Bohai Rim Low-Air Meteorological Safety Engine Artificial Intelligence System Based on Historical Reanalysis and Multi-Model Data

Abstract

This invention belongs to the field of meteorological safety technology and discloses an artificial intelligence system for a Bohai Rim low-altitude meteorological safety engine based on historical reanalysis and multi-model data. The Bohai Rim low-altitude meteorological safety engine artificial intelligence algorithm includes the following modules: meteorological variable reconstruction and extraction module, extreme event identification module, model training module, model fusion module, and safety engine construction module. The Bohai Rim low-altitude meteorological safety engine artificial intelligence algorithm of this invention has the following advantages: (1) improved data fusion and timeliness; (2) improved accuracy in identifying small-scale extreme weather at low altitudes; (3) a multi-dimensional meteorological risk assessment system and standardized grading; and (4) dynamic adaptive optimization and real-time adjustment.

Description

Technical Field

[0001] This invention belongs to the field of meteorological safety technology and relates to an artificial intelligence system for low-altitude meteorological safety engines in the Bohai Rim region based on historical reanalysis and multi-model data. Background Technology

[0002] With the rapid development of emerging industries such as the low-altitude economy (e.g., drone transportation, short-haul air transport, urban air mobility), smart city management, and coastal energy development, the importance of low-altitude (0–3 km) meteorological safety assurance is becoming increasingly prominent. Especially in the Bohai Rim region (the coastal areas of Beijing, Tianjin, Hebei, Shandong, and Liaoning), due to complex terrain, significant land-sea interaction, and frequent monsoon activity, sudden meteorological disasters such as localized severe convection, wind shear, low visibility, low-altitude jet streams, and icing occur frequently, posing a significant threat to air transport, port shipping, and urban operations. Traditional meteorological support methods mainly rely on numerical weather prediction (NWP) models (such as GFS and ECMWF), ground automatic weather station observations, wind profiler radar, and satellite remote sensing data. However, due to limitations:

[0003] (1) Insufficient mode resolution;

[0004] (2) The spatial coverage of single observation data is incomplete;

[0005] (3) Localized, small-scale sudden weather events are difficult to capture accurately;

[0006] (4) The real-time identification and accurate early warning of severe low-altitude weather are not ideal.

[0007] With the rapid development of new-generation technologies such as artificial intelligence (especially deep learning), big data fusion analysis, and reinforcement learning, the meteorological field is gradually exploring the application of AI methods in weather forecasting and risk assessment, aiming to break through the bottleneck of traditional numerical forecasting and improve the ability of intelligent identification and dynamic early warning of low-altitude meteorological conditions.

[0008] In the Bohai Rim region, severe low-altitude weather has the following characteristics:

[0009] (1) Small spatial scale (within tens of kilometers) and rapid temporal evolution (from minutes to hours);

[0010] (2) Traditional ground stations and conventional meteorological radars have sparse coverage, making it difficult to form a continuous and detailed observation network;

[0011] (3) Although the reanalysis data is stable, it lacks the ability to record real-time events that are sudden and extreme.

[0012] (4) Single-model forecasts have significant uncertainties and biases, especially in forecasting elements such as low-level wind field, humidity field, visibility, and icing conditions.

[0013] Therefore, it is necessary to organically integrate historical meteorological evolution experience (summarized through reanalysis data) with current multi-model ensemble forecasts (complementary information from different models), and combine artificial intelligence technology for spatiotemporal feature mining and risk decision optimization, in order to truly achieve intelligent identification and efficient early warning of severe weather in the low-altitude environment around the Bohai Sea.

[0014] Currently, there are several technical solutions that attempt to use meteorological data and machine learning algorithms for weather event prediction, including the following types:

[0015] (1) Statistical research on meteorological characteristics based on reanalysis data: By using global reanalysis data such as ERA5 and MERRA-2, the evolution patterns of historical meteorological elements are studied, and statistical analysis of meteorological events is conducted. For example: calculating the multi-year trends of wind shear, turbulence, and visibility changes in a specific region; identifying extreme meteorological events (such as the number of days with low visibility and the frequency of severe winds); and using it for meteorological disaster risk assessment and planning guidance. Main problems: lack of real-time capability, unable to meet the needs of dynamic early warning; only statistical inference can be performed, and it cannot be directly used for real-time decision-making; the resolution of reanalysis data is limited (ERA5 has a resolution of 31 km or higher and 0.25°), making it difficult to analyze fine-scale changes. (Hersbach, H. et al. (2020). The ERA5 globalreanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730),1999–2049.)

[0016] (2) Multi-model fusion method based on numerical forecast model ensemble output: This method uses ensemble forecast data from different forecast centers (such as ENS from ECMWF and GEFS from GFS) to fuse multi-model outputs. Through weighted averaging, Bayesian fusion, and other methods, the stability and reliability of forecasts are improved. For example: multi-model weighted fusion of wind field, humidity field, and temperature field; quantification of forecast uncertainty (probabilistic forecast); simple rule judgment of hazardous weather conditions such as wind shear and turbulence. Main problems: The fusion methods are mostly linear or simple statistical rules, which cannot effectively extract complex nonlinear correlation features; there is a lack of mechanisms for intelligent learning and dynamic optimization based on actual observations or historical events; the risk assessment index system is singular and it is difficult to quantify the comprehensive impact of multiple meteorological risks. (Leutbecher, M., & Palmer, TN (2007). Ensemble forecasting. Journal of Computational Physics, 227(7), 3515-3539.)

[0017] (3) Local weather event identification models based on machine learning. In recent years, some studies have begun to use machine learning (such as support vector machine (SVM), random forest (RF), and convolutional neural network (CNN)) to identify local weather phenomena such as wind shear, low visibility, and short-term heavy precipitation. For example, classifiers are trained using observational data to determine whether extreme weather events have occurred; meteorological characteristic variables (temperature, humidity, and wind speed change rate) are used as model inputs; and the probability or level of extreme weather events is directly given. Main problems: Most models are limited to single-region or small-sample training, resulting in weak generalization ability; there is a lack of systematic design that integrates large-scale historical data and multi-model forecast data; the warning results lack physical interpretation, leading to low user trust. (Lagerquist, R., McGovern, A. & Gagne, DJ, II. (2019) Deep learning forspatially explicit prediction of synoptic-scale fronts. Weather and Forecasting, 34, 1137-1160.)

[0018] Despite the continuous development of technologies for low-altitude weather warnings, existing technologies still have the following main problems:

[0019] (1) Insufficient data fusion capability and poor real-time performance;

[0020] Existing systems mainly rely on a single source (such as a single numerical model output or a single observation data), lacking the ability to integrate and process historical data and real-time data from multiple models. This results in a time lag in capturing the evolution of severe local weather, a slow response speed, and an inability to meet the actual needs of low-altitude economic and emergency response for minute-level rapid early warning.

[0021] (2) Low spatial resolution, resulting in insufficient accuracy in identifying extreme weather conditions;

[0022] Limited by the spatial resolution of traditional models (usually above 25km) and simple statistical or rule-based methods, existing technologies are unable to identify extreme weather phenomena such as small-scale strong convection and wind shear within 5km, resulting in a high identification error rate and affecting the accuracy of early warnings.

[0023] (3) The risk assessment system is too simplistic and lacks standardized quantitative indicators;

[0024] Most existing low-altitude weather warning systems can only provide qualitative risk warnings and lack structured, multi-dimensional quantitative indicators (such as turbulence index, wind shear index, low visibility risk level, etc.), which cannot provide downstream application units (such as airports, ports, and drone management platforms) with intuitive and graded decision-making basis.

[0025] (4) The early warning model lacks adaptive capability and is prone to false alarms and false alarms;

[0026] Most existing systems use static rules to set thresholds and lack a mechanism to dynamically adjust warning parameters based on real-time weather changes. This results in high false alarm and missed alarm rates during periods of drastic weather fluctuations, reducing the reliability and practicality of meteorological service systems. Summary of the Invention

[0027] This invention provides an artificial intelligence system for low-altitude meteorological safety in the Bohai Rim region based on historical reanalysis and multi-mode data, aiming to achieve high-precision, real-time risk identification and early warning of severe weather on a small scale at low altitudes.

[0028] The technical solution of this invention:

[0029] (1) Meteorological variable reconstruction and extraction module:

[0030] Historical reanalysis data and real-time observation data are fused from multiple sources to ensure consistent resolution across spatial and temporal distributions. Historical reanalysis data includes data acquired from ERA5, JRA-55, and MERRA-2, while real-time observation data includes ground station wind and temperature data, LiDAR-derived wind profiles, radar data, and microwave radiometer observations. Boundary layer height, wind shear, eddy current, inversion intensity, gradient, and disturbances are treated as meteorological variables. Interpolation methods are employed to improve the resolution and accuracy of these meteorological variables across temporal and spatial distributions, followed by dimensionality reduction and compression.

[0031] (2) Extreme event identification module:

[0032] Based on historical extreme meteorological events in the study area, a standardized label rule base is constructed, which includes pollution accumulation events, wind shear events, and local strong wind events. Historical extreme meteorological events are collected to establish a sample event set. Long short-term memory networks are used to reconstruct the observation data sequence of meteorological variables for error detection, which is used to identify extreme meteorological events. The observation data sequence of meteorological variables is the observation value of one or more meteorological variables at consecutive time points.

[0033] A deep autoencoder is used to automatically detect and label historical extreme meteorological events; finally, a composite index of multiple meteorological variables is constructed to improve the sensitivity of extreme meteorological event identification.

[0034] (3) Model training module:

[0035] Based on a time-series forecasting model, the model is trained using time-series data of historical meteorological variables and labeled historical extreme meteorological events. A combination of convolutional neural networks and recurrent neural networks is used to enhance the learning ability of local and long-term trends. To better consider the synergistic effects between modeling regions, a graph neural network is used to represent the interaction of meteorological models between modeling regions. Nodes represent modeling regions, and edges represent the correlation between modeling regions, thereby enhancing the spatiotemporal data modeling capability. An integrated framework of multiple models is constructed to enhance the adaptability and robustness to different meteorological models.

[0036] The modeling region is a spatial unit or a uniformly divided high-resolution grid within the area to be studied, and each modeling region serves as a node in the graph neural network.

[0037] Meteorological models model the spatiotemporal evolution of meteorological variables and their interactions within a modeling region. In graph neural networks, edges represent the similarity or physical transmission relationships between modeling regions.

[0038] (4) Pattern fusion module:

[0039] Perform unified spatiotemporal gridding processing on data from multiple sources;

[0040] The processed output is averaged using a Bayesian model, and the results of multiple models in the model training module are fused using a weighted average method.

[0041] The results are dynamically corrected using an error correction algorithm.

[0042] (5) Security Engine Building Module:

[0043] The security engine building block supports real-time rolling forecasts, updating inference results every 10 minutes to ensure rapid response to sudden extreme weather events; it also updates early warnings for extreme weather events every 30 minutes, providing the latest event identification results, level assessments, and risk maps.

[0044] The security engine building module outputs risk level maps, probability heat maps, and event time sequence maps for extreme weather events, helping relevant departments monitor weather changes in real time.

[0045] The safety engine building module provides a decision support interface, which connects with port scheduling, airport meteorological platforms and environmental monitoring platforms to promptly push meteorological risk warnings and preventive measures.

[0046] The beneficial effects of this invention are:

[0047] (1) Data fusion and improved timeliness;

[0048] Existing technologies often rely on a single model or a single data source, failing to fully utilize the combination of historical reanalysis data and multi-model real-time forecast data, resulting in poor early warning timeliness. This invention, by combining historical ERA5 and other reanalysis data with multi-model ensemble forecasts, fully utilizes multi-source heterogeneous data to provide high-frequency real-time dynamic monitoring of meteorological conditions, thereby achieving rapid response and timely early warning, significantly improving the response speed to sudden extreme weather events in low-altitude environments.

[0049] (2) Improve the accuracy of identifying small-scale extreme weather at low altitudes;

[0050] Existing technologies largely rely on numerical weather prediction (NWP) models, which have low spatial resolution and struggle to accurately capture small-scale severe weather phenomena at low altitudes (such as wind shear, turbulence, and low visibility). This invention combines deep learning and physical reasoning methods, particularly through the integration of CNN and Transformer models, to extract local small-scale weather features from high-resolution data. This enhances the ability to accurately identify low-altitude weather and reduces false alarms and missed alarms.

[0051] (3) Multidimensional meteorological risk assessment system and standardized classification;

[0052] Traditional weather warning systems often rely on qualitative descriptions or simple rules, lacking standardized quantitative risk assessments. This invention constructs a multi-dimensional risk index system, including pollution accumulation index, wind shear index, and low visibility risk level, providing users with more scientific and standardized risk assessment results, facilitating specific operational decisions in practical applications.

[0053] (4) Dynamic adaptive optimization and real-time adjustment;

[0054] Most existing technologies use fixed rules or thresholds for early warning, which cannot be dynamically adjusted according to real-time weather changes, resulting in significant false alarms and missed alarms. This invention, through a reinforcement learning mechanism, adjusts the early warning threshold based on real-time weather data, achieving adaptive optimization of the early warning strategy. This significantly reduces the probability of false alarms and missed alarms, further improving the robustness and reliability of the system. Detailed Implementation

[0055] The specific embodiments of the present invention will be further described below in conjunction with the technical solution.

[0056] An artificial intelligence system for low-altitude meteorological safety in the Bohai Rim region, based on historical reanalysis and multi-mode data, includes the following modules:

[0057] (1) Meteorological variable reconstruction and extraction module:

[0058] Historical reanalysis data and real-time observation data are fused from multiple sources to ensure consistent resolution across spatial and temporal distributions. Historical reanalysis data includes data acquired from ERA5, JRA-55, and MERRA-2, while real-time observation data includes ground station wind and temperature data, LiDAR-derived wind profiles, radar data, and microwave radiometer observations. Boundary layer height, wind shear, vorticity, inversion intensity, gradient, and disturbances are treated as meteorological variables. Interpolation methods are employed to improve the resolution and accuracy of these meteorological variables across temporal and spatial distributions. Dimensionality reduction and compression are then performed as follows: First, data integrity is achieved through various interpolation methods (such as Kriging and IDW). Second, principal component analysis (PCA) and empirical orthogonal function (EOF) are combined in the dimensionality reduction stage to extract principal modal features. Finally, a deep autoencoder is used for compression.

[0059] (1.1) Multi-source data fusion:

[0060] Historical reanalysis data (ERA5:0.25°, JRA-55:1.25°, MERRA-2:0.5°) are uniformly resampled to a high-resolution grid (e.g., 5km).

[0061] After quality control, the real-time observation data is integrated into the high-resolution grid.

[0062] Quality control involves processing real-time observation data, including integrity checks, physical rationality verification, temporal and spatial consistency checks, and outlier removal, to ensure the accuracy and reliability of real-time observation data.

[0063] (1.2) Processing of meteorological variables:

[0064] Boundary layer height (BLH): The mixed layer height diagnostic formula (such as Richardson number method, Holzworth method) is fused with the ERA5 diagnostic value;

[0065] Wind shear and vorticity: Calculate wind shear and vorticity index (VWS, wind shear level) based on the wind speed profile of 925-850 hPa layer to identify low-altitude instability conditions;

[0066] Inversion intensity: Calculate the near-surface temperature lapse rate and use strong inversion events as triggering conditions for pollution accumulation events or low visibility extreme events;

[0067] Gradients and disturbances: The central difference method is used to construct temperature, humidity and wind speed gradient fields and extract local disturbance characteristics within the boundary layer;

[0068] (1.3) Feature reduction and deep compression modeling:

[0069] First, use various interpolation algorithms (such as Kriging, IDW, etc.) to spatially interpolate missing or irregular data points to ensure data integrity;

[0070] Principal component analysis (PCA) and empirical orthogonal function (EOF) analysis were then used to reduce the dimensionality of the meteorological variables and extract the most representative features.

[0071] Finally, a deep autoencoder is used to compress the high-dimensional meteorological data, extract potential high-order features, and reduce computational complexity. The high-dimensional meteorological data is a three-dimensional meteorological data tensor composed of multiple meteorological variables, multiple vertical height layers, and multiple time steps.

[0072] (2) Extreme event identification module:

[0073] Based on historical extreme meteorological events in the study area, a standardized label rule base was constructed, which includes pollution accumulation events, wind shear events, and local strong wind events. Historical extreme meteorological events were collected to establish a sample event set. Long Short-Term Memory (LSTM) networks were used to detect reconstruction errors in the observed data sequences of meteorological variables for the identification of extreme meteorological events. The observed data sequences of meteorological variables are the observation values ​​of one or more meteorological variables at consecutive time points.

[0074] A deep autoencoder is used to automatically detect and label historical extreme meteorological events. Finally, a composite index of multiple meteorological variables, such as a joint index of temperature gradient and wind speed, is constructed to improve the sensitivity of extreme meteorological event identification. The specific steps are as follows:

[0075] (2.1) Construction of a standardized tag rule base:

[0076] Based on historical extreme meteorological events in the sample event set, the following criteria are defined:

[0077] Pollution accumulation event: BLH < 300m + wind speed < 2m / s + PM2.5 > 100μg / m³ + presence of inversion layer;

[0078] Wind shear event: ΔV / ΔZ>0.015s⁻¹; where ΔV is the vertical height difference and ΔZ is the change in wind speed between the two height layers;

[0079] Localized strong wind events: 10-minute wind speed > 12 m / s + turbulent kinetic energy > threshold;

[0080] Historical extreme meteorological events are labeled in the sample event set to construct training and validation sets;

[0081] (2.2) Integration of anomaly detection methods:

[0082] LSTM reconstruction error method: The Long Short-Term Memory (LSTM) network is trained with normal time period data. When the input data of the LSTM network includes data of extreme meteorological events, the error between the actual observation value and the predicted value of the LSTM network increases significantly. The normal time period data refers to time series data that reflect the normal change pattern of meteorological variables collected during periods when no extreme meteorological events occur.

[0083] Deep autoencoder identification: Training a deep autoencoder to compress and reconstruct meteorological variables; a significant increase in reconstruction error indicates a candidate extreme meteorological event.

[0084] Local Outlier Factor (Isolation Forest & LOF): Used for outlier detection in a low-dimensional compressed space; where the low-dimensional compressed space is the space after dimensionality reduction by a deep autoencoder or PCA.

[0085] (2.3) Design of multivariate composite indicators:

[0086] Construct mathematical models for Pollution Weather Potential Index (PAI) and Wind Shear Index (WSI);

[0087] The mathematical model for the Pollution Weather Potential Index is as follows: PAI = w1 * BLH −1 +w2*wind speed −1 +w3* Temperature inversion intensity +w4* Humidity stability;

[0088] Where w1-w4 are the weights of the corresponding meteorological variables;

[0089] The mathematical model for the wind shear index (WSI) is as follows: ;

[0090] The severity of extreme weather events is classified into four categories: light, moderate, severe, and extremely severe, using fuzzy logic rules or decision tree methods (CART).

[0091] (3) Model training module:

[0092] Based on time-series forecasting models (LSTM, Transformer, Temporal Fusion Transformer, etc.), the model is trained on time-series data of historical meteorological variables and labeled historical extreme meteorological events. A combination of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) is used to enhance the learning ability for local and long-term trends. To better consider the synergistic effects between modeling regions, Graph Neural Networks (GCN) are used to represent the interaction of meteorological models between modeling regions. Nodes represent modeling regions, and edges represent the correlations between modeling regions, enhancing the spatiotemporal data modeling capability. An integrated framework for multiple models (such as Bagging, Boosting, Stacking, etc.) is constructed to enhance the adaptability and robustness to different meteorological models.

[0093] The modeling region is a spatial unit or a uniformly divided high-resolution grid within the area to be studied, and each modeling region serves as a node in the graph neural network.

[0094] Meteorological models model the spatiotemporal evolution of meteorological variables and their interactions within a modeling region. In graph neural networks, edges represent the similarity or physical transmission relationships between modeling regions.

[0095] The specific steps are as follows:

[0096] (3.1) Integration framework of multiple models:

[0097] The integrated framework of multiple models is a multi-model collaborative spatiotemporal feature learning system. This system comprises five modules operating in parallel: a local time series modeling module, a key variable-driven modeling module, a local spatial structure extraction module, a regional collaborative modeling module, and a convolutional + recurrent hybrid temporal modeling module. Details are as follows:

[0098] (a) Local time series modeling module (LSTM or GRU):

[0099] Input: Time-series data consisting of meteorological variables (temperature, humidity, wind speed, wind direction, BLH, etc.);

[0100] Task: To study the historical evolution patterns of meteorological variables and predict the continuous value changes over the next N hours;

[0101] Output: Probability score of extreme weather events;

[0102] (b) Key variable-driven modeling module (Transformer or TFT):

[0103] Input: Time series data consisting of meteorological variables (temperature, humidity, wind speed, wind direction, BLH, etc.) and the weighting mechanism of meteorological variables;

[0104] Task: Use attention mechanisms to identify which meteorological variables are more likely to drive extreme weather events in the current period;

[0105] Output: Driver attention weights;

[0106] (c) Local Spatial Structure Extraction Module (CNN-LSTM):

[0107] Input: An image (2D matrix) of meteorological variables in a high-resolution grid.

[0108] Task: Use CNN to extract spatial gradient information and shape features (such as fronts and wind shear zones); then use LSTM to further process their temporal evolution trends.

[0109] Output: Spatial gradient recognition score;

[0110] (d) Regional collaborative modeling module, select GCN or GAT:

[0111] Input: Construct a regional graph structure. The regional graph structure abstracts each modeling region into a node in a graph neural network within the scope of the study, and represents the relationship between the modeling regions through the edges between the nodes. The modeling region is a divided spatial unit or a uniformly divided high-resolution grid. Each modeling region node contains boundary layer height, wind speed, temperature and humidity, inversion intensity, and vortex feature vector.

[0112] Task: Utilize graph convolutional networks or graph attention networks to mine the risk propagation relationships and mutual influence mechanisms between modeled regions on the regional graph structure;

[0113] Output: Risk score for the coordinated propagation of modeling regions, used to describe the spatial diffusion trend and risk linkage intensity of extreme meteorological events among different modeling regions;

[0114] (e) Convolutional + Recurrent Hybrid Temporal Modeling Module, select CNN-RNN (such as CNN-BiLSTM):

[0115] Input: Historical meteorological variable time series, single meteorological variable or combination of multiple meteorological variables;

[0116] Task: Use CNN to extract local pattern features (e.g., short-term abrupt changes, local fluctuations) from time series of meteorological variables, and then use RNN (LSTM or GRU) to capture long-term dependencies and trends;

[0117] Output: The fused time-series prediction results are used to identify precursor features of extreme meteorological events;

[0118] (3.2) Multi-model integration method:

[0119] (a) Multi-model ensemble structure:

[0120] The first-level model (Base Learners) combines the outputs of the five modules in step (3.1) at the feature level to form an integrated feature vector, and then standardizes it.

[0121] The second-level fusion model (Meta Learner) uses lightweight ensemble models such as logistic regression, random forest, and XGBoost to fuse the outputs of the first-level model; or it uses a deep stacking neural network to learn the combined weights of the first-level model output.

[0122] (b) Output content:

[0123] Probability values ​​for extreme meteorological events: output probability of pollution accumulation, probability of local strong winds, and probability of wind shear;

[0124] Uncertainty estimation: The standard deviation and consistency between outputs are calculated by using multi-model ensemble methods (such as Stacking, Bayesian model averaging, etc.) to provide a confidence range;

[0125] Event severity prediction results: Based on the integrated score, combined with rules and fuzzy logic, four levels of event severity are output: mild, moderate, severe, and extremely severe.

[0126] Key factor explanation information: Tracing back the weights and contributions of important features in multi-model fusion to form a traceable decision path.

[0127] (c) Deployment method:

[0128] Based on multi-model deployment as a scalable model API interface, it can receive multi-variable inputs at any time period and automatically complete inference, judgment and explanation output.

[0129] (4) Pattern fusion module:

[0130] A unified spatiotemporal gridding process is applied to data from multiple sources (reanalysis, observation, and future scenarios) to ensure comparability between data. Bayesian model averaging (BMA) is performed on the processed output, and the results from different models are fused using a weighted averaging method to reduce the error of a single model. Error correction algorithms (such as Kalman filtering) are applied to dynamically correct the results, improving the model's accuracy and predictive ability. Based on short-term warnings of 0–7 days, and combined with future climate scenarios in the 2030s, long-term trends and the probability of extreme events are assessed, providing assessment and decision support for future meteorological risks. The specific implementation steps are as follows:

[0131] (4.1) Spatiotemporal standardization processing:

[0132] Establish a unified spatial grid (e.g., 0.05°×0.05°) and temporal resolution (10 minutes to 1 hour) to resample multi-source data (ERA5 / observations / CMIP6 data).

[0133] (4.2) Bayesian Model Averaging (BMA) Fusion:

[0134] The system performs a weighted average of the outputs from multiple models, with the weights adaptively updated based on the models' performance during the historical validation period; it also provides the probability density function for the fused extreme weather events.

[0135] (4.3) Dynamic error correction:

[0136] Short-term forecast biases are corrected using dynamic linear models (DLM) or Kalman filtering methods; long-term future climate (CMIP6 SSP scenario) is aligned with current observations using bias correction functions (such as Quantile Mapping).

[0137] (5) Security Engine Building Module:

[0138] The safety engine module supports real-time rolling forecasts, updating inference results every 10 minutes to ensure rapid response to sudden extreme weather events. It updates extreme weather event warnings every 30 minutes, providing the latest event identification results, level assessments, and risk maps. The module outputs risk level maps, probability heatmaps, and event time series diagrams for extreme weather events, helping relevant departments monitor weather changes in real time. The module also provides decision support interfaces, connecting with port scheduling, airport meteorological platforms, and environmental monitoring platforms to promptly push meteorological risk warnings and preventative measures.

[0139] The specific implementation steps are as follows:

[0140] (5.1) Real-time inference and rolling mechanism:

[0141] The system accesses real-time meteorological data streams and triggers a security engine building module inference every 10 minutes; it also aggregates and updates extreme weather event identification results, level assessments, and risk maps every 30 minutes.

[0142] (5.2) Output content design:

[0143] Risk Level Map: Displays the spatial distribution of event severity levels (mild to severe).

[0144] Probabilistic map: Displays the probability of an event occurring using contour lines or heat maps;

[0145] Event Timeline Plot: Shows the changes in the extreme event index over the next 24 hours.

[0146] (5.3) System Integration and Deployment:

[0147] Develop interfaces with port scheduling, airport meteorology, and environmental monitoring platforms; output API interface services to provide machine-to-machine (M2M) calling capabilities and enable connection with city and regional meteorological safety operation platforms.

[0148] Through the organic combination of the above modules, this system can achieve high-precision, real-time risk identification and early warning of small-scale severe weather in the low-altitude region of the Bohai Rim, providing effective technical support for regional meteorological safety.

Claims

1. An artificial intelligence system for low-altitude meteorological safety in the Bohai Rim region based on historical reanalysis and multi-mode data, characterized in that, Includes the following modules: (1) Meteorological variable reconstruction and extraction module: Historical reanalysis data and real-time observation data are fused from multiple sources to ensure consistent resolution across spatial and temporal distributions. Historical reanalysis data includes data acquired from ERA5, JRA-55, and MERRA-2, while real-time observation data includes ground station wind and temperature data, LiDAR-derived wind profiles, radar data, and microwave radiometer observations. Boundary layer height, wind shear, eddy current, inversion intensity, gradient, and disturbances are treated as meteorological variables. Interpolation methods are employed to improve the resolution and accuracy of these meteorological variables across temporal and spatial distributions, followed by dimensionality reduction and compression. (2) Extreme event identification module: Based on historical extreme meteorological events in the study area, a standardized label rule base is constructed, which includes pollution accumulation events, wind shear events, and local strong wind events. Historical extreme meteorological events are collected to establish a sample event set. Long short-term memory networks are used to reconstruct the observation data sequence of meteorological variables for error detection, which is used to identify extreme meteorological events. The observation data sequence of meteorological variables is the observation value of one or more meteorological variables at consecutive time points. A deep autoencoder is used to automatically detect and label historical extreme meteorological events; finally, a composite index of multiple meteorological variables is constructed to improve the sensitivity of extreme meteorological event identification. (3) Model training module: Based on the time series prediction model, the model is trained on time series data of historical meteorological variables and labeled historical extreme meteorological events; a combination of convolutional neural networks and recurrent neural networks is used to enhance the learning ability of local and long-term trends; a graph neural network is used to interact with meteorological models between modeling regions, with nodes representing modeling regions and edges representing the correlation between modeling regions, thereby enhancing the spatiotemporal data modeling capability; and an integrated framework of multiple models is constructed. The modeling region is a spatial unit or a uniformly divided high-resolution grid within the area to be studied, and each modeling region serves as a node in the graph neural network. Meteorological models model the spatiotemporal evolution of meteorological variables and their interactions within a modeling region. In graph neural networks, edges represent the similarity or physical transmission relationships between modeling regions. (4) Pattern fusion module: Perform unified spatiotemporal gridding processing on data from multiple sources; The processed output is averaged using a Bayesian model, and the results of multiple models in the model training module are fused using a weighted average method. The results are dynamically corrected using an error correction algorithm. (5) Security Engine Building Module: The security engine building block supports real-time rolling forecasts, updating inference results every 10 minutes to ensure rapid response to sudden extreme weather events; it also updates early warnings for extreme weather events every 30 minutes, providing the latest event identification results, level assessments, and risk maps. The security engine building module outputs risk level maps, probability heat maps, and event time sequence maps for extreme weather events, helping relevant departments monitor weather changes in real time. The safety engine building module provides a decision support interface, which connects with port scheduling, airport meteorological platforms and environmental monitoring platforms to promptly push meteorological risk warnings and preventive measures.

2. The Bohai Rim Low-Air Meteorological Safety Engine Artificial Intelligence System based on historical reanalysis and multi-mode data as described in claim 1, characterized in that, The specific implementation process of the meteorological variable reconstruction and extraction module is as follows: (1.1) Multi-source data fusion: Historical reanalysis data is uniformly resampled to a high-resolution grid; After quality control, the real-time observation data is integrated into the high-resolution grid. (1.2) Processing of meteorological variables: Boundary layer height: The mixed layer height diagnostic formula is fused with the ERA5 diagnostic value; Wind shear and vorticity: Calculate wind shear and vorticity index based on wind speed profiles at 925-850 hPa levels to identify low-altitude instability conditions; Inversion intensity: Calculate the near-surface temperature lapse rate and use strong inversion events as triggering conditions for pollution accumulation events or low visibility extreme events; Gradients and disturbances: The central difference method is used to construct temperature, humidity and wind speed gradient fields and extract local disturbance characteristics within the boundary layer; (1.3) Feature reduction and deep compression modeling: First, use multiple interpolation algorithms to spatially interpolate missing or irregular data points to ensure the integrity of meteorological variable data; Then, principal component analysis and empirical orthogonal function analysis are used to reduce the dimensionality of meteorological variables and extract the most representative features; Finally, a deep autoencoder is used to compress the high-dimensional meteorological data, extract potential high-order features, and reduce computational complexity. The high-dimensional meteorological data is a three-dimensional meteorological data tensor composed of multiple meteorological variables, multiple vertical height layers, and multiple time steps.

3. The Bohai Rim Low-Air Meteorological Safety Engine Artificial Intelligence System based on historical reanalysis and multi-mode data as described in claim 1, characterized in that, The specific implementation process of the extreme event identification module is as follows: (2.1) Construction of a standardized tag rule base: Based on historical extreme meteorological events in the sample event set, the following criteria are defined: Pollution accumulation event: BLH < 300m + wind speed < 2m / s + PM2.5 > 100μg / m³ + presence of inversion layer; Wind shear event: ΔV / ΔZ>0.015s⁻¹; where ΔV is the vertical height difference and ΔZ is the change in wind speed between the two height layers; Localized strong wind events: 10-minute wind speed > 12 m / s + turbulent kinetic energy > threshold; Historical extreme meteorological events are labeled in the sample event set to construct training and validation sets; (2.2) Integration of anomaly detection methods: LSTM reconstruction error method: The Long Short-Term Memory (LSTM) network is trained with normal time period data. When the input data of the LSM network includes data of extreme meteorological events, the error between the actual observation value and the predicted value of the LSM network increases significantly. The normal time period data refers to time series data that reflect the normal change pattern of meteorological variables collected during periods when no extreme meteorological events have occurred. Deep autoencoder identification: Training a deep autoencoder to compress and reconstruct meteorological variables; a significant increase in reconstruction error indicates a candidate extreme meteorological event. Local outlier factor: used for outlier detection in a low-dimensional compressed space; where the low-dimensional compressed space is the space after dimensionality reduction by a deep autoencoder or PCA. (2.3) Design of multivariate composite indicators: Construct mathematical models for Pollution Weather Potential Index (PAI) and Wind Shear Index (WSI); The mathematical model for the Pollution Weather Potential Index is as follows: PAI = w1 * BLH −1 +w2*wind speed −1 +w3* Temperature inversion intensity +w4* Humidity stability; Where w1-w4 are the weights of the corresponding meteorological variables; The mathematical model for the wind shear index (WSI) is as follows: ; The severity of extreme meteorological events is classified into four categories: light, moderate, severe, and extremely severe, using fuzzy logic rules or decision tree methods.

4. The Bohai Rim Low-Air Meteorological Safety Engine Artificial Intelligence System based on historical reanalysis and multi-mode data as described in claim 1, characterized in that, The specific implementation process of the model training module is as follows: (3.1) Integration framework of multiple models: The integrated framework of multiple models is a multi-model collaborative spatiotemporal feature learning system. This system comprises five modules operating in parallel: a local time series modeling module, a key variable-driven modeling module, a local spatial structure extraction module, a regional collaborative modeling module, and a convolutional + recurrent hybrid temporal modeling module. Details are as follows: (a) Local time series modeling module, select LSTM or GRU: Input: Time-series data composed of meteorological variables; Task: To study the historical evolution patterns of meteorological variables and predict the continuous value changes over the next N hours; Output: Probability score of extreme weather events; (b) Key variable-driven modeling module, select Transformer or TFT: Input: Time-series data composed of meteorological variables and the weighting mechanism of meteorological variables; Task: Use attention mechanisms to identify which meteorological variables are more likely to drive extreme weather events in the current period; Output: Driver attention weights; (c) Local spatial structure extraction module, which is CNN-LSTM: Input: An image composed of meteorological variables in a high-resolution grid; Task: Use CNN to extract spatial gradient information and shape features; then use LSTM to further process the temporal evolution trend. Output: Spatial gradient recognition score; (d) Regional collaborative modeling module, select GCN or GAT: Input: Construct a regional graph structure. The regional graph structure abstracts each modeling region into a node in a graph neural network within the scope of the study, and represents the relationship between the modeling regions through the edges between the nodes. The modeling region is a divided spatial unit or a uniformly divided high-resolution grid. Each modeling region node contains boundary layer height, wind speed, temperature and humidity, inversion intensity, and vortex feature vector. Task: Utilize graph convolutional networks or graph attention networks to mine the risk propagation relationships and mutual influence mechanisms between modeled regions on the regional graph structure; Output: Risk score for the coordinated propagation of modeling regions, used to describe the spatial diffusion trend and risk linkage intensity of extreme meteorological events among different modeling regions; (e) Convolutional + Recurrent Hybrid Temporal Modeling Module, select CNN-RNN: Input: Time series data of historical meteorological variables, either a single meteorological variable or a combination of multiple meteorological variables; Task: Use CNN to extract local pattern features from time series data of meteorological variables, and then use RNN to capture long-term dependencies and trends; Output: The fused time-series prediction results are used to identify precursor features of extreme meteorological events; (3.2) Multi-model integration method: (a) Multi-model ensemble structure: First-level model: The outputs of the five modules in step (3.1) are combined at the feature level to form an integrated feature vector, which is then standardized. Secondary fusion model: Uses logistic regression, random forest or XGBoost to fuse the output of the primary model; or uses a deep stacking neural network to learn the combined weights of the primary model output. (b) Output content: Probability values ​​for extreme meteorological events: output probability of pollution accumulation, probability of local strong winds, and probability of wind shear; Uncertainty estimation: The standard deviation and consistency between outputs are calculated using an ensemble method of multiple models, thereby providing a confidence range; Event severity prediction results: Based on the comprehensive score after multi-model fusion, combined with rules and fuzzy logic, four levels of event severity are output: mild, moderate, severe, and extremely severe. Key factor explanation information: Tracing back the weights and contributions of important features in multi-model fusion to form a traceable decision path; (c) Deployment method: Based on multi-model deployment as a scalable model API interface, it can receive multi-variable inputs at any time period and automatically complete inference, judgment and explanation output.

5. The Bohai Rim Low-Air Meteorological Safety Engine Artificial Intelligence System based on historical reanalysis and multi-mode data as described in claim 1, characterized in that, The specific implementation process of the pattern fusion module is as follows: (4.1) Spatiotemporal standardization processing: Establish a unified spatial grid and temporal resolution to resample multi-source data; (4.2) Bayesian model average fusion: The outputs of multiple models are weighted and averaged, with the weights adaptively updated based on the performance of each model during the historical validation period; a probability density function for the fused extreme weather events is provided. (4.3) Dynamic error correction: Short-term forecast biases are corrected using dynamic linear models or Kalman filtering methods; long-term future climate is aligned with current observations using a bias correction function.

6. The Bohai Rim Low-Air Meteorological Safety Engine Artificial Intelligence System based on historical reanalysis and multi-mode data as described in claim 1, characterized in that, The specific implementation process of the security engine building module is as follows: (5.1) Real-time inference and rolling mechanism: It accesses real-time meteorological data streams and triggers a security engine building module inference every 10 minutes; it aggregates and updates the identification results, level assessments, and risk maps of extreme weather events every 30 minutes. (5.2) Output content design: Risk level map: Displays the spatial distribution of event levels; Probability heatmap: Displays the probability of an event occurring using contour lines or heat maps; Event timeline graph: Shows the changes in the extreme event index over the next 24 hours; (5.3) System Integration and Deployment: Develop interfaces with port scheduling, airport meteorology, and environmental monitoring platforms; output API interface services to provide machine-to-machine calling capabilities and enable connection with city and regional meteorological safety operation platforms.

Images