Atmospheric laser radar data analysis method based on deep learning

Abstract

The invention discloses an atmosphere laser radar data analysis method based on deep learning, and the method comprises the following steps: S1, collecting and preprocessing atmosphere laser radar echo signal data, and generating a time series data set; s2, constructing a long sequence prediction model based on Transform, and forming a preliminary prediction result sequence; s3, optimizing model structure parameters and training parameters by using a particle filtering algorithm to obtain an optimized model; s4, using the optimized model to predict current and future moment data, and generating a final prediction sequence; s5, performing residual analysis on the final prediction sequence and the real-time observation data, and identifying abnormal points; and S6, sudden disturbance is detected, particle filtering optimization is restarted, and the prediction model is dynamically updated. According to the invention, by fusing the deep learning long sequence prediction model and the particle filter optimization algorithm, high-precision prediction and abnormal dynamic identification of atmospheric laser radar data are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and environmental monitoring technology, and in particular to an atmospheric lidar data analysis method based on deep learning. Background Art

[0002] In the field of environmental monitoring and air pollution control, LiDAR (Light Detection and Ranging) has been widely used in a variety of application scenarios, including aerosol distribution monitoring, atmospheric pollution concentration assessment, and cloud structure identification, due to its high resolution, long-range detection, and real-time echo imaging capabilities. As urban air quality issues become increasingly severe, atmospheric LiDAR systems, as an important data acquisition tool, play a key role in urban environmental monitoring, ecological early warning, traffic control, and emergency response. However, the raw echo signal data obtained by LiDAR is often high-dimensional, noisy, non-stationary, and highly time-varying, making effective modeling and high-precision prediction a challenging task.

[0003] Traditional atmospheric lidar data processing methods are mostly based on inversion algorithms, empirical modeling, or statistical analysis. While methods such as the Klett inversion and Fernald algorithm can perform basic aerosol inversion and pollution concentration estimation, they exhibit significant limitations in nonlinear and highly dynamic atmospheric systems. These models lack robustness and responsiveness, particularly when faced with sudden pollution spreads, rapid meteorological changes, or complex terrain disturbances. Furthermore, most of these methods rely on single-step predictions or short-term estimates, making them difficult to support high-confidence predictions for long-term future periods and unable to meet the needs of forward-looking, dynamic decision-making in urban management and pollution control.

[0004] To improve forecasting capabilities, artificial intelligence methods, particularly neural networks, have been introduced in recent years for atmospheric data modeling. Recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) have been used to process time-series environmental data, capable of capturing temporal delays and trends to a certain extent. However, these models still suffer from limited modeling capabilities for very long sequences, low computational efficiency, and long training times. With the introduction of the Transformer model, the self-attention mechanism has demonstrated superior performance in sequence modeling compared to traditional RNN structures, ushering in a new stage of development for deep learning models in time series forecasting. Transformer-based long-sequence forecasting models, such as the Informer and Autoformer, have been successfully applied to fields such as power load forecasting and traffic flow prediction, demonstrating their powerful modeling capabilities and parallel computing advantages.

[0005] Although the Transformer architecture has significantly improved model performance in atmospheric time series processing, current research focuses primarily on innovations in the model architecture itself, with less attention paid to optimizing parameter configuration and dynamic adaptation during actual deployment. Atmospheric lidar data not only exhibits high spatial and temporal dimensionality but is also frequently subject to interference from sudden changes (such as dust storms, severe convection, and pollution events). In practical applications, statically setting model parameters such as the number of layers, sparsity, embedding structure, and regularization strength makes it difficult to achieve optimal results across diverse scenarios. Furthermore, existing optimization strategies often rely on fixed grid searches or simple gradient optimization, lacking flexibility and global search capabilities, and are unable to effectively adapt to the dynamic changes and nonlinear characteristics of lidar data.

[0006] The particle filter algorithm, a sequential Monte Carlo method for nonlinear and non-Gaussian state estimation, has recently been introduced into the fields of neural network optimization and structure search. It samples multiple individual particles to represent different parameter states and iteratively approximates the optimal solution through a combination of importance weights, resampling mechanisms, and perturbation propagation. It possesses strong global search capabilities and dynamic adaptability. Although some studies have attempted to apply particle filtering to neural network parameter search, most of these studies remain at the static hyperparameter level, lacking a closed-loop mechanism for jointly optimizing model structure parameters and training parameters, and even less so, the design of optimization paths and the integration of mechanisms tailored to the temporal characteristics of atmospheric lidar data.

[0007] Therefore, how to provide an atmospheric lidar data analysis method based on deep learning is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a deep learning-based atmospheric lidar data analysis method. This method leverages the modeling capabilities of the Transformer architecture within deep learning for processing very long time series, and combines it with a particle filter optimization strategy to achieve dynamic joint optimization of model structure and training parameters. The method describes in detail the entire process of high-precision prediction and anomaly identification using multi-source radar echo data. This method boasts high prediction accuracy, strong model adaptability, sensitive response to sudden environmental disturbances, and flexible system deployment, making it suitable for scenarios such as atmospheric pollution early warning, spatiotemporal analysis of aerosols, and environmental emergency response.

[0009] A method for analyzing atmospheric lidar data based on deep learning according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect echo signal data obtained by the atmospheric lidar system at multiple time points, and preprocess the echo signal data to generate a time series data set;

[0011] S2. Build a long sequence prediction model based on the Transformer structure, input the time series dataset into the encoder of the long sequence prediction model to form a preliminary prediction result sequence;

[0012] S3. Dynamically optimize the structural parameters and training parameters of the long sequence prediction model based on the particle filter optimization algorithm. Input each set of parameters into the long sequence prediction model and perform training and verification in combination with the time series dataset. Obtain the prediction error as the particle weight, update the particle group and iterate the optimization to obtain the optimized long sequence prediction model.

[0013] S4. Use the optimized long sequence prediction model to predict the echo signal data at the current moment and multiple moments in the future to generate a final prediction result sequence;

[0014] S5. Perform residual analysis on the final prediction result sequence and the real-time observation data, identify prediction anomalies based on the residual threshold rule, mark them as anomalies when the residual exceeds the set threshold, and output the anomaly identification results;

[0015] S6. When a sudden environmental disturbance is identified, the particle filter optimization mechanism is restarted to update the structure and parameters of the long sequence prediction model online to maintain the dynamic adaptability and prediction accuracy of the long sequence prediction model.

[0016] Optionally, the echo signal data specifically includes a multi-dimensional echo signal including time, altitude and intensity information, which is used to characterize the spatial distribution and temporal changes of targets such as aerosols, pollutants or clouds in the atmosphere.

[0017] Optionally, the preprocessing of the echo signal data specifically includes denoising, missing value filling, normalization and format unification operations on the echo signal data, which are used to improve the integrity of the time series data set and the stability of the long sequence prediction model input.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Construct a long sequence prediction model based on the Transformer structure, wherein the long sequence prediction model includes an encoder and a decoder. The encoder is composed of a multi-layer stacked sparse attention module, a position embedding module, and a feedforward network, and is used to extract high-dimensional feature representations of atmospheric lidar echo time series data;

[0020] S22. Introducing an adjustable cross-layer sparse attention compression rate parameter α into the encoder structure to control the sparsity of attention connections in each encoder layer, so that the long sequence prediction model has differentiated attention perception capabilities at different depths;

[0021] S23. Set an adjustable temperature coefficient τ in the local window attention mechanism of the encoder to adjust the response strength of the attention distribution and improve the adaptability of the long sequence prediction model when processing dynamic change patterns of different types of radars;

[0022] S24. Construct a multidimensional temporal embedding representation of the input data, including clock time, day type, season label, and solar altitude, and introduce a temporal semantic combination factor vector γ to control the fusion ratio of various temporal semantic channels in the final embedding representation;

[0023] S25. In the long sequence prediction model training process, the heterogeneous time channel suppression coefficient λ is introduced as a regularization hyperparameter to suppress the dependence on inefficient time semantic channels and prevent the long sequence prediction model from overfitting to a single time pattern.

[0024] S26. Using the above structure and the input long sequence prediction model as the core network, the generated time series data set is input into the encoder module of the long sequence prediction model to extract multi-dimensional embedding and feature representation information, and the decoder generates a preliminary prediction result sequence;

[0025] S27. The preliminary prediction result sequence is passed as the output result to the optimization step, and the long sequence prediction model structure parameters α, γ and training parameters τ, λ are used as optimization targets for the particle filter optimization algorithm to perform dynamic iterative adjustment.

[0026] Optionally, the S3 specifically includes:

[0027] S31, suppose the particle individual set of the particle filter optimization algorithm is P = {p i |i=1,2,...,N}, where each particle individual p i Represents a set of parameter combinations, including the structural parameter α i , γ i With the training parameter τ i ,λ i , which correspond to the encoder sparse compression rate vector, the time semantic combination factor vector, the attention temperature coefficient and the regularization suppression coefficient, respectively, where N is the total number of individual particles;

[0028] S32, initialize the particle individual set P, randomly generate initial values ​​for each particle within the predefined parameter range, and set the maximum number of iterations to T max ;

[0029] S33, load the parameter combination of each particle into the long sequence prediction model, combine it with the generated time series data set D for training, and record the predicted output sequence

[0030] S34, calculate the prediction error and weight of each particle, where the particle error ei Defined as:

[0031]

[0032] Particle weight w i for:

[0033]

[0034] Among them, P is the number of prediction steps, is the predicted value of the i-th particle at time t, Y t is the actual observation value, ∈ is a positive constant, μ is the gradient penalty factor, Estimate the magnitude of the gradient of the prediction results to the model parameters;

[0035] S35, perform resampling operation based on the particle weight set, select high-weight particles and eliminate low-weight particles, and generate a new generation of particle group P ′ ;

[0036] S36, the structural parameters (α i ,γ i ) performs residual direction guided perturbation, the perturbation form is:

[0037]

[0038] Among them, α′ i and γ′ i The structural parameters after perturbation update, δ i is the perturbation control factor of the i-th particle, sign( ) is the sign function, and is a normally distributed random variable, and is the variance of the normal disturbance;

[0039] S37, training parameters for each particle (τ i ,λ i ) performs uniformly distributed perturbation updates, the perturbation form is:

[0040] τ′ i =τ i +δ i ·U(-η τ ,η τ );

[0041] λ′ i =λ i +δ i ·U(-η λ ,η λ );

[0042] Among them, τ′ i and λ′ i is the training parameter after disturbance update, U(-η τ ,η τ ) and U(-n λ ,η λ ) is a uniformly distributed random variable, η τ and η λ is the disturbance range factor;

[0043] S38. Dynamic calculation of particle disturbance control factor δ i (t), jointly adjust the perturbation intensity of individual particles, and the perturbation factor is defined as:

[0044]

[0045] Among them, δ0 is the initial disturbance amplitude, γ1 is the time control coefficient, and γ2 is the error change rate adjustment coefficient. represents the error change rate, and t is the current iteration round;

[0046] S39, re-inputting the updated particle parameters into the long sequence prediction model for training and verification, calculating the updated prediction error and weight, and repeating steps S33 to S38 to iteratively update the particle population;

[0047] S310, when the global optimal error e * When the following convergence conditions are met:

[0048] e * <∈ target and

[0049] Among them, ∈ target is the preset error threshold, κ is the error stability threshold, and the parameter combination corresponding to the current optimal particle is configured to the long sequence prediction model to complete the parameter optimization process.

[0050] Optionally, the S4 specifically includes:

[0051] S41, obtaining atmospheric lidar echo signal data at the current moment, and using the atmospheric lidar echo signal data as the starting point of the latest input window, preparing for prediction input splicing;

[0052] S42, splicing the current atmospheric lidar echo signal data and the generated time series data set in chronological order to form a complete input sequence for prediction;

[0053] S43, inputting the input sequence into the optimized long sequence prediction model, and using an encoder to perform feature extraction and modeling on the time series input;

[0054] S44, inferring the echo signals of multiple future time steps through the decoder module, sequentially generating prediction outputs for each future time point, and forming a continuous prediction result sequence;

[0055] S45. Restoring the prediction output of the optimized long sequence prediction model to the same physical quantity representation as the original radar data, including atmospheric element data of pollution concentration, aerosol inversion value, and echo intensity;

[0056] S46. The generated final prediction result sequence is cached or imported into the abnormality analysis and early warning module to provide data support for real-time analysis and scheduling.

[0057] Optionally, the S5 specifically includes:

[0058] S51. Obtain the final prediction result sequence output by the optimized long sequence prediction model in the current time period as the current round of prediction benchmark data;

[0059] S52, synchronously calling the real-time observation data uploaded by the sensor device, which corresponds to the predicted sequence in the time dimension;

[0060] S53. For each time point, the predicted value is compared with the corresponding actual observed value one by one, and the residual difference value is calculated to form a complete residual sequence;

[0061] S54, judging the residual sequence point by point according to a preset residual threshold rule, and identifying the time point or segment where the error is greater than the threshold;

[0062] S55. Mark the data points identified as abnormal and record the corresponding time, location, and residual features for abnormal information tracing and analysis;

[0063] S56. Send the abnormality identification result to the system monitoring module or early warning module as a basis for triggering scheduling control, parameter update or long sequence prediction model correction.

[0064] The beneficial effects of the present invention are:

[0065] This paper deeply integrates a Transformer-based long-sequence prediction model with a particle filter optimization algorithm, addressing the technical bottlenecks of existing atmospheric lidar data analysis methods, such as weak long-series modeling capabilities, fixed model parameters, and unstable prediction accuracy. By constructing a multidimensional adjustable parameter space and designing a particle filter optimization mechanism to dynamically optimize both structural and training parameters, the model maintains high prediction accuracy and strong robustness in the face of changing atmospheric environments, complex pollution trends, and the high noise characteristics of radar data.

[0066] In terms of specific implementation, the long-sequence prediction model is responsible for extracting features from historical echo signals and predicting future trends at multiple moments. The particle filter algorithm adjusts the model configuration in real time based on prediction error feedback to ensure that the prediction structure is always optimal. By introducing structural parameters such as the cross-layer sparse compression rate and the temporal semantic combination factor, as well as training parameters such as the attention temperature coefficient and the regularization suppression coefficient, and adopting a directional perturbation and staged perturbation attenuation mechanism, the model's convergence speed and global search capabilities are effectively improved. At the same time, the proposed outlier identification mechanism based on residual trend and error stability also enhances the system's ability to perceive sudden abnormal pollution events.

[0067] The technical solution implemented in this invention can autonomously adapt to changes in data distribution and dynamically adjust its internal structure without relying on manually set rules, significantly improving the generalization and deployment adaptability of atmospheric prediction models. Compared with traditional methods, this solution not only improves the accuracy of pollution concentration predictions but also demonstrates significant advantages in model optimization efficiency, parameter update mechanisms, and anomaly warning response speed, demonstrating promising engineering application prospects and widespread application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flowchart of an atmospheric lidar data analysis method based on deep learning proposed by the present invention;

[0070] Figure 2 This is a working mechanism diagram of the particle filter optimization process of the atmospheric lidar data analysis method based on deep learning proposed in this invention. DETAILED DESCRIPTION

[0071] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0072] refer to Figure 1 and Figure 2 , a deep learning-based atmospheric lidar data analysis method, comprising the following steps:

[0073] S1. Collect echo signal data obtained by the atmospheric lidar system at multiple time points, and preprocess the echo signal data to generate a time series data set;

[0074] S2. Build a long sequence prediction model based on the Transformer structure, input the time series dataset into the encoder of the long sequence prediction model to form a preliminary prediction result sequence;

[0075] S3. Dynamically optimize the structural parameters and training parameters of the long sequence prediction model based on the particle filter optimization algorithm. Input each set of parameters into the long sequence prediction model and perform training and verification in combination with the time series dataset. Obtain the prediction error as the particle weight, update the particle group and iterate the optimization to obtain the optimized long sequence prediction model.

[0076] S4. Use the optimized long sequence prediction model to predict the echo signal data at the current moment and multiple moments in the future to generate a final prediction result sequence;

[0077] S5. Perform residual analysis on the final prediction result sequence and the real-time observation data, identify prediction anomalies based on the residual threshold rule, mark them as anomalies when the residual exceeds the set threshold, and output the anomaly identification results;

[0078] S6. When a sudden environmental disturbance is identified, the particle filter optimization mechanism is restarted to update the structure and parameters of the long sequence prediction model online to maintain the dynamic adaptability and prediction accuracy of the long sequence prediction model.

[0079] In this embodiment, the echo signal data specifically includes multi-dimensional echo signals including time, height and intensity information, which are used to characterize the spatial distribution and temporal changes of targets such as aerosols, pollutants or clouds in the atmosphere.

[0080] In this embodiment, the preprocessing of the echo signal data specifically includes denoising, missing value filling, normalization and format unification operations on the echo signal data, which are used to improve the integrity of the time series data set and the stability of the long sequence prediction model input.

[0081] In this embodiment, S2 specifically includes:

[0082] S21. Construct a long sequence prediction model based on the Transformer structure, wherein the long sequence prediction model includes an encoder and a decoder. The encoder is composed of a multi-layer stacked sparse attention module, a position embedding module, and a feedforward network, and is used to extract high-dimensional feature representations of atmospheric lidar echo time series data;

[0083] S22. Introducing an adjustable cross-layer sparse attention compression rate parameter α into the encoder structure to control the sparsity of attention connections in each encoder layer, so that the long sequence prediction model has differentiated attention perception capabilities at different depths;

[0084] S23. Set an adjustable temperature coefficient τ in the local window attention mechanism of the encoder to adjust the response strength of the attention distribution and improve the adaptability of the long sequence prediction model when processing dynamic change patterns of different types of radars;

[0085] S24. Construct a multidimensional temporal embedding representation of the input data, including clock time, day type, season label, and solar altitude, and introduce a temporal semantic combination factor vector γ to control the fusion ratio of various temporal semantic channels in the final embedding representation;

[0086] S25. In the long sequence prediction model training process, the heterogeneous time channel suppression coefficient λ is introduced as a regularization hyperparameter to suppress the dependence on inefficient time semantic channels and prevent the long sequence prediction model from overfitting to a single time pattern.

[0087] S26. Using the above structure and the input long sequence prediction model as the core network, the generated time series data set is input into the encoder module of the long sequence prediction model to extract multi-dimensional embedding and feature representation information, and the decoder generates a preliminary prediction result sequence;

[0088] S27. The preliminary prediction result sequence is passed as the output result to the optimization step, and the long sequence prediction model structure parameters α, γ and training parameters τ, λ are used as optimization targets for the particle filter optimization algorithm to perform dynamic iterative adjustment.

[0089] In this embodiment, S3 specifically includes:

[0090] S31, suppose the particle individual set of the particle filter optimization algorithm is P = {p i |i=1,2,...,N}, where each particle individual p i Represents a set of parameter combinations, including the structural parameter α i , γ i With the training parameter τ i ,λ i , which correspond to the encoder sparse compression rate vector, the time semantic combination factor vector, the attention temperature coefficient and the regularization suppression coefficient, respectively, where N is the total number of individual particles;

[0091] S32, initialize the particle individual set P, randomly generate initial values ​​for each particle within the predefined parameter range, and set the maximum number of iterations to T max ;

[0092] S33, load the parameter combination of each particle into the long sequence prediction model, combine it with the generated time series data set D for training, and record the predicted output sequence

[0093] S34, calculate the prediction error and weight of each particle, where the particle error e i Defined as:

[0094]

[0095] Particle weight w i for:

[0096]

[0097] Among them, P is the number of prediction steps, is the predicted value of the i-th particle at time t, Y t is the actual observation value, ∈ is a positive constant, μ is the gradient penalty factor, Estimate the magnitude of the gradient of the prediction results to the model parameters;

[0098] S35, perform resampling operation based on the particle weight set, select high-weight particles and eliminate low-weight particles, and generate a new generation of particle group P ′ ;

[0099] S36, the structural parameters (α i ,γ i ) performs residual direction guided perturbation, the perturbation form is:

[0100]

[0101]

[0102] Among them, α′ i and γ′ i The structural parameters after perturbation update, δ i is the perturbation control factor of the i-th particle, sign() is the sign function, and is a normally distributed random variable, and is the variance of the normal disturbance;

[0103] S37, training parameters for each particle (τ i ,λ i ) performs uniformly distributed perturbation updates, the perturbation form is:

[0104] τ′ i =τ i +δ i ·U(-η τ ,η τ );

[0105] λ′ i =λ i +δi ·U(-η λ ,η λ );

[0106] Among them, τ′ i and λ′ i is the training parameter after disturbance update, U(-η τ ,η τ ) and U(-n λ ,η λ ) is a uniformly distributed random variable, η τ and η λ is the disturbance range factor;

[0107] S38. Dynamic calculation of particle disturbance control factor δ i (t), jointly adjust the perturbation intensity of individual particles, and the perturbation factor is defined as:

[0108]

[0109] Among them, δ0 is the initial disturbance amplitude, γ1 is the time control coefficient, and γ2 is the error change rate adjustment coefficient. represents the error change rate, t is the current iteration round;

[0110] S39, re-inputting the updated particle parameters into the long sequence prediction model for training and verification, calculating the updated prediction error and weight, and repeating steps S33 to S38 to iteratively update the particle population;

[0111] S310, when the global optimal error e * When the following convergence conditions are met:

[0112] e * <∈ target and

[0113] Among them, ∈ target is the preset error threshold, κ is the error stability threshold, and the parameter combination corresponding to the current optimal particle is configured to the long sequence prediction model to complete the parameter optimization process.

[0114] In this embodiment, the S4 specifically includes:

[0115] S41, obtaining atmospheric lidar echo signal data at the current moment, and using the atmospheric lidar echo signal data as the starting point of the latest input window, preparing for prediction input splicing;

[0116] S42, splicing the current atmospheric lidar echo signal data and the generated time series data set in chronological order to form a complete input sequence for prediction;

[0117] S43, inputting the input sequence into the optimized long sequence prediction model, and using an encoder to perform feature extraction and modeling on the time series input;

[0118] S44, inferring the echo signals of multiple future time steps through the decoder module, sequentially generating prediction outputs for each future time point, and forming a continuous prediction result sequence;

[0119] S45. Restoring the prediction output of the optimized long sequence prediction model to the same physical quantity representation as the original radar data, including atmospheric element data of pollution concentration, aerosol inversion value, and echo intensity;

[0120] S46. The generated final prediction result sequence is cached or imported into the abnormality analysis and early warning module to provide data support for real-time analysis and scheduling.

[0121] In this embodiment, the S5 specifically includes:

[0122] S51. Obtain the final prediction result sequence output by the optimized long sequence prediction model in the current time period as the current round of prediction benchmark data;

[0123] S52, synchronously calling the real-time observation data uploaded by the sensor device, which corresponds to the predicted sequence in the time dimension;

[0124] S53. For each time point, the predicted value is compared with the corresponding actual observed value one by one, and the residual difference value is calculated to form a complete residual sequence;

[0125] S54, judging the residual sequence point by point according to a preset residual threshold rule, and identifying the time point or segment where the error is greater than the threshold;

[0126] S55. Mark the data points identified as abnormal and record the corresponding time, location, and residual features for abnormal information tracing and analysis;

[0127] S56. Send the abnormality identification result to the system monitoring module or early warning module as a basis for triggering scheduling control, parameter update or long sequence prediction model correction.

[0128] Example 1:

[0129] To verify the feasibility of this invention, the center applied it to a city's ecological monitoring center, continuously deploying the proposed deep learning-based atmospheric lidar data analysis method to improve pollution forecasting lead time and anomaly identification capabilities. This area, located at the urban-rural border, is frequently affected by meteorological factors such as temperature inversion, low pressure, and low wind speeds. Pollution frequently accumulates, placing extremely high demands on the real-time and accuracy of pollution warnings.

[0130] The previous prediction system, based on an LSTM architecture, provided short-term forecasts, capable of estimating pollution trends within three hours. However, it relied heavily on manual configuration of model parameters and experienced significant delays in abnormal responses during pollution emergencies. After several consecutive warning failures, the environmental protection department decided to introduce the method described in this paper for testing and verification.

[0131] During testing, the system collected multi-channel LiDAR echo data at high frequency and, in combination with a particle filter optimization strategy, dynamically updated model structure parameters (such as the attention compression ratio and the temporal embedding control factor) and training parameters (such as the regularization coefficient and the attention temperature). The model was automatically trained once a day, with a 12-hour forecast window, and updated hourly forecasts. The system also performed real-time residual analysis of anomalies against observed data.

[0132] The test period was December 1 to December 7, 2024, covering seven consecutive days of forecasting tasks under varying weather conditions and pollution intensities. Results showed that the proposed method outperformed the original system in core metrics such as daily average forecast error, anomaly identification accuracy, and early warning lead time. In particular, during the high-pollution event on the evening of December 3, the system accurately predicted sudden pollution changes three hours in advance, avoiding widespread misjudgment of pollution and significantly improving the timeliness of management and control.

[0133] The solution of the present invention successfully solves the problems of traditional models such as parameter insensitivity, error instability, and abnormality recognition delay through the dynamic optimization of long sequence model structure and the introduction of a highly robust prediction mechanism. It realizes the transformation of the pollution monitoring system from "static inference" to "intelligent adaptive prediction" and has significant engineering application value.

[0134] Table 1 Comparison of prediction performance between the method of the present invention and traditional models

[0135]

[0136] According to the analysis of the data results in Table 1, it can be clearly seen that the atmospheric lidar data analysis method based on the fusion of the long sequence prediction model and the particle filter optimization algorithm proposed in the present invention is superior to the traditional prediction model in terms of pollution concentration prediction accuracy, anomaly recognition accuracy and early warning capability.

[0137] In the actual monitoring and prediction test for 7 consecutive days, the method of the present invention showed a lower average prediction error under all levels of pollution conditions. For example, under the "severe pollution" condition on December 3, 2024, the measured PM2.5 peak was 168μg / m 3 The original model prediction error was as high as 47.8 μg / m 3 , and the method of the present invention reduces the error to 21.6μg / m 3The error reduction rate was 54.8%. This significant error compression capability was also reflected on other dates: on December 1st, under the condition of "moderate pollution", the original model error was 28.6μg / m 3 , while the model of the present invention is only 15.1μg / m 3 , the accuracy is significantly improved.

[0138] In terms of anomaly identification capabilities, the proposed method maintained a stable accuracy rate of over 89% over a seven-day test, reaching a maximum of 94.8%. This represents an improvement of over 20% compared to traditional models, demonstrating that the system can more effectively capture sudden changes in pollution levels, enhancing the timeliness and accuracy of pollution event identification. This improvement is particularly important for emergencies such as nighttime temperature inversion and rapid pollution accumulation, significantly improving the response efficiency of the regulatory platform.

[0139] Furthermore, the key indicator of "average warning lead time" in the table also demonstrates its superiority in terms of early warning response. While the original model often had a delayed response, the proposed method was able to issue warnings at the first sign of pollution trends, with an average lead time exceeding 35 minutes. On December 3rd, a day of heavy pollution, the lead time reached 65 minutes, effectively creating a window for proactive adjustments to pollution dispatch plans.

[0140] Based on the above data comparison and analysis, the present invention not only achieves superior performance in model prediction error control, but also builds a highly stable and sensitive intelligent analysis mechanism for atmospheric monitoring in terms of anomaly identification and response speed, fully demonstrating its technical practicality and engineering promotion value in actual urban air quality control scenarios.

[0141] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for analyzing atmospheric lidar data based on deep learning, characterized in that: The steps include: S1. Collect echo signal data obtained by the atmospheric lidar system at multiple time points, and preprocess the echo signal data to generate a time series data set; S2. Build a long sequence prediction model based on the Transformer structure, input the time series dataset into the encoder of the long sequence prediction model to form a preliminary prediction result sequence; S3. Dynamically optimize the structural parameters and training parameters of the long sequence prediction model based on the particle filter optimization algorithm. Input each set of parameters into the long sequence prediction model and perform training and verification in combination with the time series dataset. Obtain the prediction error as the particle weight, update the particle group and iterate the optimization to obtain the optimized long sequence prediction model. S4. Use the optimized long sequence prediction model to predict the echo signal data at the current moment and multiple moments in the future to generate a final prediction result sequence; S5. Perform residual analysis on the final prediction result sequence and the real-time observation data, identify prediction anomalies based on the residual threshold rule, mark them as anomalies when the residual exceeds the set threshold, and output the anomaly identification results; S6. When a sudden environmental disturbance is identified, the particle filter optimization mechanism is restarted to update the structure and parameters of the long sequence prediction model online to maintain the dynamic adaptability and prediction accuracy of the long sequence prediction model.

2. The atmospheric lidar data analysis method based on deep learning according to claim 1, characterized in that: The echo signal data specifically includes multi-dimensional echo signals of time, altitude and intensity information, which are used to characterize the spatial distribution and temporal changes of targets such as aerosols, pollutants or clouds in the atmosphere.

3. The atmospheric lidar data analysis method based on deep learning according to claim 1, characterized in that: The preprocessing of the echo signal data specifically includes denoising, missing value filling, normalization and format unification operations on the echo signal data, which are used to improve the integrity of the time series data set and the stability of the long sequence prediction model input.

4. The atmospheric lidar data analysis method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. Construct a long sequence prediction model based on the Transformer structure, wherein the long sequence prediction model includes an encoder and a decoder. The encoder is composed of a multi-layer stacked sparse attention module, a position embedding module, and a feedforward network, and is used to extract high-dimensional feature representations of atmospheric lidar echo time series data; S22. Introducing an adjustable cross-layer sparse attention compression rate parameter α into the encoder structure to control the sparsity of attention connections in each encoder layer, so that the long sequence prediction model has differentiated attention perception capabilities at different depths; S23. Set an adjustable temperature coefficient τ in the local window attention mechanism of the encoder to adjust the response strength of the attention distribution and improve the adaptability of the long sequence prediction model when processing dynamic change patterns of different types of radars; S24. Construct a multidimensional temporal embedding representation of the input data, including clock time, day type, season label, and solar altitude, and introduce a temporal semantic combination factor vector γ to control the fusion ratio of various temporal semantic channels in the final embedding representation; S25. In the long sequence prediction model training process, the heterogeneous time channel suppression coefficient λ is introduced as a regularization hyperparameter to suppress the dependence on inefficient time semantic channels and prevent the long sequence prediction model from overfitting to a single time pattern. S26. Using the above structure and the input long sequence prediction model as the core network, the generated time series data set is input into the encoder module of the long sequence prediction model to extract multi-dimensional embedding and feature representation information, and the decoder generates a preliminary prediction result sequence; S27. The preliminary prediction result sequence is passed as the output result to the optimization step, and the long sequence prediction model structure parameters α, γ and training parameters τ, λ are used as optimization targets for the particle filter optimization algorithm to perform dynamic iterative adjustment.

5. The atmospheric lidar data analysis method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, suppose the particle individual set of the particle filter optimization algorithm is P = {p i |i=1,2,...,N}, where each particle individual p i Represents a set of parameter combinations, including the structural parameter α i , γ i With the training parameter τ i ,λ i , which correspond to the encoder sparse compression rate vector, the time semantic combination factor vector, the attention temperature coefficient and the regularization suppression coefficient, respectively, where N is the total number of individual particles; S32, initialize the particle individual set P, randomly generate initial values ​​for each particle within the predefined parameter range, and set the maximum number of iterations to T max ; S33, load the parameter combination of each particle into the long sequence prediction model, combine it with the generated time series data set D for training, and record the predicted output sequence S34, calculate the prediction error and weight of each particle, where the particle error e i Defined as: Particle weight w i for: Among them, P is the number of prediction steps, is the predicted value of the i-th particle at time t, Y t is the actual observation value, ∈ is a positive constant, μ is the gradient penalty factor, Estimate the magnitude of the gradient of the prediction results to the model parameters; S35, perform resampling operation based on the particle weight set, select high-weight particles and eliminate low-weight particles, and generate a new generation of particle group P ′ ; S36, the structural parameters (α i ,γ i ) performs residual direction guided perturbation, the perturbation form is: Among them, α′ i and γ′ i The structural parameters after perturbation update, δ i is the perturbation control factor of the i-th particle, sign( ) is the sign function, and is a normally distributed random variable, and is the variance of the normal disturbance; S37, training parameters for each particle (τ i ,λ i ) performs uniformly distributed perturbation updates, the perturbation form is: the i =t i +d i ·U(-η τ ,or τ ); λ′ i =λ i +d i ·U(-η λ ,or λ ); Among them, τ′ i and λ′ i is the training parameter after disturbance update, U(-η τ ,η τ ) and U(-n λ ,η λ ) is a uniformly distributed random variable, η τ and η λ is the disturbance range factor; S38. Dynamic calculation of particle disturbance control factor δ i (t), jointly adjust the perturbation intensity of individual particles, and the perturbation factor is defined as: Among them, δ0 is the initial disturbance amplitude, γ1 is the time control coefficient, and γ2 is the error change rate adjustment coefficient. represents the error change rate, t is the current iteration round; S39, re-inputting the updated particle parameters into the long sequence prediction model for training and verification, calculating the updated prediction error and weight, and repeating steps S33 to S38 to iteratively update the particle population; S310, when the global optimal error e * When the following convergence conditions are met: Among them, ∈ target is the preset error threshold, κ is the error stability threshold, and the parameter combination corresponding to the current optimal particle is configured to the long sequence prediction model to complete the parameter optimization process.

6. The atmospheric lidar data analysis method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41, obtaining atmospheric lidar echo signal data at the current moment, and using the atmospheric lidar echo signal data as the starting point of the latest input window, preparing for prediction input splicing; S42, splicing the current atmospheric lidar echo signal data and the generated time series data set in chronological order to form a complete input sequence for prediction; S43, inputting the input sequence into the optimized long sequence prediction model, and using an encoder to extract features and model the time series input; S44, inferring the echo signals of multiple future time steps through the decoder module, sequentially generating prediction outputs for each future time point, and forming a continuous prediction result sequence; S45. Restoring the prediction output of the optimized long sequence prediction model to the same physical quantity representation as the original radar data, including atmospheric element data of pollution concentration, aerosol inversion value, and echo intensity; S46. The generated final prediction result sequence is cached or imported into the abnormality analysis and early warning module to provide data support for real-time analysis and scheduling.

7. The atmospheric lidar data analysis method based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. Obtain the final prediction result sequence output by the optimized long sequence prediction model in the current time period as the benchmark data for this round of prediction; S52, synchronously calling the real-time observation data uploaded by the sensor device, which corresponds to the predicted sequence in the time dimension; S53. For each time point, the predicted value is compared with the corresponding actual observed value one by one, and the residual difference value is calculated to form a complete residual sequence; S54, judging the residual sequence point by point according to a preset residual threshold rule, and identifying the time point or segment where the error is greater than the threshold; S55. Mark the data points identified as abnormal and record the corresponding time, location, and residual features for abnormal information tracing and analysis; S56. Send the abnormality identification result to the system monitoring module or early warning module as a basis for triggering scheduling control, parameter update or long sequence prediction model correction.

Images