Meteorological disaster risk estimation method and system based on historical observation and forecast fusion

Abstract

The invention discloses a historical observation and forecast fusion-based meteorological disaster risk estimation method and system, and belongs to the field of meteorological disaster risk assessment, and the method specifically comprises the steps: obtaining long-time series rainstorm observation data and rainstorm forecast data of a target region through a dual-thread data collection channel, and carrying out the preprocessing of the data, and then carrying out the prediction of the rainstorm observation data and the rainstorm forecast data; constructing a rainstorm disaster initial estimation model by using a machine learning algorithm; during a medium-and-long-term forecast effective period, according to a meteorological monitoring network layout and disaster sensitivity, setting a time interval, collecting and comparing rainstorm observation data and forecast data in real time, and once a deviation is found, immediately starting an intelligent fusion and dynamic correction algorithm to correct a pre-estimation result; and at the end period of forecasting, performing multi-dimensional visual output on the corrected rainstorm disaster estimation result, and generating a text report, thereby realizing dynamic optimization.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological disaster risk assessment, and specifically relates to a method and system for predicting meteorological disaster risks based on the integration of historical observations and forecasts. Background Art

[0002] In today's society, meteorological disasters occur frequently. Among them, rainstorm disasters, due to their suddenness and strong destructiveness, have had an extremely serious impact on human production and life. Most of the previous methods for predicting rainstorm disasters have obvious shortcomings. On the one hand, traditional predictions overly rely on a single meteorological forecasting model. However, the meteorological system is inherently unpredictable, and many complex factors are intertwined, resulting in forecast results often deviating from the actual situation and having a high degree of uncertainty. On the other hand, simply relying on the forecast information in the medium and long term completely ignores the development trajectory and evolution law of disasters under similar meteorological conditions in history, and cannot comprehensively and accurately reflect the true trend of disasters, resulting in limited accuracy and reliability of the assessment results.

[0003] For example, Chinese Patent with Publication No. CN103177301A discloses a method for predicting typhoon disaster risks, including: statistically analyzing the loss data caused by typhoon disasters in a specified monitoring area, selecting hazard of disaster-causing factors, sensitivity of disaster-bearing environment, vulnerability of disaster-affecting bodies, and disaster prevention and mitigation capabilities as the typhoon disaster risk assessment index system, establishing a typhoon disaster risk prediction model using the fuzzy transformation theory, taking the typhoon forecast result as the starting condition and input condition of the prediction model, and through the calculation and analysis of the prediction model, obtaining whether the area to be predicted will be disaster-stricken and the disaster risk level of the disaster in the future, so as to improve the early warning ability of meteorological disasters.

[0004] For example, Chinese Patent with Authorization Announcement No. CN118365147B discloses a method for predicting rainstorm and flood disaster risks based on the correction of underlying surface characteristics, including: using intelligent grid precipitation forecast data to calculate the intensity index of the rainstorm process expected to occur, based on the influence coefficient of the disaster-bearing environment of underlying surface characteristics such as terrain, river network system, and soil geology conditions, and combining the exposure and vulnerability information of disaster-affecting bodies, conducting risk prediction on high-resolution disaster-affecting bodies such as population, gross domestic product (GDP), and rice. This technical solution can effectively make up for the service limitations of relying solely on meteorological element forecasts, enrich meteorological service information such as the possibility of disaster occurrence and the degree of harm of rainstorm and flood disaster weather processes in different time periods, industries, and regions, provide a reference for the prevention and decision-making of rainstorm and flood disasters, and effectively play the role of the first line of defense in meteorological disaster prevention and mitigation.

[0005] The above existing technologies all have the following problems: relying on the accuracy of intelligent grid precipitation forecast data, if there are errors in the precipitation forecast data, the calculated intensity index of the rainstorm process may also be inaccurate, thus affecting the subsequent risk assessment results; the applicability is limited; and the dynamic correction ability is insufficient. Summary of the Invention

[0006] In view of the deficiencies of the existing technologies, the present invention proposes a method and system for meteorological disaster risk assessment based on the fusion of historical observations and forecasts. A dual-thread data collection channel is used to obtain long-time series rainstorm observation data and rainstorm forecast data of the target area. After preprocessing, a machine learning algorithm is used to construct an initial rainstorm disaster assessment model; during the effective period of medium- and long-term forecasts, according to the layout of the meteorological monitoring network and the sensitivity of disaster evolution, a time interval is set. The monitoring stations distributed in each grid unit are driven to collect the latest rainstorm observation data in real time and compare the newly collected rainstorm observation data with the medium- and long-term forecast data on which the initial assessment is based point by point to obtain the deviation of the corresponding indicators. At the same time, the deviation situation is fed back to the initial rainstorm disaster assessment model; at the end period of the forecast, the corrected rainstorm disaster assessment result is output in multiple dimensions and a text report is generated, realizing dynamic optimization.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for meteorological disaster risk assessment based on the fusion of historical observations and forecasts, including:

[0009] Step S1: Use a dual-thread data collection channel to collect long-time series historical rainstorm observation data of the target area and rainstorm forecast data issued by the meteorological department for the target area, and perform preprocessing;

[0010] Step S2: Based on the preprocessed historical rainstorm observation data, use a machine learning algorithm to construct an initial rainstorm disaster assessment model, and input the information corresponding to the model input variables in the real-time obtained rainstorm forecast data into the initial rainstorm disaster assessment model to obtain the initial assessment result of the rainstorm disaster;

[0011] Step S3: During the entire effective period of the medium- and long-term forecast, set a time interval according to the density of the meteorological monitoring network layout in the target area and the sensitivity of disaster evolution. In each time interval, drive the monitoring stations distributed in each grid unit to re-collect the latest rainstorm observation data, and compare the newly collected rainstorm observation data with the medium- and long-term forecast data on which the initial assessment is based point by point to obtain the deviation of the corresponding indicators. At the same time, feedback the deviation situation to the initial rainstorm disaster assessment model;

[0012] Step S4: Monitor the deviation situation in real time. Once it is found that the actual situation deviates from the forecast, immediately start the preset intelligent fusion and dynamic correction algorithm to correct the initial assessment result;

[0013] Step S5: After cyclic dynamic correction, at the end of the medium- and long-term forecast period, the corrected heavy rain disaster prediction results obtained are output in multi-dimensional visualization, and a text report including data indicators is generated at the same time. The text report includes the prediction results of heavy rain disasters, the correction process, deviation conditions, and possible disaster impact information.

[0014] In step S1, the heavy rain observation data and heavy rain forecast data include heavy rain events and their process data at different time scales; the time scales include months, seasons, and years.

[0015] During the construction of the initial heavy rain disaster prediction model, it is necessary to meet:

[0016] For predictions within 30 days, monthly-scale forecast data is used. The monthly-scale forecast data includes forecast data for the next ten days and S2S model forecast data.

[0017] Specifically, the specific steps of step S2 include:

[0018] S2.1: Obtain the preprocessed historical heavy rain observation data and heavy rain forecast data;

[0019] S2.2: Load the pre-constructed LSTM neural network model architecture, and use the preprocessed historical heavy rain observation data as input to train the pre-constructed LSTM neural network model to generate a trained initial heavy rain disaster prediction model;

[0020] S2.3: Load a linear regression model as the base model, and use all the features of the preprocessed heavy rain forecast data as the initial feature set X = {x 1 ,..., x n}, where x i , i ∈ [1, n] represents the i-th initial feature in the preprocessed heavy rain forecast data, and n represents the number of initial features;

[0021] S2.4: Input the initial feature set into the base model for prediction, calculate the feature importance Imp i according to the absolute value of the generated base model coefficient β i , and remove the initial features corresponding to the absolute values of the M smallest feature importances. i represents the base model coefficient index value;

[0022] S2.5: Repeat S2.4 until the predetermined number of features is reached. At the same time, during the recursive elimination process, record the order in which each feature is retained;

[0023] S2.6: Sort the features according to the order in which they are retained during the recursive elimination process, and extract the screened and sorted features from the preprocessed heavy rain forecast data to generate medium- and long-term heavy rain forecast feature data;

[0024] S2.7: Input the medium- and long-term heavy rain forecast feature data into the trained initial heavy rain disaster prediction model. The initial heavy rain disaster prediction model makes predictions based on the input feature data and outputs the initial prediction results of the heavy rain disaster.

[0025] Specifically, the specific steps of step S3 include:

[0026] S3.1: Obtain the location information of meteorological monitoring stations in the target area. Divide the target area into N grids according to the size, shape, and meteorological characteristics of the target area. In each grid cell, select a monitoring station as a representative according to the station distribution and meteorological monitoring requirements, and use the formula to evaluate the density of monitoring stations in each grid cell. Here, f(z, s) represents the density estimate value at location z and under condition variable s, m represents the total number of monitoring stations in the current grid cell, b with (s) represents the bandwidth function, w j represents the importance weight of the j-th station, K(·) represents the kernel function, and z j represents the location of the j-th monitoring station;

[0027] S3.2: Set the time interval ΔT according to the sensitivity of disaster evolution and real-time meteorological conditions. In each time interval, use the monitoring station locations determined in step S3.1 to drive the monitoring stations distributed in each grid cell to re-collect the latest heavy rain observation data.

[0028] Specifically, the specific steps of step S3 further include:

[0029] S3.3: Organize the newly collected heavy rain observation data X′ = {x 1 ′,..., x′ u} and the medium- and long-term forecast data Y = {y 1 ,..., y u} used for the initial prediction. Here, x′ u represents the u-th newly collected heavy rain observation data, y u represents the u-th medium- and long-term forecast data, and u represents the number of newly collected heavy rain observation data or medium- and long-term forecast data;

[0030] S3.4: Set the time window Δt. If the geographical locations and timestamps of the newly collected heavy rain observation data and the medium- and long-term forecast data both fall within the time window Δt, it is considered a successful match;

[0031] S3.5: Traverse the rainstorm observation data. For each piece of rainstorm observation data, check whether there is a matching record in the medium- and long-term forecast data. If a matching record is found, perform association and save the matching result.

[0032] S3.6: Perform point-by-point comparison on the matched data, calculate the difference e between the newly collected rainstorm observation data and the medium- and long-term forecast data l , and analyze the reasons for the deviation, where l represents the data index value.

[0033] Specifically, the specific steps of step S4 include:

[0034] S4.1: Obtain the deviation value e between the newly collected rainstorm observation data and the medium- and long-term forecast data l , and set the deviation threshold h;

[0035] If e l ≤h, continue monitoring;

[0036] If e l >h, perform multi-dimensional feature extraction on the newly collected rainstorm observation data, and use an intelligent fusion algorithm to perform multi-source fusion on the extracted multi-dimensional features to obtain the fused feature set δ c , where δ c represents the c-th type of fused feature;

[0037] S4.2: Use the fused feature set to correct the initial estimation result of the rainstorm disaster, and provide the corrected estimation result to decision-makers and emergency management personnel in a visual manner.

[0038] Specifically, the model input variables in step S2 include meteorological elements, geographical and topographical feature parameters, and the cumulative effect index of previous rainfall in the same grid cell in the historical same period; the meteorological elements include the daily average and extreme values of temperature, humidity, air pressure, and wind direction and speed; the geographical and topographical feature parameters include altitude, slope, distance to water system, and soil permeability; the cumulative effect index of previous rainfall includes the cumulative rainfall and wet days in the previous week and the previous month.

[0039] Specifically, the visualized content in step S5 includes rainstorm disaster events and their processes at monthly, seasonal, and annual scales.

[0040] A meteorological disaster risk estimation system based on the fusion of historical observations and forecasts includes: a data collection module, a model construction module, a monitoring and comparison module, a dynamic correction module, and a report generation module;

[0041] The data collection module is used to collect long-term historical rainstorm observation data and rainstorm forecast data released by the meteorological department in the target area, and perform preprocessing;

[0042] The model construction module is used to construct an initial prediction model for rainstorm disasters based on the pre - processed historical rainstorm observation data by using machine learning algorithms, and perform initial prediction by using the rainstorm forecast data obtained in real - time;

[0043] The monitoring and comparison module is used to set a time interval according to the density of the meteorological monitoring network layout in the target area and the sensitivity of disaster evolution during the entire period when the medium - and long - term forecast is in effect, monitor and compare the newly collected rainstorm observation data with the medium - and long - term forecast data on which the initial prediction is based in real - time, obtain the deviation situation, and feedback the deviation to the prediction model;

[0044] The dynamic correction module is used to monitor the deviation situation in real - time. Once it is found that the actual situation deviates from the forecast, it immediately starts the preset intelligent fusion and dynamic correction algorithm to correct the initial prediction result;

[0045] The report generation module is used to perform multi - dimensional visual output of the corrected rainstorm disaster prediction result at the end period of the medium - and long - term forecast, and generate a text report containing data indicators.

[0046] Specifically, the monitoring and comparison module includes: a time - interval setting unit, a data collection unit, a data comparison unit, and a deviation feedback unit;

[0047] The time - interval setting unit is used to intelligently set the time interval according to the meteorological monitoring network layout and disaster evolution sensitivity;

[0048] The data collection unit is used to drive the monitoring stations to re - collect the latest rainstorm observation data within each time interval;

[0049] The data comparison unit is used to compare the newly collected data with the medium - and long - term forecast data point by point and calculate the deviation;

[0050] The deviation feedback unit is used to timely feedback the deviation situation to the initial prediction model of rainstorm disasters.

[0051] Specifically, the dynamic correction module includes: a deviation monitoring unit and a dynamic correction unit;

[0052] The deviation monitoring unit is used to monitor the deviation situation in the comparison result in real - time;

[0053] The dynamic correction unit is used to correct the initial prediction result by using the dynamic correction algorithm according to the deviation situation.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] 1. The present invention proposes a method for predicting meteorological disaster risks based on the integration of historical observations and forecasts. Through a dual-thread data collection channel, the system can efficiently integrate historical heavy rain observation data and medium- and long-term forecast data, providing a solid foundation for constructing an accurate initial prediction model for heavy rain disasters. At the same time, during the effective period of the medium- and long-term forecast, the system can flexibly set the time interval to re-collect data according to the layout of the actual monitoring network and disaster sensitivity, and compare it with the forecast data in real time, so as to timely detect and correct deviations.

[0056] 2. The present invention proposes a method for predicting meteorological disaster risks based on the integration of historical observations and forecasts. Once it is found that the actual situation deviates from the forecast, the system can immediately activate the intelligent integration and dynamic correction algorithm to adjust the initial prediction result in real time to ensure the accuracy and reliability of the prediction result. At the end of the medium- and long-term forecast, the system can also output the corrected heavy rain disaster prediction result in a multi-dimensional visualization manner and generate a text report containing detailed data indicators, providing intuitive and comprehensive disaster risk assessment information for decision-makers. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the method for predicting meteorological disaster risks based on the integration of historical observations and forecasts of the present invention;

[0058] Figure 2 Flow chart of the principle of the method for predicting meteorological disaster risks based on the integration of historical observations and forecasts of the present invention;

[0059] Figure 3 Flow chart of the implementation of the initial prediction result of heavy rain disaster of the method for predicting meteorological disaster risks based on the integration of historical observations and forecasts of the present invention;

[0060] Figure 4 System architecture diagram of the method for predicting meteorological disaster risks based on the integration of historical observations and forecasts of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] Example 1

[0062] Please refer to Figure 1 and Figure 2 , an embodiment provided by the present invention: A method for predicting meteorological disaster risks based on the integration of historical observations and forecasts, comprising the following steps:

[0063] Step S1: Use a dual-thread data collection channel to collect long-term historical heavy rain observation data of the target area and heavy rain forecast data issued by the meteorological department for the target area, and perform preprocessing;

[0064] The heavy rain observation data and heavy rain forecast data in step S1 include heavy rain events and their process data on different time scales; the time scales include months, seasons, and years.

[0065] During the construction of the initial prediction model for rainstorm disasters, the following requirements should be met:

[0066] For predictions within 30 days, monthly-scale forecast data is used. The monthly-scale forecast data includes forecast data for the next ten days and subseasonal to seasonal (S2S) model forecast data.

[0067] Furthermore, the specific steps of step S1 include:

[0068] (1) Data collection, including: 1) Long-time series rainstorm observation data: Usually sourced from observation devices such as meteorological observation stations, meteorological satellites, and radars, which record detailed data of multiple historical rainstorm events in the target area, including rainfall amount, rainfall intensity, rainfall duration, geographical location, topography, and meteorological conditions; 2) Rainstorm forecast data: Issued by the meteorological department, usually based on advanced numerical weather prediction models and real-time observation data, to predict rainstorm conditions in the future for a certain period.

[0069] (2) Set up two independent data collection channels, respectively used to collect long-time series rainstorm observation data and rainstorm forecast data, ensuring that the data collection processes of the two channels do not interfere with each other, so as to improve the efficiency and accuracy of data collection;

[0070] (3) Preprocess the collected data, including data cleaning, data format conversion, and data interpolation.

[0071] Step S2: Based on the preprocessed historical rainstorm observation data, use machine learning algorithms to construct an initial prediction model for rainstorm disasters, and input the information corresponding to the model input variables in the real-time obtained rainstorm forecast data into the initial prediction model for rainstorm disasters to obtain the initial prediction results of rainstorm disasters;

[0072] The model input variables in step S2 include meteorological elements, geographical and topographical characteristic parameters, and the cumulative effect index of previous rainfall in the same grid cell in the same period of history; the meteorological elements include the daily change means and extreme values of temperature, humidity, air pressure, wind direction and wind speed; the geographical and topographical characteristic parameters include altitude, slope, distance to water system, and soil permeability; the cumulative effect index of previous rainfall includes the cumulative rainfall amount and wet days in the previous week and the previous month.

[0073] Step S3: During the entire period when the medium- and long-term forecast is in effect, set the time interval according to the density of the meteorological monitoring network layout in the target area and the sensitivity of disaster evolution. During each time interval, drive the monitoring stations distributed in each grid cell to re-collect the latest rainstorm observation data, and compare the newly collected rainstorm observation data with the medium- and long-term forecast data based on the initial estimation point by point to obtain the deviation of the corresponding indicators. At the same time, feedback the deviation to the initial estimation model of rainstorm disasters;

[0074] Step S4: Monitor the deviation situation in real time. Once it is found that the actual situation deviates from the forecast, immediately start the preset intelligent fusion and dynamic correction algorithm to correct the initial estimation result;

[0075] Exemplarily, when positive deviation of rainfall is observed and the growth rate of waterlogging depth accelerates, quickly expand the boundary of the disaster risk area, raise the flood warning level, and dynamically adjust the priority of population evacuation; conversely, if the actual situation eases, reasonably shrink the risk area and lower the warning level to ensure that the estimation result is closely fitted to the reality and achieve dynamic optimization.

[0076] Step S5: After cyclic dynamic correction, at the end period of the medium- and long-term forecast, output the corrected rainstorm disaster estimation result in multi-dimensional visualization, and generate a text report containing data indicators at the same time.

[0077] The visualized content in Step S5 includes rainstorm disaster events and their processes at monthly, seasonal, and annual scales, specifically including high-risk disaster areas, flood warning areas at different levels, personnel evacuation route information, as well as rainstorm intensity and duration indicators.

[0078] Furthermore, the specific steps of Step S5 include:

[0079] (1) Utilize the real-time observation data and the initial estimation result of rainstorm disasters output by the initial estimation model of rainstorm disasters, combine historical data, geographical information, and meteorological conditions, and continuously correct the estimation result through an iterative manner until the end period of the medium- and long-term forecast to obtain the finally corrected rainstorm disaster estimation result;

[0080] (2) Select the visualization tool FineReport, input the finally corrected rainstorm disaster estimation result into the FineReport tool for multi-dimensional data display and analysis, and visually express information such as the intensity, scope, and change trend of rainstorm disasters through visual elements such as colors, sizes, and shapes. Among them, the display methods select visualization chart types such as maps and heat maps;

[0081] (3) Compile a text report containing key data indicators according to the estimation result and the result of visual analysis.

[0082] It should be understood that the visualization output of the final heavy rain disaster estimation results is selected at the end period of the medium- and long-term forecast, mainly based on the following considerations:

[0083] First of all, the accuracy requirement determines this period selection. The entire estimation system is built on the basis of the integration of historical data and medium- and long-term forecasts, and is continuously optimized through a dynamic monitoring and correction mechanism. During the effective period of the medium- and long-term forecast, the real-time collected observation data and forecast data are continuously compared and corrected. Only near the end of the forecast can all feedback information be utilized to the greatest extent, making the estimation results closest to the real disaster situation. If the output is too early, the subsequent new deviation situations cannot be taken into account, which will lead to inaccuracies in the visualized high-risk areas, flood warning areas, etc., and thus mislead decision-making and rescue operations;

[0084] Secondly, from the perspective of the coherence of decision-making and implementation, the output at the end period can seamlessly connect with subsequent response measures. The disaster prevention and mitigation system with multi-department linkage such as the emergency command center and traffic control needs to quickly deploy resources and arrange personnel evacuation based on accurate and final information. At this time, each department can just obtain comprehensive information at one time based on this complete and reliable visualized map and text report, directly allocate emergency rescue forces according to the marked high-risk disaster areas, prepare corresponding rescue materials according to different levels of flood warning areas, and organize evacuations in an orderly manner along the planned evacuation routes for personnel, avoiding implementation chaos caused by multiple changes in information;

[0085] Furthermore, the rationality of technical implementation and resource allocation prompts the selection of this period. Frequent visualization output and report generation, especially during the disaster evolution process, require a large amount of computing resources and data transmission bandwidth. By concentrating the key visualization output at the end of the medium- and long-term forecast, it can not only ensure the provision of complete results at critical moments, but also concentrate computing power in the early stage for data collection, model operation and dynamic correction, optimize the operation efficiency of the entire system, ensure that limited technical resources play the greatest effectiveness, support the smooth operation of the entire estimation system until the end of the forecast, and provide a solid guarantee for disaster prevention.

[0086] Exemplarily, the target area is set as a prosperous city by the river. First of all, the research team deeply explores the local meteorological database, systematically sorts out the detailed observation files of all heavy rain weather processes in the past few decades, and divides the urban area into multiple sub-areas such as the core business district with high-rise buildings, the riverside area with low-lying terrain, and the suburban mountainous area with dense vegetation according to the unique topography of the city, and conducts refined sorting of the data of each area. At the same time, it seamlessly connects with the meteorological department's advanced forecasting system to capture the dynamic heavy rain forecast of this city within the next 72 hours in real time.

[0087] With the help of the complete historical data sorted out in the early stage, carefully train the initial prediction model. Once the latest heavy rain forecast is received, quickly input key data such as the current real-time temperature, air humidity, and the expected rainfall in the next 36 hours into the model, and initially lock in potential high-incidence points of waterlogging and crisis sections such as the shorelines at risk of river water backflow in various urban areas. In the subsequent 72 hours, collect the measured rainfall values and real-time dynamics of wind direction and wind speed transmitted back by each meteorological station in the city every 45 minutes, closely compare the differences between the forecast and the reality, flexibly and precisely dynamically adjust the waterlogging risk warning range, and update the predicted water depth values in real time until the 72-hour forecast period expires. Finally, output a detailed and highly practical heavy rain disaster risk prediction report, and be equipped with a visualized geographic information map to provide rock-solid decision-making support for multiple fields such as urban emergency management and people's livelihood security, and maximize the protection of people's lives and property and reduce disaster losses.

[0088] Example 2

[0089] Please refer to Figure 3 , the specific steps of step S2 in this embodiment include:

[0090] S2.1: Obtain the preprocessed historical heavy rain observation data and heavy rain forecast data;

[0091] S2.2: Load the pre-constructed LSTM neural network model architecture, and use the preprocessed historical heavy rain observation data as input to train the pre-constructed LSTM neural network model to generate a trained initial heavy rain disaster prediction model;

[0092] Among them, LSTM is the Long Short-Term Memory (LSTM).

[0093] Further, the specific steps of S2.2 include:

[0094] (1) Environment preparation: Ensure that the necessary deep learning frameworks have been installed and the computing resources have been configured to accelerate model training;

[0095] (2) Load the defined LSTM neural network architecture from a file or model, usually including the number of LSTM layers, the number of neurons in each layer, and the dimensions of the input and output, and ensure that the model architecture matches the format and dimensions of the preprocessed historical heavy rain observation data;

[0096] (3) Divide the preprocessed historical heavy rain observation data into a training set and a validation set for model training and evaluation, and ensure that the data has been normalized or standardized to improve the training efficiency of the model;

[0097] (4) Set the loss function of the model, such as the mean squared error (MSE) loss, and select the Adam optimizer. Set parameters such as the learning rate. Among them, the mean squared error (MSE) loss and the Adam optimizer are prior art in this field and not the creative solutions of this application, so they will not be elaborated here;

[0098] (5) Input the training data into the LSTM model for forward propagation calculation. Calculate the loss value of the model according to the loss function, calculate the gradient through the backpropagation algorithm, and update the weights of the model. Among them, the forward propagation calculation formula and the backpropagation algorithm are prior art in this field and not the creative solutions of this application, so they will not be elaborated here;

[0099] (6) Repeat the above steps until the predetermined number of training epochs is reached or the loss value converges;

[0100] (7) Use the validation set to evaluate the performance of the model and calculate evaluation metrics such as accuracy, recall, and F1-score to understand the generalization ability of the model. Among them, the calculation formulas for accuracy, recall, and F1-score are prior art in this field and not the creative solutions of this application, so they will not be elaborated here;

[0101] (8) Save the trained model to a file for subsequent use.

[0102] S2.3: Load a linear regression model as the base model and use all the features of the preprocessed rainstorm forecast data as the initial feature set \(X = \{x 1 ,\ldots,x n \}\), where \(x i , i\in[1,n]\) represents the \(i\)-th initial feature in the preprocessed rainstorm forecast data, and \(n\) represents the number of initial features. Among them, the linear regression model is prior art in this field and not the creative solution of this application, so it will not be elaborated here;

[0103] S2.4: Input the initial feature set into the base model for prediction, and calculate the feature importance according to the absolute value of the generated base model coefficients and remove the initial features corresponding to the absolute values of the smallest \(M\) feature importances. Among them, \(y\) represents the target variable, \(Imp i represents the importance of the \(i\)-th initial feature, \(\beta 0 represents the intercept term, \(\beta i represents the \(i\)-th base model coefficient, \(\varepsilon\) represents the error term, and \(i\) represents the base model coefficient index value;

[0104] S2.5: Repeat S2.4 until the predetermined number of features is reached. At the same time, during the recursive elimination process, record the order in which each feature is retained;

[0105] S2.6: Sort the features according to the retention order of the features in the recursive elimination process, and extract the screened and sorted features from the preprocessed rainstorm forecast data to generate medium- and long-term rainstorm forecast feature data;

[0106] Further, the specific steps of S2.6 include:

[0107] (1) Obtain the retention order of each feature recorded in S2.5 and sort the features;

[0108] (2) Obtain the preprocessed rainstorm forecast data, and extract the sorted features from the preprocessed rainstorm forecast data. When extracting the sorted features, a feature selection algorithm is used. The feature selection algorithm is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here;

[0109] (3) Standardize and normalize the extracted feature data. The standardization and normalization processes are the prior art content in this field and are not the creative solution of this application, so it will not be elaborated here.

[0110] S2.7: Input the medium- and long-term rainstorm forecast feature data into the trained initial rainstorm disaster prediction model. The initial rainstorm disaster prediction model makes predictions based on the input feature data and outputs the initial prediction results of the rainstorm disaster.

[0111] The specific steps of step S3 include:

[0112] S3.1: Obtain the location information of the meteorological monitoring stations in the target area. Divide the target area into N grids according to the size, shape, and meteorological characteristics of the target area. In each grid cell, randomly select a monitoring station as a representative according to the station distribution and meteorological monitoring requirements, and use the formula to evaluate the density of the monitoring stations in each grid cell, where f(z, s) represents the density estimate value of the monitoring station at location z and under condition variable s, m represents the total number of monitoring stations in the current grid cell, b with (s) represents the bandwidth function, w j represents the importance weight of the j-th station, K(·) represents the kernel function, and z j represents the location of the j-th monitoring station;

[0113] In the present invention, by introducing the condition variable s and b with (s) can be dynamically adjusted according to the environmental conditions. Exemplarily, in areas with complex meteorological conditions or sparse data, the bandwidth can be increased to smooth the data, while in areas with dense data and relatively simple meteorological conditions, the bandwidth can be decreased to capture more fine structures.

[0114] Among them, the location information of the meteorological monitoring stations includes geographical coordinates such as longitude and latitude; the size and shape of the grid cells are usually selected as square or rectangular grids.

[0115] S3.2: Set the time interval ΔT according to the sensitivity of the disaster evolution and the real-time meteorological conditions, and within each time interval, use the positions of the monitoring stations determined in step S3.1 to drive the monitoring stations distributed in each grid cell to re-collect the latest heavy rain observation data;

[0116] S3.3: Organize the newly collected heavy rain observation data X′ = {x 1 ′,..., x′ u} and the medium- and long-term forecast data Y = {y 1 ,..., y u} on which the initial prediction is based, where x′ u represents the u-th newly collected heavy rain observation data, y u represents the u-th medium- and long-term forecast data, and u represents the number of newly collected heavy rain observation data or medium- and long-term forecast data;

[0117] S3.4: Set the time window Δt. If the geographical locations and timestamps of the newly collected heavy rain observation data and the medium- and long-term forecast data both fall within the time window Δt, it is considered a successful match;

[0118] S3.5: Traverse the heavy rain observation data. For each piece of heavy rain observation data, check whether there is a matching record in the medium- and long-term forecast data. If a matching record is found, perform an association and save the matching result;

[0119] Further, the specific steps of S3.5 include:

[0120] (1) Collect and organize the heavy rain observation data, which usually includes key information such as the time, location, and precipitation amount of the heavy rain. At the same time, collect and organize the medium- and long-term forecast data, which usually includes the forecast time period, location, precipitation prediction, etc.;

[0121] (2) Preprocess the heavy rain observation data and the medium- and long-term forecast data, including data cleaning, format unification, etc., to ensure the consistency and accuracy of the data;

[0122] (3) Use a loop or iterative method to traverse each piece of heavy rain observation data;

[0123] (4) For each piece of heavy rain observation data, check whether there is a matching record in the medium- and long-term forecast data, where the matching criterion is set as the time window Δt in S3.4;

[0124] (5) If a matching record is found, associate the rainstorm observation data with the medium- and long-term forecast data, record the matching result, and save the matching result to a database or file for subsequent analysis and use.

[0125] S3.6: Perform point-by-point comparison on the matched data, and calculate the difference e between the newly collected rainstorm observation data and the medium- and long-term forecast data l = x l ′ - y l , l ∈ [1, u], and analyze the reasons for the deviation. Among them, e l represents the deviation value of the l-th data.

[0126] Among them, the reasons for the deviation include observation errors and forecast model errors.

[0127] The specific steps of step S4 include:

[0128] S4.1: Obtain the deviation value e of the newly collected rainstorm observation data and the medium- and long-term forecast data l , and set the deviation threshold h;

[0129] If e l ≤ h, continue monitoring;

[0130] If e l > h, perform multi-dimensional feature extraction on the newly collected rainstorm observation data, and use an intelligent fusion algorithm to perform multi-source fusion on the extracted multi-dimensional features to obtain a fused feature set Among them, δ c represents the c-th type of fused feature, represents the weight of the r-th feature data in the c-th type of feature, and g c,r represents the r-th feature data in the c-th type of feature in the newly collected rainstorm observation data;

[0131] Among them, in the multi-dimensional feature extraction of the newly collected rainstorm observation data, the extracted multi-dimensional features include precipitation, precipitation intensity, precipitation duration, wind speed, wind direction, air pressure, temperature, and humidity data.

[0132] S4.2: Use the fused feature set to correct the initial prediction result of the rainstorm disaster, and provide the corrected prediction result to decision-makers and emergency management personnel in a visual manner.

[0133] Further, the specific steps of S4.2 include:

[0134] (1) Obtain the initial prediction result of the rainstorm disaster and the fused feature set obtained through the intelligent fusion algorithm;

[0135] (2) Load the pre-trained regression model, and use the fused feature set as the input. The pre-trained regression model corrects the initial estimation result of the rainstorm disaster. Based on these input features, the pre-trained regression model combines the knowledge it has learned before, that is, the relationship between the features learned from the training data and the target variable, to correct the initial estimation result;

[0136] (3) According to the needs of decision-makers and emergency management personnel, select maps, charts or dashboards to visually display the corrected estimation results;

[0137] (4) Present the corrected estimation results to decision-makers and emergency management personnel in a visual way, including information such as the possible impact areas, intensity, and duration of the rainstorm disaster, as well as relevant warnings and suggestions;

[0138] (5) Decision-makers and emergency management personnel formulate corresponding decisions and emergency response plans based on the visual results, including measures such as evacuation, rescue, and material allocation.

[0139] Embodiment 3

[0140] Please refer to Figure 4 , another embodiment provided by the present invention: A meteorological disaster risk estimation system based on the fusion of historical observations and forecasts, including:

[0141] A data collection module, a model construction module, a monitoring and comparison module, a dynamic correction module, and a report generation module;

[0142] The data collection module is used to collect long-term historical rainstorm observation data in the target area and rainstorm forecast data released by the meteorological department, and perform preprocessing, including cleaning, denoising, interpolation, etc. on the data to ensure the accuracy, integrity and consistency of the data, and provide a high-quality data basis for subsequent modeling and analysis;

[0143] The model construction module is used to construct an initial estimation model for rainstorm disasters using machine learning algorithms based on the preprocessed historical rainstorm observation data, and perform initial estimation using the real-time obtained rainstorm forecast data;

[0144] The monitoring and comparison module is used to set the time interval according to the density of the meteorological monitoring network layout in the target area and the sensitivity of the disaster evolution during the entire period when the medium- and long-term forecast takes effect, and monitor and compare the newly collected rainstorm observation data with the medium- and long-term forecast data on which the initial estimation is based in real time, obtain the deviation situation, and feedback the deviation to the estimation model;

[0145] A dynamic correction module, which is used to monitor the deviation situation in real time. Once it is found that the actual situation deviates from the forecast, it immediately starts the preset intelligent fusion and dynamic correction algorithm to correct the initial estimation result and improve the accuracy of the estimation;

[0146] A report generation module, which is used to perform multi-dimensional visual output of the corrected heavy rain disaster estimation result at the end period of the medium- and long-term forecast, and generate a text report containing data indicators to provide decision-making support for the public.

[0147] The model construction module includes: a model training unit and an initial estimation unit;

[0148] The model training unit is used to train a machine learning algorithm using historical data to construct an initial heavy rain disaster estimation model;

[0149] The initial estimation unit is used to obtain the heavy rain forecast data structure in real time, and extract and map the information segments that are precisely corresponding to the input variables of the initial heavy rain disaster estimation model to obtain the preliminary result of the heavy rain disaster.

[0150] The monitoring and comparison module includes: a time interval setting unit, a data collection unit, a data comparison unit, and a deviation feedback unit;

[0151] The time interval setting unit is used to intelligently set the time interval according to the meteorological monitoring network layout and the sensitivity of disaster evolution;

[0152] The data collection unit is used to drive the monitoring stations to re-collect the latest heavy rain observation data within each time interval;

[0153] The data comparison unit is used to compare the newly collected data with the medium- and long-term forecast data point by point and calculate the deviation;

[0154] The deviation feedback unit is used to timely feedback the deviation situation to the initial heavy rain disaster estimation model.

[0155] The dynamic correction module includes: a deviation monitoring unit and a dynamic correction unit;

[0156] The deviation monitoring unit is used to monitor the deviation situation in the comparison result in real time;

[0157] The dynamic correction unit is used to correct the initial estimation result according to the deviation situation by using the dynamic correction algorithm.

[0158] The report generation module includes: a multi-dimensional visualization unit and a text report generation unit;

[0159] The multi-dimensional visualization unit is used to present the estimation result to relevant departments and the public in an intuitive and easy-to-understand way;

[0160] A text report generation unit for automatically generating a text report including data metrics, warning information, recommended measures, etc.

[0161] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make changes, modifications, substitutions, and variations to the above embodiments without departing from the purpose and scope of the present invention. All of these are within the protection scope of the present invention.

Claims

1. A meteorological disaster risk estimation method based on the fusion of historical observations and forecasts, characterized in that: include: Step S1: Use a dual-thread data collection channel to collect long-term historical series of heavy rain observation data in the target area and heavy rain forecast data released by the meteorological department for the target area, and perform preprocessing; Step S2: Based on the pre-processed historical rainstorm observation data, a machine learning algorithm is used to construct an initial prediction model for rainstorm disasters, and the information corresponding to the model input variables in the real-time rainstorm forecast data is input into the initial prediction model for rainstorm disasters to obtain an initial prediction result for rainstorm disasters; Step S3: During the whole period when the medium- and long-term forecast is effective, the time interval is set according to the density of the meteorological monitoring network layout in the target area and the sensitivity of the disaster evolution. In each time interval, the monitoring stations distributed in each grid unit are driven to re-collect the latest rainstorm observation data, and the newly collected rainstorm observation data are compared point by point with the medium- and long-term forecast data based on the initial estimation to obtain the deviation of the corresponding indicators. At the same time, the deviation is fed back to the initial estimation model of rainstorm disasters; Step S4: monitor the deviation in real time. Once the actual situation deviates from the forecast, immediately start the preset intelligent fusion and dynamic correction algorithm to correct the initial estimation result; Step S5: After cyclic dynamic correction, at the end of the medium- and long-term forecast period, the corrected rainstorm disaster prediction results are output in a multi-dimensional visualization, and a text report is generated. The text report includes the rainstorm disaster prediction results, correction process, deviations and possible disaster impact information.

2. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 1 is characterized in that: The rainstorm observation data and rainstorm forecast data in step S1 include rainstorm events and process data of different time scales; the time scales include month, season and year.

3. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 1 is characterized in that: The initial prediction model for rainstorm disasters must meet the following requirements during the construction process: For predictions within a 30-day range, monthly scale forecast data are used, which include forecast data for the next ten days and S2S model forecast data.

4. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 1 is characterized in that: The specific steps of step S2 include: S2.1: Obtain pre-processed historical rainstorm observation data and rainstorm forecast data; S2.2: Load the pre-built LSTM neural network model architecture, and use the pre-processed historical rainstorm observation data as input to train the pre-built LSTM neural network model to generate a trained initial prediction model for rainstorm disasters; S2.3: Load a linear regression model as the base model and use all the features of the preprocessed rainstorm forecast data as the initial feature set X = {x1,...,x n }, where x i ,i∈[1,n] represents the i-th initial feature in the preprocessed heavy rain forecast data, and n represents the number of initial features; S2.4: Input the initial feature set into the base model for prediction, and then generate the base model coefficient β i The absolute value of Imp is used to calculate the feature importance i , and remove the initial features corresponding to the smallest M feature importance absolute values, i represents the base model coefficient index value; S2.5: Repeat S2.4 until the predetermined number of features is reached. At the same time, during the recursive elimination process, the order in which each feature is retained is recorded; S2.6: sorting the features according to the order in which they are retained during the recursive elimination process, and extracting the screened and sorted features from the preprocessed rainstorm forecast data to generate medium- and long-term rainstorm forecast feature data; S2.7: The medium- and long-term rainstorm forecast feature data is input into the trained rainstorm disaster initial prediction model. The rainstorm disaster initial prediction model makes predictions based on the input feature data and outputs the initial prediction results of the rainstorm disaster.

5. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 4 is characterized in that: The specific steps of step S3 include: S3.1: Obtain the location information of meteorological monitoring stations in the target area, divide the target area into N grids according to the size, shape and meteorological characteristics of the target area, and select a monitoring station as a representative in each grid unit according to the station distribution and meteorological monitoring requirements, using the formula Evaluate the density of monitoring sites in each grid cell, where f(z,s) represents the density estimate at position z and conditional variable s, m represents the total number of monitoring sites in the current grid cell, and b with (s) represents the bandwidth function, w j represents the importance weight of the jth site, K(·) represents the kernel function, z j represents the location of the jth monitoring station; S3.2: According to the sensitivity of disaster evolution and real-time meteorological conditions, set the time interval ΔT, and in each time interval, use the monitoring station locations determined in step S3.1 to drive the monitoring stations distributed in each grid unit to re-collect the latest heavy rain observation data.

6. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 5 is characterized in that: The specific steps of step S3 also include: S3.3: For the newly collected rainstorm observation data X′={x1′,...,x′ u } and the medium- and long-term forecast data Y={y1,...,y u } to sort, where x′ u represents the uth newly collected rainstorm observation data, y u represents the u-th medium- and long-term forecast data, and u represents the number of newly collected heavy rain observation data or medium- and long-term forecast data; S3.4: Set a time window Δt. If the geographical location and timestamp of the newly collected rainstorm observation data and medium- and long-term forecast data fall within the time window Δt, the match is considered successful. S3.5: traverse the rainstorm observation data, and for each piece of rainstorm observation data, check whether there is a matching record in the medium- and long-term forecast data. If a matching record is found, associate it and save the matching result; S3.6: Compare the matched data point by point and calculate the difference between the newly collected heavy rain observation data and the medium- and long-term forecast data. l , and analyze the reasons for the deviation, where l represents the data index value.

7. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 6, characterized in that: The specific steps of step S4 include: S4.1: Obtain the deviation value e between the newly collected heavy rain observation data and the medium- and long-term forecast data l , and set the deviation threshold h; If e l ≤h, continue monitoring; If e l >h, then multi-dimensional features are extracted from the newly collected rainstorm observation data, and the extracted multi-dimensional features are fused from multiple sources using an intelligent fusion algorithm to obtain the fused feature set δ c , where δ c represents the c-th type of fusion features; S4.2: Use the fused feature set to revise the initial prediction results of rainstorm disasters, and provide the revised prediction results to decision makers and emergency management personnel in a visual manner.

8. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 7, characterized in that: The model input variables in step S2 include meteorological elements, geographical terrain characteristic parameters and cumulative effect indicators of previous rainfall in the same historical period and the same grid unit; the meteorological elements include the daily variation mean and extreme values ​​of temperature, humidity, air pressure, wind direction and wind speed; the geographical terrain characteristic parameters include altitude, slope, water system distance, and soil permeability; the cumulative effect indicators of previous rainfall include the accumulated rainfall and the number of wet days in the previous week and the previous month.

9. The meteorological disaster risk estimation method based on historical observation and forecast fusion according to claim 8, characterized in that: The visualization content in step S5 includes rainstorm disaster events and their processes at monthly, seasonal and annual scales.

10. A meteorological disaster risk prediction system based on the fusion of historical observations and forecasts, which is used to implement the meteorological disaster risk prediction method based on the fusion of historical observations and forecasts as described in any one of claims 1 to 9, characterized in that: include: Data collection module, model building module, monitoring and comparison module, dynamic correction module, report generation module; The data collection module is used to collect the historical long-term series of rainstorm observation data in the target area and the rainstorm forecast data issued by the meteorological department, and perform preprocessing; The model building module is used to build an initial prediction model for rainstorm disasters based on pre-processed historical rainstorm observation data using a machine learning algorithm, and to make an initial prediction using real-time acquired rainstorm forecast data; The monitoring and comparison module is used to set time intervals according to the density of the meteorological monitoring network layout in the target area and the sensitivity of disaster evolution during the entire period when the medium- and long-term forecast is effective, monitor and compare the newly collected rainstorm observation data with the medium- and long-term forecast data based on the initial estimation in real time, obtain the deviation, and feed the deviation back to the estimation model; The dynamic correction module is used to monitor the deviation in real time. Once the actual situation deviates from the forecast, the preset intelligent fusion and dynamic correction algorithm is immediately activated to correct the initial estimation result. The report generation module is used to output the corrected rainstorm disaster prediction results in a multi-dimensional visual manner at the end of the medium- and long-term forecast period, and to generate a text report containing data indicators.

11. The meteorological disaster risk prediction system based on historical observation and forecast fusion according to claim 10, characterized in that: The monitoring and comparison module includes: a time interval setting unit, a data acquisition unit, a data comparison unit, and a deviation feedback unit; The time interval setting unit is used to intelligently set the time interval according to the layout of the meteorological monitoring network and the sensitivity of disaster evolution; The data collection unit is used to drive the monitoring station to re-collect the latest rainstorm observation data at each time interval; The data comparison unit is used to compare the newly collected data with the medium- and long-term forecast data point by point and calculate the deviation; The deviation feedback unit is used to timely feed back the deviation situation to the initial prediction model of rainstorm disaster.

12. The meteorological disaster risk prediction system based on historical observation and forecast fusion according to claim 11, characterized in that: The dynamic correction module includes: a deviation monitoring unit and a dynamic correction unit; The deviation monitoring unit is used to monitor the deviation in the comparison result in real time; The dynamic correction unit is used to correct the initial estimation result using a dynamic correction algorithm according to the deviation.

Images