Heavy fog visibility prediction system and method based on deep learning

Abstract

The invention discloses a heavy fog visibility prediction system and method based on deep learning, and relates to the technical field of data processing, and the method comprises the steps: determining a particle size modal category representing a heavy fog particle size distribution type; screening a historical particle size modal sequence matched with the particle size modal category of the current region based on an internal corresponding relation between the meteorological environment state and the particle size modal category; based on the category transfer complexity reflected by the transfer path, performing sequence reconstruction on typical particle size modal categories in the historical particle size modal sequence; inputting the reconstructed historical particle size modal sequence and the current particle size modal distribution characteristics into a long short-term memory network model optimized by a particle size modal attention mechanism, and outputting a predicted heavy fog visibility grade of the current region at a future moment; according to the method, the fog particle size modal physical characteristics and the deep learning model are fused, so that the fog visibility grade prediction accuracy and reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically a fog visibility prediction system and method based on deep learning. Background Technology

[0002] Accurate visibility forecasting during heavy fog is crucial for the normal operation of transportation, safety production, aviation, and maritime sectors. Most existing fog visibility forecasting methods are based on statistical or numerical simulation models of traditional meteorological parameters, relying primarily on conventional meteorological observation data such as temperature, humidity, and wind speed. These methods fail to adequately consider the microphysical characteristics and dynamic evolution mechanisms of fog particle size modes, resulting in poor adaptability to changes in fog type and limited accuracy in practical applications.

[0003] In recent years, with the development of lidar technology, particle size distribution feature extraction based on lidar backscattering spectrum has been gradually applied to the field of fog monitoring. However, existing solutions usually only perform simple peak identification on the backscattering data obtained by lidar, without conducting in-depth analysis on the long-term stability of particle size mode distribution and its thermodynamic formation mechanism. This makes it difficult to effectively identify different particle size mode types and their stable existence conditions, resulting in difficulty in accurately predicting the development trend of fog and the corresponding visibility level in the future.

[0004] Furthermore, although existing technologies have attempted to use deep learning models for visibility prediction, most of these models directly use standard structures such as traditional Long Short-Term Memory (LSTM) networks or Recurrent Neural Networks (RNNs), without incorporating the evolution of fog particle size modal features in time series and their physical generation mechanisms. This makes it difficult to accurately capture the impact of particle size modal features on future visibility changes, resulting in poor generalization performance of the prediction models under complex environmental conditions, and the prediction accuracy still needs to be improved. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a fog visibility prediction method based on deep learning.

[0006] To achieve the above objectives, in one aspect, the present invention provides a method for predicting visibility in heavy fog based on deep learning, comprising:

[0007] Based on the backscattering intensity spectrum detected by lidar, the particle size mode distribution characteristics of the current area are extracted to determine the particle size mode category that characterizes the particle size distribution type of fog.

[0008] Based on the measured humidity and aerosol concentration data, the current meteorological environment status of the region is determined; and based on the inherent correspondence between the meteorological environment status and particle size mode category, historical particle size mode sequences that match the particle size mode category of the current region are selected.

[0009] Based on the transition patterns of mode categories in historical particle size modal sequences, the transition paths of particle size modal categories are extracted; and based on the complexity of category transitions reflected by the transition paths, the typical particle size modal categories in historical particle size modal sequences are reconstructed.

[0010] The reconstructed historical particle size mode sequence and the current particle size mode distribution features are input into a long short-term memory network model optimized by a particle size mode attention mechanism, and the output is the predicted fog visibility level for the current region at future times.

[0011] On the other hand, the present invention provides a fog visibility prediction system based on deep learning, implemented based on the aforementioned deep learning-based fog visibility prediction method, including:

[0012] The modality recognition module is used to extract the particle size modality distribution characteristics of the current area based on the backscatter intensity spectrum detected by lidar, and to determine the particle size modality category that characterizes the particle size distribution type of fog.

[0013] The sequence filtering module is used to determine the meteorological environment status of the current area based on the measured humidity and aerosol concentration data; and based on the inherent correspondence between the meteorological environment status and particle size mode category, to filter historical particle size mode sequences that match the particle size mode category of the current area.

[0014] The path reconstruction module is used to extract the transfer paths of particle size modes based on the transfer rules of modes in the historical particle size mode sequence; and to reconstruct the sequence of typical particle size modes in the historical particle size mode sequence based on the class transfer complexity reflected by the transfer paths.

[0015] The visibility prediction module is used to input the reconstructed historical particle size mode sequence and the current particle size mode distribution features into a long short-term memory network model optimized by the particle size mode attention mechanism, and output the predicted fog visibility level of the current area at future time.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] This invention extracts the particle size mode distribution characteristics of the current region based on the backscattering intensity spectrum of lidar, and combines the stability of particle size peak within a continuous time window with the probability determination of nonlinear thermodynamic stability. This effectively avoids the problem of misjudgment of particle size modes caused by ignoring the dynamic change characteristics of particle size in traditional methods, and significantly improves the accuracy and robustness of particle size mode category identification in fog.

[0018] This invention overcomes the problem of insufficient historical sequence adaptability caused by existing technologies that do not consider the physical formation mechanism of historical data by strictly screening and reconstructing historical particle size mode sequences that match the current particle size mode category based on the physical causal relationship between the current regional meteorological environment state and particle size mode category, thus ensuring that the historical data used for visibility prediction has good physical rationality and consistency.

[0019] This invention inputs the reconstructed historical particle size modality sequence and the current particle size modality distribution characteristics into a long short-term memory network model optimized by a particle size modality attention mechanism. It comprehensively considers the spatial characteristics and long-term temporal change trends of particle size modality categories, solves the problem that existing prediction models cannot accurately capture the evolution characteristics of particle size modality categories, resulting in insufficient prediction accuracy, and achieves a more accurate and reliable prediction effect for visibility levels in heavy fog. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart of the method of the present invention;

[0022] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 The first aspect of this invention provides a method for predicting visibility in heavy fog based on deep learning, comprising:

[0025] S101: Based on the backscattering intensity spectrum detected by lidar, extract the particle size mode distribution characteristics of the current area and determine the particle size mode category that characterizes the particle size distribution type of fog.

[0026] It should be noted that the particle size modal distribution characteristics described in this invention refer to the particle size interval concentration and peak particle size stability characteristics exhibited by atmospheric fog droplet particle size distribution within a specific region, used to quantitatively characterize different types of fog (such as single-mode, dual-mode, or multi-mode). The particle size modal category corresponds to the classification result of different numbers of concentrated peak regions of particle size. For example, when the particle size modal distribution characteristics in the region exhibit a single stable peak particle size interval, it is defined as a single-mode category; when it exhibits two stable peak particle size intervals, it is defined as a dual-mode category; and so on.

[0027] For example, a specific implementation of extracting the particle size mode distribution characteristics of the current region based on the backscattering intensity spectrum detected by lidar includes:

[0028] A lidar emits laser pulses of a specific wavelength into the atmosphere. The laser pulses are backscattered by atmospheric particles and return to the receiver. By measuring the intensity of the received backscattered light signal and establishing a mathematical relationship between atmospheric particle size and scattering intensity based on Mie scattering theory, a continuous spectrum of backscattering intensity as a function of particle size can be obtained, which is the backscattering intensity spectrum.

[0029] Specifically, the backscattering intensity spectrum can be expressed as a functional relationship: In the formula, D is the particle size; This represents the backscattering intensity corresponding to the particle size;

[0030] Furthermore, based on the backscattering intensity spectrum obtained above, the peak positions of each particle size range on the spectrum are identified by the local extremum determination algorithm to determine the corresponding peak particle size.

[0031] It should be noted that the particle size range is a preset, non-continuous division of particle size ranges, and each particle size range contains only a single peak particle size. For example, the particle size range can be... Divided into multiple particle size ranges, such as And within each interval, a unique peak particle size is identified and recorded.

[0032] In a specific implementation, determining the particle size mode category characterizing the particle size distribution type of fog includes:

[0033] Based on the backscattering intensity spectrum detected by lidar, the position of the spectral peak in each particle size range is identified, and the corresponding peak particle size is determined.

[0034] In practice, the first derivative determination method is applied to the backscattering intensity spectrum to determine the position where the derivative on the spectral line is zero, and the derivative changes positive and negative before and after the position, thereby identifying the position as a local maximum, i.e. the spectral peak position, and the particle size corresponding to the position is recorded as the peak particle size.

[0035] For example, the particle size range is set as Then the peak particle size in this range satisfy:

[0036]

[0037] The above method is used to obtain the set of peak particle sizes corresponding to each particle size range.

[0038] The candidate dominant particle size range is determined based on the stable occurrence frequency of peak particle size within a continuous time window;

[0039] It should be noted that, in order to accurately determine the stability of the particle size modal distribution, this step uses multiple lidar measurements within a continuous time window (e.g., 10 minutes per window, with 3 consecutive sampling windows for a total of 30 minutes), counts the number of times each peak particle size appears in all windows, and calculates the relative frequency. The frequency calculation method is as follows:

[0040]

[0041] In the formula: The number of times the peak particle size appears in all measurement windows; The total number of windows for measurement;

[0042] The particle size range corresponding to the peak particle size with a relative frequency greater than a preset stability threshold (e.g., 0.8) is determined as the candidate dominant particle size range.

[0043] Based on the nonlinear thermodynamic threshold condition of the hygroscopic growth and dissipation of fog particles, the probability of stable existence of candidate dominant particle size ranges under the current environmental humidity conditions is calculated.

[0044] It should be understood that the size of fog droplets in the atmosphere is not constant, but changes with the change of ambient humidity through hygroscopic growth or evaporation. This process has a clear thermodynamic threshold condition. This invention uses nonlinear thermodynamic relationships to accurately calculate the stable existence probability of candidate dominant particle size ranges, thereby eliminating the influence of occasional peak particle size on the determination of particle size mode distribution characteristics.

[0045] In specific implementation, the nonlinear thermodynamic threshold condition can be used The equation is expressed as:

[0046]

[0047] In the formula, S represents supersaturation, i.e., water vapor partial pressure. With saturated water vapor pressure The ratio; Indicates the droplet size; parameter These are related to droplet surface tension, ambient temperature, and aerosol solubility characteristics, and can be determined experimentally based on actual environmental conditions. For example, under typical conditions (such as an ambient temperature of 10℃), It usually takes a value close to 1. Then take (Specific values ​​were obtained through environmental experiments).

[0048] Furthermore, by measuring the actual humidity conditions of the current environment, the supersaturation of the current environment is obtained. The current environmental supersaturation is compared with the theoretical critical supersaturation threshold corresponding to the peak particle size within the particle size range calculated by the Köhler equation above. Comparisons were made to determine the probability of stable existence of each candidate dominant particle size range. The calculation formula is:

[0049]

[0050] In the formula, This is an environmental humidity fluctuation parameter, which can be obtained through statistical methods based on historical environmental humidity fluctuation data observed in actual environments. The typical value range is 0.01 to 0.1 (the specific value is obtained through actual environmental measurement data).

[0051] It should be noted that the range of values ​​for the above-mentioned stable existence probability is [missing information]. The closer the probability is to 1, the stronger the thermodynamic stability of the peak particle size in that particle size range under the current environmental humidity conditions, and vice versa.

[0052] Based on the stable existence probability satisfying the preset probability threshold condition, the target particle size range that satisfies thermodynamic stability is determined, and based on the peak particle size correspondence of the target particle size range, the particle size mode category characterizing the fog particle size distribution type is determined.

[0053] In specific implementation, the preset probability threshold condition is determined through experimental data, for example, 0.85, and is used to determine whether each candidate dominant particle size interval has the thermodynamic conditions for long-term stable existence; when the stable existence probability of a certain candidate dominant particle size interval is greater than this probability threshold, the particle size interval is determined as the target particle size interval.

[0054] It should be understood that the number of target particle size intervals obtained in this step determines the type of particle size mode category; for example, when there is only one target particle size interval, the particle size distribution type of the fog is determined to be single-mode; when there are two target particle size intervals, it is determined to be dual-mode; and so on.

[0055] For example, when the stable existence probability calculated above is 0.88 (particle size range) ) and 0.90 (particle size range) When the above two particle size ranges are simultaneously the target particle size ranges, the particle size mode category is determined to be a dual-mode type based on a preset probability threshold condition (e.g., 0.85).

[0056] S102: Based on the measured humidity and aerosol concentration data, determine the current meteorological environment status of the area; and based on the inherent correspondence between the meteorological environment status and particle size modal category, screen historical particle size modal sequences that match the particle size modal category of the current area.

[0057] It should be noted that the meteorological environment state mentioned in this step refers to the combined characteristics of environmental humidity conditions and aerosol concentration that have a significant impact on the formation and maintenance of fog. The meteorological environment state can be classified into typical meteorological conditions, such as "high humidity and high concentration", "high humidity and low concentration", "low humidity and high concentration" and other types. There is a clear physical causal relationship between the particle size modality category and the above-mentioned meteorological environment state. For example, a high humidity and high concentration environment usually corresponds to a particle size modality category with larger particle size and significant condensation growth.

[0058] In a specific implementation, the step of filtering historical particle size mode sequences that match the current region's particle size mode category includes:

[0059] Based on the measured humidity and aerosol concentration data, determine the typical meteorological conditions to which the current regional meteorological environment belongs;

[0060] In practice, the humidity and aerosol concentration data collected in the current area are first analyzed, and a quantitative threshold classification method is used to define typical meteorological conditions.

[0061] For example, meteorological conditions can be defined as follows based on experimental statistical data:

[0062] When relative humidity And aerosol concentration At that time, the current meteorological environment was classified as "high humidity and high concentration";

[0063] When relative humidity And aerosol concentration At that time, the current meteorological environment was classified as "high humidity and low concentration";

[0064] When relative humidity And aerosol concentration At that time, the current meteorological environment was classified as "medium humidity and high concentration";

[0065] Using the methods described above, the typical meteorological conditions of the current area can be clearly defined based on the measurement data.

[0066] Based on the causal correspondence between typical meteorological conditions and the physical formation mechanism of particle size modal categories, the formation path of particle size modal categories under typical meteorological conditions is established;

[0067] It should be understood that the formation path of particle size modal categories refers to the continuous physical process that fog particle size modal categories undergo from the initial stage to the stable formation stage under specific typical meteorological conditions. This process includes three stages: condensation nucleus activation, hygroscopic growth, and particle size distribution stabilization. It has clear stage-specific physical characteristics and continuity requirements.

[0068] For example, when the typical meteorological conditions are "high humidity and high concentration", the formation path of the corresponding particle size mode category is specifically manifested as follows:

[0069] Initial stage: Aerosol particles are affected by the high concentration environment, and a large number of condensation nuclei are activated, forming the initial particle size mode;

[0070] Growth stage: Driven by a high humidity environment, the initial particle size rapidly absorbs moisture and grows, forming a clear particle size growth trend;

[0071] Stable phase: After particle size growth reaches thermodynamic equilibrium, the particle size mode stabilizes within a relatively large particle size range (e.g., ...). );

[0072] Based on the above mechanism, a path model for particle size modal category formation corresponding to typical meteorological conditions is established.

[0073] Based on the physical continuity of the formation path, candidate historical sequences that match the physical formation path of the current particle size mode category are selected from the historical observed particle size mode sequences;

[0074] Specifically, the screening of candidate historical sequences that match the physical formation path of the current particle size modality category includes:

[0075] Based on the hygroscopic growth and condensation processes of particles corresponding to different particle size modal categories, the formation stage of the particle size modal category in the historical observation sequence is determined;

[0076] It should be noted that the formation stage of particle size mode category is defined as follows:

[0077] Initial stage: The peak particle size appears stably for the first time and lasts for more than the preset duration (e.g., 10 minutes).

[0078] Growth phase: The peak particle size of the particle size modality category increases significantly in subsequent time windows (e.g., the increase exceeds...). );

[0079] Stable phase: The peak particle size remains within a stable range and the variation is less than a preset range (e.g., );

[0080] Based on the above definition, the particle size modal categories in the historical observation sequence are labeled and classified according to their formation stages.

[0081] Based on the physical matching relationship between the formation stage of particle size modal categories and typical meteorological conditions, a stage transition path between each particle size modal category in the historical sequence is established;

[0082] In practice, based on continuous observation data from each stage in historical data, a stage transition probability matrix M for particle size modality categories is established, and its elements are defined as follows: In the formula, Let be the probability that the particle size mode category transitions from stage i to stage j in the historical sequence. This represents the frequency of the actual observed transition from stage i to stage j. The sum of all transition frequencies in stage i; the stage transition path is quantified using the matrix described above.

[0083] Based on the physical transfer path of the particle size modality formation stage and the physical process corresponding to the formation stage of the particle size modality in the current region, candidate historical sequences with matching stage transfer paths are screened.

[0084] It should be noted that in practical prediction applications, the "currently forming stage sequence" mentioned in this step refers to the clearly observed stage transition sequence within a certain time window tracing back from the currently observed stage node (e.g., The specific comparison method is as follows:

[0085] If the currently observed formation stage sequence is Then, from the historical observation sequence, all sequences ending at stage C and moving forward through stages are selected. Historical sequences that are completely identical or highly similar (e.g., similarity greater than 90%, defined as having the same stage order and a difference of less than 10% in the duration of each stage) are identified as candidate historical sequences.

[0086] The candidate historical sequence set is obtained through the above method.

[0087] Based on the completeness of the particle size modality formation process in the candidate historical sequence of the stage transition path matching, a candidate historical sequence matching the physical formation path of the current particle size modality is determined.

[0088] It should be noted that this step quantitatively assesses the integrity of the formation process of candidate historical sequences, thereby further screening historical sequences that highly match the physical formation path of the current particle size modality category.

[0089] In specific implementation, the integrity evaluation of the particle size modality formation process includes the following criteria:

[0090] The initial stage, growth stage and stable stage of particle size modality in the candidate historical sequence should all be clearly present. Sequences lacking any of the three key stages are considered to have an incomplete formation process.

[0091] Each phase should meet the minimum duration requirement (e.g., initial phase no less than 10 minutes, growth phase no less than 20 minutes, and stabilization phase no less than 30 minutes); phases with a duration shorter than the set threshold will be considered incomplete.

[0092] The sequence should exhibit a clear physical transfer trend between its stages, such as a smaller peak particle size in the initial stage, a significant increase in peak particle size in the growth stage, and a relatively stable peak particle size within a range in the steady stage (e.g., a fluctuation range of less than 100 mm). );

[0093] Specifically, each sequence in the candidate historical sequence set is evaluated according to the three criteria mentioned above, and the number of sequences that meet the criteria is calculated; the final integrity score is then calculated. The calculation formula is: In the formula, The value ranges from 0 to 1; when the integrity score reaches a preset integrity threshold (e.g., 0.8), the historical sequence is identified as a candidate historical sequence that matches the physical formation path of the current particle size modality category.

[0094] Based on whether the formation path of the candidate historical sequence conforms to the physical generation mechanism of fog particle size mode, determine the historical particle size mode sequence that matches the current region particle size mode category;

[0095] It should be noted that this step performs a final physical generation mechanism consistency determination on the candidate historical sequences to ensure that the selected historical particle size mode sequences have clear physical rationality and interpretability.

[0096] In practice, the conditions for determining the consistency of the physical generation mechanism of particle size modes include, but are not limited to:

[0097] The environmental humidity conditions required for each formation stage corresponding to the particle size modality should be consistent with the humidity conditions observed simultaneously in the historical sequence (e.g., the humidity required for the initial stage). Humidity required during the stabilization phase );

[0098] The particle size growth trend at each formation stage should show a positive correlation with the aerosol concentration change trend during the same period. For example, the increase in particle size during the growth stage should be significantly synchronized with the increase in aerosol concentration.

[0099] For example, the evaluation function for the consistency of physical generation mechanisms is defined as follows:

[0100]

[0101] In the formula, A consistency score is assigned to the physical generation mechanism, with a value ranging from 0 to 1; This represents the total length of time in the historical sequence where humidity conditions meet the humidity requirements of the above-mentioned stages. This represents the total observation duration of the historical sequence. The correlation coefficient (e.g., Pearson correlation coefficient) represents the relationship between the peak particle size growth trend and the aerosol concentration change trend. The value ranges from 0 to 1, and the closer the value is to 1, the higher the correlation.

[0102] It should be understood that when the physical generation mechanism consistency score is... When the value exceeds a preset threshold (e.g., 0.75), the candidate historical sequence is finally determined as the historical particle size mode sequence that matches the current region particle size mode category, for use in subsequent sequence reconstruction and prediction.

[0103] S103: Based on the transfer pattern of mode categories in the historical particle size mode sequence, extract the transfer path of particle size mode category; and based on the complexity of category transfer reflected by the transfer path, reconstruct the sequence of typical particle size mode categories in the historical particle size mode sequence;

[0104] It should be noted that the transfer path of particle size modality categories mentioned in this step refers to the time series transfer relationship of different particle size modality categories in the historical particle size modality sequence as meteorological environmental conditions change. For example, the transfer process of particle size modality categories from single-mode to dual-mode, dual-mode to single-mode, or between single modes in different particle size ranges. This step ensures that the subsequent sequence reconstruction has clear physical continuity and rationality by extracting and analyzing the transfer paths of historical particle size modality sequences.

[0105] In a specific implementation, the step of extracting the transfer path of particle size mode categories based on the transfer patterns of mode categories in historical particle size mode sequences includes:

[0106] For a historical particle size mode sequence, the particle size mode categories observed at consecutive times in the sequence are defined as follows:

[0107]

[0108] In the formula, Indicates the first The observed particle size modal category (e.g., single-mode, dual-mode, multi-mode, etc.) is determined at each observation time. Further, the number of particle size modal category transitions between adjacent observation times is counted, forming a statistical matrix of particle size modal category transitions, specifically expressed as:

[0109]

[0110] In the formula: This represents the probability that the particle size modality category changes from the i-th type to the j-th type; This represents the number of times the particle size modality category shifts from type i to type j in the historical observation sequence, with the denominator being the sum of the total number of shifts from type i to all types.

[0111] It is understandable that, through the aforementioned transition probability statistical matrix It can quantify the transfer patterns of historical particle size mode categories, and then extract the transfer paths.

[0112] In a specific implementation, the step of reconstructing the typical particle size mode categories in the historical particle size mode sequence includes:

[0113] Based on the transfer paths of mode categories in the historical particle size mode sequence, determine the physical evolution process of fog particle size modes corresponding to each transfer path;

[0114] It should be noted that the physical evolution process of particle size modes refers to the entire process from initial formation to stabilization or dissipation, which is determined by the physical formation mechanism of particle size mode categories. In this invention, it includes processes such as initial particle agglomeration, particle size growth and stabilization, and particle size dissipation.

[0115] In specific implementation, for a given transfer path, such as the transfer from a single mode (small particle size) to a dual mode (including larger particle size), this invention determines the physical evolution process corresponding to the transfer path based on the microphysical processes of particle condensation nucleus activation and hygroscopic growth. Typical steps are as follows:

[0116] Initial stage of particle size mode: Small particle size mode corresponds to the initial activation stage of condensation nuclei, and the particle size range is generally small (e.g., );

[0117] Particle size growth stage: Affected by the increase in environmental humidity, the condensation nuclei absorb moisture and grow rapidly, resulting in larger particle size modal peaks;

[0118] Particle size stabilization stage: reaching a specific particle size range (e.g.) When the particle size growth stabilizes, a bimodal characteristic is formed;

[0119] By using the above method, the physical evolution process of particle size modes corresponding to each transfer path is determined one by one to ensure the rationality of subsequent reconstruction.

[0120] Based on the thermodynamic criteria for particle size growth and dissipation during the physical evolution of particle size modes, the categories of abnormal particle size modes in the transfer path are identified;

[0121] It should be understood that the abnormal particle size mode categories described in this invention refer to particle size mode categories that appear briefly or abnormally during the physical evolution of particle size modes due to failure to meet thermodynamic criteria; these abnormal modes are usually characterized as particle size mode categories that are discontinuous in the physical evolution path or have a short duration and are not supported by clear environmental conditions.

[0122] Specifically, the identification of abnormal particle size modal categories in the transfer path includes:

[0123] Determine the dominant particle size range and the thermodynamic critical humidity required for hygroscopic growth of particles corresponding to the particle size mode categories in the historical particle size mode sequence.

[0124] Specifically, the dominant particle size range is determined based on the particle size modality category definition (e.g., single-mode or dual-mode); for example, for the single-mode category, its dominant particle size range is... or etc.; further, through The equation determines the thermodynamic critical humidity required for stable hygroscopic growth of particles within this particle size range, which is the theoretical critical supersaturation threshold for the corresponding particle size. .

[0125] Based on the humidity evolution trajectory of the corresponding particle size modality in the region during the historical observation period, the percentage of time that the humidity conditions for the corresponding particle size modality meet the critical humidity is extracted.

[0126] In practical implementation, the formula for calculating the percentage of time when humidity conditions meet the critical humidity level is defined as follows:

[0127]

[0128] In the formula: This refers to the ambient humidity (expressed as supersaturation) reaching or exceeding the critical humidity threshold during the historical observation period. The cumulative length of time, This represents the total duration of the historical observation period.

[0129] Based on the time proportion and the preset humidity threshold condition for stable and continuous existence of particle size modal categories, particle size modal categories that do not meet the stable humidity condition are identified and determined as abnormal particle size modal categories.

[0130] Specifically, the stable and persistent humidity threshold condition is determined by historical experimental observations, for example, 0.75. That is, if the proportion of time during which the humidity of a certain particle size modality meets the critical threshold during the historical observation period is less than 0.75, then the particle size modality is determined to be an abnormal category.

[0131] For example, if single-mode (particle size) If the percentage of time during which the humidity meets the critical condition during the historical observation period is only 0.6, which is significantly less than the threshold of 0.75, then this particle size mode category is identified as an abnormal particle size mode category.

[0132] Based on the discontinuity of the physical formation mechanism corresponding to the abnormal particle size mode category, the abnormal particle size mode category is sequentially eliminated;

[0133] It should be noted that the purpose of this step is to remove abnormal particle size mode categories that lack physical rationality in order to ensure the physical continuity and reliability of the historical particle size mode sequence. In practice, particle size mode categories marked as abnormal in the historical particle size mode sequence are directly deleted.

[0134] Based on the particle size mode sequence after removing abnormal particle size mode categories, the typical particle size mode category sequence is reconstructed under the condition of continuity in the physical evolution process of fog particle size modes, and the reconstructed historical particle size mode sequence is output.

[0135] It should be noted that the reconstruction of the typical particle size mode category sequence described in this step is based on ensuring that the particle size mode category has a clear and continuous thermodynamic stable state and particle size growth trend during the physical formation and evolution process, so as to avoid prediction errors caused by the existence of abnormal categories in the sequence.

[0136] In practice, the reconstruction of typical particle size mode category sequences includes the following:

[0137] First, the remaining particle size modal categories in the historical particle size modal sequence after removing abnormal particle size modal categories as described above are reordered. The order is based on the actual physical evolution stage of the particle size modal category, namely the initial stage, the growth stage, and the stable stage.

[0138] Secondly, based on the requirement for continuity of the dominant particle size range between adjacent stages in the reconstructed particle size mode category sequence, specifically, the upper limit of the dominant particle size in the previous stage must be continuous or slightly overlapped with the lower limit of the dominant particle size in the next stage (e.g., the particle size difference is less than or equal to...). This ensures the smoothness of the physical evolution process.

[0139] For example, when the initial stage particle size mode category dominates the range of the reconstructed sequence, At that time, the dominant particle size range in subsequent growth stages should be adjacent to or slightly overlap with the range of the grain size. In order to maintain a continuous evolutionary trend.

[0140] Finally, the reconstructed historical sequence of particle size modality category sequence is labeled and output as the data basis for subsequent visibility prediction model input.

[0141] S104: Input the reconstructed historical particle size mode sequence and the current particle size mode distribution features into the long short-term memory network model optimized by the particle size mode attention mechanism, and output the predicted fog visibility level of the current area at future time.

[0142] It should be noted that the Long Short-Term Memory (LSTM) network model optimized by the particle size modality attention mechanism described in this step is a dedicated neural network model built by leveraging the advantages of deep learning technology in time series prediction problems and combining it with the physical characteristics of fog particle size modality. By precisely capturing the intrinsic relationship between historical sequences and current particle size modality features, it obtains accurate predictions of fog visibility at future times.

[0143] Specifically, the long short-term memory network model optimized by the particle size modal attention mechanism includes:

[0144] The particle size modal attention coding layer includes a bidirectional multi-head self-attention sublayer based on the physical transfer characteristics of particle size modal categories;

[0145] It should be understood that this layer contains a bidirectional multi-head self-attention sublayer based on the physical transfer characteristics of particle size modal categories, which is used to capture the dynamic interaction relationships between various modal categories in the historical particle size modal sequence and enhance the contribution weight of key particle size modal categories in the historical sequence to subsequent predictions.

[0146] For example, the specific processing procedure of the bidirectional multi-head self-attention sub-layer is as follows:

[0147] First, construct the input historical particle size mode sequence. and The three matrices represent different dimensions of particle size modal features in the historical sequence.

[0148] Next, the attention weight matrix A is calculated using the self-attention mechanism:

[0149]

[0150] In the formula: Represents the vector dimension. The function ensures that the sum of the weights is 1;

[0151] Then, through a multi-head attention mechanism, multiple different sub-attention weight matrices are used to simultaneously focus on the relationship between different granularity modal categories of the sequence, and finally merged into a whole attention encoding output sequence.

[0152] The cross-modal information interaction layer includes dynamic routing capsule units for fusing historical particle size modal sequences with current particle size modal distribution characteristics;

[0153] It should be noted that this layer is used to fuse historical particle size modality sequence features with current particle size modality distribution features to ensure that the model can fully consider the impact of the current real-time observed particle size modality state on future prediction results; this layer includes dynamic routing capsule units, which dynamically capture the feature matching and interaction relationship between historical sequences and current modalities through iteration.

[0154] For example, the specific implementation process of the dynamic routing capsule unit includes:

[0155] First, the historical sequence features and the current particle size modal distribution features are input into the capsule unit to form an initial vector set;

[0156] Then, through an iterative routing mechanism, the coupling coefficient between different vector sets is calculated, and the fusion vector is iteratively updated to finally output a cross-modal fusion sequence.

[0157] The particle size modality time-dependent capture layer comprises at least three layers of a hybrid layer structure consisting of convolutionally gated recurrent units (ConvGRU) and long short-term memory network units alternately connected;

[0158] It is understood that the particle size modality temporal dependence capture layer described in this embodiment is a hybrid deep neural network structure specifically designed for the temporal variation characteristics of fog particle size modal features. This structure includes a hybrid layer structure with at least three layers of convolutionally gated recurrent units (ConvGRU) and long short-term memory network units (LSTM) connected alternately. The convolutionally gated recurrent unit (ConvGRU) integrates a convolutional structure on the basis of the traditional gated recurrent unit (GRU). Specifically, a one-dimensional convolution operation is introduced in the input and state update process of the gate unit to effectively capture local spatial features in the sequence data, while retaining the GRU gating mechanism to enhance the ability to capture changes in time series features. The long short-term memory network unit (LSTM) further captures the particle size modality evolution trend over a longer time scale.

[0159] For example, the structure of this layer is as follows:

[0160] The first layer is a convolutional gate unit, which extracts local features from the input cross-modal fusion sequence through one-dimensional convolution;

[0161] The second layer is a long short-term memory network unit, which inputs the local features extracted by convolution into the LSTM unit to capture temporal dependencies;

[0162] The above units are repeated alternately at least three times to ensure that particle size modal features are effectively captured in both the local space and long-term temporal dependence dimensions.

[0163] The sequence prediction output layer includes a linear mapping unit and a nonlinear thresholding unit;

[0164] It should be understood that the sequence prediction output layer is used to map the high-dimensional feature vector processed by the particle size mode temporal dependency capture layer to the actual predicted fog visibility level output. This layer specifically includes linear mapping units and nonlinear thresholding units.

[0165] In specific implementation, the output of the linear mapping unit is calculated as follows:

[0166]

[0167] In the formula: To output the weight matrix, The final output features of the temporally dependent capture layer, For output layer bias terms;

[0168] Nonlinear threshold unit utilizes or Nonlinear functions such as functions are used to activate the output of the linear mapping unit to obtain the final predicted fog visibility level.

[0169] In specific implementation, the method for obtaining the long short-term memory network model optimized by the particle size modal attention mechanism includes:

[0170] Based on historical particle size modal sequences and corresponding visibility observation levels, multimodal time series training samples are constructed.

[0171] It should be noted that this step involves pairing historical observed particle size modality category sequences with actual visibility level observation data from the same period to form a training sample set, where the historical particle size modality sequence is the input feature and the visibility level is the model prediction target.

[0172] The fog visibility level is a specific quantitative standard that classifies atmospheric visibility in the current area at future times based on the range of fog visibility values; specifically, it is defined according to the visibility level classification standard widely used in the field of traffic meteorology.

[0173] For example, the fog visibility level can be specifically divided into the following levels:

[0174] Level 1 (Dense Fog): Visibility less than 200m;

[0175] Level 2 (Moderate Fog): Visibility is 200–500m;

[0176] Level 3 (Light Fog): Visibility is 500–1000m;

[0177] Level 4 (Light Fog): Visibility is 1000–3000m;

[0178] Level 5 (Clear): Visibility greater than 3000m.

[0179] It should be understood that the specific numerical classification criteria for the levels can be adjusted according to the specific needs of the application scenario. The above values ​​are provided only as examples to illustrate the specific implementation of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0180] A bidirectional multi-head self-attention sublayer is used to calculate attention weights for historical particle size mode sequences by utilizing the physical transfer characteristics of particle size mode categories.

[0181] For example, the historical sequence is input into a bidirectional multi-head self-attention layer. The attention weight matrix is ​​calculated comprehensively from the perspectives of particle size change trend, particle size growth rate and modality category transfer frequency through the multi-head self-attention mechanism to highlight key historical modality features.

[0182] A dynamic routing capsule unit is used to fuse historical sequence features calculated with attention weights with current particle size modal distribution features to generate a cross-modal fusion sequence;

[0183] It should be understood that this step utilizes dynamic routing capsule units to iteratively fuse historical and current modal features to ensure that the influence of historical sequences on the current particle size state is fully reflected.

[0184] A hybrid layer structure is used to capture the temporal dependency of cross-modal fusion sequences, and the visibility prediction level is obtained through linear mapping units and nonlinear thresholding units. An improved loss function is used to optimize the network through backpropagation, and the model parameters are trained. The output is a long short-term memory network model optimized by particle size modal attention mechanism.

[0185] Specifically, an improved loss function is used to optimize the network through backpropagation, complete the training of model parameters, and finally output the optimized particle size modal attention mechanism LSTM model.

[0186] The improved loss function is an asymmetric penalty mechanism, expressed as follows:

[0187]

[0188] In the formula: The loss function value is N; the training sequence length is N. The actual observed visibility level at time i; Let i be the predicted visibility level at time i; Let be the predicted visibility level at time i−1; The basic loss weighting coefficient is determined based on specific experimental data, for example, a value of 0.5; This is the asymmetric penalty coefficient, determined based on specific experimental data, for example, a value of 1; This is the time-series fluctuation sensitivity coefficient, which is determined based on specific experimental data, for example, a value of 0.1.

[0189] Please see Figure 2 Based on the same inventive concept, a second aspect of this invention provides a fog visibility prediction system based on deep learning. For details not covered in this embodiment, please refer to the relevant sections of Embodiment 1. The system includes:

[0190] The modality recognition module 201 is used to extract the particle size modality distribution characteristics of the current area based on the backscatter intensity spectrum detected by the lidar, and determine the particle size modality category that characterizes the particle size distribution type of fog.

[0191] The sequence filtering module 202 is used to determine the meteorological environment status of the current area based on the measured humidity and aerosol concentration data; and to filter historical particle size mode sequences that match the particle size mode category of the current area based on the inherent correspondence between the meteorological environment status and particle size mode category.

[0192] The path reconstruction module 203 is used to extract the transfer path of particle size mode category according to the transfer law of mode category in the historical particle size mode sequence; and to reconstruct the sequence of typical particle size mode categories in the historical particle size mode sequence based on the category transfer complexity reflected by the transfer path.

[0193] The visibility prediction module 204 is used to input the reconstructed historical particle size mode sequence and the current particle size mode distribution features into a long short-term memory network model optimized by the particle size mode attention mechanism, and output the predicted fog visibility level of the current area at future time.

[0194] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0195] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0196] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0197] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0198] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0199] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for predicting visibility in heavy fog based on deep learning, characterized in that, include: Based on the backscattering intensity spectrum detected by lidar, the particle size mode distribution characteristics of the current area are extracted to determine the particle size mode category that characterizes the particle size distribution type of fog. Based on the measured humidity and aerosol concentration data, the current meteorological environment status of the region is determined; and based on the inherent correspondence between the meteorological environment status and particle size mode category, historical particle size mode sequences that match the particle size mode category of the current region are selected. Based on the transition patterns of mode categories in historical particle size mode sequences, the transition paths of particle size mode categories are extracted; Based on the complexity of category transfer reflected by the transfer path, the typical particle size mode categories in the historical particle size mode sequence are reconstructed. The reconstructed historical particle size mode sequence and the current particle size mode distribution features are input into a long short-term memory network model optimized by a particle size mode attention mechanism, and the output is the predicted fog visibility level for the current region at future times.

2. The method for predicting visibility in heavy fog based on deep learning according to claim 1, characterized in that, The determination of particle size mode categories characterizing the particle size distribution type of fog includes: Based on the backscattering intensity spectrum detected by lidar, the position of the spectral peak in each particle size range is identified, and the corresponding peak particle size is determined. The candidate dominant particle size range is determined based on the stable occurrence frequency of peak particle size within a continuous time window; Based on the nonlinear thermodynamic threshold condition of the hygroscopic growth and dissipation of fog particles, the probability of stable existence of candidate dominant particle size ranges under the current environmental humidity conditions is calculated. Based on the stable existence probability satisfying the preset probability threshold condition, the target particle size range that satisfies thermodynamic stability is determined, and based on the peak particle size correspondence of the target particle size range, the particle size mode category characterizing the fog particle size distribution type is determined.

3. The method for predicting visibility in heavy fog based on deep learning according to claim 2, characterized in that, The process of filtering historical particle size mode sequences that match the current region's particle size mode category includes: Based on the measured humidity and aerosol concentration data, determine the typical meteorological conditions to which the current regional meteorological environment belongs; Based on the causal correspondence between typical meteorological conditions and the physical formation mechanism of particle size modal categories, the formation path of particle size modal categories under typical meteorological conditions is established; Based on the physical continuity of the formation path, candidate historical sequences that match the physical formation path of the current particle size mode category are selected from the historical observed particle size mode sequences; Based on whether the formation path of the candidate historical sequence conforms to the physical generation mechanism of fog particle size mode, the historical particle size mode sequence that matches the current region particle size mode category is determined.

4. The method for predicting visibility in heavy fog based on deep learning according to claim 3, characterized in that, The screening of candidate historical sequences that match the physical formation path of the current particle size modality category includes: Based on the hygroscopic growth and condensation processes of particles corresponding to different particle size modal categories, the formation stage of the particle size modal category in the historical observation sequence is determined; Based on the physical matching relationship between the formation stage of particle size modal categories and typical meteorological conditions, a stage transition path between each particle size modal category in the historical sequence is established; Based on the physical transfer path of the particle size modality formation stage and the physical process corresponding to the formation stage of the particle size modality in the current region, candidate historical sequences with matching stage transfer paths are screened. Based on the completeness of the particle size modality formation process in the candidate historical sequence of the stage transition path matching, a candidate historical sequence matching the physical formation path of the current particle size modality is determined.

5. The method for predicting visibility in heavy fog based on deep learning according to claim 4, characterized in that, The sequence reconstruction of typical particle size mode categories in historical particle size mode sequences includes: Based on the transfer paths of mode categories in the historical particle size mode sequence, determine the physical evolution process of fog particle size modes corresponding to each transfer path; Based on the thermodynamic criteria for particle size growth and dissipation during the physical evolution of particle size modes, the categories of abnormal particle size modes in the transfer path are identified; Based on the discontinuity of the physical formation mechanism corresponding to the abnormal particle size mode category, the abnormal particle size mode category is sequentially eliminated; Based on the particle size mode sequence after removing abnormal particle size mode categories, the typical particle size mode category sequence is reconstructed under the condition of continuity in the physical evolution process of particle size modes in fog, and the reconstructed historical particle size mode sequence is output.

6. The method for predicting visibility in heavy fog based on deep learning according to claim 5, characterized in that, The identification of anomalous particle size modal categories in the transfer path includes: Determine the dominant particle size range and the thermodynamic critical humidity required for hygroscopic growth of particles corresponding to the particle size mode categories in the historical particle size mode sequence. Based on the humidity evolution trajectory of the corresponding particle size modality in the region during the historical observation period, the percentage of time that the humidity conditions for the corresponding particle size modality meet the critical humidity is extracted. Based on the time proportion and the preset humidity threshold condition for stable and continuous existence of particle size modal categories, particle size modal categories that do not meet the stable humidity condition are identified and determined as abnormal particle size modal categories.

7. The method for predicting visibility in heavy fog based on deep learning according to claim 6, characterized in that, The long short-term memory network model optimized by the particle size modal attention mechanism includes: The particle size modal attention coding layer includes a bidirectional multi-head self-attention sublayer based on the physical transfer characteristics of particle size modal categories; The cross-modal information interaction layer includes dynamic routing capsule units for fusing historical particle size modal sequences with current particle size modal distribution characteristics; The particle size mode temporally dependent capture layer comprises at least three hybrid layer structures consisting of alternating convolutional gated recurrent units and long short-term memory network units; The sequence prediction output layer includes a linear mapping unit and a nonlinear thresholding unit.

8. The method for predicting visibility in heavy fog based on deep learning according to claim 7, characterized in that, The method for obtaining the long short-term memory network model optimized by the particle size modal attention mechanism includes: Based on historical particle size modal sequences and corresponding visibility observation levels, multimodal time series training samples are constructed. A bidirectional multi-head self-attention sublayer is used to calculate attention weights for historical particle size mode sequences by utilizing the physical transfer characteristics of particle size mode categories. A dynamic routing capsule unit is used to fuse historical sequence features calculated with attention weights with current particle size modal distribution features to generate a cross-modal fusion sequence; A hybrid layer structure is used to capture the temporal dependencies of cross-modal fusion sequences, and the visibility prediction level is obtained through linear mapping units and nonlinear thresholding units. An improved loss function is used to optimize the network through backpropagation, and the model parameters are trained. The output is a long short-term memory network model optimized by particle size modal attention mechanism.

9. A method for predicting visibility in heavy fog based on deep learning according to claim 8, characterized in that, The improved loss function is an asymmetric penalty mechanism, expressed as follows: In the formula: The loss function value is N; the training sequence length is N. The actual observed visibility level at time i; Let i be the predicted visibility level at time i; Let be the predicted visibility level at time i−1; The basic loss weighting coefficient; The asymmetric penalty coefficient; This is the time series fluctuation sensitivity coefficient.

10. A fog visibility prediction system based on deep learning, implemented based on the fog visibility prediction method based on deep learning according to any one of claims 1-9, characterized in that, include: The modality recognition module is used to extract the particle size modality distribution characteristics of the current area based on the backscatter intensity spectrum detected by lidar, and to determine the particle size modality category that characterizes the particle size distribution type of fog. The sequence filtering module is used to determine the meteorological environment status of the current area based on the measured humidity and aerosol concentration data; and based on the inherent correspondence between the meteorological environment status and particle size mode category, to filter historical particle size mode sequences that match the particle size mode category of the current area. The path reconstruction module is used to extract the transfer path of particle size mode category based on the transfer pattern of mode category in the historical particle size mode sequence; Based on the complexity of category transfer reflected by the transfer path, the typical particle size mode categories in the historical particle size mode sequence are reconstructed. The visibility prediction module is used to input the reconstructed historical particle size mode sequence and the current particle size mode distribution features into a long short-term memory network model optimized by the particle size mode attention mechanism, and output the predicted fog visibility level of the current area at future time.

Images