PM2.5 concentration prediction method, equipment, medium and product

Abstract

The invention discloses a PM2.5 concentration prediction method and device, a medium and a product, and relates to the technical field of atmospheric environment monitoring and deep learning prediction.The method comprises the steps that laser radar vertical profile data containing extinction coefficients and depolarization ratio vertical profiles and a scalar time series data sequence are obtained; converting the laser radar vertical profile data sequence into a vertical sequence containing position information at any moment; adopting a multi-channel spatial feature extraction model, and obtaining a multi-channel spatial feature vector based on the vertical sequence; and after the scalar time sequence data sequence and the multi-channel spatial feature vector are spliced to obtain a fused feature vector, a PM2.5 concentration prediction value is obtained by adopting a time sequence prediction model. According to the method, the technical pain point caused by the fact that aerosol types cannot be distinguished and'high altitude-ground 'cross-layer physical association cannot be captured in the prior art is solved, so that accurate and robust prediction of the PM2.5 concentration on the ground is realized under complex weather conditions such as sand and dust transportation and boundary layer evolution.

Description

Technical Field

[0001] This application relates to the field of atmospheric environmental monitoring and deep learning prediction technology, and in particular to a method, device, medium and product for predicting PM2.5 concentration. Background Technology

[0002] With the increasing demand for air pollution monitoring and regional air quality assessment, accurately estimating air quality, especially the time-series prediction of near-surface PM2.5 (fine particulate matter) concentrations, has become a significant technical challenge in the field of atmospheric remote sensing.

[0003] Currently, the mainstream PM2.5 prediction technologies can be mainly divided into the following categories: 1. Deep learning prediction based on ground data: mainstream methods (such as Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Convolutional Neural Networks-Long Short-Term Memory (CNN-LSTM), etc.) usually only use historical PM2.5 concentration and ground meteorological data (such as temperature, humidity, and wind speed) as input.

[0004] Technical limitations: These methods rely entirely on ground-based scalar data, completely losing information about the vertical structure of the atmosphere. However, the evolution of PM2.5 (such as explosive growth or sudden drops) is often strongly constrained by vertical physical processes such as upper-level pollutant transport, boundary layer height evolution, and temperature inversion suppression. Due to the lack of this crucial vertical information, existing technologies suffer from severely insufficient accuracy and poor robustness in predicting complex pollution processes (such as upper-level dust deposition).

[0005] 2. Based on statistical predictions of AOD (aerosol optical thickness), (such as Chinese invention patent application CN110046771A), attempts were made to introduce satellite AOD data.

[0006] Technical limitations: AOD is an integral scalar reflecting the total amount of aerosols in the entire atmosphere. It still cannot distinguish whether aerosols are concentrated within 500 meters of the ground or suspended at an altitude of 2000 meters. Therefore, it also cannot provide the model with key vertical distribution characteristics, and its role in improving the accuracy of ground-based PM2.5 prediction is limited.

[0007] 3. Prediction based on a single radar parameter or conventional time series model: With the application of lidar, some studies have begun to try to introduce the extinction coefficient profile (alpha) retrieved from radar into the prediction model.

[0008] Question 1: The relationship between PM2.5 mass concentration (M) and extinction coefficient (alpha) is M = alpha / eta ext Among them, the mass extinction efficiency eta ext The extinction coefficient varies drastically depending on the aerosol's properties (type / shape). For example, non-spherical dust particles and spherical local pollution particles may have the same extinction coefficient value, but their corresponding PM2.5 concentrations are drastically different. Most existing technologies only input alpha (a characterization factor), ignoring the distinction between particle shape (characterization factor). This leads to models being unable to determine the mass extinction efficiency eta when dealing with complex aerosol types such as dust and high humidity. ext This can lead to significant prediction errors.

[0009] Question 2: In existing AI models for processing radar data (such as Chinese invention patent application CN119993303A), convolutional neural networks (CNNs) or recurrent neural networks (RNNs) are typically used to slide along the time axis (t-axis) to extract temporal features. However, atmospheric vertical profiles contain complex spatial global correlations (e.g., a dust layer at 2000 meters directly affects air quality at 100 meters through deposition), making it difficult to capture such global physical correlations across different altitudes.

[0010] In summary, existing technologies are simply unable to achieve accurate and robust prediction of ground-level PM2.5 concentrations under complex weather conditions such as dust transport and boundary layer evolution. Summary of the Invention

[0011] To address the aforementioned problems in the existing technology, this application provides a PM2.5 concentration prediction method, device, medium, and product.

[0012] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting PM2.5 concentration, comprising: Acquire a vertical profile data sequence and a scalar time series data sequence from a lidar system; the vertical profile data sequence from the lidar system includes an extinction coefficient vertical profile and a deflection ratio vertical profile. The vertical profile data sequence of the lidar is converted into a vertical sequence containing position information at any given time. A multi-channel spatial feature extraction model is used to obtain multi-channel spatial feature vectors based on the vertical sequence; The scalar time-series data sequence and the multi-channel spatial feature vector are concatenated to obtain a fused feature vector; The PM2.5 concentration prediction value is obtained by using a time-series prediction model based on the fused feature vector.

[0013] Optionally, the multi-channel spatial feature extraction model is a neural network built on the Transformer architecture.

[0014] Optionally, a multi-channel spatial feature extraction model is employed to obtain a multi-channel spatial feature vector based on the vertical sequence, including: A learnable global aggregation marker is added to the beginning of the vertical sequence; after processing by the Transformer encoder layer, the output vector corresponding to the global aggregation marker is directly extracted as the multi-channel spatial feature vector.

[0015] Optionally, the process of the Transformer encoder layer includes: The multi-head self-attention mechanism in the Transformer architecture is used to capture the physical correlation strength between any two altitude layers in the altitude layer position sequence, obtain the influence weight of the upper-level layer on the near-surface layer, and extract the meteorological interface features that restrict the diffusion of PM2.5 to obtain the feature sequence; the near-surface layer refers to the spatial layer at a first predetermined distance from the ground; the upper-level layer refers to the spatial layer at a second predetermined distance from the ground.

[0016] Optionally, the position encoding vector is defined using a sine / cosine function, expressed as: ; In the formula, Indicates the position index is pos The numerical value of the position encoding vector in even-numbered dimensions. Indicates the position index is pos The numerical values ​​of the position encoding vector in odd-numbered dimensions. pos Indices representing the index of the vertical height layer, d model Represents the total dimension. i The dimension index is sin(), which represents the sine function, and cos(), which represents the cosine function.

[0017] Optionally, the time-series prediction model is constructed using a gated recurrent unit based on an attention mechanism.

[0018] Optionally, a time-series prediction model is used to obtain predicted PM2.5 concentration values ​​based on the fused feature vector, including: The update gate and reset gate mechanism in the gated loop unit are used to capture the time trend of aerosol evolution in the fused feature vector, output the hidden state of each time step, and use an attention mechanism to obtain the importance weight of each time step.

[0019] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the PM2.5 concentration prediction method provided above.

[0020] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the PM2.5 concentration prediction method provided above.

[0021] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the PM2.5 concentration prediction method provided above.

[0022] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a PM2.5 concentration prediction method, device, medium, and product. By converting a lidar vertical profile data sequence containing both extinction coefficient and depolarization ratio vertical profiles into a vertical sequence with location information at any given time, and employing a multi-channel spatial feature extraction model, a multi-channel spatial feature vector is obtained based on this vertical sequence. This allows for in-depth mining of the synergistic features of the extinction coefficient (quantity) and depolarization ratio (quality) in the vertical spatial dimension, addressing the technical limitations of existing technologies that cannot distinguish aerosol types or capture the cross-layer physical relationship between upper atmosphere and ground level. By splicing scalar time-series data sequences and multi-channel spatial feature vectors to obtain a fused feature vector, and then using a time-series prediction model to obtain predicted PM2.5 concentration values, accurate and robust prediction of ground-level PM2.5 concentrations is achieved. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A schematic flowchart illustrating a PM2.5 concentration prediction method provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the implementation process of a PM2.5 concentration prediction method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Based on the physical premise that the vertical structure of the atmosphere determines the evolution of ground pollution, this paper proposes a novel research and development concept. Instead of treating radar profiles merely as simple multivariate time series, it views them as a set of vertical physical quantities with clear physical meaning. Based on this concept, this application provides a PM2.5 concentration prediction method, device, medium, and product. This aims to address the technical limitations of existing technologies, such as the inability to distinguish aerosol types and capture the cross-layer physical relationship between the upper atmosphere and the ground, by deeply exploring the synergistic characteristics of the extinction coefficient (quantity) and the deflection ratio (quality) in the vertical spatial dimension. This will enable accurate and robust prediction of ground PM2.5 concentrations under complex weather conditions such as dust transport and boundary layer evolution.

[0027] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] In one exemplary embodiment, this application provides a PM2.5 concentration prediction method. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is described using a server as an example. Figure 1 As shown, the method includes: Step 100: Acquire the lidar vertical profile data sequence and scalar time series data sequence. The lidar vertical profile data sequence includes the extinction coefficient vertical profile and the depolarization ratio vertical profile. In the scalar time series data sequence, humidity is a crucial physical quantity that must be included. The model uses humidity data to implicitly learn the hygroscopic growth factor of aerosols through a deep neural network, thereby automatically correcting the artificially high extinction coefficient caused by increased humidity during the feature fusion stage and restoring the true dry mass concentration of particulate matter.

[0029] Step 101: Convert the lidar vertical profile data sequence into a vertical sequence that contains position information at any given time. For example, at any single time t, the lidar vertical profile data sequence is considered as a vertical sequence containing position information.

[0030] Step 102: Use a multi-channel spatial feature extraction model to obtain multi-channel spatial feature vectors based on vertical sequences.

[0031] Step 103: Concatenate the scalar time series data sequence and the multi-channel spatial feature vector to obtain the fused feature vector.

[0032] The concatenation here is achieved through feature fusion using a gated fusion unit. Specifically, the low-dimensional scalar time-series data is first mapped to a high-dimensional model with the same spatial feature vectors via a fully connected layer; then, a sigmoid activation function is used to generate a gate, dynamically calculating the fusion ratio between the two (e.g., ...). ); final output fusion vector .

[0033] Step 104: Using a time-series prediction model, the predicted PM2.5 concentration is obtained based on the fused feature vector.

[0034] By implementing steps 100-104 above, this application utilizes multi-channel vertical profile data (extinction coefficient and depolarization ratio) and ground scalar time series data (i.e. scalar time series data sequence) provided by polarization lidar to achieve accurate time series prediction of ground PM2.5 concentration.

[0035] In another exemplary embodiment of this application, in order to eliminate the physical ambiguity of aerosol types (such as dust and local pollution) caused by the use of only a single optical parameter in the prior art, and to solve the problem that the prediction accuracy of the prior art is limited under complex weather conditions due to the lack of vertical structure information or confusion of spatiotemporal dimensions, this application introduces the dual-channel profiles of extinction coefficient and depolarization ratio as combined inputs into PM2.5 prediction, so as to explore the synergistic pattern of these two profiles with different physical meanings in vertical space. Based on this, in practical applications, step 100 above mainly acquires at least two types of historical data before the target prediction time. Among them, the first type of historical data is the lidar vertical profile data sequence, which includes the extinction coefficient vertical profile alpha(z, t) used to characterize the vertical distribution of aerosol quantity, and the depolarization ratio vertical profile delta(z, t) used to characterize the vertical distribution of aerosol quality / type. The second type of historical data is the scalar time series data sequence, which mainly includes historical PM2.5 concentration time series data obtained from the station and historical surface meteorological time series data (such as temperature, humidity, wind speed, etc.) obtained from the station. Among them, the scalar time series data sequence is denoted as V scalar Its dimensions are d scalar .

[0036] In another exemplary embodiment of this application, in order to adopt a spatial feature extraction-temporal prediction fusion architecture (i.e., a multi-channel spatial feature extraction model + a temporal prediction model), this application can construct a vertical spatial encoder based on the Transformer architecture (i.e., a multi-channel spatial feature extraction model) to directly capture the global collaborative spatial structure features of aerosol quantity (extinction coefficient) and quality (debias ratio) in the vertical height dimension (z-axis) (i.e., to directly capture global physical associations in the vertical height dimension (z-axis)) and compress them into a high-information-content spatial feature vector. Based on this, the implementation process of step 102 provided above in this application can be described as follows: a learnable global aggregation marker is added to the beginning of the vertical sequence; after processing by the Transformer encoder layer, the output vector corresponding to the global aggregation marker is directly extracted as the multi-channel spatial feature vector.

[0037] In practical applications, to enable the multi-channel spatial feature extraction model to perceive the absolute position at different altitudes (such as near-ground and high altitude), the input vertical sequence X needs to be added to the position encoding matrix P. The position encoding is defined using a sine / cosine function, expressed as: .

[0038] In the formula, Indicates the position index is pos The numerical value of the position encoding vector in even-numbered dimensions. Indicates the position index is pos The numerical values ​​of the positional encoding vector in odd-numbered dimensions. pos Indicates the index of the vertical height layer (i.e., its position in the vertical sequence). pos =0 represents the height layer closest to the ground (i.e., the near-ground layer). pos =N represents the highest altitude layer (i.e., the upper altitude layer). d model This represents the total dimension, which is the size of the hidden layer inside the multi-channel spatial feature extraction model. i This is a dimension index, and its value ranges from 0 to... d model / 2–1. sin() represents the sine function, and cos() represents the cosine function. 10000 is a preset constant basis used to control the frequency of change of different wavelengths, ensuring that the multi-channel spatial feature extraction model can perceive various spatial dependencies from short distances (adjacent height layers) to long distances (spanning high altitudes and the ground).

[0039] The Transformer encoder layer processing includes: employing the multi-head self-attention mechanism in the Transformer architecture to capture the physical correlation strength between any two altitude layers in the altitude layer position sequence, obtaining the influence weight of the upper atmosphere on the near-surface layer, and extracting the meteorological interface features that constrain PM2.5 diffusion to obtain a feature sequence. The near-surface layer refers to the spatial layer at a first predetermined distance from the ground. The upper atmosphere refers to the spatial layer at a second predetermined distance from the ground.

[0040] In practical applications, the multi-head self-attention mechanism in the Transformer architecture is used to directly capture the physical association strength between any two height layers by calculating the interaction between the query matrix Q, the key matrix K, and the value matrix V. The formula for calculating the attention output is: .

[0041] In the formula, Q The query matrix represents the feature representation of the target height layer. K The key matrix represents the feature indices of all height layers. V It is a value matrix that represents the actual feature content of all height layers. d k Let be the dimension of the key vector. Used as a scaling factor to prevent the gradient from vanishing due to an excessively large inner product. T This represents the transpose of the matrix. `softmax` is a normalized exponential function used to transform the dot product result into a probability distribution (i.e., attention weights).

[0042] By using the calculation formula for attention output, the multi-channel spatial feature extraction model can identify the influence weight of the upper atmosphere (such as the dust layer) on the near-surface layer.

[0043] Based on the above description, the feature extraction described above is not merely a simple dimensionality reduction of the input data, but rather, through the self-attention mechanism of the Transformer, it extracts the implicit atmospheric vertical structure features and cross-level physical correlations from the original extinction coefficient and debiasing ratio profiles. Specifically, the multi-channel spatial feature extraction model extracts the following key physical information (1) and (2) and encodes it into the feature vector to obtain the feature sequence: (1) The direct numerical correlation between the upper-level pollution layer and the near-ground layer characterizes the physical trend of pollutant deposition or diffusion.

[0044] (2) The meteorological interface features that implicitly constrain PM2.5 diffusion, representing the boundary layer height and cloud structure, are implicitly extracted from the profile gradient changes.

[0045] Among them, the key physical information in (1) and (2) is linked through the self-attention weight matrix in the Transformer. Whether a high-altitude pollution layer (e.g., dust at 3 km) will settle and affect the ground depends on the physical transport mechanism between the two.

[0046] In terms of weights, the attention weights calculated by Transformer represent the information contribution of a high-altitude layer, such as 3km, to another high-altitude layer, such as the ground. If the attention weight ({(ground, 3km)}) calculated by the model is large, this corresponds to a strong physical deposition trend (i.e., upper-air pollutants are significantly affecting the ground). This large weight value will sum the features of upper-air alpha and delta into the feature vector of the ground. After receiving this vector with upper-air features, the subsequent time-series prediction model will predict the future upward trend of PM2.5 concentration at the ground.

[0047] The lower the altitude at which the extinction coefficient profile undergoes a gradient abrupt change, the more difficult it is for pollutants to disperse, resulting in higher ground-level concentrations. At the altitude where the extinction coefficient profile undergoes a gradient abrupt change (i.e., PBLH), the distribution pattern of attention weights is truncated. There is high mutual attention between altitude layers below the altitude where the extinction coefficient profile undergoes a gradient abrupt change, while the attention of altitude layers below the altitude where the extinction coefficient profile undergoes a gradient abrupt change to altitude layers above the altitude where the extinction coefficient profile undergoes a gradient abrupt change decreases sharply.

[0048] The feature vectors generated by this weight distribution pattern implicitly encode information about poor diffusion conditions (low PBLH) or good diffusion conditions (high PBLH). After the time-series prediction model identifies the feature encoding of poor diffusion conditions, it will output a higher predicted PM2.5 concentration value.

[0049] Step 3: Perform global average pooling on the feature sequence to obtain a multi-channel spatial feature vector. Where is denoted as and is denoted as . V spatial Its dimensions are d model 。 The purpose of this step is to compress the feature sequence into a fixed-length feature vector (i.e., a multi-channel spatial feature vector).

[0050] Based on the above description, steps 1-3 provided in this application utilize a self-attention mechanism to extract global deep physical features containing atmospheric vertical structure from the collaborative information of alpha and delta.

[0051] In another exemplary embodiment of this application, the specific splicing (i.e. fusion) method of step 103 above can be: performing the operation {Concat}( V spatial , Vscalar The concatenation operation is performed along the feature dimension, meaning that while keeping the time step constant, the number of feature channels is expanded. Based on this, the dimension of the resulting fused feature vector is... D fused = d model + d scalar The fused feature vector contains both high-dimensional vertical spatial features and low-dimensional ground scalar features at that moment, serving as the basic input unit for subsequent time-series predictions.

[0052] In another exemplary embodiment of this application, based on the multi-channel spatial feature extraction model constructed above, and combined with an attention-based gated recurrent unit (Attention-GRU), temporal deduction can be performed on the sequence of spatial feature vectors obtained (i.e., fused feature vectors) in the time dimension (t-axis), thereby strictly distinguishing and effectively fusing vertical spatial structure and temporal evolution patterns. Based on this, the temporal prediction model used in step 104 of this application can be constructed by an attention-based gated recurrent unit. Therefore, in practical applications, the implementation process of step 104 can be described as follows: Step (1) GRU layer deduction: The GRU layer receives the fused feature vector, captures the temporal trend of aerosol evolution through its internal update gate and reset gate mechanism, and outputs the hidden state at each time step. Hidden state It contains the physical state information after spatial feature extraction at that moment.

[0053] Step (2) Temporal Attention Weighting: The time series prediction model dynamically focuses on moments with key physical significance in history (such as the moment a sandstorm passes through), and an attention mechanism is introduced to calculate the importance weight of each historical time step. alpha t .

[0054] In practical applications, the specific implementation process of steps (1) and (2) above can be described as follows: Calculate the first alignment using a learnable alignment model. t Hidden state at each time step energy fraction : .

[0055] In the formula, , and These are the weight matrix and bias vector that the alignment model automatically learns during training. T represents the total length of the input time series (e.g., the past 24 hours).

[0056] The energy scores of all historical time steps are normalized using the Softmax function to obtain the weight of each time step. (i.e., importance weight) alpha t ): .

[0057] In the formula, Let be the energy fraction at time k.

[0058] Based on the calculated importance weight alpha t The historical hidden states are weighted and summed to generate the final context vector used for prediction. c : For example, if at a certain moment t When a dust storm event occurs, the time-series prediction model will automatically learn to make the timeline accurate. Increases, thus affecting the importance weight. alpha t It is significantly close to 1. However, for stationary moments without critical physical events, its importance weight... alpha t It then approaches 0.

[0059] The final context vector used for prediction c The most critical physical features in the historical sequence were focused on and then fed into a fully connected layer for the final PM2.5 numerical prediction. This is based on the context vector ultimately used for prediction. c Output the predicted PM2.5 concentration at the target prediction time.

[0060] Based on the above description, the specific implementation process of the PM2.5 concentration prediction method provided in this application can be as follows: Figure 2 As shown, this invention provides a method based on a spatial feature extraction-temporal prediction architecture. It aims to address the technical limitations of existing technologies, such as the inability to distinguish aerosol types and capture the cross-layer physical relationship between upper atmosphere and ground level, by deeply mining the synergistic features of extinction coefficient (quantity) and deflection ratio (quality) in the vertical spatial dimension. This allows for accurate and robust prediction of ground-level PM2.5 concentrations under complex weather conditions such as dust transport and boundary layer evolution. Compared to existing technologies, this application also has at least the following advantages: 1) It can deeply explore the vertical physical structure of the atmosphere.

[0061] Existing technologies, such as LSTM which only inputs ground data, completely lose vertical information. Even simply introducing profile data (such as using CNN) often only captures local features. This application, however, by introducing vertical profiles, can extract implicit meteorological physical structures from the original profiles and explicitly encode these implicit physical parameters that constrain PM2.5 diffusion into the feature vector, thereby significantly improving prediction accuracy.

[0062] 2) It can eliminate physical ambiguity and improve the prediction accuracy under complex aerosol types.

[0063] Existing technologies fail to predict aerosol types when faced with complex weather conditions such as dust storms (non-spherical) or high humidity (spherical aerosols that grow by absorbing moisture), leading to prediction failures. This application, however, can accurately distinguish between the quality and quantity of aerosols, improving prediction accuracy in complex processes such as dust transport and haze outbreaks.

[0064] 3) It can achieve spatial and temporal fusion prediction.

[0065] Existing technologies, such as using CNNs to process multivariate data along the time axis, result in the dilution of vertical spatial information in the time series. However, the spatial feature extraction-temporal prediction architecture of this application ensures that the feature vectors entering the temporal model are true mappings of the atmospheric vertical structure at that moment.

[0066] Furthermore, while several alternative solutions exist in the prior art to the method provided in this application, these methods in the prior art are insufficient in terms of physical interpretability and prediction under complex weather conditions, and cannot simultaneously achieve the technical effects of this application. For example: Alternative Option 1: Conventional time-series forecasting methods based solely on ground-based monitoring data (such as standard LSTM / BiLSTM). This type of method is currently the mainstream approach for air quality forecasting, utilizing only historical PM2.5 concentrations and surface meteorological elements (temperature, humidity, etc.) as input. However, this approach completely loses information about the vertical structure of the atmosphere. When upper-level pollutant transport or drastic changes in boundary layer height occur, ground-based data often exhibits lag or cannot characterize these events, leading to a significant decrease in the model's prediction accuracy when facing sudden, imported pollution events. It cannot capture upper-level pollution signals in advance through vertical profiles, as is the approach described in this application.

[0067] Alternative Solution Two: Alternative Solution One plus aerosol optical depth (AOD) prediction methods. This approach attempts to incorporate AOD products from satellite remote sensing to supplement optical information. However, AOD is essentially a scalar quantity integral over the entire atmosphere, and it cannot distinguish whether aerosols are concentrated near the ground or suspended at high altitudes. Therefore, although this approach incorporates optical information, it still lacks crucial vertical resolution and cannot identify fine structures such as inversion layers or upper-level transport layers, thus limiting its effectiveness in improving the accuracy of ground-based PM2.5 predictions.

[0068] Alternative Solution 3: Prediction methods based on a single extinction coefficient profile. While this approach utilizes the extinction coefficient profile of the lidar, it suffers from the following problems: 1. Using only the extinction coefficient without the depolarization ratio makes the model unable to distinguish between aerosol types with the same extinction coefficient but completely different physical properties, resulting in physical ambiguity.

[0069] 2. Alternatively, after introducing the debiasing ratio, CNN can be used directly to process the profile data sequence on the time axis, which confuses the vertical spatial dimension and the time dimension.

[0070] Clearly, the processing method provided by Alternative Solution 3 makes it difficult to extract independent, global vertical structure features from a single frame profile, resulting in insufficient ability of the model to represent atmospheric physical processes.

[0071] In summary, the PM2.5 concentration prediction method provided in this application can truly achieve accurate and robust prediction of ground-level PM2.5 concentrations compared to existing technologies.

[0072] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores PM2.5 concentration prediction data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements a PM2.5 concentration prediction method.

[0073] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0074] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0075] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0076] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0077] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0078] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0079] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0080] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting PM2.5 concentration, characterized in that, include: Acquire a vertical profile data sequence and a scalar time series data sequence from a lidar system; the vertical profile data sequence from the lidar system includes an extinction coefficient vertical profile and a deflection ratio vertical profile. The vertical profile data sequence of the lidar is converted into a vertical sequence containing position information at any given time. A multi-channel spatial feature extraction model is used to obtain multi-channel spatial feature vectors based on the vertical sequence; The scalar time-series data sequence and the multi-channel spatial feature vector are concatenated to obtain a fused feature vector; The PM2.5 concentration prediction value is obtained by using a time-series prediction model based on the fused feature vector.

2. The PM2.5 concentration prediction method according to claim 1, characterized in that, The multi-channel spatial feature extraction model is a neural network built on the Transformer architecture.

3. The PM2.5 concentration prediction method according to claim 2, characterized in that, A multi-channel spatial feature extraction model is employed to obtain multi-channel spatial feature vectors based on the vertical sequence, including: A learnable global aggregation marker is added to the beginning of the vertical sequence; after processing by the Transformer encoder layer, the output vector corresponding to the global aggregation marker is directly extracted as the multi-channel spatial feature vector.

4. The PM2.5 concentration prediction method according to claim 3, characterized in that, The process of the Transformer encoder layer includes: The multi-head self-attention mechanism in the Transformer architecture is used to capture the physical correlation strength between any two altitude layers in the altitude layer position sequence, obtain the influence weight of the upper-level layer on the near-surface layer, and extract the meteorological interface features that restrict the diffusion of PM2.5 to obtain the feature sequence; the near-surface layer refers to the spatial layer at a first predetermined distance from the ground; the upper-level layer refers to the spatial layer at a second predetermined distance from the ground.

5. The PM2.5 concentration prediction method according to claim 4, characterized in that, The position encoding vector is defined using a sine / cosine function and is expressed as follows: ； In the formula, Indicates the position index is pos The numerical value of the position encoding vector in even-numbered dimensions. Indicates the position index is pos The numerical values ​​of the position encoding vector in odd-numbered dimensions. pos Indices representing the index of the vertical height layer, d model Represents the total dimension. i The dimension index is sin(), which represents the sine function, and cos(), which represents the cosine function.

6. The PM2.5 concentration prediction method according to claim 1, characterized in that, The time series prediction model is constructed using a gated recurrent unit based on an attention mechanism.

7. The PM2.5 concentration prediction method according to claim 6, characterized in that, Using a time-series prediction model, PM2.5 concentration prediction values ​​are obtained based on the fused feature vector, including: The update gate and reset gate mechanism in the gated loop unit are used to capture the time trend of aerosol evolution in the fused feature vector, output the hidden state of each time step, and use an attention mechanism to obtain the importance weight of each time step.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the PM2.5 concentration prediction method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the PM2.5 concentration prediction method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the PM2.5 concentration prediction method according to any one of claims 1-7.

Images