The statistical matrix has a spatial location index, and the rows and columns of the matrix can accurately correspond to each spatial grid cell. The above statistical matrix can clearly show the spatial distribution of the mean and dispersion of the basic meteorological parameters of each spatial grid cell in the target area.

[0055] Step S124: constructing a spatial distribution heat map as a spatial distribution feature based on the mean statistic and the variance statistic, wherein the color gradient of the spatial distribution heat map corresponds to the distribution density of the basic meteorological parameters.

[0056] Using the previously obtained statistical matrix with spatial location indexes, a spatial distribution heat map is constructed. A spatial distribution heat map is an intuitive visualization tool that can show the spatial distribution pattern of basic meteorological parameters within the target area.

[0057] When constructing a spatial distribution heat map, the color corresponding to each spatial grid cell is determined based on the mean and variance statistics. The mean statistic reflects the average level of the basic meteorological parameter within that grid cell, while the variance statistic reflects the dispersion of the basic meteorological parameter within that grid cell. By combining these two statistics, a corresponding relationship can be established between the distribution density of the basic meteorological parameter and the color gradient.

[0058] For example, a color mapping rule can be set. When the mean statistic is in a certain range and the variance statistic is small, the corresponding color is darker, indicating that the basic meteorological parameters in the area are densely distributed and the values ​​are relatively concentrated; when the mean statistic is in another range and the variance statistic is large, the corresponding color is lighter, indicating that the basic meteorological parameters in the area are more dispersed.

[0059] Each spatial grid cell is filled with a corresponding color to form a spatial distribution heat map. By observing the spatial distribution heat map, we can intuitively understand the distribution pattern of basic meteorological parameters in the target area, which areas have dense and scattered distribution of parameters.

[0060] Step S125: normalizing the temporal variation feature and the spatial distribution feature to obtain a spatiotemporal feature set.

[0061] In order to make temporal variation characteristics and spatial distribution characteristics comparable and consistent, they need to be standardized. Standardization can eliminate the dimensional differences between different characteristics and make them within the same scale range.

[0062] For time-varying features, that is, the time evolution curve generated earlier, normalization can be used. The purpose of normalization is to map the numerical range of the time evolution curve to a set interval, such as [0, 1]. Normalization can be achieved by calculating the ratio of the difference between each data point in the time evolution curve and the maximum and minimum values ​​of the curve.

[0063] For spatial distribution features, i.e., spatial distribution heat maps, the color values ​​in the spatial distribution heat map can be standardized. The color value range can be determined first, and then the color value corresponding to each grid cell can be linearly transformed to be within a uniform value range.

[0064] After normalization, the temporal variation characteristics and spatial distribution characteristics are integrated to obtain a spatiotemporal feature set, which contains information on the temporal evolution and spatial distribution patterns of meteorological data in the target area.

[0065] Step S130: Generate regional association features based on the correlation between the time-varying features and the spatial distribution features, wherein the regional association features include a description of the influence of the time evolution law on the spatial distribution pattern, wherein the time key features in the time-varying features and the spatial key areas in the spatial distribution features are extracted, and after establishing a timestamp mapping, the Pearson correlation coefficient is calculated to construct an association relationship description vector, and is fused with the time key features and the spatial key areas to generate regional association features.

[0066] Step S131: extracting the time evolution curve in the time variation feature and the spatial distribution heat map in the spatial distribution feature.

[0067] When generating regional correlation features, we first need to extract key information from the temporal variation and spatial distribution features obtained previously. From the temporal variation features, we extract a time evolution curve. This curve reflects the temporal evolution of basic meteorological parameters, including information such as the rate and direction of change per unit time.

[0068] The spatial distribution heat map is extracted from the spatial distribution characteristics. The spatial distribution heat map intuitively shows the spatial distribution pattern of basic meteorological parameters in the target area and reflects the distribution density of basic meteorological parameters through color gradient.

[0069] Step S132: Perform feature sampling processing on the time evolution curve, extract the change rate parameters and change direction parameters of the key time nodes as time key features, and perform region segmentation processing on the spatial distribution heat map to identify the boundary coordinates of the concentrated distribution area and the discrete distribution area as the spatial key area.

[0070] Feature sampling is performed on the time evolution curve to extract information about key time nodes. Key time nodes are those points in the time evolution curve where significant changes occur. A change rate threshold can be set. When the change rate parameter at a time point in the time evolution curve exceeds the threshold, that time point is identified as a key time node.

[0071] For each key time node, the corresponding change rate parameter and change direction parameter are extracted, which constitute the time key feature. The time key feature can more prominently reflect the changes of basic meteorological parameters at key time points.

[0072] Perform regional segmentation on the spatial distribution heat map to identify the boundary coordinates of concentrated and discrete distribution areas. Image segmentation algorithms can be used to distinguish different regions based on the color gradient of the heat map. Darker areas typically indicate dense distribution of basic meteorological parameters, representing concentrated distribution areas (e.g., areas with a color gradient greater than a first set value); lighter areas indicate dispersed distribution of basic meteorological parameters, representing discrete distribution areas (e.g., areas with a color gradient less than a second set value). The boundary coordinates of these regions are determined using an image segmentation algorithm, and these boundary coordinates constitute the spatial key regions.

[0073] Step S133: Establish a mapping relationship between the timestamp of the time key feature and the acquisition timestamp of the spatial key area, so that the time reference of the time dimension is consistent with the time reference of the spatial dimension.

[0074] To accurately analyze the impact of temporal evolution on spatial distribution patterns, it is necessary to establish a mapping between the timestamps of temporal key features and the collection timestamps of spatial key areas. Since temporal key features are extracted from temporal evolution curves, their timestamps reflect the temporal changes in basic meteorological parameters. Spatial key areas are identified from spatial distribution heat maps, and their collection timestamps record the moment when the spatial distribution information was collected.

[0075] By establishing a mapping relationship, the timestamps of the time-critical features are matched with the acquisition timestamps of the spatial key regions, ensuring that the time bases of the temporal and spatial dimensions are consistent. For example, if a time-critical feature has a timestamp of t1, the corresponding region in the spatial key region with the acquisition timestamp closest to t1 is found and a mapping relationship is established between the two. This ensures that the temporal and spatial information are consistent in the temporal dimension when subsequently analyzing the impact of temporal evolution on spatial distribution patterns, avoiding erroneous analysis caused by time inconsistencies.

[0076] Step S134: For the change rate parameter and change direction parameter corresponding to each time key feature, the mean statistical change of the spatial key area under the corresponding time stamp is counted, and the Pearson correlation coefficient of the change rate parameter and the mean statistical change is calculated. The absolute value of the Pearson correlation coefficient reflects the impact intensity, and the sign of the correlation coefficient reflects the impact direction.

[0077] Each time-critical feature includes a rate of change parameter and a direction of change parameter. Based on the established mapping between the time-critical feature's timestamp and the spatial key region's acquisition timestamp, the change in the mean statistic for the spatial key region at the corresponding timestamp is calculated. The change in mean statistic refers to the difference between the mean statistic for the spatial key region at that timestamp and the mean statistic at the previous related timestamp.

[0078] Taking the temperature parameter as an example, if the timestamp of the time key feature is t, the temperature mean statistic of the corresponding spatial key area at timestamp t is M1, and the temperature mean statistic at the previous related timestamp t-1 is M0, then the change in the temperature mean statistic ΔM=M1-M0.

[0079] For each temporal key feature's rate of change parameter and the corresponding spatial key region's mean statistic change, the Pearson correlation coefficient is calculated. The Pearson correlation coefficient measures the degree of linear correlation between two variables. By calculating the Pearson correlation coefficient between the rate of change parameter and the mean statistic change, we can analyze the impact of the rate of change of the underlying meteorological parameters in the temporal evolution pattern on the mean statistic change in the spatial distribution pattern.

[0080] The larger the absolute value of the correlation coefficient, the stronger the linear relationship between the change rate parameter and the change in the mean statistic, that is, the greater the influence of the temporal evolution law on the spatial distribution pattern; a positive sign of the correlation coefficient indicates that the change rate parameter is positively correlated with the change in the mean statistic, that is, when the change rate increases, the mean statistic also tends to increase, reflecting a positive influence direction; a negative sign of the correlation coefficient indicates that the two are negatively correlated, that is, when the change rate increases, the mean statistic tends to decrease, reflecting a negative influence direction.

[0081] Step S135: Perform trend analysis on the Pearson correlation coefficient of the continuous time key features to identify the stable influence relationship and the fluctuating influence relationship, extract the correlation coefficient of the stable influence relationship as the basic influence intensity, extract the correlation coefficient variation range of the fluctuating influence relationship as the influence intensity fluctuation interval, and comprehensively generate a description of the influence direction and influence intensity of the time evolution law on the spatial distribution pattern.

[0082] After obtaining the Pearson correlation coefficient corresponding to each time key feature, trend analysis is performed on the correlation coefficients of continuous time key features. By observing the changes in the correlation coefficients over time, stable influence relationships and fluctuating influence relationships can be identified.

[0083] A stable influence relationship refers to a situation where the correlation coefficient remains relatively stable over a period of time, with minimal fluctuations. In this case, the correlation coefficient within this stable period is extracted as the base influence strength. The base influence strength reflects the degree to which the temporal evolution law has a relatively stable impact on the spatial distribution pattern during this period.

[0084] A fluctuating influence relationship refers to a situation where the correlation coefficient fluctuates significantly over a period of time. For this type of relationship, the range of the correlation coefficient—the interval between the maximum and minimum values—is extracted as the influence intensity fluctuation interval. This influence intensity fluctuation interval reflects the range of influence intensity of the temporal evolution law on the spatial distribution pattern over that period.

[0085] Combining the basic impact intensity and the impact intensity fluctuation range, as well as the positive and negative signs of the correlation coefficient, we can generate a description of the impact direction and impact intensity of the time evolution law on the spatial distribution pattern. It details how the time evolution law affects the spatial distribution pattern, including whether the impact direction is positive or negative, and whether the impact intensity is stable or fluctuates within a set range.

[0086] Step S136: constructing an association relationship description vector based on the impact direction and impact intensity, fusing the association relationship description vector with the time key feature and the space key area to generate a regional association feature containing the spatiotemporal impact relationship.

[0087] Based on the previously obtained description of the direction and intensity of the impact of the temporal evolution on the spatial distribution pattern, a correlation description vector is constructed. This correlation description vector contains information about the direction and intensity of the impact and can be represented using multiple dimensions. For example, one dimension represents the direction of the impact (e.g., positive is 1, negative is -1), while other dimensions represent information such as the baseline impact intensity and the range of impact intensity fluctuations.

[0088] The association description vector is fused with temporal key features and spatial key regions. Temporal key features include the rate and direction of change parameters of basic meteorological parameters at key time points, while spatial key regions include information such as the boundary coordinates of concentrated and discrete spatial distribution areas of basic meteorological parameters. This fusion process can be performed using a splicing approach, combining the association description vector, temporal key features, and spatial key regions in a predetermined order to form regional association features that reflect spatiotemporal influence relationships. Regional association features integrate information about temporal evolution patterns and spatial distribution patterns, as well as their mutual influence.

[0089] Step S140: calling a pre-trained meteorological disaster prediction model, inputting the regional correlation features into the meteorological disaster prediction model for disaster risk assessment processing, and generating a disaster prediction result indicating the potential probability of occurrence and impact range of meteorological disasters in the target area.

[0090] Step S141: Standardize the regional correlation features, input the standardized regional correlation features into the input layer of the meteorological disaster prediction model, perform feature recombination processing on the regional correlation features through a fully connected network, and generate an intermediate feature vector with a model adaptation dimension.

[0091] To make regional correlation features more suitable for the input requirements of meteorological disaster prediction models, we first normalize them. This normalization process uses a similar approach to the normalization of temporal variation features and spatial distribution features. This process eliminates dimensional differences between different dimensions in the regional correlation features and maps their numerical range to a uniform interval, such as [0, 1].

[0092] The standardized regional correlation features are fed into the input layer of the meteorological disaster prediction model. After receiving the regional correlation features, the input layer undergoes feature reorganization processing through a fully connected network. Each neuron in the fully connected network is connected to all neurons in the input layer. Through a series of linear transformations and nonlinear activation functions, the regional correlation features are recombined and transformed.

[0093] During the feature reorganization process, the fully connected network adjusts the dimensions of the regional correlation features to meet the requirements of subsequent processing within the model, generating an intermediate feature vector with model-adaptive dimensions. This intermediate feature vector incorporates the transformed and reorganized regional correlation features, making it more useful for the model's subsequent assessment of meteorological disaster risk.

[0094] Step S142: performing nonlinear transformation processing on the intermediate feature vector through the hidden layer of the meteorological disaster prediction model to extract deep semantic features containing spatiotemporal impact relationships.

[0095] The intermediate eigenvectors enter the hidden layer of the meteorological disaster prediction model. The hidden layer is the core of the model and contains multiple neurons and a multi-layer structure. In the hidden layer, the intermediate eigenvectors undergo nonlinear transformations. Nonlinear transformations are achieved using activation functions, such as the ReLU function.

[0096] Through nonlinear transformations, the hidden layer can tap into more complex and in-depth information within the intermediate feature vectors, extracting deep semantic features that encompass temporal and spatial influence relationships. These deep semantic features are not simply transformations of regional correlation features; rather, they capture the more subtle interrelationships between temporal evolution patterns and spatial distribution patterns, as well as their potential links to meteorological disaster occurrence. The multi-layered structure of the hidden layer allows for the progressive extraction of higher-level features, enabling meteorological disaster prediction models to better understand the complex patterns in meteorological data.

[0097] Step S143: calling the probability prediction layer of the meteorological disaster prediction model to perform probability distribution calculation processing on the deep semantic features to generate potential occurrence probability values ​​of different meteorological disaster types in the target area.

[0098] Deep semantic features are incorporated into the probabilistic prediction layer of the meteorological disaster prediction model. The probabilistic prediction layer calculates the potential occurrence probabilities of different meteorological disaster types within the target area based on the deep semantic features. Through a series of calculations and processing, this layer maps the deep semantic features into the probability space of different meteorological disaster types.

[0099] The probability prediction layer may use classification algorithms or probability distribution models for calculations. For example, it can determine the degree of correlation between each feature and different meteorological disaster types based on the feature values ​​of deep semantic features. Then, through statistical analysis and other methods, it calculates the potential probability of each meteorological disaster type. Ultimately, it outputs the potential probability of occurrence of different meteorological disaster types within the target area, such as heavy rain, high winds, and lightning.

[0100] Step S144: performing spatial range inference processing on the deep semantic features through the range prediction layer of the meteorological disaster prediction model to generate boundary coordinates of the impact range corresponding to different meteorological disaster types.

[0101] Step S1441: input the deep semantic features into the spatial mapping sublayer of the range prediction layer, perform spatial position feature enhancement processing on the deep semantic features through a convolutional neural network, and generate a feature map with spatial position information.

[0102] The deep semantic features are fed into the spatial mapping sublayer of the range prediction layer of the meteorological disaster prediction model. This sublayer processes the deep semantic features using a convolutional neural network, which has powerful feature extraction and spatial information capture capabilities.

[0103] In the spatial mapping sublayer, the convolutional neural network performs sliding convolutions on the deep semantic features using convolution kernels to extract the spatial positional features within them. The convolution kernels learn local patterns of features at different spatial locations, and through multiple convolution and pooling operations, they strengthen the spatial positional information within the deep semantic features.

[0104] After processing by the convolutional neural network, a feature map with spatial location information is generated. Each element in the feature map corresponds to a specific spatial location in the target area and contains relevant feature information at that location.

[0105] Step S1442: performing threshold segmentation processing on the feature map, and extracting areas where the feature values ​​exceed a preset threshold as candidate disaster impact areas.

[0106] After obtaining a feature map with spatial location information, it is subjected to threshold segmentation. The preset threshold is a feature value boundary determined based on extensive historical data and experience. By comparing the value of each element in the feature map with the preset threshold, areas with feature values ​​exceeding the preset threshold are extracted as candidate disaster impact areas.

[0107] The purpose of threshold segmentation is to identify areas with high eigenvalues ​​that may be associated with meteorological disasters. In a feature map, areas with high eigenvalues ​​generally indicate a greater likelihood of being affected by a meteorological disaster. Using threshold segmentation, we can initially determine the approximate range of areas potentially affected by a meteorological disaster.

[0108] Step S1443: performing morphological processing on the candidate disaster-affected regions, eliminating isolated small regions and connecting adjacent regions to generate continuous disaster-affected connected regions.

[0109] Morphological processing is performed on the candidate disaster-affected areas. Morphological processing includes operations such as dilation and erosion. The dilation operation can expand the scope of the candidate disaster-affected area and connect adjacent small areas; the erosion operation can eliminate isolated small areas in the candidate disaster-affected area.

[0110] By combining dilation followed by erosion or other appropriate morphological operations, we eliminate isolated small regions within candidate disaster-affected areas and connect adjacent regions to form continuous, connected regions. Continuous, connected disaster-affected regions better reflect the impact range of actual meteorological disasters and avoid interference from discontinuous regions caused by noise or local feature fluctuations.

[0111] Step S1444: extracting the coordinates of the outer contour boundary points of the disaster-affected connected area, performing polygon fitting processing on the boundary point coordinates, and generating a polygon boundary coordinate sequence that approximately represents the affected range.

[0112] After obtaining the continuous disaster-affected connected region, the coordinates of its outer contour boundary points are extracted. An edge detection algorithm can be used to find the outer contour of the disaster-affected connected region. The edge detection algorithm calculates the gradient change of pixel values ​​in the feature map to identify the boundary points between the disaster-affected connected region and the surrounding area.

[0113] After obtaining the coordinates of the outer contour boundary points, polygon fitting is performed on these boundary points. The purpose of polygon fitting is to approximate the disaster impact range using a polygon. Methods such as the least squares method can be used to find a polygon whose boundary is as close as possible to the coordinates of the outer contour boundary points. Ultimately, a polygon boundary coordinate sequence is generated that approximates the impact range and clearly defines the approximate boundaries of the meteorological disaster impact range.

[0114] Step S1445: Calculate the area parameters and center coordinate parameters of the influence range according to the polygon boundary coordinate sequence, and use the polygon boundary coordinate sequence, area parameters and center coordinate parameters as description information of the boundary coordinates of the influence range.

[0115] Based on the polygon boundary coordinate sequence, calculate the area parameters and center coordinate parameters of the meteorological disaster impact range. The area parameters can be obtained by calculating the area of ​​the region enclosed by the polygon. The total area can be calculated by splitting the polygon into multiple triangles, calculating the area of ​​each triangle, and summing them.

[0116] The center coordinate parameters can be obtained by calculating the average value of the coordinates of all vertices of the polygon, which represents the approximate center position of the meteorological disaster impact range.

[0117] The polygon boundary coordinate sequence, area parameters, and center coordinate parameters are combined together to form the descriptive information of the boundary coordinates of the impact range. This descriptive information details the impact range corresponding to different meteorological disaster types, including the boundary, size, and center location of the range.

[0118] Step S145: performing an association matching process on the potential occurrence probability value and the boundary coordinates of the impact range to generate a disaster prediction result including the disaster type, potential occurrence probability and the boundary coordinates of the impact range.

[0119] The potential occurrence probability values ​​of different meteorological disaster types obtained from the probability prediction layer are associated and matched with the boundary coordinates of the impact range obtained from the range prediction layer. For each meteorological disaster type, its corresponding potential occurrence probability value and the boundary coordinates of the impact range are correspondingly combined.

[0120] For example, for a rainstorm disaster, its potential probability of occurrence is combined with the boundary coordinates of the rainstorm disaster's impact range (including the polygon boundary coordinate sequence, area parameters, and center coordinate parameters). This association matching process is performed for all meteorological disaster types, ultimately generating a disaster prediction result that includes the disaster type, potential probability of occurrence, and the boundary coordinates of the impact range. This disaster prediction result comprehensively reflects the potential occurrence and possible impact range of different meteorological disaster types within the target area.

[0121] Step S150: Generate a dynamic monitoring strategy including a monitoring equipment deployment plan and early warning information push rules based on the disaster prediction results, and feed the dynamic monitoring strategy back to the meteorological monitoring platform to trigger a monitoring resource allocation operation.

[0122] Step S151: parsing the potential occurrence probability values ​​and the boundary coordinates of the impact range in the disaster prediction results, and extracting the target disaster type and its target impact range corresponding to the potential occurrence probability greater than a preset probability threshold.

[0123] After receiving the disaster forecast results, they are analyzed. The disaster forecast results include the potential probability of occurrence of different meteorological disaster types and the coordinates of the impact range boundaries. The preset probability threshold is a probability limit determined based on actual needs and historical experience.

[0124] By comparing the potential probability of each meteorological disaster type with a preset probability threshold, the meteorological disaster types with a potential probability greater than the preset probability threshold are selected as target disaster types. The boundary coordinates of the impact range corresponding to these target disaster types are also extracted as the target impact range. The target disaster type and target impact range will be key areas of focus in the subsequent development of monitoring equipment deployment plans and warning information push rules.

[0125] Step S152: determining the target area to be monitored according to the boundary coordinates of the target influence range, and calculating the number of monitoring devices to be deployed based on the area size and geographical distribution density of the target area.

[0126] Determine the target area to be monitored based on the boundary coordinates of the target impact range. The target area is the area potentially affected by the target disaster type. Analyze the size and geographic distribution density of the target area. Size reflects the scale of the target area, while geographic distribution density reflects the complexity of the meteorological conditions within the target area.

[0127] For target areas with large areas and high geographic density, more monitoring equipment may be needed to comprehensively monitor meteorological conditions. For target areas with smaller areas and lower geographic density, a relatively small number of monitoring devices may be required. The number of monitoring devices required can be calculated based on the target area's size and geographic density using a pre-established empirical model or algorithm. This calculation takes into account the actual conditions of the target area to ensure that the monitoring equipment effectively covers the target area.

[0128] Step S153: combining the location distribution information of existing monitoring devices in the target area, determining the deployment location of the newly added monitoring devices through an optimization algorithm so that the coverage of the newly added monitoring devices maximizes overlap with the target influence range.

[0129] For example, step S1531: obtaining location distribution information of existing monitoring devices in the target area, where the location distribution information includes the geographic coordinates of each existing monitoring device.

[0130] To determine the deployment location of new monitoring equipment, we first need to obtain the location distribution information of existing monitoring equipment within the target area. This information is typically stored in the database of the meteorological monitoring platform. The location distribution information includes the geographic coordinates of each existing monitoring device, which can be used to accurately determine the specific location of existing monitoring equipment within the target area.

[0131] Step S1532: performing regional sensitivity division on the target impact range to generate a multi-level sensitive area set including a core sensitive area and an extended sensitive area.

[0132] Conduct regional sensitivity analysis of the target impact area. Based on the characteristics of meteorological disasters and historical data, analyze the extent to which different areas within the target impact area are affected by meteorological disasters. Divide the target impact area into core sensitive areas and extended sensitive areas, forming a multi-level sensitive area set.

[0133] Core sensitive areas are those within the target impact area that are most severely affected by meteorological disasters and most sensitive to weather changes. Extended sensitive areas are areas surrounding the core sensitive areas that are less affected by meteorological disasters but still require attention. The multi-level sensitive area cluster facilitates more targeted deployment of monitoring equipment, prioritizing the monitoring needs of core sensitive areas.

[0134] Step S1533: Calculate the overlap between the coverage area of ​​each existing monitoring device and the multi-level sensitive area set to generate an existing coverage sensitive overlapping area.

[0135] Based on the location and monitoring range of existing monitoring devices, the coverage area of ​​each existing monitoring device is calculated. Then, the coverage area of ​​each existing monitoring device is compared with the multi-level sensitive area set to calculate the overlap between them.

[0136] For example, for an existing monitoring device, the overlap between its coverage area and the core sensitive area and the extended sensitive area is calculated. The overlap between the coverage areas of all existing monitoring devices and the multi-level sensitive area set is summarized to generate the existing coverage sensitive overlap area. The existing coverage sensitive overlap area reflects the sensitive areas within the current target impact range that are already covered by existing monitoring devices.

[0137] Step S1534: determining uncovered sensitive areas in the multi-level sensitive area set that are not covered by the existing covered sensitive overlapping areas, and generating an initial set of candidate positions within the uncovered sensitive areas.

[0138] By comparing the multi-level sensitive area set with the existing covered sensitive overlap area, the portion of the multi-level sensitive area set not covered by the existing covered sensitive overlap area is identified as the uncovered sensitive area. The uncovered sensitive area is the area that requires additional monitoring equipment to focus on.

[0139] Within the uncovered sensitive area, an initial set of candidate locations is generated based on predefined rules. For example, the uncovered sensitive area can be divided into a grid, and the nodes in the grid are used as candidate locations to form the initial set of candidate locations. This initial set of candidate locations includes locations that may be suitable for deploying new monitoring equipment.

[0140] Step S1535: For each candidate position in the initial set of candidate positions, calculate the overlap ratio between its coverage area and the core sensitive area in the uncovered sensitive area and the overlap ratio with the extended sensitive area, and calculate the overlap ratio between its coverage area and the coverage areas of all existing monitoring devices.

[0141] For each candidate location in the initial set of candidate locations, its coverage area is calculated based on its location and the monitoring range of the monitoring equipment. The overlap ratio of this coverage area with the core sensitive area of ​​the uncovered sensitive areas, the overlap ratio with the extended sensitive area, and the overlap ratio with the coverage area of ​​all existing monitoring equipment are calculated.

[0142] The core sensitive area overlap ratio reflects the degree to which the candidate's monitoring equipment can cover the core sensitive area; the extended sensitive area overlap ratio reflects the degree to which the extended sensitive area is covered; and the overlap ratio with existing monitoring equipment reflects the degree to which the candidate's monitoring equipment overlaps with existing equipment. By calculating these ratios for each candidate in the initial set of candidate locations, we can quantify each candidate's performance in covering sensitive areas and avoiding overlap with existing equipment.

[0143] Step S1536: Apply the optimization algorithm to iteratively adjust the candidate locations, with the goal of maximizing the overlap ratio of the core sensitive area, the overlap ratio of the secondary extended sensitive area, and minimizing the overlap ratio with the existing coverage, and select the location with the highest comprehensive overlap ratio after optimization as the deployment location of the new monitoring equipment.

[0144] After obtaining the overlap ratios for each candidate location, an optimization algorithm is applied to iteratively adjust the candidate locations. The goal of the optimization algorithm is to find an optimal set of new monitoring device deployment locations that maximizes the overlap ratio in the core sensitive area, minimizes the overlap ratio in the extended sensitive area (i.e., covers as many extended sensitive areas as possible), and minimizes the overlap ratio with the existing coverage.

[0145] During the iteration process, candidate locations are continuously adjusted, and a comprehensive evaluation metric is calculated based on the overlap ratios of the current candidate location. The comprehensive evaluation metric can be a weighted combination of the overlap ratio of the core sensitive area, the overlap ratio of the extended sensitive area, and the overlap ratio with existing coverage. For example, the overlap ratio of the core sensitive area can be given a higher weight because ensuring monitoring of the core sensitive area is more important; the overlap ratio of the extended sensitive area can be given a moderate weight; and the overlap ratio with existing coverage can be given a negative weight to encourage reducing overlap with existing equipment.

[0146] After each iteration, the candidate locations with the best performance are selected based on the comprehensive evaluation indicators and further adjusted until the preset stopping conditions are met, such as reaching the maximum number of iterations or no significant improvement in the comprehensive evaluation indicators. Finally, the location with the highest comprehensive overlap ratio after optimization is selected as the deployment location for the new monitoring equipment.

[0147] Step S1537: Based on the newly added deployment locations and the existing monitoring equipment location distribution information, a monitoring equipment deployment plan including the geographic coordinates of all monitoring equipment is generated.

[0148] After determining the deployment locations for the new monitoring devices, integrate the geographic coordinates of the new locations with the location distribution information of existing monitoring devices. Organize and record the geographic coordinates of all monitoring devices (both existing and new) in a predefined format, and generate a monitoring device deployment plan that includes the geographic coordinates of all monitoring devices. This monitoring device deployment plan specifies the specific locations of all monitoring devices within the target area.

[0149] Step S154: Classify warning levels according to the magnitude of the potential occurrence probability value, and the warning levels include a regular warning level and an emergency warning level.

[0150] Warning levels are categorized based on the potential probability of occurrence in the disaster prediction results. A probability threshold is set as the boundary between the regular and emergency warning levels. When the potential probability of a meteorological disaster is less than the threshold, it is classified as a regular warning level; when the potential probability is greater than or equal to the threshold, it is classified as an emergency warning level.

[0151] The regular warning level indicates that there is a certain possibility of a meteorological disaster, but the risk is relatively low; the emergency warning level indicates that the possibility of a meteorological disaster is high and more urgent and effective response measures are required. By classifying warning levels, corresponding warning information push rules can be formulated according to different risk levels.

[0152] Step S155: setting corresponding push rules for different warning levels, wherein the push rules include a push object set, a push time interval, and a push information content template.

[0153] Set push notification rules for different warning levels. For regular warning levels, the push recipients can include general groups of meteorological concern, such as ordinary residents and agricultural practitioners. Push notifications can be sent at relatively long intervals, such as once a day or every few days, to avoid frequent interruptions. The push notification content template can be concise and clear, mainly including the type of meteorological disaster, its potential probability of occurrence, and basic prevention recommendations.

[0154] For emergency alert levels, the target audience for notifications needs to be expanded to include not only general meteorological concern groups but also relevant emergency management departments and rescue teams. The notification interval should be shortened, for example, to once an hour or even less, to ensure timely delivery of the latest meteorological disaster information. The notification content template should be more detailed and urgent, including not only the type of meteorological disaster and its potential probability, but also specific response measures and emergency contact information.

[0155] Step S156: Integrate the number of monitoring devices, deployment locations, warning levels, and push rules to generate a dynamic monitoring strategy including device deployment parameters and warning rule parameters.

[0156] The previously determined number of monitoring devices, deployment locations, warning levels, and push rules are integrated. The number of monitoring devices and deployment locations serve as device deployment parameters, while the warning levels and push rules serve as warning rule parameters. These warning rule parameters are combined according to the specified structure and format to generate a dynamic monitoring strategy that includes both the device deployment parameters and the warning rule parameters.

[0157] The dynamic monitoring strategy comprehensively considers the potential risks and response measures of meteorological disasters. It includes both the rational deployment of monitoring equipment to ensure effective monitoring of meteorological conditions in the target area and push rules based on different warning levels to promptly convey meteorological disaster information to relevant personnel. This dynamic monitoring strategy is fed back to the meteorological monitoring platform, which can then allocate monitoring resources based on the strategy, such as arranging the installation and commissioning of new monitoring equipment, and promptly push warning information according to the push rules, thus achieving dynamic monitoring and effective response to meteorological disasters.

[0158] Furthermore, the meteorological disaster prediction model is trained through the following steps.

[0159] Step S211: Acquire the regional correlation feature samples corresponding to each historical meteorological data unit and the disaster real label data of each historical meteorological data unit, wherein the disaster real label data includes the type of meteorological disaster that actually occurred, the probability of occurrence and the boundary coordinates of the affected range.

[0160] To train a meteorological disaster prediction model, we first need to obtain historical meteorological data. This data is collected for the target area over a period of time. This data consists of multiple historical meteorological data units, each of which contains meteorological status information at a specific timestamp.

[0161] These historical meteorological data units are processed similarly to the real-time meteorological data previously described, including spatiotemporal feature extraction and regional correlation feature generation, yielding regional correlation feature samples corresponding to each historical meteorological data unit. These regional correlation feature samples contain information about the relationship between the temporal evolution patterns and spatial distribution patterns of the historical meteorological data.

[0162] At the same time, we collect real-world disaster label data corresponding to each historical meteorological data unit. By consulting historical meteorological records, disaster reports, and other information, we determine the type, probability, and boundary coordinates of the actual meteorological disaster corresponding to each historical meteorological data unit. This real-world disaster label data serves as supervisory information for model training, guiding the meteorological disaster prediction model in learning the relationship between meteorological data and meteorological disasters.

[0163] Step S212: The regional associated feature samples and the corresponding real disaster label data are combined into training sample pairs to construct a model training data set containing multiple training sample pairs.

[0164] Each regional feature sample is paired with the corresponding real-world disaster label data to form a training sample pair. Each training sample pair contains a regional feature sample and its corresponding real-world disaster label data, reflecting the correspondence between the features in historical meteorological data and the actual meteorological disasters that occurred.

[0165] Multiple training sample pairs are integrated to construct a model training dataset. This dataset is the foundation of model training and contains a wealth of historical meteorological data and corresponding disaster information. This provides sufficient learning material for the model, enabling it to learn the complex patterns and regularities between meteorological data and meteorological disasters.

[0166] Step S213: Initialize the network parameters of the meteorological disaster prediction model, which include the weight matrix of the input layer, the nonlinear transformation parameters of the hidden layer, and the probability calculation bias of the prediction layer. The prediction layer includes a probability prediction layer and a range prediction layer.

[0167] Before training the meteorological disaster prediction model, the network parameters must be initialized. The weight matrix in the input layer is used to linearly transform the input regional correlation feature samples, mapping them to intermediate feature vectors. The initial values ​​of the weight matrix elements can be randomly initialized, such as using a Gaussian distribution to randomly generate weight values.

[0168] The nonlinear transformation parameters of the hidden layer include the parameters of the activation function. The activation function is used to introduce nonlinear factors and enhance the expressiveness of the model. The initial values ​​of the nonlinear transformation parameters can also be randomly initialized.

[0169] The probability calculation bias of the prediction layer is used to adjust the calculation results of the probability prediction layer and the range prediction layer. The initial value of the probability calculation bias can be set to zero or a small random value. By initializing these network parameters, an initial state is provided for model training, allowing the meteorological disaster prediction model to gradually learn and optimize these parameters during subsequent training.

[0170] Step S214: inputting the regional correlation feature samples in the model training data set into the initialized meteorological disaster prediction model, performing dimension adaptation processing on the regional correlation feature samples through the input layer, and generating intermediate feature vector samples.

[0171] The regional correlation feature samples from the model training dataset are sequentially fed into the input layer of the initialized meteorological disaster prediction model. After receiving the regional correlation feature samples, the input layer undergoes a linear transformation using the input layer's weight matrix. This linear transformation involves performing matrix multiplication on the regional correlation feature samples and the weight matrix to produce intermediate feature vector samples with model-adaptive dimensions.

[0172] Step S215: performing nonlinear transformation processing on the intermediate feature vector samples through the hidden layer to extract deep semantic feature samples containing spatiotemporal influence relationships.

[0173] The intermediate feature vector samples enter the hidden layer of the meteorological disaster prediction model. In the hidden layer, the intermediate feature vector samples are nonlinearly transformed using an activation function. The activation function nonlinearly maps the elements of the intermediate feature vector samples, enabling the model to learn more complex and in-depth features.

[0174] The multi-layered structure of hidden layers allows for the extraction of progressively higher-level features. Each layer of neurons processes and transforms the input features, capturing the spatiotemporal influence relationships contained in intermediate feature vector samples and extracting deep semantic feature samples that incorporate these relationships. These deep semantic feature samples capture more subtle and complex patterns in meteorological data and are crucial for accurately predicting meteorological disasters.

[0175] Step S216: Call the prediction layer to perform probability distribution calculation processing on the deep semantic feature samples to generate a potential probability prediction value of the disaster type, and at the same time perform spatial range inference processing on the deep semantic feature samples to generate a coordinate prediction value of the impact range boundary.

[0176] The deep semantic feature samples enter the prediction layer, which consists of a probability prediction layer and a range prediction layer. The probability prediction layer calculates the probability distribution of the deep semantic feature samples. Through calculation and mapping, the deep semantic feature samples are converted into predicted values ​​for the potential occurrence probability of different meteorological disaster types. The probability prediction layer can implement this calculation process using common classification algorithms or probability distribution models in related technologies.

[0177] The range prediction layer infers the spatial range of samples from deep semantic features. Similar to the range prediction layer in the previous real-time application, it generates coordinate predictions for the impact range boundaries through a spatial mapping sublayer, threshold segmentation, morphological processing, and polygon fitting. These predictions represent the potential impact range boundaries predicted by the model for different meteorological disaster types.

[0178] Step S217: performing a difference calculation between the potential occurrence probability prediction value of the disaster type and the occurrence probability of the actual meteorological disaster type in the corresponding disaster real label data to generate a probability prediction loss value.

[0179] The potential probability predictions for disaster types generated by the probability prediction layer are compared with the actual probability of meteorological disaster types in the real-world disaster label data. A loss function is used to calculate the difference between the two. Common loss functions include cross-entropy loss.

[0180] The result calculated using the loss function is the probabilistic predicted loss value. This value reflects the degree of error in the model's prediction of the probability of a meteorological disaster. The smaller the loss value, the closer the model's prediction is to the actual situation.

[0181] Step S218: Calculate the spatial position difference between the predicted coordinate value of the impact range boundary and the impact range boundary coordinate in the corresponding disaster real label data to generate a range prediction loss value.

[0182] The spatial position difference between the predicted coordinate values ​​of the impact range boundary generated by the range prediction layer and the coordinates of the impact range boundary in the actual disaster label data is calculated. For example, some distance measurement methods such as Euclidean distance and Hausdorff distance can be used to calculate the difference between the coordinate prediction value and the actual boundary coordinate.

[0183] The range-based loss prediction value is calculated by calculating the difference between the boundary coordinates of the impact ranges of all meteorological disaster types, summarizing and averaging the results. The range-based loss prediction value reflects the degree of error in the meteorological disaster prediction model in predicting the impact range of the meteorological disaster. The smaller the loss value, the closer the model's predicted range is to the actual impact range.

[0184] Step S219: The probability prediction loss value and the range prediction loss value are normalized and then weighted summed to generate a total model loss value.

[0185] In order to comprehensively consider the errors of probability prediction and range prediction, the probability prediction loss value and the range prediction loss value are normalized. The purpose of normalization is to eliminate the dimensional difference between the two loss values ​​and make them fall within the same scale range.

[0186] The standardized probabilistic and range-based loss predictions are weighted and summed. Different weights are assigned to the probabilistic and range-based loss predictions, respectively. The weights are determined based on actual needs and the model's focus. For example, if the accuracy of the predicted probability of a meteorological disaster is more important, a higher weight can be assigned to the probabilistic loss prediction; if the accuracy of the predicted impact range is more important, a higher weight can be assigned to the range-based loss prediction. The weighted probabilistic and range-based loss predictions are added together to obtain the model's total loss value. The model's total loss value comprehensively reflects the model's overall error in both probability and range predictions.

[0187] Step S220: Based on the total loss value of the model, a gradient descent optimization algorithm is used to adjust the weight matrix of the input layer, the nonlinear transformation parameters of the hidden layer, and the prediction bias of the prediction layer, and the above steps are repeated until the total loss value of the model meets the preset convergence condition.

[0188] Based on the total loss value of the model, the gradient descent optimization algorithm is used to adjust the network parameters of the model. The core idea of ​​the gradient descent optimization algorithm is to update the network parameters along the negative gradient direction of the loss function to gradually reduce the total loss value of the model.

[0189] Calculate the gradient of the model's total loss with respect to the weight matrix of the input layer, the nonlinear transformation parameters of the hidden layer, and the prediction bias of the prediction layer. The gradient represents the rate of change of the loss function under the current parameter values. Updating the parameters along the negative gradient direction can reduce the value of the loss function.

[0190] Based on the calculated gradient, the network parameters are updated according to the set learning rate. The learning rate controls the step size of each parameter update. Too large a learning rate may cause the model to fail to converge, while too small a learning rate may result in slow convergence.

[0191] Repeat steps S214-S219, i.e., continuously input regional correlation feature samples into the meteorological disaster prediction model, perform feature extraction, prediction, loss calculation, and parameter adjustment, until the total loss value of the model meets the preset convergence condition. The preset convergence condition can be that the total loss value of the model is less than a certain threshold or the loss value does not decrease significantly within a set number of iterations.

[0192] Step S221: Use a model verification data set containing historical meteorological data units that have not participated in the training to perform performance verification on the trained meteorological disaster prediction model to determine whether the prediction accuracy of the meteorological disaster prediction model reaches a preset threshold.

[0193] After training is complete, the performance of the meteorological disaster prediction model needs to be verified. A model validation dataset containing historical meteorological data units that were not used in training is prepared. Regional correlation feature samples from the validation dataset are input into the trained meteorological disaster prediction model. Prediction is performed following the same steps as the training process to obtain predicted values ​​for the potential occurrence probability of meteorological disaster types and the coordinates of the impact area boundaries.

[0194] The prediction results are compared with the corresponding real-world labeled data of disasters in the model validation dataset, and prediction accuracy indicators such as precision, recall, and F1 value are calculated. These prediction accuracy indicators comprehensively reflect the accuracy of the meteorological disaster prediction model in predicting the probability and impact range of meteorological disasters.

[0195] The calculated prediction accuracy index is compared with a preset threshold to determine whether the model's prediction accuracy meets the requirements. If the prediction accuracy reaches the preset threshold, it indicates that the model has good performance and can be used for actual meteorological disaster prediction. If it does not reach the preset threshold, further adjustment and optimization of the model is required.

[0196] Step S222: If the prediction accuracy reaches a preset threshold, the trained meteorological disaster prediction model is used as the pre-trained meteorological disaster prediction model.

[0197] When the model's prediction accuracy reaches a preset threshold, the trained meteorological disaster prediction model is designated as the pre-trained meteorological disaster prediction model. This meteorological disaster prediction model, trained and validated with extensive historical data, demonstrates good performance and reliability, making it suitable for dynamic meteorological disaster monitoring using real-time meteorological data. In subsequent meteorological disaster monitoring, regional correlation features generated by real-time meteorological data are input into the meteorological disaster prediction model to generate accurate disaster prediction results.

[0198] Figure 2 A schematic diagram illustrates exemplary hardware and software components of a system 100 for dynamically monitoring meteorological disasters for real-time meteorological data, provided in some embodiments of the present application, that can implement the concepts of the present application. For example, a processor 120 can be used in the system 100 for dynamically monitoring meteorological disasters for real-time meteorological data and perform the functions described in the present application.

[0199] The meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data can be a general-purpose server or a special-purpose server, both of which can be used to implement the meteorological disaster dynamic monitoring method applied to real-time meteorological data of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0200] For example, the meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data also includes an I / O interface 150 between the computer and other input and output devices.

[0201] For ease of explanation, only one processor is described in the meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data. However, it should be noted that the meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the meteorological disaster dynamic monitoring system 100 applied to real-time meteorological data executes step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0202] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned method for dynamic monitoring of meteorological disasters applied to real-time meteorological data is implemented.

[0203] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for dynamic monitoring of meteorological disasters using real-time meteorological data, characterized in that: The method comprises: Acquire a real-time meteorological data set of meteorological data units containing meteorological status information of a target area at multiple different time stamps; Performing spatiotemporal feature extraction processing on the real-time meteorological data set to generate temporal variation features and spatial distribution features of each meteorological data unit, wherein the temporal variation features reflect the evolution law of basic meteorological parameters at adjacent timestamps, and the spatial distribution features reflect the distribution pattern of basic meteorological parameters at different locations at the same timestamp; Generate a regional correlation feature based on the correlation between the time-varying feature and the spatial distribution feature, wherein the regional correlation feature includes a description of the influence of the time evolution law on the spatial distribution pattern, wherein the regional correlation feature is generated by extracting the time key features from the time-varying feature and the spatial key areas from the spatial distribution feature, establishing a timestamp mapping, and then calculating the Pearson correlation coefficient to construct a correlation relationship description vector, and fusing it with the time key features and the spatial key areas to generate the regional correlation feature; Calling a pre-trained meteorological disaster prediction model, inputting the regional correlation features into the meteorological disaster prediction model to perform disaster risk assessment processing, and generating a disaster prediction result indicating the potential probability of occurrence and impact range of meteorological disasters in the target area; A dynamic monitoring strategy including a monitoring equipment deployment plan and early warning information push rules is generated based on the disaster prediction results, and the dynamic monitoring strategy is fed back to the meteorological monitoring platform to trigger a monitoring resource allocation operation.

2. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 1, characterized in that: The performing of spatiotemporal feature extraction processing on the real-time meteorological data set to generate temporal variation features and spatial distribution features of each meteorological data unit includes: Performing time dimension alignment processing on the real-time meteorological data set to obtain a time-aligned meteorological data sequence; Performing time series analysis on the time-aligned meteorological data sequence, calculating a change rate parameter and a change direction parameter of each basic meteorological parameter at consecutive time stamps, and constructing a time evolution curve as a time change feature based on the change rate parameter and the change direction parameter; Performing spatial gridding processing on the real-time meteorological data set, dividing the target area into evenly distributed spatial grid cells, and extracting mean statistics and variance statistics of basic meteorological parameters at the same time stamp in each spatial grid cell; Constructing a spatial distribution heat map as a spatial distribution feature based on the mean statistic and the variance statistic, wherein the color gradient of the spatial distribution heat map corresponds to the distribution density of the basic meteorological parameters; The temporal variation feature and the spatial distribution feature are normalized to obtain a spatiotemporal feature set.

3. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 2, wherein: The time series analysis processing is performed on the time-aligned meteorological data sequence, the change rate parameter and the change direction parameter of each basic meteorological parameter at continuous time stamps are calculated, and a time evolution curve is constructed as a time change feature based on the change rate parameter and the change direction parameter, including: Extracting basic meteorological parameter values ​​of adjacent time stamps in the time-aligned meteorological data sequence, and calculating the difference between the parameter values ​​of adjacent time stamps as the absolute change; Calculate the rate of change parameter per unit time according to the absolute change amount and the time interval length; Determine a change direction parameter by comparing the magnitude relationship of adjacent timestamp parameter values, wherein the change direction parameter includes an upward trend indicator and a downward trend indicator; The change rate parameters and change direction parameters of continuous timestamps are subjected to sliding window smoothing, and the smoothed change rate parameters and change direction parameters are arranged in the order of timestamps to generate a time evolution curve reflecting the evolution law of basic meteorological parameters over time. The abscissa of the time evolution curve is the timestamp, the ordinate is the change rate parameter, and the slope sign of the curve is determined by the change direction parameter.

4. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 2, wherein: The spatial location gridding processing is performed on the real-time meteorological data set, the target area is divided into evenly distributed spatial grid units, and the mean statistics and variance statistics of the basic meteorological parameters at the same time stamp in each spatial grid unit are extracted, including: Determine the number of horizontal and vertical divisions of the spatial grid unit based on the geographic boundary range of the target area and the preset grid division accuracy; Performing equidistant grid division processing on the target area based on the number of horizontal divisions and the number of vertical divisions to generate a set of spatial grid units covering the target area; The basic meteorological parameter values ​​at the same time stamp in each spatial grid cell are collected, and the arithmetic mean of all basic meteorological parameter values ​​in the grid cell is calculated as the mean statistic; Calculate the average of the squares of the differences between the basic meteorological parameter values ​​and the mean statistic in the grid cell as the variance statistic, where the variance statistic reflects the degree of dispersion of the basic meteorological parameter in the grid cell; The mean statistics and variance statistics are arranged according to the position coordinates of the spatial grid units to generate a statistics matrix with spatial position indexes.

5. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 1, wherein: Generating the regional correlation feature according to the correlation between the temporal variation feature and the spatial distribution feature includes: Extracting a time evolution curve from the time variation feature and a spatial distribution heat map from the spatial distribution feature; Performing feature sampling processing on the time evolution curve to extract the change rate parameters and change direction parameters of key time nodes as time key features, and performing region segmentation processing on the spatial distribution heat map to identify the boundary coordinates of concentrated distribution areas and discrete distribution areas as spatial key areas; Establish a mapping relationship between the timestamp of the time-critical feature and the acquisition timestamp of the spatial key area to make the time basis of the time dimension consistent with the time basis of the spatial dimension; For each time-critical feature, the change rate parameter and the change direction parameter are counted, and the mean statistical change of the spatial key area at the corresponding time stamp is calculated. The Pearson correlation coefficient of the change rate parameter and the mean statistical change is calculated. The absolute value of the Pearson correlation coefficient reflects the impact intensity, and the sign of the correlation coefficient reflects the impact direction. Perform trend analysis on the Pearson correlation coefficient of key features of continuous time to identify stable and fluctuating influence relationships. Extract the correlation coefficient of the stable influence relationship as the basic influence intensity, and extract the variation range of the correlation coefficient of the fluctuating influence relationship as the influence intensity fluctuation range. Comprehensively generate a description of the influence direction and intensity of the time evolution law on the spatial distribution pattern. An association relationship description vector is constructed based on the impact direction and impact intensity, and the association relationship description vector is fused with the time key feature and the space key area to generate a regional association feature containing a spatiotemporal impact relationship.

6. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 1, wherein: The calling of a pre-trained meteorological disaster prediction model, inputting the regional correlation features into the meteorological disaster prediction model for disaster risk assessment processing, and generating a disaster prediction result indicating the potential probability of occurrence and impact range of meteorological disasters in the target area, includes: Standardizing the regional correlation features, inputting the standardized regional correlation features into the input layer of the meteorological disaster prediction model, and performing feature recombination processing on the regional correlation features through a fully connected network to generate an intermediate feature vector with a model adaptation dimension; Performing nonlinear transformation processing on the intermediate feature vector through the hidden layer of the meteorological disaster prediction model to extract deep semantic features including spatiotemporal impact relationships; Calling the probability prediction layer of the meteorological disaster prediction model to perform probability distribution calculation processing on the deep semantic features to generate potential occurrence probability values ​​of different meteorological disaster types in the target area; Performing spatial range inference processing on the deep semantic features through the range prediction layer of the meteorological disaster prediction model to generate boundary coordinates of the impact range corresponding to different meteorological disaster types; The potential occurrence probability value is associated with the boundary coordinates of the impact range and matched to generate a disaster prediction result including the disaster type, potential occurrence probability and boundary coordinates of the impact range.

7. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 6, characterized in that: The spatial range inference processing of the deep semantic features by the range prediction layer of the meteorological disaster prediction model to generate the boundary coordinates of the impact range corresponding to different meteorological disaster types includes: Inputting the deep semantic features into the spatial mapping sublayer of the range prediction layer, performing spatial position feature enhancement processing on the deep semantic features through a convolutional neural network to generate a feature map with spatial position information; Performing threshold segmentation processing on the feature map, and extracting areas where the feature values ​​exceed a preset threshold as candidate disaster-affected areas; Performing morphological processing on the candidate disaster-affected areas, eliminating isolated small areas and connecting adjacent areas to generate continuous disaster-affected connected areas; Extracting the coordinates of the outer contour boundary points of the disaster-affected connected area, performing polygon fitting processing on the boundary point coordinates, and generating a polygon boundary coordinate sequence that approximately represents the affected range; The area parameter and the center coordinate parameter of the influence range are calculated according to the polygon boundary coordinate sequence, and the polygon boundary coordinate sequence, the area parameter and the center coordinate parameter are used as description information of the boundary coordinate of the influence range.

8. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 1, wherein: The generating of a dynamic monitoring strategy including a monitoring equipment deployment plan and early warning information push rules based on the disaster prediction results includes: Analyze the potential occurrence probability value and the boundary coordinates of the impact range in the disaster prediction result, and extract the target disaster type and its target impact range corresponding to the potential occurrence probability greater than the preset probability threshold; Determine the scope of the target area to be monitored based on the boundary coordinates of the target impact range, and calculate the number of monitoring devices that need to be deployed based on the area size and geographical distribution density of the target area; Based on the location distribution information of existing monitoring equipment in the target area, the deployment location of the new monitoring equipment is determined through an optimization algorithm so that the coverage of the new monitoring equipment maximizes overlap with the target influence range; According to the magnitude of the potential probability value, the warning level is divided into a regular warning level and an emergency warning level; Set corresponding push rules for different warning levels, including push object set, push time interval and push information content template; The number of monitoring devices, deployment locations, warning levels, and push rules are integrated and processed to generate a dynamic monitoring strategy including device deployment parameters and warning rule parameters.

9. The method for dynamic monitoring of meteorological disasters using real-time meteorological data according to claim 1, wherein: The meteorological disaster prediction model is trained by the following steps: Obtaining regional correlation feature samples corresponding to each historical meteorological data unit and real disaster label data of each historical meteorological data unit, wherein the real disaster label data includes the type of meteorological disaster that actually occurred, the probability of occurrence, and the boundary coordinates of the affected range; The regional correlation feature samples and the corresponding disaster real label data are combined into training sample pairs, and a model training data set containing multiple training sample pairs is constructed; Initializing network parameters of the meteorological disaster prediction model, the network parameters including the weight matrix of the input layer, the nonlinear transformation parameters of the hidden layer, and the probability calculation bias of the prediction layer, the prediction layer including the probability prediction layer and the range prediction layer; Inputting the regional correlation feature samples in the model training data set into the initialized meteorological disaster prediction model, performing dimension adaptation processing on the regional correlation feature samples through the input layer to generate intermediate feature vector samples; Performing nonlinear transformation processing on the intermediate feature vector samples through the hidden layer to extract deep semantic feature samples containing spatiotemporal influence relationships; Calling the prediction layer to perform probability distribution calculation processing on the deep semantic feature samples to generate a potential probability prediction value of the disaster type, and at the same time performing spatial range inference processing on the deep semantic feature samples to generate a coordinate prediction value of the impact range boundary; Calculate the difference between the potential probability prediction value of the disaster type and the probability of the actual meteorological disaster type in the corresponding disaster real label data to generate a probabilistic predicted loss value; Calculate the spatial position difference between the predicted coordinate value of the impact range boundary and the coordinate of the impact range boundary in the corresponding disaster real label data to generate a range prediction loss value; The probability prediction loss value and the range prediction loss value are normalized and then weighted summed to generate a total model loss value; Based on the total loss value of the model, a gradient descent optimization algorithm is used to adjust the weight matrix of the input layer, the nonlinear transformation parameters of the hidden layer, and the prediction bias of the prediction layer, and the above steps are repeated until the total loss value of the model meets the preset convergence condition; Using a model validation dataset containing historical meteorological data units that did not participate in the training to perform performance verification on the trained meteorological disaster prediction model to determine whether the prediction accuracy of the meteorological disaster prediction model reaches a preset threshold; If the prediction accuracy reaches a preset threshold, the trained meteorological disaster prediction model is used as the pre-trained meteorological disaster prediction model.

10. A meteorological disaster dynamic monitoring system applied to real-time meteorological data, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the dynamic monitoring method of meteorological disasters applied to real-time meteorological data as described in any one of claims 1 to 9.

Images