Visibility time sequence prediction method based on DeepAR model and weather prediction field

Abstract

The invention discloses a visibility time sequence prediction method based on a DeepAR model and a weather prediction field, mainly relates to the technical field of visibility time sequence prediction, and is used for solving the problems that in an existing scheme, weather prediction field feature fusion is insufficient, only a single value is output, and weather prediction field features cannot be spliced by an existing DeepAR model. Comprising the steps of generating a comprehensive feature vector according to a feature result, historical visibility time sequence observation data and a preset additional covariable, inputting the comprehensive feature vector into a hidden layer of an initial DeepAR model, and completing construction of a future meteorological field extraction module in the DeepAR model; a DeepAR model with a future meteorological field extraction module is trained and optimized based on historical visibility time sequence observation data, and a trained DeepAR model is obtained; and performing visibility time sequence prediction by using the trained DeepAR model.

Description

Technical Field

[0001] This application relates to the field of visibility time series prediction technology, and in particular to a visibility time series prediction method based on a DeepAR model and a meteorological forecast field. Background Technology

[0002] Accurate visibility forecasting is a crucial technological support for safe aviation operations, the development of the low-altitude economy, and maritime activities. In the aviation sector, frequent low-visibility weather (such as fog and haze) is increasingly causing problems such as flight diversions, delays, and even cancellations, especially during takeoff and landing phases in airport terminal areas, where sudden drops in visibility directly threaten flight safety. According to civil aviation operating regulations, a low visibility warning must be issued when visibility drops to 1000 meters or the cloud base is below 90 meters and showing a downward trend; when visibility further drops to 800 meters, low visibility operational procedures are activated. In the maritime sector, sea fog, a weather phenomenon with atmospheric visibility below 1000 meters, is common at sea, around islands, and in target sea areas, severely impacting navigation safety and production operations. Therefore, accurate visibility forecasting is of great significance for ensuring the safety of civil aviation flights, maritime navigation, and related production activities.

[0003] The following are the core problems existing in current visibility prediction technologies: 1. Insufficient fusion of meteorological forecast field features: Traditional deep learning methods mainly rely on time-series modeling of historical observation data, which fails to effectively integrate the gridded spatial information of future high-resolution meteorological forecast fields. As a result, the two-dimensional and three-dimensional distribution features of key meteorological elements such as temperature field, humidity field, and wind field are not fully utilized, which limits the prediction accuracy.

[0004] 2. Outputting only a single numerical value: Existing technologies mostly employ deterministic models such as LSTM, GRU, and Transformer, which only output a single numerical prediction. However, the strong randomness and suddenness of low-visibility events place higher demands on probabilistic prediction.

[0005] 3. Limitations of DeepAR model application: Although DeepAR models have the ability to generate probabilities, they have not yet achieved effective information interaction with the characteristics of weather forecast fields in the field of visibility prediction, and have failed to fully realize their potential in complex weather scenarios. Summary of the Invention

[0006] This application provides a visibility time series prediction method based on DeepAR model and meteorological forecast field to solve the problems of insufficient fusion of meteorological forecast field features, outputting only a single value, and existing DeepAR model being unable to predict meteorological forecast field features in existing schemes.

[0007] Firstly, this application provides a visibility time series prediction method based on a DeepAR model and a meteorological prediction field, the method comprising: Acquire historical visibility time-series observation data and historical weather forecast field data; The historical weather forecast data is aligned in the time dimension to ensure that the forecast lead time of the historical visibility time series observation data is the same as that of the historical weather forecast data. Preprocessing of historical visibility time-series observation data and historical weather forecast data; Construct an initial DeepAR model based on the LSTM model; Spatial correlation features are extracted from the preprocessed historical weather forecast data; the historical visibility time series observation data and spatial correlation features are calculated using a multi-channel self-attention mechanism to obtain the spatial characteristics of the number of channels of each meteorological element involved in the historical weather forecast data. Using spatial features, the attention of each meteorological element channel is calculated; based on the channel attention, the spatial correlation features of each meteorological element channel are weighted and unified to obtain the feature results. Based on the feature results, historical visibility time series observation data, and preset additional covariates, a comprehensive feature vector is generated. The comprehensive feature vector is then input into the hidden layer of the initial DeepAR model to complete the construction of the future weather field extraction module in the DeepAR model. A DeepAR model with a future weather field extraction module is trained and optimized based on historical visibility time-series observation data to obtain a trained DeepAR model; the trained DeepAR model is then used to perform visibility time-series prediction.

[0008] In one implementation of this application, obtaining historical visibility time-series observation data and historical weather forecast field data specifically includes: Input to obtain historical visibility time-series observation data and historical weather forecast field data; Among them, historical weather forecast data It consists of the number of meteorological element channels C, the start time t, the forecast lead time T, and the spatial grid resolution H and W.

[0009] In one implementation of this application, historical weather forecast field data is aligned along the time dimension to ensure that the forecast lead time of historical visibility time-series observation data is the same as that of historical weather forecast field data. Specifically, this includes: Forecast lead time by acquiring historical weather forecast field data and historical visibility time series observation data; Interpolation processing is performed on historical weather forecast data in the time dimension to ensure that the forecast lead time of historical weather forecast data and historical visibility time series observation data is the same.

[0010] In one implementation of this application, the data collection scope of historical weather forecast fields is: Meteorological field data centered on the latitude and longitude of the station and within a preset distance.

[0011] In one implementation of this application, spatial correlation feature extraction is performed on the preprocessed historical weather forecast data, specifically including: Historical weather forecast field data of four-dimensional tensor Time slicing is performed to obtain the three-dimensional tensor corresponding to each time t. Where C represents the number of meteorological element channels, T represents the forecast lead time, and H and W represent the spatial grid resolution. Then, 2D convolution kernels were used to extract spatial features from the three-dimensional tensors of different meteorological element channels: To obtain the spatial correlation characteristics of the c-th meteorological element channel. ; in, For the convolution kernel targeting the c-th meteorological element channel, For GLEU activation function, Let b represent the three-dimensional tensor of the c-th meteorological element channel, and b represent the preset bias term.

[0012] In one implementation of this application, historical visibility time-series observation data and spatial correlation features are calculated using a multi-channel self-attention mechanism to obtain the spatial characteristics of the number of channels for each meteorological element involved in the historical weather forecast field data, specifically including: Spatial flattening was performed on the spatial correlation characteristics of the number of channels for each meteorological element to obtain the characteristic results; pass The Q value is calculated using historical visibility time-series observation data Z; pass: , Utilizing feature results Calculate the K and V values; in, , , This represents the preset weight matrix; Through the formula: Calculate the attention weight A; Through the formula: Calculate the spatial characteristics of the number of channels for each meteorological element.

[0013] In one implementation of this application, spatial features are utilized to calculate the attention of each meteorological element channel, specifically including: Through the formula: Calculate the attention of the current meteorological element channel. ; in, , This represents the preset parameter matrix, where O represents the spatial characteristics of the current meteorological element channel. This indicates a modified linear unit activation function. This represents the reshaping operation function.

[0014] In one implementation of this application, the spatial correlation features of each meteorological element channel are weighted and unified based on channel attention to obtain feature results, specifically including: Through the formula: Obtain the characteristic results of the c-th meteorological element channel. ; in, This indicates the current focus of the meteorological element channel. This indicates the spatial correlation characteristics of current meteorological element channels.

[0015] In one implementation of this application, a comprehensive feature vector is generated based on the feature results, historical visibility time-series observation data, and preset additional covariates, specifically including: Through the formula: Generate a comprehensive feature vector ; in, The result represents the feature; Z represents the historical visibility time-series observation data; and X represents the pre-defined additional covariate. This represents the concatenation function.

[0016] In one implementation of this application, a DeepAR model with a future weather field extraction module is trained and optimized based on historical visibility time-series observation data to obtain a trained DeepAR model, specifically including: Historical visibility time-series observation data were divided into training set and validation set; Using historical visibility time-series observation data from the training set, a DeepAR model with a future weather field extraction module is trained; the visibility variation pattern of the historical visibility time-series observation data in the training set is learned through a loss function, and the hyperparameters are adjusted using a validation set.

[0017] As can be seen from the above technical solutions, this application has the following advantages: 1. Multi-dimensional meteorological feature fusion improves forecast accuracy: By aligning historical observation data with the timeliness of meteorological forecast fields through alignment processing, and combining spatial correlation feature extraction technology, this application achieves for the first time the effective fusion of three-dimensional grid data such as temperature, humidity, and wind fields. Compared with traditional single-time-series modeling, this application utilizes a multi-channel self-attention mechanism to dynamically weight the spatial distribution characteristics of different meteorological elements, enabling the model to capture the nonlinear effects of physical processes within the atmospheric boundary layer (such as inversion formation and water vapor distribution) on visibility. This feature fusion mechanism incorporates the two-dimensional horizontal distribution characteristics of the meteorological forecast field with vertical profile information into the calculation, improving the sensitivity of capturing low-visibility events.

[0018] 2. Probabilistic output enhances decision reliability: Leveraging the probabilistic generation characteristics of the DeepAR model, this application can output a probability distribution range for visibility predictions instead of a single, definitive value. This feature overcomes the prediction blind spot of traditional LSTM / Transformer models in sudden low visibility events, providing a quantitative basis for risk assessment in applications such as traffic control and flight scheduling. By jointly modeling the spatial characteristics of the meteorological field with historical observation data, the model can dynamically assess the probability of visibility deterioration under different combinations of meteorological elements. For example, when a temperature inversion structure appears and the humidity field is saturated, the system will automatically generate a high-probability warning range, compensating for the "either / or" decision-making deficiency of deterministic models.

[0019] 3. Improve the generalization ability of the meteorological field interaction module: The innovative future weather field extraction module utilizes a channel attention mechanism to dynamically filter features, overcoming the limitation of the original DeepAR model, which only processes time-series data. This module's differentiated weighting of meteorological element channels (e.g., assigning higher weight to near-surface wind fields) enables the model to adapt to the prediction needs of different geographical environments (e.g., foggy coastal areas, dusty inland areas). Furthermore, the introduction of pre-defined additional covariates allows for the integration of external factors, improving the model's generalization ability. Attached Figure Description

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

[0021] Figure 1 This is a flowchart of a visibility time series prediction method based on a DeepAR model and a meteorological forecast field provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0024] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0025] Accurate visibility forecasting is a crucial technological support for safe aviation operations, the development of the low-altitude economy, and maritime activities. In the aviation sector, frequent low-visibility weather (such as fog and haze) is increasingly causing problems such as flight diversions, delays, and even cancellations, especially during takeoff and landing phases in airport terminal areas, where sudden drops in visibility directly threaten flight safety. According to civil aviation operating regulations, a low visibility warning must be issued when visibility drops to 1000 meters or the cloud base is below 90 meters and showing a downward trend; when visibility further drops to 800 meters, low visibility operational procedures are activated. In the maritime sector, sea fog, a weather phenomenon with atmospheric visibility below 1000 meters, is common at sea, around islands, and in target sea areas, severely impacting navigation safety and production operations. Therefore, accurate visibility forecasting is of great significance for ensuring the safety of civil aviation flights, maritime navigation, and related production activities.

[0026] This application proposes a visibility time series prediction method based on a DeepAR model and meteorological forecast fields. This method, by deeply mining the temporal patterns of historical observation data, innovatively integrates high-resolution meteorological forecast field data. Through constructing a synergistic mechanism between meteorological field convolutional features and visibility probability prediction, it fully explores the impact of the spatial distribution characteristics of multiple factors such as temperature, humidity, and wind field on visibility, thereby improving the prediction accuracy and timeliness of low-visibility weather and providing more reliable safety guarantees for aviation, marine, and other fields.

[0027] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0028] The embodiment provides a visibility time series prediction method based on a DeepAR model and a meteorological forecast field, such as... Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps: Step 110: Obtain historical visibility time-series observation data and historical weather forecast field data; perform time dimension alignment processing on the historical weather forecast field data to ensure that the forecast lead time of the historical visibility time-series observation data and the historical weather forecast field data is the same; perform preprocessing on the historical visibility time-series observation data and the historical weather forecast field data.

[0029] It should be noted that the dates covered by the historical visibility time-series observation data are the same as those covered by the historical weather forecast field data. The historical weather forecast field data is the forecast data for the historical date prior to the historical event. "Prior to the historical event" can be the day before the event.

[0030] In some embodiments, acquiring historical visibility time-series observation data and historical weather forecast field data specifically includes: Input to obtain historical visibility time-series observation data and historical weather forecast field data; Among them, historical weather forecast data It consists of the number of meteorological element channels C, the start time t, the forecast lead time T, and the spatial grid resolution H and W.

[0031] It should be noted that the input data includes historical visibility time-series observation data. Compared with historical weather forecast data (C represents the number of meteorological element channels, such as temperature, humidity, sea level pressure, wind speed, etc.; t represents the start time; T represents the forecast lead time, such as 00, 03, 06...21), H and W represent the spatial grid resolution. Since the objective of this application is to provide visibility forecast services with both spatiotemporal accuracy and reliability, a high temporal resolution meteorological forecast field is essential. Currently, mainstream meteorological forecast data, such as ECMWF, provides forecast data at 3-hour intervals, requiring temporal interpolation processing, such as linear interpolation or cubic spline interpolation.

[0032] This includes aligning historical weather forecast data along the time dimension to ensure that historical visibility time-series observation data and historical weather forecast data have the same forecast lead time. Specifically, this includes: Forecast lead time by acquiring historical weather forecast field data and historical visibility time series observation data; Interpolation processing is performed on historical weather forecast data in the time dimension to ensure that the forecast lead time of historical weather forecast data and historical visibility time series observation data is the same.

[0033] It should be noted that the historical visibility time series data corresponding to the start time of the meteorological field data field are matched one by one. In order to meet the timeliness requirements of meteorological forecast field data acquisition, for example, the historical visibility data is from 12:00 on January 1, 2025 to 11:00 on January 2, 2025 (1-hour interval). The meteorological field data is selected from the 12, 18, 24, and 30 times of the start time of January 1, 2025, as well as other time-series interpolated multi-channel meteorological field data (interpolated data with 1-hour intervals), so as to achieve successful acquisition of meteorological forecast field data and time alignment between sample data and historical time series data.

[0034] The data collection scope for historical weather forecast fields is as follows: Meteorological field data centered on the latitude and longitude of the station and within a preset distance.

[0035] It should be noted that for the latitude and longitude points of the station data, meteorological field data within a radius of approximately 300km centered on the station's latitude and longitude can be selected. Since feature extraction from meteorological field data takes time, the number of meteorological field grids directly affects the model's running time and efficiency.

[0036] This includes extracting spatial correlation features from the preprocessed historical weather forecast data, specifically: Historical weather forecast field data of four-dimensional tensor Time slicing is performed to obtain the three-dimensional tensor corresponding to each time t. Where C represents the number of meteorological element channels, T represents the forecast lead time, and H and W represent the spatial grid resolution. Then, 2D convolution kernels were used to extract spatial features from the three-dimensional tensors of different meteorological element channels: To obtain the spatial correlation characteristics of the c-th meteorological element channel. ; in, For the convolution kernel targeting the c-th meteorological element channel, For GLEU activation function, Let b represent the three-dimensional tensor of the c-th meteorological element channel, and b represent the preset bias term.

[0037] It should be noted that the above describes the extraction of spatial correlation features of the meteorological field using a 3D convolutional neural network. First, the future weather forecast field is defined as a four-dimensional tensor. ,in This is a weather forecast field data set at a specific forecast time. T represents the forecast time step, totaling 24 (only one step at a given time is used in the calculation), C represents the number of meteorological element channels (e.g., temperature, humidity, wind speed, sea level pressure, etc.), and H and W represent the spatial grid resolution. The weather forecast field data is a multi-time-step forecast field, with time slices processed independently; that is, each time step t corresponds to an independent three-dimensional tensor. Then, 2D convolution kernels are used to extract spatial features of meteorological elements from different channels in the 3D weather forecast field.

[0038] Based on the above description, this step first adjusts the 3-hour interval of the weather forecast data to a 1-hour interval through time-dimensional interpolation (such as linear interpolation or cubic spline interpolation), achieving time scale matching with historical visibility observation data. This processing method ensures the temporal consistency of the input data and avoids prediction bias caused by time resolution mismatch. Simultaneously, selecting meteorological field data within a 300km radius centered on the station satisfies both the spatial correlation requirement and avoids the redundant data problem caused by an excessively large computational range.

[0039] Secondly, a 3D convolutional neural network is used to extract features from the four-dimensional weather forecast field data. Each forecast time is independently processed into a three-dimensional tensor (C×H×W) through time slicing, and then spatial features are extracted channel-by-channel using 2D convolutional kernels. This method can effectively capture the spatial correlation characteristics of different meteorological elements (such as temperature, humidity, and wind speed), providing more discriminative input features for visibility prediction. The application of the GLEU activation function further enhances the expressive power of nonlinear features.

[0040] Finally, computational efficiency was improved while maintaining prediction accuracy. Time slicing and channel-specific processing reduced the data dimensionality of individual computational units, and the preset grid range control (300km) effectively balanced feature completeness with computational resource consumption. The entire solution was implemented based on existing standard meteorological data such as ECMWF in the operational system, and could be deployed without modifying observation equipment, demonstrating high engineering feasibility.

[0041] Step 120: Construct the initial DeepAR model based on the LSTM model.

[0042] Here, the hidden layer state in the DeepAR model is used to extract temporal features using an LSTM model, and the current hidden layer state is used as the input data for obtaining the next hidden layer state.

[0043] Understandably, this step utilizes the memory unit characteristics of recurrent neural networks to dynamically capture the temporal features of visibility, based on the DeepAR model constructed using LSTM. The model uses the current hidden layer state as the input for the next time step, forming a state propagation mechanism in the time dimension. This structure can automatically learn long-term dependencies in visibility data. It retains probabilistic prediction capabilities and can output the probability distribution of future visibility.

[0044] Step 130: Extract spatial correlation features from the preprocessed historical weather forecast data; calculate the spatial number of channels for each meteorological element involved in the historical weather forecast data by using a multi-channel self-attention mechanism with the historical visibility time series observation data and spatial correlation features.

[0045] Specifically, a multi-channel self-attention mechanism is used to calculate the spatial characteristics of the number of channels for each meteorological element involved in historical weather forecast data by combining historical visibility time-series observation data with spatial correlation features. This includes: Spatial flattening was performed on the spatial correlation characteristics of the number of channels for each meteorological element to obtain the characteristic results. ; pass The Q value is calculated using historical visibility time-series observation data Z; pass: , Utilizing feature results Calculate the K and V values; in, , , This represents the preset weight matrix; Through the formula: Calculate the attention weight A; Through the formula: Calculate the spatial characteristics of the number of channels for each meteorological element.

[0046] Based on the above description, this step, in terms of feature fusion, achieves intelligent association between historical observation data and meteorological elements through a multi-channel self-attention mechanism: First, the flattened spatial features are used as key-value pairs (K / V) and attention is calculated with the query vector (Q) generated based on visibility observation data. This design enables the model to automatically identify the differences in the impact of different meteorological elements (such as temperature, humidity, wind speed, etc.) on visibility. The application of a preset weight matrix ensures the standardization of feature interaction and avoids the subjective bias of traditional manual weighting. Experiments show that this method can effectively capture complex patterns such as the nonlinear relationship between pressure gradient and visibility. At the computational optimization level, the spatial flattening operation reduces the dimensionality of the three-dimensional meteorological field (C×H×W) to two-dimensional features, thereby reducing the memory footprint of subsequent attention calculations. Step 140: Utilize spatial features to calculate the attention of each meteorological element channel; weight and unify the spatial correlation features of each meteorological element channel based on the channel attention to obtain feature results; generate a comprehensive feature vector based on the feature results, historical visibility time series observation data, and preset additional covariates; input the comprehensive feature vector into the hidden layer of the initial DeepAR model to complete the construction of the future meteorological field extraction module in the DeepAR model.

[0047] In some embodiments, spatial features are utilized to calculate the attention of each meteorological element channel, specifically including: Through the formula: Calculate the attention of the current meteorological element channel. ; in, , This represents the preset parameter matrix, where O represents the spatial characteristics of the current meteorological element channel. This indicates a modified linear unit activation function. This represents the reshaping operation function.

[0048] Specifically, the spatial correlation features of each meteorological element channel are weighted and unified based on channel attention to obtain feature results, including: Through the formula: Obtain the characteristic results of the c-th meteorological element channel. ; in, This indicates the current focus of the meteorological element channel. This indicates the spatial correlation characteristics of current meteorological element channels.

[0049] Based on the feature results, historical visibility time-series observation data, and preset additional covariates, a comprehensive feature vector is generated, which specifically includes: Through the formula: Generate a comprehensive feature vector ; in, The result represents the feature; Z represents the historical visibility time-series observation data; and X represents the pre-defined additional covariate. This represents the concatenation function.

[0050] Based on the above description, this paper optimizes the feature representation capability of meteorological elements without using channel attention mechanisms and feature fusion. First, attention weights are calculated separately for each meteorological element channel, enabling the model to automatically identify the differences in importance of different meteorological elements to visibility prediction. Through the ReLU activation function and linear transformation of the parameter matrix, this mechanism effectively captures the saliency of features in each channel, avoiding the subjectivity of manually setting weights. This channel-specific processing method preserves the independence of meteorological elements while achieving feature selection through attention weighting.

[0051] Secondly, the unified feature weighting process enhances the representation capability of key meteorological elements. Multiplying the attention weights by the original spatial features effectively achieves adaptive filtering at the feature level, suppressing the influence of noise channels and highlighting meteorological factors that substantially contribute to visibility prediction. This approach is particularly suitable for multi-source meteorological data fusion scenarios because different meteorological elements (such as temperature, humidity, and wind speed) naturally have varying degrees of influence on visibility, and a unified weighting strategy cannot achieve optimal feature representation.

[0052] Finally, the construction of the comprehensive feature vector perfected the fusion of spatiotemporal features. By splicing meteorological field features, historical observation data, and covariates, a unified representation containing multi-dimensional information was formed, providing a more comprehensive input for the subsequent DeepAR model. This fusion method not only preserves the original characteristics of various types of data but also achieves automatic feature combination and optimization through neural networks. Particularly noteworthy is the introduction of pre-defined covariates, which expands the model's adaptability, enabling it to be compatible with other influencing factors besides meteorological data (such as geographical features or pollution data), enhancing the practical application flexibility of the solution.

[0053] Step 150: Train and optimize a DeepAR model with a future weather field extraction module based on historical visibility time series observation data to obtain a trained DeepAR model; use the trained DeepAR model to perform visibility time series prediction.

[0054] A DeepAR model with a future weather field extraction module is trained and optimized based on historical visibility time-series observation data to obtain a well-trained DeepAR model, specifically including: Historical visibility time-series observation data were divided into training set and validation set; Using historical visibility time-series observation data from the training set, a DeepAR model with a future weather field extraction module is trained; the visibility variation pattern of the historical visibility time-series observation data in the training set is learned through a loss function, and the hyperparameters are adjusted using a validation set.

[0055] Based on the above description, this step optimizes the visibility prediction model through a structured training process. A clear division between training and validation sets ensures the controllability of the model learning process: the training set captures the temporal variation patterns of visibility, while the validation set is responsible for hyperparameter tuning. This dual-track training mechanism effectively avoids model overfitting. During training, direct optimization of the loss function enables the model to accurately reflect the nonlinear relationship between visibility and meteorological elements, while the integration of the future meteorological field extraction module ensures feature alignment between weather forecast data and observational data, forming a spatiotemporally consistent modeling foundation. The entire training process does not rely on human experience intervention and is entirely data-driven in parameter optimization.

[0056] The model demonstrates two core advantages in its application phase: First, it inherits the inherent advantages of DeepAR models in processing time-series data, automatically learning the periodic and trend characteristics of visibility data. Second, through a pre-built meteorological field extraction module, it achieves end-to-end fusion of meteorological forecast data, avoiding the cumbersome process of manual feature engineering in traditional methods. Predictions are made by directly calling the trained model parameters, maintaining computational efficiency while ensuring the stability of prediction results. In practical deployment, this method only requires periodic updates of observation data for retraining, without modifying the model architecture, exhibiting good operational sustainability.

[0057] Furthermore, this application can also evaluate the model. Specifically, while dividing the training and validation sets, a test set is also created. The results of the trained and optimized DeepAR model are evaluated based on the test set. Evaluation metrics include mean squared error, root mean square error, mean absolute error, and visibility prediction results within different visibility intervals. Since the DeepAR model is a probabilistic model, the results can be evaluated using the expected value minus the standard deviation.

[0058] As described above, this embodiment aligns historical observation data with the timeliness of meteorological forecast fields through alignment processing. Combined with spatial correlation feature extraction technology, it achieves the effective fusion of three-dimensional grid data such as temperature, humidity, and wind fields for the first time. Compared with traditional single-time-series modeling, this application utilizes a multi-channel self-attention mechanism to dynamically weight the spatial distribution characteristics of different meteorological elements, enabling the model to capture the nonlinear effects of physical processes within the atmospheric boundary layer (such as inversion formation and water vapor distribution) on visibility. This feature fusion mechanism incorporates the two-dimensional horizontal distribution characteristics of the meteorological forecast field with vertical profile information into the calculation, improving the sensitivity of capturing low-visibility events.

[0059] Leveraging the probabilistic generation characteristics of the DeepAR model, this application can output a probability distribution range for visibility predictions instead of a single, definitive value. This feature overcomes the prediction blind spot of traditional LSTM / Transformer models in sudden low visibility events, providing a quantitative basis for risk assessment in applications such as traffic control and flight scheduling. By jointly modeling the spatial characteristics of the meteorological field with historical observation data, the model can dynamically assess the probability of visibility deterioration under different combinations of meteorological elements. For example, when a temperature inversion structure appears and the humidity field is saturated, the system will automatically generate a high-probability warning range, compensating for the "either / or" decision-making deficiency of deterministic models.

[0060] The innovative future weather field extraction module utilizes a channel attention mechanism to dynamically filter features, overcoming the limitation of the original DeepAR model, which only processes time-series data. This module's differentiated weighting of meteorological element channels (e.g., assigning higher weight to near-surface wind fields) enables the model to adapt to the prediction needs of different geographical environments (e.g., foggy coastal areas, dusty inland areas). Furthermore, the introduction of pre-defined additional covariates allows for the integration of external factors, improving the model's generalization ability.

[0061] This application improves the reliability, timeliness, and scene adaptability of low visibility forecasts by deeply fusing spatial feature extraction from high-resolution weather forecast fields with a DeepAR probabilistic temporal prediction framework. Compared to existing technologies, its advantages are reflected in the following aspects: 1. By constructing a three-dimensional convolutional feature extraction network for meteorological fields, we can overcome the limitations of traditional time-series models in representing the spatial correlation of gridded meteorological elements. This allows us to accurately capture the two-dimensional and three-dimensional distribution characteristics of elements such as sea level pressure, temperature field, humidity field, and wind field, and deeply analyze their triggering mechanism for visibility attenuation, thereby significantly improving prediction accuracy.

[0062] 2. By embedding the spatiotemporal characteristics of meteorological fields into the latent variable generation process of the DeepAR model, dynamic modeling of the visibility probability distribution is achieved, effectively quantifying the prediction uncertainty caused by the spatial heterogeneity of meteorological elements, and providing a reliable risk assessment for the randomness and suddenness of low visibility events.

[0063] 3. A cross-modal feature interaction mechanism is introduced to adaptively analyze the contribution weights of different meteorological field features to visibility prediction, which significantly enhances the model's generalization ability in complex geographical environments (such as oceans and mountains) and under sudden meteorological conditions, and improves the robustness of prediction.

[0064] This application provides technical support with both spatiotemporal resolution and risk quantification capabilities for high-precision application scenarios such as aviation safety early warning, low-altitude flight route planning, and marine navigation. It effectively enhances the practical value of low visibility prediction and provides important support for aviation safety assurance and low-altitude economic development.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A visibility time series prediction method based on a DeepAR model and a meteorological forecast field, characterized in that, The method includes: Acquire historical visibility time-series observation data and historical weather forecast field data; perform time-dimensional alignment processing on the historical weather forecast field data to ensure that the forecast lead time of the historical visibility time-series observation data and the historical weather forecast field data is the same; perform preprocessing on the historical visibility time-series observation data and the historical weather forecast field data; Construct an initial DeepAR model based on the LSTM model; Spatial correlation features are extracted from the preprocessed historical weather forecast data; the historical visibility time series observation data and spatial correlation features are calculated using a multi-channel self-attention mechanism to obtain the spatial characteristics of the number of channels of each meteorological element involved in the historical weather forecast data. Using spatial features, the attention of each meteorological element channel is calculated; the spatial correlation features of each meteorological element channel are weighted and unified according to the channel attention to obtain the feature results; based on the feature results, historical visibility time series observation data and preset additional covariates, a comprehensive feature vector is generated, and the comprehensive feature vector is input into the hidden layer of the initial DeepAR model to complete the construction of the future meteorological field extraction module in the DeepAR model. A DeepAR model with a future weather field extraction module is trained and optimized based on historical visibility time-series observation data to obtain a trained DeepAR model; the trained DeepAR model is then used to perform visibility time-series prediction.

2. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, Acquire historical visibility time-series observation data and historical weather forecast field data, specifically including: Input to obtain historical visibility time-series observation data and historical weather forecast field data; Among them, historical weather forecast data It consists of the number of meteorological element channels C, the start time t, the forecast lead time T, and the spatial grid resolution H and W.

3. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, Historical weather forecast data undergoes time-dimensional alignment processing to ensure that historical visibility time-series observation data and historical weather forecast data have the same forecast lead time. Specifically, this includes: Forecast lead time by acquiring historical weather forecast field data and historical visibility time series observation data; Interpolation processing is performed on historical weather forecast data in the time dimension to ensure that the forecast lead time of historical weather forecast data and historical visibility time series observation data is the same.

4. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, The data collection scope for historical weather forecast fields is as follows: Meteorological field data centered on the latitude and longitude of the station and within a preset distance.

5. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, Spatial correlation feature extraction is performed on the preprocessed historical weather forecast data, specifically including: Historical weather forecast field data of four-dimensional tensor Time slicing is performed to obtain the three-dimensional tensor corresponding to each time t. Where C represents the number of meteorological element channels, T represents the forecast lead time, and H and W represent the spatial grid resolution. Then, 2D convolution kernels were used to extract spatial features from the three-dimensional tensors of different meteorological element channels: To obtain the spatial correlation characteristics of the c-th meteorological element channel. ; in, For the convolution kernel targeting the c-th meteorological element channel, For GLEU activation function, Let b represent the three-dimensional tensor of the c-th meteorological element channel, and b represent the preset bias term.

6. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, By applying a multi-channel self-attention mechanism to historical visibility time-series observation data and spatial correlation characteristics, the spatial characteristics of the number of channels for each meteorological element involved in historical weather forecast data are obtained, specifically including: Spatial flattening was performed on the spatial correlation characteristics of the number of channels for each meteorological element to obtain the characteristic results; pass The Q value is calculated using historical visibility time-series observation data Z; pass: , Utilizing feature results Calculate the K and V values; in, , , This represents the preset weight matrix; Through the formula: Calculate the attention weight A; Through the formula: Calculate the spatial characteristics of the number of channels for each meteorological element.

7. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, Utilizing spatial characteristics, the attention of each meteorological element channel is calculated, specifically including: Through the formula: Calculate the attention of the current meteorological element channel. ; in, , This represents the preset parameter matrix, where O represents the spatial characteristics of the current meteorological element channel. This indicates a modified linear unit activation function. This represents the reshaping operation function.

8. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, Based on channel attention, the spatial correlation characteristics of each meteorological element channel are weighted and unified to obtain the feature results, specifically including: Through the formula: Obtain the characteristic results of the c-th meteorological element channel. ; in, This indicates the current focus of the meteorological element channel. This indicates the spatial correlation characteristics of current meteorological element channels.

9. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, Based on the feature results, historical visibility time-series observation data, and preset additional covariates, a comprehensive feature vector is generated, which specifically includes: Through the formula: Generate a comprehensive feature vector ; in, The result represents the feature; Z represents the historical visibility time-series observation data; and X represents the pre-defined additional covariate. This represents the concatenation function.

10. The visibility time series prediction method based on DeepAR model and meteorological forecast field according to claim 1, characterized in that, A DeepAR model with a future weather field extraction module is trained and optimized based on historical visibility time-series observation data to obtain a well-trained DeepAR model, specifically including: Historical visibility time-series observation data were divided into training set and validation set; Using historical visibility time-series observation data from the training set, a DeepAR model with a future weather field extraction module is trained; the visibility variation pattern of the historical visibility time-series observation data in the training set is learned through a loss function, and the hyperparameters are adjusted using a validation set.

Images